Integrated Intelligent Green Scheduling of Predictive Maintenance for Complex Equipment based on Information Services

As an important link to guarantee normal industrial production, equipment maintenance plays an increasingly key role in enhancing the competitiveness of enterprises and supporting green smart manufacturing. This paper aims to promote the implementation of predictive maintenance for complex equipment and improve the green performance of the maintenance service process. A structural framework of information sharing and service network is introduced to build a ubiquitous state data awareness environment for predictive maintenance service. Subsequently, an integrated mathematical problem model that consists of carbon emission objective and extended maintenance cost objective is constructed. Then an improved NSGA-II algorithm is utilized to solve this complicated two-objective optimization problem. In response to deal with the uncertainties of maintenance service environment and inaccuracy of prediction, a data-driven dynamic adjustment strategy is applied. A grinding roll fault case of a large vertical is used to demonstrate the effectiveness of this proposed approach.


I. INTRODUCTION
The complex equipment with characteristics, such as complex product structure, a large number of components, multidisciplinary technologies and long product life cycle, is usually the core of industrial manufacturing process. As the competition among enterprises is becoming cumulatively fierce caused by external market environment changes including globalization and individualized customer needs, timely and effective complex equipment maintenance security is increasingly valued by enterprises. Meanwhile, confronted with mounting energy and environmental pressure, governments around the world have increased their focus on green growth and low carbon development of the industrial sector. Driven by increasing competitive pressures and stricter The associate editor coordinating the review of this manuscript and approving it for publication was Guanjun Liu . environmental regulations, green manufacturing has been the inevitable trend [1]. Green manufacturing requires to take measures to minimize negative environmental impacts (such as carbon emissions) in three related levels of manufacturing enterprises, the product, the process and the system [2]- [4]. Equipment maintenance is an important link to guarantee the normal industrial production process and also a potential optimized object of process planning for green manufacturing as well. By providing scientific and timely maintenance, the equipment state can be guaranteed and energy efficiency can be improved. By avoiding non-value added and redundancy maintenance operations, it can reduce the consumption of resources such as spare parts and consumables. By combining maintenance planning with green manufacturing optimization objectives in different industrial scenarios, resource and energy consumption, and pollutant emission can be further optimized. VOLUME 8, 2020 This work is licensed under a Creative Commons Attribution 4.0 License. For more information, see https://creativecommons.org/licenses/by/4.0/ In the field of equipment maintenance, predictive maintenance is gradually replacing the passive maintenance and preventive maintenance [5], [6]. With the help of modern technologies like sensors, communication networks, and data mining, equipment current health state and the predictive information for a period of future time can be obtained to support more scientific and more flexible maintenance planning [7]. A lot of research has been done on key support technology and application technology of predictive maintenance which include equipment condition monitoring and signal processing technologies [8]- [12], equipment fault diagnosis and prediction technologies [13]- [15] and predictive maintenance scheduling optimization technologies [6], [16], [17]. In the field of maintenance scheduling, the maintenance opportunity and the maintenance resource scheduling scheme are often considered as two different optimization problems in [18]- [20]. These researches neglected the development of detailed maintenance service resource scheduling solutions when choosing the appropriate maintenance service execution opportunity. The selected maintenance time node will be inapplicable if constraints of maintenance service resources are not considered. On the theme of green manufacturing, some valuable research results have also demonstrated the positive impact of predictive maintenance. Rødseth et al. developed the profit loss indicator as a key performance criterion to measure green manufacturing state and demonstrated that the data-driven predictive maintenance strategy has a positive impact on green manufacturing by a case study of the sawmill industry [5], [21]. Yao et al. proposed a joint maintenance and energy management method and the saving of power cost is verified in an automotive assembly line [22]. Reasonable maintenance strategies can be implemented to reduce the hydraulic oil consumption of machine tools based on monitoring and prediction [23]. In the semiconductor industry, predictive maintenance strategies for lenses can be utilized to reduce waste of energy and materials and hence increase the sustainability [24]. However, the existing research focuses more on the indirect reduction of energy consumption and material consumption by planning appropriate maintenance time. The optimization of energy consumption and carbon emission caused by the maintenance scheduling and service process is ignored. For complex equipment that involve multidisciplinary technologies and various types of components from different sources, maintenance services for many fault types require long maintenance cycle and multiple resource scheduling and utilization, the optimization of energy consumption and carbon emission of maintenance service process has important significance for promoting green manufacturing. There is a research gap in integrated optimization method for maintenance opportunity and maintenance resource scheduling scheme of predictive maintenance for complex equipment under the background of green manufacturing.
Currently, some factors limit the successful implementation proportion of predictive maintenance for complex equipment [25], [26]. To monitor and analyze all potential fault types of the target complex equipment will bring tremendous cost pressure and technical pressure for individual user enterprises. Moreover, acquiring every possible fault mode from one asset is very unlikely, and thus the fault sample data of a single enterprise is often inadequate and the prediction accuracy is limited. Meanwhile, a dynamic production environment also requires improving the adaptability of predictive maintenance decision-making. Hence, timely and reliable environment information of maintenance scheduling are needed. Fortunately, the new generation of information technology trigger changes in many related fields include predictive maintenance scheduling [27]. Equipment and available maintenance resources state data can be timely and effectively collected with the help of technologies like wireless sensor networks, internet of things, digital twin, and cyber-physical systems and data mining. Moreover, manufacturing enterprises, user enterprises and service enterprises with different spatial distribution are organized into a service network. It supports information sharing of the fault knowledge and collaborative predictive maintenance service in a larger schedulable range. Under this background, based on timely and adequate information, detailed predictive maintenance plan can be further optimized, green indexes of the scheduling process can be measured. The popularization and application predictive maintenance can be promoted as well.
Based on above analysis, some related technical/ methodological contributions have been carried out in this paper, which are summarized as follows.
• A structural framework of information sharing and service network for predictive maintenance scheduling of complex equipment is discussed. Compared with the traditional way of implementing predictive maintenance, this framework is designed to support the efficient organization and flexible configuration of dispersed user enterprises and service resource suppliers, and conduct the environment for the integration and comprehensive analysis of data. The major advantages are described as follows: (1) The interconnection of user enterprises will break the information barrier, provide the opportunity of peer-to-peer evaluation, and achieve the accumulation of fault and prediction data. Based on the shared information resource, the accuracy of fault diagnosis and prediction information will be improved for these enterprises, and the technical difficulty will be reduced; (2) The construction of a service network that includes user enterprises and maintenance service resource suppliers from different spatial distribution will promotes the sharing of maintenance resources, and expands the schedulable range. With the help of effective collaborative management and allocation methods, the efficiency and the quality of maintenance services are guaranteed and improved, inventory cost pressure and technical pressure of user enterprises are further alleviated as well; (3) Based on the information perception of the whole process of resource scheduling, carbon emission prediction of maintenance scheduling process is supported. 45798 VOLUME 8, 2020 These positive effects will be a huge driving force for the implementation of predictive maintenance with low carbon emissions in enterprises.
• An integrated intelligent green scheduling optimization method of predictive maintenance for complex equipment is proposed. A mathematical problem description model under the ubiquitous state data awareness environment is constructed for the optimization solution of predictive maintenance plans. In order to improve the feasibility, the maintenance opportunity and the maintenance resource scheduling scheme are combined in the optimization object. Two optimization objectives are provided. The global maintenance cost is the first one and the added additional economic benefits obtained by improving the efficiency and quality of subsequent production tasks are considered in it. Carbon emissions during maintenance resource scheduling process is the second one. An improved NSGA-II algorithm is put forward to solve this complicated two-objective optimization problem. A data-driven dynamic adjustment strategy for predictive maintenance decisions is utilized to respond to uncertainties of maintenance service environment and inaccuracy of prediction. Driven by realtime data, the maintenance service plan which includes necessary maintenance resources and the appropriate maintenance service time node is updated at the right time based on this strategy. Through the data-driven dynamic adaptive adjustment strategy, the joint optimization solution of the maintenance opportunity and the scheduling scheme of shared maintenance resources, and the optimization objective function which more matches the actual circumstances, accuracy, performability, and flexibility of predictive maintenance decision can be improved. The performance potential of green manufacturing can be further tapped at the process level as well.
• A case problem is solved by applying the proposed method. The better adaptability of predictive maintenance decisions under this framework to various uncertainty of maintenance service environment is verified. The remainder of this paper is organized as follows. Section 2 introduces the structural framework of information sharing and service network. Section 3 constructs the integrated mathematical problem description model for predictive maintenance scheduling, and an improved NSGA-II and a data-driven dynamic adaptive adjustment strategy are utilized to solve the final predictive maintenance plan. Then, the case analysis is illustrated in Section 4. Finally, Section 5 concludes this paper.

II. AN STRUCTURAL FRAMEWORK OF INFORMATION SHARING AND SERVICE NETWORK FOR PREDICTIVE MAINTENANCE
The corresponding maintenance scheduling and service process is complicated since complex equipment involves multidisciplinary technologies and various types of components from different sources [28]. In this section, a structural framework of information sharing and service network for target equipment and distributed maintenance resources is introduced to create a ubiquitous maintenance decision and scheduling environment, and types of data need to be collected which impact maintenance service performance and corresponding carbon emissions is described. This framework supports a more accurate prediction of performance for predictive maintenance plans based on the real-time interactive capability between the cyberspace layer and the physical space layer. Meanwhile, it also promotes the interactive sharing of data across organizational boundaries and supports the efficient organization and flexible configuration of dispersed user enterprises and service resource suppliers.
The proposed framework is conducted from three function layers, i.e., the physical space function layer, local function layer of the cyberspace, and cloud function layer of the cyberspace, as shown in Fig.1.

A. THE PHYSICAL SPACE FUNCTION LAYER
In the physical space function layer, state awareness of target equipment objects and related maintenance service resource objects are independently completed by the corresponding object owners. With the wave of automation, digitalization and intellectualization upgrading in recent years, some data collection and analysis functions (such as processing parameter acquisition and product quality parameter acquisition) have been implemented by service suppliers and manufacturing enterprises. By reusing and enhancing these functions, the physical space function layer can be constructed. Data acquisition objects include target equipment, production environment of target equipment, necessary maintenance resources (such as maintenance service tools, technicians, consumable and components, etc.), and scheduling environment of maintenance resources. Therefore, four types of data need to be collected, i.e., the operating data of target equipment objects, the working environment data, the state data of service resource objects, and the scheduling environment data. Data sources include sensors, controllers, energy monitors, enterprise information systems, internet, etc. Implementing sensor-based state monitoring of equipment or manufacturing processes will encounter many issues like hardware and software investment cost, sensor selection, and sensor placement. Some indirect methods have been displayed to collect state feature parameters of the target equipment object under the insufficient sensor condition [29]. The key supporting technologies of this layer include sensor selection and deployment technologies, internet of things technology, collection and transmission technologies for different data types.

B. LOCAL FUNCTION LAYER OF THE CYBERSPACE
The collected original data in the physical space function layer are isolated and littery. These data will be stored in a temporary database. The local function layer of the cyberspace offers data preprocessing capacity and plays as the data buffer and filter. The key supporting technologies of VOLUME 8, 2020 this layer include data cleansing technology, data partition technology, and feature extraction technology. This data processing for each target object is completed independently by their owner. After preprocessing, these filtered data are uploaded to the cloud function layer of the cyberspace.

C. CLOUD FUNCTION LAYER OF THE CYBERSPACE
Related state feature parameters are extracted from the filtered data through data analysis and mining technologies in this layer. It is necessary to extract different state feature parameters for different data sources. For predictive maintenance target equipment, the extractable state parameters include fault state indexes, the degradation state indexes of production efficiency, the stability state indexes of process quality, remaining service life indexes, etc. For the production environment of target equipment, related production task information for a specific period time and their relative importance should be extracted. For maintenance service tools and device resources, the extractable state parameters include spatial location, available quantity, functional state, scheduling cost, scheduling time, etc. For consumption category maintenance resources like maintenance service spare parts, the extractable state parameters include attribution state, spatial location, available quantity, scheduling time, scheduling cost, functional state, etc. For maintenance service technicians, the extractable state parameters include attribution state, spatial location, available quantity, scheduling time, scheduling cost, skill state, etc.
Based on these data, fault state prediction information of target equipment is obtained based on corresponding equipment fault diagnosis and prediction technologies, available maintenance resource collection is pushed in time, and then predictive maintenance decisions can be supported. Moreover, these state information of different objects related to predictive maintenance services are integrated and stored in a could database. Self-comparison capabilities among the same or similar equipment objects are provided to aid in object performance evaluation and state prediction [30], [31]. Based on the state information of these same or similar service resource objects which is collected and uploaded by different service suppliers, the available service resource objects range for complex equipment maintenance is expanded.

III. INTEGRATED INTELLIGENT GREEN SCHEDULING OPTIMIZATION STRATEGY FOR COMPLEX EQUIPMENT A. MATHEMATICAL PROBLEM MODELING FOR PREDICTIVE MAINTENANCE DECISION
Based on the framework proposed above, associated knowledge data come from real-time fault state of target equipment, the working environment of target equipment and scheduling environment of service resources can be obtained timely to support predictive maintenance decisions. In this section, both the economic performance and the environmental performance are concerned, an integrated mathematical model for predictive maintenance decision is established to minimize maintenance cost and the carbon emissions of maintenance scheduling process simultaneously. The final decision result includes the maintenance time node and the corresponding optimal scheduling scheme of maintenance resources.

1) MODEL DESCRIPTION AND PARAMETER DEFINITION
Some illustrations of this mathematical model are described below.
• A target complex equipment may occur several different fault types during its entire life cycle. Different maintenance service tasks can be combined to make a synthetic maintenance plan under certain conditions [17], [32]. This paper only discusses the predictive maintenance decision of a single fault type. A fault type is divided into several fault levels according to the fault degree and several feasible maintenance modes are distinguished. The applicability and expected service execution time of different maintenance modes are different for different fault levels. Different maintenance modes under different fault levels will result in different equipment improvements. Its effects will be reflected in the extension of equipment life and the restoration of equipment state. Some models for the equipment fault probability have been provided [17], [33]. The equipment fault probability for a specific time node can be calculated.
• Based on existing scheduling information and historical experience information, the job task sequence associated with the maintenance target equipment between the scheduling decision time node and the expected equipment fault time node can be obtained by the user enterprise. Each job task is uninterrupted and maintenance service can only be performed between two adjacent job tasks. The importance of different job tasks is distinguished. The delay in job tasks caused by maintenance will lead to the delay penalty cost.
• Predictive maintenance is beneficial to improve the production efficiency and the overall product quality. The probability of defective products can also be reduced. These improvements and the flexible manufacture ability of user enterprises can offset part of the delay penalty cost. The equipment reliability will gradually decrease and the equipment maintenance cost will present a concave curve with the degradation of the device equipment state. Considering the degradation mode and the coupling relationship of complex functional structures, to restore the equipment state through the maintenance service in the early stage of equipment performance degradation can create more service revenue.
• Monitoring and forecasting of the target equipment fault state can help to obtain more explicit requirements of type and quantity for maintenance service resources. The service resource type includes spare parts, professional equipment tools, and technicians. There have demand relationships between maintenance modes and needed resource types. Based on the framework proposed above, mot only the schedulable scope of maintenance resources is extended, but also more real-time and accurate maintenance service resource state information can be obtained. Each resource type has several different feasible scheduling paths which correspond to different scheduling time and cost. To describe the decision problem of predictive maintenance clearly, some related parameters are defined in Table 1.

2) OBJECTIVE FUNCTION FOR PREDICTIVE MAINTENANCE DECISION
There are two different optimization objectives are proposed in this mathematical model. The first optimization objective attempts to minimize a global maintenance cost results from resource consumption, production delay, and fault risk, etc. The second optimization objective seeks to minimize the carbon emission during maintenance resource scheduling.
The direct maintenance cost and the system fault risk cost are usually considered as common optimization objectives [34]. However, maintenance service revenue is neglected. After the maintenance service, the equipment trouble-free running time is extended and the production efficiency is improved. There is a nonlinear relation between maintenance cost and maintenance revenue under the complex and polytropic manufacturing environment. Therefore, the more precise predictive maintenance plan should simultaneously emphasize the direct maintenance cost, the risk cost, and the maintenance service revenue. A global maintenance cost is proposed in this section. This global cost is divided into four parts, resource scheduling cost, delay cost of job tasks, maintenance service economic benefits, and system cumulative fault risk cost.
The resource scheduling cost includes the scheduling cost of various service resources such as professional tools, technicians and spare parts. The scheduling cost CR is defined as: The execution of the maintenance plan at different time node will cause the delay of subsequent job tasks. The delay cost CD is defined as:  The increased equipment trouble-free running time, the restoration for the level of qualified product rate and the improvement of production efficiency will bring additional economic benefits to user enterprises. The maintenance service economic benefits IF is defined as: Xz k ×Ex k × In aver +In amp ×Til ij ×m i ×st j (6) The equipment state change is uncertain. It is necessary to consider the equipment fault risk cost when making maintenance decisions. The system cumulative fault risk cost FR is defined as: where The cost IF should be increased, while the cost CR, CD, and FR should be reduced. User enterprises may pay different attention to these four parts. This paper combines these four cost parts into a global maintenance cost by weighting. Thus, the optimization objective for the maintenance cost is defined as follows: where the vector a = (a 1 , a 2 , a 3 , a 4 ) denotes the weight distribution of four cost parts with 4 i=1 a i = 1. Their value can be confirmed by user enterprises through expert estimation and comprehensive assessment.
Two carbon emission sources of the predictive maintenance service process are included in this optimization objective for the environmental impacts.
The first source are carbon emissions come from the maintenance resource distribution process. The corresponding maintenance service resource scheduling becomes very complicated since complex equipment involves multidisciplinary technologies and various types of components from different sources [5], [35]. Different operation links (for example, disassembly, inspection, maintenance, assembly, etc.) need to depend on different maintenance resources (such as professional testing equipment, hoisting tools, assembly tools, professional operators, etc.) during the process of maintenance service. These maintenance resources may belong to different maintenance service resources suppliers of different spatial distribution. From the perspective of environmental protection, different distribution methods and routes will produce different fuel consumption or carbon emissions. The carbon emissions CE rd caused by fuel consumption can be calculated as: The carbon conversion coefficient of different common energy resources can be obtained by converting the current coefficient of standard coal equivalent [36], [37]. The carbon conversion coefficient Re of the electricity is 0.781 kg/kwh and the carbon conversion coefficient Rg of gasoline is 2.26 kg/L [38]. Fuel consumption of the distribution process is related to vehicle type, driving speed, vehicle weight, load, and distance [39]. Based on the existing research [40], the total fuel consumption Gs jk of the k-th path for the j-th resource type is expressed as: where Fr jk denotes fuel consumption constant of the k-th path, Fv jk is fuel consumption constant of the corresponding distribution vehicles, L jk equals to the sum of no-load weight of distribution vehicle and vehicle load, V m jk is the average velocity of distribution vehicle, and D jk is the distribution distance of the k-th path.
The second source are carbon emissions come from the maintaining operation process. Maintenance resource tools with different types and states generate different unit energy consumption. The carbon emissions CE mo caused by electricity consumption during maintenance service can be calculated as: The optimization objective for environmental impacts is to reduce the total carbon emissions of maintenance service. Therefore, the optimization objective for environmental impacts is defined as follows:  [44]. Function optimization problems of the manufacturing field have been typical application goals for these intelligent optimization algorithms [45]. The existing research provides a good reference for solving the function optimization problem proposed in this paper.
In the model of optimization problem established above, a feasible predictive maintenance service scheduling scheme includes the maintenance time node, the maintenance mode and the corresponding optimal allocation scheme of maintenance resources. The time interval between the scheduling decision time node and the expected equipment fault time node is relatively long in a normal condition. The variable n that is the total number of job tasks associated with the predictive maintenance target equipment in this time interval may take a relatively large integer value. The variable b that is the total number of fault maintenance modes may take a relatively large integer value under the method of detailed partition. The multisupplier collaboration based maintenance service mode extends the schedulable scope of maintenance resources. The variable k i that is the number of feasible scheduling paths for the i-th resource type may take a relatively large integer value, too. Therefore, predictive maintenance service scheduling for the complex equipment under the ubiquitous state data awareness environment is a complicated scheduling problem, the number of possible predictive maintenance service scheduling schemes is h i=1 n × b × k i , and this number is enormous. For the complex scheduling problem, when the number of possible scheduling schemes becomes very large, the efficiency of traditional calculation methods drops considerably, and thus, the actual application cannot be satisfied. The heuristic algorithm can be an appropriate method for addressing this problem since it can find a feasible solution in an acceptable time compared with the traditional optimal solution algorithms [46].
The proposed optimization problem of predictive maintenance has two objectives. There are interactions between different objectives in a multiobjective optimization problem, and the improvement of an objective may cause the performance reduction of the other objectives. To solve different multiobjective optimization models of specific engineering problems, various of multiobjective optimization algorithms have been proposed and utilized by researchers. For example, Guo et al. considered multiobjective optimization problems of the product disassembly planning with resource-constrained and sequence-dependent relationship and proposed a lexicographic multiobjective scatter search method [47], [48]. Li et al. proposed a multiobjective optimization configuration method for maintenance service resources of complex equipment based on an improved NSGA-II algorithm [46]. Other classic algorithms include PESA, SPEA2, etc [49], [50]. Among them, NSGA-II shows better convergence stability and faster convergence rate for VOLUME 8, 2020 the optimization problem with less than three objectives [46], [51]- [53]. In addition, an optimization problem with a similar type was solved successfully by utilizing NSGA-II [46]. Hence, this paper solves the proposed two-objective optimization problem based on the improved NSGA-II.
The final optimization result is a set of Pareto optimal solutions. As the Pareto front of feasible solution space, there has a non-dominating relationship between any two different solutions of this set. Then, from the point of practical control requirements of enterprise for the maintenance cost and the carbon emission index, a proper solution can be chosen to guide the final predictive maintenance plan that can be motivated by the results in discrete event systems [54]- [56]. The flowchart of NSGA-II is shown in Fig. 2.

1) CHROMOSOME CODING METHOD FOR MAINTENANCE PLANS
The coding structure of a feasible predictive maintenance plan is shown in Fig. 3. The maintenance plan chromosome is coded with integer. It consists of three parts, the execution time node, the maintenance mode, and the corresponding maintenance service resource allocation scheme. The code length of a chromosome is 2+h. The Gene 1 and Gene 2 of chromosome code denote the selected execution time node of maintenance service and the type of selected maintenance mode, respectively. Namely, this service plan would be executed before the a-th job task when Gene 1 = a, the b-th maintenance mode type is selected when Gene 2 = b. The selected scheduling paths for corresponding h types of service resources are represented through Gene 3 to Gene 2+h . When Gene i = c, it indicates the c-th path of the i-th resource type is selected. Corresponding chromosome code of a feasible maintenance plan code can be obtained through the conversion and extraction of mathematical model parameters.

2) THE CROSSOVER AND MUTATION METHOD
As shown in Fig. 4, a small-scale whole interference crossover method is used to assist in enhancing population diversity and global search capabilities. Judgment of singlepoint mutation is carried out for each chromosome after the whole interference crossover process.

3) THE NON-DOMINANT SORTING METHOD
Assume that there are M optimization objectives. We use f k (B) and f k (C) to denote the calculated value of the k-th optimization objective for feasible maintenance plans B and C, respectively. When f i (B)≤ f i (C) is met for arbitrary i (i = 1, 2, . . . , M ) and there exists j (j ∈ {1, 2, . . . , M }) such that f j (B)< f j (C) comes into existence, the maintenance plan C is dominated by maintenance plan B. When there exist i (i ∈ {1, 2, . . . , M }) and j(j ∈ {1, 2, . . . , M }) such that f i (B)< f i (C) and f j (B)> f j (C) come into existence, respectively, there exists a non-dominated relationship between B and C. The non-dominated sorting method is shown in Fig. 5.

4) CALCULATION OF THE CROWDING DISTANCE
When several chromosomes with the same chromosome non-dominated level should be to be compared, the calculation of crowding distance is utilized to distinguish them and maintain the diversity of the individual population. Fig. 6 shows the calculation process of the crowding distance.

C. DATA-DRIVEN DYNAMIC ADJUSTMENT STRATEGY
Things are constantly changing and full of uncertainties. In addition to the accuracy of prediction information for fault state of target equipment, dynamic characteristics of maintenance service decision environment should also be concerned. There may be unexpected state changes for related service resources, such as changes in scheduling time or cost. In order to adapt to external environment pressures (such as the increasingly fierce market competition, customers individualized demands, environmental protection requirements, etc.), more flexible and dynamic production scheduling mode may lead to frequent changes of job task sequence associated with target equipment, constraint conditions and optimization objectives of maintenance decision may also be adjusted according to the subjective desire of enterprise.
To improve the responsiveness of predictive maintenance decision-making to uncertain factors, a data-driven dynamic maintenance decision strategy is adopted in this section. This dynamic decision process for an unexecuted maintenance plan is described as follows. Assume that a corresponding maintenance plan for a maintenance task is determined at time t i . The T s (t i ) and T E (t i ) denote respectively the expected maintenance execution time node and expected termination time node of dynamic correction process that is calculated at time t i . When t i+1 ≤T E (t i ) is satisfied and the (i + 1)th dynamic adjustment operation is triggered, rescheduling calculation which based on the latest data for this unexecuted maintenance task will be performed at ti+1 and then the original maintenance plan will be updated. The updated content of maintenance plans includes three parts, the execution time node, the maintenance mode which affects the maintenance resource requirement, and the configuration scheme of corresponding maintenance service resources. The resource configuration scheme includes different resource types and corresponding suppliers.
We do not restrict the driving and termination conditions for dynamic adjustment operations for an unexecuted maintenance plan and set it flexibly. Feasible driving conditions include the error threshold of prediction accuracy for fault index, the change of production plan, the state change of service resources and the maintenance service environment, the adjustment of optimization objectives for the maintenance plan, etc. The time node which satisfies the max scheduling time of necessary maintenance resources before the execution time node can be one of the feasible termination conditions.

IV. CASE STUDY
The large vertical mill with the complex product structure and a long lifecycle is the key equipment of the cement production system. The grinding roll is one of the key parts and the bearing is its core component. To guarantee the operation performance of vertical mills, working status monitoring of the bearing has been implemented in some enterprises. In this section, relevant monitoring and prediction data of bearing fault state, common maintenance mode data and available maintenance resource data which come from a user enterprise of the large vertical mill equipment are collected as the case data. The objective is to validate and analyze the impact of the proposed integrated intelligent green scheduling optimization method on predictive maintenance decisions.

A. PREPARATION OF THE CASE DATA
In this case, a predictive maintenance requirement is generated when the bearing fault state indicator of the target equipment reaches its critical threshold in this case. The corresponding bearing fault state that beyond this critical threshold are divided into five levels. There are three different maintenance modes to be chosen which include the simple non-replacement maintenance action (e.g., cleaning, lubrication, etc.), the complex non-replacement maintenance action (e.g., the repair of damaged parts) and the replacement maintenance action. Different maintenance modes can be implemented to deal with the same fault state. A total of ten maintenance service resource types are needed to support maintenance operations under different maintenance modes. From the generation of maintenance requirements to the final implementation of maintenance service, there are 24 job tasks associated with the target equipment in the acquisition interval of case data.
The initial fault state monitoring and prediction data of the generation time node of maintenance requirement and other relevant parameter data are shown in Table 2. The corresponding parameter definitions are shown in Table 1 of section 3. The data has been preprocessed to eliminate the impact of dimension and order of magnitude. The corresponding initial predictive maintenance plan can be obtained based on them.
The real-time state of relevant target objects is changing constantly under the dynamic maintenance service environment. Considering the uninterruptible assumption of the job task, the latest relevant data of the completion time node for each job task are all collected. These related data include the monitoring and prediction data of bearing fault state and the state data of available maintenance resources. They will be used to analyze the impact of the data-driven dynamic adjustment strategy on predictive maintenance decisions. Other parameters such as job task sequence and its weight do not change with time in this case.
These dynamically changing parameters include the equipment fault probability vector Pl, the corresponding fault level vector Cr, the correction factor vector Rt of the resource scheduling time, the service resources scheduling time matrix T hxMK , the service resources scheduling cost matrix C hxMK , and fuel consumption matrix Gs hxMK ,. Due to a large amount of data, the changes of vector Pl and vector Cr are shown in Figs. 7(a) and 7(b), respectively. Data change of the matrix T hxMK , Gs hxMK and C hxMK are not listed.

B. COMPARISON AND ANALYSIS
The case data in Table 2 is utilized to test the effectiveness and superiority of the proposed improved NSGA-II in solving the   integrated intelligent green scheduling problem. For comparison, the NSGA-II proposed in [46] is also used. The same encoding method and decoding rules are used to have a fair comparison. These algorithms run on Intel I5 (3.20 GHz/8.0G RAM) PC with a Windows 10 operating system. In order to obtain a suitable parameter combination of the improved NSGA-II and compare the performance of the two algorithms, we have conducted some preliminary experiments with different parameter combinations. The experimental set of the initial population is {50, 100, 200, 350, 500}. The experimental set of the maximum evolutionary generation is {50, 100, 200, 350}. The experimental set of the crossover probability is {0.85, 0.9, 0.95}. The experimental set of the mutation probability is {0.05, 0.1, 0.15}. The results of multiple optimization calculations show that both algorithms can effectively solve the integrated intelligent green scheduling problem. However, under the same parameter conditions, the average run time of the proposed NSGA-II in this paper is reduced by approximately 8% compared with the algorithm in [46]. The improvement of algorithm efficiency has important significance for timely maintenance decision. When both the size of initial population and the maximum evolutionary generation are small, for example, both are 50, the Pareto front of the proposed algorithm contains better solutions compared with the algorithm in [46]. When both the size of initial population and the maximum evolutionary generation are large enough, for example, both are 350, the Pareto optimal front of the case problem can be found by these two algorithms. Nevertheless, the probability of falling into the local optimal solution of the proposed algorithm is greatly reduced. Besides, the time complexity of the improved NSGA-II is O(maxgen * obn * popsize 2 ). The maximum evolutionary generation maxgen, the number of objectives obn and the population size popsize are the key influencing factors of the algorithm performance. The adopted whole interference crossover method of this improved NSGA-II can effectively reduce the value of maxgen and popsize while improving the optimization speed.
Based on experiments and analysis, an optimal combination of algorithm parameters with better performance is selected.
The selected algorithm parameters are as follows: the size of the initial population is 100; the maximum evolutionary generation is 100; the crossover probability is 0.9; the mutation probability is 0.1. The detailed influence of parameters on the optimization quality and the further analysis of the relationship between the problem size and algorithm parameters will be discussed in future work.
The case problem is solved by the improved NSGA-II. Fig. 8(a) shows the proportion change process of non-inferior solutions. Fitness change of two different optimization objectives are shown in Figs. 8(b) and 8(c), respectively. Fig. 8(d) shows the final Pareto front.
Several feasible predictive maintenance plans come from the final Pareto optimal solution set are listed in Table 3. To build a benchmark, an optimal maintenance plan which considering the single maintenance cost objective is obtained based on the general single objective genetic algorithm and its corresponding carbon emission is calculated. The optimal cost fitness is 26.861536, a feasible maintenance plan chromosome is [5,1,6,3,8,2,1,6,2,2,8,6], and the corresponding carbon emission fitness is 133.9262. Based on theoretical analysis, this solution should be dominated or included in the Pareto front. As verified in Fig. 8(d), this single objective optimization result is included in the Pareto front for the optimization problem. The Pareto front solution set also includes some more balanced feasible predictive maintenance plans. Compared with the single objective optimization which only  considering maintenance cost, multiobjective optimization for predictive maintenance decision enables user enterprise to minimize carbon emissions of maintenance process while controlling the maintenance cost.
The detailed maintenance plan information represented by chromosome code can be obtained by decoding. For a feasible maintenance plan chromosome [5,1,1,3,8,8,1,6,7,1,8,6], the first code 5 represents that maintenance service should be executed before the fifth job task. The second code 1 indicates that the first maintenance mode should be adopted. The remaining code string represents the necessary maintenance resource types and their corresponding scheduling paths. According to matrix O bxh which describes the demand relationships between maintenance modes and maintenance resource types, necessary maintenance resource types for the first maintenance mode can be confirmed. These resources types include the first, the second, the fifth and the eighth. Therefore, the first code, the second code, the fifth code, and the eighth code in [1,3,8,8,1,6,7,1,8,6] are valid codes. For these four types of necessary maintenance resources, their scheduling paths correspond to the first path, the third path, the first path, and the first path, respectively. From these feasible maintenance plans of the Pareto front, enterprises can identify the most appropriate one based on their latest preferences. Carbon tax and carbon credits can be an important decision basis. In this case, chromosome [5,1,1,3,8,8,1,6,7,1,8,6] is chosen as the initial predictive maintenance plan. The expected resource scheduling time is 3.5 according to Table 2. Compared to only optimizing the resource allocation scheme in literature [50] and only optimizing the maintenance time in literature [18], the proposed integrated optimization method can obtain a more feasible maintenance plan.
In response to uncertainties and dynamic characteristics of the maintenance service environment, dynamic adjustment of the unexecuted predictive maintenance plan is implemented. Completion of each job task is chosen as the trigger condition of rescheduling adjustment in this case. It generates a new maintenance plan based on the latest data to replace the original plan. This iteration process is stopped when the time difference between decision time node and corresponding expected execution time node is less than the demand time for resource scheduling. As shown in Fig. 9(a), rescheduling iteration process is terminated after the sixth adjustment. The corresponding Pareto front solutions which are calculated based on the latest forecast data is listed in Table 4. They show that the relatively optimal maintenance time node is before the eighth job task. Analyzing the dynamic rescheduling process, it is found that the execution time node of the maintenance service and the resource configuration scheme are optimized and updated according to the latest dynamic data. The maintenance cost fitness and carbon emission fitness fluctuate constantly even though with the same fault level and maintenance mode. It is because of the dynamic change of the maintenance resource state. Besides, the bearing of grinding roll is expensive and the loss of unit downtime is relatively large for the large vertical mill. Hence, it is more inclined to implement lowcomplexity maintenance in early fault levels. Fig. 9(b) shows the change of Pareto front for different rescheduling time nodes. With the dynamic adjustment strategy, the opportunity of obtaining predictive maintenance plans with lower maintenance costs and carbon emissions can be improved, and the adaptability of the maintenance plan is guaranteed as well.

V. CONCLUSION
Scientific and efficient complex equipment maintenance security can not only improve the equipment production efficiency but also play a key role in the path towards green manufacturing. Predictive maintenance further expands this potential. However, prediction accuracy and adaptability of maintenance plans limit the implementation of predictive maintenance. Meanwhile, researchers pay less attention to the environmental impact of the maintenance scheduling process. To address this problem, this paper focuses on the precise decision of predictive maintenance under a ubiquitous create maintenance scheduling environment. A structural framework of information sharing and service network for predictive maintenance scheduling of complex equipment is introduced to achieve real-time perception of equipment objects and maintenance service resource objects. This framework supports the interactive sharing of data across organizational boundaries and expands the schedulable range of distributed maintenance service resources. Carbon emission can also be monitored and controlled based on scheduling process awareness. Then, a mathematical problem model of predictive maintenance decision is established. Carbon emission is introduced as one of the objectives. Through the improved NSGA-II algorithm, a Pareto front solution set corresponding to different preferences is provided to user enterprises. The maintenance time node, the maintenance mode, and the configuration scheme of required maintenance service resources are all integrated into this optimization problem. Considering uncertainties of the maintenance service environment, a data-driven dynamic predictive maintenance decision strategy is applied. Finally, a case study is provided to validate the proposed dynamic predictive maintenance decision method. The proposed integrated intelligent VOLUME 8, 2020 green scheduling optimization method can improve the reliability of a predictive maintenance plan and mining the green performance potential of maintenance scheduling process.
In the next phase of our study, some constraints and assumptions of the mathematical model will be improved to adapt to the actual manufacturing environment. More sources of carbon emissions, and other pollutant emissions during the whole maintenance service process will be considered. Green predictive maintenance with high precision is an important research direction. The research in this paper and future work will provide feasible technical supports in this filed. He is currently a specially appointed Professor in Zhejiang University, a Doctorial Supervisor, and an Academician of the Chinese Academy of Sciences, the Dean of mechanical engineering, the Associate Supervisor of the CAD&CG State Key Laboratory, the Head of the Institute of Engineering and Computer Graphics, Zhejiang University, and the Chief Supervisor of engineering graphics state fundamental courses. His research focuses on mechanical designing and theory, digitalized designing, and manufactory. He has published eight pieces of monograph or compile, 142 pieces of articles indexed by SCI/EI, with more than 1600 times. He gathered his 15 years of working experience in manufactory and his theory in science, and proposed the technique in combination of batch and customization which is used for multitudinous customization, the status in engineering transition, the fuzzy status, the technology simulation of random status modeling and digitalized prototype integration, the combination of figures and geometry in multicomponent correlation in intricate equipment, and the analysis of multilevel disposing and multiparameter matching. He received the National Prize for four times, which includes the second prize in National Technology Progress twice and the first prize in National Excellent Education Achievement. He has also received the first prize in Provincial Level Technology Progress for six times. He transforms his technology into the software and got 12 copyrights of computer software, which achieved a lot in manufacturing enterprises.