Intelligent Deployment Solution for Tabling Adapting Deep Learning

This article presents an intelligent deployment solution for tabling that utilizes deep learning techniques. The solution involves adapting deep learning semantic segmentation algorithms with DeepLab V3+ to extract multi-dimensional image features, enabling the mapping of relationships between mineral ore belt characteristics and operating parameters using a multi-output support vector regression model optimized using a sparrow search algorithm (ssa-msvr). The proposed solution integrates image recognition software and data processing method, which significantly improves the efficiency and effectiveness of mineral processing, providing a promising avenue for further research and development in this field.


I. INTRODUCTION
To perceive the operating state of the tabling, the traditional operation controls of the tabling are up to the operator to sense the mineral ore belt of the tabling and understand the characteristic information such as the shape and color of the ore belt. Control parameters such as horizontal flushing water, bed slope, stroke, and stroke rate are adjusted accordingly to complete. Since the difference of everyone's experience, technical proficiency, and sense of responsibility, there are unavoidable errors and uncertainties in determining whether the operating state of the tabling is normal, if only by observing the mineral ore belt of the tabling with the naked eye. Therefore an objective, high-precision and fast method is urgently required to obtain the characteristic parameters of the mineral ore belt and to quickly and accurately determine the real-time operating status and indicators of the tabling.
Nowadays, the overwhelming majority of researchers attempted to optimize the operating parameters of the tabling and predict indicators using artificial neural networks (ANN) The associate editor coordinating the review of this manuscript and approving it for publication was Mohammad Shorif Uddin . based on industrial data [1], [2]. However, these studies achieved low accuracy and reliability in the optimization process. In recent years, with the development of machine vision technology, it has been widely applied in various fields [3], [4], [5], [6], including feature extraction of mineral ore belt images [7]. The gray-scale calculus image processing algorithm is utilized to calculate the boundary points of the ore belt, obtain the characteristics of the boundary line, and ultimately determine the width and color characteristics of the ore belt, enabling complete segmentation of the mineral ore belt [8]. Building on this research, the Beijing General Research Institute of Mining and Metallurgy (BGRMM) has developed an automatic mining system for gravity beneficiation tabling [9], which is mainly composed of a mining device and an inspection robot with a camera and image processing software, which improves the efficiency of mineral separation by the tabling, instead of workers intercepting concentrates, and realizing automatic interception of concentrates on the gravity beneficiation tabling [10], [11]. These studies start from the method of mineral ore belt image processing to achieve the optimization of mineral interception operations, opening new directions for later optimization studies and VOLUME 11, 2023 This work is licensed under a Creative Commons Attribution 4.0 License. For more information, see https://creativecommons.org/licenses/by/4.0/ intelligent deployment. And then multi-threshold color image segmentation algorithm based on krill optimization, and multi-threshold color image segmentation algorithm based on improved firefly algorithm, etc. which are applied to the image segmentation of the mineral ore belt [12], [13]. Increasingly effective deep convolutional neural network models and image processing algorithms are applied to the process of mineral identification and separating [15], However, all researchers currently use deep learning only to obtain the boundary separation points of the ore zone, ignoring the characteristics of the line and surface of the mineral ore belt, and all the feature detection algorithms based on deep learning serve to replace the manual realization of accurate concentrate interception, only to achieve the optimization of the interception operation, ignoring the optimization of multiple operating parameters, such as the optimization of the performance of the equipment itself of stroke and stroke times, so it is difficult to achieve the intelligent deployment.
To obtain multi-scale effective geometric features compared with other image recognition algorithms, which specifically includes all extractable features of points, lines, and surface from the entire mineral zone area. A semantic segmentation algorithm based on the excellent performance of DeepLab V3+ series is suitably applied to the detection of such rich features, which is of great significance for constructing the ML model of relationship between image features and parameters of the horizontal flushing water, slope angle, stroke, and stroke rate and other operating parameters, as well as the beneficiation indicators of the recovery rate and concentrate grade [16], Because rich and multi-scale image features can provide maximum dimensionality for the input of ML models, thus effectively guaranteeing the accuracy of relationship model mapping [8]. Traditional ML models have some shortcomings in accuracy and most of them are only suitable for single-output prediction. In this study, a multioutput support vector regression model optimized by a sparrow search algorithm (SSA-MSVR) is used to construct a mapping model between the image features of the mineral ore belt and the attributes of the tabling equipment, whose biggest advantage is that the SSA is used to optimize two important hyperparameters of the SVR to obtain the maximum prediction accuracy of the MSVR. Finally, combined with advanced deep learning and data mining technologies, and friendly optimization strategies, we have realized the intelligent deployment solution of the mineral processing of tabling.

II. MULTI-SCALE FEATURE EXTRACTION OF MINERAL ORE BELT IMAGE THROUGH DEEP LEARNING
In this study, feature extraction of mineral ore belt images is the core element to achieve the goal of intelligent mineral processing. In order to deploy the research results to practical applications as soon as possible, we need to consider many application issues, such as, whether the algorithm is stable and reliable enough, whether the software is mature, whether the communication is feasible, etc.

A. ACQUISITION AND PRODUCTION OF DATASETS
In order to generate high-quality datasets with composite model characteristics, we preprocessed the raw images generated from industrial, mainly through methods of denoising and image enhancement [17]. In this study, the image enhancement method of gray-scale transformation and the denoising method of Gaussian filtering is mainly used. During image preprocessing, we performed uniform scaling of the image, the image gray value was normalized to a reasonable gray value range, and through the reasonable allocation of pixel value weights, a large number of image preprocessing parameter combinations are obtained, which will be applied to the training of deep learning segmentation models.

B. TRAINING AND INFERENCE EVALUATION OF DEEP LEARNING MODEL
The advanced deep learning semantic segmentation with DeepLab v3+ continues to improve on the architecture of the DeepLab v3 by introducing Encoder-Decoder, which is commonly used for semantic segmentation, in order to fuse multi-scale information. As shown in FIGURE 1 the resolution of the features extracted by the encoder can be arbitrarily controlled, and the accuracy and time consumption are balanced by null convolution. Two pre-trained deep learning neural network models are provided, namely, dl_segmentation_enhanced and dl_segmentation_compact with DeeLab V3+. The result of semantic segmentation is an output image where the pixel values represent the specified category of corresponding pixels in the input image, for general deep learning (DL) networks, deeper feature mappings representing more complex features are often smaller than the input images. To obtain an output of the same size as the input, it uses a split network consisting of two components: an encoder and a decoder. The encoder has determining the characteristics of the input image, and since the information is ''encoded'' in a compressed format, the decoder needs to reconstruct the information into the desired result, in which case each pixel is assigned to a class, but when classifying the pixels, overlapping instances of the same class are not distinguished as distinct. Edge extraction is a special case of semantic segmentation, in which the trained model is designed to distinguish between two categories: ''edge'' and ''background.'' As shown in FIGURE 2, independent of the deep learning method used, the data has to be provided to the model following certain conventions. The model receives a dictionary for each input image, such a dictionary contains images and, in the case of training and evaluation, also contains information such as basic fact annotations, and as output, the model returns a dictionary with the results. Although the model itself is not required, dictionaries are used during training and evaluation.
As shown in Table 1, the training of the model is done on a personal notebook, using a professional indus-  trial camera and data acquisition software for datasets acquisition, and a server environment with running ability of deep learning is built. In order to facilitate the later model deployment and the intelligent application of mineral separation process, HALCON is used as our development script.
At present, the most commonly-used evaluation index in semantic segmentation is the mIOU. The mIOU is a standard evaluation index for image segmentation work, which calculates the overlap ratio of the intersection of two sets and their union. It calculates the difference between the real segmentation and the segmentation predicted by models, the per-class intersection over union (IOU) gives for a specific class the ratio of correctly predicted pixels to the union of annotated and predicted pixels. In image semantic segmentation of DL, many criteria are usually used to evaluate the accuracy of the model's inferred results. In addition to IOU and its variants, pixel accuracy and its variants also are included. Pixel Accuracy (PA) is the simplest measure, which is the percentage of pixels that are correctly labeled. Mean Pixel Accuracy (MPA) is a simple improvement of PA, which calculates the proportion of correctly classified pixels in each class, and then calculates the average of all classes.
As shown in Table 2, the training parameters of the DL model are set uniformly, and the final training results of the model are visualized in FIGURE 3. It can be seen from FIGURE 3 that the datasets with two scenarios have very similar segmentation effects in the DL semantic segmentation model. The training effects of the two DL semantic segmentation models on datasets I and datasets II are excellent, and there is no over-fitting or unqualified mIOU value. From the statistical results in Table 3 and Table 4, in the results of datasets I, dl_segmentation_enhanced is slightly more time-consuming, but the mIOU value during training is higher than dl_segmentation_compact. And for the datasets II, dl_segmentation_enhanced has a slight advantage in time consumption, but the mIOU value is lower than dl_segmentation_compact by more than 10%, which is enough to show that scene images similar to datasets II are more suitable for the deep semantic segmentation model using dl_segmentation_compact.    Thus, we compared two sets of different styles of image datasets from the mineral ore belt, combined with the two deep learning semantic segmentation models, and we conducted four sets of experiments. From the overall experimental effect, in the training process, since dl_segmentation_enhanced has a more complex neural network structure than dl_segmentation_compact, the loss value is always more. For application scenarios similar to the image features of Datasets I, the enhanced dl_segmentation_ is recommended as the model to deployment and practical application, and for application scenarios similar to the image features of Datasets II, the dl_segmentation_compact is recommended as the model to deployment and practical application.

III. INTELLIGENT DEPLOYMENT SOLUTION
Modeling is an important and integral part of optimization [21], the adaptive adjustment method for multi-operating parameters combines the results of the MV, ML, automatic control technologies and it is an optimization algorithm proposed through the exploration of big data modelling, and the deep learning image processing algorithm as a whole, which is an important foundation for the intelligent control of the beneficiation process, and an effective idea to realize the monitoring and intelligent process of tabling beneficiation. ML with optimization methods have been widely and maturely applied in mining equipment [22]. With the help of these mature research results and applications, this study also designs a more reasonable optimization method by the corresponding applications [23].

A. RELATIONSHIP MAPPING BETWEEN MINERAL ORE BELT AND ATTRIBUTES OF THE TABLING EQUIPMENT
To more realistically map the relationship between image of ore belt and the attributes of tabling, in addition to the need for rich image features as input, the prediction accuracy and of the model should be considered. Traditional ML is difficult to achieve high accuracy with multiple inputs and multiple outputs [24], so this paper proposes a ML learning method based on SSA-MSVR.
As shown in FIGURE 4, the goal of SVM is to maximize the ''distance'' to the nearest sample point in the hyperplane, while the goal of SVR is to minimize the ''distance'' to the farthest sample point in the hyperplane, and the larger the interval in SVR, the more low-performance it is. SVR is one of the excellent, accurate and robust regression algorithms, insensitive to dimensionality, can handle linearly separable and linearly indistinguishable data, has high stability, fast algorithm updates, and generally chooses RBF as the kernel function [25]. There are two key optimization parameters affecting the SVR model, one is the penalty coefficient (c) and the other is the kernel coefficient (g), the larger the value of c, the greater the penalty and the worse the error tolerance, its influence on the range of Gaussian action corresponding to each support vector, and the larger the g, the worse the generalization performance.
As shown in FIGURE 5, the sparrow search algorithm (SSA) first divides the population into discoverers and joiners in the process of finding the optimal SVR hyperparameters. Suppose a population is set up as a matrix X.
where X ij is the position of the i-th sparrow in the j-th dimension, n is the population size, and d is the dimensionality of the objective function variable space. The discoverer provides directions to capture food for the entire population, so it has a larger search range and the location iterations are updated as follows.
where t is the current number of iterations, T is the set maximum number of iterations, X t and X t+1 are the positions of the sparrow before and after the iteration, and Q is the normal random number. When the discoverer is aware of the  danger, it warns the population, which optimizes its position in order to find a safe area.
X t P is the optimal position of the discoverer, X L is the global worst position, L is a matrix with all elements 1, and A is a random matrix with 1 or −1. A + = A T (AA T ). The stability of the mapping relationship determines the effect of the system on the monitoring of the mineral ore belt and processing status of the tabling, which in turn affects the accuracy of the adjustment requirements of the system's parameters, and ultimately affects the system's optimization results for processing indicators and operating parameters. Thus the analysis of the ML model in the system is of great significance for the simulation results of operating parameters and processing indicators. VOLUME 11, 2023    As shown in FIGURE 6 to FIGURE 11, which are the simulation results of the system for the four operating parameters of the tabling, feed water flow, stroke, stroke rate, slope angle, and the concentrate grade and recovery rate. The 20 groups of data from actual operating parameters and processing indicators of concentrators and the data predicted by the system simulation are obtained and are compared to calculate the accuracy and error of the simulation. As shown in Table 5   Percentage Error (MAPE) and coefficient of R 2 , respectively, which are used to evaluate the performance of ML models for regression predictions, and the smaller the MSE and RMSE, the better the fitting effect of the model, the higher the R 2 , the more realistic the model, when the R 2 is greater than 0.8, the model has a very high authenticity. The smaller the MAPE, MAE and MAPE are, the better accuracy of the model. It can be seen that the larger the simulated value, the larger the simulated error value, but the better the fitting effect, and The higher the MAPE, such as the value of stroke rate, recovery rate, stroke and concentration rate, which indicates that the ML model in the system has a good ability to learn the of the larger values. But on the whole, the ML model in the system has very high accuracy to meet the industrial demand when predicting the operating parameters and processing indicators, and it has extremely high authenticity and reliability.

IV. LIGHTWEIGHTING DEPLOYMENT
In view of the complex working principle and large computational power consumption of deep learning models, lightweight deployment of deep learning models is required in practical applications. In deep learning models with mainly neural network connections, shearing becomes one of the most efficient and simple lightweighting methods.
As shown in FIGURE 12, the general idea of pruning is to cut off the unimportant connections between neurons and set the unimportant parameters in the weight matrix to zero, combined with a sparse matrix for storage and computation [26]. In order to ensure the performance of the model itself, it is necessary to iteratively prune one small step at a time, and the specific practice is to sort the weight parameters according to their numerical size, and to set the last k% of the size ranking to zero, where k% is the compression rate.

A. IMAGE PROCESSING SYSTEM
To realize intelligent deployment solution of the tabling, independently developed software plays a crucial role. An image recognition system of mineral ore belt feature for the beneficiation process is developed in this study. The entire function of this system is developed in the VS programming environment, combined with inference module of HALCON's deep learning model. As shown in FIGURE 13, before the image recognition software on the desktop is opened, the backend server and PLC control system need to be started first. After the software is opened, the communication with the server and the PLC can be automatically realized, and the upper right corner of the software will display that the system works normally and the PLC connection is successful. From the effect of recognition, the mineral ore belt in the process of beneficiation is correctly identified, the characteristics of the strip are transmitted to the server in real-time, and the server sends out the requirements of adjusting the operating parameters such as feed water flow, slope angle, stroke, and stroke rate, to the PLC according to the multi-parameter adaptive adjustment method, after the operating parameters are adjusted, the geometric characteristics of the mineral ore belt are constantly fluctuating. Since the trained model is compressed by parameter pruning and sharing when it is embedded in the system, only a few milliseconds are occupied in the recognition of mineral ore belt.

V. CONCLUSION
This paper presents a novel intelligent deployment solution for tabling in mineral processing using deep learning semantic segmentation algorithms with DeepLab V3+. The authors obtained two sets of image datasets from a real concentrator with different mineral ore belt characteristics and evaluated the algorithm's extraction effectiveness using evaluation indicators of mIOU and PA of image segmentation. They proposed a mapping relationship model between the characteristics of the mineral ore belt and operating parameters using SSA-MSVR and achieved considerable performance in fitting effect and prediction accuracy for the intelligent beneficiation system. The authors also developed a real-time monitoring software for mineral processing using lightweight deployment of pruning. The research results demonstrate that this intelligent deployment solution greatly improves the intelligence of the tabling and lays a solid foundation for the intelligent realization of the beneficiation plant.

ACKNOWLEDGMENT
The authors sincerely appreciates the data and technical support provided by related companies and researchers and especially thanks to his supervisor for all kinds of help and Jiangxi Provincial Department of Science and Technology and Jiangxi Provincial support of key research and development projects.

COMPLIANCE WITH ETHICAL STANDARDS
Ethical approval and Informed Consent: Written informed consent for the publication of this article has been obtained from Jiangxi University of Science and Technology and all authors.
The authors declare that they have no known competing financial interests or personal relationships that could have appeared to influence the work reported in this article. LIU HUIZHONG was born in Xinfeng, Jiangxi, China, in 1969. He is currently a Professor and the Doctoral Supervisor with the School of Mechanical and Electrical Engineering, Jiangxi University of Technology, and the Director of Jiangxi Mining and Metallurgy Mechanical and Electrical Engineering Technology Research Centre. He has been engaged in the research and development of high-efficiency mineral processing equipment, mineral processing process simulation and control technology for more than thirty years. He has presided over the development of ''BL series spiral chute,'' ''BPF series automatic high efficiency filter press,'' ''flotation machine level control system,'' ''BXF cyclone particle size analyzer'' and other products and technologies. His research achievements have won one National Science and Technology Progress Second Class Award, one Hubei Science and Technology Progress First Class Award, one Beijing Science and Technology Progress Second Class Award, and three China Nonferrous Metals Industry Science and Technology Progress Second Class Awards. At present, he is mainly engaged in teaching and research on high-efficiency mineral processing equipment, process simulation of mineral processing and mineral processing intelligence.