Experimental Assessment of a Tiny AI-Empowered Output Filter Parameter Extraction Framework for Digital Power

An experimental assessment of a ‘‘plug-and-play’’ tiny artificial-intelligence-empowered output filter parameter extraction framework for digital power is presented. The framework can be incorporated into an existing digital controller to perform online parameter extraction without adding extra sensors or modifying the power conversion stage. The idea is based on firstly transferring the control from the default control law to the framework, which consists of a predefined control law to perform output regulation for a few switching cycles, then introducing a small-signal perturbation into the control signal, and finally utilizing a long short-term memory (LSTM) network to recognize the dynamic response of the control signal to perform either regression or classification of filter parameters. The LSTM network is trained with a reconfigurable output filter. The proposed framework has been successfully evaluated on a 240W, 100V/48V buck DC/DC converter prototype. The framework’s performance is studied by extracting the parameters of a second-order output filter and a fourth-order output filter. For the second-order output filter, the root-mean-square errors (RMSEs) in performing the filter inductor and capacitor regressions are 2.35% and 2.25%, respectively, and the F1-scores in classifying the inductance and capacitance are 0.805 and 0.815, respectively. The framework occupies 0.93% of the memory space of the controller with 512kB flash memories. The extraction time is 17.3ms. For the fourth-order filter, the maximum RMSEs in performing the regression of the filter inductors and capacitors are 2.04% and 6.49%, respectively. The framework occupies 3.62% of the memory space and the extraction time is 328ms.


I. INTRODUCTION
The latest developments in microelectronics and embedded system technology have accelerated the penetration of Digital Power (DP) to meet the increasingly complex power electronic systems [1]. DP offers the merits of high flexibility, high reliability, and system miniaturization. Apart from the monolithic integration of the power conversion stage (PCS) The associate editor coordinating the review of this manuscript and approving it for publication was Ali Raza . and controller, such as proposing sophisticated digital pulsewidth modulation (PWM) techniques to improve power efficiency in [2] and enhance transient response in [3], mainstream research also emphasizes dealing with the effects of sampling and computation delays and nonlinearities [2], [3], enhancing the characteristics of individual functional block and control law [4], [5], and optimizing system performance [6], [7]. Reference [2] proposes a digital PWM and low-power self-tracking zero current detector that can achieve high linearity and low power consumption. Reference [3] proposes a digitally controlled buck converter that utilizes multiple-outputs bang-bang phase detector in a digital phase locked loop to reduce design complexity and improve transient response. Reference [4] proposes an FPGA-based hysteretic multimode control scheme with two PWM modes and pulse frequency modulation (PFM) modes for mixed conduction mode is proposed. Reference [5] proposes a digital predictive control algorithm for boost-type power factor corrector with the peak-to-peak inductor current estimated by a triangular modulation-based predictive control and a known value of output inductor. In [6], a digital average voltage and digital average current predictive control achieves a high output voltage accuracy, wide stability range, and fast transient response. In [7], the effective resolution of the PWM is enhanced by using a dithering technique on an ''SoC FPGA'' that can reduce peak output voltage ripple and lower switching frequency subharmonics. The control laws are typically designed by taking circuit parameters into account [5], [6]. For example, in [5], the peak-to-peak inductor is estimated by taking the value of inductor into account. As discussed in [6], the control scheme proposed is robust and can keep converter operate in stable when the circuit parameters have small variation.
In principle, if the filter parameters can be identified, the digital controller can adaptively tune the parameters in the control law to maintain optimal system performance [8], [9]. For examples, the optimal sliding surface function in the sliding mode control for dynamic voltage restorers in [8] is dependent on the output filter parameters. The switching functions in the trajectory prediction control proposed in [9] requires using output filter parameters to aim at reaching the steady state in one switching cycle after supply/load disturbances.
The identified filter parameters can also be used as information for monitoring the condition of the filter components, thereby preventing possible catastrophic faults [10], [11], [12], [13] and/or parametric faults [14], [15], [16], and thus reducing the failure rate. The entire system will not break down unexpectedly and create unplanned downtime. In [11], different fault conditions are identified by using a Rogowski coil to sense inductor voltage and a voltage sensor to sensor diode voltage. The determination of the inductor voltage is obtained by considering the inductance of the output inductor and the coupling inductance. A detection method is proposed in [12] to classify transient fault current and switching inrush current. An overview of different condition monitoring techniques for DC-link capacitors is given in [13], because the capacitor is known to be a reliability critical component [14]. A degradation model of capacitors is presented in [15]. In [16], a reference submodule-based capacitor monitoring method for capacitance estimation in the modular multilevel converter is proposed. The idea is based on utilizing the relationships between one capacitor and the others to estimate the capacitance ratios.
The converter output filter is composed of passive components, including inductors, capacitors, and resistors. Among them, capacitors have the shortest life expectancy and are susceptible to parametric faults. A large body of literature is devoted to extracting the intrinsic parameters of output capacitor, including capacitance and equivalent series resistance (ESR) [13], [16]. The ESR is sometimes chosen as an indicator to indirectly determine the health condition of the capacitor as degraded capacitors have their capacitance reduced and ESR increased.
In theory, the value of each filter component can be computed by processing the sensed voltage and current waveforms associated with the component under test [17], [18], [19]. In addition to requiring extra sensors and increasing cost, the reliability of the entire system is also negatively affected. In [17], the idea is based on firstly introducing a low-frequency signal to the control signal, and then calculating the capacitor value by sampling the low-frequency voltage and current associated with the capacitor. In [18], odd harmonic frequencies are injected by the inverter during the night time. The capacitor impedances at different frequencies are obtained using a tuned bandpass filter. The ESR and capacitance are estimated by using Least Mean Square algorithm. In [19], the ESR is calculated by using high-frequency harmonics in capacitor voltage and current and the capacitance is calculated by using low-high frequency harmonics. The harmonics of interest are extracted by the Goertzel algorithm.
Practically, the operation bandwidth of sensors limits the measurement of high-frequency voltage and current waveforms, such as ripple voltage and ripple current. For example, reference [20] requires sensing the inductor current peak and output voltage jump. This is particularly critical for future power electronic systems that will operate at a high switching frequency.
Another extraction approach is based on first devising a circuit model to describe the system dynamics, and then applying deterministic or stochastic techniques to determine filter parameters [21], [22], [23]. Since the accuracy depends on the level of abstraction and sophistication of the model, characterization of high-order filters and nonlinear behavior will pose challenges in mathematical modeling and guaranteeing the numerical convergence.
To cope with the above issues, this paper gives another perspective by proposing a ''plug-and-play'' tiny artificial intelligence (AI)-empowered output filter parameter extraction framework for digital power. The idea is based on extracting the features of the dynamics of the duty cycle with a long short-term memory (LSTM) network to perform either 1) classification or 2) regression of the output filter parameters, after the system is subject to a small-signal perturbation in the control signal. The proposed framework has the following features, and a comparison with other methods has been given in Table 1:

1.
No additional voltage and current sensors are needed. Many online parameter identification methods require multiple sensors. Besides increasing system complexity and cost, additional sensors have a detrimental effect on system reliability.
2. The framework is a standalone software module. It can be incorporated into existing digital controllers to perform online extraction of output filter parameters.
3. No detailed mathematical models are needed, as the framework is purely data-driven. Conversely, many parameter identification methods require sophisticated system modeling.
4. The weights and parameters in the LSTM network are trained under all operational conditions. They are unnecessary to be trained for individual converters.
Supervised learning for determining the parameters of the LSTM network is adopted. The training and testing datasets are obtained under pre-determined output filter parameters, realized by a reconfigurable output filter (ROF). The parameters of the LSTM network are trained by a stochastic gradient descent algorithm. Data pre-processing, network operation, post-processing of the network output, and network training are given.
The proposed framework has been experimentally evaluated on a 240W, 100V/48V buck DC/DC converter with its output filter configured as either a second-order LC or fourth-order LCLC filter. The perturbation causes the output voltage with ±3.3% variation only.  The proposed framework performs the same objective as the default controller to regulate the output voltage. Upon estimating the filter parameters, the controller will transfer the output voltage regulation function from the default controller to the framework temporarily via S sel . The framework ''Controller'' in Fig. 1 implements a PI control with the proportional gain K p,f and integral gain K i,f . It generates a control signal with the duty cycle d f in the n-th switching cycle. Mathematically,

II. PROPOSED FRAMEWORK
and 14730 VOLUME 11, 2023  where T int is the sampling time interval generated by the timer interrupt and is the same as the switching period of the main switch.
The duty cycle of S 1 , d, is superimposed by a perturbed signal d p , which is defined as where A p is the magnitude of the perturbation.
The duration of the perturbation lasts 2NT int . After engaging the framework, v o starts increasing at n = 0, where d p is changed from zero to A p . Conversely, v o starts reducing at n = N , where d p is changed from A p to zero. The magnitude of A p is chosen such that the peak output voltage overshoot V o,max and undershoot V o,min should be less than that defined in the specification of the converter. Therefore, when S sel = 1,   of L and C without using extra sensors to measure the voltage and current waveforms associated with the components. Moreover, the transfer function G f in (1) exhibits a low-pass filtering function that can attenuate high-frequency noise in v o . To make the signature recognize easily, the values of K p,f and K i,f are chosen, such that the entire converter is operated in a lightly damped condition, i.e., with damping factor close to zero. It is worth noting that the stability of the proposed framework can be confirmed as long as K p,f and K i,f are designed according to the phase margin under all possible combinations of inductance and capacitance [24], [25], [26].

III. PARAMETER IDENTIFICATION USING LSTM
Two LSTM networks having the output using regression layer and classification layer, respectively, are described. The regression model is used to estimate the value of each filter component. It is suitable for providing the default controller with information to adjust control parameters. The classification model is used to classify the value of each filter parameter to a predefined bin. The number of bins and bin size are userdefined. It is suitable for one to estimate the condition of the output filter. Fig. 3 shows the LSTM network utilizing a sequence-toone model [27]. It consists of an input layer, a hidden layer, and an output layer. The hidden layer contains LSTM cell.

A. PRE-PROCESSING OF SAMPLED DUTY CYCLE
The sampled duty cycle is filtered, down-sampled, normalized to the steady-state characteristics and performed differential from the time series of the best-fit averaged duty cycle response in the training process. The duty cycle d f is firstly filtered by a moving average filter. The time series of the averaged duty cycled f where W df is the window size of the moving average filter.
Since the natural frequency of the output filter is significantly less than the switching frequency, the input to the LSTM network is down-sampled. The averaged duty cycle is taken for the network in every W DS samples. The time series of the down-sampled duty cycled f ,DS is for 0 ≤ n ≤ 2N -1. Thus, the data length of the time series of d f andd f equals 2 N W DS .d f ,DS is then normalized by multiplying the ratio between the nominal steady-state duty cycle D norm in the training process and the actual steady-state duty cycle D act in the operation. The normalized duty cycled f ,DS,norm is Its difference, denoted byd f , from the time series of the averaged duty cycled train , which is obtained from the a-th 14732 VOLUME 11, 2023 dataset in the training process, is calculated bŷ The a-th dataset is chosen to be the one that has minimum Euclidean distance betweend f ,DS,norm andd train among ρ datasets in the training process. Mathematically,

B. ALIGNMENT OF DATA FOR THE INPUT LAYER OF THE LSTM
The input layer records the time series ofd f at every step. However, to assist the LSTM to extract the features from the input data effectively, the input data x is composed of aligned time series ofd f under different perturbations. As shown in Fig. 4, each perturbation section ofd f , as defined in (3), is considered as an individual feature. Based on Fig. 4, there are perturbations, including ''Perturbation 1'' with d p = A p and ''Perturbation 2'' with d p = 0. The input data x is constructed as a matrix with each row corresponding to one feature. Mathematically,

C. OPERATION OF THE LSTM CELL
The structure of the LSTM cell is shown in Fig. 5 where and w * and b * represent the network weight and bias, respectively. For example, w xi is the weight of x for σ i , w hi is the weight of h for σ i , and b i is the bias of σ i . Element-wise sigmoid functions σ * is used as an activation function of each gate, hyperbolic tangent tanh is used as an input activation and cell state activation and ⊙ is an element-wise product. Training of w * and b * is described in Sec. IV.

D. OUTPUT LAYER
After obtaining the output of the LSTM cell at the last step, i.e., h[N ], the output of the framework is determined by two possible output layers. They are regression layer and classification layer. Their corresponding models are described as follows.

1) USING REGRESSION LAYER
The output of the regression model, y r , is calculated by where w r and b r are the weight and bias of the output layer of the regression model, respectively. The dimension of y r equals the number of filter parameters to be identified. For example, for a second-order LC filter, the dimension of y r is two with one of its elements representing the value of L and another element representing the value of C.

2) USING CLASSIFICATION LAYER
For classification model, the dimension of the output vector a is firstly transformed by the following equation, where w c and b c are the weight and bias of the classification layer, respectively. The dimension of a equals the number of bins in the classification model. Softmax function is used to calculate the probability of the estimated parameter in each bin. The probability of the i-th bin y c,i is calculated by   where a i is the i-th element of a and Q is the total number of bins.
The bin with the highest probability is the one that the considered filter parameter is classified. Fig. 6 shows the allowable variation of a filter parameter, ranging from -Z % to +Z %. The range is divided into seven sub-ranges. Each sub-range is called ''bin''. Depending on the classification requirement, the bin size can be evenly or unevenly distributed. Fig. 6 shows an example that the bin size around the nominal value is smaller than the one away from the nominal value. Thus, the system will recognize slight change or degradation of the filter parameters at the early stage. It is considered as the point of interest (POI).
If the value of a filter parameter is close to the boundary of two adjacent bins, the estimated probabilities falling into the two bins are similar and thus the classification becomes uncertain. A post-processing technique is used. Q-1 fictitious bins, namely F-Bins, are introduced to represent filter parameter values around bin boundaries. The considered filter parameter will be classified into bin b new with the following criteria (22), as shown at the bottom of the page, where b orig is the original bin predicted without post-processing and α is the threshold to perform the post-processing. The performance of the post-processing technique is evaluated by comparing the predicted bin b new (using eq. (22)) with the actual bin and/or actual F-Bin. Table 3 shows the evaluation criteria. The principle is that the evaluation is correct if the calculated bin number, no matter originally defined bin or fictitious one, correctly matches with the corresponding bin type.

E. ROBUSTNESS AGAINST SAMPLING NOISE
After Considering that sampling noise might influence the estimation performance of the LSTM network. As discussed in [29] and [30], input-to-state stability (ISS) can be used to verify the robustness of the trained LSTM network against sampling noise. With the LSTM network defined in (11)-(18), the ISS of a trained LSTM network can be assured if the following inequalities are met, 14734 VOLUME 11, 2023 where σ denotes the sigmoid activator defined in (17) and |·| n denotes the induced n-norm of the matrix.

IV. NETWORK TRAINING
The LSTM network is trained by using a ROF that can be programmed to give different combinations of the values of the output inductance and output capacitance. The parameters in the LSTM network are trained by a stochastic-gradientdescent-based algorithm [31].

A. ROF
As shown in Fig. 7, the ROF is constructed by an inductor bank (IB) and a capacitor bank (CB). The IB consists of several series-connected basic units, namely (IBBUs), with a series inductor L F . The CB consists of two series-connected capacitor bank basic units (CBBUs) and a fixed capacitor C F connected across them. The numbers of IBBU and CBBU used in the IB and CB, respectively, depends on the required resolution of the combinations. Each IBBU is constructed by K (or K + 1) relays, S L,1 , S L,2 , . . . , S L,K , and/or S LP and K fixed-value inductors, L 1 , L 2 , . . . , L K . A relay is connected in series with each inductor, forming a branch between the two terminals of the IBBU. S LP is connected between the two terminals of the IBBU.
Each CBBU is constructed by M (or M + 1) relays, S C,1 , S C,2 , . . . , S C,M , and/or S CP and M fixed-value capacitors, C 1 , C 2 , . . . , C M . A relay is connected in series with capacitor, forming a branch between the two terminals of the CBBU. S CP is connected between the two terminals of the CBBU.
With multiple IBBUs and CBBUs connected in series, the number of combinations of equivalent output inductance and capacitance can be increased. An output filter with different frequency characteristics can then be configured. Fig. 8 illustrates how the IB and CB are connected to a buck converter. The state of the relays is controlled by a low-speed microcontroller because the operating frequency of the relays is very low.

B. TRAINING OF THE LSTM NETWORK
The time responses of the duty cycle upon subjecting the same perturbation method, as depicted in Sec. III, under different combinations of the output filter parameters with the ROF and the output powers are conducted to train, validate, and test the LSTM network parameters in Fig. 5. Let N LC be the number of combinations of the output filter parameters and ρ be the number of output power levels in the entire training process. The time responses are divided into ρ datasets. Each dataset represents the system response at one power level. It contains N LC time series of the duty-cycle response d f corresponding to N LC output filter parameter combinations. In other words, there are N LC × ρ time series. They are randomly classified into three sets. 70% of the time series is used to train the LSTM network, 20% of the time series is used to validate the hyperparameters, and 10% of the time series is used to test the LSTM network.
Let N T be the number of time series chosen for the training in one batch. At the start of the training process, the network VOLUME 11, 2023 parameters are randomly chosen. Then, they are determined through a stochastic-gradient-decent-based algorithm. The procedure is shown in Fig. 9. By using the ROF, D norm is determined and d f are collected and divided into ρ datasets.
For each d f ,d f ,DS,norm are calculated by (5)- (8). Then, ρd train are obtained. No matter in the training, validation or testing, each d f is being evaluated with the following procedure. First, D act is taken from the steady-state duty cycle before the framework is engaged.d train is chosen among ρ datasets, where the a-th dataset is selected by using (9). Then, the network input x is formulated with (5) - (8) and (10), as described in Sec. III-A and III-B. Finally, the output filter parameters are estimated with the LSTM using (11)- (22).
During the training process, the estimated output filter parameters are compared to the actual filter parameters. The difference is called ''loss''. For the regression model, the loss ℓ of the N T estimated output filter parameters is defined as where R is the number of components in the output filter, and y r,e,i,j and y r,t,i,j are the estimated and actual value of the j-th component value associated with the i-th time series, respectively. y r,e,i,j is obtained by (19).  For the classification model, the loss ℓ is defined as where y c,e,i,j is the estimated probability of the j-th bin of the i-th time series, and y c,t,i,j is the actual probability of the j-th bin of the i-th time series. y c,e,i,j is obtained by (21). y c,t,i,j is either 0 or 1. It is equal to 1 if the j-th bin is the actual class of the i-th time series.
The network parameter set θ is updated by using gradient descent algorithm where λ is the learning rate, and θ ∈ {w xi , w hi , w xf , w hf , w xo , After an epoch has finished, the entire training set is reused for the next epoch. The training process is terminated if the loss remains over a few epochs, implying that the network performance has been maximized. The entire validation set is evaluated after a number of epochs. The loss obtained with the validation set is used to 1) monitor the training process, 2) prevent the network from overfitting the training data, and 3) tune the hyperparameters used in the training process. Upon completing the training process, the performance of the framework is evaluated with the testing set. The training process will be terminated manually if the losses obtained with the training set and the validation set diverge.

V. EXPERIMENTAL VERIFICATIONS
The effectiveness of the proposed framework has been evaluated on a 240W, 100V / 48V DC-DC buck converter. Two types of output filters have been tested. They are secondorder and fourth-order filters. Fig. 10(a) and 10(b) show the schematics of the converter with the two types of filters, respectively. The nominal values of the components for the two types of filters are tabulated in Table 4.
The default controller is of Type-II and the control parameters are given in Table 2. The switching frequency and the sampling frequency of the controller are both 100kHz. Fig. 11 shows the photo of the ROF, which was built in the laboratory. Each IBBU consists of multiple fixed-value inductors and CBBU consists of multiple fixed-value capacitors. Two IBBUs and two CBBUs are used. The nominal (Nom), minimum (Min) and maximum (Max) values of the inductors and capacitors in the output filters in Fig. 10(a) and (b) are tabulated in Table 4.
The training process is performed on a computer using AMD Ryzen 3950X CPU and Nvidia RTX3080Ti GPU, the average training time is less than 13 minutes for the second-order output filter and less than 60 minutes for the fourth-order output filter. The network and training parameters are given in Table 5. During the training process, 64 samples are randomly sampled as a batch to update the network and the entire training set is processed 100 times (Epoch).
After the training process on a PC is completed, the weights and parameters in the LSTM network are downloaded to the digital controller to control the PCS and identify output filter parameters. Since they have already included all operational conditions in the PCS, it is unnecessary to train the LSTM network for individual converters. Thus, in the following discussions, the resources used in the training process are not 14736 VOLUME 11, 2023  included. The additional computational resources, including storage space and computation time, required to perform the proposed parameter estimation framework (''framework'') on the digital controller are briefly given in Table 6.
Nowadays, microcontrollers contain a lot of memory. For example, STM32F4 has 512kB flash memories. The framework only occupies 0.93% (4.768 / 512) of the memory space in executing the framework for a converter with LC filter, and 3.62% (18.54 / 512) of memory space in executing the framework for a converter with LCLC filter. It can be easily integrated into a microcontroller. Even if the operation of the framework is non-time critical, it only takes 9.3ms and 320ms to extract LC and LCLC filter parameters, respectively, in the LSTM computation, and the two perturbations take 8ms. The overall time taken to execute the framework is 17.3ms and 328ms to extract LC and LCLC filter parameters, respectively.
By substituting the trained weights and bias of the LSTM obtained in the experiment into (23) and (24), both inequalities are satisfied. It implies that if the sampling noise is bounded, the state of the LSTM would also be bounded and global asymptotic stability, demonstrating the robustness of the trained LSTM network to sampling noise. Fig. 12 shows the waveforms of the output voltage v o and inductor current i L with the second-order output filter. Fig. 13 shows the waveforms of capacitor voltage v C 4,1 , the    for the converter with the second-order output or 2% for the converter with the fourth-order output filter. Then in the time interval [t 2 , t 3 ], the injected signals are removed. It causes v o having a peak-to-peak variation of 2.38V and 3.0V, respectively for the two types of output filters. The performance in extracting the two types of output filters is given below.

A. SECOND-ORDER OUTPUT FILTER 1) REGRESSION
The results are obtained from evaluating the testing set that includes datasets of the output power changing from 20% of the rated power to 100% of the rated power and input voltage changing from 80V to 120V. The histograms of the normalized root-mean-square error (NRMSE) of each sample in the testing set in 20 runs are shown in Fig. 14. Table 7 shows the mean RMSE, NRMSE, and the 95% confidence interval (CI) of the results. The results show that the NRMSE of the estimated values of L 2,1 and C 2,1 are 2.35% and 2.25%,

2) CLASSIFICATION
The statistical populations of the F1-scores of the L 2,1 classifiers and the C 2,1 classifiers trained in 20 runs are shown in Fig. 15. Identical to the regression model, the testing set contains the same output power and input voltage variations. Table 8 shows the mean F1-scores and the 95% CI of the results. The results show that the F1-scores of the L 2,1 classifiers and the C 2,1 classifiers are 0.805 and 0.815, respectively. By introducing the F-Bin, the F1-scores of both L 2,1 classifiers and C 2,1 classifiers are improved by more than 0.078.

B. FOURTH-ORDER OUTPUT FILTER
For the fourth-order output filter, regression models are trained to estimate the four filter parameters. It is because, on the one hand, the capabilities of the proposed framework have been demonstrated in the second-order case. On the other hand, the regression model provides better understanding of the filter parameter. The histograms of the NRMSE of each output filter parameter are shown in Fig. 16. Table 9 shows the mean RMSE, NRMSE, and the 95% CI of the results. The results show that the NRMSE of the estimated values of L 4,1 , L 4,2 , C 4,1 and C 4,2 are 0.28%, 2.04%, 6.49% and 2.80%, respectively, which is generally acceptable except there is room for improvement in estimating C 4,1 .

VI. CONCLUSION
A ''plug-and-play'' tiny artificial-intelligence-empowered output filter parameter extraction framework for digital power has been presented. The framework is a software module that can be incorporated into a digital controller to extract output filter parameters without adding extra sensors or modifying the circuit. An LSTM network extracts the filter parameters by recognizing the response of the duty cycle through the framework after subjecting a perturbation into the duty cycle. Detailed descriptions on the pre-processing of sampled duty cycle, alignment of data for the input layer of the LSTM, the operation of the LSTM cell, the training of the LSTM network, and method to test the robustness of the LSTM against sampling noise have been given. The proposed framework has been tested and evaluated successfully on a 240W 100V/48V buck DC/DC converter with its output filter configured as either a second-order LC or a fourth-order LCLC filter. Regression models and classification models are both trained to estimate the filter parameter with acceptable results in practice. The proposed framework can be applied in industrial applications where the online identification of filter parameters is required especially if no additional sensors are expected to be added. These applications include but are not limited to fault diagnosis, aging monitoring, model predictive control, EMI design, etc. In the future, the proposed framework can be further improved by incorporating software simulator training instead of building the ROF for data collection.