Knowledge distillation-based lightweight deformable neural network battery life prediction method

Abstract

The invention relates to a battery life prediction method of a lightweight deformable neural network based on knowledge distillation, and belongs to the field of artificial intelligence. According to the method, efficient and accurate prediction is realized through three innovative modules: firstly, a student network of deep deformable convolution is designed to dynamically capture battery degradation characteristics, and a channel attention mechanism and channel shuffling operation are combined; secondly, proposing a diffusion knowledge distillation framework, and realizing efficient knowledge migration of the heterogeneous teacher network and the student network through a noise adding-denoising process; and finally, introducing a game-based mutual distillation strategy to improve the student network generalization ability. The performance of the model is effectively enhanced, the efficiency of the model is improved, compared with a traditional method, the number of model parameters is effectively reduced, and the reasoning time of the model is shortened on the premise that precision is not lost.

Description

Technical Field

[0001] This invention belongs to the field of artificial intelligence and relates to a battery life prediction method based on a lightweight deformable neural network using knowledge distillation. Background Technology

[0002] In recent years, deep neural networks (DNNs) have made significant progress in predicting remaining battery life. The main reason for DNN's great success is its ability to handle massive amounts of data and manage billions of model parameters. However, these heavy-duty DNNs typically require substantial computational and memory capacity, limiting their application in resource-constrained edge devices such as vehicle battery management systems (BMS). A widely adopted solution is to develop lightweight models that are simple and effective.

[0003] In the field of battery remaining life prediction, designing lightweight models is an emerging research direction, mainly focusing on three aspects: manual design of lightweight models, Neural Architecture Search (NAS), and model compression. Manual design requires deep expertise, which limits the widespread development of lightweight models. NAS provides a method for automatically optimizing network architecture, but it requires significant computational resources during the initial training phase. In contrast, model compression is a feasible method that can reduce model size and computational requirements without significantly degrading performance. Among various model compression techniques…

[0004] Knowledge distillation stands out for its ability to transfer knowledge from large, complex teacher networks to small, efficient student networks. Besides improving the performance of neural networks, knowledge distillation simplifies the training process and reduces reliance on large amounts of computational resources. Based on knowledge distillation, we propose a lightweight framework for predicting the remaining lifespan of vehicle power batteries. This framework trains a compact yet powerful student network, aiming to improve learning efficiency and prediction accuracy. However, directly utilizing knowledge distillation to design lightweight neural networks presents two main challenges.

[0005] (1) Difficulty in knowledge transfer: The key to knowledge distillation lies in transferring knowledge from the teacher network to the student network by aligning output features. Teacher networks are typically large and complex, having undergone extensive training, while student networks are designed to be smaller and more efficient. High-level features in the teacher network are often difficult to transfer directly to the simpler student network, leading to information loss and reduced effectiveness. Therefore, state-of-the-art knowledge distillation methods primarily focus on narrowing the differences in feature extraction between the teacher and student models.

[0006] (2) The knowledge lag problem of distillation models: In existing research, it is quite challenging for distillation models to maintain their advanced status. Specifically, as basic modeling techniques develop, teacher models must be constantly updated to maintain their leading position. This means that student models also need to be updated and redistilled regularly to maintain their performance. Summary of the Invention

[0007] In view of this, the purpose of this invention is to provide a battery life prediction method based on a lightweight deformable neural network of knowledge distillation, which extracts features from battery charge and discharge data and combines them with deep learning technology to achieve an effective solution for predicting the life of automotive batteries.

[0008] To achieve the above objectives, the present invention provides the following technical solution:

[0009] A battery lifetime prediction method based on a lightweight deformable neural network using knowledge distillation, comprising the following steps:

[0010] Obtain the charging and discharging dataset of vehicle power batteries and preprocess the dataset;

[0011] A student network is established, the sampling position of the convolution kernel is dynamically adjusted through deformable depthwise convolution, and the feature map is redistributed through channel shuffling. Global information is extracted using the channel attention mechanism, and channel description vectors are generated by global average pooling.

[0012] A teacher network is established, and key features of various battery degradation data are extracted through pre-trained LSTM, GRU, TCN, BiLSTM and Transformer networks. Knowledge distillation technology is used to transfer the knowledge of the teacher network to the student network, providing the student network with learning benchmarks and goals.

[0013] A diffusion model is used to add and remove noise to the student and teacher models. The output features of the teacher network are progressively added to the forward direction, while the student network learns to remove noise in the reverse direction. The output feature distribution of the student and teacher networks is realized by minimizing the maximum distribution difference loss.

[0014] The two student networks are trained by simulating a "multi-expert voting mechanism" through mutual distillation, minimizing the difference between the output features of the two student networks to learn from each other;

[0015] The above optimization training was performed, and the trained student model was used for real-time battery life prediction.

[0016] Furthermore, a publicly available battery cycle charge-discharge dataset is used as the charge-discharge dataset for vehicle power batteries. The dataset contains charging data for several electric vehicle battery packs, including voltage, charging current, maximum battery voltage, minimum battery voltage, maximum temperature, minimum temperature, available energy, and available capacity. At the same time, the dataset is numbered based on each electric vehicle.

[0017] The preprocessing of the dataset includes label capacity calculation and dataset partitioning; specifically, in the label capacity calculation process, the ampere-hour integral method is used to calculate the label capacity from the known data, and then normalization is performed.

[0018]

[0019] Where t1 and t2 are the start and end times of the selected charging segment. The average current is given by Δt, the data sampling frequency is given by SOC, and the battery state of charge is given by SOC.

[0020] Splitting the dataset refers to randomly dividing the preprocessed battery data into training and test sets according to a preset ratio.

[0021] Furthermore, in the student network, the process of dynamically adjusting the sampling position of the convolution kernel through deformable depthwise convolution is as follows:

[0022] Set input feature map The dimensions are T×F×C in Where T represents the time step, F is the number of features, and C in =n is the number of channels;

[0023] Apply deformable depthwise convolution to compute the output feature map At position p k Value at: in the grid Increase offset by {Δp} k |k=1,...,K}, where For each channel (n = 1, ..., N), apply convolution kernels independently, and pass them through offset Δp k Adjusting the position yields:

[0024] Δp k =OffsetConv(u ij )

[0025] Where OffsetConv(.) is a small convolutional neural network that predicts offsets, u ij ∈U is a sample of battery data; for each position p k Through offset Δp kTo adjust the position of the convolution kernel, at each adjusted position p = p0 + p k +Δp k Above, the feature map is represented as:

[0026]

[0027] for If the adjusted position p has non-integer coordinates, then for any Calculated using bilinear interpolation Eigenvalues:

[0028]

[0029] Where G(·,·) is the bilinear interpolation kernel, and q represents the input feature map. Integer position in This represents the set of all such locations surrounding p;

[0030] Then, G(q,p) is decomposed into two one-dimensional kernels g(q) x ,p x ) and g(q y ,p y If ), then the following relationship is satisfied:

[0031] G(q,p)=g(q x ,p x )·g(q y ,p y )

[0032] Where q x and q y q is the coordinate of the integer position, while p x and p y These are the coordinates of the adjusted position p in the input and output feature maps, respectively.

[0033] In conclusion, we have:

[0034]

[0035] After processing with deformable depthwise convolution, the feature maps input by the student network can adapt to changes in the teacher network's data, and the feature maps can more accurately represent key information.

[0036] Furthermore, in the student network, the process of redistributing feature maps through channel shuffling is as follows:

[0037] Given an input feature map First, its channel number is divided into C in Divide into G groups, each group containing Channel. (Set) Each of them It is a set of channels;

[0038] Through channel shuffling, features from different groups are mixed together to generate a single output feature map. The number of channels remains constant, but each channel feature contains information from other groups:

[0039]

[0040] in This indicates the shuffle operator applied to the i-th channel group;

[0041] Simultaneously, through an adaptive channel selector, global information is extracted from each channel of the feature map output by the multi-scale feature extractor using global average pooling:

[0042] y G =vec(GAP(y))

[0043] in This represents a vectorization operation, which flattens a 1×1×N tensor into an N-dimensional vector; The GAP operation applied to y is represented as:

[0044]

[0045] In the formula, The resulting feature vector contains global information for each channel in the output feature map y;

[0046] Then, the extracted feature vector y G Nonlinear transformation and scaling are performed using two fully connected layers and a sigmoid activation function:

[0047] S=W2·λ(W1·y G +b1)+b2

[0048] in, and These are the weight matrices of two fully connected layers, where This represents the compressed channel dimension, where r is a compression ratio;

[0049] The final output features of the student network are represented as follows:

[0050] y←y⊙Sigmoid(S)

[0051] Here, ⊙ represents element-wise multiplication, applying the scaling factor to each channel.

[0052] Furthermore, in the teacher network, LSTM, GRU, TCN, BiLSTM, and Transformer networks are used for sequence modeling. LSTM captures long-term dependencies in sequence data; Gated Recurrent Unit (GRU) captures long-term dependencies in sequence data based on a simpler architecture; Temporal Convolutional Network (TCN) models long-range temporal dependencies through causal convolution; Bidirectional Long Short-Term Memory (BiLSTM) network simultaneously captures past and future contextual information in sequence data; and Transformer utilizes a self-attention mechanism to model long-range dependencies with higher parallelization efficiency. These network models are used to extract temporal features from battery data and provide knowledge to the student network.

[0053] Furthermore, in the knowledge transfer process between teacher and student networks, a nonlinear mapping function is first identified. The nonlinear mapping function takes the battery sequence as input and outputs the desired capacity sequence Y, while minimizing the mapping error of the input data sample U, and the data sample U follows a natural distribution.

[0054] Nonlinear mapping function Configure a set of structured hyperparameters Θ and employ a learning algorithm. In a limited battery dataset Through iterative optimization, an optimal hyperparameter Θ is eventually found. * Make the loss function If minimized, the prediction paradigm for the teacher network and student network is defined as follows:

[0055]

[0056] In the formula, Indicates hard loss;

[0057] In a knowledge distillation process, knowledge from a pre-trained teacher network is transferred to a student network given a teacher network Y. T And student network Y S When the output is obtained, the KD loss of knowledge distillation is defined as:

[0058]

[0059] in, Indicates soft loss. Let represent the distance function that measures the difference between the outputs of the teacher network and the student network, and let be the mean squared error of the regression output.

[0060] Furthermore, in the knowledge transfer process between teacher and student networks, a knowledge transfer technique based on a diffusion model is introduced, the process of which is as follows:

[0061] Treating the output of the student network as a noisy version of the teacher network output, for the input sample Battery sequence data was used to simultaneously train both the teacher and student networks to extract features. and

[0062] Then, during the forward noise process, repeatedly to Add Gaussian noise:

[0063]

[0064] Where, q θ This indicates the diffusion process of the teacher network. This represents the primitive characteristics of the teacher network at the initial time step. The transition noise characteristics at time step t This indicates the use of the noise variance scheduling scheme β. s The sign of the feature is directly sampled at any time step;

[0065] In the given conditions, The distribution follows a Gaussian distribution.

[0066]

[0067] but The mean is The covariance matrix is

[0068] Gaussian distribution rewritten as If it is equal to the mean plus noise, then it represents... for:

[0069]

[0070] Where, ∈ t It's Gaussian noise.

[0071] Introduce a scaling factor This makes it have the desired variance. Therefore:

[0072]

[0073] in Standard Gaussian noise, i.e.

[0074] final, Represented as and noise variables Linear combination:

[0075]

[0076] For the teacher network, a nonlinear mapping function is trained. To predict Equivalent to The noise in the data, i.e., minimizing the L2 loss between them:

[0077]

[0078] For student networks, the initial characteristics are: By using nonlinear mapping functions Perform an iterative denoising process to reconstruct features.

[0079]

[0080] in Indicates transition variance. Represents the conditional probability of a student, given features at time t. Below, student characteristics The conditional probability at time t-1;

[0081] Modeled as a mean The Gaussian distribution, where It is a function defined by parameter θ, used to extract noisy characteristics from time step t. And predict clean features The variance of the Gaussian distribution is It represents the noise level at time t; and I is the identity matrix, indicating that the noise is independent and identically distributed.

[0082] The characteristics of teacher networks are represented by time. Gaussian noise was gradually added to obtain the data; the student network was obtained by gradually adding Gaussian noise over time. Starting with the noisy features, we learn to remove the noise in reverse order to recover the clean features.

[0083] Furthermore, the output features generated by the student network are considered as a noisy version of the features generated by the teacher network, and the noise level representing the difference between the features of the teacher network and the student network is unknown.

[0084] Therefore, the maximum mean difference method is used to align the feature distributions of the noisy teacher network and the noisy-free student network. Maximum mean difference is a distribution metric based on kernel embedding, defined as:

[0085]

[0086] in, Let q and p represent the feature space of kernel function k, respectively, and let k(·) denote the application of kernel function k to the feature vector. and

[0087] Let features and Implicitly mapped to through kernel representation In this context, k is chosen as the Gaussian kernel function, i.e.:

[0088]

[0089] Where σ represents the kernel bandwidth parameter;

[0090] Output characteristics of noisy teacher network Denoising Student Network Output Features Consider them as two independent distributions, i.e., q θ and p θ Calculate the mean of the feature maps of the two distributions, and use the maximum mean difference to measure the difference between them. Minimize the difference to push the output feature distribution of the student network closer to the output feature distribution of the teacher network. Then:

[0091]

[0092] Kernel functions are used to measure the differences between different distributions;

[0093] In the feedforward noise process, the diffusion model is first trained using the feature outputs of the teacher network, and noise is gradually added to generate a series of noise features. In the denoising process, a diffusion model is trained to learn a nonlinear mapping function. Thus, it is possible to obtain noise characteristics The original characteristics were restored in the middle. feature This represents the denoised version, obtained through a trained model. Recovered It is applied to noisy features The output of the post-model yields the original clean features at approximately time step t. Therefore, the reconstruction loss is defined as:

[0094]

[0095] In this process, the model is trained to recover clean features and perform a denoising process by minimizing the reconstruction loss.

[0096] It also defines the maximum mean difference loss.

[0097]

[0098] in, and In the regenerating nucleus Hilbert space It plays a role in the process, and k(·) is related to... Related kernel functions, norm Defined as an inner product exist The norm induced in the middle;

[0099] By losing and loss In combination, the knowledge transfer loss from the teacher model to the student model is defined as:

[0100]

[0101] Where λ∈[0,1] are weight coefficients; through reverse gradient training, the diffusion model recovers the original features from the features contaminated by noise, and uses the maximum mean difference to align the feature distribution to minimize the maximum mean difference (MMD) loss, thereby realizing the knowledge transfer between the student network and the teacher network.

[0102] Furthermore, in the mutual distillation process between the two student networks, a game-based mutual distillation strategy is adopted, the process of which is as follows:

[0103] Two teacher networks T1 and T2 are randomly selected, and two student networks S1 and S2 are trained using an MMD-based diffusion model. The two student networks are treated as independent learning entities, and they learn from each other by minimizing the difference between their output distributions.

[0104] The task of the student network is to predict continuous values ​​of battery capacity. The output of the student network is set to... and In the remaining useful life prediction task, S is used to train i The objective function is defined as the prediction capacity. With tag capacity y (j) Mean square error between:

[0105]

[0106] in It is the prediction of the i-th student network for the j-th sample, and M is the total number of samples;

[0107] If the objective is to minimize the difference between the predictions of the two student networks, then:

[0108]

[0109] For each student network, the final loss function considers both its own error relative to the true value and its mutual learning with other student networks. Therefore, the combined loss for each network is:

[0110]

[0111] Here, φ∈[0,1] is a hyperparameter used to control the trade-off between minimizing prediction error and maximizing consistency between student networks.

[0112] Furthermore, during the training of the above process, the hyperparameters of the teacher network are configured using a Bayesian hyperparameter optimization algorithm, while the hyperparameters of the student network are fine-tuned during the knowledge distillation process; the trained model is then used for real-time prediction of remaining battery life.

[0113] The beneficial effects of this invention are as follows:

[0114] The battery life prediction method based on a lightweight deformable neural network driven by knowledge distillation proposed in this invention significantly improves the accuracy and efficiency of battery life prediction through three innovative mechanisms: a lightweight student network, a knowledge distillation framework, and a deformable feature extraction module.

[0115] The deformable deep convolutional module, composed of a dynamic offset prediction network, adaptive sampling computation units, and a channel attention mechanism, is responsible for extracting local abrupt changes and global degradation trends, and forming a highly discriminative representation through feature fusion. This design can comprehensively capture key patterns in the battery degradation process from both temporal and spatial dimensions, enhancing the model's adaptability to non-uniform degradation data. By dynamically adjusting the receptive field and feature weights, the model can more accurately locate nodes of abnormal changes in battery performance, providing more robust feature representations for subsequent lifetime prediction, thereby significantly improving the stability and accuracy of predictions.

[0116] The knowledge distillation framework adopts a three-stage architecture of "multi-teacher collaboration - dynamic transfer - game optimization". It constructs a multi-scale feature knowledge base through a heterogeneous teacher network, innovatively introduces a diffusion model to achieve noise-robust feature transfer (Diffusion-based Knowledge Distillation, Diff-KD), and designs a two-student game mechanism to enhance the model's generalization ability. This significantly breaks through the performance bottleneck of traditional distillation methods in time-series prediction tasks. Its core innovation lies in combining the theoretical advantages of the diffusion model with a lightweight network design, providing a high-precision, low-latency battery life prediction solution for edge computing scenarios.

[0117] Other advantages, objectives, and features of the invention will be set forth in part in the description which follows, and in part will be apparent to those skilled in the art from the following examination, or may be learned from practice of the invention. The objectives and other advantages of the invention can be realized and obtained through the following description. Attached Figure Description

[0118] To make the objectives, technical solutions, and advantages of the present invention clearer, the preferred embodiments of the present invention will be described in detail below with reference to the accompanying drawings, wherein:

[0119] Figure 1 This is a schematic diagram of the overall process of the battery life prediction method based on knowledge distillation of a lightweight deformable neural network according to an embodiment of the present invention.

[0120] Figure 2 This is a schematic diagram of the overall architecture of the battery life prediction method based on knowledge distillation of a lightweight deformable neural network according to an embodiment of the present invention.

[0121] Figure 3 This is a schematic diagram illustrating the knowledge transfer process between the teacher network and the student network according to an embodiment of the present invention;

[0122] Figure 4 This is a schematic diagram of a game-based mutual distillation structure according to an embodiment of the present invention. Detailed Implementation

[0123] The following specific examples illustrate the implementation of the present invention. Those skilled in the art can easily understand other advantages and effects of the present invention from the content disclosed in this specification. The present invention can also be implemented or applied through other different specific embodiments, and various details in this specification can be modified or changed based on different viewpoints and applications without departing from the spirit of the present invention. It should be noted that the illustrations provided in the following embodiments are only schematic representations of the basic concept of the present invention. Unless otherwise specified, the following embodiments and features can be combined with each other.

[0124] The accompanying drawings are for illustrative purposes only and are schematic diagrams, not actual pictures. They should not be construed as limiting the invention. To better illustrate the embodiments of the invention, some parts in the drawings may be omitted, enlarged, or reduced, and do not represent the actual product dimensions. It is understandable to those skilled in the art that some well-known structures and their descriptions may be omitted in the drawings.

[0125] In the accompanying drawings of the embodiments of the present invention, the same or similar reference numerals correspond to the same or similar components. In the description of the present invention, it should be understood that if terms such as "upper," "lower," "left," "right," "front," and "rear" indicate the orientation or positional relationship based on the orientation or positional relationship shown in the drawings, they are only for the convenience of describing the present invention and simplifying the description, and do not indicate or imply that the device or element referred to must have a specific orientation, or be constructed and operated in a specific orientation. Therefore, the terms used to describe positional relationships in the drawings are only for illustrative purposes and should not be construed as limiting the present invention. For those skilled in the art, the specific meaning of the above terms can be understood according to the specific circumstances.

[0126] Please see Figures 1-4 This is a battery life prediction method based on a lightweight deformable neural network using knowledge distillation.

[0127] This embodiment describes in detail the process of the battery lifetime prediction method based on knowledge distillation and lightweight deformable networks of the present invention, such as... Figure 1 As shown, it specifically includes:

[0128] S1. Obtain the charging and discharging data of the vehicle power battery (such as voltage, current, temperature, etc.) and preprocess the data.

[0129] S2 provides a student network that dynamically adjusts the sampling position of the convolution kernel through deformable depthwise convolution to adapt to changes in teacher features. Through channel shuffling, feature maps are redistributed to promote cross-channel information sharing. Global information is extracted through a channel attention mechanism to obtain global average pooling (GAP) to generate channel description vectors, which strengthens the battery degradation features with strong correlation and suppresses irrelevant noise.

[0130] S3. Select LSTM, GRU, TCN, BiLSTM, and Transformer as teacher networks. After pre-training, these networks can extract key features of battery degradation data and generate high-quality prediction results. Through knowledge distillation technology, the knowledge of the teacher network is transferred to the student network, providing the student network with learning benchmarks and goals.

[0131] S4. Through knowledge distillation, the teacher features are subjected to a forward progressive noise-adding process, and the student network is subjected to a learning reverse noise-de-noising process, so that the output features of the two are effectively aligned. Based on the Maximum Mean Discrepancy (MMD) method, a Gaussian kernel function is designed to measure the overall distribution difference between teacher and student features, so as to minimize the MMD loss and achieve feature distribution alignment between the student network and the teacher network.

[0132] S5. To further improve the performance of the student networks, a Game-based Mutual Distillation (GMD) method is proposed. This method simulates a "multi-expert voting" mechanism, using GMD loss to allow the two networks to mutually guide each other during training. The two student networks learn from each other by minimizing the difference between their output distributions, thereby improving the consistency and robustness of the student network's predictions.

[0133] S6. Optimize and train the above process, and use the training results for real-time battery life prediction.

[0134] In step S1 of this embodiment, a publicly available battery cycle charge / discharge dataset is obtained as the vehicle power battery charge / discharge dataset for this embodiment. In this embodiment, the dataset contains charging data from 20 commercial electric vehicle battery packs, including voltage, charging current, maximum battery voltage, minimum battery voltage, maximum temperature, minimum temperature, available energy, and available capacity. These are labeled #1, #2, #3, ..., #20, respectively. The sampling frequency is 8 seconds. The dataset contains 854,591 rows and 10 columns of data, totaling 1,070 charging cycles.

[0135] The preprocessing of the dataset includes label capacity calculation and dataset partitioning. Data from electric vehicles #1 to #16 in the power battery dataset is used as the training set, while data from electric vehicles #17 to #20 is used as the validation set. To ensure interpretability of the analysis and focus on the most relevant feature relationships, this embodiment uses the XGBoost method for feature selection, followed by cluster analysis. The top 5 vehicles from the training set (electric vehicles #1 to #16) with the most similar feature patterns to the vehicles in the validation set (electric vehicles #17 to #20) were identified, thereby reducing noise, highlighting the most important relationships, and striking a balance between prediction accuracy and model complexity. Specifically, in the label capacity calculation process, the ampere-hour integral method is used to calculate the label capacity from the known data and then normalize it.

[0136]

[0137] Where t1 and t2 are the start and end times of the selected charging segment. The average current is given by Δt, the data sampling frequency is given by SOC, and the battery state of charge is given by SOC.

[0138] Splitting the dataset refers to randomly dividing the preprocessed battery data into training and test sets according to a preset ratio.

[0139] In the student network established in S2, this embodiment proposes a lightweight deformable network for knowledge distillation specifically designed for the task of predicting remaining useful life. This embodiment sets the input feature map. The dimensions are T×F×C in Where T represents the time step, F is the number of features, and C in =n is the number of channels. The goal is to use deformable depthwise convolution to compute the output feature map. At position p k The value at that location. In the grid. Increase offset by {Δp} k |k=1,...,K}, where For each channel (n = 1, ..., N), apply convolution kernels independently, and pass them through offset Δp k Adjusting the position yields:

[0140] Δp k =OffsetConv(u ij )

[0141] Where OffsetConv(.) is a small convolutional neural network for predicting offsets, u ij ∈U is a sample of battery data. For each position p k Through offset Δp k To adjust the position of the convolution kernel. At each adjusted position, p = p0 + p k +Δp k The formula becomes:

[0142]

[0143] because The adjusted position p may result in non-integer coordinates. Therefore, for any Calculated using bilinear interpolation Eigenvalues:

[0144]

[0145] Where G(·,·) is the bilinear interpolation kernel, and q represents the input feature map. Integer position in Let G(q,p) represent the set of all such locations around p. Then, decompose G(q,p) into two one-dimensional kernels g(q,p) and g(q,p). x ,p x ) and g(q y ,p y If ), then the following relationship is satisfied:

[0146] G(q,p)=g(q x ,p x )·g(q y ,p y )

[0147] Where q x and q y q is the coordinate of the integer position, while p x and p y These are the coordinates of the adjusted position p in the input and output feature maps, respectively.

[0148] In conclusion, we have:

[0149]

[0150] The above describes the principle of deformable depthwise convolutional units. After processing by deformable depthwise convolution, the input feature map can adapt to changes in data more flexibly, enabling the output feature map to more accurately represent key information.

[0151] Feature channel shuffling is performed in the student network of S2, such as Figure 2 As shown, each channel shuffling redistributes the feature map, thereby promoting information sharing across channels. Given an input feature map... First, its channel number is divided into C in Divide into G groups, each group containing Channel. (Set) Each of them It is a set of channels.

[0152] Channel shuffling rearranges channels from different groups. This operation ensures that features from different groups are blended together to generate a single output feature map. The number of channels remains the same, but each channel now receives information from other groups. This can be represented as:

[0153]

[0154] in This indicates the shuffle operator applied to the i-th channel.

[0155] To effectively reduce model parameters, in step S2 channel attention, we propose an adaptive channel selector that uses global average pooling to extract global information from each channel of the multi-scale feature extractor output feature map. Therefore:

[0156] y G =vec(GAP(y))

[0157] in This represents a vectorization operation, which flattens a 1×1×N tensor into an N-dimensional vector. This indicates the GAP operation applied to y. Specifically, GAP calculates the spatial average value for each channel as follows:

[0158]

[0159] therefore, It is the obtained feature vector, which contains global information for each channel in the output feature map y.

[0160] Then, the extracted feature vector y G Two fully connected layers and a sigmoid activation function are applied for nonlinear transformation and scaling. Therefore:

[0161] S=W2·λ(W1·y G +b1)+b2

[0162] in, and These are the weight matrices of two fully connected layers, where This represents the compressed channel dimension, and r is a compression ratio (usually chosen to make d...). <n)。

[0163] Therefore, we can conclude that:

[0164] y←y⊙Sigmoid(S)

[0165] Where ⊙ represents element-wise multiplication, applying the scaling factor to each channel.

[0166] In the teacher network established in step S3, such as Figure 2 As shown, LSTM, GRU, TCN, BiLSTM, and Transformer are selected as teacher networks. These networks, after pre-training, can extract key features from battery degradation data and generate high-quality prediction results. Knowledge distillation technology is used to transfer the knowledge from the teacher networks to the student networks, providing the student networks with learning benchmarks and goals. Among the teacher networks, LSTM, GRU, TCN, BiLSTM, and Transformer each represent a significant method in sequence modeling. LSTM is renowned for its ability to capture long-term dependencies in sequence data; Gated Recurrent Unit (GRU) has a simpler architecture and comparable performance to LSTM; Temporal Convolutional Network (TCN) excels in modeling long-range temporal dependencies through causal convolution; Bidirectional Long Short-Term Memory (BiLSTM) can simultaneously capture past and future context in sequence data; and Transformer utilizes self-attention mechanisms to model long-range dependencies with higher parallelization efficiency. These mainstream sequence modeling methods are used to extract temporal features from battery data and provide knowledge to the student networks.

[0167] In S4, to achieve efficient knowledge transfer between the student and teacher networks, a knowledge distillation process is used to progressively add noise to the teacher's features, while the student network undergoes a reverse denoising process, effectively aligning the output features of both networks. For example... Figure 3 As shown, when applying teacher or student networks to predict Remaining Useful Life (RUL), a nonlinear mapping function is sought. This function takes a battery sequence as input and outputs a desired capacity sequence Y, while minimizing the mapping error of the input data samples U (e.g., voltage, current, and capacity) from a natural distribution. for Configure a set of structured hyperparameters Θ, and a carefully designed learning algorithm. In a limited battery dataset Perform iterative optimization.

[0168] Eventually, an optimal hyperparameter Θ will be found. * It makes the loss function To reach a minimum. The teacher or student network prediction paradigm for remaining useful life (RUL) can be defined as:

[0169]

[0170] In the knowledge distillation paradigm This is commonly referred to as hard loss. The main idea behind knowledge distillation is to minimize the teacher network. Student Network The distributional differences between outputs transfer the knowledge of the pre-trained teacher network to the student network given the teacher network Y. T And student network Y S The KD loss can be defined as the output of .

[0171]

[0172] in Indicates soft loss. Let represent the distance function that measures the difference between the outputs of the teacher network and the student network. In the remaining useful life prediction task, define . The mean squared error (MSE) of the regression output.

[0173] To narrow the performance gap between teacher and student networks, a new knowledge transfer technique based on a diffusion model is introduced, which uses the diffusion principle to replace the traditional feature alignment method.

[0174] Treating the output of the student network as a noisy version of the teacher network output, for the input sample Battery sequence data was used to train both the teacher and student networks simultaneously, and features were extracted. and Then, the forward noise process will repeatedly... Adding Gaussian noise, we have:

[0175]

[0176] Where q θ This indicates the diffusion process of the teacher network. This represents the primitive characteristics of the teacher network at the initial time step. The transition noise characteristics at time step t This indicates the use of the noise variance scheduling scheme β. s The sign of the feature is directly sampled at any time step. Given... conditions, The distribution follows a Gaussian distribution.

[0177]

[0178] This means The mean is The covariance matrix is

[0179] The Gaussian distribution can be written as Therefore, it can be expressed as for:

[0180]

[0181] Where ∈ t It's Gaussian noise.

[0182] To ensure noise ∈ t With the correct variance, a scaling factor is introduced. This makes it have the desired variance. Therefore, we have:

[0183]

[0184] in Standard Gaussian noise, i.e.

[0185] therefore, It can be represented as and noise variables Linear combination:

[0186]

[0187] For the teacher network, a nonlinear mapping function is trained. To predict Equivalent to The noise in the data, i.e., minimizing the L2 loss between them:

[0188]

[0189] For student networks, the initial characteristics are: By using nonlinear mapping functions Perform an iterative denoising process to reconstruct features.

[0190]

[0191] in Indicates transition variance. Represents the conditional probability of a student, given features at time t. Below, student characteristics The conditional probability at time t-1. This is modeled as a mean of The Gaussian distribution, where It is a function defined by parameter θ, used to extract noisy characteristics from time step t. And predict clean features The variance of the Gaussian distribution is This corresponds to the noise level at time t, and I is the identity matrix, indicating that the noise is independent and identically distributed.

[0192] The characteristics of teacher networks are represented by time. Gaussian noise was gradually added to obtain the result. In contrast, the student network was obtained by gradually adding Gaussian noise over time. Starting with the noisy features, the noise is removed in reverse order to recover clean features.

[0193] The output features generated by the student network can be viewed as a noisy version of the features generated by the teacher network, and the noise level representing the difference between the teacher and student network features is unknown. Therefore, the maximum mean difference method is used to align the feature distributions of the noisy teacher network and the noise-free student network.

[0194] MMD is a distribution metric based on kernel embedding, defined as:

[0195]

[0196] in, Let represent the feature space of the kernel function k, and q and p represent the probability distributions with and without noise, respectively. In the above formula, the term k(·) indicates that the kernel function k is applied to the feature vector. and It is important to note that kernel functions implicitly map the features in the network to a high-dimensional Hilbert space. In this space, the kernel function k takes effect. Although the characteristics and These features are all obtained directly from the network, but the kernel trick ensures that they are treated as Hilbert space. The elements in the map. Specifically, the choice of kernel function determines the properties of the mapping, assuming the characteristics... and This kernel representation implicitly maps to middle.

[0197] Choose k as the Gaussian kernel function, i.e.

[0198]

[0199] Where σ represents the kernel bandwidth parameter. To achieve feature representation alignment, the noisy teacher network output features... Denoising Student Network Output Features Both are considered as two independent distributions, i.e., q θ and p θ The mean of the feature maps for these two distributions is calculated, and the difference between them is measured using the maximum mean difference. The goal is to minimize this difference, thereby pushing the output feature distribution of the student network closer to that of the teacher network. Therefore, we have:

[0200]

[0201] Kernel functions can be used to measure the differences between different distributions without explicitly calculating the mean of the distribution.

[0202] In the feedforward noise process, the diffusion model is first trained using the feature outputs of the teacher network, and noise is gradually added to generate a series of noise features. In the denoising process, a diffusion model is trained to learn a nonlinear mapping function. Thus, it is possible to obtain noise characteristics The original characteristics were restored in the middle. feature This represents the denoised version, obtained through a trained model. Recovered It is applied to noisy features The output of the post-model is intended to approximate the original clean features at time step t. Therefore, the reconstruction loss is defined as:

[0203]

[0204] in The input is denoised over time. By minimizing the reconstruction loss, the model is trained to recover clean features and perform the denoising process.

[0205] Then define the reconstruction loss. It prompts the model to extract noisy features. The original features were recovered. However, in addition to noise reduction, to ensure student network access... The feature distribution should be as close as possible to the teacher network Therefore, the maximum mean difference loss is also defined based on the characteristic distribution of the distribution.

[0206]

[0207] in and In the regenerating nucleus Hilbert space It plays a role in the process, and k(·) is related to... Related kernel functions, norm Defined as an inner product exist The norm induced in the middle.

[0208] By losing and loss Combined, the knowledge transfer loss from the teacher model to the student model can be defined as:

[0209]

[0210] Where λ∈[0,1] are weight coefficients. Through backpropagation gradient training, the diffusion model successfully learns to recover the original features from the noise-contaminated features, aligns the feature distribution using the maximum mean difference, minimizes the MMD loss, and achieves efficient knowledge transfer between the student network and the teacher network.

[0211] Subsequently, in S5, to further improve the performance of the student network, a game-based mutual distillation strategy was proposed, such as... Figure 4 As shown, two teacher networks, T1 and T2, were randomly selected, and two student networks, S1 and S2, were trained using an MMD-based diffusion model. These two student networks are treated as independent learning entities. Therefore, they learn not only from their own data but also from feedback from the other student network to optimize their parameters. In the GMD policy, the two student networks learn from each other by minimizing the difference between their output distributions.

[0212] The student network's task is to predict continuous values ​​of battery capacity. The network output is set to... and In the remaining useful life prediction task, S is used to train i The objective function is defined as the prediction capacity. With tag capacity y (j) The mean square error between them.

[0213]

[0214] in It represents the prediction of the i-th student network for the j-th sample, where M is the total number of samples.

[0215] To enable GMD, we want to minimize the difference between the predictions of the two student networks, therefore:

[0216]

[0217] For each student network, the final loss function should simultaneously consider its own error relative to the true value and its mutual learning with other student networks. Therefore, the combined loss for each network is...

[0218]

[0219] Here, φ∈[0,1] is a hyperparameter used to control the trade-off between minimizing prediction error and maximizing consistency between student networks. A larger φ focuses more on GMD loss, while a smaller φ tends to minimize individual prediction error.

[0220] Finally, in step S6 of this embodiment, a Bayesian hyperparameter optimization algorithm is used to configure the hyperparameters of the teacher model, while the hyperparameters of the student model are fine-tuned during knowledge distillation. The batch size and number of training epochs of the student model are the same as those of the teacher model. Furthermore, the temperature hyperparameter of knowledge distillation is set to T = 2, the student model specifies multiple convolutional kernel sizes (1, 3), the learning rate is dynamically adjusted using RMSprop with an initial value of 0.001, and the number of channels in the convolutional layers is set to [8, 16, 32]. For the model parameters, λ = 0.69 is set. All unspecified parameters use their default values. Estimate using the following two metrics:

[0221]

[0222] Where k represents the number of charging cycles. and y k These represent the predicted capacity and the actual capacity, respectively.

[0223] To effectively verify and evaluate the method proposed in this invention, all networks were trained on an NVIDIA RTX 3080 GPU (10GB VRAM), a 12th generation Intel Core i7-12700F (20 CPU cores), approximately 2.1GHz, DDR4 32GB memory, operating system / Microsoft Windows 10 Professional (64-bit, version 19044), programming language / Python 3.9.13, and deep learning framework / PyTorch 1.12.1. (2) All networks were tested on GPU, CPU and NVIDIA Jetson Tegra X2 (integrated nvgpu), AR-MVB processor rev3 (v8l)×4, memory / 7.7GiB, operating system / Ubuntu 18.04LTS, programming language / Python 2.7.17, and deep learning framework / PyTorch 1.6.0. EVs#17, EVs#18, EVs#19 and EVs#20 were used as validation sets. After training, teacher networks were selected as benchmarks to compare and test our method, namely Long Short-Term Memory Network (LSTM), Gated Recurrent Unit (GRU), Temporal Convolutional Network (TCN), Bidirectional Long Short-Term Memory Network (BiLSTM), and Transformer Network.

[0224] In the above embodiments, a trained capacity prediction system is deployed on the battery management system via an embedded device, and data packets generated by the lithium battery are collected and input into the capacity prediction system for detection.

[0225] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention and are not intended to limit it. Although the present invention has been described in detail with reference to preferred embodiments, those skilled in the art should understand that modifications or equivalent substitutions can be made to the technical solutions of the present invention without departing from the spirit and scope of the present invention, and all such modifications or substitutions should be covered within the scope of the claims of the present invention.

Claims

1. A battery life prediction method based on a lightweight deformable neural network using knowledge distillation, characterized in that: The method includes the following steps: Obtain the charging and discharging dataset of vehicle power batteries and preprocess the dataset; The student network is designed to dynamically adjust the sampling position of the convolution kernel through deformable depthwise convolution, redistribute feature maps through channel shuffling, extract global information using channel attention mechanism, and generate channel description vectors through global average pooling. A teacher network is established, and key features of various battery degradation data are extracted through pre-trained LSTM, GRU, TCN, BiLSTM and Transformer networks. Knowledge distillation technology is used to transfer the knowledge of the teacher network to the student network, providing the student network with learning benchmarks and goals. A diffusion model is used to add and remove noise to the student and teacher models. The output features of the teacher network are progressively added to the forward direction, while the student network learns to remove noise in the reverse direction. The output feature distribution of the student and teacher networks is realized by minimizing the maximum distribution difference loss. The two student networks are trained by simulating a "multi-expert voting mechanism" through mutual distillation, minimizing the difference between the output features of the two student networks to learn from each other; The above optimization training was performed, and the trained student model was used for real-time battery life prediction.

2. The battery life prediction method based on a lightweight deformable neural network using knowledge distillation as described in claim 1, characterized in that: The publicly available battery cycle charge and discharge dataset is used as the charge and discharge dataset for vehicle power batteries. The dataset contains charging data for several electric vehicle battery packs. The charging data includes voltage, charging current, maximum battery voltage, minimum battery voltage, maximum temperature, minimum temperature, available energy, and available capacity. At the same time, the dataset is numbered based on each electric vehicle; The preprocessing of the dataset includes label capacity calculation and dataset partitioning; specifically, in the label capacity calculation process, the ampere-hour integral method is used to calculate the label capacity from the known data, and then normalization is performed. Where t1 and t2 are the start and end times of the selected charging segment. The average current is given by Δt, the data sampling frequency is given by SOC, and the battery state of charge is given by SOC. Splitting the dataset refers to randomly dividing the preprocessed battery data into training and test sets according to a preset ratio.

3. The battery life prediction method based on a lightweight deformable neural network using knowledge distillation as described in claim 1, characterized in that: In a student network, the process of dynamically adjusting the sampling position of the convolution kernel through deformable depthwise convolution is as follows: Set input feature map The dimensions are T×F×C in Where T represents the time step, F is the number of features, and C in =n is the number of channels; Apply deformable depthwise convolution to compute the output feature map At position p k Value at: in the grid Increase offset by {Δp} k |k=1,...,K}, where For each channel Apply the convolution kernel independently, and through the offset Δp k Adjusting the position yields: Δp k =OffsetConv(u ij ) Where OffsetConv(.) is a small convolutional neural network that predicts offsets, u ij ∈U is a sample of battery data; for each position p k Through offset Δp k To adjust the position of the convolution kernel, at each adjusted position p = p0 + p k +Δp k Above, the feature map is represented as: for If the adjusted position p has non-integer coordinates, then for any Calculated using bilinear interpolation Eigenvalues: Where G(·,·) is the bilinear interpolation kernel, and q represents the input feature map. Integer position in This represents the set of all such locations surrounding p; Then, G(q,p) is decomposed into two one-dimensional kernels g(q) x ,p x ) and g(q y ,p y If ), then the following relationship is satisfied: G(q,p)=g(q x ,p x )·g(q y ,p y ) Where q x and q y q is the coordinate of the integer position, while p x and p y These are the coordinates of the adjusted position p in the input and output feature maps, respectively. In conclusion, we have: After processing with deformable depthwise convolution, the feature maps input by the student network can adapt to changes in the teacher network's data, and the feature maps can more accurately represent key information.

4. The battery life prediction method based on a lightweight deformable neural network using knowledge distillation as described in claim 3, characterized in that: In a student network, the process of redistributing feature maps through channel shuffling is as follows: Given input feature map First, its channel number is divided into C in Divide into G groups, each group containing Channel. Let x = x1, x2, ..., x n Each of them It is a set of channels; Through channel shuffling, features from different groups are mixed together to generate a single output feature map. The number of channels remains constant, but each channel feature contains information from other groups: in This indicates the shuffle operator applied to the i-th channel group; Simultaneously, through an adaptive channel selector, global information is extracted from each channel of the feature map output by the multi-scale feature extractor using global average pooling: y G =vec(GAP(y)) Where vec(·): This represents a vectorization operation, which flattens a 1×1×N tensor into an N-dimensional vector; The GAP operation applied to y is represented as: In the formula, The resulting feature vector contains global information for each channel in the output feature map y; Then, the extracted feature vector y G Nonlinear transformation and scaling are performed using two fully connected layers and a sigmoid activation function: S=W2·λ(W1·y G +b1)+b2 in, and These are the weight matrices of two fully connected layers, where This represents the compressed channel dimension, where r is a compression ratio; The final output features of the student network are represented as follows: y←y⊙Sigmoid(S) Here, ⊙ represents element-wise multiplication, applying the scaling factor to each channel.

5. The battery life prediction method based on a lightweight deformable neural network using knowledge distillation as described in claim 4, characterized in that: In the teacher network, LSTM, GRU, TCN, BiLSTM, and Transformer networks are used for sequence modeling. LSTM captures long-term dependencies in sequence data; Gated Recurrent Unit (GRU) captures long-term dependencies in sequence data based on a simpler architecture; Temporal Convolutional Network (TCN) models long-range temporal dependencies through causal convolution; Bidirectional Long Short-Term Memory (BiLSTM) network captures both past and future contextual information in sequence data; and Transformer uses a self-attention mechanism to model long-range dependencies with higher parallelization efficiency. These network models are used to extract temporal features from battery data and provide knowledge to the student network.

6. The battery life prediction method based on a lightweight deformable neural network using knowledge distillation as described in claim 5, characterized in that: In the knowledge transfer process between teacher and student networks, a nonlinear mapping function is first identified. The nonlinear mapping function takes the battery sequence as input and outputs the desired capacity sequence Y, while minimizing the mapping error of the input data sample U, and the data sample U follows a natural distribution. Nonlinear mapping function Configure a set of structured hyperparameters Θ and employ a learning algorithm. In a limited battery dataset Through iterative optimization, an optimal hyperparameter Θ is eventually found. * Make the loss function If minimized, the prediction paradigm for the teacher network and student network is defined as follows: In the formula, Indicates hard loss; In a knowledge distillation process, knowledge from a pre-trained teacher network is transferred to a student network given a teacher network Y. T And student network Y S When the output is obtained, the KD loss of knowledge distillation is defined as: in, Indicates soft loss. Let represent the distance function that measures the difference between the outputs of the teacher network and the student network, and let be the mean squared error of the regression output.

7. The battery life prediction method based on a lightweight deformable neural network using knowledge distillation as described in claim 6, characterized in that: In the knowledge transfer process between teacher and student networks, a knowledge transfer technique based on a diffusion model is introduced. The process is as follows: Treating the output of the student network as a noisy version of the teacher network output, for the input sample Battery sequence data was used to simultaneously train both the teacher and student networks to extract features. and Then, during the forward noise process, repeatedly to Add Gaussian noise: Where, q θ This indicates the diffusion process of the teacher network. This represents the primitive characteristics of the teacher network at the initial time step. The transition noise characteristics at time step t This indicates the use of the noise variance scheduling scheme β. s The sign of the feature is directly sampled at any time step; In the given conditions, The distribution follows a Gaussian distribution: but The mean is The covariance matrix is Gaussian distribution rewritten as If it is equal to the mean plus noise, then it represents... for: Where, ∈ t It's Gaussian noise. Introduce a scaling factor This makes it have the desired variance. Therefore: in Standard Gaussian noise, i.e. final, Represented as and noise variables Linear combination: For the teacher network, a nonlinear mapping function is trained. To predict Equivalent to The noise in the data, i.e., minimizing the L2 loss between them: For student networks, the initial characteristics are: By using nonlinear mapping functions Perform an iterative denoising process to reconstruct features. in Indicates transition variance. Represents the conditional probability of a student, given features at time t. Below, student characteristics The conditional probability at time t-1; Modeled as a mean The Gaussian distribution, where It is a function defined by parameter θ, used to extract noisy characteristics from time step t. And predict clean features The variance of the Gaussian distribution is It represents the noise level at time t; and I is the identity matrix, indicating that the noise is independent and identically distributed. The characteristics of teacher networks are represented by time. Gaussian noise was gradually added to obtain the data; the student network was obtained by gradually adding Gaussian noise over time. Starting with the noisy features, we learn to remove the noise in reverse order to recover the clean features.

8. The battery life prediction method based on a lightweight deformable neural network using knowledge distillation as described in claim 7, characterized in that: The output features generated by the student network are considered as a noisy version of the features generated by the teacher network, and the noise level representing the difference between the features of the teacher network and the student network is unknown. Therefore, the maximum mean difference method is used to align the feature distributions of the noisy teacher network and the noisy-free student network. Maximum mean difference is a distribution metric based on kernel embedding, defined as: in, Let q and p represent the feature space of kernel function k, respectively, and let k(·) denote the application of kernel function k to the feature vector. and Let features and Implicitly mapped to through kernel representation In this context, k is chosen as the Gaussian kernel function, i.e.: Where σ represents the kernel bandwidth parameter; Output features of the noisy teacher network Denoising Student Network Output Features Consider them as two independent distributions, i.e., q θ and p θ Calculate the mean of the feature maps of the two distributions, and use the maximum mean difference to measure the difference between them. Minimize the difference to push the output feature distribution of the student network closer to the output feature distribution of the teacher network. Then: Kernel functions are used to measure the differences between different distributions; In the feedforward noise process, the diffusion model is first trained using the feature outputs of the teacher network, and noise is gradually added to generate a series of noise features. In the denoising process, a diffusion model is trained to learn a nonlinear mapping function. Thus, it is possible to obtain noise characteristics The original characteristics were restored in the middle. feature This represents the denoised version, obtained through a trained model. Recovered It is applied to noisy features The output of the post-model yields the original clean features at approximately time step t. Therefore, the reconstruction loss is defined as: In this process, the model is trained to recover clean features and perform a denoising process by minimizing the reconstruction loss. It also defines the maximum mean difference loss. in, and In the regenerating nucleus Hilbert space It plays a role in the process, and k(·) is related to... Related kernel functions, norm Defined as an inner product exist The norm induced in the middle; By losing and loss In combination, the knowledge transfer loss from the teacher model to the student model is defined as: Where λ∈[0,1] are weight coefficients; through reverse gradient training, the diffusion model recovers the original features from the features contaminated by noise, and uses the maximum mean difference to align the feature distribution to minimize the maximum mean difference (MMD) loss, thereby realizing the knowledge transfer between the student network and the teacher network.

9. The battery life prediction method based on a lightweight deformable neural network using knowledge distillation as described in claim 8, characterized in that: In the mutual distillation process between two student networks, a game-based mutual distillation strategy is adopted, and the process is as follows: Two teacher networks T1 and T2 are randomly selected, and two student networks S1 and S2 are trained using an MMD-based diffusion model. The two student networks are treated as independent learning entities, and they learn from each other by minimizing the difference between their output distributions. The task of the student network is to predict continuous values ​​of battery capacity. The output of the student network is set to... and In the remaining useful life prediction task, S is used to train i The objective function is defined as the prediction capacity. With tag capacity y (j) Mean square error between: in It is the prediction of the i-th student network for the j-th sample, and M is the total number of samples; If the objective is to minimize the difference between the predictions of the two student networks, then: For each student network, the final loss function considers both its own error relative to the true value and its mutual learning with other student networks. Therefore, the combined loss for each network is: Here, φ∈[0,1] is a hyperparameter used to control the trade-off between minimizing prediction error and maximizing consistency between student networks.

10. A battery life prediction method based on a lightweight deformable neural network using knowledge distillation, as described in claim 9, characterized in that: In the training of the above process, the hyperparameters of the teacher network are configured using the Bayesian hyperparameter optimization algorithm, and the hyperparameters of the student network are fine-tuned during the knowledge distillation process. The trained model is used for real-time prediction of remaining battery life.

Images