Power battery attenuation prediction method based on depth auto-encoder

Abstract

The invention belongs to the technical field of power batteries, and particularly relates to a power battery attenuation prediction method based on a depth auto-encoder, which comprises the following steps: firstly, collecting new energy automobile power battery sample data, pre-processing, and constructing a depth auto-encoder model; adjusting the parameters of the depth auto-encoder model according to the gradient information of the loss function by adopting an optimization algorithm until the training of the depth auto-encoder model is completed; then collecting real-time operation data of a current target new energy automobile as original data, inputting the original data into the depth auto-encoder model, outputting reconstruction data, calculating reconstruction errors, associating the reconstruction errors with the health state of the power battery, and representing the reconstruction errors as a health index of the power battery; and finally, based on the time sequence data of the power battery health index, predicting the decline trend and the residual life of the power battery by adopting a bidirectional LSTM + Attention mechanism. The problem of low accuracy of a power battery life prediction technology in the prior art can be solved.

Description

Technical Field

[0001] The present invention belongs to the technical field of power batteries, and particularly relates to a method for predicting the attenuation of power batteries based on a deep autoencoder. Background Technique

[0002] With the rapid development of the global energy structure transformation and intelligent transportation technology, new energy vehicles have become an important development direction of the automotive industry; as the core component of new energy vehicles, the performance attenuation of power batteries directly affects the vehicle's endurance, safety, and usage cost. Therefore, accurately predicting the remaining useful life (RUL) of power batteries is of great significance for optimizing battery management strategies, extending battery life cycles, and enhancing user experiences.

[0003] However, the attenuation process of power batteries is affected by the coupling effect of multiple physical fields, and its capacity attenuation law shows highly nonlinear, time-varying, and environment-dependent characteristics, posing great challenges to life prediction; the existing power battery life prediction technologies are mainly divided into three categories: physical and chemical methods based on empirical models, model methods based on equivalent circuits, and machine learning methods based on data-driven. Among them, the physical and chemical methods establish an internal electrochemical reaction model of the battery and deduce the capacity attenuation law by combining thermodynamic theories such as the Arrhenius equation. Its advantage is that the physical meaning is clear, but the model parameter identification is complex and relies on a large amount of experimental data, making it difficult to adapt to the dynamic changes under actual working conditions; the equivalent circuit model method simulates the dynamic characteristics of the battery through an RC network and estimates the remaining capacity by combining the ampere-hour integration method. This method has high calculation efficiency, but the selection of model order and parameter optimization have empirical dependencies and it is difficult to accurately reflect the nonlinear characteristics of the battery aging process; the data-driven method predicts by mining the attenuation patterns in historical monitoring data, such as support vector machines, recurrent neural networks, etc. Its advantage is that it does not require an accurate mathematical model, but there are problems such as high data annotation costs and weak model generalization capabilities.

[0004] Therefore, the limitations in the existing technology include: 1. The problem of single data source is prominent: existing methods are mostly based on the cell aging data under laboratory standard cycling conditions, ignoring the synergistic effects of multi-source heterogeneous data such as vehicle operating parameters (such as vehicle speed, load), environmental factors (such as temperature, humidity), and user behavior (such as charge and discharge habits) on battery attenuation in actual usage scenarios; 2. Insufficient model adaptability: Traditional empirical models and equivalent circuit models are difficult to adapt to the differential attenuation characteristics of different brands and models of batteries, and prediction models based on shallow machine learning have significant defects in generalization ability across working conditions and battery types; 3. Insufficient depth of feature mining: The battery attenuation process involves the non-linear coupling of multi-dimensional parameters such as voltage, current, temperature, and internal resistance. Existing methods have limited ability to extract implicit features in high-dimensional data and lack effective data dimensionality reduction and feature fusion mechanisms, resulting in the prediction accuracy being difficult to meet the requirements of engineering applications.

[0005] Therefore, due to the above limitations, the accuracy of the attenuation prediction of power batteries will ultimately be reduced, and it is urgent to solve this problem. Summary of the Invention

[0006] The technical problem solved by the present invention is to provide a power battery attenuation prediction method based on a deep autoencoder to solve the problem of low accuracy in the existing power battery life prediction technology.

[0007] The basic solution provided by the present invention: A power battery attenuation prediction method based on a deep autoencoder, including: S1: Collect the operation data of the new energy vehicle throughout its life cycle and the offline charging detection data to obtain sample data, and after preprocessing the sample data, generate preprocessed sample data; S2: Construct a deep autoencoder model, divide the preprocessed sample data into a training set, a validation set, and a test set according to a preset ratio, input the training set into the deep autoencoder model, add a sparse constraint in the deep autoencoder model, and after outputting the reconstructed data, by constructing a loss function based on the reconstruction error and the sparse constraint, use an optimization algorithm to adjust the parameters of the deep autoencoder model according to the gradient information of the loss function until the deep autoencoder model is trained, and then use the validation set for validation and the test set for testing respectively; S3: Collect the real-time operation data of the current target new energy vehicle as the original data, preprocess it and then input it into the deep autoencoder model that has been verified and tested to output the reconstructed data; associate the reconstruction error with the health state of the power battery and characterize it as a power battery health index; S4: Based on the time series data of the power battery health index, use the bidirectional LSTM + Attention mechanism to predict the decline trend and remaining life of the power battery.

[0008] Further, the S1 includes: S1-1: Collect the charging and discharging currents, voltages, temperatures, humidities, operating environment parameters of the load of the power battery of the new energy vehicle, as well as the vehicle state data such as the driving mileage and driving mode of the new energy vehicle, and the offline charging detection data of the new energy vehicle, and generate sample data according to the operating environment parameters, vehicle state data, and offline charging detection data; S1-2: Perform missing value filling, feature normalization, and time alignment processing on the sample data to obtain preprocessed sample data.

[0009] Further, the S2 includes: S2-1: Determine the dimension of the input layer of the deep autoencoder model according to the features in the training set; S2-2: Construct the encoder of the deep autoencoder model. In the encoder, compress the data in the input training set into a low-dimensional feature space through a multi-layer neural network, and the number of neurons in each layer decreases according to a downward trend; and add a sparse constraint to the encoder; S2-3: Construct the decoder of the deep autoencoder model, and restore and denoise the compressed low-dimensional features in the encoder to obtain reconstructed data; S2-4: Output the reconstructed data through the output layer of the deep autoencoder model, and calculate the reconstruction error that can characterize the health index of the power battery according to the reconstructed data to generate a complete deep autoencoder model; S2-5: Train and optimize the deep autoencoder model based on a preset loss function and optimization algorithm to obtain the final deep autoencoder model.

[0010] Further, adding a sparse constraint to the encoder in S2-2 specifically means: Define the target sparsity , and limit the hidden layer activation degree to be less than 15%; Then the actual sparsity calculation formula is:

[0011] Use the divergence penalty term to constrain the sparsity, The divergence penalty term expression is:

[0012] Among them, represents the sparsity estimate value of the th layer, represents the number of features, represents the number of neurons, represents the hidden layer state in the th layer and the th neuron.

[0013] Further, training and optimizing the deep autoencoder model based on a preset loss function and optimization algorithm in S2-5 to obtain the final deep autoencoder model specifically means: Determine that the total loss function of the deep autoencoder model consists of an initial loss function and a sparse constraint; The reconstruction error is used as the initial loss function to measure the difference between the initial data of the input validation set and the reconstructed data. The model parameters are optimized by minimizing the loss function. The expression of the reconstruction error is as follows:

[0014] Among them, represents the initial data, represents the reconstructed data; Then the expression of the total loss function is:

[0015] Among them, represents the weight coefficient of the sparse loss.

[0016] Call the optimization algorithm to adjust the weights and biases of the deep autoencoder model according to the gradient information of the total loss function; and adjust the hyperparameters of the deep autoencoder model during the training process, determine the optimal hyperparameter combination through cross-validation, and then call the test set for testing until the final deep autoencoder model that meets the requirements is obtained.

[0017] Furthermore, the S3 includes: S3-1: Collect the real-time operation data of the current target new energy vehicle as the original data, and after preprocessing the original data, obtain the preprocessed original data; S3-2: Input the preprocessed original data into the final deep autoencoder model, extract and reconstruct the combined original signal features and differential entropy as low-dimensional features to obtain the reconstructed data; the expression of the differential entropy is:

[0018] Among them, represents the probability distribution of the differential of the feature with respect to time at the time point ; the original signal features are charging / discharging current, total voltage, temperature, cell voltage, and battery collection point temperature; S3-3: Calculate the reconstruction error between the original data and the reconstructed data, and associate the reconstruction error with the state of health of the power battery, which is characterized as the health index of the power battery. The association expression is:

[0019] Among them, represents the state of health of the power battery; It is a fully connected layer autoencoder. The association with the state of health of the power battery is as follows: Group the collected battery sample data according to the state of health levels; for each battery group with a specific state of health level, calculate the statistical features of the corresponding reconstruction error, including the mean and standard deviation, and establish an association mapping relationship between the reconstruction error and the battery state of health level; this is characterized as the power battery health indicator.

[0020] Fully connected layer autoencoder The expression is:

[0021] Where represents the original signal features, is the reconstructed data output by the autoencoder, is the weight matrix of the encoder, is the weight matrix of the decoder, , , , are bias vectors, , are both ReLu activation functions.

[0022] Taking the battery group in good condition as an example, the average value of its reconstruction error is about 0.1, and the standard deviation is 0.05; for the battery group in general condition, the average value of the reconstruction error is between 0.2 - 0.3, and the standard deviation is 0.08 - 0.1; for the battery group in poor condition, the average value of the reconstruction error is greater than 0.4, and the standard deviation is greater than 0.15. Therefore, an association mapping between the reconstruction error and the battery state of health level can be obtained.

[0023] Furthermore, the specific process of extracting and combining the original signal features and differential entropy as low - dimensional features and performing reconstruction to obtain the reconstructed data in S3 - 2 is as follows: Concatenate the original signal features and the differential entropy features into a joint feature;

[0024] Where and are their respective dimensions; here, the original signal features specifically refer to the features of charging / discharging current, total voltage, temperature, cell voltage, and the temperature signal at the battery acquisition point, and the differential entropy features refer to the differential entropy features formed by the probability distributions of the total current, total voltage, cell voltage, and temperature signals after taking the time derivative.

[0025] Compress the joint high - dimensional input into a low - dimensional latent space, and the expression is:

[0026] Among them, represents the encoder function, represents the combined multi-source features after normalization, is the encoder weight, is the encoder bias, represents the weight matrix of the first fully connected layer of the encoder, represents the weight matrix of the second fully connected layer of the encoder, represents the bias vector of the first layer of the encoder, represents the bias vector of the second layer of the encoder, is the activation function, is the low-dimensional feature ; Finally, the reconstruction term and the reconstruction error are obtained through the deep autoencoder constraint function.

[0027] Furthermore, the said S4 includes: S4-1: Extract the time series data in the battery health indicators; S4-2: Construct a and a network, and call the network to process the forward time series data and the reverse time series data respectively; The network expression is: The network expression is:

[0028]

[0029]

[0030] Among them, represents the network, represents the forward network, represents the reverse network, represents the feature dataset, represents the time step of the sequence input data, represents the hidden state at time step , represents the hidden state of the subsequent steps, represents the single-layer hidden state dimension.

[0031] S4-3: Introduce the attention mechanism, and calculate the scores of the hidden states of each time series through the fully connected layer. The expression of the attention score is:

[0032] Among them, , represents the scientific system parameter, represents the transpose of matrix V, is the attention dimension, represents the bias of the hidden layer, represents the hidden state; S4-4: Use The activation function generates attention weights to weight the concatenated hidden states to change the importance. The expression is: , ; Among them, is the attention weight, is the attention score, represents the attention score at time step k, w is the length of the time series; and then the context vector with importance is obtained Input the fully connected layer for prediction , and its expression is:

[0033] Among them, represents the weight of the fully connected layer, represents the bias term of the fully connected layer; softmax is a commonly used activation function, especially often used as the activation function of the output layer in multi-classification problems. Its role is to convert a real number vector into a probability distribution, that is, map each element in the input vector to the interval (0, 1), and ensure that the sum of all elements is 1. In this example, the softmax function is used to convert the attention score into the attention weight . These weights represent the relative importance of each time step. The larger the weight, the more important the time step is for the final prediction target.

[0034] S4-5: Set a sliding window to realize the dynamic adjustment and online update of model parameters. Arrange the power battery health indicators in the order of time series, and increase the weight allocation according to the key attenuation stages of the power battery.

[0035] An electronic device, including a processor and a memory. Programs or instructions are stored in the processor. The processor executes the method for predicting the attenuation of a power battery based on a deep autoencoder described in any one of the above by calling the programs and instructions stored in the memory.

[0036] A computer-readable storage medium stores a program or instructions, and the program or instructions cause a computer to execute a power battery attenuation prediction method based on a deep autoencoder as described in any one of the above.

[0037] The principle and advantages of the present invention are as follows: In the power battery attenuation prediction method based on a deep autoencoder in this application, the full-life cycle operation data and offline charging detection data of new energy vehicles collected are used as sample data. The sample data has multi-dimensional characteristics of the power battery, such as the charge and discharge cycles of the battery, temperature fluctuations, load changes, etc., and the decline process of the power battery has strong non-linear and time-series characteristics; Based on the above content, in order to better analyze the collected sample data effectively to obtain the power battery health index and predict the attenuation of the power battery, a deep autoencoder model is constructed. The deep autoencoder model includes an input layer, an encoder, a decoder, and an output layer. A sparse constraint is introduced in the encoder, and denoising processing is performed in the decoder to realize dimensionality reduction reconstruction and non-linear mapping of the multi-dimensional characteristics in the sample data, obtaining more robust and more discriminative features. At the same time, in the process of predicting the attenuation of the power battery, the original signal features are combined with time-frequency domain features such as differential entropy. For the first time, the differential entropy of the charge and discharge curve is used as a frequency domain feature input, and a bidirectional LSTM + Attention mechanism is introduced, so that unsupervised learning can be carried out from the multi-dimensional sample data of the power battery to predict the decline trend and remaining life of the power battery.

[0038] Therefore, the advantages of this application are as follows: 1. The multi-dimensional data of the power battery is processed through a deep autoencoder model to achieve a high-dimensional compression ratio and information retention rate, and more effectively predict the decline trend and remaining life of the power battery; 2. In the process of predicting the attenuation of the power battery, the original signal features are combined with time-frequency domain features such as differential entropy, which can improve the sensitivity of features that are crucial for battery capacity attenuation, such as lithium dendrite growth; 3. The deep autoencoder model constructed in this application uses sparse constraint + denoising processing training to improve the feature quality, achieving a hidden layer activation degree < 15% and an improvement in the feature reconstruction signal-to-noise ratio; 4. A sliding window attention mechanism is introduced into the bidirectional LSTM network. The model parameters are dynamically adjusted and updated online through the sliding window, and the weights in the critical attenuation stage of the power battery can be improved, enhancing the prediction accuracy of the remaining life and decline trend of the power battery. BRIEF DESCRIPTION OF THE DRAWINGS

[0039] Figure 1 It is a flowchart of an embodiment of the present invention; Figure 2It is a schematic structural diagram of the deep autoencoder model in the embodiments of the present invention; Figure 3 It is a schematic structural diagram of the electronic device in the present invention. Specific embodiments

[0040] The following is a further detailed description through specific embodiments: The marks in the attached drawings of the specification include: electronic device 400, processor 401, memory 402, input device 403, output device 404.

[0041] With the wide application of new energy vehicles, the life prediction of power batteries is of great significance for ensuring vehicle performance, reducing costs, and improving user experience. The following problems or defects exist in the prior art: 1. The data source is single. Usually, it only predicts based on the cyclic working conditions set in the laboratory, using the stable time series data of cell aging, ignoring the influence of multi-source data such as vehicle information and driving environment on battery life; 2. The accuracy and generalization ability of the prediction model are limited, and it is difficult to meet the prediction requirements of power battery life under different brands, models, and usage scenarios; 3. Insufficient capture of non-linear features. The capacity attenuation is non-linearly coupled with multiple factors such as temperature and charge-discharge rate; the battery monitoring data has a high dimension, such as including voltage, current, temperature, internal resistance, etc., there is information overlap, and there is a lack of in-depth mining and fusion of data features, resulting in inaccurate prediction results and unable to provide users with reliable battery life assessment.

[0042] Therefore, based on the above problems, the technical solution proposed in this application, its embodiments are basically as shown in the appendix Figure 1 shown: A power battery attenuation prediction method based on a deep autoencoder, including: S1: Collect the operation data of the new energy vehicle throughout its life cycle and the offline charging detection data, obtain the sample data, and generate the preprocessed sample data after preprocessing the sample data; in this embodiment, the collected sample data are respectively: 1. The operation data based on the new energy vehicle throughout its life cycle, including the charging and discharging current and voltage of the battery; operation environment parameters such as temperature, humidity, and load; 2. Vehicle state data such as driving mileage and driving mode; 3. The offline charging detection data based on the new energy vehicle, including battery health state data such as the current capacity, internal resistance, charging voltage, current, and detection temperature of the battery.

[0043] Subsequently, using the collected data as sample data, preprocess it. The preprocessing in this application includes: 1. Missing value processing, where interpolation is used to process missing data and outliers to ensure the continuity and stability of the data; 2. Normalize data such as voltage, current, and temperature so that they fall within a unified scale to avoid the impact of scale differences between features on model training; 3. Align all data according to timestamps to ensure the temporal consistency of vehicle data and battery health data.

[0044] S2: Construct a deep autoencoder model. Divide the preprocessed sample data into a training set, a validation set, and a test set according to a preset ratio. Input the training set into the deep autoencoder model, and add a sparsity constraint to the deep autoencoder model. After outputting the reconstructed data, by constructing a loss function based on the reconstruction error and the sparsity constraint, use an optimization algorithm to adjust the parameters of the deep autoencoder model according to the gradient information of the loss function until the deep autoencoder model is trained. Then, use the validation set for validation and the test set for testing respectively; where S2 includes: S2-1: Determine the dimension of the input layer of the deep autoencoder model according to the features in the training set; S2-2: Construct the encoder of the deep autoencoder model. In the encoder, compress the data of the input training set into a low-dimensional feature space through a multi-layer neural network, and the number of neurons in each layer decreases according to a downward trend; and add a sparsity constraint to the encoder; S2-3: Construct the decoder of the deep autoencoder model. Restore and denoise the compressed low-dimensional features in the encoder through the decoder to obtain the reconstructed data; S2-4: Output the reconstructed data through the output layer of the deep autoencoder model, and calculate the reconstruction error that can characterize the health index of the power battery based on the reconstructed data to generate a complete deep autoencoder model; S2-5: Train and optimize the deep autoencoder model based on a preset loss function and optimization algorithm to obtain the final deep autoencoder model.

[0045] In this embodiment, as Figure 2 shown, the deep autoencoder model includes an input layer, an encoder, a decoder, and an output layer. There is also a low-dimensional feature layer between the encoder and the decoder; where the input layer is used to receive data, and the dimension of the input layer is determined according to the number of features of the input data; in the encoder, there are multiple layers of encoders, and a neural network is set for each layer of the encoder. The input data is processed through the multi-layer neural network of the encoder, and the number of neurons in each layer gradually decreases according to a downward trend, and the activation function of the encoder uses the ReLU non-linear function.

[0046] Through multi-layer encoding, the data is compressed into a low-dimensional space, i.e., the low-dimensional feature layer. Core features are extracted from the low-dimensional feature layer, such as features related to the health indicators of power batteries, while redundant and noisy information is removed. A concise and crucial feature representation is provided for the decoder to facilitate subsequent data restoration and processing. Subsequently, the decoder gradually restores the low-dimensional features to obtain the reconstructed data.

[0047] In addition, a sparsity constraint is added to the encoder during training to limit the activation degree of the hidden layer to less than 15%, so as to improve the quality of features in the input data and the generalization ability of the model. Specifically: Define the target sparsity , , restricting the activation degree of the hidden layer to less than 15%; Then the actual sparsity calculation formula is:

[0048] Use the divergence penalty term to constrain sparsity, The divergence penalty term expression is:

[0049] Among them, represents the sparsity estimation value of the th layer, represents the number of features, represents the number of neurons, represents the th layer's th hidden layer state in the

[0050] The decoder is to restore the low-dimensional features compressed by the encoder back to a form close to the original data. Specifically, the number of neurons gradually increases until the original dimension , and Gaussian noise with is added to the input data. Through denoising training, the deep autoencoder model learns a more robust feature representation and improves the signal-to-noise ratio of feature reconstruction. The expression is:

[0051] Among them, the noise variance is calculated according to SNR: , ; Among them, represents the noisy feature data, represents the original signal feature data, represents the noise data.

[0052] The original signal features have been described above, specifically referring to the features of the battery charging / discharging current, total voltage, temperature, single-cell voltage, and the temperature signal at the battery acquisition point. The feature data is specifically the feature data related to the health of the power battery extracted from the charging / discharging current, total voltage, temperature, single-cell voltage, and the temperature signal at the battery acquisition point.

[0053] The output layer then outputs the reconstructed data, which is used to calculate the reconstruction error and serves as an indicator to measure the change in the health state of the power battery.

[0054] Subsequently, the deep autoencoder model is trained and optimized. In this application, double regularization is used to improve the feature quality, that is, the reconstruction error and sparse constraints are used to construct the total loss function. Specifically: The preprocessed sample data is divided into a training set, a validation set, and a test set according to a ratio of 7:2:1; It is determined that the total loss function of the deep autoencoder model consists of an initial loss function and a sparse constraint; The reconstruction error is used as the initial loss function to measure the difference between the initial data of the input validation set and the reconstructed data. The model parameters are optimized by minimizing the loss function. Among them, the expression of the reconstruction error is:

[0055] Among them, represents the initial data, represents the reconstructed data; Then the expression of the total loss function is:

[0056] Among them, represents the weight coefficient of the sparse loss, which determines the importance of sparsity in the overall loss .

[0057] Call the optimization algorithm, and adjust the weights and biases of the deep autoencoder model according to the gradient information of the total loss function. Accelerate the model convergence speed and improve the prediction accuracy. Among them, for each parameter ; is the weight parameter, is the bias parameter, and its gradient calculation and update formula are: 1. Calculate the current gradient:

[0058] Among them, represents the gradient value, represents the loss function with respect to the model parameter gradient, represents the total loss function; 2. Update the first moment estimate:

[0059] where represents the current momentum, represents the decay rate of the first moment estimate, represents the first moment estimate at time step 3. Update the second moment estimate:

[0060] where represents the current adaptive learning rate, represents the decay rate of the second moment estimate, represents the second moment estimate at time step 4. Bias correction:

[0061]

[0062] where represents the bias correction of represents the bias correction of 5. Parameter update:

[0063] where represents the learning rate, represents the time step of the network parameter represents the time step of the network parameter

[0064] And adjust the hyperparameters of the deep autoencoder model during the training process, and determine the optimal hyperparameter combination through cross-validation. Specifically, the hyperparameters to be adjusted include the learning rate, batch size, the number of neurons in the hidden layer, etc. Determining the optimal hyperparameter combination through cross-validation can improve the generalization ability and prediction performance of the model.

[0065] Then call the test set for testing until the final deep autoencoder model that meets the requirements is obtained.

[0066] In this embodiment, for the constructed deep autoencoder model, the number of neurons in the input layer is 100, corresponding to the dimensionality of the processed multi-source data features. The encoder consists of three layers with the number of neurons being 80, 50, and 30 respectively, and the ReLU activation function is adopted. The decoder consists of three layers with the number of neurons being 30, 50, and 80 respectively, and the number of neurons in the final output layer is 100, corresponding to the input layer.

[0067] Among its related parameters, the learning rate is set to 0.001, the batch size is 64, and the number of training epochs is 100. The optimal hyperparameter combination is determined through grid search and cross-validation to ensure that the model has good performance on both the training set and the validation set.

[0068] S3: Collect the real-time operation data of the current target new energy vehicle as the original data. After preprocessing, input it into the deep autoencoder model, output the reconstructed data, and calculate the reconstruction error between the original data and the reconstructed data. Associate the reconstruction error with the state of health of the power battery and characterize it as the power battery health index; where S3 includes: S3-1: Collect the real-time operation data of the current target new energy vehicle as the original data, and after preprocessing the original data, obtain the preprocessed original data; S3-2: Input the preprocessed original data into the final deep autoencoder model, extract and combine the original signal features and differential entropy as low-dimensional features and perform reconstruction to obtain the reconstructed data; the expression of the differential entropy is:

[0069] where represents the probability distribution of the differential of the feature at the time point of the feature

[0070] In S3-2, extracting and combining the original signal features and differential entropy as low-dimensional features and performing reconstruction to obtain the reconstructed data is specifically: concatenate the original signal features and the differential entropy features into joint features;

[0071] where and are the respective dimensions; Compress the joint high-dimensional input into the low-dimensional latent space, and the expression is:

[0072] where represents the encoder function, Represents the standardized multi-source feature combination, is the encoder weight, is the encoder bias, represents the weight matrix of the first fully-connected layer of the encoder, represents the weight matrix of the second fully-connected layer of the encoder, represents the bias vector of the first layer of the encoder, represents the bias vector of the second layer of the encoder, is the activation function, is the low-level feature ; Finally, the reconstruction term and the reconstruction error are obtained through the deep autoencoder constraint function.

[0073] S3-3: Calculate the reconstruction error between the original data and the reconstructed data, associate the reconstruction error with the state of health of the power battery, and characterize it as the power battery health index. The association expression is:

[0074] where, represents the state of health of the power battery; as the power battery decays, the reconstruction error between the encoded, decoded data and the original data will gradually increase, thereby reflecting the change in the state of health of the power battery.

[0075] Fully-connected layer autoencoder The expression is:

[0076] where, represents the original signal feature, is the reconstructed data output by the autoencoder, is the weight matrix of the encoder, is the weight matrix of the decoder, , , , are bias vectors, , are both ReLu activation functions.

[0077] S4: Based on the time series data of the power battery health index, use the bidirectional LSTM+Attention mechanism to predict the degradation trend and remaining life of the power battery. Among them, S4 includes: S4-1: Extract the time series data in the power battery health index; S4-2: Construct a forward and a backward In the architecture of the network, there are an input gate, a forget gate, a cell state gate, a state update gate, and an output gate. The input data will be calculated sequentially through these gates, and then The network processes the forward time series data and the reverse time series data separately; The network expression is:

[0078]

[0079]

[0080] Among them, represents the network, represents the forward network, represents the reverse network, represents the feature data set, represents the time step of the sequential input data, represents the hidden state at time step , represents the hidden state of the subsequent steps, represents a single-layer hidden state dimension.

[0081] S4-3: Introduce the attention mechanism. Calculate the scores for the hidden states of each of the above time series through a fully connected layer. The expression is:

[0082] Among them, , represents the scientific system parameter, is the attention dimension, represents the bias of the hidden layer, represents the hidden state; S4-4: Use activation function to generate attention weights and weight the concatenated hidden states to change the importance. The expression is: , ; Among them, is the attention weight, is the attention score, represents the attention score at time step k, w is the length of the time series; and then obtain the context vector with importance Input into the fully connected layer for prediction , and its expression is:

[0083] Among them, represents the weight of the fully connected layer, represents the bias term of the fully connected layer.

[0084] Softmax is a commonly used activation function, especially often used as the activation function of the output layer in multi-classification problems. Its role is to convert a real number vector into a probability distribution, that is, map each element in the input vector to the interval (0,1), and ensure that the sum of all elements is 1. In this example, the softmax function is used to convert the attention scores into attention weights . These weights represent the relative importance of each time step. The larger the weight, the more important the time step is for the final prediction target.

[0085] S4-5: Set the sliding window Implement dynamic adjustment and online update of model parameters. Arrange the battery health indicators of power batteries in chronological order. The arranged sequence , where T is the total number of cycles; set the window size to 5 cycles to generate input-output pairs. The input is: , and the output is: ; Among them , represents the current time position, that is, the index of the time step, represents the size of the sliding window, that is, the number of time steps included in the input sequence; each window corresponds to the remaining useful life value ; and improve the weight allocation according to the key attenuation stages of power batteries. For example, when diagnosed as the lithium plating and capacity diving stages, by assigning higher weights to the data points in the key stages, the sensitivity of the model to the key stages can be enhanced to provide accurate battery life assessment.

[0086] For example Figure 3 As shown, in another embodiment of this embodiment, there is also an electronic device. The electronic device 400 includes one or more processors 401 and a memory 402.

[0087] The processor 401 can be a central processing unit (CPU) or other forms of processing units with data processing capabilities and / or instruction execution capabilities, and can control other components in the electronic device 400 to perform desired functions.

[0088] The memory 402 may include one or more computer program products, and the computer program products may include various forms of computer-readable storage media, such as volatile memory and / or non-volatile memory. The volatile memory may include, for example, random access memory (RAM) and / or cache memory, etc. The non-volatile memory may include, for example, read-only memory (ROM), hard disk, flash memory, etc. One or more computer program instructions may be stored on the computer-readable storage media, and the processor 401 may run the program instructions to implement a method for predicting power battery attenuation based on a deep autoencoder and / or other desired functions of any embodiment of the present invention described above. Various contents such as initial external parameters, thresholds, etc. may also be stored in the computer-readable storage media.

[0089] In one example, the electronic device 400 may further include: an input device 403 and an output device 404, and these components are interconnected through a bus system and / or other forms of connection mechanisms (not shown). The input device 403 may include, for example, a keyboard, a mouse, etc. The output device 404 may output various information to the outside, including warning prompt information, braking force, etc. The output device 404 may include, for example, a display, a speaker, a printer, and a communication network and its connected remote output devices, etc.

[0090] Of course, for simplicity, Figure 3 only some of the components related to the present invention in the electronic device 400 are shown, and components such as buses, input / output interfaces, etc. are omitted. In addition, according to specific application scenarios, the electronic device 400 may further include any other appropriate components.

[0091] In addition to the above methods and devices, an embodiment of the present invention may also be a computer program product, which includes computer program instructions that cause the processor to execute the steps of a method for predicting power battery attenuation based on a deep autoencoder provided by any embodiment of the present invention when the computer program instructions are run by the processor.

[0092] The computer program product may be written in any combination of one or more programming languages to write program code for performing the operations of the embodiments of the present invention. The programming languages include object-oriented programming languages, such as Java, C++, etc., and also include conventional procedural programming languages, such as the "C" language or similar programming languages. The program code may be executed entirely on the user computing device, partially on the user device, executed as an independent software package, partially on the user computing device and partially on a remote computing device, or entirely on a remote computing device or server.

[0093] In addition, an embodiment of the present invention may also be a computer-readable storage medium storing computer program instructions, which, when run by a processor, cause the processor to execute the steps of a method for predicting power battery attenuation based on a deep autoencoder provided in any embodiment of the present invention.

[0094] The above are only embodiments of the present invention. Specific structures and common knowledge such as characteristics well known in the art are not described in detail here. Those of ordinary skill in the art know all the common general technical knowledge in the technical field to which the invention pertains before the filing date or the priority date, can know all the prior art in this field, and have the ability to apply conventional experimental means before this date. Those of ordinary skill in the art can, with the inspiration given in this application and combined with their own abilities, complete and implement this solution. Some typical well-known structures or well-known methods should not be an obstacle for those of ordinary skill in the art to implement this application. It should be noted that for those skilled in the art, without departing from the structure of the present invention, several deformations and improvements can be made, which should also be regarded as the protection scope of the present invention, and these will not affect the implementation effect of the present invention and the practicality of the patent. The protection scope required by this application should be based on the content of its claims, and the specific implementation manners and the like described in the specification can be used to interpret the content of the claims.

Claims

1. A method for predicting the attenuation of power batteries based on a deep autoencoder, characterized in that: Including: S1: Collect the operation data of the new energy vehicle throughout its life cycle and the offline charging detection data to obtain sample data. After preprocessing the sample data, generate the preprocessed sample data. S2: Construct a deep autoencoder model. Divide the preprocessed sample data into a training set, a validation set, and a test set according to a preset ratio. Input the training set into the deep autoencoder model, and add a sparse constraint to the deep autoencoder model. After outputting the reconstructed data, by constructing a loss function based on the reconstruction error and the sparse constraint, use an optimization algorithm to adjust the parameters of the deep autoencoder model according to the gradient information of the loss function until the deep autoencoder model is trained. Then, use the validation set for validation and the test set for testing respectively. S3: Collect the real-time operation data of the current target new energy vehicle as the original data. After preprocessing, input it into the deep autoencoder model that has been verified and tested to output the reconstructed data. At the same time, calculate the reconstruction error between the original data and the reconstructed data, and associate the reconstruction error with the state of health of the power battery, which is characterized as the power battery health index. S4: Based on the time series data of the power battery health index, use the bidirectional LSTM+Attention mechanism to predict the degradation trend and remaining life of the power battery.

2. The method for predicting the attenuation of power batteries based on a deep autoencoder according to claim 1, characterized in that: The S1 includes: S1-1: Collect the charging and discharging currents, voltages, temperatures, humidities, and operating environment parameters of the load of the power battery of the new energy vehicle, as well as the vehicle state data such as the driving mileage and driving mode of the new energy vehicle, and the offline charging detection data of the new energy vehicle. Generate sample data according to the operating environment parameters, vehicle state data, and offline charging detection data. S1-2: Perform missing value filling processing, feature normalization processing, and time alignment processing on the sample data to obtain the preprocessed sample data.

3. The method for predicting the attenuation of power batteries based on a deep autoencoder according to claim 1, characterized in that: The S2 includes: S2-1: Determine the dimension of the input layer of the deep autoencoder model according to the features in the training set. S2-2: Construct the encoder of the deep autoencoder model. In the encoder, compress the data of the input training set into a low-dimensional feature space through a multi-layer neural network, and the number of neurons in each layer decreases according to the trend; and add a sparse constraint in the encoder; S2-3: Construct the decoder of the deep autoencoder model, and restore and denoise the compressed low-dimensional features in the encoder through the decoder to obtain the reconstructed data. S2-4: Output the reconstructed data through the output layer of the deep autoencoder model, and calculate the reconstruction error that can characterize the power battery health index according to the reconstructed data to generate a complete deep autoencoder model. S2-5: Train and optimize the deep autoencoder model based on a preset loss function and optimization algorithm to obtain the final deep autoencoder model.

4. The method for predicting the attenuation of power batteries based on a deep autoencoder according to claim 3, characterized in that: The specific method of adding a sparse constraint to the encoder in S2-2 is: Define the target sparsity , and limit the activation degree of the hidden layer to be less than 15%; Then the actual sparsity calculation formula is: Usage The divergence penalty term constrains sparsity, The expression of the divergence penalty term is: Among them, represents the sparsity estimation value of the th layer, represents the number of features, represents the number of neurons, represents the th layer's th hidden layer state in the 5. The method for predicting the attenuation of power batteries based on a deep autoencoder according to claim 4, characterized in that: The specific method of training and optimizing the deep autoencoder model based on a preset loss function and optimization algorithm in S2-5 to obtain the final deep autoencoder model is: Determine that the total loss function of the deep autoencoder model consists of an initial loss function and a sparse constraint. Use the reconstruction error as the initial loss function to measure the difference between the initial data of the input validation set and the reconstructed data. Optimize the model parameters by minimizing the loss function. Among them, the reconstruction error expression is: Among them, represents the initial data, represents the reconstructed data; Then the total loss function expression is: ， Among them, represents the weight coefficient of the sparse loss; Invoke the optimization algorithm to adjust the weights and biases of the deep autoencoder model according to the gradient information of the total loss function; and adjust the hyperparameters of the deep autoencoder model during the training process, determine the optimal hyperparameter combination through cross-validation, and then call the test set for testing until the final deep autoencoder model that meets the requirements is obtained.

6. A method for predicting power battery attenuation based on a deep autoencoder according to claim 5, characterized in that: The S3 includes: S3-1: Collect the real-time operation data of the current target new energy vehicle as the original data, and after preprocessing the original data, obtain the preprocessed original data; S3-2: Input the preprocessed original data into the final deep autoencoder model, extract and reconstruct the combined original signal features and differential entropy as low-dimensional features to obtain the reconstructed data; the expression of the differential entropy is: wherein represents the probability distribution of the time differential of the feature at the time point ; S3-3: Calculate the reconstruction error between the original data and the reconstructed data, and associate the reconstruction error with the state of health of the power battery, which is characterized as the power battery health index. The association expression is: Among them, represents the state of health of the power battery, is a fully connected layer autoencoder.

7. A method for predicting power battery attenuation based on a deep autoencoder according to claim 6, characterized in that: In the S3-2, extracting and reconstructing the combined original signal features and differential entropy as low-dimensional features to obtain the reconstructed data specifically includes: Concatenate the original signal features and differential entropy features into joint features; wherein and are respective dimensions; Compress the joint high-dimensional features into a low-dimensional latent space. The expression is: Among them, represents the encoder function, represents the combined multi-source features after normalization, is the encoder weight, is the encoder bias, represents the weight matrix of the first fully-connected layer of the encoder, represents the weight matrix of the second fully-connected layer of the encoder, represents the bias vector of the first layer of the encoder, represents the bias vector of the second layer of the encoder, is the activation function, is the low-level feature ; Finally, obtain the reconstruction term and the reconstruction error through the deep autoencoder constraint function.

8. A method for predicting power battery attenuation based on a deep autoencoder according to claim 7, characterized in that: The S4 includes: S4-1: Extract the time series data in the power battery health index; S4-2: Construct a forward and a backward network, and call the network to process the forward time series data and the backward time series data respectively; The network expression is: Among them, denotes the network, denotes the forward network, denotes the reverse network, denotes the feature data set, denotes the sequence input data of time step ; denotes the hidden state of time step ; denotes the hidden state of the subsequent step, denotes the dimension of the single-layer hidden state; S4-3: Introduce the attention mechanism, calculate scores for the hidden states of each of the above time series through a fully connected layer, and the attention score is expressed as: Among them, , represents the scientific system parameter, represents the transpose of matrix V, is the attention dimension, represents the bias of the hidden layer, represents the hidden state; S4-4: Use The activation function generates attention weights to weight the concatenated hidden states to change the importance. The expression is as follows: ， ； Among them, is the attention weight, is the attention score, represents the attention score at time step k, where w is the length of the time series; and then the context vector with importance is obtained Input to the fully connected layer for prediction , and its expression is: Among them, represents the weight of the fully connected layer, represents the bias term of the fully connected layer; S4-5: Set a sliding window to realize dynamic adjustment and online update of the model parameters, arrange the power battery health index in the order of time series, and increase the weight allocation according to the key attenuation stage of the power battery.

9. An electronic device, characterized in that: It includes a processor and a memory. Programs or instructions are stored in the processor. The processor executes a method for predicting the attenuation of a power battery based on a deep autoencoder according to any one of claims 1-8 above by calling the programs and instructions stored in the memory.

10. A computer-readable storage medium, characterized in that: The computer-readable storage medium stores programs or instructions, and the programs or instructions cause the computer to execute a method for predicting the attenuation of a power battery based on a deep autoencoder according to any one of claims 1-8 above.

Images