Lithium battery life prediction method based on DCAE-Transformer
By combining convolutional denoising autoencoder and Transformer network, the problem of complex data acquisition and difficult long-sequence data processing in SOH estimation of lithium-ion batteries is solved, and the accuracy and robustness of lithium-ion battery life prediction is improved, providing a An efficient method for monitoring health status.
Patent Information
- Application Number
- CN202411167255.7
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2024-08-23
- Publication Date
- 2024-11-29
AI Technical Summary
The capacity and performance of lithium-ion batteries are attenuated after charging and discharging cycles. The existing SOH estimation methods are difficult to maintain stable predictive performance in complex data environments, and the long-sequence data processing efficiency is low, resulting in a decrease in model training efficiency.
Using the lithium battery life prediction method based on DCAE-Transformer, the convolutional denoising autoencoder and Transformer network combine to remove data noise, extract key features, and process time series data through self-attention mechanism and feedforward neural network to improve the The robustness and prediction accuracy of the model.
Improve the accuracy and robustness of lithium battery life prediction, reduce interference during training, enhance the ability to capture real signals of battery performance, solve the problems of complex data acquisition and difficult long-sequence data processing, and provide a An efficient and reliable health status monitoring method.
Smart Images

Figure CN119044778A_ABST
Abstract
Description
Technical Field
[0001] The present invention relates to the technical field of state of health (SOH) monitoring of lithium-ion batteries, and in particular to a lithium battery life prediction method based on DCAE-Transformer. Background Art
[0002] Lithium-ion batteries are widely used in portable electronic devices, electric vehicles, and energy storage systems due to their high energy density, long cycle life, and low self-discharge rate. As these applications continue to increase the requirements for battery performance and life, the state of health (SOH) monitoring of lithium-ion batteries has become a key task in battery management systems (BMS). However, after multiple charge and discharge cycles, the capacity and performance of lithium-ion batteries will gradually decay, directly affecting the service life and safety of the battery. Therefore, how to accurately estimate the SOH of lithium-ion batteries and keep abreast of the battery's operating conditions is of great practical significance.
[0003] SOH is an important indicator for evaluating the degree of battery aging, usually expressed as the ratio of the current capacity of the battery to its initial capacity. Accurate estimation of SOH helps to timely warn the battery's usage status, avoid waste of resources caused by premature battery replacement, or safety hazards caused by delayed replacement. SOH estimation is of great significance for extending battery life and improving system reliability. Therefore, it is particularly important to study effective SOH estimation methods to improve battery efficiency and safety. Although existing methods have solved the SOH estimation problem to a certain extent, there are still the following challenges:
[0004] Complex data collection: The data collection process of lithium batteries is complex, involving multiple influencing factors such as temperature, charge and discharge rate, current, voltage, etc. These factors will interfere with the model's accurate prediction ability, making it difficult for the model to maintain stable prediction performance under various working conditions. Therefore, it is necessary to establish a model that can improve anti-interference capabilities to improve the accuracy of SOH estimation.
[0005] Long sequence data processing is difficult: The life cycle of lithium batteries is long, and the amount of monitoring data is large and complex. Traditional deep learning networks such as recurrent neural networks (RNNs) and long short-term memory networks (LSTMs) often face the problem of gradient vanishing or gradient exploding when processing such long-term dependent data. These problems will lead to low training efficiency and reduced model prediction performance. Therefore, how to effectively process long sequence data and improve the training efficiency and prediction performance of the model is an urgent problem to be solved in SOH estimation. Summary of the invention
[0006] In order to solve the above technical problems, the present invention proposes a lithium battery life prediction method based on DCAE-Transformer, which can effectively remove noise in the data, improve the prediction accuracy and robustness of the model, and can efficiently and reliably predict the life of lithium batteries.
[0007] The technical solution adopted by the present invention is:
[0008] The lithium battery life prediction method based on DCAE-Transformer includes the following steps:
[0009] Step 1: Collect the capacity degradation dataset of lithium batteries:
[0010] Step 2: preprocessing the capacity degradation data set;
[0011] Step 3: Build a prediction model based on DCAE-Transformer;
[0012] Step 4: Use the trained DCAE-Transformer prediction model to predict the health status of the lithium battery and output the remaining life of the lithium battery.
[0013] Described step 1 comprises the following steps:
[0014] S1.1. Data acquisition: Use data acquisition equipment to record battery capacity degradation data, including charging voltage, discharge current, and temperature parameters.
[0015] S1.2. Data recording: The collected battery capacity degradation data shall be stored in the database to ensure the integrity and security of the data. Each battery capacity degradation data record shall contain a timestamp and relevant operating condition information.
[0016] Described step 2 comprises the following steps:
[0017] S2.1. Data cleaning: Check the battery capacity degradation data in the database and delete missing values and outliers. Specifically, use interpolation to fill missing data and linear interpolation to ensure data continuity. Identify and remove abnormal data, identify outliers through the statistical analysis method Z-Score method, and remove values that exceed three times the standard deviation;
[0018] S2.2, Data standardization: Standardize the cleaned data using the Z-Score standardization method. The specific steps are as follows:
[0019] Calculate the mean μ and standard deviation σ of the data. The calculation formula for the mean μ is:
[0020]
[0021] n represents the number of data points in the data set, that is, the total number of samples; x i It represents the specific value of the i-th data point in the data set; the calculation formula of the standard deviation σ is:
[0022]
[0023] Each data point x is standardized as follows:
[0024]
[0025] In the above formula, z represents the standardized data value, also known as the standard score;
[0026] S2.3. Data set division: Divide the standardized data into training set and test set to ensure the randomness and representativeness of the data during division.
[0027] In step 3, the DCAE-Transformer prediction model specifically includes:
[0028] 3.1: Convolutional Denoising Autoencoder DCAE:
[0029] 3.1.1: Noise addition to input data: To improve the robustness of the model, Gaussian noise is added to the original input data to obtain the noisy data.
[0030] 3.1.2: Convolutional layer feature extraction:
[0031] a: Design convolutional layer:
[0032] Select the size and number of convolution kernels: Specifically, select 3 convolution layers with kernel sizes of 3×3, 5×5, and 7×7, and numbers of convolution kernels of 32, 64, and 128 respectively; initialize the weight matrix and bias vector.
[0033] The noisy data Input the convolution layer and calculate the hidden feature h. The formula is as follows:
[0034]
[0035] In the above formula, It represents the hidden features extracted by the convolution layer. These hidden features are the input data after adding noise. The output obtained after the convolution operation and nonlinear activation function processing. This process is to extract a high-dimensional feature representation of the input data to capture important patterns and structures in the data.
[0036] ReLU is a commonly used activation function that converts negative values in the input to 0 and keeps positive values unchanged, thereby introducing nonlinearity. The definition of ReLU is as follows:
[0037] ReLU(z)=max(0,z)
[0038] ReLU activation function maps negative values in input z to 0, and positive values remain unchanged; z represents the linear combination result after convolution operation, that is, the input data The value after the convolution kernel operation and the addition of the bias term; max(0,z) represents the operation of the ReLU activation function.
[0039] b: Deconvolution layer reconstruction:
[0040] Design the deconvolution layer, specifically select 3 deconvolution layers, the convolution kernel sizes are 7×7, 5×5 and 3×3, and the number of convolution kernels are 128, 64 and 32 respectively; initialize the deconvolution weight matrix W′ and the bias vector b′. Input the hidden feature h into the deconvolution layer and reconstruct the denoised output data The formula is as follows:
[0041]
[0042] In the above formula, Represents the reconstruction result of the model for the denoised data.
[0043] 3.1.3: Model training: Using the cleaned and standardized training set data, the gradient descent method is used to optimize the convolutional denoising autoencoder DCAE model parameters as follows:
[0044] By calculating the partial derivative of the loss function with respect to the model parameters, i.e. the gradient, the direction and magnitude of the parameter adjustment are indicated, thereby gradually reducing the value of the loss function. The gradient is the rate of change of the loss function relative to the parameter, indicating how to adjust the parameter to reduce the loss under the current parameter value. A positive gradient indicates that the loss increases and the parameter needs to be reduced, while a negative gradient indicates that the loss decreases and the parameter needs to be increased.
[0045] The parameters include the weight matrix W and bias b of the convolution layer, the weight matrix W′ and bias b′ of the deconvolution layer, and the parameter update formula is as follows:
[0046] Weight parameter W is updated:
[0047]
[0048] Bias b update:
[0049]
[0050] For the weight matrix W′ and bias b′ of the deconvolution layer:
[0051] Weight matrix W′ is updated:
[0052]
[0053] Bias b′ update
[0054]
[0055] Among them, the mean square error MSE is used as the loss function, and MSE can be defined as:
[0056]
[0057] In the above formula, n represents the number of data samples, x i represents the i-th original noise-free input data; represents the i-th output data reconstructed by the model. η represents the learning rate.
[0058] By repeating, each iteration: Calculate the output of the model using the current parameters Calculate the loss function Loss, calculate the gradient of each parameter through back propagation, and update the model parameters.
[0059] Through multiple iterations, the model gradually learns how to denoise the output data. As close as possible to the original input data x, an optimized denoising autoencoder model is finally obtained.
[0060] 3.2: Transformer network model:
[0061] 3.2.1: Positional encoding:
[0062] Add position encoding to each time series data point to introduce time dimension information. Use sine and cosine functions for encoding. The specific formula is as follows:
[0063]
[0064]
[0065] In the above formula, PE is the position encoding matrix; d model is the model dimension; t is the time index; k is the dimension index; PE(t,2k) represents the value of the 2kth dimension at the time index t in the position encoding matrix; PE(2,2k+1) represents the value of the 2k+1th dimension at the time index t in the position encoding matrix.
[0066] 3.2.2: Multi-head self-attention mechanism:
[0067] Initialize the weight matrix W of the query vector Q, key vector K and value vector V Q , W E and W F , calculate the attention weights, and perform weighted summation to get the output of each head:
[0068]
[0069] In the above formula, head i Represents the attention calculation result of the i-th attention head on the input. It represents the weight matrix used to calculate the query vector A of the i-th head. It represents the weight matrix used to calculate the key vector B of the i-th head. It represents the weight matrix used to calculate the value vector C of the i-th head. Attention() represents the calculation process of the attention mechanism. The attention weight is calculated by the interaction of the query vector and the key vector, and then the value vector is weighted and summed to obtain the output of the attention.
[0070] Among them, the attention weight is calculated: the attention score is calculated by the dot product between the query vector A and the key vector B, and then these scores are scaled and normalized by the softmax function.
[0071]
[0072] In the above formula, d8 represents the dimension of the key vector, which is used to scale the dot product result to prevent the value from being too large.
[0073] For a vector z=[z1,z2,…,z n ], the definition of Softmax function:
[0074]
[0075] In the above formula, z i Represents the i-th element in vector z. is the natural exponential function, which means z i The exponential value, denominator It is the sum of the exponential values of all elements and is used to normalize the exponential value of each element.
[0076] The outputs of multiple heads are concatenated and the final output is obtained through linear transformation:
[0077] MultiHead(A,B,C)=Concat(head1,…,head h )W O
[0078] In the above formula, MultiHead(A,B,C) represents the final output of the multi-head self-attention mechanism, which is the result of concatenating and linearly transforming the outputs of multiple attention heads; head1,…,head h Represents the output of each independent attention head. Each head calculates attention and outputs features in different feature spaces; Concat(head1,…,head h ) means concatenating the outputs of all attention heads along a specific dimension to form a larger comprehensive feature vector; W b Represents the weight matrix used for the final linear transformation, mapping the concatenated feature vector to the output space of the model.
[0079] 3.2.3: Design of Feedforward Neural Network:
[0080] Design a feedforward neural network, including two linear layers and a ReLU activation function, initialize the linear transformation matrices: W3 and W4, and the bias vectors: b3 and b4;
[0081] The output of the multi-head self-attention mechanism is input into the feedforward neural network for nonlinear transformation:
[0082] FFN(x)=max(0,xW3+b3)W4+b4
[0083] In the above formula, FFN(x) represents the output obtained after the feedforward neural network processes the input, which has undergone two linear transformations and a ReLU activation function; it is the data input to the feedforward neural network, which comes from the output of the multi-head self-attention mechanism. W3 represents the weight matrix of the first linear layer, which is used to map the input to the hidden layer; b3 represents the bias vector of the first linear layer, which is used to adjust the output of the hidden layer; W4 represents the weight matrix of the second linear layer, which is used to map the output of the hidden layer to the final output layer; b4 represents the bias vector of the second linear layer, which is used to adjust the result of the final output layer. Through this feedforward neural network, the Transformer model can further process and transform features from the attention mechanism, enhancing the model's ability to capture complex patterns.
[0084] The depth of the Transformer network model is set to d = 1, and the hidden layer size is set to h = 32.
[0085] 3.2.4: Model training:
[0086] The standardized training set data is used to train the Transformer network model using the Adam optimization algorithm. The Adam optimization algorithm uses the first-order momentum (exponential moving average of the gradient) and the second-order momentum (exponential moving average of the square of the gradient) to update the model parameters. By introducing these momentums, Adam can reduce oscillation while maintaining a faster convergence speed, thereby more stably finding the global optimal solution in a complex loss function space.
[0087] In each iteration, the Adam optimization algorithm updates the parameters of the model by following these steps:
[0088] ① Calculate the gradient:
[0089] First, calculate the loss function L(θ t ) with respect to the parameter θ:
[0090]
[0091] In the above formula, t represents the current time step, that is, the number of iterations; g t It represents the loss function L(θ t ) for the current model parameters θ t The gradient of It represents the gradient operator after the model parameter θ is differentiated.
[0092] ② Calculation of first-order momentum (exponential moving average of gradient):
[0093] Calculate the first-order momentum m t , which is the exponential moving average of the gradient:
[0094] m t =β1·m t-1 +(1-β1)·g t
[0095] In the above formula, β1 is the decay rate of the first-order momentum, which is 0.9; m t-1 It represents the value of the first-order momentum at the t-1th iteration, that is, the exponential moving average of the gradient in the previous time step.
[0096] ③. Calculation of second-order momentum (exponential moving average of squared gradient):
[0097] Calculate the second-order momentum v t , which is the exponential moving average of the squared gradient:
[0098]
[0099] In the above formula, β2 is the decay rate of the second-order momentum, which is 0.999. t-1It represents the value of the second-order momentum at the t-1th iteration, that is, the exponential moving average of the square of the gradient in the previous time step.
[0100] ④. Deviation correction:
[0101] Because the estimated value of momentum may be biased towards zero in the initial stage, in order to correct the deviation of the first-order momentum and the second-order momentum, the following correction is made:
[0102]
[0103]
[0104] In the above formula, represents the first-order momentum (exponential moving average of the gradient) after bias correction at the tth iteration;
[0105] represents the bias-corrected second-order momentum (exponential moving average of the squared gradient) at the tth iteration;
[0106] It represents the t-th power of the first-order momentum decay rate β1, reflecting the accumulation of the decay effect as the number of iterations t increases;
[0107] It represents the t-th power of the second-order momentum decay rate β2, reflecting the accumulation of the decay effect as the number of iterations t increases.
[0108] ⑤、Parameter update:
[0109] Update the model parameters θ using the corrected momentum value:
[0110]
[0111] In the above formula, τ represents the learning rate, and ∈ is a very small number used to prevent division by zero errors, with a value of 10 -8 θ t+1 It represents the updated model parameters after the t+1th iteration.
[0112] When using the Adam optimization algorithm to train the Transformer model, the specific steps include: Step 1: Initialize parameters: Initialize the model parameters θ, first-order momentum m0, and second-order momentum v0.
[0113] Step 2: Calculate the loss: Use the standardized training set data to calculate the loss function value through forward propagation, and use the mean absolute error MAE here;
[0114] Step 3: Calculate the gradient: Calculate the gradient g of the loss function for each parameter through back propagationt ;
[0115] Step 4: Update momentum: Use gradient to update first-order and second-order momentum m t and v t ;
[0116] Step 5: Correct the momentum deviation: Calculate the corrected momentum and Step 6: Update parameters: Update the parameters θ according to the corrected momentum value and learning rate τ.
[0117] Step 7: Monitor loss: After each iteration, monitor the change of the loss function and determine whether to use an early stopping strategy to prevent overfitting.
[0118] Learning rate τ = 0.0001, learning regularization parameter α = 10 -4 , the loss function uses the mean absolute error MAE. During the training process, the loss function value is monitored and an early stopping strategy is implemented to prevent overfitting.
[0119] 3.3: Combining the convolutional denoising autoencoder DCAE with the Transformer network model includes the following steps:
[0120] 3.3.1: Data preprocessing: The denoised data of the convolutional denoising autoencoder DCAE is used as input and provided to the Transformer network model for processing.
[0121] 3.3.2: Model training:
[0122] Jointly train the convolutional denoising autoencoder DCAE and Transformer network model to optimize the DCAE-Transformer prediction model parameters. The joint training steps are:
[0123] (1) Forward propagation:
[0124] 1) DCAE processes input data:
[0125] First, the convolutional denoising autoencoder DCAE receives the original input data, removes noise and extracts important features, and outputs the denoised data; This is the denoised data output by the convolutional denoising autoencoder DCAE.
[0126] 2) Transformer processes DCAE output:
[0127] Next, the denoised data Input into the Transformer network model. The Transformer network further processes the data through its self-attention mechanism and feedforward neural network, and outputs the final prediction result y. The whole process is continuous, and the convolutional denoising autoencoder DCAE and the Transformer network model are closely combined. That is, the input of the Transformer is the output of the DCAE.
[0128] (2) Calculate the loss function:
[0129] The loss function is calculated using the difference between the actual remaining life (RUL) value of the lithium battery and the life value predicted by the model. (3). Back propagation and parameter update:
[0130] 1) Back propagation: Based on the value of the loss function, the back propagation algorithm is used to calculate the gradients of all parameters in the DCAE-Transformer prediction model. These parameters include: the convolutional layer weights and biases of the DCAE; all weight matrices of the Transformer network, and W in the attention mechanism. Q , W E and W F , W3 and W4 in the feedforward neural network, and the bias vectors: b3 and b4.
[0131] 2) Parameter Update: All these parameters are updated using the optimization algorithm Adam to minimize the loss function.
[0132] Through joint training, the parameters of the DCAE and Transformer networks are jointly optimized so that the model can better denoise and accurately predict the remaining life (RUL) of lithium batteries. The entire process is end-to-end, which means that the model can dynamically adjust the denoising and feature extraction strategies based on the feedback of the prediction task, thereby improving the accuracy and robustness of the prediction.
[0133] 3.4: Model stability and training efficiency:
[0134] 3.4.1: Residual connection: Add residual connection between each layer of the Transformer network model to alleviate the gradient disappearance problem. The specific formula is as follows:
[0135] Output=LayerNorm(x+Sublayer(x))
[0136] In the above formula, Output represents the final output of the current layer after residual connection and layer normalization; LayerNorm() represents the layer normalization operation, which is used to standardize the input to ensure the training stability of the model; represents the original input data; Sublayer(x) represents the processing result of the current layer.
[0137] 3.4.2: Layer Normalization: Normalize the output of each layer to stabilize the training process and accelerate model convergence. The formula is as follows:
[0138]
[0139] In the above formula, LayerNorm(x) represents the result of the layer normalization operation, which aims to standardize each feature of the input and provide additional flexibility through scaling and translation parameters γ and β; μ and σ represent the mean and standard deviation respectively; γ and β are represented as learnable parameters.
[0140] Described step 4 comprises the following steps:
[0141] 4.1: In order to ensure the consistency and accuracy of the prediction, the test set data also needs to be standardized in the same way, using the same mean u as the training data tra and standard deviation σ tra :
[0142]
[0143] In the above formula, x text represents the original test set data, Represents the standardized test set data.
[0144] 4.2: Input the standardized test set data into the trained DCAE-Transformer prediction model;
[0145] 4.2.1: Denoising by DCAE:
[0146] The test data is first denoised by the trained DCAE; DCAE uses convolutional and deconvolutional layers to extract features and remove noise to obtain denoised data.
[0147]
[0148] 4.2.2: Remaining life prediction through Transformer network model:
[0149] Next, the denoised data The data is input into the Transformer network model for processing. The Transformer network model uses the self-attention mechanism and feedforward neural network to perform deep feature extraction and prediction on the denoised data, and finally outputs the remaining life prediction value of each test sample.
[0150]
[0151] 4.3: Get the remaining life prediction value of each test sample:
[0152] Finally, the model output That is, the remaining life prediction value for each test sample. These prediction values represent the remaining life (RUL) of the lithium battery estimated by the model based on the input standardized test data.
[0153] 4.4: Evaluation indicators:
[0154] The prediction performance of the model is evaluated by the mean absolute error MAE, mean absolute percentage error MAPE, root mean square error RMSE and determination coefficient R2 evaluation indicators. The formula is as follows:
[0155]
[0156]
[0157]
[0158]
[0159] In the above formula, y i represents the true value, represents the predicted value, represents the mean of the true value, and n represents the number of samples.
[0160] The present invention provides a lithium battery life prediction method based on DCAE-Transformer, and the technical effects are as follows:
[0161] (1) The present invention improves the overall accuracy and reliability of predictions by reducing potential noise in the data, enhances the ability to capture real signals of battery performance, and reduces interference during the training process, thereby improving the accuracy of the model in predicting battery performance and its life.
[0162] (2) The present invention uses an absolute position encoding method to encode the time series data using sine and cosine functions in step 3.2.1, so that the Transformer network can accurately capture and express the time information in the data to accurately capture and express the time dimension in the data. Since the health state of lithium-ion batteries is closely related to the time series, absolute position encoding can more accurately depict the time attribute, which is crucial for estimating the health state of the battery, because the characteristics of battery performance changing over time are the key to making accurate predictions.
[0163] (3) The present invention optimizes the architecture of the original Transformer network model to better meet the needs of time series data analysis. The model architecture is simplified, and the robustness of the model in processing time series data is enhanced, thereby improving the reliability of prediction.
[0164] (4) The present invention proposes an improved DCAE-Transformer prediction model for health status monitoring of lithium-ion batteries. The model combines the convolutional denoising autoencoder DCAE and the Transformer network model architecture to reduce the noise in the data, improve the ability to capture the real signal of battery performance, and reduce interference during the training process, thereby improving the accuracy of the model in predicting battery performance and its life.
[0165] (5) In the present invention, the convolutional denoising autoencoder DCAE effectively removes noise from the data through convolution operation in the data preprocessing stage, extracts key features, and improves the clarity and effectiveness of the data;
[0166] The Transformer network model can better capture the long-term dependencies in time series data through the self-attention mechanism, thus improving the prediction performance of the model. In order to better meet the needs of time series data analysis, the Transformer network model has also been optimized, using the absolute position encoding method to more accurately express the time dimension information in the data. This optimization enables the model to better understand and utilize time information when processing time series data, thereby improving the robustness and prediction reliability of the model. The proposed DCAE-Transformer prediction model solves the problems of complex data collection and difficult long sequence data processing in lithium battery SOH estimation by combining the convolutional denoising autoencoder DCAE and the optimized Transformer network model architecture, improving the prediction accuracy and robustness of the model, and providing an efficient and reliable method for health status monitoring of lithium-ion batteries. BRIEF DESCRIPTION OF THE DRAWINGS
[0167] The present invention will be further described below in conjunction with the accompanying drawings and embodiments:
[0168] Figure 1 This is a flow chart of a lithium battery life prediction method based on DCAE-Transformer of the present invention.
[0169] Figure 2 This is the DCAE-Transformer model structure diagram.
[0170] Figure 3(a) is a graph showing the capacity decay curves of B0005, B0006, B0007 and B0018 lithium-ion batteries;
[0171] Figure 3(b) is the capacity decay curve of CS2_35, CS2_36, CS2_37, and CS2_38 lithium-ion batteries.
[0172] Figure 4(a) is a fitted curve of the predicted capacity for NASA B0005, with 60% training data and 40% test data;
[0173] Figure 4(b) is a fitted curve of the predicted capacity for NASA B0005, with 70% training data and 30% test data;
[0174] Figure 4(c) is a fitted curve of the predicted capacity for NASA B0005 with 80% training data and 20% test data;
[0175] Figure 4(d) is a graph of the predicted capacity error for NASA B0005 with 60% training data and 40% test data;
[0176] Figure 4(e) is a graph of the predicted capacity error for NASA B0005 with 70% training data and 30% test data;
[0177] Figure 4(f) is a graph of the predicted capacity error for NASA B0005 with 80% training data and 20% test data;
[0178] Figure 4(g) is a fitted curve of the predicted capacity of CALCECS2_35 with 60% training data and 40% test data;
[0179] Figure 4(h) is a fitted curve of the predicted capacity of CALCECS2_35 with 70% training data and 30% test data;
[0180] Figure 4(i) is a fitted curve of the predicted capacity of CALCECS2_35 with 80% training data and 20% test data;
[0181] Figure 4(j) is a graph of the predicted capacity error for CALCECS2_35 with 60% training data and 40% test data;
[0182] Figure 4(k) is a graph of the predicted capacity error for CALCECS2_35 with 70% training data and 30% test data;
[0183] Figure 4(l) is a graph of the predicted capacity error for CALCECS2_35 with 80% training data and 20% test data.
[0184] Figure 5(a) is a bar chart comparing the RMSE, MAPE, MAE, and R2 values of the three models for NASA B0006;
[0185] Figure 5(b) is a bar chart comparing the RMSE, MAPE, MAE, and R2 values of the three models for NASA B0007;
[0186] Figure 5(c) NASA B0018, RMSE, MAPE, MAE, and R2 numerical comparison bar chart of the three models;
[0187] Figure 5(d) CALCE CS2_36, RMSE, MAPE, MAE, and R2 numerical comparison bar chart of the three models;
[0188] Figure 5 (e) CALCE CS2_37, RMSE, MAPE, MAE, and R2 numerical comparison bar chart of the three models;
[0189] Figure 5(f) CALCE CS2_38, RMSE, MAPE, MAE, and R2 numerical comparison bar chart of the three models. DETAILED DESCRIPTION
[0190] The specific implementation modes of the present invention are described below in conjunction with the accompanying drawings so that those skilled in the art can better understand the present invention.
[0191] Figure 1 It is a flow chart of a method for predicting the life of a lithium battery of a DCAE-Transformer according to the present invention.
[0192] In this embodiment, if Figure 1 As shown, the present invention provides a DCAE-Transformer lithium battery life prediction method, comprising the following steps:
[0193] Step 1: Collect the capacity degradation dataset of lithium batteries:
[0194] Determine the data source: The present invention takes the public battery data set of NASA and CALCE as an example. The four batteries selected by NASA are B0005, B0006, B0007, and B0018, and the four batteries selected by CALCE are CS2_35, CS2_36, CS2_37, and CS2_38. The capacity decay curves of B0005, B0006, B0007, B0018 and CS2_35, CS2_36, CS2_37, and CS2_38 lithium-ion batteries are as follows: Figure 2 shown.
[0195] Data acquisition: Use data to collect battery capacity degradation data, including parameters such as charging voltage, discharge current, temperature, etc. The acquisition frequency should be high enough to capture subtle changes in battery performance.
[0196] Data recording: The collected data is stored in the database to ensure the integrity and security of the data. Each data record should contain a timestamp and relevant operating condition information.
[0197] Step 2: Data Preprocessing:
[0198] Data cleaning: Check the data in the database and delete missing values and outliers. Use interpolation to fill missing data and linear interpolation to ensure data continuity. Identify and remove outliers. Use the statistical analysis method Z-Score method to identify outliers and remove values that exceed three times the standard deviation.
[0199] Data standardization: The cleaned data is standardized using the Z-Score standardization method. The specific steps are as follows:
[0200] Calculate the mean μ and standard deviation σ of the data and standardize each data point. The formula is as follows:
[0201]
[0202] Dataset division: Divide the standardized data into training set and test set to ensure the randomness and representativeness of the data during division.
[0203] Step 3: Build a remaining life prediction model based on DCAE-Transformer:
[0204] Sub-step 3.1: Convolutional Denoising Autoencoder DCAE:
[0205] Input data noise: To improve the robustness of the model, Gaussian noise is added to the input data x to obtain the noisy data
[0206] Convolutional layer feature extraction:
[0207] Design the convolutional layers and select the size and number of convolution kernels. Specifically, select three convolutional layers with kernel sizes of 3×3, 5×5, and 7×7, and the number of convolution kernels of 32, 64, and 128. Initialize the weight matrix W and the bias vector b.
[0208] The noisy data Input the convolution layer and calculate the hidden feature h. The formula is as follows:
[0209]
[0210] Deconvolution layer reconstruction:
[0211] Design the deconvolution layer, specifically select 3 deconvolution layers, the convolution kernel sizes are 7×7, 5×5 and 3×3, and the number of convolution kernels are 128, 64 and 32. Initialize the deconvolution weight matrix W′ and the bias vector b′.
[0212] Input the hidden feature h into the deconvolution layer to reconstruct the denoised output The formula is as follows:
[0213]
[0214] Model training: Using the cleaned and standardized training set data, the DCAE model parameters are optimized using the gradient descent method to make the denoised output as close as possible to the original input data x.
[0215] Sub-step 3.2: Transformer network model:
[0216] Positional encoding:
[0217] Add position encoding to each time series data point to introduce time dimension information. Use sine and cosine functions for encoding. The specific formula is as follows:
[0218]
[0219]
[0220] Among them, PE is the position encoding matrix, d model is the model dimension, t is the time index, and k is the dimension index.
[0221] Multi-head self-attention mechanism:
[0222] Initialize the weight matrix W of the query vector Q, key vector K and value vector V Q , W E and W F .
[0223] Calculate the attention weights and perform weighted summation to get the output of each head:
[0224]
[0225] The outputs of multiple heads are concatenated and the final output is obtained through linear transformation:
[0226] MultiHead(A,B,C)=Concat(head1,…,head h )W O
[0227] Feedforward Neural Network:
[0228] Design a feedforward neural network consisting of two linear layers and a ReLU activation function. Initialize the linear transformation matrices: W3 and W4, and the bias vectors: b3 and b4.
[0229] The output of the multi-head self-attention mechanism is input into the feedforward neural network for nonlinear transformation:
[0230] FFN(x)=max(0,xW3+b3)W4+b4
[0231] The depth of the Transformer network is set to d = 1 and the hidden layer size is set to h = 32.
[0232] Model training: Using the standardized training set data, the Transformer model was trained using the Adam optimization algorithm, with a learning rate of τ = 0.0001 and a learning regularization parameter of α = 10 -4 , the loss function uses the mean absolute error MAE. During the training process, the loss function value is monitored and an early stopping strategy is implemented to prevent overfitting.
[0233] Sub-step 3.3: Combine DCAE with Transformer network model:
[0234] Data preprocessing: The DCAE denoised data is used as input and provided to the Transformer model for processing.
[0235] Model training: Jointly train the DCAE and Transformer models to optimize the overall model parameters so that the final output of the remaining life prediction value of the lithium battery has high accuracy and robustness.
[0236] Sub-step 3.4: Model stability and training efficiency:
[0237] Residual connection: Add residual connection between each layer of the Transformer network model to alleviate the gradient disappearance problem. The specific formula is as follows:
[0238] Output=LayerNorm(x+Sublayer(x))
[0239] Among them, (x) is the input, and Sublayer(x) is the processing result of the current layer.
[0240] Layer normalization: Normalize the output of each layer to stabilize the training process and accelerate model convergence. The formula is as follows:
[0241]
[0242] Among them, and are the mean and standard deviation respectively, and γ and β are learnable parameters.
[0243] Figure 2 This is the DCAE-Transformer prediction model structure diagram:
[0244] Step 4: Model prediction and evaluation
[0245] Model prediction: Input the standardized test set data into the trained DCAE-Transformer model to obtain the remaining life prediction value of each test sample.
[0246] Performance evaluation:
[0247] Mean absolute error (MAE), mean absolute percentage error (MAPE), root mean square error (RMSE) and coefficient of determination (R2) evaluation indicators were used to assess the prediction performance of the model.
[0248] The formula is as follows:
[0249]
[0250]
[0251]
[0252]
[0253] Among them, y i is the true value, is the predicted value, is the mean of the true value and is the number of samples.
[0254] To illustrate the technical effect of the present invention, a complete data set with a ratio of 60%, 70% and 80% is used for training, and the remaining data sets (40%, 30% and 20%) are used for testing. CNN and CNN-LSTM are used to estimate the battery SOH through the same training strategy. The SOH estimates of B0005 and CS2_35 are listed.
[0255] Table 1 Comparison of prediction results of DCAE-Transformer based on B0005 and CS2_35 batteries and the other three models
[0256]
[0257] As can be seen from Table 1, the RMSE and MAE values of DCAE-Transformer are significantly lower than those of other models, indicating that its prediction error is smaller. 2 The value is close to 1, especially in the case of battery B0005 (60%, 70%, 80%), the R 2The values are all above 0.99, indicating that the model's prediction of battery life is very close to the actual value. For different types of batteries (B0005 and CS2_35), the DCAE-Transformer model performs well, especially for the CS2_35 battery, which is more difficult to predict, the DCAE-Transformer still performs significantly better than CNN and CNN-LSTM. This shows that the model has strong adaptability and robustness to different battery types.
[0258] Compared with the other two methods, the DCAE-Transformer prediction model has higher accuracy in SOH prediction. Figure 4(a) to Figure 4(l) Fit and error plots, NASA B0005: 60%, 70%, 80% fit (ac) and error (df), CALCE CS2_35: 60%, 70%, 80% fit (gi) and error (jl).
[0259] B0006, B0007, B0018, CS2_36, CS2_37, and CS2_38 are further used for testing and verification. 60% of the complete dataset is used for training, and the remaining 40% of the dataset is used for testing.
[0260] Table 2 Comparison of prediction results of DCAE-Transformer and other three models based on B0006, B0007, B0018 and CS2_36, CS2_37, CS2_38 batteries
[0261]
[0262] As can be seen from Table 2, in the tests of all batteries, the DCAE-Transformer model performed well in four indicators (RMSE, MAPE, MAE, R 2 ) and significantly outperformed the CNN and CNN-LSTM models. The RMSE (root mean square error) and MAE (mean absolute error) values show that the prediction error of DCAE-Transformer is significantly lower than that of the other two models, indicating that it has a stronger ability to fit the test data. 2 The (determination coefficient) value is close to 1, indicating that the DCAE-Transformer model's prediction results for lithium battery life are very close to the actual values and have extremely high accuracy. Whether it is B0006, B0007, B0018 or CS2_36, CS2_37, CS2_38, DCAE-Transformer has shown stable high performance in the prediction task of each battery. Especially for the B series batteries, the R 2 The values are all above 0.999, indicating that the model fits the data almost perfectly.
[0263] On the more challenging CS2 series cells, DCAE-Transformer still maintains a high accuracy, R 2 The value is between 0.9862 and 0.9964, far exceeding other models. DCAE-Transformer still maintains high prediction accuracy and stability when processing more complex and challenging battery data. This shows the strong adaptability and robustness of the DCAE-Transformer model in complex data scenarios.
[0264] Figure 5(a) to Figure 5(f) This is a bar chart comparing the RMSE, MAPE, MAE, and R2 values of the three models.
[0265] The present invention combines the denoising encoder and the convolution mechanism to effectively perform the feature extraction and denoising tasks. The performance of three models, CNN, CNN-LSTM and DCAE-Transformer, is compared. By using eight battery datasets (B0005, B0006, B0007, B0018, CS2_35, CS2_36, CS2_37 and CS2_38) provided by NASA and CALCE, the performance of these models under different training data ratios (60%, 70%, 80%) is evaluated.
[0266] The results show that the DCAE-Transformer model produces lower loss values during training and outperforms other models in multiple evaluation metrics, including lower mean absolute error (MAE), mean absolute percentage error (MAPE), and root mean square error (RMSE), as well as a higher coefficient of determination (R). 2 The error indicators MAE, MAPE and RMSE of the DCAE-Transformer prediction model on NASA battery samples are all less than 1%, and the R2 value exceeds 99.5%; on CALCE samples, the error indicator is less than 5%, and the R2 value exceeds 98%. The results show that the present invention has high accuracy and generalization in lithium battery SOH estimation.
Claims
1. A lithium battery life prediction method based on DCAE-Transformer, characterized by The following steps are involved: Step 1: Collect the capacity degradation dataset of lithium batteries: Step 2: preprocessing the capacity degradation data set; Step 3: Build a prediction model based on DCAE-Transformer; Step 4: Use the trained DCAE-Transformer prediction model to predict the health status of the lithium battery and output the lithium The remaining battery life.
2. The lithium battery life prediction method based on DCAE-Transformer according to claim 1 is characterized in that: Place The step 1 comprises the following steps: S1.
1. Data collection: Use data collection equipment to record battery capacity degradation data, including charging voltage, discharge current, and temperature parameters; S1.2, Data recording: The collected battery capacity degradation data is stored in the database. Each battery capacity degradation data The records contain a timestamp and relevant operating condition information.
3. The lithium battery life prediction method based on DCAE-Transformer according to claim 1 is characterized in that: Place The step 2 comprises the following steps: S2.1, Data cleaning: Check the battery capacity degradation data in the database and delete missing values and outliers; S2.2, Data standardization: The cleaned data is standardized using the Z-Score standardization method; The steps are as follows: Calculate the mean μ and standard deviation σ of the data. The calculation formula for the mean μ is: It represents the number of data points in the data set, that is, the total number of samples; it represents the specific value of the i-th data point in the data set; the calculation formula of the standard deviation σ is: Each data point is standardized using the following formula: In the above formula, represents the standardized data value; S2.
3. Dataset division: Divide the standardized data into training set and test set.
4. The lithium battery life prediction method based on DCAE-Transformer according to claim 1 is characterized in that: In step 3, the DCAE-Transformer prediction model specifically includes: 3.1: Convolutional Denoising Autoencoder DCAE: 3.1.1: Noise addition to input data: Add Gaussian noise to the original input data to obtain the noisy data 3.1.2: Convolutional layer feature extraction: a: Design convolutional layer: Select the size and number of convolution kernels: Specifically select 3 convolution layers, initialize the weight matrix and bias vector; The noisy data Input the convolution layer and calculate the hidden feature h. The formula is as follows: In the above formula, It represents the hidden features extracted by the convolution layer. These hidden features are the input data after adding noise. The output obtained after convolution operation and nonlinear activation function processing; ReLU represents the activation function, which is used to change the negative value in the input to 0 and keep the positive value unchanged, thereby introducing nonlinearity. The definition of ReLU is as follows: ReLU z=max 0,z ReLU activation function maps negative values in input z to 0, and positive values remain unchanged; z represents the linear combination result after convolution operation, that is, the input data The value after the convolution kernel operation and the addition of the bias term; max 0,z represents the operation of the ReLU activation function; b: Deconvolution layer reconstruction: Design the deconvolution layer, specifically select 3 deconvolution layers, initialize the deconvolution weight matrix % and the bias vector %; input the hidden feature h into the deconvolution layer, and reconstruct the denoised output data The formula is as follows: In the above formula, Represents the reconstruction result of the model on the denoised data; 3.1.3: Model training: Using the cleaned and standardized training set data, the gradient descent method is used to optimize the convolutional denoising autoencoder DCAE model parameters as follows: By calculating the partial derivative of the loss function with respect to the model parameters, i.e. the gradient, the direction and magnitude of the parameter adjustment are indicated, thereby gradually reducing the value of the loss function; the parameters include the weight matrix W and bias b of the convolution layer, the weight matrix % and bias % of the deconvolution layer, and the parameter update formula is as follows: Weight parameter W is updated: Bias b update: Weight matrix % and bias % for the deconvolution layer: Weight matrix % update: Bias % Update Among them, the mean square error MSE is used as the loss function, and MSE is defined as: In the above formula, n represents the number of data samples, and represents the / th original noise-free input data; represents the output data after the i-th model reconstruction; (represents the learning rate; By repeating, each iteration: Calculate the output of the model using the current parameters Calculate the loss function Loss, calculate the gradient of each parameter through back propagation, and update the model parameters; Through multiple iterations, the model gradually learns how to denoise the output data. As close as possible to the original input data, an optimized denoising autoencoder model is finally obtained; 3.2: Transformer network model: 3.2.1: Positional encoding: Add position encoding to each time series data point to introduce time dimension information; use sine and cosine functions for encoding, the specific formula is as follows: In the above formula, 01 is the position encoding matrix; @ model is the model dimension; 2 is the time index; 4 is the dimension index; 01 2,24 represents the value of the 24th dimension at the time index 2 in the position coding matrix; 01 2,24+1 represents the value of the 24+1th dimension at the time index 2 in the position coding matrix; 3.2.2: Multi-head self-attention mechanism: Initialize the weight matrices of query vector A, key vector B, and value vector C Q , E and F , calculate the attention weights, and perform weighted summation to get the output of each head: In the above formula, head i Represents the attention calculation result of the / th attention head on the input; It represents the weight matrix used to calculate the query vector A of the / th head; It represents the weight matrix used to calculate the key vector B of the / th head; It represents the weight matrix used to calculate the value vector C of the / th head; Calculate attention weights: Calculate attention scores by taking the dot product between the query vector A and the key vector B, and then scale and normalize these scores using the softmax function. In the above formula, @8 represents the dimension of the key vector, which is used to scale the dot product result to prevent the value from being too large; For a vector = U,,…,W, the definition of Softmax function is: In the above formula, represents the / th element in the vector; is the natural exponential function, which represents the exponential value and the denominator is the sum of the exponential values of all elements, used to normalize the exponential value of each element; The outputs of multiple heads are concatenated and the final output is obtained through linear transformation: MultiHead A,B,C=Concat head,…,head a b In the above formula, MultiHead A, B, C represents the final output of the multi-head self-attention mechanism, which is the result of concatenating and linearly transforming the outputs of multiple attention heads; head,…,head a Represents the output of each independent attention head. Each head calculates attention and outputs features in different feature spaces; Concat head,…,head a It means concatenating the outputs of all attention heads along a specific dimension to form a larger comprehensive feature vector; b represents the weight matrix used for the final linear transformation, mapping the concatenated feature vector to the output space of the model; 3.2.3: Design of Feedforward Neural Network: Design a feedforward neural network, including two linear layers and a ReLU activation function, and initialize the linear transformation matrix: c and d , and the bias vector: c and 4; The output of the multi-head self-attention mechanism is input into the feedforward neural network for nonlinear transformation: FFN(x)=max 0,xW c +b3)W d +b4 In the above formula, FFN represents the output obtained by the feedforward neural network after processing the input, which has undergone two linear transformations and a ReLU activation function; is the data input to the feedforward neural network, which comes from the output of the multi-head self-attention mechanism; c represents the weight matrix of the first linear layer, used to map the input to the hidden layer; c represents the bias vector of the first linear layer, which is used to adjust the output of the hidden layer; 4 represents the weight matrix of the second linear layer, which is used to map the output of the hidden layer to the final output layer; 4 represents the bias vector of the second linear layer, which is used to adjust the result of the final output layer; Through this feedforward neural network, the Transformer model can further process and transform the features from the attention mechanism; 3.2.4: Model training: Using the standardized training set data, the Adam optimization algorithm is used to train the Transformer network model.
5. The lithium battery life prediction method based on DCAE-Transformer according to claim 4 is characterized in that: The Adam optimization algorithm uses first-order momentum and second-order momentum to update the parameters of the model; in each iteration, the Adam optimization algorithm updates the parameters of the model through the following steps: ① Calculate the gradient: First, calculate the loss function *g h Gradient with respect to parameter θ: i h =j k *g h ; In the above formula, 2 represents the current time step, that is, the number of iterations; i h It represents the loss function *g at the tth iteration h For the current model parameter g h The gradient of k It represents the gradient operator after the model parameter θ is differentiated; ② Calculation of first-order momentum: Calculate the first-order momentum O h , which is the exponential moving average of the gradient: SHE h =l·O hn +1-l·i h In the above formula, β is the decay rate of the first-order momentum; hn It represents the value of the first-order momentum at the t-1th iteration; ③、Calculation of second-order momentum: Calculate the second-order momentum o h , which is the exponential moving average of the squared gradient: In the above formula, β is the decay rate of the second-order momentum; hn It represents the value of the second-order momentum at the t-1th iteration; ④. Deviation correction: Because the estimated value of momentum may be biased towards zero in the initial stage, in order to correct the deviation of the first-order momentum and the second-order momentum, the following correction is made: It represents the first-order momentum after bias correction at the tth iteration; It represents the second-order momentum after bias correction at the tth iteration; It represents the t-th power of the first-order momentum decay rate β, reflecting the accumulation of the decay effect as the number of iterations t increases; It represents the t-th power of the second-order momentum decay rate β, reflecting the accumulation of the decay effect as the number of iterations t increases; ⑤, parameter update: Update the model parameters θ using the corrected momentum value: In the above formula, t represents the learning rate, u is a very small number used to prevent division by zero errors; g hs It represents the updated model parameters after the t+1th iteration.
6. The lithium battery life prediction method based on DCAE-Transformer according to claim 5 is characterized in that: When using the Adam optimization algorithm to train the Transformer model, the specific steps include: Step 1: Initialize parameters: Initialize the model parameters θ, first-order momentum m0 and second-order momentum v0; Step 2: Calculate the loss: Use the standardized training set data to calculate the loss function value through forward propagation, and use the mean absolute error MAE here; Step 3: Calculate the gradient: Calculate the gradient of the loss function for each parameter i through back propagation h ; Step 4: Update momentum: Use gradient to update first-order and second-order momentum m t and v t ; Step 5: Correct the momentum deviation: Calculate the corrected momentum and Step 6: Update parameters: Update the parameter θ according to the corrected momentum value and learning rate t; Step 7: Monitor loss: After each iteration, monitor the change of the loss function and determine whether to use an early stopping strategy to prevent overfitting.
7. The lithium battery life prediction method based on DCAE-Transformer according to claim 4 is characterized in that: Combining the convolutional denoising autoencoder DCAE with the Transformer network model includes the following steps: 3.3.1: Data preprocessing: The denoised data of the convolutional denoising autoencoder DCAE is used as input and provided to the Transformer network model for processing; 3.3.2: Model training: Jointly train the convolutional denoising autoencoder DCAE and Transformer network model to optimize the DCAE-Transformer prediction model parameters. The joint training steps are: (1) Forward propagation: 1) DCAE processes input data: First, the convolutional denoising autoencoder DCAE receives the original input data, removes noise and extracts important features, and outputs the denoised data; It is the denoised data output by the convolutional denoising autoencoder DCAE; 2) Transformer processes DCAE output: Next, the denoised data Input into the Transformer network model; the Transformer network further processes the data through its self-attention mechanism and feedforward neural network, and outputs the final prediction result&; (2) Calculate the loss function: The loss function is calculated using the difference between the actual remaining life (RUL) value of the lithium battery and the life value predicted by the model; (3) Back propagation and parameter update: 1) Back propagation: Based on the value of the loss function, the back propagation algorithm is used to calculate the gradients of all parameters in the DCAE-Transformer prediction model; these parameters include: the convolutional layer weights and biases of the DCAE; all weight matrices of the Transformer network, and the attention mechanism. Q , E and F , in a feedforward neural network c and 4, and the bias vector: c and 4; 2) Parameter Update: All these parameters are updated using the optimization algorithm Adam to minimize the loss function.
8. The lithium battery life prediction method based on DCAE-Transformer according to claim 7 is characterized in that: The stability and training efficiency of the model specifically include: Residual connection: Add residual connection between each layer of the Transformer network model to alleviate the gradient disappearance problem. The specific formula is as follows: Output=LayerNormJ+Sublayer K In the above formula, Output represents the final output of the current layer after residual connection and layer normalization; LayerNorm() represents the layer normalization operation, which is used to standardize the input; represents the original input data; Sublayer represents the processing result of the current layer; Layer normalization: Normalize the output of each layer to stabilize the training process and accelerate model convergence. The formula is as follows: In the above formula, LayerNorm represents the result of layer normalization operation; μ and σ represent mean and standard deviation respectively; γ and β represent learnable parameters.
9. The lithium battery life prediction method based on DCAE-Transformer according to claim 1 is characterized in that: Described step 4 comprises the following steps: 4.1: In order to ensure the consistency and accuracy of the prediction, the test set data is also standardized in the same way, using the same mean u as the training data train and standard deviation σ train : In the above formula, x text represents the original test set data, Represents the standardized test set data; 4.2: Input the standardized test set data into the trained DCAE-Transformer prediction model: 4.2.1: Denoising by DCAE: The test data is first denoised by the trained DCAE; DCAE uses convolutional and deconvolutional layers to extract features and remove noise to obtain denoised data. 4.2.2: Remaining life prediction through Transformer network model: Next, the denoised data The data is input into the Transformer network model for processing. The Transformer network model uses the self-attention mechanism and feedforward neural network to perform deep feature extraction and prediction on the denoised data, and finally outputs the remaining life prediction value of each test sample. 4.3: Get the remaining life prediction value of each test sample: Finally, the model output That is, the remaining life prediction value for each test sample; these prediction values represent the remaining life (RUL) of the lithium battery estimated by the model based on the input standardized test data; 4.4: Evaluation indicators: The prediction performance of the model is evaluated by the mean absolute error MAE, mean absolute percentage error MAPE, root mean square error RMSE and determination coefficient R2 evaluation indicators. The formula is as follows: In the above formula, & represents the true value, represents the predicted value, represents the mean of the true value, and represents the number of samples.
Citation Information
Patent Citations
Underwater acoustic signal enhancing method based on automatic encoder
CN110456332A
Short-term daily load curve prediction method based on DCAE-LSTM
CN112785088A
Lithium battery residual life probability prediction method based on time convolution attention mechanism
CN116381517A
Lithium ion battery life prediction method based on integrated Transform model
CN117148197A
Cited By
Method, device and equipment for predicting remaining service life of battery, medium and product
CN119575222A
Electric ship battery health state prediction method and device
CN119885022A
Lithium ion battery health state estimation method and device based on Transform model
CN119902111A
Method for identifying and positioning pitting corrosion and cracks of solar photovoltaic panel under deep and far sea
CN119964017A
Battery charge state prediction method, system and device and storage medium
CN121027865A
