Echelon utilization battery SOH estimation method based on improved convolutional network

Abstract

The invention discloses a cascade utilization battery SOH estimation method based on an improved convolutional network, and the method comprises the steps: firstly collecting the constant-current charging stage operation data of a cascade utilization battery, and carrying out the preprocessing, and obtaining a standardized multi-source operation data set; secondly, on the basis of the standardized multi-source operation data, an energy accumulation model is constructed, and a denoised SOH tag sequence is obtained after processing; and then, based on the standardized multi-source operation data and the de-noised SOH tag sequence, extracting an energy slope, and carrying out feature fusion through a wavelet hierarchical patch mixer to obtain a multi-scale fusion feature. And finally, inputting the multi-scale fusion features into a lightweight ResCNN network, and mapping the multi-scale fusion features to SOH output, thereby realizing prediction of the state of health of the echelon utilization battery. According to the method, the efficient estimation of the SOH can be realized by fully utilizing the voltage curve and the multi-source operation data in the charging stage, and the test cost and the data acquisition difficulty are greatly reduced.

Description

TECHNICAL FIELD

[0001] The application belongs to the technical field of new energy, and relates to power battery gradient utilization technology, in particular to a gradient utilization battery SOH estimation method based on an improved convolutional network. BACKGROUND

[0002] After the rapid power battery of the new energy vehicle industry reaches the retirement standard (usually the rated capacity is attenuated to below 80%), it still has gradient utilization value in energy storage systems, low-speed electric vehicles and other low-power applications. Gradient utilization not only prolongs the overall life cycle of the battery and reduces the use cost, but also effectively reduces resource waste and environmental pollution, which is an important link for sustainable development of the new energy industry. The state of health (SOH) of the battery is an important indicator for evaluating its remaining value and safety. Traditional SOH measurement methods usually rely on capacity testing or complete charge-discharge cycles, which have problems such as long testing time, complex operation and difficulty in real-time monitoring, which seriously restricts the rapid evaluation and large-scale application of gradient utilization batteries. In the prior art, the prediction methods for battery SOH mainly include analytical methods based on equivalent circuit models and regression methods based on traditional machine learning. The equivalent circuit model method needs to accurately establish the internal mechanism model of the battery, but the gradient utilization battery has different aging degrees and complex attenuation mechanisms, and the model parameters are difficult to adapt, resulting in insufficient prediction accuracy. Although the traditional machine learning method can use limited features for prediction, it is difficult to effectively capture the complex nonlinear relationship between the charge curve, energy change and SOH, and has limited generalization ability under multiple working conditions. In recent years, deep learning methods have been widely used in battery performance prediction, but existing schemes mostly rely on full-cycle charge-discharge data as model input, which has high data collection cost and low training efficiency, and is not optimized for the attenuation characteristics of gradient utilization batteries, so the prediction robustness in the retired battery scenario is not ideal. SUMMARY

[0003] In view of the deficiencies of the prior art, the application provides a new SOH index and deep estimation method for gradient utilization batteries, which extracts voltage curve segments and voltage slopes in the charging stage as input, combines a wavelet layered patch mixer and a lightweight ResCNN network, and realizes efficient and accurate SOH prediction, thereby providing technical support for rapid evaluation of gradient utilization batteries.

[0004] A gradient utilization battery SOH estimation method based on an improved convolutional network includes the following steps:

[0005] S1, collect the constant current charging stage voltage curve and temperature, current and other operating data of the gradient utilization battery, and perform filtering, normalization and time alignment and other pretreatments to obtain a standardized multi-source operating data set, so as to ensure the consistency and availability of the data;

[0006] S2, based on the standardized multi-source operation data obtained in S1, an energy accumulation model is constructed, the accumulated energy of the voltage in the key voltage interval in the charging stage is calculated, and a new SOH estimation label based on energy accumulation is generated, and the label sequence is processed through Gaussian smoothing to obtain a denoised SOH label sequence, so as to reduce the influence of noise and enhance the trend characteristics;

[0007] S3, based on the standardized multi-source operation data obtained in S1 and the denoised SOH label sequence in S2, a multi-scale feature extraction mechanism is adopted in the time domain and the frequency domain, the energy slope is extracted, and local and global feature fusion is performed through a wavelet hierarchical patch mixer to obtain an optimized multi-scale fusion feature, realizing adaptive enhancement of key features and optimization of overall feature expression;

[0008] S4, the optimized multi-scale fusion feature obtained in S3 is input into a lightweight ResCNN network, deep time sequence features are captured through a residual convolution module, and are mapped to SOH output through global pooling and a fully connected layer, realizing rapid and accurate prediction of the battery state of health through gradient utilization, and improving the model convergence speed and prediction stability through batch training and loss function optimization strategy;

[0009] S5, in the test stage, the model performance is verified by calculating the mean absolute error (MAE), root mean square error (RMSE) and mean absolute percentage error (MAPE) of the predicted SOH and the real SOH output by S4, and the feature extraction parameters of S3 and the network parameters of S4 are iteratively optimized according to the error feedback to further improve the prediction accuracy and robustness.

[0010] The present application has the following advantages:

[0011] The present method is aimed at the health state prediction of the gradient utilization battery, fully utilizes the voltage curve and multi-source operation data in the charging stage, and does not need to rely on complex or special test equipment, so as to realize efficient estimation of SOH, greatly reduce the test cost and data acquisition difficulty. By constructing a new SOH index based on energy accumulation, and combining multi-scale feature extraction and wavelet hierarchical patch mixer for feature enhancement, the voltage change rate, energy slope and key degradation characteristics can be adaptively captured, and the local and global characteristics of the battery can be modeled at the same time. The lightweight ResCNN network further compresses the sequence, extracts deep time sequence features, and relieves the gradient vanishing problem through the residual structure, improves the prediction accuracy and robustness. The method is suitable for gradient utilization batteries in different aging states, and provides a reliable and generalizable technical solution for battery screening, health monitoring, life management and safety warning, and has strong practical application value. BRIEF DESCRIPTION OF DRAWINGS

[0012] Figure 1 Fig. 1 is a schematic diagram of a network structure in an embodiment of the present application. DETAILED DESCRIPTION

[0013] The present application will be further explained in connection with the accompanying drawings, wherein:

[0014] As shown in Figure 1 Fig. 1, a cascade utilization battery SOH estimation method based on an improved convolutional network comprises the following steps:

[0015] S1, collect the constant-current charging phase voltage curve and temperature, current and other operation data of the cascade utilization battery, and perform filtering, normalization and time alignment and other pretreatments to obtain a standardized multi-source operation data set, ensure the consistency and availability of the data, and provide basic data support for subsequent model construction and feature extraction;

[0016] S2, based on the standardized multi-source operation data obtained in S1, construct an energy accumulation model, calculate the accumulated energy value of the voltage in the key voltage interval in the charging phase, and generate a new SOH estimation label based on energy accumulation, and at the same time, process the label sequence through Gaussian smoothing to obtain a denoised SOH label sequence, so as to reduce the influence of noise and enhance the trend characteristics;

[0017] S2-1, calculate the instantaneous charging power according to the charging curve data in the pretreated standardized multi-source operation data:

[0018]

[0019] Where, P chg (t) is the charging instantaneous power, I(t) is the battery pack current, V pack (t) is the battery pack voltage, E chg (t) is the charging energy, is the charging energy change rate, when is greater than or equal to 0, it is the battery charging. By judging , the charging state of the battery can be judged, and the data set can be defined as k charge and discharge cycles.

[0020] S2-2, to indicate that the voltage of the maximum voltage of the single cell in the key voltage interval is the quasi-platform region, the indicator function is introduced:

[0021]

[0022] Where, U max (t) represents the maximum voltage of the single cell, when U max (t) is between V s and V m , χV(t) is 1, V s is the lower limit of the quasi-platform region, and Vm χV(t) is 0.

[0023] S2-3, integrate the instantaneous charging power in the voltage interval, i.e. the quasi-platform region, to obtain the cumulative energy of the kth charge-discharge cycle:

[0024] E k =∫χV(t)P chg (t)dt

[0025] where the integral of the instantaneous charging power in the charging phase and χV(t) is 1 is taken to obtain the cumulative energy value E k of the kth charge-discharge cycle.

[0026] S2-4, normalize the cumulative energy E k to obtain a new SOH estimation label SOH k :

[0027]

[0028] where E ref represents the nominal capacity of the battery.

[0029] S2-5, Gaussian smoothing is performed on the new SOH estimation label sequence {SOH1, SOH2,..., SOH N}, to obtain the denoised SOH label sequence:

[0030]

[0031] w j is the weight, which is given by the Gaussian kernel:

[0032]

[0033] where σ is the Gaussian coefficient, so the window half window k in the processing is approximately 3, and j is the discrete position index in the Gaussian window. Weight normalization is performed on w j to obtain

[0034]

[0035] S2-6, a new SOH label based on energy accumulation is constructed through the cumulative energy calculation and Gaussian smoothing processing, which is used for the training and prediction of the subsequent deep learning model;

[0036] S3, based on the standardized multi-source operation data obtained in S1 and the SOH label sequence after denoising in S2, multi-scale feature extraction mechanism is adopted in time domain and frequency domain to extract energy slope, and local and global feature fusion is performed through wavelet layered patch mixer to obtain optimized multi-scale fusion features, realizing adaptive enhancement of key features and optimization of overall feature expression;

[0037] S3-1, for each charge and discharge cycle, two types of health features are extracted to construct a final feature vector:

[0038] S3-1-1, in the charging stage, the voltage sequence S of the single battery voltage in the key voltage interval is extracted (k) = {V (k) |V (k) ∈ [V s ,V m ]} is extracted, then the sequence S (k) is interpolated and uniformly resampled to a sequence of length L = 23 and normalized to obtain:

[0039]

[0040] wherein, is the health feature H1 corresponding to the kth cycle.

[0041] S3-1-2, least squares linear fitting is performed on the H1 sequence to calculate the fitting slope a (k) as the voltage rising rate feature H2:

[0042]

[0043] S3-1-3, H1 and H2 are spliced to form the final feature vector of each cycle, reflecting the local platform region morphology and overall voltage rising trend information;

[0044] S3-2, for the final feature vector obtained in S3-1-3, local and multi-scale feature fusion is realized through a wavelet layered patch mixer module to obtain optimized multi-scale fusion features, which includes the following steps:

[0045] S3-2-1, multi-scale decomposition of feature sequence: multi-layer discrete wavelet transform is performed on the final feature vector of S3-1-3, and low-frequency approximation part and high-frequency detail part are extracted through low-pass filter and high-pass filter respectively; Daubechies 2 (db2) wavelet is selected as the wavelet basis function to ensure that the time-frequency characteristics are considered, the low frequency captures the global trend, and the high frequency captures the local fluctuation;

[0046] S3-2-2, Multi-scale patching: A multi-scale patching strategy is adopted for the low-frequency approximation part and the high-frequency detail part obtained in S3-2-1, and the sequence is divided into patches of different lengths to capture short-term fluctuations and long-term trend information through different scales;

[0047] S3-2-3, Patch internal mixing: Each patch internal channel is modeled, linear layers and GELU activation functions are combined with layer normalization to achieve patch internal mixing:

[0048] Z k =P k +Intra_Mixer(P k )

[0049] Intra_Mixer(·)=Linear st2 (GELU(Linear st1 (LN(·))))

[0050] wherein, P k represents the input feature tensor of the kth patch; Z k represents the output feature tensor after patch internal mixing;

[0051] LN(·) represents layer normalization; Linear st2 and Linear st1 are linear transformation layers; GELU represents GaussianError Linear Unit activation function.

[0052] S3-2-4, Patch mixing: Different patches are modeled to interact, and dimension conversion, linear layers and activation functions are used to process to achieve patch mixing:

[0053] Z′ k+1 =Z k+1 +Inter_Mixer(Z k )

[0054] Inter_Mixer(·)=Permute -1 (Linear lt2 (GELU(Linear lt1 (LN(Permute(·))))))

[0055] wherein, Z′ k+1 represents the output feature after patch mixing; Permute represents dimension conversion operation, which is used to set patch dimension as the main modeling axis; Linear lt2 and Linear lt1linear layer;

[0056] S3-2-5, output Z' of different size patches k Flatten and linear projection, and then weighted fusion through learnable weights to generate the fusion sequence Z agg . By traversing all feature dimensions in Z agg , randomly extract any two different dimensional features to form a dimensional combination, calculate the "product" and "absolute difference value" of the two at each time step, and then weighted sum the two results to form the cross feature of the dimensional pair. Finally, concatenate all the cross features of the dimensional pairs with Z agg to obtain an enhanced sequence containing feature interaction information, i.e., the optimized multi-scale fusion feature, for subsequent model input.

[0057] S4, input the optimized multi-scale fusion feature obtained in S3 into a lightweight ResCNN network, capture deep temporal features through a residual convolution module, map to SOH output through global pooling and fully connected layer, and realize rapid and accurate prediction of battery state of health through gradient utilization by combining with the denoised SOH label sequence obtained in S2, and improve model convergence speed and prediction stability through batch training and loss function optimization strategy;

[0058] S4-1, input convolution module (CBR): composed of 1D convolution layer, normalization layer and activation function, for the optimized multi-scale fusion feature x, one-dimensional convolution is used to extract local time sequence feature Y0, convolution kernel size K = 7, step s = 2:

[0059] Y0 = ReLU(BN(Conv(x)))

[0060] Where Conv(·) is the convolution layer, and ReLU(·) is the batch normalization layer. Then, the sequence length is further compressed through the max pooling layer to enhance the translation invariance:

[0061]

[0062] Where L0 represents the sequence length of the initial convolution module output, L' represents the sequence length after the max pooling layer, k is the pooling kernel size, and s is the step.

[0063] S4-2, for the local time sequence feature Y0, a residual convolution module composed of two convolution, normalization and activation functions is used, and the gradient vanishing problem is alleviated through residual connection, and the output of the residual convolution module is Y Res .

[0064] S4-3, global average pooling and fully connected layer, the sequence features are mapped to a single vector through global average pooling, and a SOH prediction value is output using a fully connected layer

[0065] Through the above structure, the ResCNN network can effectively extract local and global information of the battery feature sequence, and realize high-precision SOH prediction.

[0066] The optimized multi-scale fusion features obtained by S3 are used as model input, and the denoised SOH label sequence of S2 is used as model output; the model is trained in batch mode, and mean square error (MSE) is used as the loss function; the model performance is verified by test set data, and the average absolute error (MAE), root mean square error (RMSE) and mean absolute percentage error (MAPE) of the prediction result and the true SOH are calculated.

[0067] The test set data is input to verify the network performance, and the related model parameters and test results are shown in Table 1:

[0068] Table 1

[0069]

[0070]

[0071] Among them, when the input feature dimension is 24 and the number of residual layers of ResCNN network is 3, the performance indicators MAE, RMSE and MAPE of the orbit utilization battery SOH prediction model on the test set are 1.25%, 1.90% and 1.77% respectively.

[0072] When the input feature dimension is fixed at 24, the number of residual layers increases from 3 to 5, and the MAE, RMSE and MAPE are significantly reduced (by 5.6%, 7.9% and 8.5% respectively), which shows that appropriately increasing the network depth can enhance the model's ability to capture deep features; but when the number of residual layers is further increased to 7, the error indicators increase, which shows that too deep network will lead to overfitting (fitting too well to training data and poor generalization ability), which verifies the rationality of the design of "lightweight network".

[0073]

[0074] Among them, y i represents the true SOH value, represents the estimated SOH value, and N represents the total number of samples, i.e. the battery charge and discharge cycle period. MAE and RMSE measure the absolute size and variance level of the prediction error respectively, and MAPE reflects the size of the relative error of the prediction result.

Claims

1. A method for estimating the state of harm (SOH) of a tiered-use battery based on an improved post-convolutional network, characterized in that, Includes the following steps: S1. Collect the constant current charging stage operation data of the secondary utilization battery, and preprocess it to obtain a standardized multi-source operation dataset. S2. Based on standardized multi-source operational data, an energy accumulation model is constructed to generate a novel SOH estimation label based on energy accumulation. After processing, a denoised SOH label sequence is obtained. S3. Based on standardized multi-source running data and denoised SOH label sequences, energy slope is extracted and feature fusion is performed through wavelet layered patch mixer to obtain optimized multi-scale fusion features. S4. The optimized multi-scale fusion features are input into the lightweight ResCNN network and mapped to the SOH output to predict the health status of the battery for tiered use, while performing reverse training and testing.

2. The method for estimating the state of harm (SOH) of a tiered-use battery based on an improved convolutional network according to claim 1, characterized in that, Step S2 is implemented as follows: S2-1. Calculate the instantaneous charging power based on the charging curve data in the standardized multi-source operation data: P chg (t) represents the instantaneous charging power, I(t) represents the battery pack current, and V pack (t) represents the battery pack voltage, E chg (t) represents the charging energy. The rate of change of charging energy. When the value is greater than or equal to 0, charge the battery; S2-2. The voltage of the single-unit maximum voltage within the critical voltage range is considered a quasi-plateau region. An indicator function is introduced: Among them, U max (t) represents the maximum voltage of a single cell, when U max (t) in V s To V m When χV(t) is between, V s V is the lower limit of the quasi-platform area. m This is the upper bound of the quasi-platform region; otherwise, χV(t) is 0. S2-3. Integrate the instantaneous charging power within the voltage range, i.e., the quasi-plateau region, to obtain the accumulated energy E of the k-th charge-discharge cycle. k =∫χV(t)P chg (t)dt; S2-4, Standardized Cumulative Energy E k A novel SOH estimation label was obtained. Among them, E ref Indicates the battery's nominal capacity; S2-5. Estimating the tag sequence {SOH1,SOH2,...,SOH} for the novel SOH. N Gaussian smoothing is performed to obtain the denoised SOH tag sequence: w j The weights are given by the Gaussian kernel: Where σ is the Gaussian coefficient, j is the discrete position index within the Gaussian window, and w j Perform weight normalization to obtain 3. The SOH estimation method for tiered-use batteries based on the improved convolutional network according to claim 2, characterized in that, Step S3 is implemented as follows: S3-1. For each charge-discharge cycle, extract two types of health features and construct the final feature vector; S3-2. For the final feature vector, the local and multi-scale features are fused through the wavelet layered patch mixer module to obtain the optimized multi-scale fused features.

4. The SOH estimation method for tiered-use batteries based on the improved convolutional network according to claim 3, characterized in that, The specific implementation of step S3-1 is as follows: S3-1-1. During the charging phase, extract the voltage sequence S of individual cell voltages within the key voltage range. (k) ={V (k) |V (k) ∈[V s V m Then, for sequence S (k) Interpolation is performed, and the data is resampled into sequences of uniform length. Then normalize to get As the health feature H1 corresponding to the kth cycle; S3-1-2. Perform least-squares linear fitting on the H1 sequence and calculate the fitting slope α. (k) H2 is a characteristic of the voltage rise rate: S3-1-3. Concatenate columns H1 and H2 to form the final feature vector for each cycle.

5. The SOH estimation method for tiered-use batteries based on the improved convolutional network according to claim 4, characterized in that, The specific implementation of step S3-2 is as follows: S3-2. For the final feature vector, the local and multi-scale features are fused using the wavelet hierarchical patch mixer module, specifically including the following steps: S3-2-1. Perform multi-level discrete wavelet transform on the final feature vector, and extract the low-frequency approximation part and the high-frequency detail part through low-pass filter and high-pass filter respectively. S3-2-2, A multi-scale block strategy is adopted for the obtained low-frequency approximation part and high-frequency detail part to divide the sequence into blocks of different lengths; S3-2-3. Model the channels within each patch, using a combination of linear layers and the GELU activation function with layer normalization to achieve intra-patch blending: Z k =P k +Intra_Mixer(P k ) Intra_Mixer(·)=Linear st2 (GELU(Linear st1 (LN(·)))) Among them, P k Z is the input feature tensor of the k-th patch; k The output feature tensor after mixing within the patch; LN(·) represents layer normalization; Linear st2 and Linear st1 For linear transformation layers; GELU represents the Gaussian Error Linear Unit activation function; S3-2-4, Patch Blending: This involves interactive modeling between different patches, using dimensionality transformation, linear layers, and activation functions to achieve patch blending. WITH' k+1 =Z k+1 +Inter_Mixer(Z k ) Inter_Mixer(·)=Permute -1 (Linear lt2 (GELU(Linear lt1 (LN(Permute(·)))))) Among them, Z′ k+1 Represents the output features after mixing patches; Permute represents the dimensionality transformation operation, setting the patch dimension as the primary modeling axis; Linear lt2 and Linear lt1 Linear layer; S3-2-5, Convert the output Z′ of different patches k The sequence is flattened and linearly projected, then weighted and fused using learnable weights to generate a fused sequence Z. agg By traversing Z agg For all feature dimensions, randomly select any two different feature dimensions to form a feature dimension pair. For each feature dimension pair, calculate the product and absolute difference of the two features at each time step, and then sum these two results with weights to form the cross feature of that dimension pair. Finally, sum the cross features of all dimension pairs with Z. agg The sequences are concatenated to obtain an enhanced sequence containing feature interaction information, which is the optimized multi-scale fusion feature.

6. The SOH estimation method for tiered-use batteries based on the improved convolutional network according to claim 5, characterized in that, The specific implementation process of step S4 is as follows: S4-1, Input Convolutional Module, consists of a 1D convolutional layer, a normalization layer, and an activation function. For the optimized multi-scale fused feature x, it uses one-dimensional convolution to extract local time series features Y0, and then uses a max pooling layer to compress the sequence length and enhance translation invariance. Where L0 represents the sequence length output by the convolutional module, L′ represents the sequence length after the max pooling layer, k is the pooling kernel size, and s is the stride; S4-2. For the local time series feature Y0, a residual convolution module consisting of two convolutional layers, normalization, and activation functions is used, and the output Y is obtained through residual connections. Res ; S4-3, Output Y Res Sequence features are mapped to a single vector using global average pooling, and the SOH prediction value is output using a fully connected layer.

Images