Battery state of health estimation model training method

Abstract

A battery state of health estimation model training method includes the following steps: obtaining multiple battery multi-sensor signal data, and processing the multi-sensor signal data to form a source domain data set; decomposing the multi-sensor signal data in the source domain data set by EMD to obtain IMF components and residual components; performing trend test on the residual components, and obtaining a fusion trend signal according to the trend test result; stacking the IMF components to obtain final IMF signals; constructing a dense recurrent convolutional network, and inputting the fusion trend signal and the final IMF signal into the dense recurrent convolutional network for feature extraction to obtain time sequence features and local oscillation features respectively. The battery state of health estimation model training method realizes battery SOH estimation under multiple working conditions and small sample conditions, the model trained has wide applicability and can be popularized to different target domains, is suitable for multiple different battery working conditions, has strong generalization ability, and the prediction result is high in accuracy.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of battery management, in particular to a battery state of health estimation model training method. BACKGROUND

[0002] As a key component of smart phones, new energy vehicles and other devices, the normal operation of the battery plays a crucial role in the entire system. Once the battery is damaged and fails unexpectedly, it will have a serious impact on the entire device, and in severe cases, it will endanger the personal safety of the user. Therefore, it is necessary to accurately estimate the state of health (SOH) of the battery. Once the battery performance degradation or impending failure is found, appropriate countermeasures can be taken in time to avoid unnecessary losses.

[0003] During the cycle charging and discharging process of the battery, its SOH will gradually deteriorate with the increase of the number of charging and discharging times. The above-mentioned failure can be understood as the SOH decreasing to a specified threshold. With the increasing application scenarios of the battery, the working conditions of the battery are also generally variable, and the working condition data of the battery under some working conditions are difficult to collect, so the data will be scarce to some extent. In this case, it is necessary to better extract the features of the working condition data. For the data feature extraction in the field of batteries, many model-driven methods have been proposed by researchers, such as equivalent circuit models and electrochemical models. However, these model-driven methods are only applicable to some types of batteries, have poor generalization, and cannot be applied when the working conditions change. Therefore, under the condition of small sample, the accuracy and generalization performance of the model trained by the obtained features are poor. SUMMARY

[0004] Therefore, the purpose of the present application is to provide a battery state of health estimation model training method which is suitable for various battery working conditions, has strong generalization ability, and has high accuracy of the predicted results.

[0005] The present application provides a battery state of health estimation model training method, comprising the following steps:

[0006] Obtain the multi-sensor signal data of a plurality of batteries, and process the multi-sensor signal data to form a source domain data set;

[0007] Decompose the multi-sensor signal data in the source domain data set by EMD to obtain IMF components and residual components;

[0008] Perform trend test on the residual components, and obtain a fusion trend signal according to the trend test result; stack the IMF components to obtain final IMF signals;

[0009] A dense recurrent convolutional network is constructed, and the fused trend signal and the final IMF signal are input into the dense recurrent convolutional network for feature extraction, so as to obtain the temporal features and local oscillation features respectively.

[0010] Local oscillation features and time-series features are screened separately to obtain weighted spatial features and weighted temporal features. The weighted spatial features and weighted temporal features are then flattened and fused to obtain the first final degenerate feature.

[0011] Based on the different types of multi-sensor signal data in the source domain dataset, select the corresponding initial SOH estimator, and input the first final degradation feature into the corresponding initial SOH estimator for training to obtain the final SOH estimator.

[0012] In one embodiment, the step of performing a trend test on the residual components and obtaining a fusion trend signal based on the trend test results further includes:

[0013] Perform a trend test on the residual components and calculate the weights of the residual components based on the test results;

[0014] The fusion trend signal is obtained by calculating the residual components and their weights.

[0015] In one embodiment, the step of channel stacking of the IMF components to obtain the final IMF signal further includes:

[0016] The IMF components of sensor signal data of the same category are fused to obtain fused IMF components;

[0017] The final IMF signal is obtained by stacking the channels of the fused IMF components of different types of sensor signal data.

[0018] In one embodiment, the construction of a dense recurrent convolutional network, which inputs the fused trend signal and the final residual component into the dense recurrent convolutional network for feature extraction to obtain temporal features and local oscillation features respectively, further includes:

[0019] Construct a dense recurrent convolutional network, wherein the dense recurrent convolutional network includes multi-layer BiGRU and multi-layer CNN;

[0020] A multi-layer BiGRU is used to extract features from the fused trend signal to obtain time-series features;

[0021] Multi-layer CNNs are used to extract features from the final IMF signal to obtain local oscillation features.

[0022] In one embodiment, the step of inputting the first final degradation feature into the corresponding initial SOH estimator for training to obtain the final SOH estimator further includes:

[0023] The source domain dataset includes a meta-training dataset and a meta-test set; correspondingly, the final degradation features include the final degradation features of the training set and the final degradation features of the test set.

[0024] The first final degenerate feature of the meta-training set is input into the initial SOH estimator for training, so as to output the first prediction estimate loss value;

[0025] The first loss function is calculated based on the first predicted estimated loss value and the corresponding first true loss value. The SOH estimator is then updated using the first loss function to obtain the intermediate SOH estimator.

[0026] In one embodiment, the step of inputting the final degenerate features of the meta-test set into the intermediate SOH estimator for testing, and then updating the intermediate SOH estimator again to obtain the final SOH estimator, further includes:

[0027] The final degradation features of the meta-training set are embedded into the final degradation features of the meta-test set to obtain the final degradation features of the fused test set.

[0028] The final degradation features of the fused test set are input into the intermediate SOH estimator to obtain the second prediction estimate loss value;

[0029] The second loss function is calculated based on the second predicted estimated loss value and the corresponding second true loss value. The intermediate SOH estimator is then updated again using the second loss function and the first loss function to obtain the final SOH estimator.

[0030] In one embodiment, the model training method further includes:

[0031] Acquire multi-target domain data and mix the multi-target domain data to obtain the total training dataset for the target domain;

[0032] Obtain the second final degradation feature of the total training dataset in the target domain, and input the second final degradation feature into the final SOH estimator to obtain the third prediction estimate;

[0033] The fine-tuning loss function is calculated based on the third predicted estimate and the corresponding third true loss value, and the final SOH estimator is fine-tuned using the fine-tuning loss function.

[0034] In one embodiment, the formula for calculating the fine-tuned loss function based on the third predicted estimate and the third true loss value is as follows:

[0035]

[0036] Among them, L FT To fine-tune the loss function, y t and Let L represent the third true loss value and the third predicted estimate, respectively, and L be the sample size.

[0037] In one embodiment, the formula for mixing the multi-target domain data is:

[0038] Da t =MIX{DT1,DT2,...,DT n}

[0039] Among them, Da t Let MIX{} be the total training dataset for the target domain, MIX{} denotes the data mixing operation, and DTn represent the nth target domain data.

[0040] The battery health state estimation model training method provided by this invention realizes battery SOH estimation under various operating conditions and small sample conditions. The trained model has wide applicability, can be extended to different target domains, is applicable to various battery operating conditions, has strong generalization ability, and has high accuracy in prediction results. Attached Figure Description

[0041] To more clearly illustrate the technical solutions of the embodiments of the present invention, the accompanying drawings used in the embodiments will be briefly introduced below. It should be understood that the following drawings only show some embodiments of the present invention and should not be regarded as a limitation on the scope. For those skilled in the art, other related drawings can be obtained based on these drawings without creative effort.

[0042] Figure 1 The flowchart shows the battery health state estimation model training method provided by the present invention.

[0043] Figure 2 This is a schematic diagram of the EMD decomposition of the source domain dataset provided by the present invention.

[0044] Figure 3 A schematic diagram of the constructed DCRCN provided by the present invention.

[0045] Figure 4 This is a training strategy diagram for the source domain dataset provided by the present invention.

[0046] Figure 5 This is a schematic diagram of the structure of the meta-fusion module provided by the present invention.

[0047] Figure 6 The training data table provided by this invention.

[0048] Figure 7 A line graph of the first loss function provided by the present invention.

[0049] Figure 8 A line graph of the second loss function provided by the present invention.

[0050] Figure 9 This is a graph showing the battery SOH estimation results for target domain 1 of the present invention.

[0051] Figure 10 This is a graph showing the battery SOH estimation results for target domain 2 of the present invention.

[0052] Figure 11 This is a graph showing the battery SOH estimation results for target domain 3 of the present invention.

[0053] Figure 12 The ablation experiment comparison table provided by this invention. Detailed Implementation

[0054] The specific embodiments of the present invention will now be described in detail with reference to the accompanying drawings. Obviously, the described embodiments are merely some, not all, of the embodiments of the present invention. Based on the description of the present invention, all other embodiments obtained by those skilled in the art without inventive effort are within the scope of protection of the present invention.

[0055] In the description of this invention, unless otherwise explicitly specified and limited, the terms "set," "install," "connect," etc., should be interpreted broadly. For example, they can refer to a fixed connection, a detachable connection, or an integral connection; they can refer to a mechanical connection or an electrical connection; they can refer to a direct connection or an indirect connection through an intermediate medium. Those skilled in the art can understand the specific meaning of the above terms according to the specific circumstances.

[0056] The terms “upper,” “lower,” “left,” “right,” “front,” “back,” “top,” “bottom,” “inner,” and “outer,” etc., indicate the orientation or positional relationship based on the orientation or positional relationship shown in the accompanying drawings, or the orientation or positional relationship commonly used when the product of the invention is in use. They are only for the convenience of description and simplification, and do not indicate or imply that the device or element referred to must have a specific orientation, or be constructed and operated in a specific orientation. Therefore, they should not be construed as limitations on the present invention.

[0057] The terms “first,” “second,” “third,” etc., are used merely to distinguish elements with similar properties, not to indicate or imply relative importance or a specific order.

[0058] The terms “include,” “comprising,” or any other variation thereof are intended to cover non-exclusive inclusion, which includes not only the elements listed but also other elements not expressly listed.

[0059] Please see Figure 1 The battery health state estimation model training method provided by this invention includes the following steps:

[0060] S1: Acquire multi-sensor signals from several batteries and process the multi-sensor signal data to form a source domain dataset.

[0061] It is understood that the acquired multi-sensor signal data from the battery may include two datasets. Dataset 1 contains four operating conditions: Condition 1 (discharge current 2A, charging current 1.5A, ambient temperature 24℃), Condition 2 (discharge current 4A, charging current 1.5A, ambient temperature 24℃), Condition 3 (discharge current 4A, charging current 1.5A, ambient temperature 43℃), and Condition 4 (discharge current 1A, charging current 1.5A, ambient temperature 4℃). Dataset 2 is laboratory data, containing two operating conditions: Condition 1 (discharge current 0.8A, charging current 0.4A, ambient temperature 24℃) and Condition 2 (discharge current 0.6A, charging current 0.4A, ambient temperature 24℃). Both datasets use 18650 lithium batteries. Please refer to [link / reference]. Figure 6 The working conditions in the source domain dataset can be referenced. Figure 6 Choose from the options provided.

[0062] Please see Figure 2 S2, the multi-sensor signal data in the source domain dataset is decomposed by EMD (Empirical Mode Decomposition) to obtain IMF components and residual components.

[0063] It can be known that the decomposition formula is:

[0064]

[0065] Among them, Res i and Let x represent the i-th type of sensor data respectively. i After EMD decomposition, a residual component and the j-th IMF component are obtained. Ψ() represents the EMD decomposition operation. The above decomposition method is a known algorithm and will not be explained in detail here.

[0066] S3 performs a trend test on the residual components and obtains the fused trend signal based on the trend test results; the IMF components are stacked to obtain the final IMF signal.

[0067] It is understandable that step S3 above may further include:

[0068] S301, perform a trend test on the residual components and calculate the weights of the residual components based on the test results.

[0069] It is known that the Mann-Kendall trend test algorithm can be used to test the trend of the residual components. The formula for calculating the weight of the residual components is as follows:

[0070] ω i =MK(Res i ), i = 1, 2, ..., n

[0071] Where, ω i The residual component weights for calculating the i-th type of sensor signal data are MK(), which is the Mann-Kendall trend test operation, and Res is the weights for the residual components. i These are the residual components.

[0072] S302, the fusion trend signal is obtained by calculating the residual components and their weights.

[0073] It is known that the formula for calculating the fused trend signal is:

[0074]

[0075] Where, x Trend To integrate trend signals, Res i Represents the i-th type of sensor signal data x i The residual components obtained after EMD decomposition, ω i These are the weights of the residual components.

[0076] The final residual components can be obtained by channel stacking of IMF components, which may further include:

[0077] S303, fuse the IMF components of sensor signal data of the same category to obtain fused IMF components.

[0078] Understandably, the fusion method can refer to the channel stacking of the fusion IMF components of different categories of sensor signal data described below.

[0079] The final IMF signal is obtained by stacking the channels of the fused IMF components of different types of sensor signal data.

[0080] Understandably, the final formula for obtaining the IMF signal can be expressed as:

[0081]

[0082] Where, x IMFThis represents the final IMF signal obtained, and Ξ[] indicates stacking by channel. Represents the i-th type of sensor data x i The j-th IMF component obtained after EMD decomposition.

[0083] Please see Figure 3 S4. Construct a dense recurrent convolutional network and input the fused trend signal and the final residual component into the dense recurrent convolutional network for feature extraction to obtain temporal features and local oscillation features respectively.

[0084] As can be seen, please refer to Figure 2 The above steps may further include:

[0085] S401, Construct a Dense Recurrent Convolutional Network (DCRCN), which includes multi-layer BiGRU (Bidirectional Gated Recurrent Unit) and multi-layer CNN (Convolutional Neural Network);

[0086] S402, a multi-layer BiGRU extracts features from the fused trend signal to obtain time-series features;

[0087] It is understandable that the formula for extracting time-series features can be:

[0088] F temp =Θ <BiGRU k (x Trend )>

[0089] Among them, F temp Representing time series characteristics, Θ <BiGRU k (x Trend )> represents the cascaded feature extraction of the fusion trend signal by K BiGRU layers.

[0090] S403, a multi-layer CNN extracts features from the final IMF signal to obtain local oscillation features.

[0091] It is understandable that time-series features can characterize the overall trend of battery health status, while local oscillation features can characterize local regeneration and fluctuation phenomena in battery health status. The acquisition of time-series features and local oscillation features can be parallel. The extraction formula for the aforementioned local oscillation features is as follows:

[0092]

[0093] in, Indicating local oscillation characteristics, Θ <CNN k (x IMF This can be represented as K CNN layers processing the final IMF signal x. IMF Cascaded feature extraction.

[0094] S5, the local oscillation features and time series features are screened respectively to obtain weighted spatial features and weighted time series features. The weighted spatial features and weighted time series features are then tiled and fused to obtain the first final degradation feature.

[0095] It is understandable that the selection of local oscillation features and time series features can be expressed as:

[0096] ω temp =δ(Avg) temp (F temp ))

[0097]

[0098]

[0099]

[0100] Where, ω temp Represents time series weights. Let δ() represent the spatial weights of the local oscillation features of the l-th CNN output, and let Avg() represent the sigmoid nonlinear activation function. temp This indicates that the average value is calculated based on the time series dimension, Max. spat This indicates that the maximum value is calculated based on the channel dimension. Indicates element-wise multiplication. and F represents the obtained weighted temporal features and the l-th weighted spatial features, respectively. temp Representing temporal characteristics, This indicates local oscillation characteristics.

[0101] The first final degradation feature can be obtained using the following formula:

[0102]

[0103] Among them, F out This represents the first final degenerate feature extracted by DCRCN, and Γ() represents the feature tiling operation. and These represent the weighted temporal features and the Kth weighted spatial features, respectively.

[0104] Please see Figure 4 S6. Select the corresponding initial SOH estimator according to the different multiple sensor signals in the source domain dataset, and input the first final degradation feature into the corresponding initial SOH estimator for training to obtain the final SOH estimator.

[0105] Understandably, the initial SOH estimator can be downloaded from the internet. Each initial SOH estimator corresponds to a type of sensor signal and may include a CNN layer.

[0106] Step S6 above may further include:

[0107] S601, the source domain dataset includes the meta-training dataset and the meta-test set, and correspondingly, the final degradation features include the final degradation features of the training set and the final degradation features of the test set.

[0108] It is understandable that one source domain data can be selected as the meta-test set, and the remaining source domain data can be used as the meta-training data. The data in the meta-test set and the meta-training data can be the first final degradation feature obtained through the above steps.

[0109] S602, the first final degenerate feature of the meta-training set is input into the initial SOH estimator for training, so as to output the first prediction estimate loss value.

[0110] Understandably, the first predicted estimated loss value can be expressed by the following formula:

[0111]

[0112] in, This represents the first prediction estimate loss value for the i-th source domain in the meta-training set. This represents the i-th SOH estimator. This represents the fully connected layer (CNN) of the i-th SOH estimator. It can be represented as the first final degradation feature extracted from the i-th source domain.

[0113] S603, calculate the first loss function based on the first predicted estimated loss value and the corresponding first true loss value, and update the SOH estimator through the first loss function to obtain the intermediate SOH estimator.

[0114] Understandably, the formula for calculating the first loss function can be expressed as:

[0115]

[0116] in, Let L be the first loss function of the i-th source domain in the meta-training set, and L be the number of sampling points of the sample. and Let represent the first true loss value and the first predicted estimated loss value of the i-th source domain in the meta-training set, respectively. The parameters of the fully connected layer of the SOH estimator can be adjusted using the first loss function. The above training can guide the SOH estimator to learn domain-invariant features through domain adaptation. The domain distribution difference is obtained through the maximum mean discrepancy (MMD). The calculation process of the domain distribution difference is expressed as follows:

[0117]

[0118] Among them, L DA Due to differences in domain distribution, represent and The MMD values ​​between This represents the first final degenerate feature extracted from the i-th source domain in the meta-training set. Let N represent the first final degradation feature extracted from the l-th source domain. s -1 represents the number of source domain data in the meta-training set.

[0119] S604, input the final degenerate features of the meta-test set into the intermediate SOH estimator for testing, and update the intermediate SOH estimator again to obtain the final SOH estimator.

[0120] Understandably, after the above training process is completed, the intermediate SOH estimator formed by training can be updated again using the meta-test set to improve the final SOH estimator with higher prediction accuracy.

[0121] It can be understood that step S604 above may further include:

[0122] S604a embeds the final degenerate features of the meta-training set into the final degenerate features of the meta-test set to obtain the final degenerate features of the fused test set.

[0123] Please see Figure 5 It is understandable that embedding the final degraded features of the meta-training set into the final degraded features of the meta-test set can be achieved using the meta-fusion module. Fusion can ensure the diversified distribution of the meta-test set features. The above formula for obtaining the final degraded features of the fused test set can be expressed as:

[0124]

[0125]

[0126] in, This represents a random array generated according to the feature distribution of the i-th source domain in the meta-training set. and These are the i-th first final degenerate features in the meta-training set. The mean and standard deviation of a Gaussian distribution. This represents generating a random array with a Gaussian distribution. The final degradation feature of the fusion test set of the i-th source domain output by the meta-fusion module.

[0127] S604b inputs the final degraded features of the fused test set into the intermediate SOH estimator to obtain the second predicted estimated loss value.

[0128] Understandably, the second predicted loss value can be expressed by the following formula:

[0129]

[0130] in, The second prediction estimate loss value for the meta-test set. This represents the i-th intermediate SOH estimator in the meta-training set. This represents the final degradation feature of the i-th fusion test set output by the meta-fusion module.

[0131] S604c, the second loss function is calculated based on the second predicted estimated loss value and the corresponding second true loss value. The intermediate SOH estimator is updated again using the second loss function and the first loss function to obtain the final SOH estimator.

[0132] Understandably, the formula for calculating the second loss function can be:

[0133]

[0134] in, The second loss function is y, where L is the number of sampling points in the sample. Te and These represent the values ​​of the meta-test set, the second true loss value, and the second predicted estimated loss value, respectively.

[0135] In some embodiments, after the above training is completed, the loss during the learning and training phase can also be calculated, and the calculation process can be expressed as follows:

[0136]

[0137] Among them, L ML The total loss of the meta-learning process. Let L be the second loss function for the i-th source domain, and L be the first loss function for the i-th source domain in the meta-training set. DA To account for the regional distribution differences, the parameters of the final SOH estimator are optimized using this total loss.

[0138] After the training and testing described above are completed, model fine-tuning is performed using small sample data from the multi-objective domain. In some embodiments, the fine-tuning operation may include the following steps:

[0139] Please see Figures 9 to 11 S7, acquire multi-target domain data, mix the multi-target domain data to obtain the total training dataset for the target domain.

[0140] Understandably, multi-objective data can serve as a reference. Figure 6 The example operating condition data in the example can be mixed in the following ways:

[0141] Da t =MIX{DT1,DT2,...,DT n}

[0142] Among them, Da t This represents the total training dataset of the target domain after mixing, MIX{} represents the data mixing operation, and DTn represents the nth target domain data.

[0143] The second final degradation feature of the total training dataset in the target domain is obtained, and this second final degradation feature is input into the final SOH estimator to obtain the third prediction estimate. Figures 9 to 11 The red line represents the third forecast estimate.

[0144] Understandably, the second final degradation feature can be obtained by referring to S1-S5 above, and the third prediction estimate can be expressed as:

[0145] Ft = DCRCN(x) t )

[0146]

[0147] Among them, F t x is the second final degradation feature extracted by DCRCN. t For the total training dataset Da of the target domain t The sample data in This is the third predicted estimate. Let i represent the i-th final SOH estimator in the meta-training set.

[0148] The fine-tuning loss function is calculated based on the third predicted estimate and the corresponding third true loss value. The final SOH estimator is then fine-tuned using the fine-tuning loss function. Figures 9 to 11 The black line in the middle represents the third true loss value.

[0149] Understandably, fine-tuning can be represented as:

[0150]

[0151] Among them, L FT To fine-tune the loss function, y t and Let represent the third true loss value and the third predicted estimate of the total training dataset in the target domain, respectively. The parameters of the final SOH estimator are optimized using this fine-tuned loss function to obtain the fine-tuned final SOH estimator.

[0152] Please see Figure 7 and Figure 8 The iterative process of learning the SOH estimator is as follows: Figure 8 As shown, from 7 and Figure 8 It can be clearly seen that the first and second loss functions of the SOH estimator gradually decrease with the increase of the number of iterations. The final third prediction value of the SOH estimator on the test data in the target domain 1-3 is as follows: Figures 9-11 As shown, the final SOH estimator's SOH estimation results in each target domain are very close to the actual SOH values, demonstrating the excellent performance of the final SOH estimator in SOH estimation on small sample battery data in multiple target domains.

[0153] To demonstrate the effectiveness of the training strategy, this invention employs an ablation experiment. DCRCN and the learning strategy are sequentially removed from the original model; that is, when DCRCN is removed, the learning strategy is retained, and when the learning strategy is deleted, DCRCN is retained in the original model. This experiment uses two metrics—Mean Absolute Error (MAE) and Normalized Root Mean Square Error (NRMSE)—to evaluate the model's performance on battery SOH estimation. The comparison results are as follows: Figure 12 As shown, it is easy to see that when the DCRCN or learning strategy is removed from the SOH estimator, the model's performance on both MAE and NRMSE deteriorates compared to the original model. Therefore, this proves the effectiveness of the training method in battery SOH estimation.

[0154] As described above, the battery health state estimation model training method provided by this invention can estimate battery SOH under various operating conditions and small sample conditions. Furthermore, with the support of the method proposed in this invention, the model has wide applicability and can be extended to different target domains. It is suitable for various battery operating conditions, has strong generalization ability, and can intuitively reflect the SOH estimation results. At the same time, it can achieve battery health state estimation in multiple target domains with only small sample data.

[0155] The above description is merely a specific embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any variations or substitutions that can be easily conceived by those skilled in the art within the technical scope disclosed in the present invention should be included within the scope of protection of the present invention. Therefore, the scope of protection of the present invention should be determined by the appended claims.

Claims

1. A method for training a battery health state estimation model, characterized in that, Includes the following steps: Acquire multi-sensor signal data from several batteries and process the multi-sensor signal data to form a source domain dataset; EMD is used to decompose the multi-sensor signal data in the source domain dataset to obtain the IMF component and the residual component. A trend test is performed on the residual components, and a fusion trend signal is obtained based on the trend test results; To obtain the final IMF signal by channel stacking of the IMF components, the method further includes: Perform a trend test on the residual components and calculate the weights of the residual components based on the test results; The fusion trend signal is obtained by calculating the residual components and their weights: The IMF components of sensor signal data of the same category are fused to obtain fused IMF components; The final IMF signal is obtained by stacking the channels of the fused IMF components of different types of sensor signal data; A dense recurrent convolutional network is constructed, and the fused trend signal and the final IMF signal are input into the dense recurrent convolutional network for feature extraction, so as to obtain the temporal features and local oscillation features respectively. Local oscillation features and time-series features are screened separately to obtain weighted spatial features and weighted temporal features. The weighted spatial features and weighted temporal features are then flattened and fused to obtain the first final degenerate feature. Based on the different types of multi-sensor signal data in the source domain dataset, select the corresponding initial SOH estimator, and input the first final degradation feature into the corresponding initial SOH estimator for training to obtain the final SOH estimator.

2. The battery health state estimation model training method as described in claim 1, characterized in that, The construction of a dense recurrent convolutional network, in which the fused trend signal and the final IMF signal are input into the dense recurrent convolutional network for feature extraction to obtain temporal features and local oscillation features respectively, further includes: Construct a dense recurrent convolutional network, wherein the dense recurrent convolutional network includes multi-layer BiGRU and multi-layer CNN; A multi-layer BiGRU is used to extract features from the fused trend signal to obtain time-series features; Multi-layer CNNs are used to extract features from the final IMF signal to obtain local oscillation features.

3. The battery health state estimation model training method as described in claim 2, characterized in that, The step of inputting the first final degradation feature into the corresponding initial SOH estimator for training to obtain the final SOH estimator further includes: The source domain dataset includes a meta-training dataset and a meta-test set; correspondingly, the final degradation features include the final degradation features of the training set and the final degradation features of the test set. The first final degenerate feature of the meta-training set is input into the initial SOH estimator for training, so as to output the first prediction estimate loss value; The first loss function is calculated based on the first predicted estimated loss value and the corresponding first true loss value. The SOH estimator is then updated using the first loss function to obtain the intermediate SOH estimator. The final degenerate features of the meta-test set are input into the intermediate SOH estimator for testing, and the intermediate SOH estimator is updated again to obtain the final SOH estimator.

4. The battery health state estimation model training method as described in claim 3, characterized in that, The step of inputting the final degenerate features of the meta-test set into the intermediate SOH estimator for testing, and then updating the intermediate SOH estimator again to obtain the final SOH estimator, further includes: The final degradation features of the meta-training set are embedded into the final degradation features of the meta-test set to obtain the final degradation features of the fused test set. The final degradation features of the fused test set are input into the intermediate SOH estimator to obtain the second prediction estimate loss value; The second loss function is calculated based on the second predicted estimated loss value and the corresponding second true loss value. The intermediate SOH estimator is then updated again using the second loss function and the first loss function to obtain the final SOH estimator.

5. The battery health state estimation model training method as described in claim 1, characterized in that, The model training method also includes: Acquire multi-target domain data and mix the multi-target domain data to obtain the total training dataset for the target domain; Obtain the second final degradation feature of the total training dataset in the target domain, and input the second final degradation feature into the final SOH estimator to obtain the third prediction estimate; The fine-tuning loss function is calculated based on the third predicted estimate and the corresponding third true loss value, and the final SOH estimator is fine-tuned using the fine-tuning loss function.

6. The battery health state estimation model training method as described in claim 5, characterized in that, The formula for calculating the fine-tuned loss function based on the third predicted estimate and the third true loss value is as follows: in, To fine-tune the loss function, and Let represent the third true loss value and the third predicted estimate, respectively, and L be the sample size t.

7. The battery health state estimation model training method as described in claim 5, characterized in that, The formula for mixing the multi-target domain data is: Among them, Da t Let MIX{} represent the total training dataset for the target domain, and DT represent the data mixing operation. n This represents the nth target domain data.

Images