Method and system for estimating SOH (state of health) of battery based on actual charging data

Abstract

The invention relates to a method and system for estimating the SOH of a battery based on actual charging data. The method comprises the following steps: acquiring battery charging data; constructing a deep residual shrinkage network model, and training the deep residual shrinkage network model by taking the battery charging data as the training sample as input; and inputting the actual battery charging data of the to-be-evaluated battery into the trained deep residual shrinkage network model, estimating the health state of the to-be-evaluated battery, and outputting the health state of the to-be-evaluated battery. The method is very suitable for common random incomplete charging and discharging conditions in industrial scenes, and the flexibility and adaptability of the method are greatly improved. Secondly, in order to relieve the sensitivity of a pure data driving method to noise, a deep residual shrinkage network (DRSN) is adopted in the research to automatically explore capacity attenuation characteristics hidden in a local charging curve, an attention mechanism and a soft threshold are fused, and the evaluation precision is improved;

Description

Technical Field

[0001] The present invention relates to the field of battery technology, and in particular to a method and system for estimating battery SOH based on actual charging data. Background Art

[0002] As an important energy storage device, lithium-ion batteries play an important role in many fields such as electric vehicles, mobile devices, aerospace, and renewable energy storage. Accurate battery capacity estimation is crucial for condition monitoring, performance evaluation, and the formulation of control strategies. However, due to the complexity of the internal chemical and physical processes of the battery and the uncertainty brought by external operating procedures, accurate capacity estimation or state of health (SOH) estimation remains a challenging task in actual industrial scenarios.

[0003] At present, there are two main methods for estimating the SOH of lithium-ion batteries: model-driven and data-driven. Model-driven methods include electrochemical models and equivalent circuit models. Electrochemical models are based on the working principle of batteries and establish mathematical expressions to describe the intricate electrochemical reactions and physical processes occurring between the positive and negative electrodes, separators, and electrolytes of batteries. These models have high accuracy in estimating battery capacity. However, these models require a comprehensive understanding of the physicochemical properties of electrode materials and consider the internal and external working conditions of the battery. The computational solution process of these models consumes a lot of computing resources, making them challenging for online applications. In contrast, the equivalent circuit model has a simpler structure and requires fewer parameters. However, it provides a simplified approximation of the physical and electrochemical dynamics inside the battery and relies on many assumptions. Therefore, its accuracy is lacking compared to the electrochemical model. In addition, when using filtering algorithms for model parameter identification, it is also necessary to solve difficult problems such as parameter initialization. These limitations restrict the widespread application of model-driven methods.

[0004] With the continuous development of artificial intelligence algorithms such as machine learning, data-driven methods have received increasing attention. Data-driven methods have powerful nonlinear problem fitting capabilities and can simulate and emulate battery aging and degradation processes that rely on complex electrochemical reactions and their side reactions. In addition, these data-driven methods have also shown impressive effectiveness and accuracy. The key steps of data-driven methods include extracting health indicators (HIs) and selecting machine learning models. Widely used machine learning models include support vector machines (SVM), relevance vector machines (RVM), Gaussian process regression (GPR), and artificial neural networks. HIs extraction refers to the evaluation of battery capacity or health by monitoring and analyzing specific parameters and behaviors of the battery. At present, a large number of studies have been devoted to extracting HIs. For example, the slope of the voltage curve, the minimum and maximum values ​​of the charging curve and their corresponding times, the skewness index and kurtosis index of the signal, the regional voltage and charging duration, the discharge voltage entropy, the equidistant capacity interval voltage and voltage slope, etc.

[0005] However, these methods also face corresponding challenges. First, existing methods often rely on data from complete charging and discharging or a specific voltage range. Due to the randomness of user behavior, it is unrealistic to collect all this data, so existing methods often lack flexibility and adaptability to industrial applications. Second, the above methods are usually based on laboratory data with good data quality, while real data will face a more complex and changeable noise environment, so direct modeling of real data may be susceptible to noise and incomplete data. Summary of the invention

[0006] In view of the above problems, an object of the present invention is to provide a method for estimating the health state of a lithium-ion battery using actual variable charging data based on a deep residual shrinkage network.

[0007] A method for estimating battery SOH based on actual charging data, comprising:

[0008] Get battery charging data;

[0009] Construct a deep residual shrinkage network model, use the battery charging data as training samples as input, and train the deep residual shrinkage network model;

[0010] The actual battery charging data of the battery to be evaluated is input into the trained deep residual shrinkage network model to estimate the health status of the battery to be evaluated and output it.

[0011] Furthermore, the battery charging data is data of the constant current charging stage;

[0012] After obtaining the battery charging data, the battery charging data is also preprocessed. The preprocessing process includes:

[0013] Fitting the battery voltage and battery capacity in the battery charging data to form a constant current charging curve;

[0014] A constant current charging curve segment is intercepted according to a predetermined voltage window, and a voltage sampling point and a capacity sampling point are obtained by sampling the curve segment at a preset voltage interval, and the capacity increment is calculated from the capacity sampling point to form a voltage-capacity increment data sequence;

[0015] The voltage-capacity increment data sequence formed by preprocessing is used as the input of the deep residual shrinkage network model.

[0016] Furthermore, the deep residual shrinkage network model includes an input layer, a convolution module, a residual shrinkage module and a classification recognition module connected in sequence, and the residual shrinkage module is composed of at least one residual shrinkage unit;

[0017] The input layer receives battery charging data;

[0018] The convolution module performs convolution operation on the battery charging data to form original feature data;

[0019] The residual shrinkage module performs feature extraction and noise filtering on the original feature data to obtain classification feature data;

[0020] The classification recognition module evaluates the classification feature data to obtain the health status of the battery and outputs it.

[0021] Furthermore, each residual shrinkage unit includes a residual unit, an attention mechanism unit, and a soft thresholding unit;

[0022] The input of the first residual shrinkage unit of the residual shrinkage module is the output of the convolution module, the output of the last residual shrinkage unit is used as the classification feature data, the input of each of the remaining residual shrinkage units is the output of the previous residual shrinkage unit, and the output of each of the remaining residual shrinkage units is the input of the next residual shrinkage unit;

[0023] The residual unit performs feature extraction on the feature data input into the residual shrinkage unit to obtain first feature data;

[0024] The attention mechanism unit calculates an adaptive threshold according to the first feature data;

[0025] The soft thresholding unit performs soft thresholding processing on the first feature data according to the adaptive threshold to obtain second feature data;

[0026] The residual unit adds the feature data input to the residual shrinkage unit and the second feature data, and uses the addition result as the final output feature data.

[0027] Furthermore, the convolution operations in the convolution module and the residual shrinkage module both adopt dilated convolution.

[0028] A system for estimating battery SOH based on actual charging data, used to execute the aforementioned method for estimating battery SOH based on actual charging data, comprising:

[0029] A data acquisition module, used for acquiring battery charging data;

[0030] The model training module is connected to the data acquisition module and is used to construct a deep residual shrinkage network model. The battery charging data as a training sample is used as input to train and store the deep residual shrinkage network model.

[0031] The state evaluation module is connected to the data acquisition module and is used to input the actual battery charging data of the battery to be evaluated into the trained deep residual shrinkage network model, estimate the health state of the battery to be evaluated and output it.

[0032] Furthermore, the data acquisition module includes:

[0033] An acquisition unit, used for acquiring data of a constant current charging stage of a battery as battery charging data;

[0034] A curve fitting unit, connected to the acquisition unit, is used to fit the battery voltage and battery capacity in the battery charging data to form a constant current charging curve;

[0035] A data sampling unit, connected to the curve fitting unit, is used to intercept a constant current charging curve segment according to a predetermined voltage window, and to sample the curve segment at a preset voltage interval to obtain a voltage sampling point;

[0036] A sequence forming unit, connected to the data sampling unit, is used to calculate the capacity increment from the capacity sampling points to form a voltage-capacity increment data sequence;

[0037] The voltage-capacity increment data sequence formed by the sequence forming unit is used as the input of the deep residual shrinkage network model.

[0038] Furthermore, the deep residual shrinkage network model includes an input layer, a convolution module, a residual shrinkage module and a classification recognition module connected in sequence, and the residual shrinkage module is composed of at least one residual shrinkage unit;

[0039] The input layer is used to receive battery charging data;

[0040] The convolution module is used to perform convolution operation on the battery charging data to form original feature data;

[0041] The residual shrinkage module is used to extract features and filter noise from the original feature data to obtain classification feature data;

[0042] The classification recognition module is used to evaluate the classification feature data to obtain the health status of the battery and output it.

[0043] Furthermore, each residual shrinkage unit includes a residual unit, an attention mechanism unit, and a soft thresholding unit;

[0044] The input of the first residual shrinkage unit of the residual shrinkage module is the output of the convolution module, the output of the last residual shrinkage unit is used as the classification feature data, the input of each of the remaining residual shrinkage units is the output of the previous residual shrinkage unit, and the output of each of the remaining residual shrinkage units is the input of the next residual shrinkage unit;

[0045] The residual unit is used to extract features from the feature data input into the residual shrinkage unit to obtain first feature data;

[0046] The attention mechanism unit is used to calculate and obtain an adaptive threshold according to the first feature data;

[0047] The soft thresholding unit is used to perform soft thresholding processing on the first feature data according to the adaptive threshold to obtain second feature data;

[0048] The residual unit is used to add the feature data input to the residual shrinkage unit and the second feature data, and use the addition result as the final output feature data.

[0049] Furthermore, the convolution operations in the convolution module and the residual shrinkage module both adopt dilated convolution.

[0050] The beneficial technical effect of the present invention is that compared with the existing technology, the present invention can estimate the capacity of lithium-ion batteries using random fragment voltage-capacity data obtained during real-time charging. This method is very suitable for random incomplete charging and discharging conditions commonly seen in industrial scenarios, greatly improving the flexibility and adaptability of the method. Secondly, in order to alleviate the sensitivity of pure data-driven methods to noise, this study used a deep residual shrinkage network (DRSN) with good noise resistance to automatically explore the capacity decay features hidden in the local charging curve. It combines the attention mechanism and soft threshold, and is expected to achieve higher result accuracy and stronger noise robustness, which has great application significance. BRIEF DESCRIPTION OF THE DRAWINGS

[0051] Figure 1-2 A flowchart of a method for estimating battery SOH based on actual charging data according to the present invention;

[0052] Figure 3-6 The model architecture diagram of the deep residual shrinkage network of the present invention;

[0053] Figure 7-8 It is a module schematic diagram of the deep residual shrinkage network of the present invention;

[0054] Fig. 9 The module diagram of a system for estimating battery SOH based on actual charging data of the present invention is shown in FIG. DETAILED DESCRIPTION

[0055] The following will be combined with the drawings in the embodiments of the present invention to clearly and completely describe the technical solutions in the embodiments of the present invention. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without creative work are within the scope of protection of the present invention.

[0056] It should be noted that, in the absence of conflict, the embodiments of the present invention and the features in the embodiments may be combined with each other.

[0057] The present invention will be further described below with reference to the accompanying drawings and specific embodiments, but they are not intended to limit the present invention.

[0058] See also Figure 1 The present invention provides a method for estimating battery SOH based on actual charging data, comprising:

[0059] Get battery charging data;

[0060] Construct a deep residual shrinkage network model, use the battery charging data as training samples as input, and train the deep residual shrinkage network model;

[0061] The actual battery charging data of the battery to be evaluated is input into the trained deep residual shrinkage network model to estimate the health status of the battery to be evaluated and output it.

[0062] See also Figure 2 ,Furthermore, the battery charging data is the data of the constant current charging stage;

[0063] After obtaining the battery charging data, the battery charging data is also preprocessed. The preprocessing process includes:

[0064] Fitting the battery voltage and battery capacity in the battery charging data to form a constant current charging curve;

[0065] A constant current charging curve segment is intercepted according to a predetermined voltage window, and a voltage sampling point and a capacity sampling point are obtained by sampling the curve segment at a preset voltage interval, and the capacity increment is calculated from the capacity sampling point to form a voltage-capacity increment data sequence;

[0066] The voltage-capacity increment data sequence formed by preprocessing is used as the input of the deep residual shrinkage network model.

[0067] Existing methods often rely on data from complete charging and discharging or a specific voltage range. Due to the randomness of user behavior, it is unrealistic to collect all of this data. Therefore, existing methods often lack flexibility and adaptability to industrial applications. To address this problem, the present invention can use fragmented constant current charging data as model input.

[0068] The data of the constant current charging stage is selected. Different battery sensors have different data sampling methods, so the smoothing spline fitting method is used to interpolate the data to generate a constant current charging curve. Spline fitting interpolation is used to unify the data format differences of different battery data acquisition systems.

[0069] Then, voltage segmentation is performed, and the voltage V and capacity Q data of a predetermined voltage window with a fixed voltage window length L and an equidistant voltage interval (preset voltage interval) in the constant current charging curve are selected, and the voltage-capacity increment is used as the input of the model. That is, the input data is generated by sliding window iteration, and the expression (1) of the generated voltage-capacity increment ΔQ-V data sequence is as follows:

[0070] [ΔQ1,ΔQ2,…,ΔQ n-1 ,ΔQ n ]

[0071] [V s +ΔV,V s +2ΔV,…,V s +(n-1)ΔV,V s +nΔV]

[0072] Where V s is the starting point of the predetermined voltage window, ΔQ n From V s To V s The cumulative charging capacity (i.e., capacity increment) of +nΔV is expressed as (2) as follows:

[0073]

[0074] Wherein, i(t) is the charging current at time t, ΔV is the preset voltage interval, and the length n of the voltage-capacity increment data sequence is calculated by n=L / ΔV, where L is the window length of the preset voltage window.

[0075] The input data is generated iteratively using a voltage sliding window, so the framework can be applied to any initial charging state and incomplete charging scenarios, thereby achieving flexibility in input data. Random fragment charging data is used as the input of the model to train the model, and the structure in the constructed one-dimensional DRSN is used to process random fragment charging data to obtain the battery SOH evaluation result. Specifically, the training of the DRSN model is learned through the back-propagation algorithm. During the training process, the actual health state SOH of the battery is used as a label.

[0076] Furthermore, the process of preprocessing the battery charging data also includes: normalizing the sampled voltage-capacity increment data sequence, and using the normalized voltage-capacity increment data sequence as the input of the deep residual shrinkage network model. The normalization formula is as follows:

[0077]

[0078] Among them, u represents the mean of the sample and σ represents the variance of the sample.

[0079] The definition formula of battery health state SOH is as follows:

[0080] Among them, C i It represents the maximum capacity of the battery after it is fully charged through i charge and discharge cycles, and C0 represents the initial rated capacity of the battery (such as the maximum storage capacity of the battery when it leaves the factory).

[0081] See also Figure 3 The present invention first acquires constant current charging data, then preprocesses the constant current charging data, then constructs a deep residual shrinkage network model, initializes the model structure, and then initializes the model parameters. The model weights and biases are iteratively updated using training data to obtain an optimized model. After the model is optimized, the model is tested. Specifically, during training, the battery charging data (with the preprocessed voltage-capacity value-added data sequence) used as training samples are divided into training set, validation set, and test set according to 6:2:2; and k-fold cross validation is used to achieve battery SOH evaluation, and the back propagation algorithm is used to update the model parameters and the weights in the attention mechanism. The model training rounds are set to 20,000 times, and the model is saved only when the validation set loss function decreases to alleviate overfitting. Use the test data set for testing.

[0082] See also Figure 3-6 and Figure 8 , further, the deep residual shrinkage network model includes an input layer (21), a convolution module (22), a residual shrinkage module and a classification recognition module (24) connected in sequence, and the residual shrinkage module is composed of at least one residual shrinkage unit (23);

[0083] The input layer (21) receives battery charging data;

[0084] The convolution module (22) performs a convolution operation on the battery charging data to form original feature data;

[0085] The residual shrinkage module performs feature extraction and noise filtering on the original feature data to obtain classification feature data;

[0086] The classification recognition module (24) evaluates the classification feature data to obtain the health status of the battery and outputs it.

[0087] First, feature extraction is performed through a convolution module, and then feature extraction and noise filtering are performed through multiple residual shrinkage units (RSUs). Finally, the classification and recognition module evaluates the health status of the battery.

[0088] Furthermore, each residual shrinkage unit (23) includes a residual unit (231), an attention mechanism unit (232) and a soft thresholding unit (233);

[0089] The input of the first residual contraction unit (23) of the residual contraction module is the output of the convolution module (22), the output of the last residual contraction unit (23) is used as classification feature data, the input of each of the remaining residual contraction units (23) is the output of the previous residual contraction unit (23), and the output of each of the remaining residual contraction units (23) is the input of the next residual contraction unit (23);

[0090] The residual unit (231) performs feature extraction on the feature data input to the residual contraction unit (23) to obtain first feature data;

[0091] The attention mechanism unit (232) calculates and obtains an adaptive threshold according to the first feature data;

[0092] The soft thresholding unit (233) performs soft thresholding processing on the first feature data according to the adaptive threshold to obtain second feature data;

[0093] The residual unit (231) adds the feature data input to the residual contraction unit (23) and the second feature data, and uses the addition result as the final output feature data.

[0094] The basic convolutional neural network operation in RSU is as follows (3):

[0095]

[0096] where x l is the input vector of the previous layer of network, b l is the bias parameter, is the weight parameter of the convolution kernel, n is the number of convolution filters, Relu is the nonlinear activation function, and * represents the convolution operation.

[0097] The residual shrinkage unit (RSU) mainly integrates the deep attention mechanism and integrated soft threshold on the basis of the residual unit.

[0098] The residual unit may include a plurality of sequentially connected convolution groups, for example, two sequentially connected convolution groups, each of which includes a convolution layer and a ReLU function layer. After feature extraction by the several convolution groups, the first feature data (ie, the first feature vector) is output.

[0099] In tasks such as classification and regression, the contribution of different feature channels to the overall performance may vary from sample to sample. To address this problem, the attention mechanism unit uses a compact subnetwork to learn a set of weights. The attention mechanism unit first calculates the absolute value of the first feature data, and then squeezes the intermediate feature data u from the upper layer after taking the absolute value through a global average pooling layer (GAP) to obtain the global information z. The size of the intermediate feature data is converted from C×H×W to C×1×1. Formula (4) is as follows:

[0100]

[0101] The excitation operation is performed sequentially through the fully connected layer, ReLU function layer, fully connected layer and Sigmoid function layer. Formula (5) is as follows:

[0102] s=F ex (z,W)=σ(W2δ(W1z))

[0103] The global information z is multiplied by the weight W1 through a fully connected layer (FC), and then multiplied by W2 through a ReLU activation function δ(·) and another FC layer. Finally, the weight result is obtained by a Sigmoid function σ(·).

[0104] The Sigmoid function normalizes the output result to between 0 and 1, and the intermediate feature data u is also a positive value, which ensures that the final adaptive threshold is a positive value and the value is not too large.

[0105] The adaptive threshold is obtained according to the following formula (6):

[0106]

[0107] u is the u in formula (4), that is, the intermediate feature data of GAP input, s is the calculation result in formula (5), and the weight is used to adjust the relative importance of each channel contribution. This process can be understood as applying different degrees of attention to the feature channels.

[0108] DRSN uses a soft threshold function to reduce the residual structure within the network. This technique helps suppress noise and enhance the network's robustness to outliers and interference in the data. Soft thresholding is a common technique in signal denoising algorithms and is implemented using the following formula (7):

[0109]

[0110] In formula (7), x is the input information and τ represents the adaptive threshold.

[0111] The basic idea of ​​soft thresholding is to set the noise components in the signal to zero while retaining the useful components, thereby extracting valuable information from noisy data. The adaptive threshold τ determines the selection range of the noise signal, where the signal with an absolute value less than the adaptive threshold τ is regarded as noise and set to zero, while the signal with an absolute value greater than the adaptive threshold τ is retained as a valuable signal and reduced according to its contribution to the task. Through the synergy of the attention mechanism and soft thresholding, key feature information is retained and noise is reduced, which helps the accuracy of subsequent classification and recognition.

[0112] The residual shrinkage unit (RSU) is the basic structure of DRSN integrating the attention mechanism unit and the soft thresholding unit. By introducing the soft threshold, the channel weighting operation in the SENet attention mechanism is improved. By using the ability of the neural network to automatically update the model parameters through the back propagation algorithm, the adaptive threshold τ in the soft threshold is integrated as a learnable component in the neural network training process to achieve automatic learning.

[0113] Finally, the soft thresholded output is added to the feature data obtained by the identity shortcut to realize the residual connection. The residual connection result is used as the final output of the residual shrinkage unit RSU. The last residual shrinkage unit RSU outputs the classification feature data.

[0114] Specifically, the residual shrinkage module may include 5 residual shrinkage units RSU.

[0115] The classification recognition module includes a global average pooling layer (GAP) applied to all feature channels, and then outputs the capacity estimation result through the last two fully connected layers FC. The first fully connected layer integrates the features from the previous layer, and the second fully connected layer maps the integrated features of the previous fully connected layer to the output space, and outputs the evaluated battery health state SOH.

[0116] As an implementation method, the configuration of the deep residual shrinkage network model of the present invention is shown in the following table:

[0117] DRSN Dilation rate output size Input -- L×2 <![CDATA[Conv1D a (8,3) b ]]> 1 L×8 RSU-1(8,3) 2 L×8 RSU-2(8,3) 3 L×8 RSU-3(16,3) 4 L×16 RSU-4(16,3) 4 L×16 ReLU, GAP, FC(8) -- 8 Linear,FC(1) -- 1

[0118] Conv1D and RSU-l represent the convolutional layer and the lth RSU layer, respectively;

[0119] The first number in the brackets indicates the number of convolution kernels or neurons, and the second number indicates the size of the convolution filter (if any).

[0120] Dilation rate refers to the expansion rate.

[0121] Furthermore, the convolution operations in the convolution module and the residual shrinkage module all adopt the dilated convolution method. All convolution operations in the deep residual shrinkage network model of the present invention maintain the input shape by zero padding. The shape of the original input data is small enough, so the maximum pooling layer is not used. On the contrary, while keeping the data shape unchanged, the dilated convolution strategy is adopted to increase the receptive field, and the dilated convolution method is adopted to strengthen the network structure to adapt to short-term data sequences, ensure the data length during the forward propagation of data, and avoid information loss caused by the pooling operation.

[0122] Compared with the existing technology, the present invention can estimate the capacity of lithium-ion batteries by using random fragment voltage-capacity data obtained during constant current charging. The same training model can be applied to the evaluation of lithium-ion batteries of the same specifications. This method is very suitable for random incomplete charging and discharging conditions commonly seen in industrial scenarios, greatly improving the flexibility and adaptability of the method. Secondly, in order to alleviate the sensitivity of pure data-driven methods to noise, this study used a deep residual shrinkage network (DRSN) with good noise resistance to automatically explore the capacity attenuation features hidden in the local charging curve. It combines the attention mechanism and soft threshold, and is expected to achieve higher result accuracy and stronger noise robustness, which has great application significance.

[0123] See also Fig. 9 The present invention also provides a system for estimating battery SOH based on actual charging data, which is used to execute the aforementioned method for estimating battery SOH based on actual charging data, comprising:

[0124] A data acquisition module (1), used for acquiring battery charging data;

[0125] The model training module (2) is connected to the data acquisition module (1) and is used to construct a deep residual shrinkage network model, take the battery charging data as a training sample as input, train the deep residual shrinkage network model and store it;

[0126] The state evaluation module (3) is connected to the data acquisition module (2) and is used to input the actual battery charging data of the battery to be evaluated into the trained deep residual shrinkage network model, estimate the health state of the battery to be evaluated and output it.

[0127] Furthermore, the data acquisition module (1) comprises:

[0128] An acquisition unit (11) is used to acquire data in a constant current charging stage of a battery as battery charging data;

[0129] A curve fitting unit (12), connected to the acquisition unit (11), is used to fit the battery voltage and battery capacity in the battery charging data to form a constant current charging curve;

[0130] A data sampling unit (13) is connected to the curve fitting unit (12) and is used to intercept a constant current charging curve segment according to a predetermined voltage window, and to perform sampling on the curve segment at a preset voltage interval to obtain a voltage sampling point and a capacity sampling point;

[0131] The sequence forming unit (14) is connected to the data sampling unit (13) and is used to calculate the capacity increment from the capacity sampling points to form a voltage-capacity increment data sequence; the voltage-capacity increment data sequence formed by the sequence forming unit (14) is used as the input of the deep residual shrinkage network model.

[0132] Furthermore, the deep residual shrinkage network model comprises an input layer (21), a convolution module (22), a residual shrinkage module and a classification recognition module (24) connected in sequence, and the residual shrinkage module is composed of at least one residual shrinkage unit (23);

[0133] The input layer (21) is used to receive battery charging data;

[0134] The convolution module (22) is used to perform a convolution operation on the battery charging data to form original feature data;

[0135] The residual shrinkage module is used to extract features and filter noise from the original feature data to obtain classification feature data;

[0136] The classification identification module (24) is used to evaluate the classification feature data to obtain the health status of the battery and output it.

[0137] Furthermore, each residual shrinkage unit (23) includes a residual unit (231), an attention mechanism unit (232) and a soft thresholding unit (232);

[0138] The input of the first residual contraction unit (23) of the residual contraction module is the output of the convolution module (22), the output of the last residual contraction unit (23) is used as classification feature data, the input of each of the remaining residual contraction units (23) is the output of the previous residual contraction unit (23), and the output of each of the remaining residual contraction units (23) is the input of the next residual contraction unit (23);

[0139] The residual unit (231) is used to extract features from the feature data input into the residual shrinkage unit to obtain first feature data;

[0140] The attention mechanism unit (232) is used to calculate and obtain an adaptive threshold according to the first feature data;

[0141] The soft thresholding unit (233) is used to perform soft thresholding processing on the first feature data according to the adaptive threshold to obtain second feature data;

[0142] The residual unit (231) is used to add the feature data input to the residual contraction unit (23) and the second feature data, and use the addition result as the final output feature data.

[0143] Furthermore, the convolution operations in the convolution module (22) and the residual shrinkage module both adopt dilated convolution.

[0144] Compared with the existing technology, the present invention can estimate the capacity of lithium-ion batteries by using random fragment voltage-capacity data obtained during constant current charging. This method is very suitable for random incomplete charging and discharging conditions commonly seen in industrial scenarios, greatly improving the flexibility and adaptability of the method. Secondly, in order to alleviate the sensitivity of pure data-driven methods to noise, this study uses a deep residual shrinkage network (DRSN) with good noise resistance to automatically explore the capacity decay characteristics hidden in the local charging curve. It combines attention mechanisms and soft thresholds, and is expected to achieve higher result accuracy and stronger noise robustness, alleviating the noise problem of pure data-driven methods, and has great application significance.

[0145] The above are only preferred embodiments of the present invention, and are not intended to limit the implementation methods and protection scope of the present invention. Those skilled in the art should be aware that all solutions obtained by equivalent substitutions and obvious changes made using the description and illustrations of the present invention should be included in the protection scope of the present invention.

Claims

1. A method for estimating battery SOH based on actual charging data, characterized in that: include: Get battery charging data; Constructing a deep residual shrinkage network model, taking the battery charging data as a training sample as input, and training the deep residual shrinkage network model; The actual battery charging data of the battery to be evaluated is input into the trained deep residual shrinkage network model, and the health state of the battery to be evaluated is estimated and output.

2. A method for estimating battery SOH based on actual charging data as claimed in claim 1, characterized in that: The battery charging data is data of the constant current charging stage; After acquiring the battery charging data, the battery charging data is preprocessed, and the preprocessing process includes: Fitting the battery voltage and battery capacity in the battery charging data to form a constant current charging curve; A constant current charging curve segment is intercepted according to a predetermined voltage window, and a voltage sampling point and a capacity sampling point are obtained by sampling the curve segment at a preset voltage interval, and the capacity increment is calculated from the capacity sampling point to form a voltage-capacity increment data sequence; The voltage-capacity increment data sequence formed by preprocessing is used as input of the deep residual shrinkage network model.

3. A method for estimating battery SOH based on actual charging data as claimed in claim 1, characterized in that: The deep residual shrinkage network model comprises an input layer, a convolution module, a residual shrinkage module and a classification recognition module connected in sequence, and the residual shrinkage module is composed of at least one residual shrinkage unit; The input layer receives the battery charging data; The convolution module performs a convolution operation on the battery charging data to form original feature data; The residual shrinkage module performs feature extraction and noise filtering on the original feature data to obtain classification feature data; The classification identification module evaluates the classification feature data to obtain the health status of the battery and outputs it.

4. A method for estimating battery SOH based on actual charging data as claimed in claim 3, characterized in that: Each of the residual shrinkage units includes a residual unit, an attention mechanism unit and a soft thresholding unit; The input of the first residual shrinkage unit of the residual shrinkage module is the output of the convolution module, the output of the last residual shrinkage unit is used as the classification feature data, the input of each of the remaining residual shrinkage units is the output of the previous residual shrinkage unit, and the output of each of the remaining residual shrinkage units is the input of the next residual shrinkage unit; The residual unit performs feature extraction on the feature data input to the residual shrinkage unit to obtain first feature data; The attention mechanism unit calculates and obtains an adaptive threshold according to the first feature data; The soft thresholding unit performs soft thresholding processing on the first feature data according to the adaptive threshold to obtain second feature data; The residual unit adds the feature data input to the residual shrinkage unit and the second feature data, and uses the addition result as the final output feature data.

5. A method for estimating battery SOH based on actual charging data as claimed in claim 3, characterized in that: The convolution operations in the convolution module and the residual shrinkage module both adopt dilated convolution methods.

6. A system for estimating battery SOH based on actual charging data, characterized in that: A method for estimating battery SOH based on actual charging data according to any one of claims 1 to 5, comprising: A data acquisition module, used for acquiring battery charging data; A model training module, connected to the data acquisition module, used to construct a deep residual shrinkage network model, taking the battery charging data as a training sample as input, training the deep residual shrinkage network model and storing it; A state evaluation module is connected to the data acquisition module and is used to input the actual battery charging data of the battery to be evaluated into the trained deep residual shrinkage network model, estimate the health state of the battery to be evaluated and output it.

7. A system for estimating battery SOH based on actual charging data as claimed in claim 6, characterized in that: The data acquisition module comprises: An acquisition unit, used for acquiring data of a constant current charging stage of a battery as the battery charging data; A curve fitting unit, connected to the acquisition unit, for fitting the battery voltage and battery capacity in the battery charging data to form a constant current charging curve; A data sampling unit, connected to the curve fitting unit, for intercepting a constant current charging curve segment according to a predetermined voltage window, and sampling on the curve segment at a preset voltage interval to obtain a voltage sampling point and a capacity sampling point; A sequence forming unit, connected to the data sampling unit, for calculating the capacity increment from the capacity sampling points to form a voltage-capacity increment data sequence; The voltage-capacity increment data sequence formed by the sequence forming unit is used as input of the deep residual shrinkage network model.

8. A system for estimating battery SOH based on actual charging data as claimed in claim 6, characterized in that: The deep residual shrinkage network model comprises an input layer, a convolution module, a residual shrinkage module and a classification recognition module connected in sequence, and the residual shrinkage module is composed of at least one residual shrinkage unit; The input layer is used to receive the battery charging data; The convolution module is used to perform a convolution operation on the battery charging data to form original feature data; The residual shrinkage module is used to perform feature extraction and noise filtering on the original feature data to obtain classification feature data; The classification identification module is used to evaluate the classification feature data to obtain the health status of the battery and output it.

9. A system for estimating battery SOH based on actual charging data as claimed in claim 8, characterized in that: Each of the residual shrinkage units includes a residual unit, an attention mechanism unit and a soft thresholding unit; The input of the first residual shrinkage unit of the residual shrinkage module is the output of the convolution module, the output of the last residual shrinkage unit is used as the classification feature data, the input of each of the remaining residual shrinkage units is the output of the previous residual shrinkage unit, and the output of each of the remaining residual shrinkage units is the input of the next residual shrinkage unit; The residual unit is used to extract features from the feature data input into the residual shrinkage unit to obtain first feature data; The attention mechanism unit is used to calculate and obtain an adaptive threshold according to the first feature data; The soft thresholding unit is used to perform soft thresholding processing on the first feature data according to the adaptive threshold to obtain second feature data; The residual unit is used to add the feature data input to the residual shrinkage unit and the second feature data, and use the addition result as the final output feature data.

10. A system for estimating battery SOH based on actual charging data as claimed in claim 8, characterized in that: The convolution operations in the convolution module and the residual shrinkage module both adopt dilated convolution methods.

Images