Lithium battery SOH prediction system and method based on network convergence

Abstract

The invention provides a lithium battery SOH prediction system and method based on network convergence, and relates to the technical field of lithium batteries. The system comprises a charge and discharge test data acquisition module, a charge and discharge data set construction module, an SOH prediction model creation module, a predicted value constraint module and an SOH prediction module. The method is based on the system and comprises the following steps: firstly, acquiring lithium battery charging and discharging original data and cleaning the data to obtain a charging and discharging data set; then fusing the reversible instance normalization module, the channel attention module, the Kalman filter and the multi-layer gating cycle unit prediction module, creating a battery SOH prediction model based on network fusion, and training the model by using a battery charging and discharging data set; and creating a sharpness perception minimization optimizer to train the obtained model and restrain a predicted value, and finally performing SOH prediction on the lithium battery by using the trained battery SOH prediction model based on network fusion.

Description

Technical Field

[0001] The present invention relates to the technical field of lithium batteries, and in particular to a lithium battery SOH prediction system and method based on network fusion. Background Art

[0002] As an important carrier for energy storage and power supply, lithium batteries have the advantages of long cycle life, low self-discharge rate, and low environmental pollution, and are widely used in various fields. Although lithium batteries are rapidly becoming popular, their long-term application faces severe challenges. Similar to machines, key components such as electrodes and diaphragms of lithium batteries will also degrade to varying degrees over time. These negative effects lead to inevitable performance degradation of lithium batteries throughout their life cycle, which is usually manifested as degradation of the battery's state of health (SOH). When the degradation of SOH accumulates to a critical threshold, it will lead to a series of serious consequences, such as operational failures, downtime, and even catastrophic safety issues. Given the complex mechanism of the internal aging reaction of the battery and the many influencing factors, the correlation between the battery parameters measured directly or indirectly and the SOH is still unclear, and the accurate and reliable estimation of SOH is a difficult problem in battery management technology.

[0003] SOH is usually defined as the ratio of current capacity to initial capacity. There are three main methods for predicting SOH:

[0004] The first is a model-based prediction method, which is based on the degradation mechanism of lithium-ion batteries and describes the process of lithium battery health degradation by establishing mathematical and physical models, thereby estimating the battery health status. However, the accuracy of the model-based method depends on the complexity of the model and the accuracy of the parameter setting, and its generalization in practical applications is insufficient.

[0005] The second is a data-driven prediction method. Thanks to the rapid development of data mining technology and big data analysis technology, data-driven methods can use directly or indirectly measured parameters to establish a mapping relationship between parameter characteristics and battery health status. Data-driven methods have high requirements on the quality of input data, and the data distribution drift problem between training data and test data is a long-term challenge in the task of battery health status prediction.

[0006] The third is the prediction method based on fusion, which generally includes the fusion of the same type of methods, the fusion of different types of methods, and the fusion of prediction methods and optimization techniques. The fusion-based method aims to integrate the advantages of multiple methods to eliminate the limitations of a single prediction method. Finding suitable methods and optimization techniques is the key issue in predicting lithium battery SOH based on the fusion method.

[0007] Therefore, how to tap the advantages of fusion-based methods and make full use of the characteristics of multi-dimensional battery data to provide a stable and reliable lithium battery SOH prediction method and system based on network fusion has become an urgent problem to be solved. Summary of the invention

[0008] The technical problem to be solved by the present invention is to provide a lithium battery SOH prediction system and method based on network fusion in view of the deficiencies of the above-mentioned prior art, so as to improve the accuracy and generalization ability of SOH prediction. In order to solve the above-mentioned technical problems, the technical solution adopted by the present invention is: on the one hand, the present invention provides a lithium battery SOH prediction system based on network fusion, including a charge and discharge test data acquisition module, a charge and discharge data set construction module, a SOH prediction model creation module, a prediction value constraint module and a SOH prediction module;

[0009] The charge and discharge test data acquisition module is used to acquire the original charge and discharge data of the lithium battery;

[0010] Furthermore, the charge and discharge test data acquisition module includes a charge and discharge machine and an internal resistance tester; the charge and discharge machine is used to collect real-time voltage, current, temperature, voltage change rate and charge and discharge capacity data during the battery test cycle, and the internal resistance tester is used to collect real-time internal resistance data of the battery; the data collected by the charge and discharge machine and the data collected by the internal resistance tester are merged according to the same time starting point to obtain the battery original data set.

[0011] The charge and discharge data set construction module is used to clean and process the raw data of lithium battery charge and discharge to obtain a charge and discharge data set;

[0012] The SOH prediction model creation module is used to create a battery SOH prediction model based on network fusion, and train the created model using the charge and discharge data set;

[0013] Furthermore, the battery SOH prediction model based on network fusion includes a data preprocessing module, an instance normalization module, a channel attention module, an instance denormalization module, a Kalman filter module and a multi-layer gated recurrent unit prediction module connected in sequence;

[0014] The data preprocessing module is used to perform normalization, noise addition and encoding operations on the charge and discharge data set, and convert the charge and discharge data set into a multi-dimensional feature vector;

[0015] The instance normalization module is used to normalize the multidimensional feature vector using the mean and variance of the feature vector itself, remove the non-stationary information in the multidimensional feature vector, and obtain the normalized output feature;

[0016] The channel attention module is used to assign the attention matrix as a weight to the channel of the normalized output feature to obtain the output feature of the channel attention module;

[0017] The instance denormalization module denormalizes the normalized output features and the output features of the attention module to restore the non-stationary information of the multi-dimensional feature vector and obtain the denormalized output features;

[0018] The Kalman filter module is used to filter out feature redundancy from the inverse normalized output features, calculate the correlation between the multi-dimensional features and the battery SOH change, derive the weight relationship, and obtain the fused refined features;

[0019] The multi-layer gated cyclic unit prediction module is used to solve the mapping relationship between the refined features and the battery SOH value, and obtain the predicted value of the battery SOH prediction model based on network fusion.

[0020] The prediction value constraint module is used to create a sharpness perception minimization optimizer to constrain the prediction value of the battery SOH prediction model based on network fusion, so as to obtain a trained battery SOH prediction model based on network fusion;

[0021] The SOH prediction module is used to predict the SOH of the lithium battery using the trained battery SOH prediction model based on network fusion.

[0022] On the other hand, the present invention also provides a lithium battery SOH prediction method based on network fusion, comprising the following steps:

[0023] Step 1: Use the charge and discharge test data acquisition module to collect the raw data of lithium battery charge and discharge;

[0024] Step 1.1: Fix the test battery in the customized battery fixture and connect the charger and discharger and the internal resistance tester;

[0025] Step 1.2: Set the charging and discharging machine steps according to the battery charging and discharging test requirements, carry out the charging and discharging cycle, and collect the real-time voltage, current, temperature, voltage change rate and charging and discharging capacity data during the battery test cycle;

[0026] Step 1.3: Use the internal resistance tester to collect the real-time internal resistance data of the battery, process the internal resistance data sampling points, and ensure that the data collection interval is the same as the data frequency collected by the charger and discharger;

[0027] Step 1.4: Combine the data collected by the charger and discharger and the data collected by the internal resistance tester into a battery original data set at the same starting time.

[0028] In the battery original data set, voltage data is the first dimension data, current data is the second dimension data, temperature is the third dimension data, charging capacity is the fourth dimension data, discharging capacity is the fifth dimension data, voltage change rate is the sixth dimension data, and internal resistance data is the seventh dimension data.

[0029] Step 2: Use the charge and discharge data set construction module to clean and smooth the battery raw data, and calculate the voltage change rate and temperature change rate and add them to the cleaned and smoothed battery data to obtain the battery charge and discharge data set;

[0030] Step 3: Use the SOH prediction model creation module to create a battery SOH prediction model based on network fusion, and use the battery charging and discharging data set to train the created model;

[0031] Furthermore, the battery SOH prediction model based on network fusion includes a data preprocessing module, an instance normalization module, a channel attention module, an instance denormalization module, a Kalman filter module and a multi-layer gated recurrent unit prediction module connected in sequence;

[0032] The data preprocessing module normalizes the battery charge and discharge data set, then adds Gaussian white noise to the normalized battery charge and discharge data set, and finally constructs an automatic encoder to encode the obtained data to extract multi-dimensional features;

[0033] The instance normalization module uses the mean and variance of the feature vector itself to normalize the multidimensional feature vector, removes non-stationary information in the multidimensional feature vector, and obtains normalized output features;

[0034] The channel attention module obtains the channel dependency of feature vectors in different channels, takes the normalized output features as input vectors, calculates the attention matrix corresponding to each channel information in the input vector, and assigns the attention matrix to the channel as a weight. After completing the weighting operation, the input features with weights become the output features of the channel attention module;

[0035] The instance denormalization module denormalizes the normalized output features and the output features of the attention module using the mean and variance of the feature vector itself, recovers the non-stationary information of the multi-dimensional feature vector, and obtains the denormalized output features;

[0036] The Kalman filter module is used to filter out feature redundancy from the inverse normalized features, calculate the correlation between the multi-dimensional features and the battery SOH change, derive the weight relationship, and obtain the fused refined features, which can be expressed mathematically as follows:

[0037]

[0038] in, is the refined feature after fusion, KF(·) is the Kalman filter state equation; m corresponds to the seven dimensions of the battery charge and discharge data set; and b m Represent the m-th dimension feature and the learnable weight of the m-th dimension feature respectively;

[0039] The multi-layer gated cycle unit prediction module is used to solve the mapping relationship between the refined features and the battery SOH value, and then obtain the predicted value of the battery SOH prediction model based on network fusion, which is mathematically expressed as:

[0040]

[0041] h i+1 =GRU(h i )

[0042] Among them, h 0 is the initial state of the recurrent gating unit, h i+1 is the output of the i+1th iteration of the recurrent gating unit, which is the SOH prediction value;

[0043] Step 4: Use the prediction value constraint module to create a sharpness-aware minimization optimizer, and use the sharpness-aware minimization optimizer to train the model obtained in step 3 and constrain the prediction value to obtain a trained battery SOH prediction model based on network fusion;

[0044] The sharpness-aware minimization optimizer selects a mean square error loss function as the target optimization function and adopts a cosine annealing learning rate strategy to guide the learning rate update of the optimizer;

[0045] Step 5: The SOH prediction module uses the battery SOH prediction model based on network fusion trained in step 4 to predict the SOH of the lithium battery.

[0046] The beneficial effects of adopting the above technical solution are:

[0047] (1) The present invention is based on network fusion and determines the selected method network structure according to the characteristics of the battery sequence data set, including a reversible instance normalization module, a channel attention module, a Kalman filter, and a multi-layer recurrent unit prediction module. The reversible instance normalization module is used to solve the data distribution drift problem in the battery sequence data, thereby improving the generalization ability of the SOH prediction method; the channel attention module and the Kalman filter are used to optimize the feature extraction ability of the SOH prediction method; and the multi-layer gated recurrent unit prediction module is used to solve the mapping relationship between SOH features and SOH values, thereby improving the prediction accuracy of the SOH prediction method.

[0048] (2) The present invention provides a method for multi-dimensional battery data fusion. The sharpness-aware minimization optimizer is used to train the lithium battery SOH prediction model based on network fusion, and the fusion weights of the multi-dimensional battery data features are returned according to the gradient direction of the loss function value minimization and the loss sharpness minimization. The Kalman filter module fuses the multi-dimensional battery data into a refined feature containing battery SOH information according to their respective weights, thereby improving the utilization rate of battery data and the training efficiency of the battery SOH prediction model.

[0049] (3) The present invention also designs a battery SOH prediction system suitable for the embedded computing device NVIDIA Jetson Xavier NX that matches the battery SOH prediction method. The SOH prediction method can be deployed on the embedded computing device NVIDIA Jetson Xavier NX, and SOH prediction model training and output predicted SOH values ​​can be implemented without large computing equipment, providing a better solution for the implementation of the SOH prediction method in actual application environments. BRIEF DESCRIPTION OF THE DRAWINGS

[0050] Figure 1 A schematic diagram of the structure of a lithium battery SOH prediction system based on network fusion provided in an embodiment of the present invention;

[0051] Figure 2 A schematic diagram of a battery SOH prediction model based on network fusion provided in an embodiment of the present invention;

[0052] Figure 3 A flowchart of a lithium battery SOH prediction method based on network fusion provided in an embodiment of the present invention;

[0053] Figure 4 A deployment flow chart on an embedded computing device NVIDIA Jetson Xavier NX provided by an embodiment of the present invention; DETAILED DESCRIPTION

[0054] The specific implementation of the present invention is further described in detail below in conjunction with the accompanying drawings and examples. The following examples are used to illustrate the present invention, but are not intended to limit the scope of the present invention.

[0055] In this embodiment, a lithium battery SOH prediction system based on network fusion is provided. Figure 1 As shown, it includes a charge and discharge test data acquisition module, a charge and discharge data set construction module, a SOH prediction model creation module, a prediction value constraint module and a SOH prediction module;

[0056] The charge and discharge test data acquisition module is used to acquire the original charge and discharge data of the lithium battery;

[0057] The charge and discharge test data acquisition module includes a charge and discharge machine and an internal resistance tester; the charge and discharge machine is used to collect real-time voltage, current, temperature, voltage change rate, and charge and discharge capacity data during a battery test cycle; the internal resistance tester is used to collect real-time internal resistance data of the battery, process the internal resistance data sampling points, and ensure that the data collection interval is the same as the data collected by the charge and discharge machine; the data collected by the charge and discharge machine and the data collected by the internal resistance tester are merged at the same time starting point to obtain the battery original data set.

[0058] The charge and discharge data set construction module is used to clean and process the raw data of lithium battery charge and discharge to obtain a charge and discharge data set;

[0059] The SOH prediction model creation module is used to create a battery SOH prediction model based on network fusion, and train the created model using the charge and discharge data set;

[0060] The battery SOH prediction model based on network fusion, such as Figure 2 As shown, it includes a data preprocessing module, an instance normalization module, a channel attention module, an instance denormalization module, a Kalman filter module and a multi-layer gated recurrent unit prediction module connected in sequence;

[0061] The data preprocessing module is used to perform normalization, noise addition and encoding operations on the charge and discharge data set, and convert the charge and discharge data set into a multi-dimensional feature vector;

[0062] The instance normalization module is used to normalize the multidimensional feature vector using the mean and variance of the feature vector itself, remove the non-stationary information in the multidimensional feature vector, and obtain the normalized output feature;

[0063] The channel attention module is used to obtain the channel dependency of feature vectors in different channels, take the normalized output features as input vectors, calculate the attention matrix corresponding to each channel information in the input vector, and assign the attention matrix to the channel as a weight. After completing the weighting operation, the input features with weights become the output features of the channel attention module.

[0064] The instance denormalization module denormalizes the normalized output features and the output features of the attention module using the mean and variance of the feature vector itself, recovers the non-stationary information of the multi-dimensional feature vector, and obtains the denormalized output features;

[0065] The Kalman filter module is used to filter out feature redundancy from the inverse normalized output features to reduce the computational cost of the model; the weight parameters are calculated based on the correlation between the multi-dimensional features and the battery SOH changes, and the weight parameters guide the fusion of multi-dimensional features. The fused refined features contain information related to the battery SOH in all dimensional features, and the input of the subsequent model changes from multi-dimensional features to a fused refined feature, which is conducive to the prediction value constraint module to quickly and efficiently find the model optimization direction and improve the training efficiency of the model.

[0066] The multi-layer gated cyclic unit prediction module is used to solve the mapping relationship between the refined features and the battery SOH value, and then obtain the predicted value of the battery SOH prediction model based on network fusion.

[0067] The prediction value constraint module is used to create a sharpness perception minimization optimizer to constrain the prediction value of the battery SOH prediction model based on network fusion, so as to obtain a trained battery SOH prediction model based on network fusion;

[0068] The SOH prediction module is used to predict the SOH of the lithium battery using the trained battery SOH prediction model based on network fusion.

[0069] In order to improve the accuracy and generalization of SOH prediction, the present invention provides a lithium battery SOH prediction method based on network fusion. The basic idea of ​​the method is as follows: the existing battery SOH prediction method usually uses a single model-based or data-driven method, and both the model-based and data-driven methods have certain limitations, resulting in insufficient accuracy and generalization of the final SOH prediction model. In response to this problem, the method of the present invention constructs a battery SOH prediction model that integrates a reversible instance normalization module, a channel attention module, a Kalman filter, and a multi-layer gated recurrent unit prediction module, extracts and fuses features of the input data set, and can better learn the feature representation of the multi-dimensional battery sequence. The present invention uses instance normalization to normalize the input features, eliminates the non-stationary information in the input features, inputs the normalized features into the channel attention module to learn the output, and uses inverse instance normalization on the output to explicitly restore the non-stationary information, so that the model ignores the data distribution drift problem during learning and avoids the loss of non-stationary information. Under the guidance of the objective optimization function and the sharpness-aware minimization optimizer, the Kalman filter module assigns weights to the multi-dimensional battery sequence features and fuses them into battery SOH refined features, thereby improving the utilization of input data and the training efficiency of the battery SOH prediction model.

[0070] In this embodiment, a lithium battery SOH prediction method based on network fusion is provided. Figure 3 As shown, the following steps are included:

[0071] Step 1: Use the charge and discharge test data acquisition module to collect the raw data of lithium battery charge and discharge;

[0072] Step 1.1: Fix the test battery in the customized battery fixture and connect the charger and discharger and the internal resistance tester;

[0073] Step 1.2: Set the charging and discharging machine steps according to the battery charging and discharging test requirements, carry out the charging and discharging cycle, and collect the real-time voltage, current, temperature, voltage change rate and charging and discharging capacity data during the battery test cycle;

[0074] In the embodiment, the battery charging and discharging process step setting is exemplified as follows: the battery is charged in a 24 degree Celsius environment in a 1C constant current mode until the battery voltage reaches the charging cut-off voltage, and then charged in a constant voltage mode until the current drops to 20mA; after charging is completed, it is left for 30 minutes and discharged in a 1C constant current mode until the battery voltage reaches the discharge cut-off voltage, and then discharged in a constant voltage mode until the current drops to 20mA, and a charging and discharging cycle is completed.

[0075] Step 1.3: Use the internal resistance tester to collect the real-time internal resistance data of the battery, process the internal resistance data sampling points, and ensure that the data collection interval is the same as the data frequency collected by the charger and discharger;

[0076] Step 1.4: Combine the data collected by the charger and discharger and the data collected by the internal resistance tester into a battery original data set with a CSV file suffix based on the same starting time.

[0077] In the battery original data set with the CSV file suffix, voltage data is the first dimension data, current data is the second dimension data, temperature is the third dimension data, charging capacity is the fourth dimension data, discharging capacity is the fifth dimension data, voltage change rate is the sixth dimension data, and internal resistance data is the seventh dimension data.

[0078] Step 2: Use the charge and discharge data set construction module to clean the battery raw data, delete the outliers in the data and use the interpolation method to supplement the missing values. Use the moving average method to smooth the points with large voltage, temperature, and current fluctuations caused by sampling noise; use the python tool to process the voltage and temperature data to calculate the voltage change rate and temperature change rate, and add the voltage change rate and temperature change rate data to the battery data after cleaning and smoothing to obtain the battery charge and discharge data set;

[0079] Step 3: Use the SOH prediction model creation module to create a battery SOH prediction model based on network fusion, and use the battery charging and discharging data set to train the created model;

[0080] The battery SOH prediction model based on network fusion includes a data preprocessing module, an instance normalization module, a channel attention module, an instance denormalization module, a Kalman filter module and a multi-layer gated recurrent unit prediction module connected in sequence;

[0081] The data preprocessing module performs normalization, noise addition, and encoding operations on the input charge and discharge data set in sequence. The battery charge and discharge data set is normalized to the range of [0, 1], and then Gaussian white noise is added to the normalized battery charge and discharge data set. Finally, an automatic encoder is constructed to encode the obtained data to extract multi-dimensional features;

[0082] The instance normalization module uses the mean and variance of the feature vector itself to normalize the multidimensional feature vector, removes the non-stationary information in the multidimensional feature vector, and obtains the normalized output feature, which is mathematically expressed as:

[0083]

[0084]

[0085] Among them, x′ represents the multi-dimensional input features extracted by the autoencoder; x′ j represents the input feature of time node j; E[x′] and Var[x′] represent the mean and variance of the multi-dimensional input feature respectively; represents the normalized output features; λ and β are learnable affine transformation parameters; T x The number of time nodes representing the input features; b″ is a learnable bias parameter;

[0086] The channel attention module obtains the channel dependency of feature vectors in different channels, takes the normalized output features as input vectors, calculates the attention matrix corresponding to each channel information in the input vector, and assigns the attention matrix to the channel as a weight. After completing the weighting operation, the input features with weights become the output features of the channel attention module.

[0087] The mathematical representation of the attention matrix is:

[0088]

[0089] in, So As the input attention matrix; softmax is the activation function; w Q and are all attention weight matrices; d m Represents the dimension of the channel attention module input;

[0090] The example denormalization module denormalizes the normalized output features and the output features of the attention module using the mean and variance of the feature vector itself, recovers the non-stationary information of the multi-dimensional feature vector, and obtains the denormalized output features, which can be expressed mathematically as:

[0091]

[0092] in, is the output feature after inverse normalization;

[0093] The Kalman filter module is used to filter out feature redundancy from the features after inverse normalization, reducing the computational cost of the model; the weight parameters are calculated based on the correlation between the multi-dimensional features and the battery SOH changes, and the weight parameters guide the fusion of multi-dimensional features. The fused refined features contain information related to the battery SOH in all dimensional features. The input of the subsequent model changes from multi-dimensional features to a fused refined feature, which is conducive to the prediction value constraint module to quickly and efficiently find the model optimization direction and improve the training efficiency of the model. The mathematical expression is:

[0094]

[0095] in, is the refined feature after fusion of the output, KF(·) is the Kalman filter state equation; m corresponds to the seven dimensions of the battery charge and discharge data set; and b m Represent the m-th dimension feature and the learnable weight of the m-th dimension feature respectively;

[0096] The multi-layer gated cyclic unit prediction module is used to solve the mapping relationship between the refined features and the battery SOH value, and then obtain the predicted value of the battery SOH prediction model based on network fusion.

[0097] In this embodiment, a three-layer gated recurrent unit is used to solve the mapping relationship. Given that the battery data set created by the charge and discharge data set construction module has a large amount of data, a single-layer gated recurrent unit cannot fully compress the information in the refined features. A three-layer gated recurrent unit can ensure the integrity of the information while taking into account the training efficiency. Mathematical expression:

[0098]

[0099] h i+1 =GRU(h i ) (8)

[0100] Among them, h 0 is the initial state of the recurrent gating unit, h i+1 It is the output of the i+1th iteration of the recurrent gating unit, that is, the SOH prediction value.

[0101] Step 4: Use the prediction value constraint module to create a sharpness-aware minimization optimizer. The optimizer selects the mean square error loss function as the target optimization function and uses the cosine annealing learning rate strategy to guide the learning rate update of the optimizer. The sharpness-aware minimization optimizer trains the model obtained in step 3 and constrains the prediction value to obtain a trained battery SOH prediction model based on network fusion.

[0102] The formula of the objective optimization function is:

[0103]

[0104] Where L is the loss value, l(·) represents the loss equation, and x target They represent the predicted and actual values ​​of SOH respectively, and n represents the number of SOH values.

[0105] Step 5: The SOH prediction module uses the battery SOH prediction model based on network fusion trained in step 4 to predict the SOH of the lithium battery.

[0106] In an embodiment, Figure 4 As shown, the SOH prediction module is used to realize SOH prediction on the trained battery SOH prediction model based on network fusion using the battery charge and discharge data set, and the battery SOH prediction method is deployed on the embedded computing device NVIDIA Jetson Xavier NX, providing a better solution for the implementation of the SOH prediction method in the actual application environment.

[0107] The above is a detailed introduction to a lithium battery SOH prediction system and method based on network fusion provided by the present invention. In this embodiment, specific examples are used to illustrate the principles and implementation methods of the present invention. The description of the above embodiments is only used to help understand the core idea of ​​the present invention. It should be pointed out that for ordinary technicians in this technical field, without departing from the principles of the present invention, several improvements and modifications can be made to the present invention, and these improvements and modifications also fall within the scope of protection of the claims of the present invention.

Claims

1. A lithium battery SOH prediction system based on network fusion, characterized by: It includes a charge and discharge test data acquisition module, a charge and discharge data set construction module, a SOH prediction model creation module, a prediction value constraint module and a SOH prediction module; The charge and discharge test data acquisition module is used to acquire the original charge and discharge data of the lithium battery; The charge and discharge data set construction module is used to clean and process the raw data of lithium battery charge and discharge to obtain a charge and discharge data set; The SOH prediction model creation module is used to create a battery SOH prediction model based on network fusion, and train the created model using the charge and discharge data set; The prediction value constraint module is used to create a sharpness perception minimization optimizer to constrain the prediction value of the battery SOH prediction model based on network fusion, so as to obtain a trained battery SOH prediction model based on network fusion; The SOH prediction module is used to predict the SOH of the lithium battery using the trained battery SOH prediction model based on network fusion.

2. According to claim 1, a lithium battery SOH prediction system based on network fusion is characterized in that: The charge and discharge test data acquisition module includes a charge and discharge machine and an internal resistance tester; the charge and discharge machine is used to collect real-time voltage, current, temperature, voltage change rate and charge and discharge capacity data during the battery test cycle, and the internal resistance tester is used to collect real-time internal resistance data of the battery; the data collected by the charge and discharge machine and the data collected by the internal resistance tester are merged at the same time starting point to obtain the battery original data set.

3. The lithium battery SOH prediction system based on network fusion according to claim 1 is characterized in that: The battery SOH prediction model based on network fusion includes a data preprocessing module, an instance normalization module, a channel attention module, an instance denormalization module, a Kalman filter module and a multi-layer gated recurrent unit prediction module connected in sequence; The data preprocessing module is used to perform normalization, noise addition and encoding operations on the charge and discharge data set, and convert the charge and discharge data set into a multi-dimensional feature vector; The instance normalization module is used to normalize the multidimensional feature vector using the mean and variance of the feature vector itself, remove the non-stationary information in the multidimensional feature vector, and obtain the normalized output feature; The channel attention module is used to assign the attention matrix as a weight to the channel of the normalized output feature to obtain the output feature of the channel attention module; The instance denormalization module denormalizes the normalized output features and the output features of the attention module to restore the non-stationary information of the multi-dimensional feature vector and obtain the denormalized output features; The Kalman filter module is used to filter out feature redundancy from the inverse normalized output features, calculate the correlation between the multi-dimensional features and the battery SOH change, derive the weight relationship, and obtain the fused refined features; The multi-layer gated cyclic unit prediction module is used to solve the mapping relationship between the refined features and the battery SOH value, and obtain the predicted value of the battery SOH prediction model based on network fusion.

4. A lithium battery SOH prediction method based on network fusion, implemented based on a lithium battery SOH prediction system based on network fusion according to claim 1, characterized in that: The following steps are involved: Step 1: Use the charge and discharge test data acquisition module to collect the raw data of lithium battery charge and discharge; Step 2: Use the charge and discharge data set construction module to clean and smooth the battery raw data, and calculate the voltage change rate and temperature change rate and add them to the cleaned and smoothed battery data to obtain the battery charge and discharge data set; Step 3: Use the SOH prediction model creation module to create a battery SOH prediction model based on network fusion, and use the battery charging and discharging data set to train the created network model; Step 4: Use the prediction value constraint module to create a sharpness-aware minimization optimizer, and use the sharpness-aware minimization optimizer to train the model obtained in step 3 and constrain the prediction value to obtain a trained battery SOH prediction model based on network fusion; Step 5: The SOH prediction module uses the battery SOH prediction model based on network fusion trained in step 4 to predict the SOH of the lithium battery.

5. The method for predicting SOH of a lithium battery based on network fusion according to claim 4 is characterized in that: The details of step 1 are as follows: Step 1.1: Fix the test battery in the customized battery fixture and connect the charger and discharger and the internal resistance tester; Step 1.2: Set the charging and discharging machine steps according to the battery charging and discharging test requirements, carry out the charging and discharging cycle, and collect the real-time voltage, current, temperature, voltage change rate and charging and discharging capacity data during the battery test cycle; Step 1.3: Use the internal resistance tester to collect the real-time internal resistance data of the battery, process the internal resistance data sampling points, and ensure that the data collection interval is the same as the data frequency collected by the charger and discharger; Step 1.4: Combine the data collected by the charger and discharger and the data collected by the internal resistance tester into a battery original data set at the same starting time.

6. The method for predicting SOH of a lithium battery based on network fusion according to claim 4 is characterized in that: The battery SOH prediction model based on network fusion includes a data preprocessing module, an instance normalization module, a channel attention module, an instance denormalization module, a Kalman filter module and a multi-layer gated recurrent unit prediction module connected in sequence; The data preprocessing module normalizes the battery charge and discharge data set, then adds Gaussian white noise to the normalized battery charge and discharge data set, and finally constructs an automatic encoder to encode the obtained data to extract multi-dimensional features; The instance normalization module uses the mean and variance of the feature vector itself to normalize the multidimensional feature vector, removes non-stationary information in the multidimensional feature vector, and obtains normalized output features; The channel attention module obtains the channel dependency of feature vectors in different channels, takes the normalized output features as input vectors, calculates the attention matrix corresponding to each channel information in the input vector, and assigns the attention matrix to the channel as a weight. After completing the weighting operation, the input features with weights become the output features of the channel attention module; The instance denormalization module denormalizes the normalized output features and the output features of the attention module using the mean and variance of the feature vector itself, recovers the non-stationary information of the multi-dimensional feature vector, and obtains the denormalized output features; The Kalman filter module is used to filter out feature redundancy from the inverse normalized features, calculate the correlation between the multi-dimensional features and the battery SOH change, derive the weight relationship, and obtain the fused refined features, which can be expressed mathematically as follows: in, is the refined feature after fusion, KF(·) is the Kalman filter state equation; m corresponds to the seven dimensions of the battery charge and discharge data set; and b m Represent the m-th dimension feature and the learnable weight of the m-th dimension feature respectively; The multi-layer gated cycle unit prediction module is used to solve the mapping relationship between the refined features and the battery SOH value, and then obtain the predicted value of the battery SOH prediction model based on network fusion, which is mathematically expressed as: h i+1 =GRU(h i ) Among them, h0 is the initial state of the recurrent gating unit, h i+1 It is the output of the i+1th iteration of the recurrent gating unit, that is, the SOH prediction value.

7. The method for predicting SOH of a lithium battery based on network fusion according to claim 4 is characterized in that: The sharpness-aware minimization optimizer selects a mean square error loss function as the target optimization function and adopts a cosine annealing learning rate strategy to guide the learning rate update of the optimizer.

Images