Method and system for monitoring and predicting health state of storage battery based on multi-modal feature fusion

Abstract

The invention provides a method and a system for monitoring and predicting the health state of a storage battery based on multi-modal feature fusion, and relates to the technical field of storage battery monitoring, and the method comprises the steps: obtaining multi-modal data through periodically collecting static data of the storage battery and collecting dynamic data of the storage battery in real time; and then the method is realized through the steps of data monitoring, shared bottom layer feature extraction, SOH prediction, RUL prediction, multi-target joint optimization and the like. According to the method, multi-modal feature data generated by the storage battery can be monitored and analyzed on line, a multi-modal feature fusion and multi-target joint prediction and optimization mechanism is introduced, and SOH and RUL are predicted by adopting independent branches based on an SOH prediction model and an RUL prediction model, so that the health state of the storage battery is predicted; according to the method, automatic analysis on multi-modal mass data of the battery can be realized, the workload is reduced, the analysis efficiency is improved, the accuracy of a prediction result can be improved, and meanwhile, a constructed feature sharing mechanism can reduce the calculation complexity and improve the prediction efficiency.

Description

Technical Field

[0001] The present invention relates to the technical field of battery monitoring, and particularly to a method and system for monitoring and predicting the health state of a battery by multi-modal feature fusion. Background Art

[0002] The safe and stable operation of a battery pack is one of the important guarantees for ensuring energy supply and economic development. In recent years, with the continuous progress of new energy large model technology, many shortcomings have emerged in the traditional battery pack monitoring methods that rely on manual operations in the past. For example, for the massive data generated by various important indicators of the battery such as temperature, voltage, and internal resistance, mainly the end-side devices continuously send the massive multi-dimensional data of the battery to the system through protocols, compare them with the pre-set critical values of the battery parameters, and then mainly predict the health state of the battery through manual analysis, without using algorithms to predict the health state of the battery. The main defects of this traditional method that only relies on manual analysis of massive data are as follows: 1. Analyzing the massive data generated by the battery solely by manual operation is inefficient.

[0003] 2. Operating the battery pack, such as capacity verification operation, requires a large number of parameters, the system interface is complex, and it has high requirements for the workers' power knowledge reserve and their proficiency in operating the system.

[0004] 3. The results of manually predicting the remaining life of the battery are inaccurate, which greatly wastes battery energy.

[0005] In summary, the traditional method that only relies on manual analysis of massive data is not only inefficient and has a large workload, but also cannot accurately predict the remaining life of the battery, resulting in a great waste of energy.

[0006] Based on this, the present invention is proposed. Summary of the Invention

[0007] To solve the problems of low efficiency and large error existing in the traditional manual analysis method, the purpose of the present invention is to provide a method and system for monitoring and predicting the health state of a battery by multi-modal feature fusion, which can not only automatically analyze the multi-modal massive data of the battery, reduce the workload, improve the analysis efficiency, but also improve the accuracy of the prediction results.

[0008] To achieve the above purpose, the present invention provides the following technical solutions: The present invention provides a method for monitoring and predicting the health state of a battery with multi-modal feature fusion in a first aspect, including: data acquisition, periodically acquiring static data of the battery and real-time acquiring dynamic data of the battery to obtain multi-modal data; data monitoring, aggregating and monitoring the acquired multi-modal data; shared underlying feature extraction, performing feature processing on the multi-modal data based on a shared underlying feature extraction model to obtain shared underlying features; SOH prediction, performing feature extraction and fusion on the obtained shared underlying features based on an SOH prediction model, and outputting an SOH prediction result; RUL prediction, performing feature extraction and fusion on the obtained shared underlying features based on an RUL prediction model, and outputting an RUL prediction result; multi-objective joint optimization, respectively calculating the SOH loss and the RUL loss and designing a weighted multi-objective loss function to obtain a total loss, and optimizing the prediction error.

[0009] The present invention provides a preferred solution in a first aspect, where the static data includes electrolyte concentration and self-discharge rate; the dynamic data includes single-cell voltage of the battery, charge and discharge current, ambient temperature, internal resistance parameter, and charge and discharge cycle times.

[0010] The present invention provides a preferred solution in a first aspect, where the shared underlying feature extraction model uses LSTM, SAE or an attention network.

[0011] The present invention provides a preferred solution in a first aspect, where the shared underlying features include shared underlying features derived from static data and shared underlying features derived from dynamic data; in the SOH prediction step, cross-cycle feature encoding is performed on the shared underlying features derived from static data to extract trend features including long-term aging features, and intra-cycle encoding is performed on the shared underlying features derived from dynamic data to extract temporal dynamic features including abnormal features.

[0012] The present invention provides a preferred solution in a first aspect, where the SOH prediction model includes a multi-layer perceptron encoder, a transposed convolution encoder, a splicing layer, and a fully connected layer; cross-cycle feature encoding is performed on the shared underlying features derived from static data through the multi-layer perceptron encoder to obtain a cross-cycle feature vector; intra-cycle encoding is performed on the shared underlying features derived from dynamic data through the transposed convolution encoder to obtain an intra-cycle feature vector; the cross-cycle and intra-cycle feature vectors are spliced through the splicing layer, and a dynamic weight allocation mechanism is introduced to adjust the contribution degree of each feature vector according to the battery type and usage scenario, and then output through the fully connected layer.

[0013] The present invention provides a preferred solution in a first aspect, where the fully connected layer outputs an estimated value of the capacity attenuation rate or an estimated value of the internal resistance change rate, and calculates an SOH estimated value based on the estimated value of the capacity attenuation rate or the estimated value of the internal resistance change rate as the SOH prediction result.

[0014] In the first aspect, the present invention provides a preferred solution. The RUL prediction model includes: a feature extraction module, based on a multi-layer feedforward network and a residual connection structure, sharing underlying features for feature extraction to extract a high-dimensional representation of degradation features; a multi-head attention module, calculating through parallel subspace attention to capture the non-linear coupling relationship between multi-source features; a self-supervised learning module, designing an auxiliary task based on time series prediction to jointly optimize the feature extraction and the main task of RUL prediction; an RUL prediction output module, outputting the RUL prediction value through a linear layer and a matrix fusion operation.

[0015] In the first aspect, the present invention provides a preferred solution. In the multi-objective joint optimization step, the weighted multi-objective loss function is expressed as:

[0016] Wherein, is the total loss, is the SOH loss, is the RUL loss, is the weight.

[0017] In the first aspect, the present invention provides a preferred solution. The weighted multi-objective loss function adopts a dynamically weighted multi-objective loss function, and the is the dynamic weight, which is adjusted according to the task priority.

[0018] In the second aspect, the present invention provides a multi-modal feature fusion-based battery health status monitoring and prediction system for implementing the above method, including: a data acquisition unit, used for periodically collecting the static data of the battery and real-time collecting the dynamic data of the battery to obtain multi-modal data; a data monitoring unit, used for collecting and monitoring the obtained multi-modal data; a shared underlying feature extraction unit, used for performing feature processing on the multi-modal data based on a shared underlying feature extraction model to obtain shared underlying features; an SOH prediction unit, used for performing feature extraction and fusion on the obtained shared underlying features based on an SOH prediction model and then outputting an SOH prediction result; an RUL prediction unit, used for performing feature extraction and fusion on the obtained shared underlying features based on an RUL prediction model and then outputting an RUL prediction result; a multi-objective joint optimization unit, used for respectively calculating the SOH loss and the RUL loss and designing a weighted multi-objective loss function to obtain the total loss and optimize the prediction error.

[0019] Compared with the prior art, the present invention has the following beneficial technical effects: The present invention can perform online monitoring and analysis on multi-modal feature data generated by a storage battery. At the same time, a multi-modal feature fusion and multi-objective joint prediction and optimization mechanism is introduced. Based on the SOH prediction model and the RUL prediction model, independent branches are used to predict the SOH (state of health) and RUL (remaining useful life) respectively, so as to predict the health state of the storage battery. It can not only automatically analyze the multi-modal massive data of the battery, reduce the workload, improve the analysis efficiency, but also improve the accuracy of the prediction results. Moreover, based on the shared underlying feature extraction model, the multi-modal data is characterized to obtain shared underlying features, and a feature sharing mechanism is constructed, which can reduce redundant calculations, reduce the computational complexity, and improve the prediction efficiency. Description of the Drawings

[0020] In order to more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the following will briefly introduce the drawings required for the description of the embodiments or the prior art. Obviously, the drawings in the following description are only the embodiments of the present invention. For those of ordinary skill in the art, without creative efforts, other drawings can be obtained according to the provided drawings.

[0021] Figure 1 It is a flowchart of the method for monitoring and predicting the health state of a storage battery with multi-modal feature fusion provided in Embodiment 1 of the present invention; Figure 2 It is a block diagram of the modules of the system for monitoring and predicting the health state of a storage battery with multi-modal feature fusion provided in Embodiment 1 of the present invention; Figure 3 It is an overall architecture diagram of the system for monitoring and predicting the health state of a storage battery with multi-modal feature fusion provided in Embodiment 2 of the present invention; Figure 4 It is an architecture diagram of the SOH prediction model in the method and system for monitoring and predicting the health state of a storage battery with multi-modal feature fusion provided in Embodiment 2 of the present invention; Figure 5 It is an architecture diagram of the RUL prediction model in the method and system for monitoring and predicting the health state of a storage battery with multi-modal feature fusion provided in Embodiment 2 of the present invention.

[0022] Reference numerals: data acquisition unit 1, acquisition module 11, charge and discharge module 12, data monitoring unit 2, storage battery monitoring host 21, shared underlying feature extraction unit 3, SOH prediction unit 4, RUL prediction unit 5, multi-objective joint optimization unit 6, first storage battery group 71, second storage battery group 72. Detailed Embodiments

[0023] Next, the technical solutions in the embodiments of the present invention will be clearly and completely described in conjunction with the accompanying drawings in the embodiments of the present invention. Obviously, the described embodiments are only a part of the embodiments of the present invention, rather than all the embodiments. All other embodiments obtained by those of ordinary skill in the art based on the embodiments of the present invention without creative efforts shall fall within the protection scope of the present invention. Embodiment 1

[0024] Please refer to Figure 1 , this embodiment provides a method for monitoring and predicting the health state of a battery with multi-modal feature fusion, including: S1. Data collection, periodically collect the static data of the battery and collect the dynamic data of the battery in real time to obtain multi-modal data; S2. Data monitoring, collect and monitor the obtained multi-modal data; S3. Shared underlying feature extraction, perform feature processing on the multi-modal data based on the shared underlying feature extraction model to obtain shared underlying features; S4. SOH prediction, after performing feature extraction and fusion on the obtained shared underlying features based on the SOH prediction model, output the SOH prediction result; S5. RUL prediction, after performing feature extraction and fusion on the obtained shared underlying features based on the RUL prediction model, output the RUL prediction result; S6. Multi-objective joint optimization, calculate the SOH loss and the RUL loss respectively and design a weighted multi-objective loss function to obtain the total loss and optimize the prediction error.

[0025] Please refer to Figure 2 , correspondingly, this embodiment provides a system for monitoring and predicting the health state of a battery with multi-modal feature fusion for executing the above method, which mainly consists of the following units: a data collection unit 1 for periodically collecting the static data of the battery and collecting the dynamic data of the battery in real time to obtain multi-modal data; a data monitoring unit 2 for collecting and monitoring the obtained multi-modal data; a shared underlying feature extraction unit 3 for performing feature processing on the multi-modal data based on the shared underlying feature extraction model to obtain shared underlying features; an SOH prediction unit 4 for outputting the SOH prediction result after performing feature extraction and fusion on the obtained shared underlying features based on the SOH prediction model; an RUL prediction unit 5 for outputting the RUL prediction result after performing feature extraction and fusion on the obtained shared underlying features based on the RUL prediction model; a multi-objective joint optimization unit 6 for calculating the SOH loss and the RUL loss respectively and designing a weighted multi-objective loss function to obtain the total loss and optimize the prediction error.

[0026] The battery health status monitoring and prediction method and system with multi-modal feature fusion in this embodiment can perform online monitoring and analysis on the multi-modal feature data generated by the battery. At the same time, a multi-modal feature fusion and multi-objective joint prediction and optimization mechanism is introduced. Based on the SOH prediction model and the RUL prediction model, independent branches are used to predict the SOH (health status) and RUL (remaining useful life) respectively, so as to predict the battery health status. It can not only automatically analyze the multi-modal massive data of the battery, reduce the workload, improve the analysis efficiency, but also improve the accuracy of the prediction results. Moreover, based on the shared underlying feature extraction model, the multi-modal data is characterized to obtain shared underlying features, and a feature sharing mechanism is constructed, which can reduce redundant calculations, reduce the computational complexity, and improve the prediction efficiency. Embodiment 2

[0027] Please refer to Figure 3 , on the basis of Embodiment 1, this embodiment gives a more preferable battery health status monitoring and prediction system with multi-modal feature fusion, which will be specifically described below: This embodiment takes two battery groups as an example, namely the first battery group 71 and the second battery group 72. Each battery group includes multiple single batteries. The data acquisition unit 1 used in this embodiment includes multiple acquisition modules 11 and charge and discharge modules 12. The battery state data (including dynamic parameters, dynamic parameters) is obtained through the acquisition modules. Specifically, the first battery group 71 periodically acquires the static data of the first battery group 71 and real-time acquires the dynamic data of the first battery group 71 through the corresponding acquisition module 11 to obtain multi-modal data; the second battery group 72 also periodically acquires the static data of the second battery group 72 and real-time acquires the dynamic data of the second battery group 72 through the corresponding acquisition module 11 to obtain multi-modal data. As Figure 3 shown, in one embodiment, 10 single batteries share one acquisition module. In a group of batteries, for 108 single batteries, 11 acquisition modules are required. Of course, in some other implementation schemes, actually according to the size of the battery storage cabinet, a certain number of single batteries are stored and are responsible by one acquisition module. For example, 20 batteries are placed in the same cabinet and one acquisition module is used.

[0028] Each acquisition module 11 is connected to the charge and discharge module 12. The charge and discharge module 12 is used to control the charge and discharge of the single battery groups in each battery group according to the acquisition instructions of the acquisition module 11, so as to obtain the battery state data (including dynamic parameters, dynamic parameters) of each single battery, including static data and dynamic data.

[0029] In this embodiment, the static data mainly includes: electrolyte concentration and self-discharge rate. The dynamic data mainly includes: single battery voltage of the battery, charge and discharge current, ambient temperature, internal resistance parameter, and charge and discharge cycle times.

[0030] In this embodiment, the data monitoring unit 2 adopts a battery pack monitoring host 21, which is used to collect and monitor the multimodal data obtained by each acquisition module 11. In addition, the energy storage cabinet / battery cabinet storing the battery pack is monitored online 24 hours a day through a bus-powered sensor to avoid battery power loss.

[0031] This embodiment shares the shared underlying feature extraction model adopted by the shared underlying feature extraction unit 3, preferably using LSTM, SAE or attention network. Preferably, a multi-layer LSTM network is used to extract the temporal degradation features in the battery charge and discharge data. The temporal degradation features are part of the shared underlying features and are manifested as the change of data within a certain period of time. The shared underlying features mainly consist of the shared underlying features derived from static data and the shared underlying features derived from dynamic data. The shared underlying features are manifested as charge and discharge cycle features. Through the shared underlying feature mechanism, it is convenient to compare the actual data with the predicted data to evaluate the accuracy of the model. Please refer to Figure 4 , in the SOH prediction unit 4, cross-cycle feature encoding is performed on the shared underlying features derived from static data to extract trend features including long-term aging features, and intra-cycle encoding is performed on the shared underlying features derived from dynamic data to extract temporal dynamic features including abnormal features. The cross-cycle feature encoding mainly obtains some trend features, that is, the data continuously becomes smaller or larger, etc. It can involve multiple charge and discharge cycles of multiple batteries and mainly analyzes the change law of data from a macroscopic perspective. The intra-cycle feature encoding mainly obtains some abnormal features, such as selecting the highest voltage value and the highest temperature to analyze whether they are abnormal. It involves one charge and discharge cycle of one battery (fully charged and consumed = 1 charge and discharge cycle) and mainly analyzes the data from a microscopic perspective. More specifically, the SOH prediction model based on this embodiment mainly consists of the following modules: a multi-layer perceptron encoder, a transposed convolution encoder, a splicing layer, and a fully connected layer. For feature processing, in this embodiment, the multi-layer perceptron encoder (MLP Encoder) is used to perform cross-cycle feature encoding on the shared underlying features derived from static data to obtain cross-cycle feature vectors. Through the transposed convolution encoder (TransConv Encoder), that is, the intra-cycle feature encoder, intra-cycle encoding is performed on the shared underlying features derived from dynamic data to obtain intra-cycle feature vectors.

[0032] The multi-layer perceptron (MLP) in the multi-layer perceptron encoder (MLP Encoder) performs non-linear mapping on static data / parameters (such as battery structure parameters, number of cycles, ambient temperature, etc.) to extract long-term aging features. The multi-layer perceptron encoder internally integrates fuzzy inference rules to convert expert experience into weight parameters and enhance the interpretability of the model.

[0033] The Transposed Convolution Encoder mainly includes a Transformer layer, a convolutional layer, and a generalization layer. Transformer layer: It has two deep learning models collectively called the Transformer model, which captures long-term temporal dependencies during the charge and discharge process through the self-attention mechanism, such as voltage fluctuations and current spikes. Convolutional layer and generalization layer: Extract local temporal features, such as short-term internal resistance changes and temperature gradients, suppress noise interference, and enhance the generality of the model.

[0034] For the convenience of classification, the cross-cycle and intra-cycle feature vectors output by the multi-layer perceptron encoder and the transposed convolution encoder are flattened into a single column, converting the multi-dimensional time series signal into a one-dimensional vector to adapt to downstream feature fusion. Concatenation layer: The flattened cross-cycle and intra-cycle feature vectors are concatenated through the concatenation layer, and a dynamic weight allocation mechanism is introduced to adjust the contribution degree of each feature vector according to the battery type and usage scenario. Different batteries require different feature processing, and usage scenarios include extreme environments such as fast charging, high and low temperatures. In a high-temperature environment, the weight of the internal resistance is increased. Also, the weight coefficients of each feature are different for different types of batteries. For example, ternary batteries may decay faster at high temperatures, while lithium iron phosphate batteries are more sensitive to over-discharge.

[0035] Fully connected layer: Achieve feature dimensionality reduction and non-linear combination through the fully connected layer.

[0036] In an optional implementation, the fully connected layer outputs an estimated value of the capacity attenuation rate or an estimated value of the internal resistance change rate, and calculates the SOH estimated value based on the estimated value of the capacity attenuation rate or the estimated value of the internal resistance change rate as the SOH prediction result. There are two methods for calculating the SOH estimated value. One is based on the estimated value of the capacity attenuation rate, and the other is based on the internal resistance change rate.

[0037] (1) Based on the capacity attenuation rate, the formula for the capacity attenuation rate is as follows:

[0038] Then, use the capacity attenuation rate to calculate the SOH estimated value, and the formula is as follows:

[0039] (2) Based on the internal resistance change rate, the formula for the internal resistance change rate is as follows:

[0040] Then, use the internal resistance change rate to calculate the SOH estimated value, and the formula is as follows: The SOH prediction model adopted in this embodiment first performs non - linear transformation on cross - cycle static parameters through a multi - layer perceptron encoder to extract long - term aging features, which can achieve deep mapping of static features, improve the characterization ability of static features, and greatly reduce the SOH estimation error. Secondly, through the transposed convolution encoder, it captures local mutations in the short - term charge - discharge process and suppresses high - frequency noise, enabling the extraction of time - series dynamic features. At the same time, combined with the Transformer self - attention mechanism, it captures long - cycle degradation patterns such as voltage fluctuations to achieve long - time - series dependence modeling, enhancing the dynamic feature modeling ability and enabling earlier identification of early degradation. Finally, through the splicing layer and the fully - connected layer, the static and dynamic features are spliced, combining long - term aging and short - term dynamic information to achieve the fusion of multi - modal features, improving the comprehensive characterization ability of the fused features and reducing the collaborative error of SOH and RUL predictions.

[0041] Please refer to Figure 5 , in the RUL prediction unit 5, the RUL prediction model based on this embodiment mainly consists of the following modules: Feature extraction module (left side) Composition: Residual Connection, LayerNorm, and FeedForward neural network.

[0042] Input: Shared underlying features including parameters such as voltage, current, and temperature.

[0043] Processing flow: ① Input features → LayerNorm → FeedForward neural network; ② Original input + FeedForward output → Residual Connection; ③ Repeat the above steps (double - layer structure in the figure).

[0044] Output: High - dimensional representation of degradation features , to characterize the global trend of battery performance degradation, that is, to retain the long - term degradation trend, such as linear capacity decay, and prevent the attention mechanism from over - focusing on short - term fluctuations. Mathematical expression:

[0045] Among them, is the output feature (high - dimensional representation of degradation features), is the input feature (shared underlying feature, including data such as voltage, current, and temperature), FeedForward is the fully - connected layer, and LayerNorm is the layer normalization operation.

[0046] Multi - head attention module (middle) Composition: Parallel attention heads (the "multi - head attention mechanism" in the figure), matrix multiplication operation ("matrix multiplication").

[0047] Input: Output of the feature extraction module .

[0048] Processing flow: ① Split H out into Q / K / V vectors; ② Multiple attention heads perform parallel calculations, that is, capture the non-linear coupling relationship between multi-source features (voltage / current / temperature, etc.) through parallel subspace attention calculation, and obtain their dynamic correlation.

[0049] Output: Features capturing the non-linear coupling relationship, as the output of multi-head attention .

[0050] Key function: Identify key degradation factors, such as abnormal internal resistance, enhance the sensitivity of the model to mutation conditions such as current spikes and temperature sudden changes, and avoid ignoring local anomalies.

[0051] Self-supervised learning module (middle and lower part) Composition: Auxiliary prediction head ("self-supervised" in the figure); Temporal prediction task branch. Design an auxiliary task based on temporal prediction to jointly optimize the feature extraction and the main task of RUL prediction. By constructing an auxiliary learning objective related to the main task (RUL prediction), force the model to mine more robust temporal degradation features from the original data. Use unlabeled data (or data in the same batch) for self-supervised training to reduce the dependence on labeled data. The auxiliary task is: Predict the change trend of battery dynamic parameters (voltage, current, temperature, etc.) in future time steps. The auxiliary task is essentially a self-supervised strategy. By designing a simple temporal prediction objective, such as predicting future battery parameters, guide the model to learn a more discriminative feature representation for the main task (RUL prediction). This design improves the robustness of the model in noisy environments and small sample scenarios while reducing the annotation cost.

[0052] Input: Intermediate output of the feature extraction module.

[0053] Processing flow: ① Intercept intermediate features from the feature extraction layer; ② Predict features in future time steps, such as voltage change trend; ③ Calculate the auxiliary loss, which is achieved by calculating the mean square error (MSE) loss. The specific loss function is as follows:

[0054] where is the loss value, t is the time step index, is the true value of RUL, is the predicted value of RUL, Indicates summing over the time steps that are valid only for masked tokens to avoid interference from invalid data in model training.

[0055] Output: The prediction result of the auxiliary task.

[0056] RUL prediction output module (right side), Composition: Feature fusion layer ("concatenation"), linear regression layer ("linear regression"), matrix multiplication fusion ("matrix multiplication").

[0057] Input: The output of feature extraction and the output of multi-head attention.

[0058] Processing flow: ① Matrix fusion; ② Linear regression.

[0059] Output: The predicted remaining useful life value, that is, the RUL prediction value. The RUL prediction value is output through the linear layer and matrix fusion operation, and the mathematical expression is as follows:

[0060] Among them, RUL is the RUL prediction value output by the RUL prediction model, that is, the predicted value of the remaining useful life, and are the parameters of the linear regression layer, representing the weight matrix and the bias term respectively, which are used to realize the linear mapping from the feature vector to the RUL value and compensate for the baseline offset of the prediction result, is the degradation feature fusion matrix, including the global degradation feature (the output of the residual network ) and the local correlation feature (the output of the attention ), is flattening, which compresses multi-dimensional features into a one-dimensional vector.

[0061] In a preferred embodiment, the remaining useful life (RUL) and the cell consistency index are predicted synchronously. The calculation formula of the consistency index is as follows:

[0062] Among them, is the standard deviation, and mean is the average value. Standard deviation (σ): It refers to the standard deviation of the measured values of the voltages of all individual cells in the battery pack at the same moment. Average value (mean): It refers to the arithmetic average of the voltages of all individual cells at the same moment.

[0063] The difference in the voltages of individual cells in the battery pack directly reflects the imbalance in capacity, internal resistance, and aging degree among the cells. Selecting the voltage of an individual cell directly reflects the inconsistency. By calculating the consistency index, the following purposes can be further achieved: 1. Early warning of abnormal cells: Identify cells with high / low voltage through the voltage standard deviation, and locate potential faulty batteries, such as micro-short circuits and accelerated aging.

[0064] 2. Evaluate the health state of the battery pack: Poor voltage consistency will accelerate the overall degradation. The deterioration of the indicator indicates a decrease in the battery pack life, that is, the RUL is shortened.

[0065] 3. Optimize the charge and discharge strategy: Dynamically adjust the equalization strategy, such as supplementing the charging of low-voltage single cells to delay the deterioration of consistency.

[0066] 4. Support the joint prediction of RUL: Provide the key input of the "barrel effect" for the RUL prediction model, that is, the weakest single cell determines the overall life.

[0067] For the RUL prediction model adopted in this embodiment, first, a feature extraction module with a multi-layer feedforward network + residual connection is used. The residual connection is used to alleviate the problem of gradient disappearance, enhance the network depth, and realize the extraction of complex high-dimensional degradation features during the battery degradation process, so as to improve the degradation feature representation ability and reduce the RUL prediction error. Secondly, a multi-head attention module is used to capture the dynamic correlation between multi-source features through parallel subspace attention calculation and identify key degradation driving factors. On the one hand, it can realize the capture of non-linear coupling relationships, and on the other hand, through the adaptive allocation of attention weights, it can suppress the interference of irrelevant noise and improve the anti-noise ability, thereby improving the prediction stability and reducing the false alarm rate of the present invention under complex working conditions. Then, a self-supervised learning module is used to jointly optimize the feature extraction and the main RUL prediction task by designing an auxiliary task based on time series prediction, enhance the generalization ability of the model to the time series degradation law, pre-train with unlabeled data, reduce the dependence on labeled data, improve the data utilization rate, and improve the prediction accuracy of the model in small sample scenarios. Finally, combined with the RUL prediction output module of a linear layer + matrix fusion, the feature extraction and attention output are integrated through matrix fusion, retaining global and local degradation information, and the linear layer realizes efficient regression to meet the real-time requirements, greatly shortening the prediction response time and realizing online real-time monitoring.

[0068] Generally speaking, this embodiment uses a multi-layer LSTM network to extract the time series degradation features in the battery state data (including dynamic parameters, dynamic parameters), and then on the basis of shared features, the SOH branch outputs the capacity attenuation rate through a fully connected layer, and the RUL branch combines the degradation trend to predict the remaining cycle times.

[0069] In the multi-objective joint optimization unit 6, the weighted multi-objective loss function in this embodiment is expressed as:

[0070] Among them, is the total loss, is the SOH loss, is the RUL loss, is the weight.

[0071] The weighted multi-objective loss function adopts a dynamically weighted multi-objective loss function, where the dynamic weights are adjusted according to the task priorities. In this embodiment, during the multi-objective joint optimization process, a weighted loss function is designed to dynamically adjust the weight parameters according to the prediction errors, giving priority to optimizing critical tasks. For example, when the SOH is abnormal, its weight is increased to adapt to the task priority management, avoid multi-task conflicts, and enhance the stability and accuracy of the prediction.

[0072] Those of ordinary skill in the art can understand that all or part of the processes in the methods of the above embodiments can be completed by instructing relevant hardware through a computer program. The program can be stored in a computer-readable storage medium. When the program is executed, it can include the processes of the above embodiments of each method. Among them, the storage medium can be a magnetic disk, an optical disk, a read-only memory (ROM), or a random access memory (RAM), etc.

[0073] The technical features of the above embodiments can be combined arbitrarily. For the sake of concise description, not all possible combinations of the technical features in the above embodiments are described. However, as long as there is no contradiction in the combination of these technical features, it should be considered within the scope described in this specification. Moreover, the above embodiments only represent several implementation manners of the present invention, and their descriptions are relatively specific and detailed, but they should not be construed as limiting the scope of the present invention. For those of ordinary skill in the art, without departing from the concept of the present invention, several modifications and improvements can still be made, and these all belong to the protection scope of the present invention.

Claims

1. A method for monitoring and predicting the health state of a storage battery with multi-modal feature fusion, characterized in that, Including: Data acquisition, periodically acquiring the static data of the battery and real-time acquiring the dynamic data of the battery to obtain multimodal data; Data monitoring, aggregating and monitoring the acquired multimodal data; Shared underlying feature extraction, performing feature processing on the multimodal data based on the shared underlying feature extraction model to obtain shared underlying features; SOH prediction, after performing feature extraction and fusion on the obtained shared underlying features based on the SOH prediction model, outputting the SOH prediction result; RUL prediction, after performing feature extraction and fusion on the obtained shared underlying features based on the RUL prediction model, outputting the RUL prediction result; Multi-objective joint optimization, calculating the SOH loss and the RUL loss respectively and designing a weighted multi-objective loss function to obtain the total loss and optimize the prediction error.

2. The method for monitoring and predicting the state of health of a storage battery with multi-modal feature fusion according to claim 1, characterized in that The static data includes electrolyte concentration and self-discharge rate; the dynamic data includes battery cell voltage, charge and discharge current, ambient temperature, internal resistance parameter and charge and discharge cycle times.

3. The method for monitoring and predicting the state of health of a storage battery with multimodal feature fusion according to claim 1, wherein The shared underlying feature extraction model adopts LSTM, SAE or attention network.

4. The method for monitoring and predicting the health state of a storage battery with multimodal feature fusion according to claim 1, wherein The shared underlying features include the shared underlying features derived from static data and the shared underlying features derived from dynamic data; in the SOH prediction step, cross-cycle feature encoding is performed on the shared underlying features derived from static data to extract trend features including long-term aging features, and in-cycle encoding is performed on the shared underlying features derived from dynamic data to extract temporal dynamic features including abnormal features.

5. The method for monitoring and predicting the health state of a storage battery with multimodal feature fusion according to claim 4, characterized in that, The SOH prediction model includes a multi-layer perceptron encoder, a transposed convolution encoder, a splicing layer and a fully connected layer; cross-cycle feature encoding is performed on the shared underlying features derived from static data through the multi-layer perceptron encoder to obtain cross-cycle feature vectors; in-cycle encoding is performed on the shared underlying features derived from dynamic data through the transposed convolution encoder to obtain in-cycle feature vectors; the cross-cycle and in-cycle feature vectors are spliced through the splicing layer, and a dynamic weight allocation mechanism is introduced to adjust the contribution degree of each feature vector according to the battery type and usage scenario, and then output through the fully connected layer.

6. The method for monitoring and predicting the state of health of a storage battery by multimodal feature fusion according to claim 5, characterized in that, The fully connected layer outputs an estimated value of the capacity attenuation rate or an estimated value of the internal resistance change rate, and calculates the SOH estimated value according to the estimated value of the capacity attenuation rate or the estimated value of the internal resistance change rate as the SOH prediction result.

7. The method for monitoring and predicting the state of health of a storage battery by multi-modal feature fusion according to claim 1, characterized in that, The RUL prediction model based on includes: A feature extraction module, based on a multi-layer feedforward network and a residual connection structure, performing feature extraction on the shared underlying features to extract a high-dimensional representation of the degradation features; A multi-head attention module, capturing the non-linear coupling relationship between multi-source features through parallel subspace attention calculation; A self-supervised learning module, designing an auxiliary task based on time series prediction to jointly optimize the feature extraction and the RUL prediction main task; An RUL prediction output module, outputting the RUL prediction value through a linear layer and a matrix fusion operation.

8. The method for monitoring and predicting the health state of a storage battery with multimodal feature fusion according to claim 1, characterized in that In the multi-objective joint optimization step, the weighted multi-objective loss function is expressed as: ； wherein, is the total loss, is the SOH loss, is the RUL loss, is the weight.

9. The method for monitoring and predicting the health state of a storage battery with multi-modal feature fusion according to claim 1, characterized in that The weighted multi-objective loss function adopts a dynamically weighted multi-objective loss function, and the is a dynamic weight, which is adjusted according to the task priority.

10. A battery health state monitoring and prediction system with multimodal feature fusion, which is used to execute the method described in any one of the above claims 1 to 9, is characterized in that, Including: A data acquisition unit, used for periodically acquiring the static data of the battery and real-time acquiring the dynamic data of the battery to obtain multimodal data; The data monitoring unit is used to collect and monitor the acquired multi-modal data; The shared underlying feature extraction unit is used to perform feature processing on the multi-modal data based on the shared underlying feature extraction model to obtain shared underlying features; The SOH prediction unit is used to perform feature extraction and fusion on the obtained shared underlying features based on the SOH prediction model, and then output the SOH prediction result; The RUL prediction unit is used to perform feature extraction and fusion on the obtained shared underlying features based on the RUL prediction model, and then output the RUL prediction result; The multi-objective joint optimization unit is used to calculate the SOH loss and the RUL loss respectively and design a weighted multi-objective loss function to obtain the total loss and optimize the prediction error.

Images