Lithium ion battery health state prediction system and method based on multi-scale circulating voltage data

Abstract

The invention discloses a lithium ion battery health state prediction system and method based on multi-scale circulating voltage data. The system comprises a data acquisition module, a multi-scale feature extraction module, a graph structure sample construction module, a multi-scale prediction modeling module and a prediction result output module. The method comprises the following steps: collecting circulating voltage data of a battery in a charging and discharging process, extracting multi-scale health indexes closely related to the health state of the battery, and constructing dynamically connected graph structure data to reflect degradation conditions of the battery in different circulating periods. The system carries out training and prediction by using a multi-scale image convolution network, and fuses the characteristics of different time scales through an adaptive weighting mechanism, thereby enhancing the capability of capturing the change of the health state of the battery, and remarkably improving the prediction precision. The system can accurately predict the decline trend of the battery capacity and the future health state, and has excellent cross-material system universality. According to the method, the prediction error on a public battery data set is smaller than 0.7%, excellent prediction precision and wide application applicability are shown, the method can be used for a battery management system, and real-time monitoring and early warning of the health state of the battery are achieved.

Description

Technical Field

[0001] The present invention relates to the field of battery management technology, and in particular to a lithium-ion battery health status prediction system based on multi-scale cycle voltage data and an implementation method thereof. The system and method can realize the health status prediction of lithium-ion batteries, especially the prediction of battery capacity decay, by performing multi-scale feature extraction and modeling only on the cycle voltage data of the battery. Background Art

[0002] Lithium-ion batteries are widely used in consumer electronic devices, energy storage systems, and electric vehicles, but their performance gradually degrades over time, affecting the stability and safety of the equipment. Therefore, battery state of health (SoH) prediction is crucial for battery life management. Existing SoH prediction methods usually rely on multiple data types of batteries (such as voltage, current, temperature, resistance, etc.), but this increases the complexity of data acquisition and processing. In addition, these methods are relatively limited in extracting battery degradation features, lack accurate modeling capabilities, and cannot cope well with complex working conditions in practical applications.

[0003] In recent years, data-driven prediction methods (such as machine learning and deep learning) have been widely used. However, existing methods mainly rely on multimodal data and do not fully utilize the degradation information in the single data source of battery cycle voltage. In addition, single-resolution feature extraction and modeling may ignore the degradation trends on different time scales, affecting the accuracy of prediction.

[0004] To solve the above problems, the present invention proposes a lithium-ion battery health status prediction system and method based on multi-scale cycle voltage data. By extracting multi-scale features from the battery's charge and discharge voltage data and constructing a multi-scale graph convolutional network for health status prediction, the accuracy and generalization of prediction can be improved while reducing the complexity of data collection. Summary of the invention

[0005] The present invention proposes a lithium-ion battery health status prediction system and method based on multi-scale cycle voltage data, aiming to overcome the problems of the prior art in terms of strong data dependence, high complexity and insufficient prediction accuracy. By extracting multi-scale resolution features from the voltage data collected during the battery charging and discharging process, combined with the modeling of graph structure data and the prediction model of the multi-scale graph convolutional network, a high-precision prediction of the battery's future health status and capacity decay can be achieved.

[0006] 1. Technical solution overview

[0007] The core of the present invention is to extract health indicators that are highly correlated with battery life by performing multi-scale analysis on the charge and discharge cycle voltage data of different types of lithium-ion batteries, and construct these health indicators into graph structure data, and further predict the health status through a multi-scale graph convolutional network. The system achieves efficient and accurate battery health status prediction through the collaborative work of feature extraction module, graph structure construction module, prediction model module and output module. Specifically, the system adopts three key technologies:

[0008] 1) Multi-scale feature extraction technology: Extract health indicators at different time resolutions from the battery’s charge and discharge voltage curves to capture short-term fluctuations and long-term trends during battery degradation, ensuring data integrity and feature richness;

[0009] 2) Graph structure data modeling technology: The extracted multi-scale health indicators are expressed in the form of a graph structure. The nodes represent the specific cycle of the battery, and the edges are dynamically connected according to the similarity between the voltage data, ensuring that the graph structure data can accurately reflect the time dependency and degradation trend of the battery;

[0010] 3) Multi-scale graph convolution prediction technology: By convolving graph structure data at multiple time scales, the multi-dimensional characteristics of the battery health status are captured, and the feature weights of each time scale are adaptively adjusted through a dynamic feature fusion mechanism, thereby improving the prediction accuracy.

[0011] 2. System module design

[0012] 1) Data acquisition module: This module is used to collect voltage data in real time during the charging and discharging process of lithium-ion batteries. The collected voltage data can be completed through a conventional battery management system or acquired through a dedicated voltage monitoring device. The system can record the voltage changes of the battery in each charge and discharge cycle to form a complete voltage time series data;

[0013] 2) Multi-scale feature extraction module: The multi-scale feature extraction module of the present invention performs multi-resolution sampling on the collected voltage data and extracts characteristic parameters related to the battery health status. The specific operations include:

[0014] a) Multi-resolution sampling unit: The system samples voltage data at multiple time scales, such as 1 cycle, 2 cycles, and 4 cycles, to ensure that the battery degradation process can be captured at different time resolutions;

[0015] b) Data standardization unit: De-noise the collected voltage data to eliminate noise and outliers in the data to ensure the purity and consistency of the data. Specifically, a cubic smoothing filter is used to reduce voltage fluctuation noise. The filtering formula is:

[0016]

[0017] Among them, V i Represents the voltage value during the i-th charge and discharge process. Apply linear interpolation to the original voltage sequence to fill in the missing values ​​caused by sensor noise or data loss. The formula is:

[0018]

[0019] This formula is used to estimate the i-th missing value for the data point.

[0020] c) Health feature calculation unit: This unit extracts the standard deviation, average voltage and cycle duration of each battery cycle through numerical calculation as the time domain characteristics reflecting the performance changes of battery materials and electrolytes, as well as the indicators of the rate changes of the internal reactions of the battery; This unit analyzes the abnormal inflection points in the voltage curve through Z-score transformation, extracts the mutation characteristics reflecting battery degradation, and identifies the voltage abnormal points and degradation trends in the cycle. The formula is:

[0021]

[0022] Among them, σ V is the standard deviation of the voltage series. The voltage mutation point is identified by performing two consecutive Z-score transformation analyses on the voltage curve; this unit applies wavelet transform to perform multi-scale decomposition on the voltage data, extracts fine-grained information in the voltage signal, and captures the high-frequency and low-frequency fluctuation characteristics in the battery degradation process. Specifically, Daubechies wavelet (db4) is used for multi-scale decomposition, and its decomposition formula is:

[0023]

[0024] Among them, a jtk is the approximate coefficient, is the scale function, d jtk is the detail coefficient, ψ jtk (t) is a wavelet function, which extracts detail information and approximate information at different scales by decomposing the original voltage signal into 5 layers. j The change of (t) can identify the characteristics of voltage fluctuation during battery degradation.

[0025] 3) Graph structure sample construction module: This module converts the health indicator sequence obtained from the multi-scale feature extraction module into graph structure data, thereby providing input for the training and prediction of the multi-scale graph convolutional network. This module constructs graph structure data to represent the complex relationships in the time series and the dynamic similarity between nodes as a graphical data structure, capturing the timing dependency and spatial characteristics of the battery. Specific operations include:

[0026] a) Node generation unit: This unit is responsible for treating each battery charge and discharge cycle as a node in the graph. Each node contains multi-scale health features (such as standard deviation, average voltage, Z-score inflection point, wavelet features, etc.) extracted from the voltage data of the corresponding cycle. These node features are the input data for the model to make predictions, representing the health status information of the battery during the cycle. Node generation can be performed for different time resolutions, such as 1 cycle, 2 cycles, 4 cycles, etc., and a corresponding number of nodes are generated for each resolution. These nodes are arranged in sequence to form a time series path, which can be used for graph convolutional network operations.

[0027] b) Dynamic node connection unit: After generating nodes, it is necessary to establish connection relationships between nodes. Here, a dynamic node connection mechanism is introduced, which is mainly based on the similarity of the cosine characteristics of each node for dynamic connection. The cosine similarity is used to reflect the similarity of the voltage characteristics of two cycles. The higher the value, the more similar the health status between the cycles. For node n a and n b , the calculation formula is:

[0028]

[0029] Among them, X a and X b Represents node n a and n b When establishing connections, edges with similarity below the set threshold are deleted to ensure graph sparsity and computational efficiency. Usually 70% to 80% of high-similarity connections are retained, which can reduce the interference of irrelevant information and retain important health status change information.

[0030] c) Graph construction optimization unit: Since there may be data missing or abnormal data collection during actual battery operation and data collection, the graph construction optimization unit is responsible for handling these problems to ensure the integrity and continuity of graph structure data. For missing data, interpolation methods are usually used to complete it, for example: linear interpolation of the values ​​of adjacent nodes to fill in the features of missing nodes; when there are few data points or the changes are drastic, spline interpolation methods are used for more precise interpolation; combined with machine learning or deep learning methods to predict and complete missing data.

[0031] 4) Multi-scale prediction modeling module: This module is the core part of the system. Its main function is to train and predict the constructed graph structure samples through a multi-scale graph convolution network, and use graph structure data with different time resolutions to capture the characteristic changes of the battery at multiple time scales to achieve accurate prediction of the health status. Specific operations include:

[0032] a) Multi-scale graph convolution unit: The multi-scale graph convolution network includes multiple convolution layers, each of which processes graph data at different time resolutions (such as 1 cycle, 2 cycles, and 4 cycles) to capture the degradation trend of the battery in the short and long term. For example, a resolution of 1 cycle may capture the health changes of the battery in the short term, while a resolution of 4 cycles can capture long-term trends. The convolution operation formula for each layer is:

[0033]

[0034] in, is the adjacency matrix after adding self-connection, D is the degree matrix, H (l) is the node feature matrix of the lth layer, W (l) is the trainable weight matrix of the lth layer, and σ is the activation function. Through multiple layers of convolution, the model can accumulate information about the change of battery health status at various time scales.

[0035] b) Dynamic feature fusion unit: This unit uses a gating network similar to the attention mechanism to adjust the importance of features from different time scales by dynamically weighting, ensuring that the model can focus on the most relevant time scale information. The gating scores generated by the fully connected layer are used to weight features of different scales and normalized by the Softmax function to form the final comprehensive feature representation. Its dynamic weighting mechanism is implemented through the following formula:

[0036]

[0037] in, is the weight score of the i-th scale of the l-th layer calculated by the gating network. By weighted summing of features at different scales, the model can adaptively focus on the most relevant time scale, thereby improving prediction accuracy.

[0038] c) Weight normalization unit: To ensure that features of different time scales are not dominated by a single scale during feature fusion, the multi-scale graph convolutional network uses a weight normalization unit to balance the weights of each time scale. The weight normalization unit uses the Softmax function to normalize the weights generated by the gated network to ensure that the proportion of features of each time scale in the final comprehensive features is reasonable.

[0039] d) Comprehensive feature generation unit: After dynamic feature fusion and normalization, the features of multiple time scales are linearly combined to generate a comprehensive feature representation. This comprehensive feature can effectively capture the health change information of the battery at different time scales and provide more accurate state prediction. The output features of this unit will be used as the final prediction basis of the model to generate the battery health state (SoH) and capacity decay curve.

[0040] 5) Prediction result output module: This module is responsible for displaying or outputting the battery health status prediction results. The system generates a prediction curve for future battery capacity decay by analyzing voltage data. The curve represents the capacity change trend of the battery in each charge and discharge cycle. Users can intuitively view the historical changes in battery capacity and the predicted capacity trend in the future. The curve is predicted based on the battery cycle voltage data and reflects the degree of battery degradation. This module displays the current health status of the battery in percentage, that is, the ratio of the current available capacity to the nominal capacity. The prediction module calculates the future health status through historical voltage data, allowing users to know in advance when the battery may reach a critical health state (for example, below 80%).

[0041] 3. Implementation steps

[0042] The present invention also provides a method for predicting the health status of a lithium-ion battery based on the above system, comprising the following steps:

[0043] Through the data acquisition module, the system collects the charging and discharging voltage data of different types of lithium-ion batteries in real time; the voltage data can be stored in the form of time series for subsequent multi-scale feature extraction; the system performs multi-resolution sampling on the collected voltage data and extracts multi-scale health indicators including time domain features, Z-score features and wavelet decomposition features; these features capture different aspects of battery degradation and form a rich feature set for model training; the system constructs the extracted multi-scale health indicator data sequence into a graph structure sample; each node represents a battery cycle, and the connection between nodes is dynamically adjusted according to the similarity between the voltage sequences to ensure that the graph structure can accurately reflect the relationship between different time points in the battery degradation process; through the multi-scale graph convolutional network, the system trains the graph structure samples to capture the complex dynamic characteristics of the battery health status; the convolutional network can process data at different time scales and optimize the feature combination of each time scale through a dynamic feature fusion mechanism; based on voltage data of at least 20 consecutive cycles, the system uses a single-step iterative prediction method to predict the future battery capacity decline trend and output real-time information on the battery health status.

[0044] 4.Technological advantages

[0045] Compared with the prior art, the system of the present invention only relies on the charging and discharging voltage data of the battery, and does not need to collect multimodal data such as current and temperature, which simplifies the data collection and processing process, avoids the need to collect multiple data types, and reduces the cost of sensors and data acquisition systems. On the one hand, the present invention uses multi-scale feature extraction and graph convolution network modeling, dynamic node connection and missing data processing mechanisms to accurately predict the health status of the battery, and the mean absolute percentage error (MAPE) can be controlled within 0.67%. On the other hand, it shows strong predictive ability under a variety of battery materials and working conditions, and can be applied to a variety of lithium-ion batteries such as NCA and NCM types. BRIEF DESCRIPTION OF THE DRAWINGS

[0046] Figure 1 It is a schematic diagram of the overall structure of the system of the present invention;

[0047] Figure 2 A schematic diagram of multi-scale feature extraction provided by an embodiment of the present invention;

[0048] Figure 3 A schematic diagram of a graph structure sample provided by an embodiment of the present invention;

[0049] Figure 4 A schematic diagram of the graph convolutional network structure provided by an embodiment of the present invention;

[0050] Figure 5 A schematic diagram of outputting prediction results provided by an embodiment of the present invention. DETAILED DESCRIPTION

[0051] In order to further understand the purpose, technical solutions and advantages of the present invention more clearly, the following embodiments will be described in detail with reference to the accompanying drawings:

[0052] Embodiment 1:

[0053] A lithium-ion battery health status prediction system based on multi-scale cycle voltage data, such as Figure 1 As shown, it includes: a data acquisition module, a multi-scale feature extraction module, a graph structure sample construction module, a multi-scale prediction modeling module and a prediction result output module.

[0054] The data acquisition module is used to obtain the cycle voltage data of the lithium-ion battery during the charge and discharge process. In this embodiment, the data in the lithium-ion battery cycle life data set released by NASA is used, and the data set includes the cycle voltage data of various batteries under different working conditions. The batteries in the data set undergo hundreds of charge and discharge cycles, and the charge and discharge voltage is recorded in each cycle. These cycle voltage data are used as the output results of the simulation data acquisition module in this embodiment. Batteries #5, #6, and #7 were selected for testing, where batteries #5 and #7 were used for training, and battery #6 was used for verification.

[0055] The multi-scale feature extraction module extracts health indicators that are highly correlated with the battery health status by sampling the above-mentioned raw cycle voltage data at multi-scale resolutions. The module includes a multi-resolution sampling unit, a data normalization unit, and a health feature calculation unit. The multi-resolution sampling unit uses different time resolutions (such as 1 cycle, 2 cycles, and 4 cycles) to capture coarse-grained and fine-grained features in the battery degradation process. To ensure the integrity of the data, the system denoises and interpolates the collected voltage data to eliminate missing values ​​caused by sensor noise and data loss. Figure 2 As shown in (a), direct indicators extracted from the raw voltage data include standard deviation (STD), average voltage (AV) and cycle time (CD). Figure 2 As shown in (b), this module extracts the Z-score index based on statistical methods. Through standardization, the Z-score can identify three abnormal inflection points in the battery cycle voltage curve and evaluate the degree of battery degradation. Furthermore, this module uses wavelet decomposition technology to perform time-frequency analysis on the voltage signal to capture subtle changes in the battery at different time scales. Figure 2 As shown in (c), by analyzing the detail components after multi-scale wavelet decomposition, the abnormal inflection points of the voltage curve in the frequency domain are extracted.

[0056] The graph structure sample construction module constructs graph structure data using the extracted multi-scale health indicator sequence. The module includes: node generation unit, dynamic node connection unit and graph construction optimization unit. Figure 3 As shown, the node generation unit represents each battery cycle as a node, and the edges between nodes are dynamically connected based on cosine similarity. In order to ensure the sparsity of the graph structure, for each node, the system calculates its feature similarity with other nodes, and deletes low-similarity edges based on the set threshold, retaining only the 80% edges with the highest similarity. In order to solve the problem of missing data that may occur during the sampling process, the system also uses an interpolation method to ensure the integrity and continuity of the graph structure. Figure 3 The graph structures at different resolutions and at different degradation stages are shown, which can effectively reflect the changes in battery health status.

[0057] The multi-scale prediction modeling module predicts the battery health status through a multi-scale graph convolution network. The module includes: a multi-scale graph convolution unit, a dynamic feature fusion unit, a weight normalization unit, and a comprehensive feature generation unit. Figure 4As shown in the figure, the multi-scale graph convolution architecture consists of multiple graph convolution layers, each of which processes graph structure data of different time resolutions. The network can effectively capture the degradation trend of batteries at different time scales by adaptively weighting features of different resolutions. The importance of each time scale feature is dynamically adjusted through the gating mechanism. The system calculates the gating score of each time scale and normalizes these features using the Softmax function. Finally, the linear combination generates a comprehensive feature representation, which significantly improves the prediction accuracy of the model.

[0058] The prediction result output module is used to display and output the prediction results. Figure 5 As shown in (a) and 5(b), the system generates the capacity decay curve of the battery and predicts the future health status. When the battery is close to failure, the system will also issue an alarm to remind the user to take appropriate measures. According to the experimental results, the prediction error of this system on the NASA public data set is less than 0.7%, demonstrating the excellent performance of the system in terms of prediction accuracy and applicability across material systems.

[0059] Embodiment 2:

[0060] The difference from Example 1 is that a lithium-ion battery degradation data set provided by Toyota Research Center is used, and the prediction accuracy of the test samples is above 99%.

[0061] Embodiment 3:

[0062] The difference from Example 1 is that the ternary lithium-ion battery degradation data set provided on the Zenodo website is used, and the prediction accuracy of the test samples is above 99%.

Claims

1. A lithium-ion battery health status prediction system based on multi-scale cycle voltage data, the characteristics of which include: 1) Data acquisition module: used to collect continuous voltage data generated by different types of lithium-ion batteries during the charge and discharge cycle; 2) Multi-scale feature extraction module: used to perform multi-scale resolution preprocessing on the continuous voltage data to extract health indicators that are highly correlated with battery life; 3) Graph structure sample construction module: used to construct multi-scale health indicator sequences into graph structure data, where nodes represent specific cycles of batteries and edges are dynamically connected based on the similarity of node features; 4) Multi-scale predictive modeling module: includes a multi-scale graph convolutional network for training based on multi-resolution graph structure data arrays to predict the health state of the battery, including future capacity decay; 5) Prediction result output module: used to display or output the prediction results, generate the battery capacity decay curve and indicate the health status.

2. The system according to claim 1, wherein the multi-scale feature extraction module further comprises: 1) Multi-resolution sampling unit: three resolutions are set, namely 1 cycle, 2 cycles and 4 cycles, and 20 consecutive cycles of voltage data are sampled; 2) Data standardization unit: pre-process the collected voltage data to eliminate noise and outliers in the data; 3) Health feature calculation unit: Based on statistical analysis, the voltage data including time domain features such as standard deviation and mean are extracted; the Z-score transformation is used to analyze the abnormal inflection points of the voltage data and extract the mutation features; the voltage data is decomposed into multiple scales through wavelet transformation, the frequency domain features are extracted, and the fine-grained voltage fluctuation information is captured.

3. The system according to claim 1, wherein the graph structure sample construction module further comprises: 1) Node generation unit: used to generate nodes in the graph structure according to each battery cycle, and the node features are based on the extracted multi-scale voltage features; 2) Dynamic node connection unit: For each node (representing a cycle of the battery), the cosine similarity between it and other nodes is calculated using the following formula: Among them, X a and X b is the voltage characteristic vector representing different battery cycle periods, n a and n b For different cycles; low similarity connections are deleted according to the set threshold, and only 80% of high similarity node connections are retained; 3) Graph construction optimization unit: used to use the tail elimination mechanism and interpolation method to deal with the data missing problem that may occur during the sampling process, ensuring the integrity and continuity of the graph structure.

4. The system of claim 1, wherein the multi-scale graph convolutional network further comprises: 1) Multi-scale graph convolution unit: It consists of 3 graph convolution layers and has an adaptive feature adjustment mechanism. It can adjust the convolution parameters of the network layer for graph structure data of different time resolutions (1 cycle, 2 cycles, 4 cycles) to capture the degradation trend of the battery at different time scales; 2) Dynamic feature fusion unit: weighted scores are performed on features from different time scales through a gating network, dynamically adjusting the importance of features at each time scale; 3) Weight normalization unit: Generates gated scores through the fully connected layer and normalizes each scale feature through the Softmax function to ensure the balance of features at different scales; 4) Comprehensive feature generation unit: Linearly combine features from multiple time scales according to the normalized dynamic weights to generate a comprehensive feature representation to improve the model's sensitivity to changes in battery health status.

5. The system according to claim 1, wherein the system is applicable to different types of lithium-ion batteries, including but not limited to NCA batteries, NCM batteries, lithium iron phosphate batteries, lithium cobalt oxide batteries, and lithium manganese oxide batteries, etc., and has the prediction versatility across material systems and can accurately predict the battery health status under different working environments.

6. A method for predicting the health status of a lithium-ion battery based on multi-scale cyclic voltage data of the system according to any one of claims 1 to 5, characterized in that: The following steps are involved: 1) Collect continuous battery cycle voltage data; 2) Perform multi-scale sampling on the collected voltage data to generate voltage sequences with different time resolutions; 3) Extract multi-scale health indicators from the preprocessed voltage sequence, including time domain features, Z-score transformation features, and wavelet transformation features; 4) Construct the multi-scale health indicator sequence as a graph structure data sample, where nodes represent cycle periods and edges represent dynamic connections based on cosine similarity; 5) Iteratively training the graph structure data using a multi-scale graph convolutional network to predict the health status of the battery; 6) Based on the voltage data of at least 20 consecutive cycles, a single-step iterative prediction method is used to predict the future battery capacity degradation curve and output the battery health status information.

Images