Artificial intelligence-based battery state of health online monitoring method and system

Abstract

The application provides a kind of battery health state online monitoring method and system based on artificial intelligence, by obtaining the real-time monitoring data set of battery in the real-time running process of battery;The battery real-time monitoring data set is enhanced by feature correlation, and a battery multi-domain correlation feature set is generated;Based on the battery multi-domain correlation feature set, a battery health state correlation graph is constructed;Through the battery health state correlation graph, the battery health state multi-scale evaluation processing is carried out, and the battery health state evaluation result set is obtained, to output the battery health state online monitoring result based on the battery health state evaluation result set.Through the present application, the information of multiple physical quantities and multiple time scales can be integrated, the accuracy and reliability of battery health state online monitoring can be effectively improved, and the problem of one-sided or inaccurate monitoring results caused by feature isolation and evaluation fragmentation in traditional monitoring methods is solved.

Description

Technical Field

[0001] This invention relates to the field of battery monitoring, and more specifically, to an online monitoring method and system for the health status of batteries based on artificial intelligence. Background Technology

[0002] With the widespread application of batteries in new energy and energy storage fields, online monitoring technology for battery health status has received increasing attention. This technology acquires and analyzes battery operating data to assess battery health status in real time and predict performance trends. Currently, it typically involves acquiring operating data such as battery voltage, current, and temperature, using preset thresholds or simple feature engineering to extract statistical features, and then inputting this data into traditional machine learning models for health status assessment. However, this monitoring method, which relies solely on simple dimensional data and shallow feature processing, struggles to effectively capture the complex electrochemical reactions and multi-physics coupling processes within the battery, resulting in insufficient accuracy and reliability of health status assessment results. Therefore, improving the accuracy of online battery health status monitoring has become a pressing technical problem in this field. Summary of the Invention

[0003] This invention provides an online monitoring method and system for the health status of batteries based on artificial intelligence.

[0004] In a first aspect, embodiments of the present invention provide an online monitoring method for the health status of a battery based on artificial intelligence, comprising: acquiring a set of real-time monitoring data of the battery during real-time operation; performing feature association enhancement on the set of real-time monitoring data of the battery to generate a multi-domain association feature set of the battery; constructing a battery health status association map based on the multi-domain association feature set of the battery; performing multi-scale evaluation processing of the battery health status through the battery health status association map to obtain a set of battery health status evaluation results, and outputting online monitoring results of the battery health status based on the set of battery health status evaluation results.

[0005] Secondly, embodiments of the present invention provide a computer system, including: a memory storing a computer program; and a processor for loading the computer program to implement the online monitoring method for battery health status based on artificial intelligence as described above.

[0006] The present invention provides an AI-based online monitoring method for battery health status. After acquiring a real-time monitoring data set of the battery during its operation, the method performs feature association enhancement processing on this data set to generate a multi-domain associated feature set. This processing overcomes the limitations of traditional techniques that extract or simply stitch together features from monitoring data such as voltage, current, and temperature. It actively explores the inherent coupling relationships between different physical quantities, thereby generating a feature set that integrates multi-domain associations. This makes the feature information more closely reflect the actual mechanism of electro-thermal-chemical multi-physics coupling during battery operation, avoiding the one-sidedness of single features. After obtaining the multi-domain associated feature set, a battery health status association graph containing a set of feature nodes and a set of node association weights is constructed based on this set. This graph-based modeling method abstracts multi-domain associated features into feature nodes and quantifies the degree of health impact between features, making the association between features explicit and dynamic, overcoming the shortcomings of traditional techniques that store features in low-dimensional structures such as vectors or matrices, which makes it difficult to model complex dynamic relationships and improving the accuracy of feature representation. Subsequently, the battery health status correlation map is used to perform multi-scale assessment of the battery health status, resulting in a set of battery health status assessment results including short-term health trend assessment results, medium-term performance degradation assessment results, and long-term life prediction assessment results. By using feature nodes and correlation weights in the map to convey the influence of features at different time scales, the short-term, medium-term, and long-term assessments are integrated into an organic whole based on feature correlation. This avoids the problems of traditional single-time-scale assessments or the disconnect between the assessment process and feature correlation analysis, enabling assessment results at different time scales to be correlated and verified. Through the above processing, this method can integrate information from multiple physical quantities and multiple time scales, effectively improving the accuracy and reliability of online battery health status monitoring, and solving the problems of isolated features and fragmented assessments in traditional monitoring methods that lead to one-sided or inaccurate monitoring results. Attached Figure Description

[0007] Figure 1 This is a flowchart of an online monitoring method for the health status of a battery based on artificial intelligence, provided in an embodiment of the present invention.

[0008] Figure 2 This is a schematic diagram of the composition of a computer system provided in an embodiment of the present invention. Detailed Implementation

[0009] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0010] Please see Figure 1 , Figure 1 A flowchart illustrating an online monitoring method for the health status of a battery based on artificial intelligence, provided as an embodiment of the present invention, is shown. This method can be executed by a computer system and includes the following steps:

[0011] Step S100: Obtain the real-time monitoring data set of the battery during real-time operation.

[0012] A real-time battery monitoring dataset is a collection of data reflecting the battery's operating status during real-time operation. This data reflects the battery's actual working condition at any given moment from multiple dimensions. Specifically, this dataset may include, but is not limited to, battery voltage, current, and temperature data. Voltage data reflects the potential difference across the battery terminals and is one of the indicators for measuring the battery's energy storage and release capabilities; different voltage values ​​correspond to different states of charge (SOCs). Current data reflects the magnitude and direction of the current during battery operation; monitoring the current allows us to understand the battery's charging and discharging status, as well as power consumption. Temperature data reflects the battery's heat generation during operation; temperature changes significantly affect battery performance and lifespan; excessively high or low temperatures can lead to performance degradation or even damage.

[0013] This data can be obtained through various sensors installed on the battery, such as voltage sensors, current sensors, temperature sensors, and so on.

[0014] Step S200: Enhance the feature association of the real-time monitoring data set of the battery to generate a multi-domain association feature set of the battery.

[0015] Feature association enhancement involves analyzing and processing the collected real-time battery monitoring data sets to uncover the inherent relationships and interactions between the data, thereby enhancing the expressive power and information content of the data features. Through this processing, previously independent and scattered data features are linked together to form a more comprehensive and representative feature set.

[0016] The multi-domain correlation feature set of a battery is a collection of features obtained after feature correlation enhancement processing. These features cover multiple different domains and dimensions, and can comprehensively and accurately reflect the battery's operating status and health condition. The multi-domain correlation feature set can include correlation features between voltage and current, current and temperature, and voltage and temperature, etc. These features from different domains are interconnected and mutually influential, forming a comprehensive feature system that helps to gain a deeper understanding of the battery's internal working mechanisms and health status.

[0017] As one implementation method, step S200 involves performing feature association enhancement on the real-time monitoring data set of the battery to generate a multi-domain association feature set of the battery. Specifically, this can be implemented as follows: steps S210~S240:

[0018] Step S210: Extract voltage fluctuation data and current change data from the real-time monitoring data set of the battery, synchronize the voltage fluctuation data and current change data in time, and obtain synchronized voltage fluctuation data and synchronized current change data. Input the synchronized voltage fluctuation data and synchronized current change data into a preset coupling feature extraction model. Calculate the correlation degree of the synchronized voltage fluctuation data and synchronized current change data through the time correlation layer of the preset coupling feature extraction model to generate a voltage-current correlation degree sequence. Then, perform feature fusion on the synchronized voltage fluctuation data and synchronized current change data based on the voltage-current correlation degree sequence through the feature fusion layer of the preset coupling feature extraction model to obtain voltage-current coupling features.

[0019] Voltage fluctuation data refers to the change in battery voltage over time during battery operation, reflecting the stability and dynamic characteristics of the battery voltage. Voltage fluctuations may be caused by factors such as the battery's charging and discharging process, load changes, and internal chemical reactions. Current variation data refers to the change in battery current over time, reflecting the dynamic changes in the magnitude and direction of the battery's charging and discharging current.

[0020] Time synchronization refers to aligning voltage fluctuation data and current change data along the time dimension, ensuring that the data from both have a corresponding relationship at the same point in time. Since there may be some time error or delay in the acquisition of voltage and current data, time synchronization can eliminate these errors, making subsequent correlation analysis more accurate. Time synchronization can be achieved using timestamp matching, that is, adding timestamp information to the acquired voltage and current data respectively, and then aligning the two data based on the timestamps. In practice, the acquisition time of one data point can be used as a reference to adjust the time of the other data point, so that the two data points have corresponding values ​​at the same point in time.

[0021] The pre-trained coupling feature extraction model is a pre-trained model used to extract coupling features between voltage and current. This model consists of two main parts: a time-series correlation layer and a feature fusion layer. The time-series correlation layer calculates the correlation between time-synchronized voltage fluctuation data and current change data to determine the degree of correlation and variation patterns between them. The feature fusion layer, based on the voltage-current correlation sequence, fuses the voltage fluctuation data and current change data to generate a comprehensive voltage-current coupling feature that more fully reflects the interaction between voltage and current.

[0022] The voltage-current correlation sequence is a series of correlation values, each representing the degree of correlation between voltage fluctuation data and current change data at a given moment. Analyzing this sequence reveals the changing correlation between voltage and current at different points in time. The voltage-current coupling characteristic is a comprehensive feature obtained by fusing voltage fluctuation data and current change data. It integrates voltage and current information, providing a more accurate reflection of the battery's electrical characteristics and health status during operation.

[0023] As one implementation method, in step S210, the correlation degree of the time-synchronized voltage fluctuation data and the time-synchronized current change data is calculated through the time-series correlation layer of the preset coupling feature extraction model to generate a voltage-current correlation degree sequence. Specifically, this can be implemented as the following steps S211~S214:

[0024] Step S211: The voltage fluctuation data after time synchronization is processed by time series framing through the time series correlation layer of the preset coupling feature extraction model to obtain multiple voltage framing data. The current change data after time synchronization is processed by time series framing to obtain multiple current framing data. The number of frames of the multiple current framing data and the multiple voltage framing data are the same and the frame time corresponds.

[0025] Time series framing is the process of dividing continuous time series data into several fixed-length frames. Framing can convert long-term time series data into multiple short-term subsequences. In this embodiment, the time-synchronized voltage fluctuation data and current change data are separately framed to obtain multiple voltage framed data and current framed data.

[0026] The temporal correlation layer of the pre-defined coupled feature extraction model has a frame-segmentation function, which can perform frame-segmentation on voltage fluctuation data and current change data according to preset frame-segmentation parameters. Frame-segmentation parameters include frame length and overlap rate, which can be preset according to actual needs or experience. The frame length determines the number of data points contained in each frame, while the overlap rate indicates the number of data points overlapping between two adjacent frames.

[0027] Multiple voltage frame data refers to a series of sub-data sequences obtained by framing voltage fluctuation data, with each sub-data sequence corresponding to a time interval. Similarly, multiple current frame data is a series of sub-data sequences obtained by framing current change data. The requirement that the number of frames in the multiple current frame data and the multiple voltage frame data be the same and that the frame times correspond ensures that voltage and current data have a corresponding relationship within the same time interval in subsequent correlation analysis, facilitating accurate calculation of their correlation degree.

[0028] As one implementation method, step S211 involves performing time-series framing processing on the time-synchronized voltage fluctuation data through the time-series correlation layer of a preset coupling feature extraction model to obtain multiple voltage frame data. Specifically, this can be implemented as the following steps S2111~S2113:

[0029] Step S2111: Obtain the time length of the voltage fluctuation data after timing synchronization through the timing correlation layer of the preset coupling feature extraction model, and determine the time interval of each frame based on the time length and the preset number of frames.

[0030] The duration of the voltage fluctuation data after timing synchronization refers to the time range covered by the data, that is, the time span from the start to the end of data acquisition. The preset number of frames is the number of frames into which the voltage fluctuation data is divided. The time interval of each frame refers to the duration of each frame.

[0031] The temporal correlation layer of the pre-defined coupling feature extraction model can obtain the time length by analyzing the timestamp information of voltage fluctuation data. Then, based on the time length and the pre-defined number of frames, the time interval of each frame is determined. For example, the time interval of each frame = time length / pre-defined number of frames.

[0032] Step S2112: Divide the time-synchronized voltage fluctuation data into multiple initial voltage frame data by equal intervals according to the time interval, and perform boundary smoothing processing on each initial voltage frame data to eliminate feature abrupt changes at the frame boundary.

[0033] Equal-interval framing divides the time-synchronized voltage fluctuation data into multiple frames at predetermined time intervals. Each frame contains an equal amount of time, and the time interval between any two adjacent frames is also equal. Equal-interval framing ensures that the data in each frame has the same temporal resolution.

[0034] Boundary smoothing is a data processing operation performed to eliminate abrupt changes in features at frame boundaries. Since framing involves artificially dividing continuous time-series data, discontinuities and abrupt changes in features may occur at frame boundaries, potentially affecting the accuracy of correlation analysis. Boundary smoothing can be achieved using methods such as moving average filtering or Gaussian filtering. Taking moving average filtering as an example, for the boundary portion of each initial voltage frame data, the average value of several adjacent data points is calculated, and this average value is used to replace the original data value at the boundary, thus making the data changes at the boundary smoother.

[0035] Step S2113: Use the smoothed initial voltage frame data as voltage frame data to obtain multiple voltage frame data.

[0036] The initial voltage frame data, after boundary smoothing, has eliminated abrupt changes in features at the frame boundaries, making the data smoother and more continuous. Using this smoothed initial voltage frame data as the final voltage frame data provides a more accurate and reliable data foundation for subsequent correlation analysis.

[0037] Step S212: For each voltage frame data, determine the current frame data corresponding to the voltage frame data from multiple current frame data, and form a frame data pair by combining each voltage frame data and its corresponding current frame data to obtain multiple frame data pairs.

[0038] Since multiple current frame data points and multiple voltage frame data points have the same number of frames and corresponding frame times, the corresponding current frame data points can be found for each voltage frame data point based on the time correspondence. Specifically, for each voltage frame data point, the corresponding current frame data points are current frame data points collected within the same time interval.

[0039] Combining each voltage frame data point with its corresponding current frame data point into frame data pairs facilitates the calculation of the correlation between voltage and current data within the same time interval. Each frame data pair contains voltage and current information at the same point in time. By analyzing these frame data pairs, the dynamic correlation between voltage and current can be understood more accurately. In this way, multiple frame data pairs can be obtained, each representing voltage and current data within a specific time interval.

[0040] Step S213: Calculate the correlation degree of voltage and current frame data in each frame data pair by using the temporal correlation layer of the preset coupling feature extraction model, and obtain the corresponding frame correlation degree of each frame data pair.

[0041] The temporal correlation layer of the pre-defined coupled feature extraction model calculates the correlation degree between the voltage and current frame data in each frame data pair after obtaining multiple frame data pairs. The correlation degree can be calculated using methods such as the Pearson correlation coefficient method or the mutual information method.

[0042] Taking the Pearson correlation coefficient method as an example, the correlation between two variables is measured by calculating the degree of linear correlation between them. For voltage and current frame data in each frame data pair, their means are first calculated. Then, the difference between each data point and the mean is calculated. Next, the product of these differences is calculated. Finally, all products are added together and divided by the product of the number of data points and the standard deviation to obtain the Pearson correlation coefficient. The value of this coefficient ranges from -1 to 1. The closer the value is to 1, the stronger the positive linear correlation between the two variables; the closer the value is to -1, the stronger the negative linear correlation; and the closer the value is to 0, the weaker the linear correlation between the two.

[0043] By calculating the correlation degree for each frame data pair, the corresponding frame correlation degree can be obtained. These frame correlation degrees reflect the degree of correlation between voltage and current in different time intervals.

[0044] Step S214: Arrange the corresponding frame correlation degrees of multiple frame data pairs in chronological order to generate a voltage-current correlation degree sequence.

[0045] After obtaining the frame correlation degree corresponding to each frame data pair, these frame correlation degrees are arranged in chronological order. Since each frame data pair corresponds to a time interval, the frame correlation degrees can be sorted according to the chronological order of the frames.

[0046] Arranging the frame correlation in chronological order clearly shows how the correlation between voltage and current changes over time. By analyzing the voltage-current correlation sequence, we can understand the correlation patterns and trends between voltage and current in different time periods. Finally, the arranged frame correlations are combined into a sequence, namely the voltage-current correlation sequence.

[0047] Step S220: Extract current change data and temperature distribution data from the real-time monitoring data set of the battery. Perform spatial matching on the current change data and temperature distribution data to obtain spatially matched current change data and spatially matched temperature distribution data. Input the spatially matched current change data and spatially matched temperature distribution data into a preset linkage feature extraction model. Perform spatial feature mapping on the spatially matched current change data and spatially matched temperature distribution data through the spatial mapping layer of the preset linkage feature extraction model to generate a current-temperature spatial mapping matrix. And perform feature integration on the spatially matched current change data and spatially matched temperature distribution data based on the current-temperature spatial mapping matrix through the feature integration layer of the preset linkage feature extraction model to obtain current-temperature linkage features.

[0048] Temperature distribution data refers to the temperature information of a battery at different spatial locations, reflecting the internal temperature distribution of the battery. Since different parts of the battery may generate heat differently during operation, temperature distribution data helps to understand the battery's thermal characteristics and health status. Spatial matching refers to aligning current variation data and temperature distribution data in a spatial dimension, ensuring that the data at the same spatial location correspond. Because current and temperature data may be collected from different spatial locations, spatial matching can eliminate spatial errors, making subsequent correlation analysis more accurate. Spatial matching can be achieved using coordinate matching methods, that is, adding spatial coordinate information to the collected current and temperature data respectively, and then aligning the two data based on the coordinates. In practice, the spatial location of the other data can be adjusted based on the collection location of one data point, so that the two data points have corresponding values ​​at the same spatial location.

[0049] The pre-trained linkage feature extraction model is a pre-trained model used to extract the linkage features between current and temperature. This model consists of two main parts: a spatial mapping layer and a feature integration layer. The spatial mapping layer performs spatial feature mapping on the spatially matched current variation data and spatially matched temperature distribution data, mapping the current and temperature data into a two-dimensional matrix, namely the current-temperature spatial mapping matrix. Each element of this matrix represents the correlation information between current and temperature at a spatial location.

[0050] The feature integration layer, based on the current-temperature spatial mapping matrix, integrates spatially matched current variation data and spatially matched temperature distribution data to generate a comprehensive current-temperature linkage feature. This feature can more comprehensively reflect the interaction between current and temperature, as well as the battery's thermal characteristics and health status.

[0051] Step S230: Input the voltage-current coupling features and current-temperature linkage features into the preset interactive feature extraction model, perform cross-correlation analysis on the voltage-current coupling features and current-temperature linkage features through the cross-correlation layer of the preset interactive feature extraction model to generate voltage-temperature cross-correlation features, and perform feature enhancement processing on the voltage-temperature cross-correlation features through the feature enhancement layer of the preset interactive feature extraction model to obtain voltage-temperature interactive features.

[0052] The pre-trained interaction feature extraction model is a pre-trained model used to extract interaction features between voltage, current, and temperature. This model consists of two main parts: a cross-correlation layer and a feature enhancement layer. The cross-correlation layer performs cross-correlation analysis on voltage-current coupling features and current-temperature linkage features to uncover the intrinsic connections and interactions between these two features. Through cross-correlation analysis, complex correlation patterns between voltage, current, and temperature can be discovered, generating voltage-temperature cross-correlation features.

[0053] The voltage-temperature cross-correlation characteristic reflects the relationship between voltage and temperature through the intermediate variable of current. It integrates information from voltage-current coupling characteristics and current-temperature linkage characteristics, and can reveal more deeply the internal physical processes and health status of the battery.

[0054] The feature enhancement layer then enhances the voltage-temperature cross-correlation features to highlight the more important features for assessing battery health. This enhancement improves the expressive power and discriminative power of the features, making the health assessment more accurate and reliable. The final voltage-temperature interaction features are a comprehensive feature after enhancement, which more effectively reflects the battery's health status.

[0055] As one implementation method, in step S230, the voltage-temperature cross-correlation features are enhanced by the feature enhancement layer of a preset interactive feature extraction model to obtain voltage-temperature interactive features. Specifically, this can be implemented as follows: steps S231~S235:

[0056] Step S231: The voltage-temperature cross-correlation features are decomposed into feature primitives by the feature enhancement layer of the preset interactive feature extraction model. The voltage-temperature cross-correlation features are decomposed into multiple independent feature primitives, and each feature primitive corresponds to a single physical action mode in the voltage-temperature interaction process.

[0057] Characteristic primitive decomposition refers to the process of breaking down complex voltage-temperature cross-correlation features into multiple independent characteristic primitives with corresponding physical meanings. Each characteristic primitive represents a single physical action mode in the voltage-temperature interaction process, such as the effect of voltage changes on temperature, or the feedback of temperature changes on voltage.

[0058] The feature enhancement layer of the pre-defined interactive feature extraction model can employ methods such as Principal Component Analysis (PCA) and Independent Component Analysis (ICA) to decompose feature primitives. Taking PCA as an example, this method performs a linear transformation on the voltage-temperature cross-correlation feature data, projecting it onto a new set of orthogonal coordinate axes to maximize the variance of the data on these axes. The vectors corresponding to these new coordinate axes are the principal components, and each principal component can be considered a feature primitive. By selecting the principal components with a large variance contribution rate as the main feature primitives, the voltage-temperature cross-correlation features can be decomposed into multiple independent feature primitives. This simplifies complex features, facilitating individual analysis and processing of each feature primitive.

[0059] Step S232: Calculate the sensitivity coefficient of each feature primitive in the preset health state sample library. The sensitivity coefficient characterizes the response strength of the feature primitive as the health state of the battery changes.

[0060] The pre-collected health status sample library is a database containing various characteristic data of batteries under different health states, collected and organized in advance. The data in this sample library is obtained through long-term monitoring and analysis of a large number of batteries under different usage stages and operating conditions. Each sample in the library corresponds to a battery health state, along with various characteristic data for that state, including voltage, current, temperature, and other related features.

[0061] The sensitivity coefficient is an indicator used to measure the sensitivity of each feature element to changes in the battery's health status. For each feature element, its sensitivity coefficient is calculated by analyzing its value changes under different health states in a pre-defined health status sample library. Specifically, regression analysis can be used. For example, a regression model is established with the battery's health status as the independent variable and the feature element's value as the dependent variable. By calculating regression coefficients and other statistics, the response strength of the feature element to changes in the battery's health status, i.e., the sensitivity coefficient, is obtained. A larger sensitivity coefficient indicates that the feature element is more sensitive to changes in the battery's health status and plays a more important role in health status assessment.

[0062] Step S233: Dynamically weight each feature primitive based on the sensitivity coefficient, mark the feature primitive with a sensitivity coefficient higher than the preset threshold as key primitive, and increase its weight to the preset enhancement factor.

[0063] Dynamic weighting is the process of assigning weights to each feature primitive based on its sensitivity coefficient. The sensitivity coefficient reflects the importance of the feature primitive to changes in the battery's health status; therefore, dynamic weighting can highlight the role of those feature primitives that are more critical to the health status assessment.

[0064] The preset threshold is a pre-defined standard value used to distinguish between critical and non-critical features. When the sensitivity coefficient of a feature feature is higher than this threshold, it indicates that the feature feature is highly sensitive to changes in battery health and is marked as a critical feature. The preset enhancement factor is a predetermined coefficient used to increase the weight of critical features. For example, if the preset enhancement factor is 2, then for feature features marked as critical features, their original weight is multiplied by 2, giving them a greater influence in subsequent feature reconstruction and health status assessment.

[0065] This dynamic weighting and key primitive labeling method can effectively improve the expressive power of features, enabling subsequent analysis to focus more on those features that are most sensitive to changes in battery health status, thereby improving the accuracy of health status assessment.

[0066] Step S234: Perform nonlinear correlation reconstruction on the weighted feature primitives, and construct an enhanced feature network by introducing higher-order interaction terms between feature primitives. The higher-order interaction terms are calculated based on the covariance matrix between key primitives.

[0067] Nonlinear correlation reconstruction refers to the recombination and reconstruction of weighted feature elements based on the nonlinear relationships between them, forming a feature network that better reflects the health status of the battery. In actual battery operation, the relationships between the various feature elements are often not simple linear relationships, but rather involve complex nonlinear interactions.

[0068] Higher-order interaction terms represent second-order and higher-order interactions between feature primitives. Introducing these terms allows for the capture of more complex interactions between feature primitives. These higher-order interaction terms are calculated based on the covariance matrix between key primitives. The covariance matrix reflects the linear correlation and trends among key primitives. By analyzing and processing the covariance matrix, higher-order interaction information between key primitives can be obtained. The enhanced feature network is a network structure jointly constructed from weighted feature primitives and higher-order interaction terms. This network structure can more comprehensively and accurately reflect the complex relationships between feature primitives, thus providing richer information for subsequent feature aggregation and health status assessment.

[0069] Step S235: Perform feature aggregation on the enhanced feature network, and focus on the interaction region of key primitives through a self-attention mechanism to generate the final voltage-temperature interaction features.

[0070] Feature aggregation is the process of integrating and summarizing the various feature primitives and their interaction information in a reinforced feature network. The goal is to transform complex feature network information into a simpler and more representative feature.

[0071] When performing feature aggregation on a reinforced feature network, a self-attention mechanism can help identify and focus on the interaction regions of key primitives. The interaction regions of key primitives refer to the areas of interaction between feature primitives that are most critical to assessing the battery's health status. Through the self-attention mechanism, the importance weight of each feature primitive in the interaction region can be calculated, and then the feature primitives are weighted and summed based on these weights, thereby achieving the focus and integration of key information. After feature aggregation and processing by the self-attention mechanism, the final generated voltage-temperature interaction feature is a feature that integrates the complex interaction information between voltage, current, and temperature, and can more effectively reflect the battery's health status.

[0072] Step S240: Integrate the voltage-current coupling features, current-temperature linkage features, and voltage-temperature interaction features into a multi-domain correlation feature set for the battery.

[0073] Voltage-current coupling characteristics, current-temperature linkage characteristics, and voltage-temperature interaction characteristics reflect the state information of the battery during operation from different angles and levels. Voltage-current coupling characteristics mainly reflect the interaction between voltage and current, reflecting the electrical characteristics of the battery; current-temperature linkage characteristics highlight the correlation between current and temperature, reflecting the thermal characteristics of the battery; voltage-temperature interaction characteristics integrate the complex interaction information between voltage, current, and temperature, providing a more comprehensive reflection of the overall operating state of the battery.

[0074] Integrating these three features to form a multi-domain associated feature set for the battery aims to synthesize and unify feature information from different domains. Through integration, redundant information between features can be eliminated while retaining their complementary information, resulting in a more comprehensive and accurate feature set reflecting the battery's operating status and health condition. The integration process can employ feature splicing.

[0075] Step S300: Construct a battery health status association map based on the multi-domain association feature set of the battery.

[0076] A battery health status correlation graph is a graphical representation method used to describe the correlation between various battery characteristics and their health status. It displays the interrelationships between various features in a multi-domain correlation feature set of the battery, as well as the mapping relationship between these features and the battery's health status. By constructing this correlation graph, the operating status and health status of the battery can be understood more intuitively.

[0077] As one implementation method, step S300, constructing a battery health status correlation map based on the battery multi-domain correlation feature set, can be specifically implemented as the following steps S310~S340:

[0078] Step S310: Extract voltage-current coupling features, current-temperature linkage features, and voltage-temperature interaction features from the multi-domain correlation feature set of the battery. Take the voltage-current coupling features as the first feature node, the current-temperature linkage features as the second feature node, and the voltage-temperature interaction features as the third feature node. The first feature node, the second feature node, and the third feature node form a feature node set.

[0079] Feature nodes are the basic elements in the correlation graph, with each node representing a corresponding feature. Voltage-current coupling features, current-temperature linkage features, and voltage-temperature interaction features are extracted from the battery's multi-domain correlation feature set and designated as the first, second, and third feature nodes, respectively. These feature nodes reflect the battery's operating status and health condition from different perspectives. The feature node set, composed of these three feature nodes, contains key feature information closely related to the battery's health status. By combining these feature nodes, a preliminary correlation graph structure can be constructed.

[0080] Step S320: Perform feature co-occurrence analysis on the first feature node and the second feature node to determine the co-occurrence frequency. Based on the co-occurrence frequency, calculate the initial association weight between the first feature node and the second feature node using a preset weight calculation model. Perform feature co-occurrence analysis on the first feature node and the third feature node to determine the co-occurrence frequency. Based on the co-occurrence frequency, calculate the initial association weight between the first feature node and the third feature node using a preset weight calculation model. Perform feature co-occurrence analysis on the second feature node and the third feature node to determine the co-occurrence frequency. Based on the co-occurrence frequency, calculate the initial association weight between the second feature node and the third feature node using a preset weight calculation model. Combine the initial association weight between the first feature node and the second feature node, the initial association weight between the first feature node and the third feature node, and the initial association weight between the second feature node and the third feature node to form an initial node association weight set.

[0081] Feature co-occurrence analysis refers to analyzing the frequency with which two feature nodes appear simultaneously in the same data sample. Co-occurrence analysis is performed separately for the first and second feature nodes, the first and third feature nodes, and the second and third feature nodes. The co-occurrence frequency is obtained by counting the number of times these two feature nodes appear simultaneously in a large number of data samples and then dividing by the total number of samples. The co-occurrence frequency reflects the degree of correlation between the two feature nodes; the higher the co-occurrence frequency, the stronger the correlation between the two feature nodes. The pre-trained weight calculation model is a model used to calculate the correlation weights between feature nodes based on their co-occurrence frequencies. This model can be built based on machine learning algorithms such as neural networks and decision trees. It takes the co-occurrence frequency as input and establishes a mapping relationship between the co-occurrence frequency and the correlation weights by learning from a large amount of sample data. For each pair of feature nodes, their co-occurrence frequencies are input into the pre-trained weight calculation model, and the model outputs the initial correlation weights between them.

[0082] The initial node association weight set consists of the initial association weights between the first and second feature nodes, the initial association weights between the first and third feature nodes, and the initial association weights between the second and third feature nodes. The weights in this set reflect the initial association strength between each feature node.

[0083] Step S330: Input the feature node set and the initial node association weight set into the preset graph optimization model, and adjust each initial association weight in the initial node association weight set through the weight adjustment layer of the preset graph optimization model to obtain the optimized node association weight set.

[0084] The preset graph optimization model is used to optimize and adjust the initially constructed association graph. It adjusts the initial node association weights to make the association graph more accurately reflect the true relationships between feature nodes.

[0085] The weight adjustment layer is a component of the pre-defined graph optimization model. Its function is to recalculate and adjust the initial association weights in the initial node association weight set. The weight adjustment layer can employ various methods for weight adjustment, such as considering indirect associations between feature nodes and incorporating prior knowledge. By adjusting the initial association weights, unreasonable association weights can be eliminated, while highlighting more important associations, thereby improving the accuracy and reliability of the association graph. The final optimized node association weight set is an adjusted weight set that better reflects the true association strength between feature nodes, providing a basis for constructing a more accurate battery health status association graph.

[0086] As one implementation method, in step S330, the initial association weights in the initial node association weight set are adjusted through the weight adjustment layer of the preset graph optimization model to obtain the optimized node association weight set. Specifically, this can be implemented as follows: S331~S334:

[0087] Step S331: Extract the initial association weight between the first feature node and the second feature node from the initial node association weight set through the weight adjustment layer of the preset graph optimization model. Record it as the first initial weight. Extract the initial association weight between the first feature node and the third feature node. Record it as the second initial weight. Extract the initial association weight between the second feature node and the third feature node. Record it as the third initial weight.

[0088] The weight adjustment layer of the pre-defined graph optimization model first needs to extract the initial association weights between each feature node from the initial node association weight set. For ease of calculation and processing, the initial association weights between the first and second feature nodes are labeled as the first initial weights, the initial association weights between the first and third feature nodes are labeled as the second initial weights, and the initial association weights between the second and third feature nodes are labeled as the third initial weights. These initial weights are preliminary association weights obtained based on feature co-occurrence analysis and the pre-defined weight calculation model, reflecting the initial association strength between feature nodes.

[0089] Step S332: Input the first initial weight, the second initial weight, and the third initial weight into the preset weight transfer model. Calculate the indirect association weight between the first feature node and the third feature node through the preset weight transfer model, and record it as the first indirect weight. Calculate the indirect association weight between the second feature node and the third feature node through the first feature node, and record it as the second indirect weight. Calculate the indirect association weight between the third feature node and the second feature node through the first feature node, and record it as the third indirect weight.

[0090] The pre-defined weight transfer model is used to calculate the indirect association weights between feature nodes. In an association graph, feature nodes are not only directly associated, but may also be indirectly associated through other nodes. The pre-defined weight transfer model can calculate the indirect association weights between each feature node based on the initial association weights.

[0091] For the first, second, and third feature nodes, calculate the indirect association weights between them. The first indirect weight is the indirect association weight between the first feature node and the third feature node through the second feature node, reflecting the degree of influence of the first feature node on the third feature node through the mediation of the second feature node. Similarly, the second indirect weight is the indirect association weight between the second feature node and the third feature node through the first feature node, and the third indirect weight is the indirect association weight between the third feature node and the second feature node through the first feature node.

[0092] The pre-defined weight transfer model can be constructed based on shortest path algorithms and transitive closure algorithms in graph theory. By inputting the first, second, and third initial weights into the pre-defined weight transfer model, the model calculates the weights of each indirect association based on the relationships between nodes and the weight transfer rules.

[0093] Step S333: Adjust the second initial weight based on the first indirect weight to obtain the first adjusted weight; adjust the third initial weight based on the second indirect weight to obtain the second adjusted weight; and adjust the first initial weight based on the third indirect weight to obtain the third adjusted weight.

[0094] After obtaining the indirect association weights, the initial association weights need to be adjusted based on these indirect association weights. The purpose of the adjustment is to take into account the indirect association relationships between feature nodes, so that the association weights can better reflect the true association strength between feature nodes.

[0095] For the second initial weight, the first indirect weight is used as an adjustment factor. The adjustment method can be to sum the first indirect weight and the second initial weight using a weighted average to obtain the first adjusted weight. Similarly, for the third initial weight, the second indirect weight is used as an adjustment factor to obtain the second adjusted weight; and for the first initial weight, the third indirect weight is used as an adjustment factor to obtain the third adjusted weight.

[0096] This adjustment method can incorporate indirect association information between feature nodes into the association weight, thereby making the association weight more accurately reflect the association relationship between feature nodes.

[0097] Step S334: Combine the first adjusted weight, the second adjusted weight, and the third adjusted weight to form the optimized node association weight set.

[0098] After adjusting the initial association weights, we obtained the first adjusted weight, the second adjusted weight, and the third adjusted weight. These three adjusted weights are combined into a new set, which is the optimized node association weight set.

[0099] The optimized node association weight set takes into account the indirect association relationships between feature nodes. Compared with the initial node association weight set, it can more accurately reflect the true association strength between feature nodes.

[0100] Step S340: Construct a battery health status association graph from the feature node set and the optimized node association weight set.

[0101] After obtaining the set of feature nodes and the optimized set of node association weights, a battery health status association graph can be constructed. The feature nodes in the feature node set serve as nodes in the graph, representing various features related to the battery health status; the weights in the optimized set of node association weights serve as the weights of the edges between nodes in the graph, reflecting the association strength between each feature node.

[0102] By combining the weights of feature nodes and edges, a graph structure is formed, namely the battery health status correlation graph. This graph can intuitively display the relationships between various feature nodes and their potential connections with the battery health status. Through the analysis of this correlation graph, a deeper understanding of the battery's operating status and health condition can be obtained.

[0103] Step S400: Perform multi-scale assessment of battery health status using the battery health status correlation map to obtain a set of battery health status assessment results, and output the online monitoring results of battery health status based on the set of battery health status assessment results.

[0104] Multi-scale assessment of battery health status refers to a comprehensive and in-depth evaluation of the battery's health status from different time scales and dimensions. Different time scales can include short-term, medium-term, and long-term. By conducting assessments at different time scales, a more comprehensive understanding of the battery's health status trends can be obtained.

[0105] The battery health status assessment results set is a collection of assessment results obtained at different time scales. This set of results can include short-term health trend assessments, medium-term performance degradation assessments, and long-term lifespan prediction assessments. These results reflect the battery's health status from different perspectives, providing rich information for a comprehensive assessment of battery performance and lifespan.

[0106] The online monitoring results for battery health status, based on a set of battery health status assessment results, involve synthesizing and analyzing the results obtained from multi-scale assessments, and then outputting them to users or monitoring systems in an intuitive manner. Online monitoring results can be presented in the form of reports, charts, and early warning information, enabling users to understand the current health status and future development trends of the battery in a timely manner.

[0107] As one implementation method, step S400 involves performing multi-scale assessment of battery health status using a battery health status correlation map to obtain a set of battery health status assessment results. Specifically, this can be implemented as follows: S410~S450:

[0108] Step S410: Extract the feature node set and node association weight set from the battery health status association graph, and construct a short-term feature node subgraph based on the feature node set and node association weight set. The short-term feature node subgraph includes the feature nodes that are active within a preset short-term time window and their corresponding node association weights.

[0109] A preset short-term time window is a pre-defined, relatively short time range used to define the time scale for short-term assessment. Within this time window, certain feature nodes may be more active, meaning their values ​​change more significantly and have a greater impact on the battery's short-term health status. After extracting the set of feature nodes and the set of node association weights from the battery health status correlation graph, the active feature nodes within this time window are selected based on the preset short-term time window. These active feature nodes and their corresponding node association weights form a short-term feature node subgraph. The short-term feature node subgraph is a local graph structure that focuses on feature nodes and their relationships that have a significant impact on the battery's health status in the short term.

[0110] Step S420: Input the short-term feature node subgraph into the preset short-term assessment model. Through the node influence analysis layer of the preset short-term assessment model, perform influence analysis on each feature node and its associated weight in the short-term feature node subgraph to generate a short-term feature influence degree sequence. Based on the short-term feature influence degree sequence, perform trend prediction on the short-term health status of the battery through the trend prediction layer of the preset short-term assessment model to obtain the short-term health trend assessment result.

[0111] The preset short-term assessment model is a model specifically designed for assessing and predicting the short-term health status of batteries, and includes a node impact analysis layer and a trend prediction layer.

[0112] The node influence analysis layer is used to perform in-depth analysis of each feature node and its associated weights in the short-term feature node subgraph, calculating the degree of influence of each feature node on other nodes, thereby generating a short-term feature influence sequence. The node influence analysis layer can employ methods such as graph neural networks to learn the relationships and weights between feature nodes and calculate the influence of each feature node. The influence reflects the importance and impact of that feature node in the short-term health status assessment.

[0113] The trend prediction layer predicts the short-term health status of the battery based on the short-term feature influence sequence. Methods such as time series analysis and machine learning can be used to predict the short-term health status trend of the battery based on the changing patterns of the short-term feature influence sequence. For example, by analyzing the changes in the influence of feature nodes over time, it can predict whether parameters such as battery voltage and current will experience abnormal fluctuations in the short term, thus obtaining a short-term health trend assessment result.

[0114] As one implementation method, in step S420, the influence analysis of each feature node and its associated weight in the short-term feature node subgraph is performed through the node influence analysis layer of the preset short-term evaluation model to generate a short-term feature influence degree sequence. Specifically, this can be implemented as follows: S421~S424:

[0115] Step S421: Extract each feature node from the short-term feature node subgraph using the node influence analysis layer of the preset short-term evaluation model and record it as a short-term feature node. Extract the node association weight between each short-term feature node from the short-term feature node subgraph and record it as a short-term association weight.

[0116] The node impact analysis layer of the pre-defined short-term assessment model first needs to extract each feature node and its associated weights from the short-term feature node subgraph. These feature nodes are denoted as short-term feature nodes, and the node association weights between them are denoted as short-term association weights. These short-term feature nodes and short-term association weights are the basic data for performing node impact analysis.

[0117] Step S422: For each short-term feature node, identify other short-term feature nodes that are associated with it and record them as associated nodes. Extract the short-term association weight between the short-term feature node and each associated node from the short-term association weight and record it as node pair association weight.

[0118] For each short-term feature node, it is necessary to identify other short-term feature nodes that are associated with it; these nodes are called associated nodes. By analyzing the structure of the short-term feature node subgraph, the associated nodes of each short-term feature node can be determined. Then, the association weight between the short-term feature node and each associated node is extracted from the short-term association weights, denoted as the node-pair association weight. The node-pair association weight reflects the strength of the association between each short-term feature node and its associated nodes.

[0119] Step S423: Multiply the feature value of each associated node by the corresponding node-to-association weight to obtain the influence value of the associated node on the current short-term feature node. Sum the influence values ​​of all associated nodes on the current short-term feature node to obtain the total influence of the current short-term feature node.

[0120] For each short-term feature node, the influence of its associated nodes on it is calculated. Specifically, the feature value of each associated node (such as the specific values ​​of voltage, current, etc.) is multiplied by the corresponding node-to-association weight to obtain the influence value of that associated node on the current short-term feature node. Then, the influence values ​​of all associated nodes on the current short-term feature node are summed to obtain the total influence of the current short-term feature node. The total influence reflects the combined influence of all associated nodes on the current short-term feature node.

[0121] Step S424: Sort the total influence of each short-term feature node in chronological order to generate a short-term feature influence sequence.

[0122] After obtaining the total influence of each short-term feature node, these total influence values ​​are sorted in chronological order. The purpose of sorting is to organize the influence of short-term feature nodes according to the time dimension, forming a sequence, namely the short-term feature influence sequence. This sequence can clearly show how the influence of each feature node changes over time in the short term.

[0123] Step S430: Construct a mid-term feature node subgraph based on the feature node set and the node association weight set. The mid-term feature node subgraph contains the active feature nodes and their corresponding node association weights within a preset mid-term time window. Input the mid-term feature node subgraph into a preset mid-term evaluation model. Perform attenuation analysis on each feature node and its associated weight in the mid-term feature node subgraph through the attenuation analysis layer of the preset mid-term evaluation model to generate a mid-term feature attenuation sequence. Based on the mid-term feature attenuation sequence, evaluate the mid-term performance status of the battery through the performance evaluation layer of the preset mid-term evaluation model to obtain the mid-term performance attenuation evaluation result.

[0124] The preset mid-term time window is a pre-defined, medium-length time range used to define the time scale for mid-term evaluation. Within this time window, changes in feature nodes may affect the mid-term performance of the battery. Based on the set of feature nodes and the set of node association weights, active feature nodes and their corresponding node association weights are selected within the preset mid-term time window to construct a mid-term feature node subgraph. The mid-term feature node subgraph focuses on feature nodes and their associations that have a significant impact on battery performance during the mid-term.

[0125] The pre-defined mid-term evaluation model is used to assess the mid-term performance status of a battery. It includes a degradation analysis layer and a performance evaluation layer. The degradation analysis layer performs degradation analysis on each feature node and its associated weights in the mid-term feature node subgraph. By comparing the changes in the feature values ​​of feature nodes at the initial and final moments of the pre-defined mid-term, and combining the node association weights, it calculates the degree of degradation for each feature node, generating a mid-term feature degradation sequence. The performance evaluation layer evaluates the mid-term performance status of the battery based on the mid-term feature degradation sequence. It can determine whether the battery's performance has degraded during the mid-term and the degree of degradation based on the changing patterns of the mid-term feature degradation sequence, thus obtaining the mid-term performance degradation evaluation result.

[0126] As one implementation method, in step S430, the decay analysis layer of the preset mid-term evaluation model is used to perform decay analysis on each feature node and its associated weight in the mid-term feature node subgraph to generate a mid-term feature decay sequence. Specifically, this can be implemented as the following steps S431~S434:

[0127] Step S431: Extract each feature node from the mid-term feature node subgraph through the decay analysis layer of the preset mid-term evaluation model and record it as a mid-term feature node. Extract the feature value of each mid-term feature node at the preset mid-term initial time from the mid-term feature node subgraph and record it as the initial feature value. Extract the feature value of each mid-term feature node at the preset mid-term end time and record it as the end feature value.

[0128] The attenuation analysis layer of the pre-defined mid-term evaluation model first needs to extract each feature node from the mid-term feature node subgraph, denoted as the mid-term feature node. Then, the feature values ​​of these mid-term feature nodes are extracted at the pre-defined initial and final mid-term times. The initial feature value reflects the state of each feature node at the start of the mid-term evaluation, while the final feature value reflects the state of each feature node at the end of the mid-term evaluation. These feature values ​​can be specific values ​​of relevant features such as voltage, current, and temperature.

[0129] Step S432: Calculate the difference between the end feature value and the initial feature value of each intermediate feature node to obtain the feature value change. Extract the node association weight between each intermediate feature node from the intermediate feature node subgraph and record it as the intermediate association weight.

[0130] For each intermediate-term feature node, the change in feature value is obtained by calculating the difference between its ending feature value and its initial feature value. This change in feature value reflects the change of the feature node within a preset intermediate-term time window. Simultaneously, the node association weights between intermediate-term feature nodes are extracted from the intermediate-term feature node subgraph and denoted as intermediate-term association weights. These intermediate-term association weights reflect the strength of the association between intermediate-term feature nodes.

[0131] Step S433: For each intermediate feature node, determine the change in the feature value of other intermediate feature nodes associated with it and record it as the associated feature change. Multiply each associated feature change by the corresponding intermediate association weight to obtain the association attenuation contribution value. Sum all the association attenuation contribution values ​​to obtain the total association attenuation contribution value.

[0132] For each intermediate-term feature node, identify its associated intermediate-term feature nodes and obtain the changes in their feature values, denoted as the associated feature changes. Then, multiply each associated feature change by its corresponding intermediate-term association weight to obtain the association decay contribution value of each associated node to the current intermediate-term feature node. The association decay contribution value reflects the degree of contribution of each associated node to the decay of the current intermediate-term feature node. Finally, sum the association decay contribution values ​​of all associated nodes to the current intermediate-term feature node to obtain the total association decay contribution value. The total association decay contribution value reflects the combined influence of all associated nodes on the decay of the current intermediate-term feature node.

[0133] Step S434: Add the change in the eigenvalue of the current intermediate feature node to the associated total decay contribution value to obtain the total decay of the intermediate feature node. Sort the total decay of each intermediate feature node in chronological order to generate an intermediate feature decay sequence.

[0134] The total decay of a mid-term feature node is obtained by adding the change in its eigenvalue to the total decay contribution of the associated nodes. The total decay comprehensively considers the changes in the feature node itself as well as the influence of associated nodes, thus providing a more complete picture of the decay of the feature node within the preset mid-term time window.

[0135] The total decay of each intermediate feature node is sorted chronologically to generate an intermediate feature decay sequence. This sequence clearly shows how the decay of each feature node changes over time within the intermediate period.

[0136] Step S440: Construct a long-term feature node subgraph based on the feature node set and the node association weight set. The long-term feature node subgraph contains the active feature nodes and their corresponding node association weights within a preset long-term time window. Input the long-term feature node subgraph into a preset long-term evaluation model. Perform life prediction analysis on each feature node and its association weight in the long-term feature node subgraph through the life prediction layer of the preset long-term evaluation model to generate long-term life prediction indicators. Based on the long-term life prediction indicators, predict the long-term life status of the battery through the result output layer of the preset long-term evaluation model to obtain the long-term life prediction evaluation result.

[0137] A preset long-term time window is a pre-defined, relatively long time range used to define the time scale for long-term evaluation. Within this time window, changes in certain characteristic nodes may have a critical impact on the long-term lifespan of the battery.

[0138] Based on the set of feature nodes and the set of node association weights, a long-term feature node subgraph is constructed by selecting active feature nodes and their corresponding node association weights within a preset long-term time window. This long-term feature node subgraph focuses on feature nodes and their associations that have a significant impact on battery life over the long term.

[0139] The pre-defined long-term assessment model is used to predict the long-term lifespan of batteries. It consists of two main parts: a lifespan prediction layer and a result output layer. The lifespan prediction layer performs lifespan prediction analysis on each feature node and its associated weights in the long-term feature node subgraph. By analyzing the historical data and changing trends of the feature nodes, combined with the node association weights, it predicts the impact of each feature node on the battery lifespan and generates long-term lifespan prediction indicators.

[0140] The output layer predicts the long-term lifespan of a battery based on long-term lifespan prediction indicators. It can determine the battery's long-term lifespan status, such as remaining service life and end-of-life time, based on the values ​​and trends of these indicators, thus obtaining a long-term lifespan prediction assessment result.

[0141] As one implementation method, in step S440, the long-term lifetime prediction index is generated by performing lifetime prediction analysis on each feature node and its associated weight in the long-term feature node subgraph through the lifetime prediction layer of the preset long-term evaluation model. Specifically, this can be implemented as the following steps S441~S445:

[0142] Step S441: Extract each feature node from the long-term feature node subgraph through the lifetime prediction layer of the preset long-term evaluation model and record them as long-term feature nodes. Extract the historical feature sequence of each long-term feature node from the long-term feature node subgraph. The historical feature sequence contains the feature values ​​of each long-term feature node at different times within the preset long-term time window.

[0143] The lifetime prediction layer of the pre-defined long-term assessment model first extracts each feature node from the long-term feature node subgraph, denoted as long-term feature nodes. Then, it extracts the feature values ​​of these long-term feature nodes at different times within a pre-defined long-term time window, forming a historical feature sequence. The historical feature sequence records the value changes of each long-term feature node over a long period.

[0144] Step S442: Perform trend fitting processing on the historical feature sequence of each long-term feature node to obtain the feature trend curve of the long-term feature node. Based on the feature trend curve, predict the time point when the feature value of the long-term feature node reaches the preset failure threshold and record it as the node failure time.

[0145] Trend fitting refers to the process of analyzing and fitting the historical feature sequences of long-term feature nodes using specific mathematical methods to find a suitable function to describe their changing trends. Common trend fitting methods include linear regression, multinomial regression, and exponential regression. Through trend fitting, a characteristic trend curve can be obtained for the long-term feature node, which reflects the trend of the feature node's value changes over a future period.

[0146] The preset failure threshold is a pre-defined limit for a characteristic value. When the characteristic value of a long-term characteristic node reaches this threshold, the battery performance corresponding to that node is considered to have deteriorated to the point where it can no longer function properly, i.e., it has reached a failure state. Based on the characteristic trend curve, the time point at which the characteristic value of the long-term characteristic node reaches the preset failure threshold can be predicted, and this time point is denoted as the node failure time.

[0147] As one implementation method, step S442 involves performing trend fitting processing on the historical feature sequence of each long-term feature node to obtain the feature trend curve of that long-term feature node. Specifically, this can be implemented as the following steps S4421~S4426:

[0148] Step S4421: Perform empirical mode decomposition on the historical feature sequence of each long-term feature node to obtain multiple intrinsic mode function components and one residual component.

[0149] Empirical Mode Decomposition (EMD) can decompose a complex signal into multiple intrinsic mode function (IMF) components and a residual component. For the historical feature sequence of each long-term feature node, EMD can decompose it into multiple IMF components with different oscillation characteristics and a residual component reflecting the overall trend of the signal. The IMF components have characteristics such as zero mean and symmetrical upper and lower envelopes, and they can reflect different frequency components and local variation information in the historical feature sequence.

[0150] Step S4422: Calculate the instantaneous frequency and instantaneous amplitude of each intrinsic mode function component, and select the effective intrinsic mode function components whose instantaneous frequency is within the preset physiological frequency range.

[0151] For each intrinsic mode function (IMF) component, its instantaneous frequency and instantaneous amplitude are calculated. The instantaneous frequency reflects the oscillation frequency of the signal at each moment, while the instantaneous amplitude reflects the magnitude of the signal's amplitude at each moment. The preset physiological frequency range is a pre-defined frequency interval determined based on the battery's physical characteristics and operating principle. It is used to filter out IMF components closely related to changes in battery performance. By filtering out valid IMF components whose instantaneous frequencies fall within the preset physiological frequency range, noise and irrelevant high- or low-frequency components can be removed, improving the accuracy of subsequent analysis.

[0152] Step S4423: Perform ensemble empirical mode decomposition on the effective intrinsic mode function components to obtain the quadratic decomposition components.

[0153] Ensemble Empirical Mode Decomposition (EEMD) is a method developed based on Empirical Mode Decomposition (EMD) to solve the mode aliasing problem in EMD. It involves performing EEMD on the selected effective intrinsic mode function (IMF) components, further decomposing them into more refined quadratic decomposition components. These quadratic decomposition components more accurately reflect the signal characteristics and variation patterns within the effective IMF components.

[0154] Step S4424: Reconstruct the secondary decomposition components and residual components to generate a denoised historical feature sequence.

[0155] Reconstructing the secondary decomposition components and residual components involves recombining the effective information obtained after decomposition and filtering to generate a new sequence. This new sequence removes noise and irrelevant components from the original historical feature sequence, and more smoothly and accurately reflects the true changing trend of long-term feature nodes, i.e., the denoised historical feature sequence.

[0156] Step S4425: Construct a hidden state transition model based on a long short-term memory network, input the denoised historical feature sequence into the hidden state transition model, and output the feature state evolution trajectory.

[0157] A hidden state transition model is constructed based on a Long Short-Term Memory (LSTM) network. This model can learn hidden patterns and state transition rules from denoised historical feature sequences. The denoised historical feature sequences are input into the hidden state transition model, which learns and computes to output the feature state evolution trajectory. The feature state evolution trajectory describes the state change trend of long-term feature nodes over a future period.

[0158] Step S4426: Generate the characteristic trend curve of the long-term characteristic node based on the characteristic state evolution trajectory using cubic spline interpolation.

[0159] Cubic spline interpolation fits the data by constructing a cubic polynomial between each data point, resulting in an interpolation curve with good smoothness and approximation accuracy. Based on the characteristic state evolution trajectory, cubic spline interpolation is used to interpolate discrete points on the trajectory, generating a continuous curve, namely the characteristic trend curve of the long-term characteristic node. The characteristic trend curve can intuitively show the future change trend of the long-term characteristic node, providing a basis for predicting the node failure time.

[0160] Step S443: Extract the node association weights between each long-term feature node from the long-term feature node subgraph and record them as long-term association weights. For each long-term feature node, determine the node failure time of other long-term feature nodes associated with it and record it as associated node failure time.

[0161] The node association weights between long-term feature nodes are extracted from the long-term feature node subgraph and denoted as long-term association weights. These long-term association weights reflect the strength of the associations between long-term feature nodes and are important for comprehensively considering the impact of each node on the long-term lifespan of the battery.

[0162] For each long-term characteristic node, identify other long-term characteristic nodes associated with it and obtain the failure times of these associated nodes, denoted as the associated node failure times. The associated node failure times reflect the impact of associated nodes on the current long-term characteristic node.

[0163] Step S444: Multiply the failure time of each associated node by the corresponding long-term association weight to obtain the association lifetime impact value. Sum all the association lifetime impact values ​​to obtain the total association lifetime impact value. Add the node failure time of the current long-term feature node to the total association lifetime impact value to obtain the comprehensive lifetime prediction time of the long-term feature node.

[0164] For each long-term feature node, the failure time of its associated nodes is multiplied by the corresponding long-term association weight to obtain the association lifetime impact value of each associated node on the current long-term feature node. The association lifetime impact value reflects the degree of influence of each associated node on the lifetime of the current long-term feature node.

[0165] The total lifetime impact value is obtained by summing the lifetime impact values ​​of all associated nodes on the current long-term feature node. The total lifetime impact value comprehensively considers the influence of all associated nodes on the lifetime of the current long-term feature node.

[0166] The overall lifetime prediction time of a long-term characteristic node is obtained by adding its failure time to the associated total lifetime impact value. This overall lifetime prediction time more comprehensively considers both the failure time of the long-term characteristic node itself and the impact of associated nodes.

[0167] Step S445: Calculate the average comprehensive lifetime prediction time based on the comprehensive lifetime prediction time of all long-term feature nodes, and use the average comprehensive lifetime prediction time as the long-term lifetime prediction indicator.

[0168] Statistical analysis was performed on the comprehensive lifetime prediction time of all long-term characteristic nodes, and their average values ​​were calculated to obtain the average comprehensive lifetime prediction time. The average comprehensive lifetime prediction time integrates information from all long-term characteristic nodes, providing a more comprehensive and accurate reflection of the battery's long-term lifespan. This average comprehensive lifetime prediction time is used as a long-term lifetime prediction indicator for subsequent long-term lifespan status prediction and assessment.

[0169] Step S450: Combine the short-term health trend assessment results, the medium-term performance degradation assessment results, and the long-term life prediction assessment results into a set of battery health status assessment results.

[0170] The short-term health trend assessment results reflect the changes in the battery's health status in the short term, helping users to promptly identify potential short-term abnormalities. The medium-term performance degradation assessment results show the degree of performance degradation of the battery in the medium term, providing a basis for formulating medium-term maintenance plans. The long-term life prediction assessment results predict the long-term life of the battery, providing a reference for battery replacement.

[0171] The various algorithms involved in the above descriptions of the embodiments of this invention, such as principal component analysis, interpolation algorithms, Pearson similarity algorithms, etc., can all be obtained from relevant content in the prior art. To save space, they will not be elaborated on in the embodiments of this invention. In addition, those skilled in the art can supplement the details based on common knowledge in the art when implementing the solutions of this invention. For example, they can use normalization to eliminate dimensional conflicts before feature fusion, use interpolation to eliminate dimensional differences, reasonably set thresholds based on historical data, experience or business scenario requirements, train the model based on a general model training method, set the number of layers in the model structure based on actual needs, select activation functions, etc. This invention will not provide redundant descriptions of overly detailed implementation processes.

[0172] Please see Figure 2 , Figure 2 This is a schematic diagram of a computer system provided in an embodiment of the present invention. The computer system includes at least a processor 101, a communication interface 102, and a memory 103. The processor 101, communication interface 102, and memory 103 can be connected via a bus or other means. The processor 101 (or Central Processing Unit, CPU) is the computing and control core of the computer system, capable of parsing various instructions and processing various data within the computer system. The communication interface 102 may optionally include standard wired interfaces or wireless interfaces (such as Wi-Fi, mobile communication interfaces, etc.), and can be used to send and receive data under the control of the processor 101; the communication interface 102 can also be used for data transmission and interaction within the computer system. The memory 103 is a storage device in the computer system used to store programs and data. It is understood that the memory 103 here can include the computer system's built-in memory, or it can include extended memory supported by the computer system. The memory 103 provides storage space, which stores the computer system's operating system, including but not limited to: Android system, iOS system, Windows Phone system, etc., which are not limited in this invention.

[0173] In one embodiment, the processor 101 executes the artificial intelligence-based online monitoring method for battery health status provided above in the embodiments of the present invention by running a computer program in the memory 103.

Claims

1. A method for online monitoring of battery health status based on artificial intelligence, characterized in that, include: Acquire a set of real-time monitoring data of the battery during its real-time operation; The real-time monitoring data set of the battery is subjected to feature association enhancement to generate a multi-domain association feature set of the battery; Specifically, this includes: extracting voltage fluctuation data and current change data from the real-time battery monitoring data set; performing time-series synchronization on the voltage fluctuation data and current change data to obtain time-synchronized voltage fluctuation data and time-synchronized current change data; inputting the time-synchronized voltage fluctuation data and time-synchronized current change data into a preset coupling feature extraction model; calculating the correlation degree of the time-synchronized voltage fluctuation data and time-synchronized current change data through the time-series correlation layer of the preset coupling feature extraction model to generate a voltage-current correlation degree sequence; and performing feature fusion on the time-synchronized voltage fluctuation data and time-synchronized current change data based on the voltage-current correlation degree sequence through the feature fusion layer of the preset coupling feature extraction model to obtain voltage-current coupling features; extracting current change data and temperature distribution data from the real-time battery monitoring data set; performing spatial matching on the current change data and temperature distribution data to obtain spatially matched current change data and spatially matched temperature distribution data; and inputting the spatially matched current change data and temperature distribution data into a preset coupling feature extraction model to obtain voltage-current coupling features. The spatially matched temperature distribution data is input into a preset linkage feature extraction model. The spatial mapping layer of the preset linkage feature extraction model performs spatial feature mapping on the spatially matched current change data and the spatially matched temperature distribution data to generate a current-temperature spatial mapping matrix. The feature integration layer of the preset linkage feature extraction model then integrates the features of the spatially matched current change data and the spatially matched temperature distribution data based on the current-temperature spatial mapping matrix to obtain current-temperature linkage features. The voltage-current coupling features and the current-temperature linkage features are input into a preset interactive feature extraction model. The cross-correlation layer of the preset interactive feature extraction model performs cross-correlation analysis on the voltage-current coupling features and the current-temperature linkage features to generate voltage-temperature cross-correlation features. The feature enhancement layer of the preset interactive feature extraction model then enhances the voltage-temperature cross-correlation features to obtain voltage-temperature interactive features. Finally, the voltage-current coupling features, the current-temperature linkage features, and the voltage-temperature interactive features are integrated into a multi-domain linkage feature set for the battery. The step of calculating the correlation between the time-synchronized voltage fluctuation data and the time-synchronized current change data through the time-series correlation layer of the preset coupling feature extraction model to generate a voltage-current correlation sequence includes: The voltage fluctuation data after time synchronization is divided into time series frames by the time series correlation layer of the preset coupling feature extraction model to obtain multiple voltage frame data. The current change data after time synchronization is divided into time series frames to obtain multiple current frame data. The multiple current frame data and the multiple voltage frame data have the same number of frames and corresponding frame times. For each voltage frame data, the current frame data corresponding to the voltage frame data is determined from the multiple current frame data, and each voltage frame data and its corresponding current frame data are combined into a frame data pair to obtain multiple frame data pairs. The correlation degree of voltage frame data and current frame data in each frame data pair is calculated by the temporal correlation layer of the preset coupling feature extraction model to obtain the frame correlation degree corresponding to each frame data pair. Arrange the corresponding frame correlation degrees of the multiple frame data pairs in chronological order to generate a voltage-current correlation degree sequence. A battery health status association map is constructed based on the battery multi-domain association feature set. The battery health status is assessed using a multi-scale process based on the battery health status correlation map, resulting in a set of battery health status assessment results. The online monitoring results of the battery health status are then output based on this set of results.

2. The method according to claim 1, characterized in that, The construction of the battery health status association map based on the battery multi-domain association feature set includes: Voltage-current coupling features, current-temperature linkage features, and voltage-temperature interaction features are extracted from the battery multi-domain correlation feature set. The voltage-current coupling features are used as the first feature node, the current-temperature linkage features are used as the second feature node, and the voltage-temperature interaction features are used as the third feature node. The first feature node, the second feature node, and the third feature node form a feature node set. Feature co-occurrence analysis is performed on the first feature node and the second feature node to determine the co-occurrence frequency. Based on the co-occurrence frequency, an initial association weight between the first feature node and the second feature node is calculated using a preset weight calculation model. Feature co-occurrence analysis is performed on the first feature node and the third feature node to determine the co-occurrence frequency. Based on the co-occurrence frequency, an initial association weight between the first feature node and the third feature node is calculated using the preset weight calculation model. Feature co-occurrence analysis is performed on the second feature node and the third feature node to determine the co-occurrence frequency. Based on the co-occurrence frequency, an initial association weight between the second feature node and the third feature node is calculated using the preset weight calculation model. The initial association weights between the first feature node and the second feature node, the first feature node and the third feature node, and the second feature node and the third feature node are combined to form an initial node association weight set. The feature node set and the initial node association weight set are input into a preset graph optimization model. The weight adjustment layer of the preset graph optimization model is used to adjust each initial association weight in the initial node association weight set to obtain the optimized node association weight set. A battery health status association map is constructed using the set of feature nodes and the optimized set of node association weights.

3. The method according to claim 2, characterized in that, The step of adjusting each initial association weight in the initial node association weight set through the weight adjustment layer of the preset graph optimization model to obtain the optimized node association weight set includes: The initial association weights between the first feature node and the second feature node are extracted from the initial node association weight set by the weight adjustment layer of the preset graph optimization model and recorded as the first initial weights. The initial association weights between the first feature node and the third feature node are extracted and recorded as the second initial weights. The initial association weights between the second feature node and the third feature node are extracted and recorded as the third initial weights. The first initial weight, the second initial weight, and the third initial weight are input into a preset weight transfer model. The indirect association weight generated by the first feature node through the second feature node to the third feature node is calculated through the preset weight transfer model and recorded as the first indirect weight. The indirect association weight generated by the second feature node through the first feature node to the third feature node is calculated and recorded as the second indirect weight. The indirect association weight generated by the third feature node through the first feature node to the second feature node is calculated and recorded as the third indirect weight. The first adjusted weight is obtained by adjusting the second initial weight based on the first indirect weight; the second adjusted weight is obtained by adjusting the third initial weight based on the second indirect weight; and the third adjusted weight is obtained by adjusting the first initial weight based on the third indirect weight. The first adjusted weight, the second adjusted weight, and the third adjusted weight are combined to form an optimized set of node association weights.

4. The method according to claim 1, characterized in that, The multi-scale assessment of battery health status using the battery health status correlation map yields a set of battery health status assessment results, including: Extract a set of feature nodes and a set of node association weights from the battery health status association graph. Construct a short-term feature node subgraph based on the set of feature nodes and the set of node association weights. The short-term feature node subgraph includes the feature nodes that are active within a preset short-term time window and their corresponding node association weights. The short-term feature node subgraph is input into a preset short-term assessment model. The node influence analysis layer of the preset short-term assessment model performs influence analysis on each feature node and its associated weight in the short-term feature node subgraph to generate a short-term feature influence degree sequence. Based on the short-term feature influence degree sequence, the trend prediction layer of the preset short-term assessment model performs trend prediction on the short-term health status of the battery to obtain the short-term health trend assessment result. A mid-term feature node subgraph is constructed based on the set of feature nodes and the set of node association weights. The mid-term feature node subgraph contains the feature nodes active within a preset mid-term time window and their corresponding node association weights. The mid-term feature node subgraph is input into a preset mid-term evaluation model. The decay analysis layer of the preset mid-term evaluation model performs decay analysis on each feature node and its association weight in the mid-term feature node subgraph to generate a mid-term feature decay sequence. Based on the mid-term feature decay sequence, the mid-term performance status of the battery is evaluated through the performance evaluation layer of the preset mid-term evaluation model to obtain the mid-term performance decay evaluation result. A long-term feature node subgraph is constructed based on the set of feature nodes and the set of node association weights. The long-term feature node subgraph contains the feature nodes that are active within a preset long-term time window and their corresponding node association weights. The long-term feature node subgraph is input into a preset long-term evaluation model. The life prediction layer of the preset long-term evaluation model performs life prediction analysis on each feature node and its association weight in the long-term feature node subgraph to generate a long-term life prediction index. Based on the long-term life prediction index, the long-term life status of the battery is predicted through the result output layer of the preset long-term evaluation model to obtain the long-term life prediction evaluation result. The short-term health trend assessment results, the medium-term performance degradation assessment results, and the long-term life prediction assessment results are combined to form a battery health status assessment result set.

5. The method according to claim 4, characterized in that, The step of performing influence analysis on each feature node and its associated weight in the short-term feature node subgraph through the node influence analysis layer of the preset short-term evaluation model to generate a short-term feature influence sequence includes: The node influence analysis layer of the preset short-term assessment model extracts each feature node from the short-term feature node subgraph and records it as a short-term feature node. The node association weight between each short-term feature node is extracted from the short-term feature node subgraph and recorded as a short-term association weight. For each short-term feature node, other short-term feature nodes that are associated with that short-term feature node are identified and recorded as associated nodes. The short-term association weight between that short-term feature node and each associated node is extracted from the short-term association weight and recorded as node pair association weight. The influence value of the associated node on the current short-term feature node is obtained by multiplying the feature value of each associated node with the corresponding node pair association weight. The total influence of the current short-term feature node is obtained by summing the influence values ​​of all associated nodes on the current short-term feature node. The total influence of each short-term feature node is sorted according to time sequence to generate a short-term feature influence sequence.

6. The method according to claim 4, characterized in that, The step of performing attenuation analysis on each feature node and its associated weight in the mid-term feature node subgraph through the attenuation analysis layer of the preset mid-term evaluation model to generate a mid-term feature attenuation sequence includes: The attenuation analysis layer of the preset mid-term evaluation model extracts each feature node from the mid-term feature node subgraph and records it as a mid-term feature node. The feature value of each mid-term feature node at the preset mid-term initial time is extracted from the mid-term feature node subgraph and recorded as the initial feature value. The feature value of each mid-term feature node at the preset mid-term end time is extracted and recorded as the end feature value. The difference between the end feature value and the initial feature value of each intermediate feature node is calculated to obtain the feature value change. The node association weight between each intermediate feature node is extracted from the intermediate feature node subgraph and recorded as the intermediate association weight. For each intermediate feature node, the change in the feature value of other intermediate feature nodes associated with it is determined and recorded as the associated feature change. Each associated feature change is multiplied by the corresponding intermediate association weight to obtain the association decay contribution value. All association decay contribution values ​​are summed to obtain the total association decay contribution value. The total decay of the current intermediate feature node is obtained by adding the change in the feature value of the current intermediate feature node to the associated total decay contribution value. The total decay of each intermediate feature node is sorted in chronological order to generate an intermediate feature decay sequence.

7. The method according to claim 4, characterized in that, The step of generating long-term lifetime prediction indicators by performing lifetime prediction analysis on each feature node and its associated weight in the long-term feature node subgraph through the lifetime prediction layer of the preset long-term evaluation model includes: The long-term feature node is extracted from the long-term feature node subgraph by the life prediction layer of the preset long-term assessment model and recorded as a long-term feature node. The historical feature sequence of each long-term feature node is extracted from the long-term feature node subgraph, and the historical feature sequence contains the feature values ​​of each long-term feature node at different times within the preset long-term time window. The historical feature sequence of each long-term feature node is subjected to trend fitting to obtain the feature trend curve of the long-term feature node. The time point when the feature value of the long-term feature node reaches the preset failure threshold is predicted based on the feature trend curve and recorded as the node failure time. The node association weights between each long-term feature node are extracted from the long-term feature node subgraph and recorded as long-term association weights. For each long-term feature node, the node failure time of other long-term feature nodes associated with it is determined and recorded as associated node failure time. The failure time of each associated node is multiplied by the corresponding long-term association weight to obtain the associated lifetime impact value. All associated lifetime impact values ​​are summed to obtain the total associated lifetime impact value. The node failure time of the current long-term feature node is added to the total associated lifetime impact value to obtain the comprehensive lifetime prediction time of the long-term feature node. The average comprehensive lifetime prediction time is calculated based on the comprehensive lifetime prediction time of all long-term characteristic nodes, and the average comprehensive lifetime prediction time is used as a long-term lifetime prediction index.

8. A computer system, characterized in that, include: A memory, wherein a computer program is stored; A processor is configured to load the computer program to implement the online monitoring method for battery health status based on artificial intelligence as described in any one of claims 1-7.

Images