Battery fault prediction method and system based on hierarchical health assessment model

Abstract

The invention relates to the technical field of battery cell capacity fault prediction, and particularly discloses a battery fault prediction method and system based on a hierarchical health assessment model, and the method comprises the steps: obtaining an operation historical data set and an initial capacity; performing aging stage division on the operation historical data; constructing a three-level prediction model; outputting a capacity prediction result set; constructing a dynamic residual sequence in combination with the stage sensing information and the long-short-term memory network; and outputting an alarm result. Compared with the prior art that the difference of the aging stages of the battery cell is ignored, and particularly, the technical problems that the model prediction error is large and early warning is not timely under the condition that the battery cell enters the later stage of nonlinear capacity degradation are solved. The prediction adaptability and sensitivity in each stage are effectively improved, and the accuracy and timeliness of battery cell fault early warning are enhanced.

Description

Technical Field

[0001] This invention relates to the field of battery cell capacity failure prediction technology, and in particular to a battery failure prediction method and system based on a graded health assessment model. Background Technology

[0002] Currently, with the widespread application of lithium-ion batteries in electric vehicles, energy storage power stations, and industrial power supply systems, accurately assessing cell capacity change trends and achieving early fault warnings has become a core research direction in the field of battery health management (BHM). Common capacity prediction methods in existing technologies include capacity fitting methods based on statistical regression or filtering, capacity prediction models based on end-to-end neural networks, and hybrid methods based on the fusion of physical modeling and data-driven approaches. Among these, statistical methods such as Kalman filtering and recursive least squares, while offering good real-time performance, struggle to handle the nonlinear changes during cell aging. Deep learning methods such as LSTM and GRU can automatically extract historical sequence features, but most employ a uniform structure for full-cycle training, failing to explicitly distinguish the feature evolution processes of different aging stages. This can easily lead to insufficient model learning in the early and late stages, affecting the sensitive capture of early capacity degradation and accurate tracking in the late stages. Fusion methods, while possessing some physical interpretability, rely on complex parameter calibration, resulting in high deployment and maintenance costs and poor versatility.

[0003] In actual operation, the capacity degradation of battery cells exhibits distinct phases, typically including an initial rapid decline phase, a mid-term stable operation phase, and a final rapid degradation phase. These different phases differ significantly in terms of capacity change rate, noise disturbance amplitude, and fault risk distribution. Directly using a unified prediction model for modeling can easily lead to problems such as inter-phase weight averaging and fitting error diffusion, resulting in large capacity prediction errors and delayed early warnings. Furthermore, most existing methods lack phase labels and phase awareness mechanisms, failing to achieve dynamic switching and targeted adjustments to the prediction model. They also lack effective strategies to address the non-stationarity and distribution drift of battery cell operating states, further limiting the model's generalization ability and response speed to abnormal trends. Especially in the final stages of battery cell aging, the sample size is scarce and the capacity degradation rate accelerates, often resulting in insufficient training for the unified model and significant warning blind spots. On the other hand, existing residual discrimination methods mostly use fixed thresholds or static standards to measure capacity prediction errors, failing to dynamically adjust the judgment boundaries by incorporating factors such as phase evolution patterns, time series trend derivatives, and compensation coefficients, thus reducing the accuracy of early warning judgments.

[0004] Therefore, there is an urgent need to propose a battery failure prediction method based on a graded health assessment model to improve the model's adaptability and early warning response capability at each stage of the entire life cycle. Summary of the Invention

[0005] To address the aforementioned technical shortcomings, the purpose of this invention is to propose a battery fault prediction method based on a graded health assessment model. This method aims to solve the technical problems in existing technologies that ignore the differences in cell aging stages, especially when cells enter the later stages of nonlinear capacity degradation, leading to large prediction errors and untimely warnings.

[0006] To solve the above technical problems, the present invention adopts the following technical solution: The present invention provides a battery fault prediction method based on a graded health assessment model.

[0007] The battery failure prediction method based on the hierarchical health assessment model includes:

[0008] Step S10: Obtain the set of historical operating data for the target battery cell. and initial cell capacity Running historical data sets This includes the input feature vector x and the actual cell capacity c;

[0009] Step S20: Based on the historical data set and initial cell capacity A weakly supervised aging stage segmentation labeling mechanism based on multi-source feature enhancement is used to perform the aging stage segmentation task, and a training dataset with stage labels is output. ;

[0010] Step S30: Based on the training dataset with stage labels A three-level cell fault prediction model is constructed; wherein, the three-level cell fault prediction model includes mutually independent first-level cell fault prediction sub-models. Second-level cell fault prediction sub-model and the third-level cell fault prediction sub-model ;

[0011] Step S40: Pre-construct a lightweight three-class classifier S(x), apply the lightweight three-class classifier S(x) to perform stage label prediction processing on the input feature vector x, and output the current prediction stage s; input the input feature vector x and the current prediction stage s into the three-level cell fault prediction model, and the three-level cell fault prediction model outputs a set of capacity prediction results for each level of cell fault prediction sub-model;

[0012] Step S50: Based on the capacity prediction result set, a dynamic residual evaluation mechanism combining LSTM network and stage perception is used to perform the cell fault early warning judgment task, and output the cell early capacity anomaly alarm result set.

[0013] Preferably, in step S10, the input feature vector x includes voltage feature vector, current feature vector, temperature feature vector, SOC feature vector, cumulative charge / discharge feature vector, and charge / discharge rate feature vector.

[0014] Preferably, in step S20, based on the historical data set... and initial cell capacity A weakly supervised aging stage segmentation labeling mechanism based on multi-source feature enhancement is used to perform the aging stage segmentation task, and a training dataset with stage labels is output. The steps specifically include:

[0015] Step S201: From the historical data set Extract the actual cell capacity c from the data, and then determine the cell capacity c and the initial cell capacity. Calculate the relative capacity decay rate , Based on the relative capacity decay rate Construct an aging index feature vector, and add the aging index feature vector to the input feature vector x to build an enhanced feature vector. Subsequently, the enhanced feature vectors were processed. Perform normalization processing and output the normalized enhanced feature vector. ;

[0016] Step S202: Optimize the normalized enhanced feature vector A nonlinear feature embedding transformation is performed using an adaptive kernel width Gaussian kernel mapping mechanism to obtain a spectral-preserving similarity matrix K; a low-dimensional embedding representation vector set Z is obtained by performing principal component analysis on the spectral-preserving similarity matrix K.

[0017] Step S203: Based on the low-dimensional embedding representation vector set Z, a weakly supervised spectral clustering task is performed, ultimately outputting a training dataset with stage labels. .

[0018] Preferably, in step S203, the task of weakly supervised spectral clustering based on the low-dimensional embedding representation vector set Z ultimately outputs a training dataset with stage labels. The steps specifically include:

[0019] First, construct a weighted similarity matrix based on the set of low-dimensional embedding representation vectors Z. Weighted similarity matrix Used to jointly characterize the spectral distance similarity and aging trend similarity between samples;

[0020] Based on the weighted similarity matrix The normalized Laplacian matrix is ​​calculated using the normalized graphical Laplacian spectral clustering method. , Where D is the angle matrix, used to ensure the symmetry normalization of the feature space; further, the normalized Laplace matrix is ​​normalized... Feature decomposition is performed to extract the feature vector set U corresponding to the k smallest feature values. Then, the feature vector set U is subjected to K-means clustering algorithm for preliminary stage division processing to obtain the preliminary label set S. The preliminary label set S includes a first preliminary stage cluster, a second preliminary stage cluster, and a third preliminary stage cluster. The first preliminary stage cluster is used to represent the initial stage of cell aging, the second preliminary stage cluster is used to represent the middle stage of cell aging, and the third preliminary stage cluster is used to represent the final stage of cell aging.

[0021] Further introduce relative capacity decay rate Based on relative capacity decay rate Perform interval mapping on the initial label set S to obtain unsupervised clustering labels. Further based on unsupervised clustering labels By using minimum deviation mapping, unsupervised clustering labels are... The calibration process yields final stage labels with physical semantic meaning, resulting in a training dataset with stage labels as the final output. .

[0022] Preferably, in step S30, the first-level cell fault prediction sub-model The training is performed using only a training dataset representing the initial stage of cell aging, and the loss function used is the standard mean squared error loss function; the second-level cell fault prediction sub-model The training is performed using only a training dataset representing the mid-stage of cell aging, and the loss function used is a weighted combination loss function combining absolute value error and mean square error; Level 3 Cell Fault Prediction Sub-model Training is performed using only a training dataset representing the final stage of cell aging, and the loss function used for training is an exponentially weighted loss function.

[0023] Preferably, in step S40, the lightweight three-class classifier includes an input layer for receiving an input feature vector; a hidden layer for extracting intermediate nonlinear features related to the classification of the aging stage; and an output layer for constructing a classification probability vector using a Softmax activation function and outputting the prediction probability distribution of the current prediction stage.

[0024] Preferably, step S50, which involves using a dynamic residual evaluation mechanism combining LSTM network and stage perception based on the capacity prediction result set to perform the cell fault early warning judgment task and output the cell early capacity anomaly alarm result set, specifically includes:

[0025] Step S501: Obtain the current prediction stage s from the capacity prediction result set; for the current prediction stage s, obtain the capacity prediction result sequence for T consecutive historical time periods; obtain the capacity reference value sequence for T consecutive historical time periods; construct the mean square residual term and the trend derivative term based on the capacity prediction result sequence and the capacity reference value sequence, and introduce the trend compensation coefficient. Based on the mean square residual term, trend derivative term, and trend compensation coefficient The residual sequence is calculated using a method that combines mean squared residuals and trend derivatives. ;

[0026] Step S502: Based on the residual sequence Capacity residual trend modeling is performed using an autoregressive method based on LSTM residual time series forecasting, and the output residual forecast sequence is generated. ;

[0027] Step S503: Predict the sequence based on the residual The task of capacity anomaly early warning and discrimination is performed with the preset residual threshold. When the residual prediction sequence... When the residual value at any future moment exceeds the residual threshold, an alarm is triggered, and a set of early capacity anomaly alarm results for the battery cell is output. The set of early capacity anomaly alarm results for the battery cell is in the form of an alarm triplet. The alarm triplet includes the warning trigger time, which is used to indicate the prediction starting point for the capacity anomaly to occur; the predicted anomaly stage, which is used to indicate the future stage number predicted to have a capacity anomaly; and the anomaly severity index, which is used to indicate the severity of the anomaly.

[0028] This invention also provides a battery failure prediction system based on a graded health assessment model, comprising:

[0029] The data acquisition module is used to obtain the historical operating data set of the target battery cell. and initial cell capacity Running historical data sets This includes the input feature vector x and the actual cell capacity c;

[0030] The phase division and labeling module is used to define the phases based on historical runtime data sets. and initial cell capacity A weakly supervised aging stage segmentation labeling mechanism based on multi-source feature enhancement is used to perform the aging stage segmentation task, and a training dataset with stage labels is output. ;

[0031] A multi-stage prediction model building module for use with a training dataset labeled with stage numbers. A three-level cell fault prediction model is constructed; wherein, the three-level cell fault prediction model includes mutually independent first-level cell fault prediction sub-models. Second-level cell fault prediction sub-model and the third-level cell fault prediction sub-model ;

[0032] The stage identification and prediction module is used to pre-construct a lightweight three-class classifier S(x), apply the lightweight three-class classifier S(x) to perform stage label prediction processing on the input feature vector x, and output the current prediction stage s; input the input feature vector x and the current prediction stage s to the three-level cell fault prediction model, and the three-level cell fault prediction model outputs a set of capacity prediction results for each level of cell fault prediction sub-model;

[0033] The dynamic residual assessment and early warning module is used to perform cell fault early warning judgment tasks based on the capacity prediction result set using a dynamic residual assessment mechanism that combines LSTM network and stage perception, and outputs a set of early capacity anomaly alarm results for cells.

[0034] The present invention also provides a battery failure prediction device based on a graded health assessment model, comprising: a memory, a processor, and a battery failure prediction program based on the graded health assessment model stored in the memory and executable on the processor. When the battery failure prediction program based on the graded health assessment model is executed by the processor, a battery failure prediction method based on the graded health assessment model is implemented.

[0035] The present invention also provides a computer program product, including a battery failure prediction program based on a graded health assessment model, wherein the battery failure prediction program based on the graded health assessment model implements the battery failure prediction method based on the graded health assessment model when executed by a processor.

[0036] The beneficial effects of this invention are as follows: By introducing aging stage labels and a stage perception mechanism, this invention constructs a three-level cell capacity prediction sub-model adapted to different aging stages. Combined with a lightweight identifier, it quickly classifies and calls the current stage, enabling the prediction model to have targeted feature learning capabilities at different stages. This effectively avoids the problem of decreased prediction accuracy of traditional unified models in the rapid capacity decay or stable stage, and improves the model's segmented response capability and prediction accuracy to cell degradation trends.

[0037] This invention combines current prediction phase information with a Long Short-Term Memory (LSTM) network to fuse the capacity prediction sequence and reference capacity to construct a trend residual evaluation sequence. It also introduces a joint analysis mechanism of mean square residual and trend derivative to dynamically capture abnormal capacity changes. This mechanism can identify abrupt changes in degradation trends in advance, significantly improving the ability to detect early-stage capacity anomalies in battery cells, thereby achieving earlier and more accurate capacity fault warnings. Attached Figure Description

[0038] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0039] Figure 1 This is a flowchart illustrating the first embodiment of a battery fault prediction method based on a graded health assessment model according to the present invention.

[0040] Figure 2 This is a schematic diagram of clustering using multi-source feature enhancement combined with a weakly supervised staging mechanism, representing a first embodiment of a battery fault prediction method based on a hierarchical health assessment model according to the present invention.

[0041] Figure 3 This is a schematic diagram of the traditional KMeans clustering method in the first embodiment of the battery fault prediction method based on a hierarchical health assessment model of the present invention.

[0042] Figure 4 This is a schematic diagram of a device for predicting battery failures based on a graded health assessment model, according to the present invention. Detailed Implementation

[0043] 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.

[0044] Example 1: As Figure 1 The diagram shown is a flowchart of the first embodiment of the battery fault prediction method based on the hierarchical health assessment model of the present invention, which presents the first embodiment of the battery fault prediction method based on the hierarchical health assessment model of the present invention.

[0045] In the first embodiment, the battery failure prediction method based on a graded health assessment model includes:

[0046] Step S10: Obtain the set of historical operating data for the target battery cell. and initial cell capacity Running historical data sets This includes the input feature vector x and the actual cell capacity c;

[0047] It should be noted that the "historical operational data set" refers to multi-source time-series operational information continuously collected and preprocessed during the actual service life of the battery cell. This set includes not only raw sensor sampling signals, such as operating voltage, current, temperature, charge / discharge time, and ambient temperature, but also statistical features obtained after data cleaning, time-window moving average, and normalization based on these raw signals, such as charge / discharge capacity change rate, cycle efficiency, internal resistance increment, SOC offset, and temperature rise rate. By constructing an input feature vector x, time-series variables from different sources can be transformed into a unified multi-dimensional feature representation to capture the dynamic changes of the battery cell at different operational stages; the actual battery cell capacity c is a capacity benchmark obtained from experimental measurement or online estimation, used to reflect the current aging level and provide a realistic reference for the supervision signal in model training.

[0048] Understandably, by acquiring historical datasets and merging them to construct a multi-dimensional input feature vector x, the dimensionality and expressive power of the input information for subsequent AI models can be significantly enhanced. Traditional single voltage or current inputs can only reflect the instantaneous operating state and cannot reveal the potential trend of cell capacity degradation; while the multi-source feature fusion input adopted in this step not only retains the dynamic characteristics of the original data but also integrates slow variables reflecting the aging process, thereby improving the model's adaptability under different operating states and aging stages, making subsequent stage division and capacity prediction more stable and reliable.

[0049] For example, when conducting lifespan testing on a certain type of lithium iron phosphate battery cell, the first 500 charge-discharge cycles were selected as the data acquisition phase. An embedded BMS system recorded signals such as voltage, current, temperature, state of charge (SOC), and cycle capacity. The input feature vector x, constructed after feature extraction, contains 12 feature components, including average charging voltage, charging current variance, temperature rise rate, discharge internal resistance change rate, and cumulative charge-discharge cycles. Comparative experiments show that when only voltage and current signals are used for capacity prediction, the model's prediction error after 200 cycles is approximately 8%; however, after employing the feature fusion mechanism described in this invention, the error is reduced to below 3%, indicating that multi-source feature fusion significantly improves the model's learning accuracy and its ability to capture capacity degradation trends.

[0050] Step S20: Based on the historical data set and initial cell capacity A weakly supervised aging stage segmentation labeling mechanism based on multi-source feature enhancement is used to perform the aging stage segmentation task, and a training dataset with stage labels is output. ;

[0051] It should be noted that the "weakly supervised aging stage labeling mechanism based on multi-source feature enhancement" refers to automatically identifying the stage boundaries in the aging process of battery cells by utilizing the changing trends of different feature dimensions in the historical data set, through feature fusion and adaptive clustering methods, in the absence of precise manual stage labeling. This mechanism first establishes a unified feature space based on multi-source features such as voltage, current, temperature, SOC, internal resistance, and cumulative charge / discharge. Then, it performs normalization and time window smoothing on the feature sequences to eliminate differences caused by sampling frequency and sensor drift. Subsequently, it calculates the capacity decay rate and feature change slope at each time point in the feature space, and automatically detects abrupt changes in the capacity change rate using spectral clustering or density clustering algorithms, dividing the entire life cycle into several aging stages. The entire process does not require manual specification of thresholds or labeled samples; the stage division is adaptively completed through feature distribution and dynamic trends, thereby generating a training dataset with stage labels.

[0052] Understandably, by introducing multi-source feature enhancement and weakly supervised staging mechanisms, the implicit stage patterns in cell operation data can be fully utilized, transforming the traditional coarse-grained staging based on time or cycle count into a fine-grained staging structure based on feature change trends. This method can adaptively adjust stage boundaries under different aging stages, allowing the model to establish fitting relationships for the feature differences of each stage during training, thereby improving the stage adaptability of subsequent prediction models. Through this process, the sample labels in the training dataset not only reflect the degree of capacity degradation but also carry feature evolution information, providing a more discriminative data foundation for subsequent segmented model training.

[0053] It should be understood that, compared to traditional methods that divide stages based on a fixed number of cycles or capacity decay ratio, the weakly supervised stage labeling mechanism of this invention can automatically determine stage boundaries based on changes in feature distribution without manual intervention, avoiding the stage division distortion problem caused by manual threshold selection in traditional methods. Traditional stage division methods based on time or number of cycles often fail to reflect the stage transitions in the actual performance of the battery cell, and are prone to "stage drift" or "aliasing" under conditions of unbalanced samples or frequent changes in operating conditions, leading to model training bias. This invention, through multi-source feature enhancement and spectral similarity-preserving mapping technology, makes the stage division results continuous in time and separable in feature space, significantly improving the stability and physical interpretability of stage labels, and providing a reliable foundation for stage-aware prediction.

[0054] For example, during the full-life cycle testing of a batch of ternary lithium-ion power cells, the collected voltage, current, temperature, and capacity data, after feature extraction, constitute a multi-source feature matrix. Using the weakly supervised stage labeling mechanism of this invention, the algorithm automatically identifies three stage boundary points, corresponding to abrupt changes in capacity decay rate and temperature gradient changes, respectively. Experimental results show that, compared with the traditional scheme based on equidistant division by cycle count, the stage boundaries identified by this method detect performance degradation inflection points an average of 10 to 15 cycles earlier. After stage division, the feature difference between sub-samples used for model training increases by approximately 40%, verifying that this invention can identify cell aging stage changes earlier and more accurately, thereby providing higher-quality training data for subsequent segmented modeling and fault prediction.

[0055] For example, such as Figure 2 and Figure 3 As shown, under the same normalized feature space (with capacity, internal resistance, and temperature as the main axes), the traditional KMeans clustering method is used ( Figure 3 When performing clustering, the number of clusters needs to be pre-defined, and the division results are mainly based on minimizing geometric distance. This makes it difficult to accurately capture the nonlinear evolution and abrupt boundary changes during the cell aging process, resulting in significant overlap between different aging stages. The stage division exhibits rigidity and insensitivity to physical changes. Figure 2 The multi-source feature-enhanced weakly supervised aging mechanism demonstrated in the paper does not require a preset number of categories. It can automatically identify the stage boundaries in the cell aging process based on the natural aggregation trend of capacity decay rate and multi-dimensional feature change slope, and effectively distinguish the blurred transition areas as "boundary zones" or "anomalies," which is closer to the physical process of gradual performance degradation of cells under real operating conditions. The segmentation results of this mechanism are continuous in the time dimension and separable in the feature space, which significantly improves the stability, fine granularity, and physical interpretability of the aging labels, providing a more reliable data foundation for subsequent stage-aware modeling.

[0056] Step S30: Based on the training dataset with stage labels A three-level cell fault prediction model is constructed; wherein, the three-level cell fault prediction model includes mutually independent first-level cell fault prediction sub-models. Second-level cell fault prediction sub-model and the third-level cell fault prediction sub-model ;

[0057] It should be noted that the "three-level cell failure prediction model" here refers to stage-aware prediction sub-models constructed for three typical stages in the cell aging lifecycle (i.e., the initial rapid degradation stage, the mid-term stable operation stage, and the final rapid decay stage). Each sub-model is trained and inferred only within its corresponding lifecycle stage, avoiding the model learning redundant or irrelevant features throughout the entire lifecycle. Specifically, the first-level sub-model mainly fits the rate trend of capacity decline in the initial stage, the second-level sub-model models the stable process of slow decay in the mid-term stage, and the third-level sub-model focuses on the drastic changes in behavior near failure. Each sub-model can be independently optimized in terms of parameter structure, input variable combination, and learning strategy to achieve accurate fitting of stage features and improved generalization ability.

[0058] Understandably, constructing stage-specific prediction sub-models can significantly enhance the adaptability of AI models to differences in stage-specific features, avoiding the prediction accuracy loss that traditional unified models suffer when facing the heterogeneity of lifecycle features. Especially when the cell's operating state is significantly affected by external environmental factors (temperature, current fluctuations, etc.), a single model struggles to simultaneously account for the complex feature evolution of different aging stages. Adopting a stage-based modeling strategy not only improves prediction accuracy but also reduces model training difficulty, accelerates convergence, and enhances overall prediction stability and deployment reliability.

[0059] It should be understood that, compared to the single-model or hybrid model architectures commonly found in existing technologies, the three-level prediction structure proposed in this invention has greater physical correspondence and engineering practicality. Traditional methods, due to the lack of explicit modeling of stage-specific evolution, often suffer from problems such as early underestimation, mid-term stabilization, and overfitting in the later stages. This invention, however, drives independent modeling for each stage through stage labels, focusing model parameters and feature weights on the most discriminative indicators for that stage, thereby achieving a simultaneous improvement in performance consistency and prediction interpretability. Furthermore, this architecture facilitates the future introduction of stage-specific degradation mechanisms, periodic correction mechanisms, or transfer learning strategies, laying the foundation for deployment and generalization under various operating conditions.

[0060] Step S40: Pre-construct a lightweight three-class classifier S(x), apply the lightweight three-class classifier S(x) to perform stage label prediction processing on the input feature vector x, and output the current prediction stage s; input the input feature vector x and the current prediction stage s into the three-level cell fault prediction model, and the three-level cell fault prediction model outputs a set of capacity prediction results for each level of cell fault prediction sub-model;

[0061] It should be noted that the "lightweight three-class classifier S(x)" refers to a lightweight stage recognition model built for the three typical stages of battery cell aging (initial rapid decay stage, mid-term stable decay stage, and final severe decay stage). This classifier uses a classification network with a simple structure, few parameters, and fast inference speed, such as a shallow convolutional neural network (CNN) or a lightweight fully connected neural network (such as a variant of MobileNet). Its input is a feature vector x formed by the fusion of multi-source features, and its output is the stage label of the current sample. This classifier can be deployed independently on edge computing units without a complete prediction model to achieve rapid pre-judgment of battery cell stage perception.

[0062] Understandably, by introducing a lightweight stage identifier before the complete prediction process, the original features acquire stage context information before entering the three-stage cell fault prediction model. This allows for the targeted activation of different prediction sub-models (e.g., the first-stage sub-model for fitting the initial decay trend, the second-stage sub-model for predicting the stable period, and the third-stage sub-model for monitoring the final degradation stage). This effectively avoids the problem of insufficient generalization ability of a single model and improves the stage matching and inference accuracy of the prediction model. Since each sub-model focuses on the feature patterns of its own stage, its structural parameters and loss functions can also be designed differently, thereby enhancing the model's adaptability and convergence speed.

[0063] Step S50: Based on the capacity prediction result set, a dynamic residual evaluation mechanism combining LSTM network and stage perception is used to perform the cell fault early warning judgment task, and output the cell early capacity anomaly alarm result set.

[0064] It should be noted that the "dynamic residual evaluation mechanism based on LSTM network and stage awareness" in this step refers to jointly encoding the time series structure of cell operation data with aging stage labels, and constructing a Long Short-Term Memory (LSTM) neural network to model the dynamic residual change trend between the predicted capacity and the actual capacity. During the model training phase, the LSTM network not only receives the current feature vector of the cell and the corresponding predicted capacity, but also integrates the aging stage information to which that time point belongs, in order to capture the residual evolution patterns within and between stages.

[0065] Understandably, introducing dynamic residual sequence analysis can effectively overcome the problem of insufficient perception of early capacity faults by the static error threshold method. The LSTM structure is good at capturing long-term dependencies, enabling it to extract hidden abnormal trends in historical residual patterns, and improve the model's sensitivity to error patterns at different aging stages through a stage-aware mechanism.

[0066] It should be understood that traditional fault warning methods often employ fixed threshold rules, such as "an alarm is triggered when the deviation between predicted capacity and actual capacity exceeds a set value." However, this method has the following drawbacks: it ignores the differences in capacity change rates at different aging stages, resulting in overly coarse threshold settings; and in the early decay stage, due to the slow capacity decline, the fixed threshold may not trigger an alarm at all. This invention, by constructing an LSTM residual model that integrates stage information, can adaptively adjust the sensing sensitivity at each stage, significantly improving the accuracy and lead time of warnings, especially demonstrating superior performance in the early degradation stage.

[0067] For example, in a set of measured lithium battery degradation data, the traditional static threshold method only triggers an alarm when the capacity drops below 80% of the initial capacity. However, the LSTM-stage residual mechanism in this invention detects abnormal risks through residual trend drift even when the capacity remains at 85%, achieving an early warning lead time of up to 70 cycles. This significantly improves the response window before cell failure, providing more time for scheduling and replacement strategies. This mechanism demonstrates superior ROC-AUC and average early detection rate compared to traditional methods on multiple public datasets (such as NASA-Battery and CALCE), exhibiting good generalization ability and engineering feasibility.

[0068] Example 2: Furthermore, the present invention provides a battery fault prediction system based on a hierarchical health assessment model, employing a battery fault prediction method based on a hierarchical health assessment model as described in the above embodiments, which can solve the technical problem of battery fault prediction based on a hierarchical health assessment model. Compared with the prior art, the beneficial effects of the battery fault prediction system based on a hierarchical health assessment model provided by the present invention are the same as the beneficial effects of the battery fault prediction method based on a hierarchical health assessment model provided in the above embodiments, and other technical features of the battery fault prediction system based on a hierarchical health assessment model are the same as those disclosed in the methods of the above embodiments, and will not be repeated here.

[0069] Example 3: This invention provides a battery fault prediction device based on a graded health assessment model. Please refer to... Figure 4A battery failure prediction device based on a graded health assessment model includes: at least one processor; and a memory communicatively connected to the at least one processor; wherein the memory stores instructions executable by the at least one processor, which, when executed by the at least one processor, enable the at least one processor to perform a battery failure prediction method based on a graded health assessment model as described in Embodiment 1 above. The battery failure prediction device based on a graded health assessment model in this embodiment of the invention may include, but is not limited to, mobile terminals such as mobile phones, laptops, digital radio receivers, PDAs (Personal Digital Assistants), PADs (Portable Application Description), PMPs (Portable Media Players), in-vehicle terminals (e.g., in-vehicle navigation terminals), and fixed terminals such as digital TVs and desktop computers. This battery failure prediction device based on a graded health assessment model is merely an example and should not impose any limitations on the functionality and scope of use of the embodiments of the invention. A battery failure prediction device based on a graded health assessment model may include a processing unit 1001 (e.g., a central processing unit, a graphics processing unit, etc.) that can perform various appropriate actions and processes according to a program stored in a read-only memory 1002 or a program loaded from a storage device 1003 into a random access memory 1004. The random access memory 1004 also stores various programs and data required for the operation of the battery failure prediction device based on the graded health assessment model. The processing unit 1001, the read-only memory 1002, and the random access memory 1004 are interconnected via a bus 1005. An I / O interface 1006 is also connected to the bus. Typically, the following systems can be connected to the I / O interface 1006: input devices 1007 including, for example, a touchscreen, touchpad, keyboard, mouse, image sensor, microphone, accelerometer, gyroscope, etc.; output devices 1008 including, for example, a liquid crystal display (LCD), speaker, vibrator, etc.; storage devices 1003 including, for example, magnetic tape, hard disk, etc.; and communication devices 1009. Communication device 1009 allows a battery failure prediction device based on a graded health assessment model to communicate wirelessly or wiredly with other devices to exchange data. While the figure shows one battery failure prediction device based on a graded health assessment model with various systems, it should be understood that implementation or possession of all the systems shown is not required. More or fewer systems may be implemented alternatively.

[0070] Example 4: This invention also provides a computer program product, including a computer program that, when executed by a processor, implements the steps of the battery fault prediction method based on a graded health assessment model as described above. The computer program product provided by this invention can solve a technical problem related to battery fault prediction based on a graded health assessment model. Compared with the prior art, the beneficial effects of the computer program product provided by this invention are the same as those of the battery fault prediction method based on a graded health assessment model provided in the above embodiments, and will not be repeated here.

[0071] In particular, according to the embodiments disclosed in this invention, the processes described above with reference to the flowcharts can be implemented as computer software programs. For example, embodiments of this invention include a computer program product comprising a computer program carried on a computer-readable medium, the computer program containing program code for performing the methods shown in the flowcharts. In such embodiments, the computer program can be downloaded and installed from a network via a communication device, or installed from storage device 1003, or installed from read-only memory 1002. When the computer program is executed by processing device 1001, it performs the functions defined in the methods of the embodiments disclosed in this invention.

[0072] It should be understood that the various parts disclosed in this invention can be implemented using hardware, software, firmware, or a combination thereof. In the description of the above embodiments, specific features, structures, materials, or characteristics can be combined in any suitable manner in one or more embodiments or examples.

[0073] Obviously, those skilled in the art can make various modifications and variations to this invention without departing from its spirit and scope. Therefore, if these modifications and variations fall within the scope of the claims of this invention and their equivalents, this invention also intends to include these modifications and variations.

Claims

1. A battery fault prediction method based on a graded health assessment model, characterized in that, The methods include: Step S10: Obtain the set of historical operating data for the target battery cell. and initial cell capacity Running historical data sets This includes the input feature vector x and the actual cell capacity c; Step S20: Based on the historical data set and initial cell capacity A weakly supervised aging stage segmentation labeling mechanism based on multi-source feature enhancement is used to perform the aging stage segmentation task, and a training dataset with stage labels is output. ; Step S30: Based on the training dataset with stage labels A three-level cell fault prediction model is constructed; wherein, the three-level cell fault prediction model includes mutually independent first-level cell fault prediction sub-models. Second-level cell fault prediction sub-model and the third-level cell fault prediction sub-model ; Step S40: Pre-construct a lightweight three-class classifier S(x), apply the lightweight three-class classifier S(x) to perform stage label prediction processing on the input feature vector x, and output the current prediction stage s; input the input feature vector x and the current prediction stage s into the three-level cell fault prediction model, and the three-level cell fault prediction model outputs a set of capacity prediction results for each level of cell fault prediction sub-model; Step S50: Based on the capacity prediction result set, a dynamic residual evaluation mechanism combining LSTM network and stage perception is used to perform the cell fault early warning judgment task, and output the cell early capacity anomaly alarm result set.

2. The battery fault prediction method based on a graded health assessment model as described in claim 1, characterized in that, In step S10, the input feature vector x includes voltage feature vector, current feature vector, temperature feature vector, SOC feature vector, cumulative charge / discharge feature vector, and charge / discharge rate feature vector.

3. The battery fault prediction method based on a graded health assessment model as described in claim 1, characterized in that, In step S20, based on the historical data set... and initial cell capacity A weakly supervised aging stage segmentation labeling mechanism based on multi-source feature enhancement is used to perform the aging stage segmentation task, and a training dataset with stage labels is output. The steps specifically include: Step S201: From the historical data set Extract the actual cell capacity c from the data, and then determine the cell capacity c and the initial cell capacity. Calculate the relative capacity decay rate , Based on the relative capacity decay rate Construct an aging index feature vector, and add the aging index feature vector to the input feature vector x to build an enhanced feature vector. Subsequently, the enhanced feature vectors were processed. Perform normalization processing and output the normalized enhanced feature vector. ; Step S202: Optimize the normalized enhanced feature vector A nonlinear feature embedding transformation is performed using an adaptive kernel width Gaussian kernel mapping mechanism to obtain a spectral-preserving similarity matrix K; a low-dimensional embedding representation vector set Z is obtained by performing principal component analysis on the spectral-preserving similarity matrix K. Step S203: Based on the low-dimensional embedding representation vector set Z, a weakly supervised spectral clustering task is performed, ultimately outputting a training dataset with stage labels. .

4. The battery fault prediction method based on a graded health assessment model as described in claim 3, characterized in that, In step S203, the task of weakly supervised spectral clustering based on the low-dimensional embedding representation vector set Z finally outputs a training dataset with stage labels. The steps specifically include: First, construct a weighted similarity matrix based on the set of low-dimensional embedding representation vectors Z. Weighted similarity matrix Used to jointly characterize the spectral distance similarity and aging trend similarity between samples; Based on the weighted similarity matrix The normalized Laplacian matrix is ​​calculated using the normalized graphical Laplacian spectral clustering method. , Where D is the angle matrix, used to ensure the symmetry normalization of the feature space; further, the normalized Laplace matrix is ​​normalized... Feature decomposition is performed to extract the feature vector set U corresponding to the k smallest feature values. Then, the feature vector set U is subjected to K-means clustering algorithm for preliminary stage division processing to obtain the preliminary label set S. The preliminary label set S includes a first preliminary stage cluster, a second preliminary stage cluster, and a third preliminary stage cluster. The first preliminary stage cluster is used to represent the initial stage of cell aging, the second preliminary stage cluster is used to represent the middle stage of cell aging, and the third preliminary stage cluster is used to represent the final stage of cell aging. Further introduce relative capacity decay rate Based on relative capacity decay rate Perform interval mapping on the initial label set S to obtain unsupervised clustering labels. Further based on unsupervised clustering labels By using minimum deviation mapping, unsupervised clustering labels are... The calibration process yields final stage labels with physical semantic meaning, resulting in a training dataset with stage labels as the final output. .

5. The battery fault prediction method based on a graded health assessment model as described in claim 1, characterized in that, In step S30, the first-level cell fault prediction sub-model The training is performed using only a training dataset representing the initial stage of cell aging, and the loss function used is the standard mean squared error loss function; the second-level cell fault prediction sub-model The training is performed using only a training dataset representing the mid-stage of cell aging, and the loss function used is a weighted combination loss function combining absolute value error and mean square error; Level 3 Cell Fault Prediction Sub-model Training is performed using only a training dataset representing the final stage of cell aging, and the loss function used for training is an exponentially weighted loss function.

6. The battery fault prediction method based on a graded health assessment model as described in claim 1, characterized in that, In step S40, the lightweight three-class classifier includes an input layer for receiving input feature vectors; a hidden layer for extracting intermediate nonlinear features related to the classification of aging stages; and an output layer for constructing a classification probability vector using the Softmax activation function and outputting the prediction probability distribution of the current prediction stage.

7. The battery fault prediction method based on a graded health assessment model as described in claim 1, characterized in that, Step S50, which involves using a dynamic residual evaluation mechanism combining LSTM network and stage perception based on the capacity prediction result set to perform the cell fault early warning judgment task and output the cell early capacity anomaly alarm result set, specifically includes: Step S501: Obtain the current prediction stage s from the capacity prediction result set; for the current prediction stage s, obtain the capacity prediction result sequence for T consecutive historical time periods; obtain the capacity reference value sequence for T consecutive historical time periods; construct the mean square residual term and the trend derivative term based on the capacity prediction result sequence and the capacity reference value sequence, and introduce the trend compensation coefficient. Based on the mean square residual term, trend derivative term, and trend compensation coefficient The residual sequence is calculated using a method that combines mean squared residuals and trend derivatives. ; Step S502: Based on the residual sequence Capacity residual trend modeling is performed using an autoregressive method based on LSTM residual time series forecasting, and the output residual forecast sequence is generated. ; Step S503: Predict the sequence based on the residual The task of capacity anomaly early warning and discrimination is performed with the preset residual threshold. When the residual prediction sequence... When the residual value at any future moment exceeds the residual threshold, an alarm is triggered, and a set of early capacity anomaly alarm results for the battery cell is output. The set of early capacity anomaly alarm results for the battery cell is in the form of an alarm triplet. The alarm triplet includes the warning trigger time, which is used to indicate the prediction starting point for the capacity anomaly to occur; the predicted anomaly stage, which is used to indicate the future stage number predicted to have a capacity anomaly; and the anomaly severity index, which is used to indicate the severity of the anomaly.

8. A battery fault prediction system based on a graded health assessment model, applied to the battery fault prediction method based on a graded health assessment model as described in any one of claims 1 to 7, characterized in that, The battery failure prediction system based on the hierarchical health assessment model includes: The data acquisition module is used to obtain the historical operating data set of the target battery cell. and initial cell capacity Running historical data sets This includes the input feature vector x and the actual cell capacity c; The phase division and labeling module is used to define the phases based on historical runtime data sets. and initial cell capacity A weakly supervised aging stage segmentation labeling mechanism based on multi-source feature enhancement is used to perform the aging stage segmentation task, and a training dataset with stage labels is output. ; A multi-stage prediction model building module for use with a training dataset labeled with stage numbers. A three-level cell fault prediction model is constructed; wherein, the three-level cell fault prediction model includes mutually independent first-level cell fault prediction sub-models. Second-level cell fault prediction sub-model and the third-level cell fault prediction sub-model ; The stage identification and prediction module is used to pre-construct a lightweight three-class classifier S(x), apply the lightweight three-class classifier S(x) to perform stage label prediction processing on the input feature vector x, and output the current prediction stage s; input the input feature vector x and the current prediction stage s to the three-level cell fault prediction model, and the three-level cell fault prediction model outputs a set of capacity prediction results for each level of cell fault prediction sub-model; The dynamic residual assessment and early warning module is used to perform cell fault early warning judgment tasks based on the capacity prediction result set using a dynamic residual assessment mechanism that combines LSTM network and stage perception, and outputs a set of early capacity anomaly alarm results for cells.

9. A battery fault prediction device based on a graded health assessment model, characterized in that, The battery failure prediction device based on the graded health assessment model includes: a memory, a processor, and a battery failure prediction program based on the graded health assessment model stored in the memory and executable on the processor. When the battery failure prediction program based on the graded health assessment model is executed by the processor, it implements a battery failure prediction method based on the graded health assessment model according to any one of claims 1 to 7.

10. A computer program product, characterized in that, The computer program product includes a battery failure prediction program based on a graded health assessment model, which, when executed by a processor, implements a battery failure prediction method based on a graded health assessment model as described in any one of claims 1 to 7.

Images