A mechanism-data fusion lithium ion battery safety early warning method and system

Abstract

This invention discloses a mechanism-data fusion method and system for early warning of lithium-ion battery safety. The method includes: retrieving historical multi-source time-series operational data of lithium-ion batteries through a cloud platform; extracting battery safety features by combining knowledge reasoning, electrochemical mechanism model prediction, and data-driven methods to construct a battery safety feature cluster; using Pearson correlation analysis and a weighted Relief-F importance selection algorithm to screen the relevance and rank the importance of the battery safety feature cluster, obtaining a simplified key feature cluster; building an integrated learning framework that integrates attention mechanism and multi-core extreme learning machine Adaboost, using the simplified key feature cluster as input samples to train a battery safety risk identification model; inputting the corresponding multi-source time-series operational data of lithium-ion batteries collected in real time from the cloud platform into the battery safety risk identification model, outputting the lithium-ion battery safety risk status judgment result, and realizing battery safety risk early warning and fault diagnosis.

Description

Technical Field

[0001] This invention relates to the field of electrochemical energy storage safety technology, and in particular to a mechanism-data fusion method and system for early warning of lithium-ion battery safety, applicable to lithium-ion battery management systems such as power batteries and energy storage batteries. Background Technology

[0002] Lithium-ion batteries are prone to thermal runaway under abusive or aging conditions, exhibiting characteristics such as high insidiousness, rapid development, and significant hazards. Developing timely and accurate safety early warning technologies is a core challenge and key prerequisite for achieving proactive safety protection in lithium-ion battery management systems. Currently, mainstream methods for lithium-ion battery safety early warning can be categorized into two types: purely mechanistic methods and purely data-driven methods. Purely mechanistic methods suffer from delayed warnings and are insensitive to early-stage faults; while purely data-driven methods are limited by data completeness, resulting in insufficient generalization ability and interpretability under unknown operating conditions.

[0003] The evolution of early-stage potential hazards into thermal runaway failures within lithium-ion batteries is typically accompanied by a gradual evolution of various side reactions and state parameters. Purely mechanistic or purely data-driven methods have inherent limitations in describing this complex process. Therefore, a novel lithium-ion battery safety early warning system is urgently needed to achieve earlier, more reliable, and more interpretable identification and warning of potential risks. Summary of the Invention

[0004] This invention addresses the problems of delayed early warning, insensitivity to early faults, weak generalization ability, and poor interpretability in existing technologies. It proposes a mechanism-data fusion-based lithium-ion battery safety early warning method. By integrating prior knowledge of electrochemical mechanisms with multi-source time-series operational data, this method achieves early, reliable, and interpretable identification and warning of potential risks to lithium-ion batteries. This invention also relates to a mechanism-data fusion-based lithium-ion battery safety early warning system.

[0005] The technical solution of the present invention is as follows:

[0006] A mechanism-data fusion-based method for early warning of lithium-ion battery safety, characterized by comprising the following steps:

[0007] S1. By retrieving historical multi-source time-series operating data of lithium-ion batteries, including voltage, current, and surface temperature, through the cloud platform, and jointly employing a knowledge reasoning method based on historical safety data, a battery state prediction method based on an electrochemical pseudo-two-dimensional electro-thermal coupling mechanism model, and a data-driven feature extraction method based on a time convolutional network, battery safety features are extracted from the historical multi-source time-series operating data, and duplicate features are removed to construct a battery safety feature cluster.

[0008] S2. The battery safety feature cluster is screened for correlation using Pearson correlation analysis; the feature contribution rate of the screened features is calculated and the importance is sorted by weighted Relief-F importance selection algorithm. The cumulative contribution rate of the feature weights is calculated by accumulating them in descending order of importance. The feature corresponding to the cumulative contribution rate reaching the preset threshold is selected to form a key feature cluster after dimensional simplification.

[0009] S3. Construct an integrated learning framework that combines an attention mechanism with the multi-core extreme learning machine Adaboost. The attention mechanism is used to adaptively and differentially weightedly fuse features from different sources obtained by the knowledge reasoning method, the mechanism model battery state prediction method, and the temporal convolutional network data-driven feature extraction method in the dimensionally simplified key feature cluster, taking into account the feature source, the real-time identification scenario, and the contribution differences of each feature. The dimensionally simplified key feature cluster is used as input samples to train the integrated learning framework to obtain a battery safety risk identification model with state fitting performance and feature weight parsing ability.

[0010] S4. Input the corresponding multi-source time-series operation data of lithium-ion batteries collected in real time by the cloud platform into the battery safety risk identification model, and output the lithium-ion battery safety risk status judgment result to realize battery safety risk warning and fault diagnosis.

[0011] Preferably, in step S1, the knowledge reasoning method based on historical safety data obtains historical safety data from retrieved historical multi-source time-series operational data. The historical safety data includes historical safety accident data of lithium-ion batteries, historical thermal runaway experimental data of lithium-ion batteries, and historical vehicle recall record data. Based on the historical safety data, a correspondence between voltage parameters, current parameters, and temperature parameters and battery safety risks is established. According to the correspondence, the voltage is compared with voltage threshold, current with current threshold, and surface temperature with temperature threshold in the historical multi-source time-series operational data, respectively, to extract safety features related to abnormal voltage fluctuations, over-limit current, and over-limit temperature, forming the first type of safety features.

[0012] Preferably, in step S1, the battery state prediction method based on the electrochemical pseudo-two-dimensional electro-thermal coupling mechanism model involves inputting historical multi-source time-series operating data into the electrochemical pseudo-two-dimensional electro-thermal coupling mechanism model. The electrochemical pseudo-two-dimensional electro-thermal coupling mechanism model calculates and outputs the state of charge, state of health, battery internal temperature, ohmic internal resistance, and polarization overpotential characteristics of each lithium-ion cell. The output state of charge, state of health, battery internal temperature, ohmic internal resistance, and polarization overpotential characteristics are then subjected to secondary analysis to generate thermal runaway risk warning characteristics, cell inconsistency anomaly characteristics, and abnormal internal resistance growth characteristics, forming a second type of safety characteristics.

[0013] Preferably, in step S1, the data-driven feature extraction method based on temporal convolutional networks involves temporally organizing the retrieved historical multi-source time-series running data to construct an input sequence based on a time window; inputting the input sequence into a temporal convolutional network to extract temporal features through multi-layer temporal convolution and residual connections; and performing pooling and fully connected mapping processing on the features output by the temporal convolutional network to obtain the third type of security features.

[0014] The third type of safety features are summarized with the first and second types of safety features, and features with the same or repeated meanings are removed or merged to construct a battery safety feature cluster.

[0015] Preferably, in step S2, the weighted Relief-F importance selection algorithm is used to calculate the feature contribution rate and rank the importance of the selected features. Specifically, this includes: configuring initial weights for each selected feature; randomly selecting samples as anchor points and using Euclidean distance to search for nearest neighbors among samples of the same class and samples of different classes; adjusting feature weights according to the feature's ability to distinguish samples of different classes; iterating the sampling and weight update process multiple times and taking the average weight after multiple iterations; calculating the feature contribution rate and ranking the importance of the selected features based on the average weight; and then normalizing the ranked feature weights to facilitate the subsequent calculation of the cumulative contribution rate of the normalized feature weights in descending order of importance.

[0016] Preferably, in step S3, an integrated learning framework combining the attention mechanism and the multi-core extreme learning machine Adaboost is built, specifically including:

[0017] Based on the attention mechanism, the key feature clusters obtained in step S2 are fused into multiple features, including linear mapping of each group of features in the key feature cluster to construct query vectors and key vectors; based on the query vectors and key vectors, the matching score is calculated by additive attention and normalized by Softmax to obtain attention weights, and then the fused input feature vectors are obtained by weighting based on the attention weights.

[0018] A multi-core extreme learning machine is constructed as a weak learner. The multi-core extreme learning machine replaces the basis function of the traditional extreme learning machine with a kernel function that is a weighted combination of the exponential function and the sigmoid function. By associating the kernel function weights with the adaptation requirements of the ensemble learning framework, the kernel function weights are adaptively updated according to the attention weights of the features from different sources obtained by the knowledge reasoning method, the mechanism model battery state prediction method, and the temporal convolutional network data-driven feature extraction method in the key feature cluster after dimensional simplification. This simultaneously optimizes the kernel function adaptability and the overall compatibility of the ensemble learning framework, achieving dual-objective optimization.

[0019] An Adaboost integration module was built as a strong learner, forming an integrated learning framework that combines attention mechanism, multi-core extreme learning machine and Adaboost integration module.

[0020] Preferably, in step S3, the simplified key feature clusters are used as input samples and fed into the established ensemble learning framework for training, specifically including:

[0021] The multi-core extreme learning machine built into the integrated learning framework is used as the basic learning unit. The Adaboost algorithm is used to complete the iterative training integration based on the Adaboost integration module. The iterative training process includes: initializing the input sample weight distribution, completing the sample feature fitting based on the basic learning unit and calculating the fitting deviation, determining the corresponding weight of the basic learning unit based on the fitting deviation, synchronously updating the sample weight distribution and repeatedly executing the iterative operation until the preset training termination condition is met.

[0022] An objective quantitative assignment method is used to construct a battery safety risk assessment index. This assessment index is then integrated into the training process to optimize the model parameters. After training, a battery safety risk identification model with both state fitting ability and feature weight analysis ability is obtained.

[0023] A mechanism-data fusion lithium-ion battery safety early warning system, characterized in that it comprises, in sequence, a mechanism-data fusion battery safety feature cluster construction module, a Pearson joint weighted Relief-F feature dimension simplification module, an attention fusion multi-core extreme learning machine integrated modeling module, and a battery safety risk intelligent identification and early warning execution module, wherein...

[0024] The mechanism-data fusion battery safety feature cluster construction module retrieves historical multi-source time-series operational data of lithium-ion batteries, including voltage, current, and surface temperature, through a cloud platform. It then employs a knowledge reasoning method based on historical safety data, a battery state prediction method based on an electrochemical pseudo-two-dimensional electro-thermal coupling mechanism model, and a data-driven feature extraction method based on a time convolutional network to extract battery safety features from the historical multi-source time-series operational data. After removing duplicate features, it constructs a battery safety feature cluster.

[0025] The Pearson-weighted Relief-F feature dimensionality simplification module uses Pearson correlation analysis to screen the battery safety feature cluster for correlation; it uses the weighted Relief-F importance selection algorithm to calculate the feature contribution rate of the screened features and sort them by importance, and calculates the cumulative contribution rate of the feature weights in descending order of importance, and selects the features corresponding to the cumulative contribution rate that reaches a preset threshold to form the key feature cluster after dimensional simplification.

[0026] The attention-fusion multi-core extreme learning machine integrated modeling module builds an integrated learning framework that integrates the attention mechanism and the multi-core extreme learning machine - Adaboost. The attention mechanism is used to adaptively differentiate and weightedly fuse features from different sources obtained by the knowledge reasoning method, the mechanism model battery state prediction method, and the temporal convolutional network data-driven feature extraction method in the dimensionally simplified key feature cluster, taking into account the feature source, the real-time identification scenario, and the contribution differences of each feature. The dimensionally simplified key feature cluster is used as input samples to train the integrated learning framework to obtain a battery safety risk identification model with state fitting performance and feature weight parsing ability.

[0027] The battery safety risk intelligent identification and early warning execution module inputs the corresponding multi-source time-series operation data of lithium-ion batteries collected in real time by the cloud platform into the battery safety risk identification model, and outputs the lithium-ion battery safety risk status judgment result, thereby realizing battery safety risk early warning and fault diagnosis.

[0028] Preferably, in the mechanism-data fusion battery safety feature cluster construction module, the knowledge reasoning method based on historical safety data obtains historical safety data from retrieved historical multi-source time-series operational data. This historical safety data includes historical safety accident data for lithium-ion batteries, historical thermal runaway experimental data for lithium-ion batteries, and historical vehicle recall records. Based on this historical safety data, a correspondence is established between voltage parameters, current parameters, and temperature parameters and battery safety risks. According to this correspondence, the voltage is compared with voltage thresholds, current with current thresholds, and surface temperature with temperature thresholds in the historical multi-source time-series operational data to extract safety features related to abnormal voltage fluctuations, over-limit currents, and over-limit temperatures, forming the first type of safety features.

[0029] The battery state prediction method based on the electrochemical pseudo-two-dimensional electro-thermal coupling mechanism model involves inputting historical multi-source time-series operational data into the electrochemical pseudo-two-dimensional electro-thermal coupling mechanism model. The electrochemical pseudo-two-dimensional electro-thermal coupling mechanism model calculates and outputs the state of charge, state of health, internal temperature, ohmic internal resistance, and polarization overpotential characteristics of each lithium-ion cell. The output state of charge, state of health, internal temperature, ohmic internal resistance, and polarization overpotential characteristics are then subjected to secondary analysis to generate thermal runaway risk warning characteristics, cell inconsistency anomaly characteristics, and abnormal internal resistance growth characteristics, forming a second type of safety characteristics.

[0030] The data-driven feature extraction method based on temporal convolutional networks involves temporally organizing historical multi-source time-series running data to construct an input sequence based on a time window; inputting the input sequence into a temporal convolutional network to extract temporal features through multiple layers of temporal convolution and residual connections; and performing pooling and fully connected mapping on the features output by the temporal convolutional network to obtain the third type of security features.

[0031] The third type of safety features are summarized with the first and second types of safety features, and features with the same or repeated meanings are removed or merged to construct a battery safety feature cluster.

[0032] Preferably, in the Pearson joint weighted Relief-F feature dimension simplification module, the weighted Relief-F importance selection algorithm is used to calculate the feature contribution rate and rank the importance of the selected features. Specifically, this includes: configuring initial weights for each selected feature; randomly selecting samples as anchor points and using Euclidean distance to search for nearest neighbors among samples of the same class and samples of different classes; adjusting feature weights according to the feature's ability to distinguish samples of different classes; iterating the sampling and weight update process multiple times and taking the average weight after multiple iterations; calculating the feature contribution rate and ranking the importance of the selected features based on the average weight; and then normalizing the ranked feature weights to facilitate the subsequent accumulation of the normalized feature weights in descending order of importance to calculate the cumulative contribution rate.

[0033] The technical effects of this invention are as follows:

[0034] This invention relates to a mechanism-data fusion-based lithium-ion battery safety early warning method. In step S1, it combines a knowledge reasoning method based on historical safety data, a battery state prediction method based on an electrochemical pseudo-two-dimensional electro-thermal coupling mechanism model, and a data-driven feature extraction method based on a temporal convolutional network. This approach takes into account three dimensions: historical experience, intrinsic mechanism, and real-time data. The knowledge reasoning method relies on historical safety data to ensure that the features match the actual safety early warning requirements. The electrochemical pseudo-two-dimensional electro-thermal coupling mechanism model deeply explores the internal electro-thermal coupling characteristics of the battery and extracts core features reflecting the battery's intrinsic safety state (such as internal temperature distribution, electrode reaction rate, etc.), thus overcoming the shortcomings of single data-driven methods that only look at the surface and not the essence. The temporal convolutional network (TCN) has a powerful ability to capture temporal features and can accurately extract dynamic change features from multi-source temporal operating data (voltage, current, surface temperature), adapting to the temporal characteristics of battery operating data. The integration of these three elements ensures that the extracted battery safety features encompass both surface operating conditions and internal mechanisms, are supported by historical experience, and are adaptable to real-time operating conditions. This completely resolves the issues of one-sided and poorly targeted feature extraction in existing technologies. By eliminating duplicate features, feature redundancy is effectively reduced, avoiding problems such as increased computational load, low training efficiency, and model overfitting caused by data redundancy during subsequent correlation screening and model training. This lays an efficient and reliable feature foundation for feature screening in step S2 and model training in step S3, improving the overall efficiency of the early warning method. The resulting battery safety feature cluster integrates multi-dimensional, non-redundant safety features, constructing a standardized feature cluster. This provides a unified and standardized input for subsequent feature screening and model training, solving the problems of scattered features and lack of unified standards in existing technologies, which lead to poor coordination between different steps and poor model adaptability.

[0035] In step S2 of this invention, precise feature purification is achieved through dual screening. First, Pearson correlation analysis is used to screen the battery safety feature cluster constructed in S1, eliminating highly correlated redundant features, solving the problem of multicollinearity among features, reducing subsequent computation, and avoiding a decrease in model generalization ability due to feature redundancy. Then, a weighted Relief-F importance selection algorithm is used to calculate the feature contribution rate and rank the importance of the screened features. Compared with the traditional Relief-F algorithm, the weighted optimization can more accurately quantify the contribution of each feature to battery safety warning, avoid the wrong removal of key features, and improve the accuracy of feature screening. By calculating the cumulative contribution rate of feature weights in descending order of importance, and selecting features whose cumulative contribution rate reaches a preset threshold, a scientific simplification of feature dimensions is achieved. This retains key features that contribute significantly to battery safety warnings while eliminating redundant features with low contribution rates. This solves the problem in existing technologies where dimensionality simplification either results in excessive dimensionality reduction leading to the loss of key information or insufficient dimensionality reduction leaving redundancy. Furthermore, calculating the cumulative contribution rate based on feature weights provides a clear quantitative basis for setting the screening threshold, improving the scientific rigor and repeatability of the screening process and avoiding the subjective and arbitrary threshold setting defects of existing technologies. The resulting cluster of key features with simplified dimensions balances feature relevance, importance, and simplicity, significantly reducing the training burden of the S3-step ensemble learning framework, improving model training efficiency, and avoiding interference from redundant features on model accuracy, laying the foundation for high-precision model training in subsequent steps.

[0036] In step S3 of this invention, an integrated learning framework combining an attention mechanism and Multi-Core Extreme Learning Machine (MK-ELM) – Adaboost – is constructed, along with key feature cluster training. This integrated learning framework, which integrates the attention mechanism, MK-ELM, and Adaboost algorithm, achieves synergistic effects among the three technologies. The attention mechanism can adaptively and differentially weight and fuse features from different sources within the simplified key feature cluster, considering the feature source, the real-time identification scenario, and the contribution differences of each feature. This accurately captures the importance of each feature in the key feature cluster, assigning higher weights to high-contribution features and improving the model's ability to target key features. The system focuses on key features to avoid interference from irrelevant features, addressing the problem of equal feature focus and insufficient capture of key information in existing integrated frameworks. Multi-core Extreme Learning Machines (XLMs) replace traditional single kernel functions with multi-core fusion, resulting in stronger fitting ability and better generalization performance compared to traditional XLMs. They can adapt to the nonlinear and complex characteristics of battery operating data, overcoming the poor adaptability of single kernel functions. The Adaboost algorithm iteratively trains weak learners (multi-core XLMs) and integrates them into a strong learner, effectively reducing model training errors and improving the model's robustness and generalization ability, avoiding the problems of overfitting and low accuracy associated with single learners. The trained battery safety risk identification model not only possesses excellent state fitting performance, accurately fitting the mapping relationship between battery operating states and safety risks, providing high-precision support for subsequent early warnings, but also has feature weight analysis capabilities, clarifying the contribution of each key feature to battery safety risks. This achieves the dual goals of state fitting and feature weight analysis, overcoming the shortcomings of existing technologies where models only know the "what" but not the "why," unable to analyze feature weights and struggling to trace fault origins, thus providing technical support for subsequent fault diagnosis. By using the simplified key feature clusters obtained in step S2 as input samples, precise feature-frame adaptation is achieved, avoiding problems such as low training efficiency and decreased accuracy caused by mismatch between input features and model framework in existing technologies, and significantly improving model training efficiency and training effect.

[0037] Step S4 of this invention relies on a cloud platform to collect multi-source time-series operational data of lithium-ion batteries in real time, requiring no manual intervention. This results in high data acquisition efficiency and strong real-time performance. The real-time data is directly input into a pre-trained battery safety risk identification model, which can quickly output a safety risk status judgment result. This achieves full automation of the data acquisition-real-time analysis-early warning output process, overcoming the shortcomings of existing technologies such as high warning delays and inability to monitor in real time. This provides sufficient time for timely handling of battery safety hazards and reduces the probability of battery safety accidents. Compared to existing technologies that can only output simple warning signals, this step not only achieves safety risk warnings but also, through the model's feature weight analysis capabilities combined with the risk status judgment result, achieves accurate diagnosis of battery faults, clarifying the fault type and influencing factors. This addresses the shortcomings of existing technologies that only provide warnings but not diagnoses, resulting in insufficient practicality. It provides precise guidance for battery maintenance and fault handling. Furthermore, relying on a cloud platform for data retrieval and real-time acquisition allows for adaptation to the safety warning needs of multiple lithium-ion batteries and multiple application scenarios (such as new energy vehicles and energy storage power stations), solving the problems of poor adaptability and difficulty in large-scale application of existing technologies, thus improving the practicality and promotional value of the method.

[0038] This invention's mechanism-data fusion-based lithium-ion battery safety early warning method, through the synergy of core technologies in each step, forms a complete technical chain of feature extraction, feature screening, model building and training, and real-time early warning diagnosis. 1) It solves the core pain point of existing technologies being disconnected from mechanisms and data: Through mechanism-data fusion feature extraction in S1 and integrated framework optimization in S3, it achieves deep integration of electrochemical mechanisms and data-driven approaches, enabling the early warning method to have both inherent mechanism support and adaptability to complex actual working conditions, significantly improving the robustness and accuracy of early warnings, and effectively reducing false alarm and false negative rates. 2) It improves early warning accuracy and efficiency: Through dual feature screening in S2, dimensionality is simplified, reducing the model training burden; the integrated learning framework optimization in S3 improves model fitting accuracy and training efficiency. The synergy of these two aspects can improve the training efficiency of the early warning model by more than 30% and the early warning accuracy by more than 25%. Compared with existing single models, the generalization ability and robustness are significantly enhanced. 3) Integrated Early Warning and Diagnosis: Overcoming the limitations of existing technologies that can only provide early warning but not diagnosis, this method integrates accurate early warning and fault diagnosis of battery safety risks through the feature weight analysis capability of the S3 model and the real-time analysis of S4, enhancing the practicality and engineering application value of the method. 4) Excellent Real-Time Performance and Scalability: Real-time data retrieval and collection are achieved through a cloud platform, combined with efficient model computation, enabling real-time early warning. Furthermore, the entire method is standardized and modularized, adaptable to different types and scenarios of lithium-ion batteries, facilitating large-scale application. 5) Reduced Maintenance Costs: Accurate early warning and fault diagnosis allow for timely detection of battery safety hazards, guiding maintenance personnel to perform targeted maintenance, avoiding downtime losses and safety accidents caused by battery failures, significantly reducing the maintenance costs and safety risks of lithium-ion batteries, and demonstrating significant economic and social benefits.

[0039] Furthermore, by clarifying the specific implementation process of the knowledge reasoning method based on historical safety data, defining the specific types of historical safety data (historical safety accidents, thermal runaway experiments, vehicle recall record data), and clarifying the extraction of abnormal related safety features through parameter comparison to form the first type of safety features, this addresses the shortcomings of existing technologies, such as the lack of specific data support, the absence of clear basis for feature extraction, and the weak targeting of extracted safety features. This ensures that the extracted first type of safety features have solid historical experience support, can accurately capture abnormal safety signals during battery operation, and provide a reliable experience-based feature foundation for the subsequent construction of battery safety feature clusters, thereby improving the accuracy and targeting of feature extraction.

[0040] Furthermore, by refining the battery state prediction method based on the pseudo-two-dimensional electro-thermal coupling mechanism model of electrochemistry, the model input and output parameters and the specific feature types generated by secondary analysis are clarified. This solves the defects of existing technologies, such as the fuzzy application of mechanism models, the inability to accurately extract the intrinsic safety features of batteries, and the difficulty in reflecting the inherent safety state of batteries. The model accurately outputs core intrinsic features such as battery state of charge and internal temperature, and transforms them into targeted safety features (second-type safety features) such as thermal runaway risk. This makes up for the shortcomings of single data-driven feature extraction, further enriches the dimensions of battery safety feature clusters, and improves the depth and reliability of features.

[0041] Furthermore, by clarifying the specific process of the data-driven feature extraction method based on temporal convolutional networks (temporal sorting, time window construction, convolution extraction, and pooling mapping), and defining the aggregation and deduplication merging process of the three types of safety features, this paper solves the defects of existing technologies such as non-standard temporal feature extraction, ineffective removal of feature redundancy, and messy multi-source feature fusion. Relying on the temporal capture advantage of temporal convolutional networks, dynamic temporal features (the third type of safety features) in the running data are accurately extracted. At the same time, through feature aggregation and deduplication merging, it ensures that the final battery safety feature cluster is free of redundancy, duplication, and comprehensive in dimensions, providing high-quality feature input for subsequent feature selection and model training, and improving the operating efficiency and accuracy of the entire safety early warning method.

[0042] Furthermore, by refining the specific implementation steps of the weighted Relief-F importance selection algorithm, relying on multiple iterations to obtain stable weight values ​​by averaging the weights, standardizing the feature contribution rate calculation and weight ranking process, and adding a weight normalization processing step after ranking, this approach overcomes the shortcomings of the traditional Relief-F algorithm, such as large weight fluctuations, inconsistent quantization standards, and lack of a normalization quantization step, which leads to inconsistent weight levels and large deviations in subsequent cumulative contribution rate statistical calculations. This makes the feature importance evaluation results more stable and objective, and provides a standardized quantification basis for uniformly accumulating weights and accurately selecting key features based on preset thresholds, effectively improving the scientific nature and result stability of feature dimensionality reduction and selection.

[0043] Furthermore, by refining the construction process of the ensemble learning framework integrating the attention mechanism and the multi-core extreme learning machine (XLM) – Adaboost – this paper clarifies the specific methods of multi-feature fusion, the improvement points of the multi-core XLM, and the construction logic of the Adaboost integration module. This addresses the shortcomings of existing technologies, such as non-standard ensemble learning framework construction, ambiguous application of the attention mechanism and multi-core technology, poor compatibility among framework modules, and lack of clear optimization objectives. The attention mechanism enables precise focusing of key features, and the kernel function weights of the multi-core XLM can be adaptively updated based on the attention weights of features from different sources, while simultaneously optimized according to the adaptation requirements of the ensemble learning framework. This further strengthens the inter-module linkage and overall compatibility. Kernel function optimization of the multi-core XLM improves the adaptability of weak learners, and the Adaboost integration module builds strong learners. This ensures that the constructed ensemble learning framework has a reasonable structure, strong compatibility, and clear optimization objectives, laying a stable and efficient framework foundation for subsequent model training and improving the model's fitting ability and generalization performance.

[0044] Furthermore, by refining the training process of the integrated learning framework, clarifying the specific steps of iterative training, the construction method of evaluation indicators, and the optimization logic of the model, this addresses the shortcomings of existing technologies, such as unclear model training processes, lack of clear optimization basis during training, susceptibility to overfitting or insufficient training, and inability to guarantee model performance. A standardized iterative training process ensures sufficient model training, and objective quantitative assignment is used to construct evaluation indicators to avoid subjective human interference. Integrating evaluation indicators into training achieves parameter optimization. The resulting battery safety risk identification model not only has high fitting accuracy but also possesses feature weight analysis capabilities, perfectly meeting the technical requirements of weight 1. This further enhances the reliability and practicality of the early warning model, providing strong support for subsequent real-time early warning and fault diagnosis.

[0045] This invention also relates to a mechanism-data fusion lithium-ion battery safety early warning system, corresponding to the mechanism-data fusion lithium-ion battery safety early warning method described above. It can be understood as a system implementing the aforementioned mechanism-data fusion lithium-ion battery safety early warning method. This system includes a mechanism-data fusion battery safety feature cluster construction module, a Pearson joint weighted Relief-F feature dimension simplification module, an attention fusion multi-core extreme learning machine integrated modeling module, and a battery safety risk intelligent identification and early warning execution module, all connected in sequence. These modules work collaboratively, integrating battery operation data and mechanism feature data from different sources, effectively expanding the dimensions of battery safety description, and utilizing methods such as Relief-F. Importance selection and optimization were performed to improve the interpretability of features. Redundant features were eliminated while core key features were retained, ensuring both the comprehensiveness and accuracy of feature extraction. Furthermore, core technologies such as attention mechanisms and multi-core extreme learning machines were used to enhance the rationality of feature fusion and the stability of model training. At the same time, the results of multi-feature fusion were incorporated into the data-driven evaluation system, further improving the interpretability and prediction accuracy of the data-driven method. This effectively avoided the warning bias caused by single data or single models, achieving accurate identification, real-time warning, and fault diagnosis of battery safety status. It significantly improved the reliability, stability, and practicality of the entire warning system, providing comprehensive system support for the safe and stable operation of lithium-ion batteries. Attached Figure Description

[0046] Figure 1 This is a flowchart of the lithium-ion battery safety early warning method based on mechanism-data fusion of the present invention. Detailed Implementation

[0047] The present invention will now be described with reference to the accompanying drawings.

[0048] This invention provides a mechanism-data fusion-based method for early warning of lithium-ion battery safety. First, lithium-ion battery operating data is acquired through a cloud platform. Battery safety features are extracted using a combination of knowledge reasoning, model prediction, and data-driven approaches, establishing a battery safety feature cluster. Second, Pearson correlation analysis and a weighted Relief-F importance selection algorithm are used to filter and rank the battery safety feature cluster, establishing a reduced-order battery key feature cluster. Then, an integrated learning framework combining an attention mechanism and a multi-core extreme learning machine (Adaboost) algorithm is constructed. The evaluation function during model training is optimized based on the multi-feature fusion results, correcting the model training direction and improving model fitting accuracy and feature parsing performance. The significant advantage of this invention lies in its ability to improve the interpretability of data-driven algorithms through knowledge-driven and mechanism feature extraction, and to achieve high-precision battery safety risk identification and fault diagnosis through multi-feature fusion. The method flow is as follows: Figure 1 As shown, it includes the following steps:

[0049] S1. Mechanism-Data Fusion Battery Safety Feature Cluster Construction Steps: Historical multi-source time-series operational data of lithium-ion batteries, including voltage, current, and surface temperature, are retrieved from a cloud platform. A knowledge-based reasoning method based on historical safety data, a battery state prediction method based on an electrochemical pseudo-two-dimensional electro-thermal coupling mechanism model, and a data-driven feature extraction method based on temporal convolutional networks are jointly employed to extract battery safety features from the historical multi-source time-series operational data. After removing duplicate features, a battery safety feature cluster is constructed. This step involves retrieving historical multi-source time-series operational data of lithium-ion batteries from a cloud platform and extracting battery safety features using a combination of knowledge reasoning, model prediction, and data-driven methods to establish a battery safety feature cluster.

[0050] S11. Knowledge Reasoning Method Based on Historical Security Data

[0051] Historical safety data was obtained by extracting historical multi-source time-series operational data, including historical safety accident data of lithium-ion batteries, historical thermal runaway experimental data of lithium-ion batteries, and historical vehicle recall records. Based on this historical safety data, a correspondence was established between key parameters such as voltage, current, and temperature and battery safety risks. According to this correspondence, voltage was compared with voltage thresholds, current with current thresholds, and surface temperature with temperature thresholds in the historical multi-source time-series operational data to extract safety features related to abnormal voltage fluctuations, over-limit current, and over-limit temperature, forming the first type of safety features, also known as knowledge reasoning features. The mathematical expression is:

[0052] (1)

[0053] in, Battery voltage, Battery current, The surface temperature of the battery. , , These are the voltage threshold, current threshold, and temperature threshold of a lithium-ion battery, which can be determined through statistical analysis of historical safety data or industry standards and specifications.

[0054] S12. Battery state prediction method based on electrochemical pseudo-two-dimensional electro-thermal coupling mechanism model

[0055] Historical multi-source time-series operational data (including battery terminal voltage, current, surface temperature, etc.) are input into the electrochemical pseudo-two-dimensional electro-thermal coupling mechanism model. Key parameters such as open-circuit voltage, equivalent ohmic internal resistance, and polarization overpotential are obtained by solving the solid / liquid phase diffusion equation and interface reaction kinetic equation of the model. Combined with the mass conservation equation and terminal current constraint, the state of charge (SOC), state of health (SOH), battery internal temperature, ohmic internal resistance, and polarization overpotential characteristics of each lithium-ion cell are calculated to form the original feature set of the mechanism model.

[0056] Specifically, the calculation process based on the pseudo-two-dimensional electro-thermal coupling mechanism model of electrochemistry includes:

[0057] (1) Calculation of terminal voltage:

[0058] (2)

[0059] in, Let be the terminal voltage of the j-th individual cell. Open circuit voltage, Battery current, This is the equivalent ohmic internal resistance. The polarization overpotentials are all obtained by solving the solid / liquid phase diffusion and interfacial reaction kinetic equations in the pseudo-two-dimensional electro-thermal coupling mechanism model (P2D model).

[0060] (2) Based on the matter conservation equation and terminal current constraint in the pseudo-two-dimensional electro-thermal coupling mechanism model of electrochemistry, the state of charge (SOC) of each monomer is calculated:

[0061] (3)

[0062] in, For the first The rated capacity of each individual battery cell This is the initial state of charge.

[0063] (3) Introducing health state (SOH) characterization based on the electrochemical pseudo-two-dimensional electro-thermal coupling mechanism model:

[0064] (4)

[0065] (5)

[0066] in, Let be the initial ohmic internal resistance of the battery. This represents the ohmic internal resistance of the battery as it ages over time. This is a mapping function used to integrate capacity and internal resistance degradation information. Effective capacity varies over time.

[0067] (4) Construct an electro-thermal coupling model under high-rate operating conditions to predict the internal temperature of the battery. .

[0068] (6)

[0069] In the formula, The heat capacity is known and is an inherent parameter of the battery. For ohmic heating rate, The heat power generated by the reaction overpotential, The convective heat transfer coefficient is... For heat exchange area, This represents the predicted internal temperature of a single battery cell.

[0070] Based on the pseudo-two-dimensional electro-thermal coupling mechanism model of electrochemistry, the following feature set can be output at each sampling time:

[0071] (7)

[0072] Furthermore, a secondary analysis is performed on the output characteristics of state of charge, state of health, internal battery temperature, ohmic internal resistance, and polarization overpotential to generate thermal runaway risk warning features, cell inconsistency anomaly features, and abnormal internal resistance growth features, forming a second type of safety feature, which can also be called mechanism model features. In other words, based on data such as battery voltage, current, and temperature, the state of charge, state of health, and internal battery temperature under high-rate operation of each cell are predicted and used as features. Preferably, the features predicted by the model can be further analyzed using knowledge reasoning to establish a safety rule base R based on expert experience, as shown in Table 1.

[0073] Table 1

[0074]

[0075] Based on the constructed security rule base, the following features are output:

[0076] (8)

[0077] In the formula, Represents the knowledge reasoning mapping function. R represents the number of individual battery cells in the system, and R represents the safety rule base. These are high-level security features generated from model-predicted features through knowledge reasoning.

[0078] S13. Data-driven feature extraction method based on temporal convolutional networks

[0079] The retrieved historical multi-source time-series data is processed into a time-series format to construct an input sequence based on a time window. The input sequence is then fed into a temporal convolutional network, where temporal features are extracted through multiple layers of temporal convolution and residual connections. The features output by the temporal convolutional network are then processed by pooling and fully connected mapping to obtain the third type of security features.

[0080] The retrieved historical multi-source time-series runtime data is processed and organized into a time-series format:

[0081] (9)

[0082] Select length as The time window for each moment Construct the input sequence:

[0083] (10)

[0084] Construct a temporal convolutional network based on dilated causal convolution to extract features from the above temporal input:

[0085] (11)

[0086] in, Represents the input sequence. For the first The number of channels in the layer. For the first Layer expansion coefficient, used to expand the receptive field. , These are the kernel weights and bias parameters, respectively. It is a non-linear activation function.

[0087] Residual connection structures are introduced between each convolutional layer to enhance feature representation and stabilize gradient propagation:

[0088] (12)

[0089] in: is the network layer number; k refers to the kth training time; h is the hidden feature vector; F is the nonlinear feature transformation function.

[0090] After extracting high-dimensional temporal features through multiple layers of temporal convolution and residual blocks, the output of the final layer is pooled or compressed in the temporal dimension to obtain the time-series data. The corresponding feature vectors are processed using global average pooling:

[0091] (13)

[0092] in Effective time length of the last layer This is the output of the last layer of the TCN. Then, Input one or more fully connected mappings:

[0093] (14)

[0094] In the formula, , It is a fully connected layer. For activation function, That is, at time The battery safety feature vector (third type of safety feature) extracted by the data-driven approach can also be called TCN data-driven feature.

[0095] Finally, the values ​​obtained from each individual cell at each time point are... Together with the first type of safety features obtained by the knowledge reasoning method and the second type of safety features predicted based on the P2D mechanism model, these features are summarized, and features with the same meaning or high repetition are eliminated or merged to finally form a battery safety feature cluster for characterizing the safety status of lithium-ion batteries.

[0096] S2. Pearson Joint Weighted Relief-F Feature Dimension Simplification Step: Pearson correlation analysis is used to screen the battery safety feature cluster for relevance. The weighted Relief-F importance selection algorithm is used to calculate the feature contribution rate of the screened features and rank them by importance. The cumulative contribution rate of the feature weights is calculated by accumulating them from highest to lowest importance. Features whose cumulative contribution rate reaches a preset threshold are selected to form the key feature cluster after dimensional simplification. This step utilizes the weighted Relief-F importance selection algorithm and correlation analysis method to identify and rank the key features, establishing a reduced-order (dimensionally simplified) battery key feature cluster.

[0097] Key features are identified by using Pearson correlation analysis to cluster battery safety features and select features that are strongly positively or negatively correlated with battery failure. Furthermore, a weighted Relief-F importance selection algorithm is used. First, the importance of each feature is estimated based on expert experience and assigned a certain weight value. Then, the weighted Relief-F importance selection algorithm is used to calculate the contribution rate of each feature and select features with a cumulative contribution rate of more than 90% as the final battery safety features.

[0098] S21. Initial screening based on Pearson correlation analysis.

[0099] The battery safety feature cluster obtained in step S1 is denoted as:

[0100] (15)

[0101] The feature matrix and label vector are further constructed as follows:

[0102]

[0103] In the formula, the number of samples collected is , No. The first sample Each feature is denoted as The corresponding battery operating status label is recorded as .

[0104] S22, Key Feature Identification Using the Weighted Relief-F Importance Selection Algorithm

[0105] The weighted Relief-F importance selection algorithm is used to calculate the feature contribution rate and rank the importance of the features after relevance screening. Specifically, this may include:

[0106] (1) Initialize weights:

[0107] Assign an initial weight (usually 0) to each of the filtered features.

[0108] (2) Random sampling and nearest neighbor search:

[0109] Randomly select a sample (called the "anchor"), find the nearest neighbor of the sample in the same class, and find the nearest neighbor of the sample in each different class.

[0110] (3) Weight update:

[0111] For each feature, the weight is adjusted based on its ability to distinguish between similar and dissimilar categories. The formula for calculating the feature weight is shown below:

[0112] (17)

[0113] (18)

[0114] Where T init (C) represents the initial / previous iteration weights for category C, Sample represents the sample, and H i M represents the nearest neighbor of the same type as Sample. i (C) represents the nearest neighbors of Sample from different classes, p(Sample) represents the probability that a sample belongs to its class, and p(class(Sample)) represents the probability of a class in the entire dataset. Under the premise of uniform sampling, p(Sample) = p(class(Sample)). max(C) and min(C) represent the maximum and minimum values ​​of all samples in class C, respectively. i(C) represents the data in the i-th frame under category C. k represents the maximum number of nearest neighbors of samples of the same type and the maximum number of nearest neighbors of samples of different types.

[0115] Euclidean distance is used to determine the nearest neighbors of the same sample and the nearest neighbors of different samples. Specifically, within the same class of samples, the k samples with the closest Euclidean distance to the sample are considered the same sample's nearest neighbors; within different classes of samples, the k samples with the closest Euclidean distance to the sample are considered the different sample's nearest neighbors. The formula for calculating Euclidean distance is:

[0116] (19)

[0117] Where: N represents the feature dimension, X and Y represent points of dimension N, and D(X,Y) represents the Euclidean distance.

[0118] (4) Iterative calculation:

[0119] The sampling and update process is repeated multiple times, and the final weights are averaged to improve stability.

[0120] (5) Ranking of feature importance:

[0121] The results after feature weight update are shown in Table 2.

[0122] Table 2

[0123]

[0124] Where C i Representing the i-th feature, T(C) i ) represents C i Corresponding weights.

[0125] Using the feature weights in the table above as the sorting criterion, the feature labels are sorted from largest to smallest according to their feature weights. The sorted feature weights are then normalized using the following formula:

[0126] (20)

[0127] The normalized feature weights are accumulated and calculated. Features with a weight ratio of more than 90% are used as the updated content of the feature library.

[0128] S3. Integrated Modeling Steps of Attention Fusion with Multi-Core Extreme Learning Machine: An integrated learning framework integrating an attention mechanism and the Adaboost multi-core extreme learning machine is constructed. The attention mechanism is used to adaptively differentiate and weightedly fuse features from different sources obtained by the knowledge reasoning method, the mechanistic model battery state prediction method, and the temporal convolutional network data-driven feature extraction method within the simplified key feature cluster. This is done by considering the feature source, the real-time identification scenario, and the contribution differences of each feature. The simplified key feature cluster is used as input samples to train the integrated learning framework, resulting in a battery safety risk identification model with state fitting performance and feature weight analysis capabilities. This step constructs an integrated learning framework integrating an attention mechanism and the Adaboost multi-core extreme learning machine algorithm. It optimizes the evaluation function during model training based on the feature information obtained from multi-feature fusion, further correcting the model training direction and improving model fitting accuracy and feature analysis performance.

[0129] Specifically, the integrated learning framework that combines attention mechanisms with the multi-core extreme learning machine Adaboost includes:

[0130] (1) Multi-feature fusion of the key feature cluster obtained in step S2 based on the attention mechanism, including linear mapping of each group of features in the key feature cluster to construct query vector and key vector; based on the query vector and key vector, the matching score is calculated by additive attention and the attention weight is obtained by Softmax normalization, and then the fused input feature vector is obtained by weighting based on the attention weight.

[0131] The final cluster of key battery features after order reduction (dimensional simplification) using the S2 step is assumed to contain... The first feature, denoted as the first... The feature vectors of each sample are:

[0132] (twenty one)

[0133] By using linear mapping, each set of features is mapped to a unified latent space:

[0134] (twenty two)

[0135] In the formula, For the first Group characteristics.

[0136] Constructing sample-level query vectors Key vectors for each set of features :

[0137] (twenty three)

[0138] Additive attention is used to calculate the matching score for each set of features:

[0139] (twenty four)

[0140] Where, q i No. The query vector corresponding to each input feature For key vectors, It's the match score.

[0141] Attention weights are obtained through Softmax normalization:

[0142] (25)

[0143] The calculated input feature vector after multi-feature fusion is as follows:

[0144] (26)

[0145] (2) A multi-core extreme learning machine is constructed as a weak learner. The multi-core extreme learning machine replaces the basis functions of the traditional extreme learning machine with multiple weighted kernel functions. The weighting method is selected according to the type of kernel function. Preferably, the kernel function is established by combining the exponential function and the sigmoid function, and its weight value is designed to be related to the model training residual: when the model training residual is high, the weight of the exponential function is increased; when the model training residual is low, the weight of the sigmoid function is increased, thereby improving the training speed and prediction ability of the algorithm. At the same time, the weight of the kernel function is also adaptively updated according to the attention weights of the different source features obtained by the knowledge reasoning method, the mechanism model battery state prediction method and the temporal convolutional network data-driven feature extraction method in the key feature cluster after dimensional simplification. The kernel function adaptability and the overall compatibility of the integrated learning framework are optimized simultaneously to achieve dual-objective optimization.

[0146] A dual optimization objective is adopted, corresponding to minimizing output bias and optimizing weight configuration, respectively, to eliminate randomness generated during training and minimize output bias:

[0147] (27)

[0148] In the formula, β represents the model weights, and δ i This represents the deviation between the output and the true value, and N represents the number of samples.

[0149] By introducing Lagrange multipliers, we can simultaneously satisfy these constraints and minimize the objective function during the optimization process:

[0150] (28)

[0151] In the formula, a ij Represents the Lagrange multiplier, M represents the number of neurons in the output layer, and β j It is the weight vector associated with the j-th output neuron (or output dimension), t ij It is the model's true target value or label, x i The input sample is h(x). i ) is x i The value after mapping by the kernel function This represents the error term.

[0152] By solving for the partial derivatives of the Lagrange function with respect to each variable (such as weight β, bias term δ, etc.), the condition for minimizing the Lagrange function is obtained, as shown in the equation:

[0153] (29)

[0154] In the formula: c represents the regularization parameter, which takes values ​​between [-1, 1], and H represents the hidden layer output matrix.

[0155] Because the decision function satisfies the equation

[0156] (30)

[0157] The output function can then be written as:

[0158] (31)

[0159] The kernel function h(x) is written as:

[0160] (32)

[0161] Where x represents a new input sample point, z represents a sample point in the training set, and the kernel function measures the relationship between the new input x and the training sample z in high-dimensional space by calculating the similarity between x and z; σ represents the kernel width parameter.

[0162] (3) Build an Adaboost integration module as a strong learner. The integrated learning framework that integrates the attention mechanism and the multi-core extreme learning machine-Adaboost algorithm refers to using the multi-core extreme learning machine as a weak learner. For each weak learner, the attention mechanism module is first used to process the input features. Then, the prediction results of multiple weak learners are input into the Adaboost algorithm for learning to build a strong learner. The strong learner is then combined to form an integrated learning framework that integrates the attention mechanism, the multi-core extreme learning machine and the Adaboost integration module.

[0163] The simplified key feature clusters are used as input samples to train within the established ensemble learning framework, specifically including:

[0164] The multi-core extreme learning machine built into the integrated learning framework is used as the basic learning unit. The Adaboost algorithm is used to complete the iterative training integration based on the Adaboost integration module. The iterative training process includes: initializing the input sample weight distribution, completing the sample feature fitting based on the basic learning unit and calculating the fitting deviation, determining the corresponding weight of the basic learning unit based on the fitting deviation, synchronously updating the sample weight distribution and repeatedly executing the iterative operation until the preset training termination condition is met.

[0165] A battery safety risk assessment index is constructed using an objective quantitative assignment method. Specifically, each feature is assigned a non-linear weight value. After normalizing all features, a weighted fusion is performed based on the weight values ​​of each feature to obtain the battery safety risk index predicted by the model, i.e., the battery safety risk assessment index. This assessment index is then combined with the original assessment function in the ensemble learning algorithm using a weighted combination: when the output value of the original assessment function is low, the weight of the battery safety risk assessment index is increased; when the output value of the original assessment function is high, the weight of the original assessment function is increased. In this way, the assessment index is integrated into the training process to complete the optimization of model parameters. After training, a battery safety risk identification model with both state fitting ability and feature weight analysis ability is obtained.

[0166] The weak classifier consists of a kernel extreme learning machine, which uses the predictions from multiple kELMs to train Adaboost to build a strong classifier.

[0167] (1) Initialize the training data samples and assign weights to them. The weight distribution satisfies the following formula:

[0168] (33)

[0169] Where i represents the i-th sample, and N represents the total number of samples.

[0170] (2) Train kELM and calculate the error rate of the training results. The formula for calculating the error rate is:

[0171] (34)

[0172] In the formula: This represents an indicator function, which is 1 when the classification is incorrect and 0 otherwise.

[0173] (3) Calculate the weight α of the t-th kELM classifier t For example:

[0174] (35)

[0175] (4) Update the weight distribution D of the samples t+1 (i), as shown in the formula:

[0176] (36)

[0177] In the formula: y i h represents the training label. t (x i Z represents the output of the t-th kELM classifier. t This represents the normalization factor.

[0178] Repeat steps (2) through (4) until all kELM classifiers have been trained.

[0179] S4. Battery Safety Risk Intelligent Identification and Early Warning Execution Steps: Input the corresponding multi-source time-series operation data of lithium-ion batteries collected in real time by the cloud platform into the battery safety risk identification model, output the lithium-ion battery safety risk status judgment result, and realize battery safety risk early warning and fault diagnosis.

[0180] This invention also relates to a mechanism-data fusion lithium-ion battery safety early warning system, corresponding to the mechanism-data fusion lithium-ion battery safety early warning method described above. It can be understood as a system that implements the aforementioned mechanism-data fusion lithium-ion battery safety early warning method. This system includes, in sequence, a mechanism-data fusion battery safety feature cluster construction module, a Pearson joint weighted Relief-F feature dimension simplification module, an attention fusion multi-core extreme learning machine integrated modeling module, and a battery safety risk intelligent identification and early warning execution module.

[0181] The mechanism-data fusion battery safety feature cluster construction module retrieves historical multi-source time-series operational data of lithium-ion batteries, including voltage, current, and surface temperature, through a cloud platform. It then employs a knowledge reasoning method based on historical safety data, a battery state prediction method based on an electrochemical pseudo-two-dimensional electro-thermal coupling mechanism model, and a data-driven feature extraction method based on a time convolutional network to extract battery safety features from the historical multi-source time-series operational data. After removing duplicate features, it constructs a battery safety feature cluster.

[0182] The Pearson-weighted Relief-F feature dimensionality simplification module uses Pearson correlation analysis to screen the battery safety feature cluster for correlation; it uses the weighted Relief-F importance selection algorithm to calculate the feature contribution rate of the screened features and sort them by importance, and calculates the cumulative contribution rate of the feature weights in descending order of importance, and selects the features corresponding to the cumulative contribution rate that reaches a preset threshold to form the key feature cluster after dimensional simplification.

[0183] The attention-fusion multi-core extreme learning machine integrated modeling module builds an integrated learning framework that integrates the attention mechanism and the multi-core extreme learning machine - Adaboost. The attention mechanism is used to adaptively differentiate and weightedly fuse features from different sources obtained by the knowledge reasoning method, the mechanism model battery state prediction method, and the temporal convolutional network data-driven feature extraction method in the dimensionally simplified key feature cluster, taking into account the feature source, the real-time identification scenario, and the contribution differences of each feature. The dimensionally simplified key feature cluster is used as input samples to train the integrated learning framework to obtain a battery safety risk identification model with state fitting performance and feature weight parsing ability.

[0184] The battery safety risk intelligent identification and early warning execution module inputs the corresponding multi-source time-series operation data of lithium-ion batteries collected in real time by the cloud platform into the battery safety risk identification model, and outputs the lithium-ion battery safety risk status judgment result, thereby realizing battery safety risk early warning and fault diagnosis.

[0185] Furthermore, in the mechanism-data fusion battery safety feature cluster construction module, the knowledge reasoning method based on historical safety data obtains historical safety data from retrieved historical multi-source time-series operational data. This historical safety data includes historical safety accident data for lithium-ion batteries, historical thermal runaway experimental data for lithium-ion batteries, and historical vehicle recall records. Based on this historical safety data, a correspondence is established between voltage parameters, current parameters, and temperature parameters and battery safety risks. According to this correspondence, the voltage is compared with voltage thresholds, current with current thresholds, and surface temperature with temperature thresholds in the historical multi-source time-series operational data to extract safety features related to abnormal voltage fluctuations, over-limit currents, and over-limit temperatures, forming the first type of safety features.

[0186] The battery state prediction method based on the electrochemical pseudo-two-dimensional electro-thermal coupling mechanism model involves inputting historical multi-source time-series operational data into the electrochemical pseudo-two-dimensional electro-thermal coupling mechanism model. The electrochemical pseudo-two-dimensional electro-thermal coupling mechanism model calculates and outputs the state of charge, state of health, internal temperature, ohmic internal resistance, and polarization overpotential characteristics of each lithium-ion cell. The output state of charge, state of health, internal temperature, ohmic internal resistance, and polarization overpotential characteristics are then subjected to secondary analysis to generate thermal runaway risk warning characteristics, cell inconsistency anomaly characteristics, and abnormal internal resistance growth characteristics, forming a second type of safety characteristics.

[0187] The data-driven feature extraction method based on temporal convolutional networks involves temporally organizing historical multi-source time-series running data to construct an input sequence based on a time window; inputting the input sequence into a temporal convolutional network to extract temporal features through multiple layers of temporal convolution and residual connections; and performing pooling and fully connected mapping on the features output by the temporal convolutional network to obtain the third type of security features.

[0188] The third type of safety features are summarized with the first and second types of safety features, and features with the same or repeated meanings are removed or merged to construct a battery safety feature cluster.

[0189] Furthermore, in the Pearson joint weighted Relief-F feature dimension simplification module, the weighted Relief-F importance selection algorithm is used to calculate the feature contribution rate and rank the importance of the selected features. Specifically, this includes: configuring initial weights for each selected feature; randomly selecting samples as anchor points and using Euclidean distance to search for nearest neighbors among samples of the same class and samples of different classes; adjusting feature weights according to the feature's ability to distinguish samples of different classes; iterating the sampling and weight update process multiple times and taking the average weight after multiple iterations; calculating the feature contribution rate and ranking the importance of the selected features based on the average weight; and then normalizing the ranked feature weights to facilitate the subsequent accumulation of the normalized feature weights in descending order of importance to calculate the cumulative contribution rate.

[0190] It should be noted that the specific embodiments described above enable those skilled in the art to more fully understand the present invention, but do not limit the present invention in any way. Therefore, although the present invention has been described in detail with reference to the accompanying drawings and embodiments, those skilled in the art should understand that modifications or equivalent substitutions can still be made to the present invention. In short, all technical solutions and improvements that do not depart from the spirit and scope of the present invention should be covered within the protection scope of the present invention patent.

Claims

1. A mechanism-data fusion method for early warning of lithium-ion battery safety, characterized in that, Includes the following steps: S1. By retrieving historical multi-source time-series operating data of lithium-ion batteries, including voltage, current, and surface temperature, through the cloud platform, and jointly employing a knowledge reasoning method based on historical safety data, a battery state prediction method based on an electrochemical pseudo-two-dimensional electro-thermal coupling mechanism model, and a data-driven feature extraction method based on a time convolutional network, battery safety features are extracted from the historical multi-source time-series operating data, and duplicate features are removed to construct a battery safety feature cluster. S2. The battery safety feature cluster is screened for correlation using Pearson correlation analysis; the feature contribution rate of the screened features is calculated and the importance is sorted by weighted Relief-F importance selection algorithm. The cumulative contribution rate of the feature weights is calculated by accumulating them in descending order of importance. The feature corresponding to the cumulative contribution rate reaching the preset threshold is selected to form a key feature cluster after dimensional simplification. S3. Construct an integrated learning framework that combines an attention mechanism with the multi-core extreme learning machine Adaboost. The attention mechanism is used to adaptively and differentially weightedly fuse features from different sources obtained by the knowledge reasoning method, the mechanism model battery state prediction method, and the temporal convolutional network data-driven feature extraction method in the dimensionally simplified key feature cluster, taking into account the feature source, the real-time identification scenario, and the contribution differences of each feature. The dimensionally simplified key feature cluster is used as input samples to train the integrated learning framework to obtain a battery safety risk identification model with state fitting performance and feature weight parsing ability. S4. Input the corresponding multi-source time-series operation data of lithium-ion batteries collected in real time by the cloud platform into the battery safety risk identification model, and output the lithium-ion battery safety risk status judgment result to realize battery safety risk warning and fault diagnosis.

2. The mechanism-data fusion-based lithium-ion battery safety early warning method according to claim 1, characterized in that, In step S1, the knowledge reasoning method based on historical safety data obtains historical safety data from retrieved historical multi-source time-series operational data. The historical safety data includes historical safety accident data of lithium-ion batteries, historical thermal runaway experimental data of lithium-ion batteries, and historical vehicle recall record data. Based on the historical safety data, a correspondence between voltage parameters, current parameters, and temperature parameters and battery safety risks is established. According to the correspondence, the voltage is compared with voltage threshold, current with current threshold, and surface temperature with temperature threshold in the historical multi-source time-series operational data, respectively, to extract safety features related to abnormal voltage fluctuations, over-limit current, and over-limit temperature, forming the first type of safety features.

3. The mechanism-data fusion-based lithium-ion battery safety early warning method according to claim 2, characterized in that, In step S1, the battery state prediction method based on the electrochemical pseudo-two-dimensional electro-thermal coupling mechanism model involves inputting historical multi-source time-series running data into the electrochemical pseudo-two-dimensional electro-thermal coupling mechanism model. The electrochemical pseudo-two-dimensional electro-thermal coupling mechanism model calculates and outputs the state of charge, state of health, internal temperature, ohmic internal resistance, and polarization overpotential characteristics of each lithium-ion cell. A secondary analysis is performed on the output state of charge, state of health, battery internal temperature, ohmic internal resistance, and polarization overpotential characteristics to generate thermal runaway risk warning characteristics, single-cell inconsistency abnormal characteristics, and internal resistance abnormal growth characteristics, forming a second type of safety characteristics.

4. The mechanism-data fusion-based lithium-ion battery safety early warning method according to claim 3, characterized in that, In step S1, the data-driven feature extraction method based on temporal convolutional networks involves temporally organizing the retrieved historical multi-source time-series running data to construct an input sequence based on a time window; inputting the input sequence into a temporal convolutional network to extract temporal features through multiple layers of temporal convolution and residual connections. The features output by the temporal convolutional network are subjected to pooling and fully connected mapping to obtain the third type of security features; The third type of safety features are summarized with the first and second types of safety features, and features with the same or repeated meanings are removed or merged to construct a battery safety feature cluster.

5. The mechanism-data fusion method for early warning of lithium-ion battery safety according to any one of claims 1 to 4, characterized in that, In step S2, the weighted Relief-F importance selection algorithm is used to calculate the feature contribution rate and rank the importance of the selected features. Specifically, this includes: configuring initial weights for each selected feature; randomly selecting samples as anchor points and using Euclidean distance to search for nearest neighbors among samples of the same class and samples of different classes; adjusting feature weights based on the feature's ability to distinguish samples of different classes; iterating the sampling and weight update process multiple times and taking the average weight after multiple iterations; calculating the feature contribution rate and ranking the importance of the selected features based on the average weight; and then normalizing the ranked feature weights to facilitate the subsequent calculation of the cumulative contribution rate of the normalized feature weights in descending order of importance.

6. The mechanism-data fusion method for early warning of lithium-ion battery safety according to any one of claims 1 to 4, characterized in that, In step S3, an integrated learning framework combining the attention mechanism and the multi-core extreme learning machine Adaboost is built, specifically including: Based on the attention mechanism, the key feature clusters obtained in step S2 are fused into multiple features, including linear mapping of each group of features in the key feature cluster to construct query vectors and key vectors; based on the query vectors and key vectors, the matching score is calculated by additive attention and normalized by Softmax to obtain attention weights, and then the fused input feature vectors are obtained by weighting based on the attention weights. A multi-core extreme learning machine is constructed as a weak learner. The multi-core extreme learning machine replaces the basis function of the traditional extreme learning machine with a kernel function that is a weighted combination of the exponential function and the sigmoid function. By associating the kernel function weights with the adaptation requirements of the ensemble learning framework, the kernel function weights are adaptively updated according to the attention weights of the features from different sources obtained by the knowledge reasoning method, the mechanism model battery state prediction method, and the temporal convolutional network data-driven feature extraction method in the key feature cluster after dimensional simplification. This simultaneously optimizes the kernel function adaptability and the overall compatibility of the ensemble learning framework, achieving dual-objective optimization. An Adaboost integration module was built as a strong learner, forming an integrated learning framework that combines attention mechanism, multi-core extreme learning machine and Adaboost integration module.

7. The mechanism-data fusion-based lithium-ion battery safety early warning method according to claim 6, characterized in that, In step S3, the simplified key feature clusters are used as input samples and fed into the established ensemble learning framework for training, specifically including: The multi-core extreme learning machine built into the integrated learning framework is used as the basic learning unit. The Adaboost algorithm is used to complete the iterative training integration based on the Adaboost integration module. The iterative training process includes: initializing the input sample weight distribution, completing the sample feature fitting based on the basic learning unit and calculating the fitting deviation, determining the corresponding weight of the basic learning unit based on the fitting deviation, synchronously updating the sample weight distribution and repeatedly executing the iterative operation until the preset training termination condition is met. An objective quantitative assignment method is used to construct a battery safety risk assessment index. This assessment index is then integrated into the training process to optimize the model parameters. After training, a battery safety risk identification model with both state fitting ability and feature weight analysis ability is obtained.

8. A mechanism-data fusion lithium-ion battery safety early warning system, characterized in that, The module comprises a mechanism-data fusion battery safety feature cluster construction module, a Pearson joint weighted Relief-F feature dimension simplification module, an attention fusion multi-core extreme learning machine integrated modeling module, and a battery safety risk intelligent identification and early warning execution module, which are connected in sequence. The mechanism-data fusion battery safety feature cluster construction module retrieves historical multi-source time-series operational data of lithium-ion batteries, including voltage, current, and surface temperature, through a cloud platform. It then employs a knowledge reasoning method based on historical safety data, a battery state prediction method based on an electrochemical pseudo-two-dimensional electro-thermal coupling mechanism model, and a data-driven feature extraction method based on a time convolutional network to extract battery safety features from the historical multi-source time-series operational data. After removing duplicate features, it constructs a battery safety feature cluster. The Pearson-weighted Relief-F feature dimensionality simplification module uses Pearson correlation analysis to screen the battery safety feature cluster for correlation; it uses the weighted Relief-F importance selection algorithm to calculate the feature contribution rate of the screened features and sort them by importance, and calculates the cumulative contribution rate of the feature weights in descending order of importance, and selects the features corresponding to the cumulative contribution rate that reaches a preset threshold to form the key feature cluster after dimensional simplification. The attention-fusion multi-core extreme learning machine integrated modeling module builds an integrated learning framework that integrates the attention mechanism and the multi-core extreme learning machine - Adaboost. The attention mechanism is used to adaptively differentiate and weightedly fuse features from different sources obtained by the knowledge reasoning method, the mechanism model battery state prediction method, and the temporal convolutional network data-driven feature extraction method in the dimensionally simplified key feature cluster, taking into account the feature source, the real-time identification scenario, and the contribution differences of each feature. The dimensionally simplified key feature cluster is used as input samples to train the integrated learning framework to obtain a battery safety risk identification model with state fitting performance and feature weight parsing ability. The battery safety risk intelligent identification and early warning execution module inputs the corresponding multi-source time-series operation data of lithium-ion batteries collected in real time by the cloud platform into the battery safety risk identification model, and outputs the lithium-ion battery safety risk status judgment result, thereby realizing battery safety risk early warning and fault diagnosis.

9. The mechanism-data fusion lithium-ion battery safety early warning system according to claim 8, characterized in that, In the mechanism-data fusion battery safety feature cluster construction module, the knowledge reasoning method based on historical safety data obtains historical safety data from retrieved historical multi-source time-series operational data. This historical safety data includes historical safety accident data for lithium-ion batteries, historical thermal runaway experimental data for lithium-ion batteries, and historical vehicle recall records. Based on this historical safety data, a correspondence is established between voltage parameters, current parameters, and temperature parameters and battery safety risks. According to this correspondence, voltage is compared with voltage thresholds, current with current thresholds, and surface temperature with temperature thresholds in the historical multi-source time-series operational data to extract safety features related to abnormal voltage fluctuations, over-limit currents, and over-limit temperatures, forming the first type of safety features. The battery state prediction method based on the electrochemical pseudo-two-dimensional electro-thermal coupling mechanism model involves inputting historical multi-source time-series operating data into the electrochemical pseudo-two-dimensional electro-thermal coupling mechanism model. The electrochemical pseudo-two-dimensional electro-thermal coupling mechanism model calculates and outputs the state of charge, state of health, internal temperature, ohmic internal resistance, and polarization overpotential characteristics of each lithium-ion cell. A secondary analysis is performed on the output state of charge, state of health, battery internal temperature, ohmic internal resistance, and polarization overpotential characteristics to generate thermal runaway risk warning characteristics, single-cell inconsistency abnormal characteristics, and internal resistance abnormal growth characteristics, forming a second type of safety characteristics. The data-driven feature extraction method based on temporal convolutional networks involves temporally organizing the retrieved historical multi-source time-series data to construct an input sequence based on a time window; inputting the input sequence into a temporal convolutional network to extract temporal features through multiple layers of temporal convolution and residual connections. The features output by the temporal convolutional network are subjected to pooling and fully connected mapping to obtain the third type of security features; The third type of safety features are summarized with the first and second types of safety features, and features with the same or repeated meanings are removed or merged to construct a battery safety feature cluster.

10. The mechanism-data fusion lithium-ion battery safety early warning system according to claim 8 or 9, characterized in that, In the Pearson Joint Weighted Relief-F Feature Dimension Reduction Module, the Weighted Relief-F Importance Selection Algorithm is used to calculate the feature contribution rate and rank the importance of the selected features. Specifically, this includes: configuring initial weights for each selected feature; randomly selecting samples as anchor points and using Euclidean distance to search for nearest neighbors among samples of the same class and samples of different classes; adjusting feature weights based on the feature's ability to distinguish samples of different classes; iterating the sampling and weight update process multiple times and taking the average weight after multiple iterations; calculating the feature contribution rate and ranking the importance of the selected features based on the average weight; and then normalizing the ranked feature weights to facilitate the subsequent accumulation of the normalized feature weights in descending order of importance to calculate the cumulative contribution rate.

Images