Multi-modal coupling lithium battery thermal runaway multi-stage early warning method and multi-modal coupling lithium battery thermal runaway multi-stage early warning system

Abstract

The invention provides a multi-modal coupled lithium battery thermal runaway multi-stage early warning method and system in the technical field of lithium battery safety monitoring. The method comprises the following steps: S1, collecting a large amount of historical multi-modal monitoring data in a lithium battery operation process to construct a data set; s2, creating a thermal runaway monitoring model based on a time-frequency feature extraction module, a feature fusion module, a reconstruction decoding module, a physical simulation module, a multi-source information fusion module and a report output module, and training and deploying the thermal runaway monitoring model through a data set; s3, collecting real-time multi-mode monitoring data of the lithium battery, inputting the real-time multi-mode monitoring data into the deployed thermal runaway monitoring model, and obtaining a monitoring report carrying a temperature average value, a temperature standard deviation, a voltage average value and a voltage standard deviation; and S4, performing thermal runaway early warning on the lithium battery based on the monitoring report. The lithium battery thermal runaway early warning method has the advantages that the accuracy, the working condition adaptability and the reliability of lithium battery thermal runaway early warning are greatly improved.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of lithium battery safety monitoring, and particularly discloses a multi-modal coupled lithium battery thermal runaway multi-stage early warning method and system. BACKGROUND

[0002] With the rapid development of new energy vehicles, energy storage systems and other fields, lithium batteries have been widely used due to their high energy density and long cycle life. However, lithium batteries have the risk of triggering thermal runaway due to overcharging, internal short circuit or external mechanical damage, etc., which may cause serious safety accidents. Therefore, it is of great practical significance to build an efficient, accurate and robust lithium battery thermal runaway risk detection system / method to ensure the safety of the system operation.

[0003] At present, the detection technology for lithium battery thermal runaway mainly includes the following two schemes:

[0004] 1. Monitoring method based on empirical threshold: This method uses fixed thresholds of voltage, temperature and other parameters to alarm abnormities. This method is simple to implement, but it lacks adaptability in complex working conditions and has a high risk of false positives or false negatives.

[0005] 2. Modeling method based on data-driven: This method uses machine learning algorithms or deep learning models to mine and model multi-modal data such as voltage, current, temperature, state of charge (SOC), etc. Abnormities are judged by analyzing reconstruction errors or prediction errors. This method has the advantage of not relying on complex physical mechanism models and has certain applicability to various types of lithium batteries.

[0006] However, the above existing technologies still have the following significant defects or limitations:

[0007] 1. Limitations of single-dimensional feature monitoring: Existing methods rely heavily on / focus on monitoring a single physical quantity (such as temperature, voltage). For example, when only monitoring the surface temperature of the battery, it may not capture significant temperature rise changes during the early stage of internal short circuit-induced thermal runaway. If only voltage changes are analyzed, it is difficult to effectively identify / distinguish normal working condition fluctuations from early abnormal signals.

[0008] 2. Lack of generalization of pure data-driven models: Deep learning-based time series prediction models (such as LSTM, Transformer, etc.) can capture complex nonlinear temporal relationships between data, but they lack prior knowledge constraints on the physical and chemical mechanisms of the battery. When faced with new working conditions that are not fully covered by the training data (such as extreme temperatures and large rate charging and discharging), their generalization ability is often limited, making it difficult to ensure the reliability of the detection results.

[0009] 3. Insufficient cross-modal information fusion: The thermal runaway process of lithium batteries is essentially a complex process involving the close coupling of multiple physical fields such as electrochemistry and thermodynamics. The resulting time-domain characteristics can reflect the evolution trend of the state, while the frequency-domain characteristics can reveal potential periodic or frequency-related laws. However, current technical solutions generally fail to effectively utilize or deeply fuse these multi-scale, cross-modal information correlations.

[0010] 4. Bottlenecks in the application of physical mechanism models: Although physical mechanism models such as equivalent circuit models (ECM) have the advantage of theoretical interpretability, traditional model parameter identification methods face challenges in adapting to battery dynamic aging and complex dynamic operating conditions. This leads to a significant decrease in model accuracy over time and as operating conditions evolve, affecting its predictive accuracy in online monitoring.

[0011] Therefore, how to provide a multi-modal coupled multi-level early warning method and system for lithium battery thermal runaway, so as to improve the accuracy, adaptability and reliability of lithium battery thermal runaway early warning, has become an urgent technical problem to be solved. Summary of the Invention

[0012] The technical problem to be solved by the present invention is to provide a multi-modal coupled multi-level early warning method and system for thermal runaway of lithium batteries, so as to improve the accuracy, adaptability and reliability of early warning of thermal runaway of lithium batteries.

[0013] In a first aspect, the present invention provides a multi-modal coupled, multi-level early warning method for thermal runaway in lithium batteries, comprising the following steps:

[0014] Step 1: Collect a large amount of historical multimodal monitoring data during the operation of lithium batteries, and construct a dataset after preprocessing the historical multimodal monitoring data.

[0015] Step 2: Create a thermal runaway monitoring model based on the time-frequency feature extraction module, feature fusion module, reconstruction decoding module, physical simulation module, multi-source information fusion module, and report output module. Train and deploy the thermal runaway monitoring model using the dataset.

[0016] Step 3: Collect real-time multimodal monitoring data of lithium battery, preprocess the real-time multimodal monitoring data and input it into the deployed thermal runaway monitoring model to obtain a real-time monitoring report carrying the average temperature, temperature standard deviation, average voltage and voltage standard deviation.

[0017] Step 4: Based on the real-time monitoring report, provide an early warning of thermal runaway for the lithium battery.

[0018] Furthermore, Step 1 specifically includes:

[0019] Collect a large amount of historical multi-modal monitoring data of lithium batteries during operation, wherein the historical multi-modal monitoring data at least includes voltage, current, temperature and SOC;

[0020] After preprocessing the historical multi-modal monitoring data at least including normalization, denoising and missing value filling, a data set is constructed.

[0021] Further, in the step Step2, the time-frequency feature extraction module is constructed based on a time domain feature extraction unit and a frequency domain feature extraction unit; the time domain feature extraction unit is used to encode the multi-modal monitoring data by a Transformer-based time domain encoder T-encoder to obtain time domain features E1; the frequency domain feature extraction unit is used to extract initial frequency domain features by fast Fourier transform or short-time Fourier transform on the multi-modal monitoring data, and then encode the initial frequency domain features by a Transformer-based time domain encoder F-encoder to obtain frequency domain features E2;

[0022] The feature fusion module is used to fuse the time domain features E1 and the frequency domain features E2 by a cross-attention mechanism to obtain time-frequency features E3;

[0023] The reconstruction decoding module is used to combine the time-frequency features E3, the current and the SOC in the multi-modal monitoring data as prompts, and generate the reconstructed voltage v ′ and the predicted temperature T pred corresponding to the current and the SOC by a decoder;

[0024] The physical simulation module is used to simulate the voltage response of the lithium battery according to the current and the SOC in the multi-modal monitoring data by an ECM model, and output the physical model predicted voltage v ” ;

[0025] The multi-source information fusion module is used to fuse the reconstructed voltage v ′ and the physical model predicted voltage v ” by a projection layer to obtain the final predicted voltage V pred ;

[0026] The report output module is used to calculate the temperature average M T and the temperature standard deviation S T between the predicted temperature T pred and the real temperature T true , and the voltage average M V and the voltage standard deviation S V between the predicted voltage V pred and the real voltage V true by a sliding window.and output a monitoring report carrying the temperature average value, the temperature standard deviation, the voltage average value and the voltage standard deviation;

[0027] The calculation formula of the temperature average value, the temperature standard deviation, the voltage average value and the voltage standard deviation is:

[0028]

[0029]

[0030] Wherein, N represents the time length of the sliding window; represents the predicted temperature T pred at t moment; represents the real temperature T true at t moment, that is, the temperature in the multi-modal monitoring data; represents the predicted voltage V pred at t moment; represents the real voltage V true at t moment, that is, the voltage in the multi-modal monitoring data.

[0031] Further, the step Step3 is specifically:

[0032] Collecting real-time multi-modal monitoring data of the lithium battery, pre-processing each of the real-time multi-modal monitoring data through a streaming computing engine, including at least normalization, noise reduction and missing value filling, inputting the pre-processed real-time multi-modal monitoring data into a deployed thermal runaway monitoring model, and obtaining a real-time monitoring report carrying a temperature average value, a temperature standard deviation, a voltage average value and a voltage standard deviation through hardware acceleration technology for inference of the thermal runaway monitoring model.

[0033] Further, the step Step4 is specifically:

[0034] Based on the monitored temperature during the normal charging process of the lithium battery, calculating the average value and the standard deviation as a temperature average threshold M1 and a temperature standard deviation threshold S1; based on the monitored voltage during the normal charging process of the lithium battery, calculating the average value and the standard deviation as a voltage average threshold M2 and a voltage standard deviation threshold S2;

[0035] When S T <S1, S V <S2, M T <M1 and M V <M2, it indicates that the lithium battery is running normally, and the monitoring continues;

[0036] When S T <S1, S V <S2, and M T ≥M1 or M V≥ M2, a second-level thermal runaway early warning is executed;

[0037] When S T ≥ S1 or S V ≥ S2, and M T < M1 and M V < M2, a first-level thermal runaway early warning is executed;

[0038] When S T ≥ S1 or S V ≥ S2, and M T ≥ M1 or M V ≥ M2, a third-level thermal runaway early warning is executed.

[0039] In a second aspect, the present application provides a multi-modal coupled lithium battery thermal runaway multi-level early warning system, comprising the following modules:

[0040] A data set construction module is configured to collect a large amount of historical multi-modal monitoring data of lithium batteries during operation, and construct a data set after preprocessing each of the historical multi-modal monitoring data;

[0041] A thermal runaway monitoring model training module is configured to create a thermal runaway monitoring model based on a time-frequency feature extraction module, a feature fusion module, a reconstruction decoding module, a physical simulation module, a multi-source information fusion module, and a report output module, and train and deploy the thermal runaway monitoring model through the data set;

[0042] A real-time monitoring module is configured to collect real-time multi-modal monitoring data of lithium batteries, and input the preprocessed real-time multi-modal monitoring data into the deployed thermal runaway monitoring model to obtain a real-time monitoring report carrying a temperature average value, a temperature standard deviation, a voltage average value, and a voltage standard deviation;

[0043] A thermal runaway early warning module is configured to perform thermal runaway early warning on lithium batteries based on the real-time monitoring report.

[0044] Further, the data set construction module is specifically configured to:

[0045] Collect a large amount of historical multi-modal monitoring data of lithium batteries during operation, wherein the historical multi-modal monitoring data at least includes voltage, current, temperature, and SOC;

[0046] Construct a data set after preprocessing each of the historical multi-modal monitoring data, wherein the preprocessing at least includes normalization, noise reduction, and missing value filling.

[0047] Further, in the thermal runaway monitoring model training module, the time-frequency feature extraction module is constructed based on a time domain feature extraction unit and a frequency domain feature extraction unit; the time domain feature extraction unit is configured to encode the multi-modal monitoring data by a Transformer-based time domain encoder T-encoder to obtain time domain features E1; the frequency domain feature extraction unit is configured to perform fast Fourier transform or short-time Fourier transform on the multi-modal monitoring data to extract initial frequency domain features, and then encode the initial frequency domain features by a Transformer-based time domain encoder F-encoder to obtain frequency domain features E2;

[0048] The feature fusion module is configured to fuse the time domain features E1 and the frequency domain features E2 by a cross-attention mechanism to obtain time-frequency features E3;

[0049] The reconstruction decoding module is configured to combine the time-frequency features E3, the current and the SOC in the multi-modal monitoring data as prompt words, and generate a reconstructed voltage V ′ corresponding to the current and the SOC by a decoder. pred ;

[0050] The physical simulation module is configured to simulate the voltage response of the lithium battery according to the current and the SOC in the multi-modal monitoring data by an ECM model, and output a physical model predicted voltage V ” .

[0051] The multi-source information fusion module is configured to fuse the reconstructed voltage V ′ and the physical model predicted voltage V ” by a projection layer to obtain a final predicted voltage V pred .

[0052] The report output module is configured to calculate the temperature average M pred and the temperature standard deviation S true between the predicted temperature T T and the real temperature T T , and the voltage average M pred and the voltage standard deviation S true between the predicted voltage V V and the real voltage V V by a sliding window, and output a monitoring report carrying the temperature average, the temperature standard deviation, the voltage average and the voltage standard deviation.

[0053] The calculation formulas of the temperature average, the temperature standard deviation, the voltage average and the voltage standard deviation are as follows:

[0054]

[0055]

[0056] wherein, N represents the time length of the sliding window; represents the predicted temperature T pred at time t; represents the real temperature T true at time t, i.e. the temperature in the multi-modal monitoring data; represents the predicted voltage V pred at time t; represents the real voltage V true at time t, i.e. the voltage in the multi-modal monitoring data.

[0057] Further, the real-time monitoring module is specifically configured to:

[0058] collect real-time multi-modal monitoring data of the lithium battery, pre-process each of the real-time multi-modal monitoring data through a streaming computing engine, including at least normalization, noise reduction, and missing value filling, input the pre-processed real-time multi-modal monitoring data into a deployed thermal runaway monitoring model, and obtain a real-time monitoring report carrying a temperature mean value, a temperature standard deviation, a voltage mean value, and a voltage standard deviation through inference of the thermal runaway monitoring model by using a hardware acceleration technology.

[0059] Further, the thermal runaway early warning module is specifically configured to:

[0060] based on the monitored temperature during the normal charging process of the lithium battery, calculate the mean value and the standard deviation as a temperature mean threshold M1 and a temperature standard deviation threshold S1; based on the monitored voltage during the normal charging process of the lithium battery, calculate the mean value and the standard deviation as a voltage mean threshold M2 and a voltage standard deviation threshold S2;

[0061] when S T <S1, S V <S2, M T <M1 and M V <M2, it indicates that the lithium battery is running normally, and the monitoring continues;

[0062] when S T <S1, S V <S2, M T ≥M1 or M V ≥M2, a secondary thermal runaway early warning is performed;

[0063] when S T ≥S1 or S V ≥S2, M T <M1 and M V <M2, a primary thermal runaway early warning is performed;

[0064] when S T≥S1 or S V ≥S2, while M T ≥M1 or M V When the temperature reaches ≥ M2, a Level 3 thermal runaway warning will be issued.

[0065] The advantages of this invention are:

[0066] 1. By collecting a large amount of historical multimodal monitoring data during the operation of lithium batteries, a dataset is constructed after preprocessing each historical multimodal monitoring data. Then, a thermal runaway monitoring model is created based on time-frequency feature extraction, feature fusion, reconstruction decoding, physical simulation, multi-source information fusion, and report output modules. The thermal runaway monitoring model is trained and deployed using the dataset. Next, real-time multimodal monitoring data of lithium batteries is collected, and after preprocessing each real-time multimodal monitoring data, it is input into the deployed thermal runaway monitoring model to obtain a real-time monitoring report carrying the average temperature, temperature standard deviation, average voltage, and voltage standard deviation. Based on the real-time monitoring report, thermal runaway early warning is provided for the lithium battery. That is, by fusing multimodal monitoring data of voltage, current, temperature, and SOC during the operation of lithium batteries and using time-frequency feature extraction, feature fusion, reconstruction decoding, physical simulation, multi-source information fusion, and report output modules, a thermal runaway monitoring model is created. The feature extraction module combines a Transformer encoder with a cross-attention mechanism to deeply mine the temporal evolution trend and frequency domain periodicity. At the same time, it innovatively constructs a physical-data hybrid model—using a physical simulation module (ECM model) to provide interpretable constraints, a reconstruction decoding module to capture complex nonlinear relationships, and then achieving dynamic operating condition compensation through multi-source information fusion, significantly improving the model's adaptability to new operating conditions. Finally, it generates the mean and standard deviation indices of temperature and voltage based on a sliding window, and designs a multi-level early warning mechanism (sudden fluctuations are triggered by the standard deviation threshold, continuous deviations are captured by the mean threshold, and coupled anomalies are jointly judged), systematically solving bottlenecks such as single feature underreporting, insufficient model generalization, and physical model aging failure, thereby greatly improving the accuracy, operating condition adaptability, and reliability of lithium battery thermal runaway early warning.

[0067] 2. By integrating multiple monitoring data sources such as voltage, current, temperature, and SOC (State of Charge), it covers all dimensions of lithium battery operation status information. Compared with single-mode monitoring (such as only voltage or temperature), this multi-mode coupling method can more comprehensively capture early signals of thermal runaway and reduce false alarms and missed alarms. For example, the combined analysis of temperature standard deviation and voltage standard deviation can identify abnormal fluctuations and improve the detection sensitivity of potential faults.

[0068] 3. Use Transformer-based encoders (such as T-encoder and F-encoder) for time-domain and frequency-domain feature extraction, effectively capturing the dynamic time series and frequency characteristics of lithium batteries (such as through Fast Fourier Transform or Short-Time Fourier Transform); combine cross-attention mechanism to fuse time-domain features (E1) and frequency-domain features (E2), generate time-frequency features (E3), solve the problem of feature isolation in traditional methods, improve the model's recognition ability for complex operating modes, reduce feature redundancy, and thus improve the early warning accuracy.

[0069] 4. Simultaneously use data-driven reconstruction decoding modules and physics-based ECM models, and fuse them through the projection layer in the multi-source information fusion module. This hybrid modeling approach compensates for the shortcomings of pure data-driven models being susceptible to noise interference and the shortcomings of pure physical models making simplifying assumptions, ensuring that the predicted voltage is closer to the true value and enhancing the system's stability and generalization ability.

[0070] 5. Early warning based on statistical index threshold comparison, implementing a three-level early warning strategy (normal, first level, second level, third level); for example, triggering a third-level warning when the standard deviation and mean value simultaneously exceed the threshold. This hierarchical mechanism allows different responses (such as early warning or shutdown) to be taken according to the risk level, avoiding overreaction, reducing operation and maintenance costs, and improving system safety.

[0071] 6. Real-time data processing through streaming computing engines and hardware acceleration technologies (such as GPU or FPGA), achieving low-latency inference and report output; combined with sliding window calculation, it can efficiently process high-frequency monitoring data, supporting real-time monitoring of large-scale lithium battery systems (such as electric vehicles or energy storage power stations).

[0072] 7. Preprocessing steps include normalization, noise reduction, and missing value filling, effectively cleaning data, reducing noise impact, and improving model input quality; at the same time, hardware acceleration is used in model training and deployment, optimizing computing resource usage, reducing energy consumption, and enabling efficient operation in resource-constrained embedded systems.

[0073] 8. Output report provides temperature mean value, standard deviation, and other statistical indicators to present risks in quantitative form, facilitating operators to intuitively assess battery status. This standardized output is easy to integrate into existing BMS (Battery Management System) or cloud platforms, supporting automated decision-making, reducing manual intervention, and improving system intelligence level.

[0074] 9. Through the fusion of multi-modal monitoring data (such as voltage, current, temperature, SOC) in lithium battery operation, combined with innovative time-frequency feature extraction (Transformer-based time / frequency encoder) and cross-attention fusion mechanism, and in coordination with data-driven reconstruction decoding and physical simulation model (ECM), high-precision dynamic prediction of battery state is realized; it uses sliding window statistics (mean and standard deviation of temperature / voltage) to construct a multi-level early warning threshold system, realizes real-time risk assessment through streaming computing and hardware acceleration, significantly improves the accuracy, robustness and response efficiency of thermal runaway early warning, and supports hierarchical risk control, balancing the real-time and reliability of industrial deployment.

[0075] 10. Through the fusion of time and frequency domain information, the sensitivity of the model to abnormal fluctuations is improved; through frequency domain feature extraction and cross attention fusion, the model's ability to identify abnormal data is enhanced; through the combination of data-driven and physical model dual modeling mechanism, the prediction accuracy and robustness are improved; through the multi-source information fusion mechanism, the reliability of abnormal detection is improved; through the fusion of data-driven and physical model prediction results, the limitations of a single method are reduced; through the multi-level thermal runaway early warning mechanism, the accuracy of early warning is improved. BRIEF DESCRIPTION OF DRAWINGS

[0076] The application will be further described below with reference to the accompanying drawings and embodiments.

[0077] Fig. 1 is a flowchart of a multi-modal coupled lithium battery thermal runaway multi-level early warning method of the application.

[0078] Fig. 2 is a structural schematic diagram of a multi-modal coupled lithium battery thermal runaway multi-level early warning system of the application.

[0079] Fig. 3 is a schematic diagram of a thermal runaway monitoring model of the application.

[0080] Fig. 4 is a flowchart of a multi-level early warning of the application. DETAILED DESCRIPTION

[0081] The technical solutions in the embodiments of the present application have the following general idea: by fusing the multi-modal monitoring data of voltage, current, temperature and SOC in the operation of lithium batteries, using a time-frequency feature extraction module, combining a Transformer encoder and a cross-attention mechanism to deeply mine the time-domain evolution trend and the frequency-domain periodicity rule, innovatively constructing a physical-data hybrid model to provide an interpretable constraint by an ECM model and capture complex nonlinear relationships by a reconstruction decoding module, and then realizing dynamic working condition compensation through multi-source information fusion, the adaptability of the model to new working conditions is significantly improved; finally, the mean and standard deviation indexes of temperature and voltage are generated based on a sliding window, and a multi-level early warning mechanism is designed, thereby systematically solving the bottlenecks of single feature false reporting, model generalization deficiency, and physical model aging failure.

[0082] Please refer to Figs. 1 to 4 The preferred embodiment of the multi-modal coupled lithium battery thermal runaway multi-level early warning method of the present application comprises the following steps:

[0083] Step S1, a large amount of historical multi-modal monitoring data in the operation process of lithium batteries is collected, and a data set is constructed after pre-processing each of the historical multi-modal monitoring data;

[0084] Step S2, a thermal runaway monitoring model is created based on a time-frequency feature extraction module, a feature fusion module, a reconstruction decoding module, a physical simulation module, a multi-source information fusion module and a report output module, and the thermal runaway monitoring model is trained and deployed through the data set;

[0085] Step S3, real-time multi-modal monitoring data of the lithium battery is collected, and after pre-processing each of the real-time multi-modal monitoring data, the deployed thermal runaway monitoring model is inputted to obtain a real-time monitoring report carrying the average temperature, the temperature standard deviation, the average voltage and the voltage standard deviation;

[0086] Step S4, the lithium battery is warned of thermal runaway based on the real-time monitoring report.

[0087] The step S1 specifically comprises:

[0088] A large amount of historical multi-modal monitoring data in the operation process of lithium batteries is collected, and the historical multi-modal monitoring data at least includes voltage, current, temperature and SOC;

[0089] After pre-processing each of the historical multi-modal monitoring data at least including normalization, noise reduction and missing value filling, a data set is constructed. The data format of the data set is (B, S, L), wherein B represents batch size, S represents time step, and L represents feature dimension (the value is 4, i.e. 4 dimensions of voltage, current, temperature and SOC).

[0090] In the step S2, the time-frequency feature extraction module is constructed based on a time domain feature extraction unit and a frequency domain feature extraction unit; the time domain feature extraction unit is configured to encode the multi-modal monitoring data by a Transformer-based time domain encoder T-encoder to obtain time domain features E1; the frequency domain feature extraction unit is configured to perform fast Fourier transform (FFT) or short-time Fourier transform (STFT) on the multi-modal monitoring data to extract initial frequency domain features, and then encode the initial frequency domain features by a Transformer-based time domain encoder F-encoder to obtain frequency domain features E2;

[0091] The feature fusion module is configured to fuse the time domain features E1 and the frequency domain features E2 by a cross attention mechanism to obtain time-frequency features E3; that is, the time domain features E1 are taken as query, and the frequency domain features E2 are taken as key and value for fusion:

[0092]

[0093] wherein, softmax() represents a normalized exponential function; W q represents a weight matrix of the query; W k represents a weight matrix of the key; W v represents a weight matrix of the value; T represents transposition; d represents the dimension of the feature vector, i.e., the dimension of the frequency domain features E2 after projection as the key;

[0094] The reconstruction decoding module is configured to combine the time-frequency features E3, the current and the SOC in the multi-modal monitoring data as prompts, and generate reconstructed voltages v′ and predicted temperatures T pred corresponding to the current and the SOC by a decoder.

[0095] The physical simulation module is configured to simulate the voltage response of the lithium battery according to the current and the SOC in the multi-modal monitoring data by an ECM model to output a physical model predicted voltage v″; the ECM model adopts a recursive least squares method to identify parameters (Ohmic resistance R0, polarization resistance Rp, polarization capacitance Cp, etc.) online.

[0096] The multi-source information fusion module is configured to fuse the reconstructed voltage v′ and the physical model predicted voltage v″ by a projection layer to obtain a final predicted voltage V pred .

[0097] V pred = σ(W 1 V′ + W 2 V″ + b);

[0098] wherein, σ() represents an activation function; W 1 and W 2 represent weight coefficients; and b represents a bias term;

[0099] The report output module is configured to calculate, through a sliding window, a temperature mean M pred and a temperature standard deviation S true between the predicted temperature T T and the real temperature T T , a voltage mean M pred and a voltage standard deviation S true between the predicted voltage V V and the real voltage V V , and output a monitoring report carrying the temperature mean, the temperature standard deviation, the voltage mean and the voltage standard deviation.

[0100] The calculation formulae of the temperature mean, the temperature standard deviation, the voltage mean and the voltage standard deviation are as follows:

[0101]

[0102] wherein, N represents the time length of the sliding window; represents the predicted temperature T pred at time t; represents the real temperature T true at time t, i.e. the temperature in the multi-modal monitoring data; represents the predicted voltage V pred at time t; represents the real voltage V true at time t, i.e. the voltage in the multi-modal monitoring data.

[0103] The step S3 specifically includes:

[0104] The real-time multi-modal monitoring data of the lithium battery is collected, the real-time multi-modal monitoring data is preprocessed through a streaming computing engine, at least including normalization, noise reduction and missing value filling, the preprocessed real-time multi-modal monitoring data is input into a deployed thermal runaway monitoring model, and the thermal runaway monitoring model is inferred through a hardware acceleration technology to obtain a real-time monitoring report carrying a temperature mean, a temperature standard deviation, a voltage mean and a voltage standard deviation.

[0105] The step S4 specifically includes:

[0106] Based on the temperature monitored during normal charging of the lithium battery, the average value and standard deviation are calculated as the average temperature threshold M1 and the standard deviation temperature threshold S1; based on the voltage monitored during normal charging of the lithium battery, the average value and standard deviation are calculated as the average voltage threshold M2 and the standard deviation voltage threshold S2.

[0107] When S T <S1, S V <S2, M T <M1 and M V If the value is less than M2, it indicates that the lithium battery is operating normally; continue monitoring.

[0108] When S T <S1, S V <S2, at the same time M T ≥M1 or M V When M2 is ≥2, a Level 2 thermal runaway warning will be issued;

[0109] When S T ≥S1 or S V ≥S2, while M T <M1 and M V When the temperature is less than M2, a Level 1 thermal runaway warning will be issued.

[0110] When S T ≥S1 or S V ≥S2, while M T ≥M1 or M V When the temperature reaches ≥ M2, a Level 3 thermal runaway warning will be issued.

[0111] A preferred embodiment of the multi-modal coupled lithium battery thermal runaway multi-level early warning system of the present invention includes the following modules:

[0112] The dataset construction module is used to collect a large amount of historical multimodal monitoring data during the operation of lithium batteries, and to construct the dataset after preprocessing the historical multimodal monitoring data.

[0113] The thermal runaway monitoring model training module is used to create a thermal runaway monitoring model based on the time-frequency feature extraction module, feature fusion module, reconstruction decoding module, physical simulation module, multi-source information fusion module, and report output module. The thermal runaway monitoring model is trained and deployed using the dataset.

[0114] The real-time monitoring module is used to collect real-time multimodal monitoring data of lithium batteries. After preprocessing the real-time multimodal monitoring data, it is input into the deployed thermal runaway monitoring model to obtain a real-time monitoring report carrying the average temperature, temperature standard deviation, average voltage, and voltage standard deviation.

[0115] The thermal runaway early warning module is used to provide early warning of thermal runaway of lithium batteries based on the real-time monitoring report.

[0116] The data set construction module is specifically used for:

[0117] A large amount of historical multi-modal monitoring data of lithium batteries during operation is collected, and the historical multi-modal monitoring data at least includes voltage, current, temperature and SOC;

[0118] After the historical multi-modal monitoring data is preprocessed at least including normalization, noise reduction and missing value filling, a data set is constructed. The data format of the data set is (B, S, L), wherein B represents batch size, S represents time step, and L represents feature dimension (the value is 4, i.e. 4 dimensions of voltage, current, temperature and SOC).

[0119] In the thermal runaway monitoring model training module, the time-frequency feature extraction module is constructed based on a time domain feature extraction unit and a frequency domain feature extraction unit; the time domain feature extraction unit is used for encoding multi-modal monitoring data by a time domain encoder T-encoder based on a Transformer to obtain time domain features E1; the frequency domain feature extraction unit is used for extracting initial frequency domain features by fast Fourier transform (FFT) or short-time Fourier transform (STFT) on the multi-modal monitoring data, and then encoding the initial frequency domain features by a time domain encoder F-encoder based on a Transformer to obtain frequency domain features E2;

[0120] The feature fusion module is used for fusing the time domain features E1 and the frequency domain features E2 by a cross attention mechanism to obtain time-frequency features E3; that is, the time domain features E1 are taken as query (query), and the frequency domain features E2 are taken as key (key) and value (value) for fusion:

[0121]

[0122] Wherein, softmax() represents a normalized exponential function; W q represents a weight matrix of query; W k represents a weight matrix of key; W v represents a weight matrix of value; T represents transposition; d represents the dimension of a feature vector, i.e. the dimension of the frequency domain features E2 after projection as key;

[0123] The reconstruction decoding module is used for combining the time-frequency features E3 with the current and SOC in the multi-modal monitoring data as prompts (prompts), and generating reconstructed voltage v' and predicted temperature T pred corresponding to the current and SOC by a decoder (decoder).

[0124] The physical simulation module is used to simulate the voltage response of a lithium battery based on the current and SOC data from multimodal monitoring data using an ECM model, and outputs the predicted voltage v″ from the physical model. The ECM model uses the recursive least squares method to identify parameters (ohmic internal resistance R0, polarization resistance Rp, polarization capacitance Cp, etc.) online.

[0125] The multi-source information fusion module is used to fuse the reconstructed voltage v′ and the physical model predicted voltage v″ through the projection layer to obtain the final predicted voltage V. pred ;

[0126] V pred =σ(W 1 V′+W 2 V″+b);

[0127] Where σ() represents the activation function; W 1 and W 2 All represent weighting coefficients; b represents the bias term;

[0128] The report output module is used to calculate the predicted temperature T through a sliding window. pred With the actual temperature T true The average temperature M between T and temperature standard deviation S T Predicted voltage V pred With the actual voltage V true The average voltage M between V and voltage standard deviation S V It outputs a monitoring report carrying the average temperature, standard deviation of temperature, average voltage, and standard deviation of voltage.

[0129] The formulas for calculating the average temperature, standard deviation of temperature, average voltage, and standard deviation of voltage are as follows:

[0130]

[0131] Where N represents the time length of the sliding window; The predicted temperature T at time t. pred ; T represents the actual temperature at time t. true That is, temperature in multimodal monitoring data; V represents the predicted voltage at time t. pred ; V represents the actual voltage at time t. true That is, the voltage in the multimodal monitoring data.

[0132] The real-time monitoring module is specifically used for:

[0133] Real-time multimodal monitoring data of lithium batteries are collected. The real-time multimodal monitoring data are preprocessed by a streaming computing engine, including at least normalization, noise reduction and missing value filling. The preprocessed real-time multimodal monitoring data are input into a deployed thermal runaway monitoring model. The thermal runaway monitoring model is inferred by hardware acceleration technology to obtain a real-time monitoring report carrying the average temperature, temperature standard deviation, average voltage and voltage standard deviation.

[0134] The thermal runaway early warning module is specifically used for:

[0135] Based on the temperature monitored during normal charging of the lithium battery, the average value and standard deviation are calculated as the average temperature threshold M1 and the standard deviation temperature threshold S1; based on the voltage monitored during normal charging of the lithium battery, the average value and standard deviation are calculated as the average voltage threshold M2 and the standard deviation voltage threshold S2.

[0136] When S T <S1, S V <S2, M T <M1 and M V If the value is less than M2, it indicates that the lithium battery is operating normally; continue monitoring.

[0137] When S T <S1, S V <S2, at the same time M T ≥M1 or M V When M2 is ≥2, a Level 2 thermal runaway warning will be issued;

[0138] When S T ≥S1 or S V ≥S2, while M T <M1 and M V When the temperature is less than M2, a Level 1 thermal runaway warning will be issued.

[0139] When S T ≥S1 or S V ≥S2, while M T ≥M1 or M V When the temperature reaches ≥ M2, a Level 3 thermal runaway warning will be issued.

[0140] In summary, the advantages of this invention are as follows:

[0141] 1、By collecting a large number of historical multi-modal monitoring data of lithium batteries during operation, the data set is constructed after preprocessing each historical multi-modal monitoring data; then a thermal runaway monitoring model is created based on a time-frequency feature extraction module, a feature fusion module, a reconstruction decoding module, a physical simulation module, a multi-source information fusion module and a report output module, and the thermal runaway monitoring model is trained and deployed through the data set; then the real-time multi-modal monitoring data of the lithium battery is collected, and the deployed thermal runaway monitoring model is inputted after preprocessing each real-time multi-modal monitoring data, to obtain a real-time monitoring report carrying temperature average, temperature standard deviation, voltage average and voltage standard deviation, and the lithium battery is warned of thermal runaway based on the real-time monitoring report; that is, by fusing the multi-modal monitoring data of voltage, current, temperature and SOC of the lithium battery during operation, and using the time-frequency feature extraction module combined with the Transformer encoder and the cross attention mechanism to deeply mine the time domain evolution trend and the frequency domain periodicity, a physical-data hybrid model is innovatively constructed to provide explainability constraints with the physical simulation module (ECM model), capture complex nonlinear relationships with the reconstruction decoding module, and realize dynamic working condition compensation through multi-source information fusion, significantly improving the model's adaptability to new working conditions; finally, the mean and standard deviation indicators of temperature and voltage are generated based on the sliding window, and a multi-level warning mechanism is designed (sudden fluctuations are triggered by the standard deviation threshold, continuous deviation is captured by the mean threshold, and coupled abnormalities are jointly determined), which systematically solves the bottlenecks of single feature false alarm, model generalization deficiency, and physical model aging failure, thereby greatly improving the accuracy, working condition adaptability and reliability of lithium battery thermal runaway warning.

[0142] 2、By integrating various monitoring data sources such as voltage, current, temperature, SOC (State of Charge), etc., covering the full-dimensional state information of lithium battery operation, compared with single modal monitoring (such as only voltage or temperature), this multi-modal coupling method can more comprehensively capture early signals of thermal runaway, reducing false alarms and omissions; for example, the combined analysis of temperature standard deviation and voltage standard deviation can identify abnormal fluctuations and improve the detection sensitivity of potential faults.

[0143] 3、Use the Transformer-based encoder (such as T-encoder and F-encoder) to extract time domain and frequency domain features, effectively capturing the dynamic time series and frequency characteristics of the lithium battery (such as through Fast Fourier Transform or Short-Time Fourier Transform); combine the cross attention mechanism to fuse time domain features (E1) and frequency domain features (E2), generate time-frequency features (E3), solve the problem of feature isolation in traditional methods, improve the model's recognition ability for complex operating modes, reduce feature redundancy, and thus improve the warning accuracy.

[0144] 4. Simultaneous use of data-driven reconstruction decoding module and physical-based ECM model, and fusion through projection layer in multi-source information fusion module. This hybrid modeling approach makes up for the shortcomings of pure data-driven models being susceptible to noise interference and pure physical models being deficient in simplifying assumptions, ensuring that the predicted voltage is closer to the true value, enhancing the stability and generalization ability of the system.

[0145] 5. Early warning based on statistical index threshold comparison, implementing a three-level early warning strategy (normal, first level, second level, third level). For example, when the standard deviation and mean value exceed the threshold value at the same time, the third level of early warning is triggered. This grading mechanism allows different responses (such as early warning or shutdown) to be taken according to the risk level, avoiding overreaction, reducing operation and maintenance costs, and improving system safety.

[0146] 6. Real-time data processing through streaming computing engine and hardware acceleration technology (such as GPU or FPGA), achieving low-latency inference and report output; combined with sliding window calculation, it can efficiently process high-frequency monitoring data, supporting real-time monitoring of large-scale lithium battery systems (such as electric vehicles or energy storage power stations).

[0147] 7. Preprocessing steps include normalization, noise reduction, and missing value filling, effectively cleaning data, reducing noise impact, and improving model input quality; at the same time, hardware acceleration is used in model training and deployment, optimizing computing resource usage, reducing energy consumption, and enabling efficient operation in resource-constrained embedded systems.

[0148] 8. Output report provides temperature mean value, standard deviation, and other statistical indicators in quantitative form to present risks, making it easy for operators to intuitively assess battery status. This standardized output is easy to integrate into existing BMS (Battery Management System) or cloud platforms, supporting automated decision-making, reducing manual intervention, and improving system intelligence level.

[0149] 9. Through the fusion of multi-modal monitoring data (such as voltage, current, temperature, SOC) in lithium battery operation, combined with innovative time-frequency feature extraction (Transformer-based time / frequency encoder) and cross-attention fusion mechanism, and in coordination with data-driven reconstruction decoding and physical simulation model (ECM), high-precision dynamic prediction of battery status is achieved. It uses sliding window statistics (temperature / voltage mean and standard deviation) to build a multi-level early warning threshold system, realizes real-time risk assessment through streaming computing and hardware acceleration, significantly improves the accuracy, robustness, and response efficiency of thermal runaway early warning, and supports hierarchical risk control, balancing real-time and reliability for industrial deployment.

[0150] 10. By fusing time domain and frequency domain information, the sensitivity of the model to abnormal fluctuations is improved; by frequency domain feature extraction and cross attention fusion, the model's ability to identify abnormal data is enhanced; by combining data-driven and physical model dual modeling mechanisms, the prediction accuracy and robustness are improved; by a multi-source information fusion mechanism, the reliability of anomaly detection is improved; by fusing the prediction results of data-driven and physical models, the limitations of a single method are reduced; by a multi-level thermal runaway early warning mechanism, the accuracy of early warning is improved.

[0151] Although the specific embodiments of the present application are described above, it should be understood by those skilled in the art that the specific examples described are illustrative only and are not intended to limit the scope of the present application, and equivalent modifications and variations made in accordance with the spirit of the present application by those skilled in the art should be covered within the scope of the claims of the present application.

Claims

1. A multi-modal coupled, multi-level early warning method for thermal runaway in lithium batteries, characterized in that: Includes the following steps: Step 1: Collect a large amount of historical multimodal monitoring data during the operation of lithium batteries, and construct a dataset after preprocessing the historical multimodal monitoring data. Step 2: Create a thermal runaway monitoring model based on the time-frequency feature extraction module, feature fusion module, reconstruction decoding module, physical simulation module, multi-source information fusion module, and report output module. Train and deploy the thermal runaway monitoring model using the dataset. Step 3: Collect real-time multimodal monitoring data of lithium battery, preprocess the real-time multimodal monitoring data and input it into the deployed thermal runaway monitoring model to obtain a real-time monitoring report carrying the average temperature, temperature standard deviation, average voltage and voltage standard deviation. Step 4: Based on the real-time monitoring report, provide an early warning of thermal runaway for the lithium battery.

2. The multi-modal coupled, multi-level early warning method for thermal runaway of lithium batteries as described in claim 1, characterized in that: Step 1 specifically involves: Collect a large amount of historical multimodal monitoring data during the operation of lithium batteries. The historical multimodal monitoring data includes at least voltage, current, temperature, and SOC. The dataset is constructed by preprocessing the historical multimodal monitoring data, including at least normalization, noise reduction, and missing value imputation.

3. The multi-modal coupled, multi-level early warning method for thermal runaway of lithium batteries as described in claim 1, characterized in that: In Step 2, the time-frequency feature extraction module is constructed based on a time-domain feature extraction unit and a frequency-domain feature extraction unit. The time-domain feature extraction unit is used to encode the multimodal monitoring data using a Transformer-based time-domain encoder (T-encoder) to obtain time-domain feature E1. The frequency-domain feature extraction unit is used to extract initial frequency-domain features from the multimodal monitoring data by performing a Fast Fourier Transform or Short-Time Fourier Transform, and then encode the initial frequency-domain features using a Transformer-based time-domain encoder (F-encoder) to obtain frequency-domain feature E2. The feature fusion module is used to fuse time-domain feature E1 and frequency-domain feature E2 through a cross-attention mechanism to obtain time-frequency feature E3; The reconstruction decoding module is used to combine the time-frequency feature E3 with the current and SOC in the multimodal monitoring data as prompt words, and generate the reconstruction voltage v corresponding to the current and SOC through the decoder. ′ and predicted temperature T pred ; The physical simulation module is used to simulate the voltage response of a lithium battery using an ECM model based on the current and SOC data from multimodal monitoring data, and outputs a physical model-predicted voltage V. ” ; The multi-source information fusion module is used to process the reconstructed voltage V through the projection layer. ′ And the physical model predicts the voltage V ” The fusion process yields the final predicted voltage V. pred ; The report output module is used to calculate the predicted temperature T through a sliding window. pred With the actual temperature T true The average temperature M between T and temperature standard deviation S T Predicted voltage V pred With the actual voltage V true The average voltage M between V and voltage standard deviation S V It outputs a monitoring report carrying the average temperature, standard deviation of temperature, average voltage, and standard deviation of voltage. The formulas for calculating the average temperature, standard deviation of temperature, average voltage, and standard deviation of voltage are as follows: Where N represents the time length of the sliding window; The predicted temperature T at time t. pred ; T represents the actual temperature at time t. true That is, temperature in multimodal monitoring data; V represents the predicted voltage at time t. pred ; V represents the actual voltage at time t. true That is, the voltage in the multimodal monitoring data.

4. The multi-modal coupled, multi-level early warning method for thermal runaway of lithium batteries as described in claim 1, characterized in that: Step 3 specifically involves: Real-time multimodal monitoring data of lithium batteries are collected. The real-time multimodal monitoring data are preprocessed by a streaming computing engine, including at least normalization, noise reduction and missing value filling. The preprocessed real-time multimodal monitoring data are input into a deployed thermal runaway monitoring model. The thermal runaway monitoring model is inferred by hardware acceleration technology to obtain a real-time monitoring report carrying the average temperature, temperature standard deviation, average voltage and voltage standard deviation.

5. The multi-modal coupled lithium battery thermal runaway multi-level early warning method as described in claim 1, characterized in that: Step 4 specifically involves: Based on the temperature monitored during normal charging of the lithium battery, the average value and standard deviation are calculated as the average temperature threshold M1 and the standard deviation temperature threshold S1; based on the voltage monitored during normal charging of the lithium battery, the average value and standard deviation are calculated as the average voltage threshold M2 and the standard deviation voltage threshold S2. When S T <S1, S V <S2, M T <M1 and M V If the value is less than M2, it indicates that the lithium battery is operating normally; continue monitoring. When S T <S1, S V <S2, at the same time M T ≥M1 or M V When M2 is ≥2, a Level 2 thermal runaway warning will be issued; When S T ≥S1 or S V ≥S2, while M T <M1 and M V When the temperature is less than M2, a Level 1 thermal runaway warning will be issued. When S T ≥S1 or S V ≥S2, while M T ≥M1 or M V When the temperature reaches ≥ M2, a Level 3 thermal runaway warning will be issued.

6. A multi-modal coupled multi-level early warning system for thermal runaway of a lithium battery, characterized in that: Includes the following modules: The dataset construction module is used to collect a large amount of historical multimodal monitoring data during the operation of lithium batteries, and to construct the dataset after preprocessing the historical multimodal monitoring data. The thermal runaway monitoring model training module is used to create a thermal runaway monitoring model based on the time-frequency feature extraction module, feature fusion module, reconstruction decoding module, physical simulation module, multi-source information fusion module, and report output module. The thermal runaway monitoring model is trained and deployed using the dataset. The real-time monitoring module is used to collect real-time multimodal monitoring data of lithium batteries. After preprocessing the real-time multimodal monitoring data, it is input into the deployed thermal runaway monitoring model to obtain a real-time monitoring report carrying the average temperature, temperature standard deviation, average voltage, and voltage standard deviation. The thermal runaway early warning module is used to provide early warning of thermal runaway of lithium batteries based on the real-time monitoring report.

7. A multi-modal coupled lithium battery thermal runaway multi-level early warning system as described in claim 6, characterized in that: The dataset construction module is specifically used for: Collect a large amount of historical multimodal monitoring data during the operation of lithium batteries. The historical multimodal monitoring data includes at least voltage, current, temperature, and SOC. The dataset is constructed by preprocessing the historical multimodal monitoring data, including at least normalization, noise reduction, and missing value imputation.

8. A multi-modal coupled lithium battery thermal runaway multi-level early warning system as described in claim 6, characterized in that: In the thermal runaway monitoring model training module, the time-frequency feature extraction module is constructed based on a time-domain feature extraction unit and a frequency-domain feature extraction unit. The time-domain feature extraction unit is used to encode the multimodal monitoring data using a Transformer-based time-domain encoder (T-encoder) to obtain time-domain features E1. The frequency-domain feature extraction unit is used to extract initial frequency-domain features from the multimodal monitoring data by performing a Fast Fourier Transform or Short-Time Fourier Transform, and then encode the initial frequency-domain features using a Transformer-based time-domain encoder (F-encoder) to obtain frequency-domain features E2. The feature fusion module is used to fuse time-domain feature E1 and frequency-domain feature E2 through a cross-attention mechanism to obtain time-frequency feature E3; The reconstruction decoding module uses the time-frequency feature E3 combined with the current and SOC in the multimodal monitoring data as prompt words, and generates the reconstruction voltage V corresponding to the current and SOC through the decoder. ′ and predicted temperature T pred ; The physical simulation module is used to simulate the voltage response of a lithium battery using an ECM model based on the current and SOC data from multimodal monitoring data, and outputs a physical model-predicted voltage V. ” ; The multi-source information fusion module is used to process the reconstructed voltage v through the projection layer. ′ And the physical model predicts the voltage v ” The fusion process yields the final predicted voltage V. pred ; The report output module is used to calculate the predicted temperature T through a sliding window. pred With the actual temperature T true The average temperature M between T and temperature standard deviation S T Predicted voltage V pred With the actual voltage V true The average voltage M between V and voltage standard deviation S V It outputs a monitoring report carrying the average temperature, standard deviation of temperature, average voltage, and standard deviation of voltage. The formulas for calculating the average temperature, standard deviation of temperature, average voltage, and standard deviation of voltage are as follows: Where N represents the time length of the sliding window; The predicted temperature T at time t. pred ; T represents the actual temperature at time t. true That is, temperature in multimodal monitoring data; V represents the predicted voltage at time t. pred ; V represents the actual voltage at time t. true That is, the voltage in the multimodal monitoring data.

9. A multi-modal coupled lithium battery thermal runaway multi-level early warning system as described in claim 6, characterized in that: The real-time monitoring module is specifically used for: Real-time multimodal monitoring data of lithium batteries are collected. The real-time multimodal monitoring data are preprocessed by a streaming computing engine, including at least normalization, noise reduction and missing value filling. The preprocessed real-time multimodal monitoring data are input into a deployed thermal runaway monitoring model. The thermal runaway monitoring model is inferred by hardware acceleration technology to obtain a real-time monitoring report carrying the average temperature, temperature standard deviation, average voltage and voltage standard deviation.

10. A multi-modal coupled lithium battery thermal runaway multi-level early warning system as described in claim 6, characterized in that: The thermal runaway early warning module is specifically used for: Based on the temperature monitored during normal charging of the lithium battery, the average value and standard deviation are calculated as the average temperature threshold M1 and the standard deviation temperature threshold S1; based on the voltage monitored during normal charging of the lithium battery, the average value and standard deviation are calculated as the average voltage threshold M2 and the standard deviation voltage threshold S2. When S T <S1, S V <S2, M T <M1 and M V If the value is less than M2, it indicates that the lithium battery is operating normally; continue monitoring. When S T <S1, S V <S2, at the same time M T ≥M1 or M V When M2 is ≥2, a Level 2 thermal runaway warning will be issued; When S T ≥S1 or S V ≥S2, while M T <M1 and M V When the temperature is less than M2, a Level 1 thermal runaway warning will be issued. When S T ≥S1 or S V ≥S2, while M T ≥M1 or M V When the temperature reaches ≥ M2, a Level 3 thermal runaway warning will be issued.

Images