A neural network-based lithium-sulfur battery thermal runaway monitoring method and system

Abstract

The application provides a lithium-sulfur battery thermal runaway monitoring method and system based on a neural network, comprising the following steps: S1: collecting monomer battery state data, group-level heat dissipation data, and battery cabin temperature data and performing preprocessing; S2: respectively constructing a double-branch gating network, a multi-head attention enhancement mechanism, and a multi-scale residual enhancement network, and sequentially extracting double-mode battery features, battery heterogeneity features, and deep heterogeneity features; S3: extracting a temperature field correlation feature through a temperature field correlation gating fusion mechanism; then extracting monomer-whole temperature field conjugate features; S4: extracting cross-dimension fusion features through a double-feature internal correlation dynamic weighting mechanism; calculating the abnormal heat value of each monomer battery and the overall thermal runaway risk level; and S5: performing visual display on a monitoring terminal and triggering an early warning signal. The application can solve the problem of local overheating missed judgment caused by the inability of traditional methods to fuse monomer heterogeneity and overall temperature field correlation.

Description

TECHNICAL FIELD

[0001] The application belongs to the technical field of lithium-sulfur batteries, and particularly relates to a lithium-sulfur battery thermal runaway monitoring method and system based on a neural network. BACKGROUND

[0002] The thermal runaway identification technology of the aged lithium-sulfur battery pack of an energy storage power station is a key technology for avoiding thermal runaway risks in advance by identifying abnormal characteristics of the temperature field of the battery pack. Its core requirement is to accurately capture the overheating signs of locally aged single cells and to correlate and judge the overall temperature field state of the battery pack in the scenario where the single cells are unevenly aged due to long-term cycling. However, the current technical applications in this field still have significant shortcomings.

[0003] Traditional thermal runaway identification methods mainly include mechanism modeling based on heat conduction equations, alarm systems relying on hardware thresholds, and traditional machine learning methods. The mechanism modeling method needs to rely on precise battery thermal physical parameters to build a model, which can roughly predict the overall temperature trend, but ignores the local temperature differences caused by single cell aging heterogeneity, and cannot identify local overheating. The alarm system based on hardware thresholds triggers an alarm only by preset temperature and voltage thresholds. When the overall average temperature does not reach the threshold value but some aged single cells are already overheated, it is easy to miss the judgment. Traditional machine learning methods (such as SVM and BP neural network) do not require complex physical modeling, but rely on manual feature design, which is difficult to handle the complex nonlinear relationship between single cell heterogeneity and group-level temperature field, and cannot simultaneously identify local and overall abnormalities.

[0004] In recent years, deep learning technology has been gradually applied in this field, mainly using a single neural network model to extract local features such as single cell voltage and internal resistance, or to capture the overall temperature field correlation of the battery pack. However, these solutions still have obvious limitations: they lack the capture of the conduction correlation between single cells and the overall temperature field, and cannot determine whether local overheating will trigger overall risks, ultimately leading to the fact that existing deep learning solutions cannot effectively integrate single cell heterogeneity local features and overall temperature field global correlation. When the battery pack is unevenly aged due to single cell aging, the overall temperature is normal but the local single cell is overheated, which is easy to miss the judgment of thermal runaway risk, and it is difficult to meet the actual needs of energy storage power stations for fine thermal safety identification of aged lithium-sulfur battery packs. SUMMARY

[0005] In view of the problem that the traditional method cannot integrate single cell heterogeneity and overall temperature field correlation, which leads to easy missed judgment of local overheating, the purpose of the present application is to provide a lithium-sulfur battery thermal runaway monitoring method and system based on a neural network.

[0006] To achieve the above-mentioned purpose, the technical solutions of the present application are as follows:

[0007] A lithium-sulfur battery thermal runaway monitoring method based on a neural network, comprising the following steps:

[0008] S1: Collecting monomer battery state data, group-level heat dissipation data and battery cabin temperature data, preprocessing to obtain preprocessed monomer battery state data, preprocessed group-level heat dissipation data and preprocessed cabin temperature data;

[0009] S2: According to the preprocessed monomer battery state data, a double-branch gating network is constructed to extract double-mode battery features; a multi-head attention enhancement mechanism is then constructed to calculate battery heterogeneity features; finally, a multi-scale residual enhancement network is used to extract deep heterogeneity features;

[0010] S3: According to the deep heterogeneity features, the preprocessed group-level heat dissipation data and the preprocessed cabin temperature data, the temperature field correlation gating fusion mechanism is used to extract the temperature field correlation features; then, the deep heterogeneity features are combined with the deep heterogeneity features to model the conduction correlation between the monomer and the whole temperature field, and the monomer-whole temperature field conjugate features are obtained;

[0011] The temperature field correlation features are obtained by the temperature field correlation gating fusion mechanism, which weights and fuses the deep heterogeneity features, the group-level heat dissipation features extracted from the preprocessed group-level heat dissipation data, and the cabin temperature features extracted from the preprocessed cabin temperature data, and then processes them through convolution;

[0012] S4: According to the deep heterogeneity features and the monomer-whole temperature field conjugate features, the cross-dimension fusion features are extracted through the double-feature internal correlation dynamic weighting mechanism; then, the abnormal heat value of each monomer battery and the overall thermal runaway risk level of the battery pack are calculated;

[0013] S5: According to the abnormal heat value of each monomer battery and the overall thermal runaway risk level of the battery pack, the visualization display is performed on the monitoring terminal, and the warning signal is triggered.

[0014] Further, the specific process of step S1 is as follows:

[0015] The monomer battery state data is collected by the battery management system, including the monomer cycle number and dynamic data; the type of dynamic data in the monomer battery state data is time series numerical data, including the monomer real-time voltage, the monomer real-time charging and discharging current and the monomer real-time internal resistance; after removing the abnormal values in the dynamic data in the monomer battery state data by the three times standard deviation method, the linear normalization is used to unify the data magnitude, and the preprocessed monomer battery state data is obtained;

[0016] The group-level heat dissipation data is collected by the heat dissipation system controller, and the data type is time series numerical data, including the heat dissipation fan speed and the cooling liquid circulation flow; the sliding window filtering method is used to denoise the group-level heat dissipation data, and the linear interpolation is used to complete the missing values, and the preprocessed group-level heat dissipation data is obtained;

[0017] The battery cabin temperature data is collected by an in-cabin temperature sensor, and the data type is time series numerical data; then Gaussian filtering is used for denoising, and range clipping is used to remove outliers to obtain the preprocessed in-cabin temperature data.

[0018] Further, in step S2, the specific process of extracting the dual-modal battery feature is:

[0019] According to the preprocessed single battery state data, a dual-branch gating network is constructed, one-dimensional convolution layers are used to extract the dynamic data preliminary features of each single battery, multi-layer perceptrons are used to extract the cycle number preliminary features of each single battery, a dynamic data gate is used to weight process the dynamic data preliminary features of all single batteries, a cycle number gate is used to weight process the cycle number preliminary features of all single batteries, the weighted two types of features are added element by element, and then a convolution layer is used for processing to obtain the dual-modal battery feature.

[0020] Further, in step S2, the specific process of calculating the battery heterogeneity feature is:

[0021] According to the dual-modal battery feature and the single cycle number, a multi-head attention enhancement mechanism is constructed to calculate the aging factor of each single battery; based on the aging factors of all single batteries and the dual-modal battery feature, a multi-head attention weight matrix is calculated; a weight calibration coefficient is calculated and the multi-head attention weight matrix is calibrated; the dual-modal battery feature is weighted processed through the calibrated multi-head attention weight matrix to obtain the battery heterogeneity feature.

[0022] Further, in step S2, the specific process of calculating the deep heterogeneity feature is:

[0023] The battery heterogeneity feature is input into a multi-scale residual enhancement network, the battery heterogeneity feature is first reshaped in dimension, three types of scale convolution features are extracted through three convolution layers with different size convolution kernels respectively, the three types of scale convolution features are added element by element to obtain a multi-scale fusion convolution feature, the multi-scale fusion convolution feature is processed by residual to obtain a residual feature, and the residual feature is processed by global average pooling to obtain the deep heterogeneity feature.

[0024] Further, the specific process of step S2 is:

[0025] S21: According to the preprocessed single battery state data, a dual-branch gating network is constructed, and the dual-modal battery feature is extracted, and the calculation method is:

[0026]

[0027]

[0028]

[0029]

[0030]

[0031] in, This represents the preliminary dynamic data characteristics of the i-th individual cell, where i is the index of the individual cell. It is a one-dimensional convolutional layer. This refers to the dynamic data of the i-th individual cell in the preprocessed individual cell state data. This represents a preliminary characteristic of the cycle number for the i-th individual cell. It is a multilayer perceptron. This represents the cycle number of the i-th cell in the preprocessed cell state data. For dynamic data gates, For the Sigmoid function, For continuous splicing operations, the splicing range is from i=1 to i=N, where N is the total number of individual cells. For loop count gate, It is a dual-mode battery feature. It is a convolutional layer. For Hadama accumulation, This is an element-wise addition;

[0032] S22: Based on the characteristics of the dual-mode battery and the cycle number of individual cells, a multi-head attention enhancement mechanism is constructed. Weight calibration coefficients are calculated and features are weighted to obtain the battery heterogeneity characteristics. The calculation method is as follows:

[0033]

[0034]

[0035]

[0036]

[0037] in, Let be the aging factor of the i-th individual cell. This represents the aging threshold for the number of battery cycles. This refers to the maximum number of cycles for a single lithium-sulfur battery cell. This is a multi-head attention weight matrix. As a multi-head attention mechanism, For weight calibration coefficients, This is a characteristic of battery heterogeneity. This is a transpose operation;

[0038] S23: Input the battery heterogeneity features into a multi-scale residual enhancement network to extract deep heterogeneity features. The calculation method is as follows:

[0039]

[0040]

[0041]

[0042]

[0043]

[0044]

[0045]

[0046] in, The heterogeneity characteristics of the battery after dimensional reshaping. For dimensional reshaping operations, These are convolutional features at the first, second, and third scales, respectively. These are convolutional layers with kernel sizes of 1×1, 3×1, and 5×1, respectively. To fuse convolutional features at multiple scales, For residual characteristics, This is a characteristic of deep heterogeneity. This is global average pooling.

[0047] It should be further explained that the state data of lithium-sulfur battery cells in the energy storage cabinet in this scenario includes two different types of information: dynamic data and cycle count. Dynamic data has time-series variation characteristics, while cycle count is a static feature. The feature expression forms, information dimensions, and correlation logic with thermal risk of the two types of data are significantly different. When directly fusing and extracting features from the two types of data, problems such as mutual interference of features and obscuring of key information can easily occur, making it impossible to accurately separate the features that are effective for thermal risk identification. This is the core difficulty of feature extraction in this scenario.

[0048] Step S21 of this invention constructs a dual-branch gating network, achieving efficient feature extraction through multi-stage synergy: First, considering the attribute differences between dynamic data and loop counts, two independent branches are designed to extract features separately. A one-dimensional convolutional layer can specifically capture temporal correlation features in dynamic data, while a multilayer perceptron adapts to the feature expression needs of static data such as loop counts, ensuring that the initial features of both types of data are fully extracted and avoiding the adaptation defects of a single extraction structure for different attribute data. Second, the initial features of all units are weighted through dynamic data gating and loop count gating, and the gating mechanism adaptively adjusts based on the overall feature distribution of the units. The process involves weighting the data to highlight key features related to thermal risk, suppressing irrelevant and redundant information, and optimizing the initial feature selection. Finally, the two weighted features are fused element-wise and further integrated through convolutional layers to achieve deep complementarity between dynamic and static features. Throughout the process, the branch extraction stage ensures the integrity of features with different attributes, the gating weighting stage achieves accurate feature selection, and the fusion processing stage completes the collaborative enhancement of information. The three stages are interconnected and work together to solve the adaptability problem of feature extraction for data with different attributes and avoid feature interference through dynamic selection, ultimately achieving high-quality extraction of dual-modal battery features.

[0049] It should be further explained that the thermal risk of lithium-sulfur batteries in energy storage cabinets is directly related to the aging heterogeneity of individual cells. The differences in dual-modal characteristics of individual cells with different cycle counts contribute differently to thermal risk identification. However, the traditional multi-head attention mechanism only relies on the general correlation between features to calculate weights, lacking adaptation to battery-specific aging heterogeneity information. This leads to a disconnect between weight allocation and the needs of thermal risk identification, making it difficult to accurately focus on the characteristic differences of different aged cells.

[0050] In step S22, this invention proposes a multi-head attention enhancement mechanism: First, unlike traditional multi-head attention which only uses the features themselves as computational input, this mechanism first calculates the aging factor of each cell based on the number of cycles per cell, quantifying battery aging heterogeneity into features that can participate in attention calculation, providing a specific reference for weight allocation, and making the focus of attention more aligned with the actual needs of lithium-sulfur battery thermal risk identification; Second, after processing the aging factors of all cells, they are input into the multi-head attention mechanism along with the dual-modal battery features, so that the calculation of the weight matrix not only relies on the general correlation between features, but also integrates the domain information of aging heterogeneity, realizing the differentiated allocation of dynamic weights to features of different aged cells; Finally, a weight calibration coefficient step is added, which calibrates the initially calculated attention weight matrix to correct the weight deviation caused by the lack of domain constraints in the traditional mechanism, further improving the accuracy of weight allocation. Throughout the process, the embedding of the aging factor solves the problem of insufficient domain adaptability in the traditional mechanism, and the weight calibration step makes up for the deficiency of insufficient weight optimization in the traditional mechanism. The two work together with multi-head attention calculation to ultimately achieve accurate extraction of battery heterogeneity features.

[0051] Furthermore, this invention constructs a multi-scale residual enhancement network: First, it processes the dimensionally reshaped battery heterogeneity features in parallel through convolutional layers with three different kernel sizes, capturing small-scale local heterogeneity, medium-scale correlated heterogeneity, and large-scale global heterogeneity features respectively, ensuring that no multi-scale features are missed. Second, it adds the multi-scale fused convolutional features to the features after secondary convolution through residual connections, effectively alleviating the gradient vanishing problem in deep networks, ensuring the integrity of heterogeneous features during deep transmission, and avoiding feature information decay. Finally, it performs gate enhancement on the residual features through activation functions and a dedicated multilayer perceptron, filtering out deep features key to thermal risk identification, and then further integrating and optimizing them through global average pooling. Throughout the process, multi-scale convolution achieves comprehensive feature coverage, residual structure ensures effective feature transmission, and gate enhancement and pooling operations complete the accurate selection and integration of features. The three work together to overcome the adaptation defects of traditional networks in multi-scale feature extraction and deep enhancement.

[0052] Furthermore, the specific process of step S3 is as follows:

[0053] S31: Based on the deep heterogeneity characteristics, preprocessed group-level heat dissipation data, and preprocessed cabin temperature data, a preliminary fusion of multi-source data is achieved through a temperature field correlation gating fusion mechanism to obtain temperature field correlation characteristics. The calculation method is as follows:

[0054]

[0055]

[0056]

[0057]

[0058] in, As a group-level heat dissipation feature, This is the preprocessed group-level heat dissipation data. The cabin temperature characteristics, The preprocessed cabin temperature data, For gated loop unit, For temperature field related gate, For splicing operations, The thermal conductivity coefficient of lithium-sulfur batteries, Temperature field correlation characteristics, These are the first, second, and third columns of data after tensor slicing of the temperature field correlation gate;

[0059] S32: Based on temperature field correlation characteristics and deep heterogeneity characteristics, model the conduction correlation between the individual and global temperature fields to obtain the conjugate characteristics of the individual-global temperature fields. The calculation method is as follows:

[0060]

[0061]

[0062] ,

[0063] in, This is the thermal conduction coupling matrix. For the Softmax function, For enhanced conduction coupling characteristics, These are the thermal capacity parameters for lithium-sulfur batteries. It exhibits a conjugate characteristic of the single-unit and global temperature fields.

[0064] It should be further explained that the thermal risk evolution of lithium-sulfur batteries in energy storage cabinets involves cross-scale correlation between the microscopic heterogeneity characteristics of individual cells and the overall macroscopic temperature field characteristics. Among them, deep heterogeneity characteristics reflect the differences in local thermal risk of individual cells, while group-level heat dissipation data and cabin temperature data reflect the overall temperature field regulation state. The dimensions, physical meanings and conduction logic of the two types of features are significantly different. Traditional feature extraction methods are prone to severing the heat conduction link, resulting in the extracted features failing to accurately map the synergistic relationship between the two and deviating from the actual laws of thermal diffusion of lithium-sulfur batteries.

[0065] This invention constructs a single-unit-global temperature field conjugate feature: First, in step S31, based on the temperature field correlation gating fusion mechanism, deep heterogeneity features, group-level heat dissipation features, and cabin temperature features are used as fusion inputs. The temperature field correlation gate is constrained by the thermal conductivity coefficient of the lithium-sulfur battery, enabling the gating unit to adaptively adjust the fusion weights of the three types of features, specifically highlighting key information related to heat conduction, suppressing invalid redundant interference, and achieving accurate preliminary fusion of multi-source cross-scale data to obtain temperature field correlation features that can carry basic temperature field correlation information. Subsequently, in step S32, based on the temperature field correlation features and deep heterogeneity features, a heat conduction coupling matrix is ​​constructed. This matrix quantifies the conduction correlation strength between single-unit features and the overall temperature field features, and then combines the thermal conductivity coefficient and heat capacity parameters of the lithium-sulfur battery. The process involves calculating the enhanced conduction coupling features to ensure that the conduction correlation modeling conforms to the physical laws of thermal diffusion in lithium-sulfur batteries. Finally, the enhanced conduction coupling features are fused element-by-element with the group-level heat dissipation features weighted by the thermal conduction coupling matrix to achieve deep synergy between individual heterogeneity information and overall temperature field conduction information. Throughout the process, the temperature field correlation gating fusion mechanism solves the initial adaptation and screening problem of multi-source cross-scale data, the thermal conduction coupling matrix realizes the quantification of the conduction correlation between individual cells and the overall temperature field, and the thermal physical parameters of lithium-sulfur batteries provide physical constraints for the modeling process. The synergistic effect of these three elements avoids the blindness of multi-source data fusion and achieves the organic connection between individual cell and overall temperature field information through conduction correlation modeling, ultimately obtaining conjugate features that can fully reflect the synergistic correlation between individual cell heterogeneity and overall temperature field.

[0066] Furthermore, the specific process of step S4 is as follows:

[0067] S41: Based on the deep heterogeneity characteristics and the conjugate characteristics of the individual-global temperature field, feature fusion is performed through a dynamic weighting mechanism based on the intrinsic correlation of dual features to obtain cross-dimensional fused features. The calculation method is as follows:

[0068]

[0069]

[0070]

[0071]

[0072]

[0073] in, The variance of the aging factor distribution of a single battery cell. For variance calculation, The temperature field conjugate characteristic stability coefficient is... For characteristic correlation strength factor, These are the weighting coefficients for the conjugate characteristics of the temperature field. For cross-dimensional fusion features;

[0074] S42: Based on cross-dimensional fusion characteristics and individual cell aging factors, calculate the abnormal heat value of each individual cell. The calculation method is as follows:

[0075]

[0076]

[0077]

[0078] in, For the cross-dimensional fusion features of the i-th individual cell, Tensor slicing operation for extracting features of the i-th single entity. The aging-weighted thermal characteristics of the i-th individual cell are shown below. Let be the abnormal heat value of the i-th individual cell. For batch normalization operations, The safe temperature field threshold for lithium-sulfur batteries, This is the critical temperature threshold for thermal runaway.

[0079] S43: Based on the abnormal thermal values ​​of individual cells and combined with the conjugate characteristics of the cell-to-system temperature field, calculate the overall thermal runaway risk level of the battery pack. The calculation method is as follows:

[0080]

[0081]

[0082]

[0083]

[0084] in, This represents the maximum abnormal heat value of all individual battery cells. To obtain the maximum value, This is the average of the abnormal heat values ​​of all individual cells. Calculated for the mean. For the overall thermal risk score, The overall thermal runaway risk level is as follows: This is the function for extracting the maximum index. and These are weighting coefficients, determined empirically.

[0085] It should be further explained that the cross-dimensional fusion features used for thermal risk identification must simultaneously carry two types of core information: deep heterogeneity features focus on the microscopic aging differences of individual cells, and cell-to-global temperature field conjugate features focus on the macroscopic temperature field correlation patterns. The information dimensions, core values, and correlation strength of the two types of features will dynamically change with the battery operating conditions. Traditional fusion methods use fixed weights or simple splicing, which cannot adapt to the dynamic changes in the intrinsic correlation between the two types of features. This can easily lead to an imbalance problem in the fusion features, such as microscopic information overload or macroscopic correlation weakening, and cannot accurately support the subsequent calculation of abnormal heat values.

[0086] This invention achieves cross-dimensional fusion feature extraction through a dynamic weighting mechanism based on the intrinsic correlation of dual features: First, the dispersion of individual cell aging differences is quantified by calculating the variance of the distribution of aging factors in individual cells. Simultaneously, the fluctuation state of macroscopic temperature field correlation features is quantified by the stability coefficient of the temperature field conjugate feature. These two types of quantitative indicators capture the intrinsic correlation attributes of the dual features from the two dimensions of microscopic heterogeneity and macroscopic stability, providing an objective basis for weight adjustment. Second, the two types of quantitative indicators are concatenated and a feature correlation strength factor is constructed through a multilayer perceptron to dynamically reflect the degree of synergistic correlation between aging differences and temperature field stability, avoiding the shortcomings of traditional fixed weights that cannot adapt to dynamic changes in features. Subsequently, based on the feature correlation strength factor, the temperature field conjugate feature weighting coefficients are dynamically generated through the activation function. Then, the deep heterogeneity features and the individual-to-global temperature field conjugate features are weighted separately through the Hadamard product. Finally, the two types of features are fused by element-wise addition. Throughout the process, the quantification of the aging factor distribution variance and the temperature field stability coefficient provides a reference for weight adjustment. The feature correlation strength factor realizes the dynamic mapping of the correlation attributes, and the dynamic weight allocation ensures the adaptive fusion of the two types of features. This avoids information interference between micro and macro features and achieves complementary enhancement of the two types of core information, ultimately resulting in a cross-dimensional fused feature that takes into account both individual differences and temperature field correlation.

[0087] It should be further explained that the overall thermal runaway risk assessment of lithium-sulfur battery packs in energy storage cabinets needs to take into account two core dimensions simultaneously: one is the extreme thermal risk of individual cells (abnormal heat of the most dangerous cell), and the other is the overall thermal risk baseline of the battery pack (average abnormal heat of all cells). Furthermore, the assessment results need to be corrected by combining the macroscopic temperature field correlation characteristics. Traditional assessment methods often rely on only a single indicator (such as the average heat value) or calculate the two types of indicators in isolation, ignoring the synergistic effect of extreme risks and the overall baseline, and also failing to incorporate temperature field stability information. This leads to a disconnect between the risk level assessment and the actual thermal runaway evolution law, and fails to accurately reflect the transmission logic of local extreme risks leading to overall runaway.

[0088] Step S43 of this invention first quantifies the extreme thermal risk of individual cells and the overall thermal risk baseline of the battery pack by calculating the maximum and average values ​​of abnormal thermal values ​​of all individual cells. Second, it normalizes the weighted fused maximum and average values ​​based on the thermal runaway hazard temperature threshold, and then corrects them using the Hadamard product through the temperature field conjugate characteristic stability coefficient. This ensures that the risk score reflects both the synergistic effect of the two thermal indicators and adapts to the stability state of the macroscopic temperature field—appropriately increasing the weight of extreme risks when the temperature field stability is poor, and emphasizing the reference value of the overall baseline risk when the temperature field is stable. Finally, the corrected values ​​are... The overall thermal risk score after positive correlation is concatenated with the conjugate features of the individual and overall temperature fields. Deep integration of features is achieved through multilayer perceptron and Softmax function. Then, the accurate risk level is obtained by extracting the maximum value index. In the whole process, the quantification of extreme values ​​and mean values ​​realizes the comprehensive coverage of multi-dimensional risks, the temperature field stability correction gives the assessment results macroscopic adaptability, and the feature integration and hierarchical operation complete the transformation of risk information. This not only solves the one-sidedness of single index assessment, but also improves the physical rationality of the assessment results through temperature field feature fusion, and finally achieves a level assessment that is highly matched with the actual thermal runaway risk.

[0089] Existing technologies typically calculate only the average or maximum value of abnormal heat values ​​in individual cells when calculating the overall thermal runaway risk level of a battery pack, without correcting for this by incorporating the conjugate characteristics of the cell-to-pack temperature field. This results in assessments that either underestimate the transmission impact of local extreme risks or overestimate the severity of the overall baseline risk, failing to comprehensively and accurately reflect the actual thermal risk state of the battery pack. In contrast, this invention innovatively integrates extreme values ​​and mean values ​​as a weighted risk indicator, while introducing a stability coefficient based on the conjugate characteristics of the temperature field for correction. By combining these characteristics with a deep integration and classification, the invention takes into account the synergistic impact of local extreme risks and the overall baseline risk, and improves the adaptability of the assessment through correction using macroscopic temperature field information. This solves the problem of traditional technologies being one-sided and detached from actual thermal evolution laws, resulting in a more valuable overall thermal runaway risk level.

[0090] Furthermore, the specific process of step S5 is as follows:

[0091] S51: Add the collection timestamp and battery pack number attributes to the abnormal heat values ​​of each individual battery cell and the overall thermal runaway risk level, and set the response threshold corresponding to each risk level to obtain basic thermal risk data with attribute labels.

[0092] S52: Perform format conversion on the thermal risk basic data with attribute identifiers, convert it into JSON data format compatible with the energy storage cabinet monitoring system, and obtain system-compatible standardized thermal risk data;

[0093] S53: Based on standardized thermal risk data, the local monitoring terminal of the energy storage cabinet displays the abnormal heat distribution of individual units and the overall risk level, and triggers the corresponding level of early warning signal according to the risk level.

[0094] The present invention also discloses a lithium-sulfur battery thermal runaway monitoring system based on neural networks, including a battery data acquisition module, a single cell feature extraction module, a temperature field correlation module, a risk level calculation module, and an early warning module;

[0095] The battery data acquisition module collects individual battery status data, group-level heat dissipation data, and battery compartment temperature data, and performs preprocessing to obtain preprocessed individual battery status data, preprocessed group-level heat dissipation data, and preprocessed compartment temperature data.

[0096] The single-cell feature extraction module constructs a dual-branch gating network for fusion processing based on the preprocessed single-cell state data to obtain dual-modal cell features; then it constructs a multi-head attention enhancement mechanism to calculate the cell heterogeneity features; finally, it extracts deep heterogeneity features based on a multi-scale residual enhancement network.

[0097] The temperature field correlation module extracts temperature field correlation features based on deep heterogeneity characteristics, preprocessed group-level heat dissipation data, and preprocessed cabin temperature data through a temperature field correlation gating fusion mechanism; then, combined with deep heterogeneity characteristics, it models the conduction correlation between the individual and overall temperature fields to obtain the individual-overall temperature field conjugate features.

[0098] The risk level calculation module extracts cross-dimensional fusion features based on deep heterogeneity characteristics and single-cell-to-system temperature field conjugate characteristics through a dynamic weighting mechanism of intrinsic correlation between dual features; then it calculates the abnormal heat value of each single cell and the overall thermal runaway risk level of the battery pack.

[0099] The early warning module visualizes the abnormal heat values ​​of each individual battery cell and the overall thermal runaway risk level of the battery pack on the monitoring terminal and triggers an early warning signal.

[0100] In summary, due to the adoption of the above technical solution, the beneficial effects of the present invention are:

[0101] (1) To address the problem of missed detection of local overheating caused by the inability of traditional methods to integrate the correlation between individual cell heterogeneity and overall temperature field, this invention extracts deep heterogeneity features layer by layer through a dual-branch gating network, a multi-head attention enhancement mechanism, and a multi-scale residual enhancement network to capture the differences in cell aging; then, through a temperature field correlation gating fusion mechanism, the individual cell and overall temperature field features are correlated and modeled to achieve organic connection between local features and overall temperature field information; finally, through a dual-feature intrinsic correlation dynamic weighting mechanism, the two types of core features are fused to calculate the abnormal heat value of the individual cell and the overall thermal runaway risk level of the battery pack. This avoids the defects of traditional mechanism modeling ignoring heterogeneity, hardware thresholding missing detection of local overheating, and single deep learning separating the correlation between local and overall features. It realizes the collaborative identification of local overheating and overall risk of aging lithium-sulfur battery packs, and meets the refined thermal safety protection requirements of energy storage power stations.

[0102] (2) To address the core challenge of the large differences in attributes between dynamic data and static data of cycle count of lithium-sulfur batteries in energy storage cabinets, which can easily lead to feature interference and obscuring of key information when directly fused, this invention innovatively constructs a dual-branch gating network. Features are extracted through two independent branches. A one-dimensional convolutional layer accurately captures the temporal correlation of dynamic data, and a multilayer perceptron adapts to the static features of cycle count, ensuring the integrity of both types of features. Then, the dynamic data gate and the cycle count gate are adaptively weighted and filtered to highlight key information of thermal risk and suppress redundancy. Finally, the dynamic and static features are deeply complementary, effectively solving the problem of adaptability of data extraction with different attributes, and successfully extracting high-quality dual-modal battery features.

[0103] (3) In view of the problem that traditional multi-head attention mechanisms lack adaptation to the heterogeneity information of lithium-sulfur battery aging, the weight allocation is out of sync with the needs of thermal risk identification, and it is difficult to focus on the differences in characteristics of different aged cells, this invention proposes an innovative multi-head attention enhancement mechanism. First, the aging factor is calculated based on the number of cycles of a single cell, and the heterogeneity information is quantified as the input for attention calculation, providing domain-specific reference. Then, the aging factor is processed and used together with the dual-mode battery features to participate in the weight calculation, realizing differentiated weight allocation. In addition, this invention also adds a weight calibration coefficient to correct the deviation, realizing the effective extraction of battery heterogeneity features.

[0104] (4) In view of the problem that traditional methods separate the microscopic heterogeneity characteristics of lithium-sulfur battery cells in energy storage cabinets from the overall macroscopic temperature field characteristics, resulting in the inability of extracted features to accurately map the synergistic relationship between the two and deviating from the actual laws of thermal diffusion, this invention constructs a conjugate feature of cell-overall temperature field; through the temperature field correlation gating fusion mechanism, it achieves accurate preliminary fusion of deep heterogeneity, group-level heat dissipation and cabin temperature characteristics, and uses the thermal conduction coupling matrix to quantify the conduction correlation strength between the two, and combines the thermal physical parameters of lithium-sulfur batteries to ensure the rationality of modeling, and finally realizes the organic connection of cell and overall temperature field information, fully reflects the synergistic relationship between the two, avoids the blindness of multi-source data fusion, solves the defect of traditional features deviating from the laws of thermal diffusion, and provides a more realistic core basis for thermal risk identification.

[0105] (5) In view of the problem that traditional fixed weight or simple splicing fusion methods cannot adapt to the dynamic changes of micro aging differences and single macro temperature field correlation, and are difficult to accurately support the calculation of abnormal heat value, this invention realizes cross-dimensional fusion feature extraction through a dynamic weighting mechanism of dual feature intrinsic correlation; by capturing the intrinsic correlation of two types of features through quantitative indicators, weighting coefficients adapted to the working conditions are dynamically generated, and finally fusion features that take into account both individual aging differences and overall temperature field correlation are obtained, solving the information imbalance defect of traditional fusion, and providing accurate and reliable feature support for subsequent abnormal heat value calculation.

[0106] (6) In view of the problem that traditional assessment methods rely on only a single indicator or isolated calculation, ignore the synergistic effect of individual extreme thermal risks and the overall risk baseline of the battery pack, and do not incorporate temperature field stability information, resulting in the assessment being out of touch with the actual thermal runaway evolution law, this invention optimizes the overall thermal runaway risk level assessment logic; by quantifying the maximum and average values ​​of abnormal heat values ​​to cover dual-dimensional risks, combined with the normalization of thermal runaway danger temperature threshold, and then using the temperature field conjugate feature stability coefficient for dynamic correction, and finally integrating the temperature field conjugate features for deep integration and classification; it not only solves the one-sidedness of traditional assessment, but also improves the physical rationality of the results, accurately reflects the transmission logic of local extreme risks leading to overall runaway, and achieves a refined assessment that is highly matched with the actual thermal risks. Detailed Implementation

[0107] To make the objectives, technical solutions, and advantages of the present invention clearer, the present invention will be further described in detail below with reference to the embodiments.

[0108] A neural network-based method for monitoring thermal runaway in lithium-sulfur batteries includes the following steps:

[0109] S1: The lithium-sulfur battery is a battery pack, consisting of several individual cells. Individual cell state data, pack-level heat dissipation data, and battery compartment temperature data are collected and preprocessed to obtain preprocessed individual cell state data, preprocessed pack-level heat dissipation data, and preprocessed compartment temperature data; including:

[0110] The battery management system collects individual battery status data, including the number of cycles and dynamic data. The dynamic data in the individual battery status data is time-series numerical data, including real-time voltage, real-time charge and discharge current, and real-time internal resistance of the individual battery. The dynamic data in the individual battery status data is processed by removing outliers using the three-standard-deviation method, and then the data volume is unified by linear normalization to obtain preprocessed individual battery status data.

[0111] Group-level heat dissipation data is collected by the heat dissipation system controller. The data type is time-series numerical data, including cooling fan speed and coolant circulation flow rate. The group-level heat dissipation data is denoised by sliding window filtering and then missing values ​​are filled by linear interpolation to obtain preprocessed group-level heat dissipation data.

[0112] Temperature data inside the battery compartment is collected by an in-cabin temperature sensor and is data in the form of time series numerical data. Gaussian filtering is then used to denoise the data, and outliers are removed by range clipping to obtain the preprocessed in-cabin temperature data.

[0113] S2: Based on the preprocessed single-cell state data, a dual-branch gating network is constructed to extract dual-modal cell features; then, a multi-head attention enhancement mechanism is constructed to calculate cell heterogeneity features; finally, based on a multi-scale residual enhancement network, deep heterogeneity features are extracted, including:

[0114] S21: Based on the preprocessed single-cell state data, a dual-branch gating network is constructed to extract dual-mode cell features. The calculation method is as follows:

[0115]

[0116]

[0117]

[0118]

[0119]

[0120] in, This represents the preliminary dynamic data characteristics of the i-th individual cell, where i is the index of the individual cell. It is a one-dimensional convolutional layer. This refers to the dynamic data of the i-th individual cell in the preprocessed individual cell state data. This represents a preliminary characteristic of the cycle number for the i-th individual cell. It is a multilayer perceptron. This represents the cycle number of the i-th cell in the preprocessed cell state data. For dynamic data gates, For the Sigmoid function, For continuous splicing operations, the splicing range is from i=1 to i=N, where N is the total number of individual cells. For loop count gate, It is a dual-mode battery feature. It is a convolutional layer. For Hadama accumulation, This is an element-wise addition;

[0121] S22: Based on the characteristics of the dual-mode battery and the cycle number of individual cells, a multi-head attention enhancement mechanism is constructed. Weight calibration coefficients are calculated and features are weighted to obtain the battery heterogeneity characteristics. The calculation method is as follows:

[0122]

[0123]

[0124]

[0125]

[0126] in, Let be the aging factor of the i-th individual cell. This represents the aging threshold for the number of battery cycles. This refers to the maximum number of cycles for a single lithium-sulfur battery cell. This is a multi-head attention weight matrix. As a multi-head attention mechanism, For weight calibration coefficients, This is a characteristic of battery heterogeneity. This is a transpose operation;

[0127] S23: Input the battery heterogeneity features into a multi-scale residual enhancement network to extract deep heterogeneity features. The calculation method is as follows:

[0128]

[0129]

[0130]

[0131]

[0132]

[0133]

[0134]

[0135] in, The heterogeneity characteristics of the battery after dimensional reshaping. For dimensional reshaping operations, These are convolutional features at the first, second, and third scales, respectively. These are convolutional layers with kernel sizes of 1×1, 3×1, and 5×1, respectively. To fuse convolutional features at multiple scales, For residual characteristics, This is a characteristic of deep heterogeneity. For global average pooling;

[0136] S3: Based on deep heterogeneity characteristics, preprocessed group-level heat dissipation data, and preprocessed cabin temperature data, temperature field correlation features are extracted through a temperature field correlation gating fusion mechanism; then, combined with deep heterogeneity characteristics, the conduction correlation between the individual and overall temperature fields is modeled to obtain the individual-overall temperature field conjugate features, including:

[0137] S31: Based on the deep heterogeneity characteristics, preprocessed group-level heat dissipation data, and preprocessed cabin temperature data, a preliminary fusion of multi-source data is achieved through a temperature field correlation gating fusion mechanism to obtain temperature field correlation characteristics. The calculation method is as follows:

[0138]

[0139]

[0140]

[0141]

[0142] in, As a group-level heat dissipation feature, This is the preprocessed group-level heat dissipation data. The cabin temperature characteristics, The preprocessed cabin temperature data, For gated loop unit, For temperature field related gate, For splicing operations, The thermal conductivity coefficient of lithium-sulfur batteries, Temperature field correlation characteristics, These are the first, second, and third columns of data after tensor slicing of the temperature field correlation gate;

[0143] S32: Based on temperature field correlation characteristics and deep heterogeneity characteristics, model the conduction correlation between the individual and global temperature fields to obtain the conjugate characteristics of the individual-global temperature fields. The calculation method is as follows:

[0144]

[0145]

[0146] ,

[0147] in, This is the thermal conduction coupling matrix. For the Softmax function, For enhanced conduction coupling characteristics, These are the thermal capacity parameters for lithium-sulfur batteries. It exhibits a conjugate characteristic of the single-unit and global temperature fields;

[0148] S4: Based on the deep heterogeneity characteristics and the conjugate characteristics of the individual cell-to-system temperature field, cross-dimensional fusion features are extracted through a dynamic weighting mechanism based on the intrinsic correlation of dual features; then, the abnormal heat values ​​of each individual cell and the overall thermal runaway risk level of the battery pack are calculated, including:

[0149] S41: Based on the deep heterogeneity characteristics and the conjugate characteristics of the individual-global temperature field, feature fusion is performed through a dynamic weighting mechanism based on the intrinsic correlation of dual features to obtain cross-dimensional fused features. The calculation method is as follows:

[0150]

[0151]

[0152]

[0153]

[0154]

[0155] in, The variance of the aging factor distribution of a single battery cell. For variance calculation, The temperature field conjugate characteristic stability coefficient is... For characteristic correlation strength factor, These are the weighting coefficients for the conjugate characteristics of the temperature field. For cross-dimensional fusion features;

[0156] S42: Based on cross-dimensional fusion characteristics and individual cell aging factors, calculate the abnormal heat value of each individual cell. The calculation method is as follows:

[0157]

[0158]

[0159]

[0160] in, For the cross-dimensional fusion features of the i-th individual cell, Tensor slicing operation for extracting features of the i-th single entity. The aging-weighted thermal characteristics of the i-th individual cell are shown below. Let be the abnormal heat value of the i-th individual cell. For batch normalization operations, The safe temperature field threshold for lithium-sulfur batteries, This is the critical temperature threshold for thermal runaway.

[0161] S43: Based on the abnormal thermal values ​​of individual cells and combined with the conjugate characteristics of the cell-to-system temperature field, calculate the overall thermal runaway risk level of the battery pack. The calculation method is as follows:

[0162]

[0163]

[0164]

[0165]

[0166] in, This represents the maximum abnormal heat value of all individual battery cells. To obtain the maximum value, This is the average of the abnormal heat values ​​of all individual cells. Calculated for the mean. For the overall thermal risk score, The overall thermal runaway risk level is as follows: This is the function for extracting the maximum index.

[0167] S5: Based on the abnormal heat values ​​of each individual battery cell and the overall thermal runaway risk level of the battery pack, the monitoring terminal displays the data visually and triggers early warning signals, including:

[0168] S51: Add the collection timestamp and battery pack number attributes to the abnormal heat values ​​of each individual battery cell and the overall thermal runaway risk level, and set the response threshold corresponding to each risk level to obtain basic thermal risk data with attribute labels.

[0169] S52: Perform format conversion on the thermal risk basic data with attribute identifiers, convert it into JSON data format compatible with the energy storage cabinet monitoring system, and obtain system-compatible standardized thermal risk data;

[0170] S53: Based on standardized thermal risk data, the local monitoring terminal of the energy storage cabinet displays the abnormal heat distribution of individual units and the overall risk level, and triggers the corresponding level of early warning signal according to the risk level.

[0171] Example 1

[0172] A neural network-based thermal runaway monitoring system for lithium-sulfur batteries includes:

[0173] Battery data acquisition module: Collects individual battery status data, group-level heat dissipation data, and battery compartment temperature data, and performs preprocessing to obtain preprocessed individual battery status data, preprocessed group-level heat dissipation data, and preprocessed compartment temperature data.

[0174] Single-cell feature extraction module: Based on the preprocessed single-cell state data, a dual-branch gating network is constructed for fusion processing to obtain dual-modal cell features; then, a multi-head attention enhancement mechanism is constructed to calculate the cell heterogeneity features; finally, deep heterogeneity features are extracted based on a multi-scale residual enhancement network.

[0175] Temperature field correlation module: Based on deep heterogeneity characteristics, preprocessed group-level heat dissipation data, and preprocessed cabin temperature data, temperature field correlation features are extracted through a temperature field correlation gating fusion mechanism; then, combined with deep heterogeneity characteristics, the conduction correlation between the individual and overall temperature fields is modeled to obtain the individual-overall temperature field conjugate features.

[0176] Risk level calculation module: Based on deep heterogeneity characteristics and single-cell-to-system temperature field conjugate characteristics, cross-dimensional fusion features are extracted through a dynamic weighting mechanism of intrinsic correlation between dual features; then, the abnormal heat value of each single cell and the overall thermal runaway risk level of the battery pack are calculated.

[0177] Early warning module: Based on the abnormal heat values ​​of each individual battery cell and the overall thermal runaway risk level of the battery pack, the module displays the data visually on the monitoring terminal and triggers an early warning signal.

[0178] Setting the safe temperature field threshold for lithium-sulfur batteries thermal runaway danger temperature threshold Risk levels 1-5 correspond to the following early warning rules: Level 1-2 triggers a yellow alert (attention), Level 3-4 triggers an orange alert (intervention), and Level 5 triggers a red alert (emergency response).

[0179] Scenario 1: Localized aging monomers overheat, but the overall average temperature is normal;

[0180] A certain energy storage cabinet contains 10 lithium-sulfur batteries. Due to the difference in the number of cycles (500 cycles for cells 1-4, 1200 cycles for cell 5, and 600 cycles for cells 6-10), the aging degree of cell 5 is significantly higher than that of the others.

[0181] Initial moment: Due to high aging factor, monomer No. 5 exhibits abnormal local heat generation, with its abnormal heat value... =72, the abnormal heat values ​​of other monomers are approximately 40-55, at this time, , ;

[0182] 30 minutes later: Unit 5 has poor heat dissipation. =85, other monomers remained at 42-56. , At this point, due to localized overheating, the temperature field stability decreases, and the temperature field conjugate characteristic stability coefficient... ;

[0183] Overall thermal risk score:

[0184] Overall thermal runaway risk level This triggered an orange alert.

[0185] At this point, the traditional method based on average temperature only calculates the average temperature, which is below the alarm threshold, and therefore determines that there is no risk.

[0186] Scenario 2: Due to environmental factors, the overall temperature rises synchronously, but no individual battery cells age.

[0187] The 10 lithium-sulfur batteries in a certain energy storage cabinet were all new batteries (cycle count <100 times, no aging). Due to the high temperature environment in summer, the overall heat dissipation of the energy storage cabinet was insufficient, and the thermal risk scores of all individual cells increased simultaneously. There were no individual cells with local extreme over-standard performance.

[0188] Initial time: Abnormal heat values ​​of all individuals =55 (Not aged, aging factor is low, score is within the safe range), at this point... , , (The temperature field is stable, with no local differences);

[0189] 30 minutes later: The ambient temperature continued to rise, and the abnormal thermal values ​​of all cells increased. =65, all exceeding (but without local abnormalities), at this time, , , (The temperature field remains stable, with no local extreme risks.)

[0190] Overall thermal risk score:

[0191] Overall thermal runaway risk level This triggers a yellow alert, indicating a need for overall heat dissipation optimization.

[0192] Traditional alarm methods based on fixed thresholds for individual units will trigger a full alarm if all units exceed the threshold, confusing environmental factors with individual unit failure risks and increasing the cost of misjudgment in operation and maintenance.

[0193] This invention integrates overall temperature field stability with dual-dimensional risk assessment to accurately identify non-fault attributes of overall temperature rise, avoid indiscriminate alarms, and achieve scientific and accurate risk warning.

[0194] It should be noted that the sequence numbers of the above embodiments of the present invention are merely for descriptive purposes and do not represent the superiority or inferiority of the embodiments. Furthermore, the terms "comprising," "including," or any other variations thereof are intended to cover non-exclusive inclusion, such that a process, apparatus, article, or method that comprises a list of elements includes not only those elements but also other elements not expressly listed, or elements inherent to such a process, apparatus, article, or method. Without further limitations, an element defined by the phrase "comprising one..." does not exclude the presence of other identical elements in the process, apparatus, article, or method that includes that element.

Claims

1. A method for monitoring thermal runaway in lithium-sulfur batteries based on neural networks, characterized in that, Includes the following steps: S1: Collect individual cell status data, group-level heat dissipation data, and battery compartment temperature data, perform preprocessing, and obtain preprocessed individual cell status data, preprocessed group-level heat dissipation data, and preprocessed compartment temperature data. S2: Based on the preprocessed single-cell state data, a dual-branch gating network is constructed to extract dual-modal cell features. The specific process is as follows: Based on the preprocessed individual cell state data, a dual-branch gating network is constructed. The preliminary dynamic data features of each individual cell are extracted through a one-dimensional convolutional layer, and the preliminary cycle count features of each individual cell are extracted through a multilayer perceptron. The preliminary dynamic data features of all individual cells are weighted through a dynamic data gate, and the preliminary cycle count features of all individual cells are weighted through a cycle count gate. The two types of weighted features are added element by element and then processed through a convolutional layer to obtain the dual-modal cell features. Then, a multi-head attention enhancement mechanism is constructed to calculate battery heterogeneity characteristics. The specific process is as follows: Based on the characteristics of the dual-mode battery and the number of cycles per cell, a multi-head attention enhancement mechanism is constructed, and the aging factor of each cell is calculated. Based on the aging factors of all cells and the characteristics of the dual-mode battery, the multi-head attention weight matrix is ​​calculated. Calculate the weight calibration coefficients and calibrate the multi-head attention weight matrix; The dual-modal battery features are weighted using the calibrated multi-head attention weight matrix to obtain the battery heterogeneity features. Finally, based on a multi-scale residual enhancement network, deep heterogeneity features are extracted. The specific process is as follows: The heterogeneous features of the battery are input into a multi-scale residual enhancement network. First, the dimensions of the heterogeneous features of the battery are reshaped. Then, three convolutional features of three scales are extracted by three convolutional layers with convolutional kernels of different sizes. The three scale convolutional features are added element-wise to obtain multi-scale fused convolutional features. The multi-scale fused convolutional features are subjected to residual processing to obtain residual features. The residual features are subjected to global average pooling to obtain deep heterogeneous features. S3: Based on the deep heterogeneity characteristics, preprocessed group-level heat dissipation data, and preprocessed cabin temperature data, temperature field correlation features are extracted through a temperature field correlation gating fusion mechanism; then, combined with the deep heterogeneity characteristics, the conduction correlation between the individual and overall temperature fields is modeled to obtain the individual-overall temperature field conjugate features. The temperature field correlation features are obtained by fusing deep heterogeneity features, group-level heat dissipation features extracted from preprocessed group-level heat dissipation data, and cabin temperature features extracted from preprocessed cabin temperature data through the temperature field correlation gating mechanism, weighted separately by temperature field correlation gates, and then fused by convolution processing. The specific process is as follows: S31: Based on the deep heterogeneity characteristics, preprocessed group-level heat dissipation data, and preprocessed cabin temperature data, a preliminary fusion of multi-source data is achieved through a temperature field correlation gating fusion mechanism to obtain temperature field correlation characteristics. The calculation method is as follows: ， in, As a group-level heat dissipation feature, It is a one-dimensional convolutional layer. This is the preprocessed group-level heat dissipation data. The cabin temperature characteristics, The preprocessed cabin temperature data, For gated loop unit, For temperature field related gates, For the Sigmoid function, It is a multilayer perceptron. For splicing operations, This is a characteristic of deep heterogeneity. The thermal conductivity coefficient of lithium-sulfur batteries, For Hadama accumulation, Temperature field correlation characteristics, It is a convolutional layer. To add element by element, These are the first, second, and third columns of data after tensor slicing of the temperature field correlation gate; S32: Based on temperature field correlation characteristics and deep heterogeneity characteristics, model the conduction correlation between the individual and global temperature fields to obtain the conjugate characteristics of the individual-global temperature fields. The calculation method is as follows: ， in, This is the thermal conduction coupling matrix. For the Softmax function, For enhanced conduction coupling characteristics, These are the thermal capacity parameters for lithium-sulfur batteries. It exhibits a conjugate characteristic of the single-unit and global temperature fields; S4: Based on the deep heterogeneity characteristics and the conjugate characteristics of the individual cell-to-the-system temperature field, cross-dimensional fusion features are extracted through a dynamic weighting mechanism based on the intrinsic correlation of dual features; then, the abnormal heat value of each individual cell and the overall thermal runaway risk level of the battery pack are calculated. S5: Based on the abnormal heat values ​​of each individual battery cell and the overall thermal runaway risk level of the battery pack, the monitoring terminal displays the data visually and triggers an early warning signal.

2. The method for monitoring thermal runaway of lithium-sulfur batteries based on neural networks as described in claim 1, characterized in that, The specific process of step S1 is as follows: The battery management system collects individual battery status data, including the number of cycles and dynamic data. The dynamic data in the individual battery status data is time-series numerical data, including real-time voltage, real-time charge and discharge current, and real-time internal resistance of the individual battery. The dynamic data in the individual battery status data is processed by removing outliers using the three-standard-deviation method, and then the data volume is unified by linear normalization to obtain preprocessed individual battery status data. Group-level heat dissipation data is collected through the heat dissipation system controller. The data type is time-series numerical data, including cooling fan speed and coolant circulation flow rate. The group-level heat dissipation data is denoised using a sliding window filtering method, and then missing values ​​are filled in by linear interpolation to obtain the preprocessed group-level heat dissipation data. Temperature data inside the battery compartment is collected by an in-cabin temperature sensor and is data in the form of time series numerical data. Gaussian filtering is then used to denoise the data, and outliers are removed by range clipping to obtain the preprocessed in-cabin temperature data.

3. The method for monitoring thermal runaway of lithium-sulfur batteries based on neural networks as described in claim 1, characterized in that, The specific process of step S2 is as follows: S21: Based on the preprocessed single-cell state data, a dual-branch gating network is constructed to extract dual-mode cell features. The calculation method is as follows: ， in, This represents the preliminary dynamic data characteristics of the i-th individual cell, where i is the index of the individual cell. This refers to the dynamic data of the i-th individual cell in the preprocessed individual cell state data. This represents a preliminary characteristic of the cycle number for the i-th individual cell. This represents the cycle number of the i-th cell in the preprocessed cell state data. For dynamic data gates, For continuous splicing operations, the splicing range is from i=1 to i=N, where N is the total number of individual cells. For loop count gate, This is a characteristic of dual-mode batteries; S22: Based on the characteristics of the dual-mode battery and the cycle number of individual cells, a multi-head attention enhancement mechanism is constructed. Weight calibration coefficients are calculated and features are weighted to obtain the battery heterogeneity characteristics. The calculation method is as follows: ， in, Let be the aging factor of the i-th individual cell. This represents the aging threshold for the number of cycles of a single battery cell. This refers to the maximum number of cycles for a single lithium-sulfur battery cell. This is a multi-head attention weight matrix. As a multi-head attention mechanism, For weight calibration coefficients, This is a characteristic of battery heterogeneity. This is a transpose operation; S23: Input the battery heterogeneity features into a multi-scale residual enhancement network to extract deep heterogeneity features. The calculation method is as follows: ， in, The heterogeneity characteristics of the battery after dimensional reshaping. For dimensional reshaping operations, These are convolutional features at the first, second, and third scales, respectively. These are convolutional layers with kernel sizes of 1×1, 3×1, and 5×1, respectively. To fuse convolutional features at multiple scales, For residual characteristics, This is global average pooling.

4. The method for monitoring thermal runaway of lithium-sulfur batteries based on neural networks as described in claim 3, characterized in that, The specific process of step S4 is as follows: S41: Based on the deep heterogeneity characteristics and the conjugate characteristics of the individual-global temperature field, feature fusion is performed through a dynamic weighting mechanism based on the intrinsic correlation of dual features to obtain cross-dimensional fused features. The calculation method is as follows: ， in, The variance of the aging factor distribution of a single battery cell. For variance calculation, The temperature field conjugate characteristic stability coefficient. For characteristic correlation strength factor, These are the weighting coefficients for the conjugate characteristics of the temperature field. For cross-dimensional fusion features; S42: Based on cross-dimensional fusion characteristics and individual cell aging factors, calculate the abnormal heat value of each individual cell. The calculation method is as follows: ， in, For the cross-dimensional fusion features of the i-th individual cell, Tensor slicing operation for extracting features of the i-th single entity. The aging-weighted thermal characteristics of the i-th individual cell are shown below. Let be the abnormal heat value of the i-th individual cell. For batch normalization operations, The safe temperature field threshold for lithium-sulfur batteries. This is the critical temperature threshold for thermal runaway. S43: Based on the abnormal thermal values ​​of individual cells and combined with the conjugate characteristics of the cell-to-system temperature field, calculate the overall thermal runaway risk level of the battery pack. The calculation method is as follows: ， in, This represents the maximum abnormal heat value of all individual battery cells. To obtain the maximum value, This is the average of the abnormal heat values ​​of all individual cells. Calculated for the mean. For the overall thermal risk score, The overall thermal runaway risk level is as follows: This is the function for extracting the maximum index. and These are weighting coefficients, determined empirically.

5. The method for monitoring thermal runaway of lithium-sulfur batteries based on neural networks as described in claim 1, characterized in that, The specific process of step S5 is as follows: S51: Add the collection timestamp and battery pack number attributes to the abnormal heat values ​​of each individual battery cell and the overall thermal runaway risk level, and set the response threshold corresponding to each risk level to obtain basic thermal risk data with attribute labels. S52: Perform format conversion on the thermal risk basic data with attribute identifiers, convert it into JSON data format compatible with the energy storage cabinet monitoring system, and obtain system-compatible standardized thermal risk data; S53: Based on standardized thermal risk data, the local monitoring terminal of the energy storage cabinet displays the abnormal heat distribution of individual units and the overall risk level, and triggers the corresponding level of early warning signal according to the risk level.

6. A system used in the neural network-based lithium-sulfur battery thermal runaway monitoring method as described in any one of claims 1-5, characterized in that, It includes a battery data acquisition module, a single cell feature extraction module, a temperature field correlation module, a risk level calculation module, and an early warning module; The battery data acquisition module collects individual battery status data, group-level heat dissipation data, and battery compartment temperature data, and performs preprocessing to obtain preprocessed individual battery status data, preprocessed group-level heat dissipation data, and preprocessed compartment temperature data. The single-cell feature extraction module constructs a dual-branch gating network for fusion processing based on the preprocessed single-cell state data to obtain dual-modal battery features. A multi-head attention enhancement mechanism is then constructed to calculate battery heterogeneity features; finally, deep heterogeneity features are extracted based on a multi-scale residual enhancement network. The temperature field correlation module extracts temperature field correlation features based on deep heterogeneity characteristics, preprocessed group-level heat dissipation data, and preprocessed cabin temperature data through a temperature field correlation gating fusion mechanism; then, combined with deep heterogeneity characteristics, it models the conduction correlation between the individual and overall temperature fields to obtain the individual-overall temperature field conjugate features. The risk level calculation module extracts cross-dimensional fusion features based on deep heterogeneity characteristics and single-entity-global temperature field conjugate characteristics through a dynamic weighting mechanism of intrinsic correlation between dual features; Then calculate the abnormal heat value of each individual battery cell and the overall thermal runaway risk level of the battery pack; The early warning module visualizes the abnormal heat values ​​of each individual battery cell and the overall thermal runaway risk level of the battery pack on the monitoring terminal and triggers an early warning signal.

Images