Energy storage battery pack temperature monitoring method and system

Abstract

The application relates to the technical field of battery temperature monitoring, and discloses a kind of energy storage battery pack temperature monitoring method and system, the method comprises obtaining internal temperature data, generating unified temperature field representation by global deduction;Based on temperature field, calculate local temperature gradient amplitude, locate heat distribution uneven area and obtain gradient value;If the gradient is over the threshold, simulate heat flow path, get spatial transfer characteristic description. Extract continuous time sequence samples from the transfer characteristic, quantify the cumulative effect of thermal anomaly and judge the dynamic change trend;Fusion dynamic trend and spatial characteristics generate multi-dimensional feature vector, get risk assessment grouping by density peak clustering. If the cumulative effect of a certain grouping is over the warning line, correct the feature vector with historical thermal runaway data, calculate the local high temperature evolution probability by hidden Markov model;Finally, the probability is mapped into a digital signal sequence, and the control command is generated by encoding, and the early warning notice is output. The method can effectively solve the problem of inaccurate temperature monitoring of energy storage battery pack.

Description

Technical Field

[0001] This invention relates to the field of battery temperature monitoring technology, and in particular to a method and system for monitoring the temperature of energy storage battery packs. Background Technology

[0002] Currently, in large-scale energy storage applications, temperature status is directly related to system operation safety, cycle life, and performance stability. As the core component of energy storage in new energy systems, energy storage battery packs composed of secondary batteries will significantly increase the risk of thermal runaway if the internal temperature distribution of the battery pack is uneven. Therefore, achieving accurate temperature field monitoring, enabling pre-diagnosis, and health management is the key to ensuring the reliable operation of energy storage systems.

[0003] In one existing technology, NTC thermistors are arranged on the surface of each battery module at a density of 1-2 cells per string. These thermistors are synchronously sampled by a 24-bit ADC chip in the BMS system, converting the resistance signals into temperature values, and then transmitted in real-time to the local controller via an I2C interface. The collected temperature data undergoes simple outlier filtering without deep processing such as temperature gradient calculation or spatiotemporal feature analysis; the real-time temperature values ​​are directly retained. A fixed temperature threshold is set as the judgment standard. When any sensor detects a temperature ≥55℃, an audible and visual warning is triggered; when the temperature ≥65℃, the BMS is activated to cut off the charging and discharging circuit and start the cooling system.

[0004] However, current technologies rely solely on real-time temperature values ​​for judgment, neglecting the dynamic trend of temperature changes over time. For slowly rising but continuously accumulating thermal risks, early warnings are impossible because the temperature hasn't reached a fixed threshold, potentially missing the optimal intervention window. Furthermore, they can only output single-point temperature data, failing to construct a three-dimensional temperature field for the battery pack, quantify the overall temperature distribution uniformity, and assess the distribution of thermal safety hazards from a global perspective. They only achieve "single-point alarms" rather than "system-wide warnings." In summary, current technologies suffer from insufficient accuracy in temperature monitoring. Summary of the Invention

[0005] This invention provides a method and system for monitoring the temperature of energy storage battery packs, in order to solve the problem of insufficient accuracy in temperature monitoring in the prior art.

[0006] In a first aspect, to solve the above-mentioned technical problems, the present invention provides a method for monitoring the temperature of an energy storage battery pack, comprising: Temperature data at various locations inside the battery pack is acquired, and spatial interpolation is performed on the temperature data to generate a temperature surface. The temperature surface is then mapped into a matrix to obtain a unified temperature field representation. The three-dimensional gradient feature tensor is extracted from the unified temperature field representation and the local temperature gradient amplitude is calculated to obtain the gradient of the region with uneven heat distribution. If the gradient of the uneven heat distribution area exceeds the preset thermal diffusion critical threshold, the thermal flow path between adjacent batteries is simulated and calculated, and the heat flux density vector is calculated based on the thermal flow path to obtain a description of the spatial transfer characteristics. Continuous time series samples are extracted from the spatial transfer characteristics description, and dynamic change trends are determined and cumulative effects are quantified based on the continuous time series samples. Based on the dynamic change trend and the spatial transmission characteristics, a multidimensional feature vector is generated, and the multidimensional feature vector is clustered and grouped to obtain risk assessment groups. If the cumulative effect in the risk assessment group is higher than the preset warning line, the corresponding multidimensional feature vector is corrected, and the probability of high temperature occurrence is calculated to obtain the local high temperature evolution probability. A digital signal sequence is generated based on the probability of local high temperature evolution. The digital signal sequence is then converted into an output instruction and executed to obtain the final early warning notification.

[0007] In a second aspect, the present invention provides a temperature monitoring system for an energy storage battery pack, comprising: The data acquisition module is used to acquire temperature data at various locations inside the battery pack, perform spatial interpolation on the temperature data to generate a temperature surface, and map the temperature surface into a matrix to obtain a unified temperature field representation. The amplitude calculation module is used to extract the three-dimensional gradient feature tensor from the unified temperature field representation and calculate the local temperature gradient amplitude to obtain the gradient of the uneven heat distribution region. The path simulation module is used to simulate and calculate the heat flow path between adjacent batteries if the gradient of the uneven heat distribution area exceeds a preset heat diffusion crisis threshold, and calculate the heat flux density vector based on the heat flow path to obtain a description of the spatial transfer characteristics. The data extraction module is used to extract continuous time series samples from the spatial transit characteristic description, determine the dynamic change trend and quantify the cumulative effect based on the continuous time series samples; The vector grouping module is used to generate multi-dimensional feature vectors based on the dynamic change trend and the spatial transmission characteristics, and to cluster the multi-dimensional feature vectors to obtain risk assessment groups. The vector correction module is used to correct the corresponding multidimensional feature vector if the cumulative effect in the risk assessment group is higher than the preset warning line, and to calculate the probability of high temperature occurrence to obtain the local high temperature evolution probability. The sequence generation module is used to generate a digital signal sequence based on the local high temperature evolution probability, convert the digital signal sequence into an output instruction and execute it to obtain the final early warning notification.

[0008] Compared with the prior art, the present invention has the following beneficial effects: (1) This invention obtains temperature data at various locations inside the battery pack and forms a unified temperature field representation through global simulation, which fully covers key areas such as the inside of the cell, module gaps, and cooling channels, thus completely eliminating monitoring blind spots; it quantifies the characteristics of uneven heat distribution, calculates the amplitude of local temperature gradients, and accurately locates areas with uneven heat distribution, rather than focusing only on the temperature value of a single point, thus achieving an upgrade from "single-point surface monitoring" to "global thermal state perception".

[0009] (2) This invention achieves a breakthrough through thermal flow path simulation. When the thermal unevenness gradient exceeds a threshold, it simulates the thermal flow path between adjacent batteries, clearly describing the direction, rate, and range of heat transfer in three-dimensional space. It establishes spatial risk correlation, no longer judging the temperature anomaly at a single point in isolation, but assessing the chain reaction of local overheating on surrounding cells based on thermal transfer characteristics, thus locking in the source of heat diffusion risk in advance. This solves the limitations of existing technologies that lack spatial correlation analysis between monitoring points and cannot identify the heat transfer path and diffusion trend inside the battery pack.

[0010] (3) This invention achieves a breakthrough through time series analysis, captures time dimension data from the spatial transmission characteristic description, accurately judges the dynamic change trend of temperature; quantifies the heat accumulation effect, transforms the change law of the time dimension into a calculable accumulation effect, avoids missing the slowly accumulating heat hazards due to failure to reach the static threshold, and solves the problem of early warning lag.

[0011] (4) Based on the spatial gradient calculation of convolutional neural network, this invention quickly screens out suspected areas of heat accumulation. The long short-term memory network performs continuous time series analysis on these areas to distinguish between temporary fluctuations and continuous deterioration trends. The hidden Markov model finally integrates spatial positioning and time series diagnostic information, and introduces historical thermal runaway knowledge for verification and probability inference, forming a complete closed loop from anomaly perception, trend judgment to risk quantification. This enables the system to predict the occurrence of thermal runaway and achieve accurate early warning of potential heat diffusion areas, avoiding the shortcomings of traditional methods that cause early warning lag and insufficient accuracy due to isolated judgment of single-point temperature. Attached Figure Description

[0012] Figure 1 This is a schematic flowchart of the energy storage battery pack temperature monitoring method provided in the first embodiment of the present invention; Figure 2 This is a schematic diagram of the structure of the energy storage battery pack temperature monitoring system provided in the second embodiment of the present invention. Detailed Implementation

[0013] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0014] Reference Figure 1 The first embodiment of the present invention provides temperature monitoring for energy storage battery packs, including the following steps: S11, acquire temperature data at various locations inside the battery pack, perform spatial interpolation on the temperature data to generate a temperature surface, map the temperature surface into a matrix to obtain a unified temperature field representation; S12, extract the three-dimensional gradient feature tensor from the unified temperature field representation and calculate the local temperature gradient amplitude to obtain the gradient of the uneven heat distribution region. S13, if the gradient of the uneven heat distribution area exceeds the preset thermal diffusion critical threshold, then the thermal flow path between adjacent batteries is simulated and calculated, and the heat flux density vector is calculated based on the thermal flow path to obtain a spatial transfer characteristic description. S14, extract continuous time series samples from the spatial transmission characteristic description, determine the dynamic change trend and quantify the cumulative effect based on the continuous time series samples; S15, Generate a multi-dimensional feature vector based on the dynamic change trend and the spatial transmission characteristics, and cluster the multi-dimensional feature vector to obtain risk assessment groups; S16, If the cumulative effect in the risk assessment group is higher than the preset warning line, then the corresponding multidimensional feature vector is corrected, and the probability of high temperature occurrence is calculated to obtain the local high temperature evolution probability. S17, Generate a digital signal sequence based on the local high temperature evolution probability, convert the digital signal sequence into an output instruction and execute it to obtain the final early warning notification.

[0015] In step S11, acquiring temperature data at various locations inside the battery pack, performing spatial interpolation on the temperature data to generate a temperature surface, and mapping the temperature surface into a matrix to obtain a unified temperature field representation includes: Calculate the local neighborhood numerical dispersion based on the temperature data, extract the temperature data of the region where the local neighborhood numerical dispersion is lower than a preset dispersion threshold, and generate an effective observation dataset. Based on the effective observation dataset, a covariance function is constructed and the fusion weights are calculated to obtain a set of feature temperature points; Spatial interpolation is performed on the set of characteristic temperature points to obtain a gridded temperature surface; The meshed temperature surface is mapped to a three-dimensional numerical matrix to generate a unified temperature field representation.

[0016] It should be noted that miniature NTC thermistor sensors (accuracy ±0.5℃, response time ≤100ms) are used. The sensor array is arranged according to the principle of "uniform three-dimensional coverage plus densification in key areas". Key areas include, but are not limited to, the positive and negative electrode tabs of the battery cell, module gaps, and cooling channels. A three-dimensional spatial neighborhood (radius 50mm) is defined with each sensor as the center. Each neighborhood contains 8-12 adjacent sensors. The local neighborhood numerical dispersion is equal to the standard deviation of the temperature data of all sensors within the neighborhood. The dispersion threshold is set at 1.5℃ (calibrated based on sensor accuracy and the normal temperature fluctuation range of the battery pack). If the absolute value of the difference between a sensor temperature value and the average temperature in the neighborhood is less than twice the dispersion, it is included in the valid observation dataset. If the dispersion is >1.5℃ (indicating an abnormal temperature change within the neighborhood), abnormal data that deviates from the average temperature in the neighborhood by more than 3℃ are removed, and the remaining data are recalculated for dispersion before further screening.

[0017] The covariance function used is the Gaussian covariance function (a common function in spatial interpolation), with covariance C(h) = C0 × exp(-3h² / a²), where h is the spatial straight-line distance between the two sensors (in mm); C0 is the sill value (reflecting the overall fluctuation of temperature data), usually taken as the variance of the effective observation dataset, and in this embodiment, it is taken as 1.2 times; a is the range (reflecting the range of spatial correlation of temperature), taken as 50 mm (based on the thermal conductivity characteristics of the energy storage battery pack for calibration, the spatial correlation of temperature data beyond this distance can be ignored); the fusion weight of a single effective observation data is equal to the sum of the covariances of that data with all other effective data divided by the sum of the covariances between all effective data, and the total weight is normalized to 1. Calculate the spatial distance h between any two sensors in the effective observation dataset; substitute it into the Gaussian covariance function to calculate the pairwise covariance; calculate the fusion weight of each data according to the weight formula, and select data with a weight ≥ 0.02 (total weight percentage ≥ 2%) as feature temperature points.

[0018] In this embodiment, the spatial interpolation is preferably implemented using the Kriging interpolation method. Kriging interpolation is a spatial statistical method based on variogram and structural analysis, which is particularly suitable for extrapolating physical quantities with spatial continuity and correlation, such as the temperature field of a battery pack. By fitting the spatial correlation structure of temperature data using a Gaussian variogram, Kriging interpolation can provide unbiased and optimal linear predictions, thereby generating continuous and accurate gridded temperature surfaces from discrete sensor data. Those skilled in the art will know that other spatial interpolation algorithms, such as inverse distance weighting and radial basis function methods, can also be used, but Kriging interpolation has advantages in accuracy and noise resistance in this application scenario.

[0019] A uniform three-dimensional mesh is used, with a basic mesh size of 5mm×5mm×5mm (adapting to the size of individual energy storage battery cells while considering local thermal detail capture and computational efficiency). The mesh range is limited by the actual physical boundary of the battery cell, and Kriging interpolation is used for global extrapolation, which is a standard algorithm in the field of spatial interpolation. The three-dimensional coordinates and temperature values ​​of the feature temperature point set are input into the Kriging interpolation model. Each mesh point is traversed in mesh order, and feature point data in the search neighborhood are called. The interpolation weight is calculated through the covariance function. The temperature value of each mesh point is obtained by weighted summation, and the temperature values ​​of all mesh points form a continuous meshed temperature surface. The mesh size should be determined comprehensively based on the actual size of the battery cell, the sensor layout density, and computing resources in practical applications, such as 10mm, 20mm, or larger. The core is to establish a discretized model that can reflect the spatial changes in temperature.

[0020] The Kriging interpolation model was trained based on 100 sets of measured temperature field data covering different battery types and module specifications. The training and validation sets were divided in a 7:3 ratio. Spatial correlation was fitted using a Gaussian variogram, and the nugget value, structural variance, and range were iteratively optimized using nonlinear least squares (100 iterations, convergence threshold 1e-6). An unbiased interpolation weight calculation model was constructed. After accuracy verification with MAE ≤ 0.3℃ and RMSE ≤ 0.4℃ on the validation set, it was packaged into a callable module for global temperature extrapolation. The 3D-CNN model training used 300... Zero sets (including data augmentation) of 3D gradient feature tensors and thermally uneven region annotation masks were used. The classic structure of "convolution-batch normalization-pooling" was adapted for 3D feature extraction. The Dice loss function and Adam optimizer (β1=0.9, β2=0.999) were selected. The initial learning rate was 0.001 and decayed by 10% every 10 rounds. The batch size was set to 8. The iteration was carried out until the Dice coefficient on the validation set was ≥0.85 and there was no improvement for 3 consecutive rounds (maximum 100 rounds). Finally, the training was completed and exported with an accuracy of Dice coefficient 0.88 and IOU 0.81.

[0021] The dimension of the three-dimensional numerical matrix is ​​the three-dimensional dimension of the number of grids. Each element in the matrix is ​​the interpolated temperature value of the corresponding grid point, and the element index corresponds one-to-one with the three-dimensional coordinates of the grid point. The unified temperature field is represented by a three-dimensional numerical matrix and a grid coordinate mapping table, where the coordinate mapping table records the correspondence between "matrix index - three-dimensional physical coordinates - grid size", thus obtaining the unified temperature field representation.

[0022] In step S12, the step of extracting the three-dimensional gradient feature tensor from the unified temperature field representation and calculating the local temperature gradient magnitude to obtain the gradient of the uneven heat distribution region includes: A three-dimensional numerical matrix is ​​extracted from the unified temperature field representation, and a three-dimensional gradient feature tensor is generated by performing a convolution operation on the three-dimensional numerical matrix. The three-dimensional gradient feature tensor is input into the trained convolutional neural network model to calculate the thermal features and obtain the thermal feature map. The thermal feature map is deconvolutionally reconstructed to obtain a set of pixel coordinates for heat distribution. The local temperature gradient magnitude is calculated by mapping the set of heat distribution pixel coordinates to the three-dimensional numerical matrix, and the gradient of the uneven heat distribution region is calculated and output based on the local temperature gradient magnitude.

[0023] It should be noted that the three-dimensional numerical matrix is ​​extracted from the unified temperature field representation generated by S11. The discrete differential operator adopts the three-dimensional Sobel operator (a general operator for thermal gradient extraction). The operator size is 3×3×3. A 3×3×3 neighborhood window is extracted with each voxel of the three-dimensional numerical matrix as the center. The element-wise multiplication and summation are performed with the Sobel operators of the X, Y, and Z axes respectively to obtain the gradient components (gx, gy, gz) of the voxel in the three directions. All voxels are traversed to form three independent gradient component matrices (each with a dimension of 200×100×60). These matrices are then concatenated into a three-dimensional gradient feature tensor (dimension 200×100×60×3). Each element of the tensor is a triplet of (gx, gy, gz).

[0024] The three-dimensional gradient feature tensor is input into a pre-trained convolutional neural network model to calculate thermal features. The model adopts a 3D-CNN model with 3 input channels (corresponding to the three gradient components X, Y, and Z) and 1 output channel (thermal feature response value). The three-dimensional gradient feature tensor (200×100×60×3) is normalized (gradient components are mapped to the [0,1] interval) and then input into the pre-trained 3D-CNN model. The model output is a thermal feature map (dimension 50×25×15). The response value (0-1) of each voxel in the map represents the confidence that the region is a region with uneven heat distribution. The closer the response value is to 1, the higher the confidence.

[0025] The 3D-CNN model was trained using a dataset of 3000 sets (including data augmentation) of 3D gradient feature tensors and thermal unevenness region annotation masks covering different battery types, module specifications, and thermal scenarios. This dataset was divided into training and validation sets in a 7:3 ratio. The model employed a classic "convolution-batch normalization-pooling" structure: input layer → 3 convolutional layers + batch normalization + 2 max pooling layers → output layer. The Dice loss function (Dice=1-[2×|P∩G|+ε] / [|P|+|G|+ε], where P is the predicted mask, G is the ground truth mask, and ε=1e-6) was used to mitigate the imbalance in the proportion of thermal unevenness regions. This was combined with the Adam optimizer (β1=0.9, β2=0.999, weight_de) (cay=1e-5), with an initial learning rate of 0.001 decaying by 10% every 10 rounds and a batch size of 8. The iteration continued until the Dice coefficient on the validation set was ≥0.85 and there was no improvement for 3 consecutive rounds (maximum 100 rounds). During this period, overfitting and small region omissions were optimized by adding Dropout layers and assigning loss weights to small targets. The final model achieved a Dice coefficient of 0.88, an IOU of 0.81, and a normal region recognition accuracy of 93.5% on the validation set. After being exported as a standardized module, it can accept a 200×100×60×3-dimensional gradient feature tensor and output a thermal feature map after inference of ≤0.5s / group. Voxels with a response value ≥0.7 are used as candidates for thermally uneven regions.

[0026] A 3D deconvolutional layer was used to reconstruct the thermal feature map (50×25×15). The parameters of the deconvolutional layer were symmetrically matched with the convolutional layer of the CNN model. After reconstruction, the output feature map dimension was restored to 200×100×60, which is consistent with the dimension of the original three-dimensional numerical matrix. The confidence threshold was set to 0.7 (calibrated based on 1000 sets of training data to balance recall and precision). Voxels with response values ​​≥0.7 in the reconstructed feature map were selected. The three-dimensional indices (i,j,k) of these voxels were extracted and converted into physical coordinates (X=5i,Y=5j,Z=5k) mm to form a set of pixel coordinates of heat distribution.

[0027] Traverse each voxel in the set of heat distribution pixel coordinates, and extract a 3×3×3 neighborhood window centered on that voxel (covering 27 adjacent grid points); calculate the mean of the gradient components in the three directions within the neighborhood. The local temperature gradient magnitude is equal to the sum of the squares of the mean gradient components in the X direction, the Y direction, and the Z direction, and then take the square root; count the local temperature gradient magnitudes of all voxels in the set of heat distribution pixel coordinates, and calculate their average value. This average value is the "gradient of the uneven heat distribution area".

[0028] In step S13, if the gradient of the uneven heat distribution region exceeds a preset thermal diffusion crisis threshold, the thermal flow path between adjacent batteries is simulated and calculated. Based on the thermal flow path, a heat flux density vector is calculated to obtain a spatial transfer characteristic description, including: Obtain the actual physical structure data of the battery pack; If the gradient of the uneven heat distribution region exceeds the preset thermal diffusion crisis threshold, a topology is constructed based on the actual physical structure data to obtain the geometric topology of the battery cell. A local three-dimensional finite difference mesh is constructed based on the geometric topology of the battery cell, and input into a pre-trained heat flux evolution calculation model to generate a transient thermal response matrix, and the heat flux vector field is derived. A heat flow path is generated based on the heat flux vector field, and a heat flux density vector is calculated based on the heat flow path to obtain a description of the spatial transfer characteristics.

[0029] It should be noted that the actual physical structure data of the battery pack includes geometric structure data and material thermal property data. The material thermal property data (derived from industry standard test data to ensure accuracy) includes the thermal conductivity of the cell body, the heat insulation layer, and the cooling channels (aluminum alloy). The actual physical structure data is obtained by standardizing it into a structured dataset of "component name-geometric parameters-thermal property parameters".

[0030] If the gradient of the uneven heat distribution area exceeds the preset thermal diffusion critical threshold (2.0℃ / cm, based on thermal diffusion test calibration of 500 sets of energy storage batteries), topology construction is initiated. Based on actual physical structure data, digital modeling is performed using SolidWorks. The topology nodes represent key locations such as the center, corners, tab center, and cooling channel wall of the battery cell, with a total of 1200 nodes. Node attributes include "3D coordinates (X,Y,Z), associated component, and thermal characteristic parameters". Edges connect adjacent nodes (spatial distance ≤10mm) and represent possible heat transfer paths. Edge attributes include "connection type (inter-cell / cell-cooling channel / cell-insulation layer), length, contact area, and thermal resistance". The output format is an STL file export of the topology, which simultaneously generates a "node coordinate table + edge attribute table", thus constructing the geometric topology of the battery cell.

[0031] The local three-dimensional finite difference mesh is constructed as a structured hexahedral mesh (a common mesh type for the finite difference method); the core heterogeneous region (high thermal gradient) has a mesh size of 1mm×1mm×1mm, and the surrounding transition region (low thermal gradient) has a mesh size of 3mm×3mm×3mm (balancing computational accuracy and efficiency); the heat flow evolution is calculated using a transient heat flow evolution finite difference model based on Fourier's law of heat conduction, and the transient heat conduction equation is: , where ρ is the material density (2600 kg / m³ for the battery cell and 30 kg / m³ for the insulation layer); c is the specific heat capacity of the material (1000 J / (kg·K) for the battery cell and 1400 J / (kg·K) for the insulation layer). The temperature change rate over time is given by ; k is the thermal conductivity of the material. q represents the temperature gradient; q represents the internal heat source intensity (5000W / m³ internal heat source intensity of the cell under 3C discharge conditions); the initial temperature of the grid nodes is the temperature value at the corresponding position in the S12 unified temperature field; the boundary conditions are set as follows: the cooling channel wall is a convective boundary (coolant temperature 25℃, convective heat transfer coefficient 50W / (m²·K)), the outer surface of the insulation layer is an adiabatic boundary (heat flux density is 0), and the contact surface between individual cells is a coupling boundary (considering contact thermal resistance). The convective heat transfer coefficient is referenced from the "Design Specification for Thermal Management System of Energy Storage Battery". The range of convective heat transfer coefficient for forced air cooling channels is 30-80W / (m²·K). This scheme takes the middle value of 50W / (m²·K) to adapt to the conventional air cooling design of energy storage modules. The internal heat source intensity is based on the measured value of 4000-6000W / m³ of the internal heat source intensity of ternary lithium batteries under 3C discharge conditions (refer to the "Battery Thermal Management Technical Manual"). This scheme takes 5000W / m³, which is the middle value of the measured value, to ensure the simulation is realistic.

[0032] The generated transient thermal response matrix has a dimension equal to the number of grid nodes multiplied by the number of time steps (time step size 1s). The matrix element T(i,j) represents the temperature value (°C) of the i-th grid node at the j-th time step, reflecting the dynamic evolution of the temperature of each node over time. Based on Fourier's law, the heat flux vector is equal to the negative of the material's thermal conductivity multiplied by the temperature gradient. For the transient thermal response matrix at each time step, the three-dimensional temperature gradient of each grid node is calculated using the central difference method. Combined with the thermal conductivity, the heat flux vector of each node is obtained, forming a global heat flux vector field.

[0033] The heat flow path originates at the core node (the grid node with the largest gradient) in the region of uneven heat distribution. Starting from the origin, each step selects the adjacent grid node with the largest heat flux vector magnitude as the next path point until the heat flux magnitude is ≤50W / m² (the heat flux threshold for normal areas). This path generation method aims to track the main direction and path of heat diffusion from the high-temperature core area outward, providing a representative spatial trajectory for assessing the transmission of thermal risks. This helps simplify the analysis and focus on the most critical heat diffusion channels. The path is represented by a "three-dimensional coordinate sequence," where the magnitude of the heat flux density vector is the magnitude of the heat flux vector at each node. The core of the spatial transmission characteristic description includes four types of information, forming a standardized transmission characteristic description, including basic path information (path origin / end coordinates, total length, number of nodes); directional characteristics (main direction of heat flow, contribution ratio of secondary directions); intensity characteristics (range of heat flux density vector magnitude on the path, average heat flux density); and associated components (types of components along the path (cell body → cell contact surface → cooling channel wall) and corresponding thermal resistance distribution).

[0034] In step S14, the step of extracting continuous time series samples from the spatial transfer characteristic description, determining the dynamic change trend based on the continuous time series samples, and quantifying the cumulative effect includes: Based on the spatial transfer characteristics, heat flux density vector values ​​are collected, and the time series of heat migration process is monitored to obtain continuous time series samples. The continuous time series samples are input into a pre-trained long short-term memory network model to capture temperature fluctuation patterns, and the output is the temperature fluctuation characteristics that characterize the evolution of the battery's thermal state. If the temperature fluctuation characteristics exhibit a divergent pattern, then the current dynamic change trend of the battery system is determined to be in the stage of thermal stability degradation. During the thermal stability failure stage, the core temperature of the core area of ​​the battery pack is continuously collected, and the difference between the core temperature and the preset safe temperature value is calculated. The difference is integrated to obtain the cumulative effect of thermal anomaly over time.

[0035] It should be noted that, along the heat flow path generated by S13, sampling points are selected according to the principle of "uniform distribution and dense core". The core sampling points are the midpoint of the path with the largest heat flux density vector magnitude (1), the starting point of the path (heat unevenness core area), and the ending point (heat diffusion boundary) (1 each). Ordinary sampling points are arranged every 5 mm on the path, and a total of 10-15 sampling points are selected for a single path to ensure complete coverage of the heat migration trajectory. Each sampling point collects a three-dimensional heat flux density vector. The dimension of the continuous time series sample is one time series corresponding to a single heat flow path. The sequence dimension is equal to the number of sampling points multiplied by 3 (three-dimensional vector components). Min-Max normalization processing is performed on each vector component (mapped to the [0,1] interval). Finally, a structured time series sample is output, which includes "path ID-time series matrix-sampling point coordinate lookup table".

[0036] The normalized time-series sample matrix is ​​input into the trained LSTM model in time step order; the model outputs an 8-dimensional temperature fluctuation feature vector, with each dimension corresponding to "heat flux intensity change rate, vector direction stability, local peak frequency, fluctuation amplitude, trend consistency, spatial diffusion coefficient, time decay coefficient, and synchronicity of associated sampling points", and the feature value range is [0,1]. The larger the value, the more significant the corresponding characteristic. The magnitude of the feature vector is calculated. A magnitude value ≥ 0.6 is judged as a valid feature (significantly reflecting the evolution law of thermal state), and a magnitude value < 0.6 means that the time-series data of this path is re-acquired.

[0037] The LSTM model is trained on 300 sets of time-series heat flux density data, labeled with three categories: "divergent / stable / convergent". It adopts a 3-layer LSTM + 1-layer fully connected layer structure (64 / 32 / 16 hidden units). The cross-entropy loss function (Loss=-Σ(y_i×log(p_i)), where y_i is the label (0 / 1 / 2 corresponds to stable / divergent / convergent), and p_i is the model prediction probability) is used with a dropout rate of 0.2 to suppress overfitting. Training is stopped when the accuracy of the validation set is ≥93%. It is used for temperature fluctuation feature extraction and dynamic trend judgment. The HMM model uses 1000 sets of historical thermal runaway data as training samples, initializes 3 hidden states (low / medium / high risk) and state transition matrices, and fits the observed probability distribution through 200 rounds of Baum-Welch algorithm (convergence threshold 1e-5) to finally achieve accurate calculation of the local high temperature evolution probability.

[0038] Based on the trend consistency and fluctuation amplitude dimensions of the temperature fluctuation feature vector, and combined with the dynamic changes in the feature modulus, three-dimensional judgment conditions are set. These conditions must simultaneously meet the following criteria within five consecutive time steps (5 seconds): trend consistency ≥ 0.7 (continuously unified direction of heat flow evolution, no reverse fluctuations); fluctuation amplitude ≥ 0.6 (drastic fluctuations in heat flow intensity, exceeding the normal stable range); feature modulus growth rate ≥ 10% (continuous increase in modulus, intensified thermal state evolution); and, combined with the spatial transfer characteristics of S13, the angle between the principal direction of the heat flux density vector and the vector direction of adjacent sampling points ≤ 15° (no sudden changes in divergence direction in space). If all the above conditions are met, the temperature fluctuation feature is judged to exhibit a "divergent pattern," and the current battery system enters the thermal stability destruction stage; otherwise, it is judged to be in a "stable pattern" or "convergent pattern," and subsequent cumulative effect calculations are not triggered. The starting timestamp of entering the destruction stage, the initial mean value of the initial heat flux density vector, and the initial temperature of the core sampling point are recorded to generate a "destruction stage initiation information table."

[0039] It should be noted that the judgment conditions and corresponding thresholds for the divergence pattern, such as trend consistency and fluctuation amplitude thresholds, are a set of optimized parameters derived from feature analysis, model training, and verification of a large amount of historical thermal runaway case data and normal operation data. The purpose is to extract the combination of quantitative indicators that best characterize the precursors of thermal runaway from the high-dimensional temperature fluctuation features output by the LSTM model. Those skilled in the art will understand that these thresholds can be calibrated and optimized according to different battery systems, such as ternary lithium, lithium iron phosphate, module design, or operating conditions. The core lies in using a data-driven approach to identify trends that deviate from normal steady state and evolve towards an unstable state by analyzing time-series characteristics.

[0040] The core area of ​​the thermal stability failure stage is the region consisting of the battery cell at the starting point of the S13 thermal flow path and two adjacent cells. The core temperature is the average of the temperatures at three key points within this area (cell center, near the tab, and adjacent to the cooling channel). The preset safe temperature value is based on the energy storage battery industry standard and is set at 45℃ (the lower limit of the thermal safety critical temperature for ternary lithium batteries, which is consistent with actual application scenarios). The difference between the core temperature and the preset safe temperature value is calculated at each time step (1 second). If the difference is ≤0 (the core temperature has not exceeded the safe value), the difference at that time step is recorded as 0 (no thermal anomaly accumulation). If the difference is >0, the actual difference is retained, in ℃. The cumulative effect is quantified by integral calculation. The cumulative effect is equal to the temperature difference at each time step multiplied by the sampling interval. In the discrete-time sampling system, the actual calculation is the accumulation of the temperature exceedance difference within each sampling period. From the start of the thermal stability failure stage, the integration continues until the warning is triggered or the core temperature drops below the safe temperature. The longest integration time does not exceed 60 minutes (to avoid meaningless calculations).

[0041] In step S15, the generation of multi-dimensional feature vectors based on the dynamic change trend and the spatial transmission characteristics, and the clustering of the multi-dimensional feature vectors to obtain risk assessment groups, includes: By concatenating and mapping the dynamic change trend with the heat flux density vector value in the spatial transfer characteristic description, a multidimensional feature vector is constructed. The similarity matrix is ​​calculated based on the multidimensional feature vector and the cluster center is calculated by inputting the trained density peak clustering model. The individual cells in the battery pack are merged according to the cluster center to obtain the thermal behavior region. Abnormal outliers within the thermal behavior region are removed, and the average deviation of the temperature of the remaining cells relative to a preset baseline safety value is calculated. The average deviations are then grouped to obtain risk assessment groups.

[0042] It should be noted that the heat flux density vector data corresponding to each battery cell in the S13 spatial transfer characteristic description are extracted. From the dynamic change trend output in S14, four core quantitative indicators are extracted to form a dynamic feature subset, including the temperature rise rate (taken as the average temperature rise rate in the thermal stability destruction stage in S14), fluctuation frequency (the reciprocal of the temperature fluctuation period, ≤0.1Hz in the stable region and ≥0.3Hz in the divergent region), cumulative effect (directly using the calculation results of S14), and divergence degree (obtained by normalizing the magnitude of the temperature fluctuation feature vector in S14, 0 represents stability, and 1 represents complete divergence). The three-dimensional components of the heat flux density vector of each cell are retained as a spatial feature subset, which is concatenated with the dynamic feature subset to form a multi-dimensional feature vector. The vector dimension is a 7-dimensional feature vector (4-dimensional dynamic features plus 3-dimensional spatial features); all feature dimensions are Min-Max normalized (mapped to the [0,1] interval).

[0043] The cosine similarity is used to calculate the similarity between any two individual feature vectors, and the density peak clustering model is employed. The core parameters (calibrated based on thermal behavior data from 500 sets of energy storage batteries) are as follows: the cutoff distance is set to 0.3 (a similarity threshold; cells with a similarity lower than this value are considered to have no direct association); the local density is set to the number of neighboring cells whose local density equals the number of cells with similarity greater than the cutoff distance; the cluster center determination criteria are the top 5% in local density and a similarity ≤ 0.2 with other cells of higher density (ensuring the independence of the cluster centers); the clustering process involves calculating the local density and similarity distance of 128 cells; selecting 3-5 cluster centers (taking 3 cluster centers for 128 cells as an example); and assigning each cell to the cluster center with the highest similarity to it, forming a thermal behavior region.

[0044] The density peak clustering model was trained using 500 sets of 7-dimensional feature vectors (4-dimensional dynamic features + 3-dimensional heat flux density vector features, normalized to the [0,1] interval by Min-Max) and thermal behavior labels covering different battery types, module specifications, and thermal behavior scenarios. These were divided into training and validation sets in a 7:3 ratio. A 128×128 similarity matrix was constructed using cosine similarity calculation. The initial cutoff distance dc = 0.3 was set (based on the training set similarity matrix, the distribution histogram of similarity for all samples was calculated, and the similarity value corresponding to the peak of the histogram was selected as the initial dc). Local density was defined as the number of neighboring samples with similarity > dc. The cluster center determination criteria were "the top 5% of local density and similarity ≤ 0.2 with higher density samples," with a silhouette coefficient ≥ 0.6. A RAND index ≥ 0.75 is considered acceptable. The DC value is iteratively adjusted in the 0.2-0.4 range with a step size of 0.05, ultimately determining the optimal parameter as DC = 0.3. During training, feature vectors are loaded in batches, and the local density and similarity distance of each sample are calculated. 3-5 cluster centers are selected and samples are assigned to form thermal behavior regions. The effect is evaluated on the validation set using silhouette coefficient, RAND index, and mean similarity within the region (≥ 0.7). Issues such as excessive outliers and abnormal cluster center numbers are addressed by adjusting DC or splitting / merging cluster centers. The final validation set achieves a silhouette coefficient of 0.68, a RAND index of 0.81, and thermal behavior consistency within the region ≥ 92%. After training, the core parameters are encapsulated as a callable module, which completes thermal behavior region division within 5 seconds after inputting the individual battery pack feature vector.

[0045] If the local density of a single cell is less than 10% of the local density of the corresponding cluster center, and the similarity with the cluster center is less than 0.15, it is identified as an outlier (probably due to data anomalies or accidental thermal disturbances). Outliers are screened according to the above criteria in each thermal behavior region, with the proportion of outliers in a single region being ≤3%. The remaining cells after removal constitute the effective region. A preset baseline safety temperature is set at 45℃ (the lower limit of the thermal safety critical temperature for ternary lithium batteries), based on the energy storage battery industry standard. The average deviation is equal to the arithmetic mean of the absolute values ​​of the temperature values ​​of all effective cells in the region minus the preset baseline safety temperature. Risk grouping thresholds (calibrated based on 500 sets of thermal runaway test data) are set as follows: high-risk group: average deviation ≥ 3.0℃ (significant thermal anomaly, requiring priority handling); medium-risk group: 1.0℃ ≤ average deviation < 3.0℃ (thermal anomaly exists, requiring attention); low-risk group: average deviation < 1.0℃ (normal thermal state, no intervention required). These are used to obtain risk assessment groups. The groups are sorted in descending order of average deviation, with high-risk groups having the highest priority, followed by medium-risk groups, and low-risk groups having the lowest priority, forming a clear order of action.

[0046] In step S16, if the cumulative effect in the risk assessment group is higher than a preset warning line, the corresponding multidimensional feature vector is corrected, and the probability of high temperature occurrence is calculated to obtain the local high temperature evolution probability, including: If the cumulative effect in the risk assessment group is higher than the preset warning line, historical thermal runaway evolution data matching the risk assessment group are retrieved. The current multidimensional feature vector is corrected using the historical thermal runaway evolution data to obtain a modified feature vector; The modified feature vector is input into the trained state transition probability model to calculate the probability of high-temperature evolution, thus obtaining the local high-temperature evolution probability.

[0047] It should be noted that the preset cumulative effect warning threshold is 20℃·min (calibrated based on 500 sets of measured data on thermal runaway of energy storage batteries, corresponding to the critical cumulative risk of thermal safety of ternary lithium batteries). If the average cumulative effect of a certain risk assessment group in S15 is ≥20℃·min, the historical data retrieval process will be triggered to construct a "thermal runaway evolution history database" that stores 1000 sets of thermal runaway case data under different operating conditions and different battery types (ternary lithium / lithium iron phosphate). Each set of data includes three parts: "risk group characteristics - evolution process data - final result". The core data dimensions (consistent with the current feature vector dimensions to ensure matching) include the average deviation of risk groups (corresponding to the regional average temperature deviation in S15); cumulative effect (corresponding to the thermal anomaly cumulative value in S14); average value of heat flux density vector magnitude (corresponding to the spatial transfer intensity characteristics in S13); average temperature rise rate (corresponding to the dynamic change trend characteristics in S14); and operating parameters (charge and discharge rate, ambient temperature, and SOC).

[0048] The similarity between the current risk group features and historical data can be calculated using methods such as the reciprocal of Euclidean distance or cosine similarity. Five core features of the current risk group (average deviation, cumulative effect, mean heat flux density modulus, mean temperature rise rate, and current operating parameters) are extracted. The historical database is traversed, and the similarity between the current features and each group of historical data is calculated. Historical data with a similarity ≥ 0.8 are selected (to ensure matching accuracy and avoid interference from irrelevant data) to form a matching dataset.

[0049] The current multidimensional feature vector is 7-dimensional, with the dimension definition consistent with S15. For each feature dimension, the statistical feature (mean plus standard deviation) of that dimension in the matching dataset is calculated, generating a dimension correction factor. The correction factor for a certain dimension can be calculated by subtracting the mean of that dimension in the matching dataset from the current dimension value, multiplying by the similarity weight, dividing by the standard deviation of that dimension in the matching dataset, and finally adding 1. The similarity weight is equal to the similarity between the current feature and the historical data divided by the sum of the similarities of all matching data. The correction constraint range is limited to [0.8, 1.2] to avoid over-correction leading to feature distortion. Following the above method, the correction factor is calculated and corrected for each dimension of the 7-dimensional feature vector to obtain the corrected feature vector. The values ​​of each dimension of the corrected feature vector must fall within the range of the corresponding dimension in the historical matching dataset. The lower limit of the range is the difference between the mean and the standard deviation, and the upper limit is the sum of the mean and the standard deviation, ensuring that the corrected features conform to the characteristic patterns of similar thermal runaway cases.

[0050] Hidden Markov Model (HMM) is used as the state transition probability model; the modified feature vector is standardized (mapped to the [0,1] interval, consistent with the model training data format); the trained HMM model is input, and the posterior probability of each hidden state is calculated through the forward algorithm; the local high temperature evolution probability is equal to the posterior probability of the system being in a high-risk state, or defined as the probability estimate of transitioning from the current state to the thermal runaway absorption state; the probability value range is [0,1], where 0 represents no high temperature risk and 1 represents that local high temperature will definitely occur. Usually, a probability ≥0.7 is judged as a high-risk warning trigger condition.

[0051] The Hidden Markov Model (HMM) was trained using a dataset of 1000 sets of 7-factor positive feature vectors and hot state labels (low / medium / high risk) covering different battery types, module specifications, and risk scenarios. The dataset was divided into training and validation sets in a 7:3 ratio. The feature vectors were Min-Max normalized to the [0,1] interval and a time series of length 30 was constructed. The hidden state transition matrix A was initialized as [[0.90,0.08,0.02], [0.15,0.75,0.10], [0.05,0.30,0.65]], and the initial state probability π was [0.85,0.13,0.02]. The observation probability matrix B was fitted using three Gaussian mixture components. The Gaussian mixture model parameters were initialized as follows: each hidden state corresponds to three Gaussian mixture components; the initial mean vector was taken as the mean of the corresponding state observation sequence in the training set; the covariance matrix was taken as the identity matrix; and the mixture weights were initialized according to a uniform distribution (ω=[1 / 3,1 / 3,1 / ...). 3) The core parameters were iteratively optimized using the Baum-Welch algorithm, with a maximum iteration of 100 rounds and a convergence threshold of 1e-5. L2 regularization (λ=1e-4) was added to suppress overfitting. The mean, covariance matrix, and mixing weights of the Gaussian mixture components were iteratively optimized using the EM algorithm, with 100 iterations and a convergence threshold of 1e-6 (the difference in likelihood function between two adjacent rounds ≤1e-6). In each round, the forward-backward algorithm was used to calculate the posterior probability of the state and the expected transition and update the parameters. The validation criteria were ≥88% state prediction accuracy, ≤5% probability error, and ≥90% high-risk recall. The model was optimized by adjusting the number of Gaussian components or the initial value of the transition matrix. The final validation set achieved a state prediction accuracy of 90.2%, a probability error of 3.8%, and a high-risk recall of 92.5%. After training, the parameters were encapsulated as a callable module. After inputting the corrected feature vector, the forward algorithm was used to calculate the posterior probability of each state, and the local high-temperature evolution probability was output according to the formula.

[0052] It is worth noting that the training processes of Convolutional Neural Networks (CNN), Long Short-Term Memory Networks (LSTM), Hidden Markov Models (HMM), and Density Peak Clustering (DNBC) mentioned in this invention all follow the standard procedures in the field of machine learning. These processes involve preparing and partitioning the dataset (training / validation set), initializing the model structure, selecting the loss function, applying optimization algorithms (such as Adam and Baum-Welch algorithms), and iterative training until convergence. The initial parameters of the model, such as the weights of the neural network, the initial transition matrix A of the HMM, and the emission probability B, can be randomly initialized or set empirically. The final effective parameters are all learned from the training data through the optimization algorithms. The specific network structures, number of layers, and hyperparameters, such as learning rate and number of iterations, given in the embodiments are examples of effective training and are not the only configurations.

[0053] In step S17, a digital signal sequence is generated based on the local high temperature evolution probability, the digital signal sequence is converted into an output instruction and executed to obtain the final early warning notification.

[0054] It should be noted that, based on the probability of local high-temperature evolution (range [0,1]), three discretized intervals are divided, corresponding to high / medium / low warning levels. The interval thresholds are calibrated based on 500 sets of measured data on thermal runaway warnings. The high-risk interval is 0.7 ≤ probability ≤ 1.0 (extremely high risk of thermal runaway, requiring emergency intervention); the medium-risk interval is 0.4 ≤ probability < 0.7 (exhibiting a trend of thermal anomalies, requiring close attention); and the low-risk interval is 0 ≤ probability < 0.4 (normal thermal state, no warning required). The probabilities of each interval are mapped to a fixed pulse frequency (a core feature of digital signals). The frequency range conforms to industrial control signal standards (1-10Hz), avoiding high-frequency interference and low-frequency delay: the high-risk pulse frequency is 5Hz (5 pulses per second, corresponding to emergency warning); the medium-risk pulse frequency is 2Hz (2 pulses per second, corresponding to regular warning); the low-risk pulse frequency is 0Hz (no pulse, no warning); the initial pulse frequency is equal to the difference between the high-risk frequency multiplied by the probability minus 0.7 and divided by 0.3 (high-risk range), and the difference between the medium-risk frequency multiplied by the probability minus 0.4 and divided by 0.3 (medium-risk range), with the result rounded to the nearest integer.

[0055] A square wave pulse signal is used, fixed at 10 pulse cycles (40 pulses for high-risk sequences and 20 pulses for medium-risk sequences). Structured data storage using "timestamp plus level status" is used to form the initial digital signal sequence. Manchester encoding (a standard encoding method in industrial control) is employed, with the core rule: a high-level to low-level transition represents binary "1", and a low-level to high-level transition represents binary "0". The initial pulse frequency is converted into a binary basic code ("1100" for high-risk, "1010" for medium-risk, and "0000" for low-risk), with a basic code length of 4 bits (to accommodate subsequent opcode matching). The basic code is then Manchester encoded and modulated, with each binary bit corresponding to 2 pulse cycles (200ms), forming a binary encoded stream. Based on this binary encoding... The bit width of the code stream (4 bits) matches the preset control instruction operation code. The operation code is bound to the warning level and the action to be executed. It is stored in the system control module in a structured manner. The operation code of the coded stream 1100 is 0x01, the warning level is high risk, and the corresponding action is to add audible and visual alarm, start directional cooling, and reduce the load power by 50%. The operation code of the coded stream 1010 is 0x02, the warning level is medium risk, and the corresponding action is to add audible and visual alarm and increase the real-time monitoring frequency to 2Hz. The operation code of the coded stream 0000 is 0x00, the warning level is low risk, and the corresponding action is no action.

[0056] The instruction data frame is encapsulated using the CAN bus standard data frame format and transmitted to the system control module via the CAN bus. After receiving the instruction data frame, the system control module synchronously outputs three types of early warning notifications: local audible and visual warning, activating the device's built-in audible and visual alarm (red indicator light flashes, buzzer frequency matches pulse frequency); local display warning, the device display shows "High-risk warning - area coordinates - immediate directional cooling recommended"; and remote platform warning, pushing the warning command to the energy storage system monitoring platform via 4G / Ethernet, containing structured information such as "warning level, probability of local high temperature evolution, risk area coordinates, and action to be performed", supporting remote intervention.

[0057] In summary, this invention discloses a method for monitoring the temperature of energy storage battery packs, including energy storage battery pack temperature monitoring. This invention achieves the acquisition of temperature data from various locations within the battery pack, and through global extrapolation, forms a unified temperature field representation, fully covering key areas such as the cell interior, module gaps, and cooling channels, completely eliminating monitoring blind spots. It quantifies thermal unevenness characteristics, calculates the amplitude of local temperature gradients, and accurately locates areas of uneven heat distribution, rather than focusing solely on single-point temperature values, achieving an upgrade from "surface single-point monitoring" to "global thermal state perception." A breakthrough is achieved through thermal flow path simulation. When the thermal unevenness gradient exceeds a threshold, it simulates the thermal flow path between adjacent cells, clearly describing the direction, rate, and range of heat transfer in three-dimensional space. It establishes spatial risk correlations, moving beyond isolated judgments of individual point temperature anomalies to assessing the cascading effects of local overheating on surrounding cells based on heat transfer characteristics, thus identifying the source of heat diffusion risk in advance. This overcomes the limitations of existing technologies that lack spatial correlation analysis between monitoring points and cannot identify the heat transfer path and diffusion trend within the battery pack. Breakthroughs have been achieved through time series analysis, capturing time-dimensional data from spatial transmission characteristics to accurately determine dynamic temperature change trends; quantifying the heat accumulation effect transforms the time-dimensional change patterns into calculable cumulative effects, avoiding the omission of slowly accumulating heat hazards due to failure to reach static thresholds, and solving the problem of delayed early warning.

[0058] Reference Figure 2 The second embodiment of the present invention provides a temperature monitoring system for an energy storage battery pack, comprising: The data acquisition module is used to acquire temperature data at various locations inside the battery pack, perform spatial interpolation on the temperature data to generate a temperature surface, and map the temperature surface into a matrix to obtain a unified temperature field representation. The amplitude calculation module is used to extract the three-dimensional gradient feature tensor from the unified temperature field representation and calculate the local temperature gradient amplitude to obtain the gradient of the uneven heat distribution region. The path simulation module is used to simulate and calculate the heat flow path between adjacent batteries if the gradient of the uneven heat distribution area exceeds a preset heat diffusion crisis threshold, and calculate the heat flux density vector based on the heat flow path to obtain a description of the spatial transfer characteristics. The data extraction module is used to extract continuous time series samples from the spatial transit characteristic description, determine the dynamic change trend and quantify the cumulative effect based on the continuous time series samples; The vector grouping module is used to generate multi-dimensional feature vectors based on the dynamic change trend and the spatial transmission characteristics, and to cluster the multi-dimensional feature vectors to obtain risk assessment groups. The vector correction module is used to correct the corresponding multidimensional feature vector if the cumulative effect in the risk assessment group is higher than the preset warning line, and to calculate the probability of high temperature occurrence to obtain the local high temperature evolution probability. The sequence generation module is used to generate a digital signal sequence based on the local high temperature evolution probability, convert the digital signal sequence into an output instruction and execute it to obtain the final early warning notification.

[0059] It should be noted that the energy storage battery pack temperature monitoring system provided in this embodiment of the invention is used to execute all the process steps of the energy storage battery pack temperature monitoring method in the above embodiment. The working principles and beneficial effects of the two are one-to-one, so they will not be described again.

[0060] It should be noted that the system embodiments described above are merely illustrative. The units described as separate components may or may not be physically separate, and the components shown as units may or may not be physical units; that is, they may be located in one place or distributed across multiple network units. Some or all of the modules can be selected to achieve the purpose of this embodiment according to actual needs. Furthermore, in the accompanying drawings of the system embodiments provided by this invention, the connection relationships between modules indicate that they have communication connections, which can be specifically implemented as one or more communication buses or signal lines. Those skilled in the art can understand and implement this without any creative effort.

[0061] The specific embodiments described above further illustrate the purpose, technical solution, and beneficial effects of the present invention. It should be understood that the above descriptions are merely specific embodiments of the present invention and are not intended to limit the scope of protection of the present invention. In particular, it should be noted that any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the scope of protection of the present invention for those skilled in the art.

Claims

1. A method for monitoring the temperature of an energy storage battery pack, characterized in that, include: Temperature data at various locations inside the battery pack is acquired, and spatial interpolation is performed on the temperature data to generate a temperature surface. The temperature surface is then mapped into a matrix to obtain a unified temperature field representation. The three-dimensional gradient feature tensor is extracted from the unified temperature field representation and the local temperature gradient amplitude is calculated to obtain the gradient of the region with uneven heat distribution. If the gradient of the uneven heat distribution area exceeds the preset thermal diffusion critical threshold, the thermal flow path between adjacent batteries is simulated and calculated, and the heat flux density vector is calculated based on the thermal flow path to obtain a description of the spatial transfer characteristics. Continuous time series samples are extracted from the spatial transfer characteristics description, and dynamic change trends are determined and cumulative effects are quantified based on the continuous time series samples. Based on the dynamic change trend and the spatial transmission characteristics, a multidimensional feature vector is generated, and the multidimensional feature vector is clustered and grouped to obtain risk assessment groups. If the cumulative effect in the risk assessment group is higher than the preset warning line, the corresponding multidimensional feature vector is corrected, and the probability of high temperature occurrence is calculated to obtain the local high temperature evolution probability. A digital signal sequence is generated based on the probability of local high temperature evolution. The digital signal sequence is then converted into an output instruction and executed to obtain the final early warning notification.

2. The method for monitoring the temperature of an energy storage battery pack according to claim 1, characterized in that, The process of acquiring temperature data at various locations within the battery pack, performing spatial interpolation on the temperature data to generate a temperature surface, and mapping the temperature surface into a matrix to obtain a unified temperature field representation includes: Calculate the local neighborhood numerical dispersion based on the temperature data, extract the temperature data of the region where the local neighborhood numerical dispersion is lower than a preset dispersion threshold, and generate an effective observation dataset. Based on the effective observation dataset, a covariance function is constructed and the fusion weights are calculated to obtain a set of feature temperature points; Spatial interpolation is performed on the set of characteristic temperature points to obtain a gridded temperature surface; The meshed temperature surface is mapped to a three-dimensional numerical matrix to generate a unified temperature field representation.

3. The method for monitoring the temperature of an energy storage battery pack according to claim 1, characterized in that, The step of extracting the three-dimensional gradient feature tensor from the unified temperature field representation and calculating the local temperature gradient magnitude to obtain the gradient of the uneven heat distribution region includes: A three-dimensional numerical matrix is ​​extracted from the unified temperature field representation, and a three-dimensional gradient feature tensor is generated by performing a convolution operation on the three-dimensional numerical matrix. The three-dimensional gradient feature tensor is input into the trained convolutional neural network model to calculate the thermal features and obtain the thermal feature map. The thermal feature map is deconvolutionally reconstructed to obtain a set of pixel coordinates for heat distribution. The local temperature gradient magnitude is calculated by mapping the set of heat distribution pixel coordinates to the three-dimensional numerical matrix, and the gradient of the uneven heat distribution region is calculated and output based on the local temperature gradient magnitude.

4. The method for monitoring the temperature of an energy storage battery pack according to claim 1, characterized in that, If the gradient of the uneven heat distribution region exceeds a preset thermal diffusion crisis threshold, the heat flow path between adjacent batteries is simulated and calculated. Based on the heat flow path, a heat flux density vector is calculated to obtain a spatial transfer characteristic description, including: Obtain the actual physical structure data of the battery pack; If the gradient of the uneven heat distribution region exceeds the preset thermal diffusion crisis threshold, a topology is constructed based on the actual physical structure data to obtain the geometric topology of the battery cell. A local three-dimensional finite difference mesh is constructed based on the geometric topology of the battery cell, and input into a pre-trained heat flux evolution calculation model to generate a transient thermal response matrix, and the heat flux vector field is derived. A heat flow path is generated based on the heat flux vector field, and a heat flux density vector is calculated based on the heat flow path to obtain a description of the spatial transfer characteristics.

5. The method for monitoring the temperature of an energy storage battery pack according to claim 1, characterized in that, The step of extracting continuous time series samples from the spatial transfer characteristic description, determining the dynamic change trend based on the continuous time series samples, and quantifying the cumulative effect includes: Based on the spatial transfer characteristics, heat flux density vector values ​​are collected, and the time series of heat migration process is monitored to obtain continuous time series samples. The continuous time series samples are input into a pre-trained long short-term memory network model to capture temperature fluctuation patterns, and the output is the temperature fluctuation characteristics that characterize the evolution of the battery's thermal state. If the temperature fluctuation characteristics exhibit a divergent pattern, then the current dynamic change trend of the battery system is determined to be in the stage of thermal stability degradation. During the thermal stability failure stage, the core temperature of the core area of ​​the battery pack is continuously collected, and the difference between the core temperature and the preset safe temperature value is calculated. The difference is integrated to obtain the cumulative effect of thermal anomaly over time.

6. The method for monitoring the temperature of an energy storage battery pack according to claim 1, characterized in that, The process involves generating multidimensional feature vectors based on the dynamic change trend and spatial transmission characteristics, and then clustering these multidimensional feature vectors to obtain risk assessment groups, including: By concatenating and mapping the dynamic change trend with the heat flux density vector value in the spatial transfer characteristic description, a multidimensional feature vector is constructed. The similarity matrix is ​​calculated based on the multidimensional feature vector and the cluster center is calculated by inputting the trained density peak clustering model. The individual cells in the battery pack are merged according to the cluster center to obtain the thermal behavior region. Abnormal outliers within the thermal behavior region are removed, and the average deviation of the temperature of the remaining cells relative to a preset baseline safety value is calculated. The average deviations are then grouped to obtain risk assessment groups.

7. The method for monitoring the temperature of an energy storage battery pack according to claim 1, characterized in that, If the cumulative effect in the risk assessment group is higher than a preset warning line, the corresponding multidimensional feature vector is corrected, and the probability of high temperature occurrence is calculated to obtain the probability of local high temperature evolution, including: If the cumulative effect in the risk assessment group is higher than the preset warning line, historical thermal runaway evolution data matching the risk assessment group are retrieved. The current multidimensional feature vector is corrected using the historical thermal runaway evolution data to obtain a modified feature vector; The modified feature vector is input into the trained state transition probability model to calculate the probability of high-temperature evolution, thus obtaining the local high-temperature evolution probability.

8. A temperature monitoring system for an energy storage battery pack, characterized in that, include: The data acquisition module is used to acquire temperature data at various locations inside the battery pack, perform spatial interpolation on the temperature data to generate a temperature surface, and map the temperature surface into a matrix to obtain a unified temperature field representation. The amplitude calculation module is used to extract the three-dimensional gradient feature tensor from the unified temperature field representation and calculate the local temperature gradient amplitude to obtain the gradient of the uneven heat distribution region. The path simulation module is used to simulate and calculate the heat flow path between adjacent batteries if the gradient of the uneven heat distribution area exceeds a preset heat diffusion crisis threshold, and calculate the heat flux density vector based on the heat flow path to obtain a description of the spatial transfer characteristics. The data extraction module is used to extract continuous time series samples from the spatial transit characteristic description, determine the dynamic change trend and quantify the cumulative effect based on the continuous time series samples; The vector grouping module is used to generate multi-dimensional feature vectors based on the dynamic change trend and the spatial transmission characteristics, and to cluster the multi-dimensional feature vectors to obtain risk assessment groups. The vector correction module is used to correct the corresponding multidimensional feature vector if the cumulative effect in the risk assessment group is higher than the preset warning line, and to calculate the probability of high temperature occurrence to obtain the local high temperature evolution probability. The sequence generation module is used to generate a digital signal sequence based on the local high temperature evolution probability, convert the digital signal sequence into an output instruction and execute it to obtain the final early warning notification.

Images