A battery warehouse management method and system

Abstract

The application relates to the technical field of warehouse management, and discloses a battery warehouse management method and system, the method comprising the following steps: constructing a space topology and a dynamic adjacency matrix based on a real-time position of a battery, and mapping temperature, voltage and gas concentration to a node characteristic sequence; performing message transmission by using a graph neural network, and integrating a directional diffusion mechanism of gas concentration along an edge in a propagation process to update a node embedding vector; constructing a space-time characteristic tensor along a time dimension, inputting a preset time sequence prediction model, and outputting an abnormal probability of each node; if the probability exceeds a preset threshold value, tracing a high-influence source and a potential path along a reverse propagation path, and locating a thermal runaway risk area. The method significantly improves risk identification accuracy and early warning timeliness through dynamic topology modeling and multi-feature time sequence fusion, and provides effective technical support for battery warehouse safety.

Description

Technical Field

[0001] This invention relates to the field of warehouse management technology, and in particular to a battery warehouse management method and system. Background Technology

[0002] Currently, battery storage management is a key area for ensuring the safe operation of large-scale energy storage systems, and risk warning for battery storage is of great practical significance.

[0003] In one existing technology, temperature and voltage sensors, powered by sensor network chips, are installed at key locations on battery modules or shelves to collect and upload data. These sensors periodically collect static values ​​at single moments and set uniform alarm thresholds in a data center, such as triggering an alarm when the temperature exceeds 60 degrees Celsius or the voltage falls below the nominal value by 20%. However, this existing technology suffers from limitations due to its single monitoring dimension and statically delayed judgment logic. It cannot detect the multi-physical coupling effects such as heat conduction and gas diffusion caused by the proximity of batteries. Because alarms are triggered only based on isolated threshold exceedances, the system often only responds when the fault is already clearly apparent. It lacks the ability to warn of early anomalies such as localized temperature rise trends and gradual changes in gas concentration, and it cannot pinpoint the source and direction of risk diffusion, easily leading to missed or false alarms. This makes it difficult to achieve pre-emptive risk intervention and precise handling, resulting in limited overall warning timeliness and accuracy.

[0004] In summary, existing technologies suffer from the problem of being unable to dynamically perceive the early gradual risks and spatial propagation paths under the coupling of multiple physical fields in the battery, making it difficult to accurately locate the source of the risk. Summary of the Invention

[0005] This invention provides a battery storage management method and system to achieve safe management of battery storage.

[0006] In a first aspect, to solve the above-mentioned technical problems, the present invention provides a battery storage management method, comprising: Acquire battery temperature, battery voltage, gas concentration data, raw signal strength and unique identifier; calculate the real-time position coordinates of the battery based on the raw signal strength; construct a spatial topology based on the real-time position coordinates of the battery and the unique identifier to obtain the warehouse space topology. Based on the storage space topology, the set of adjacent batteries for a single battery is determined, and a dynamic adjacency matrix is ​​obtained. The battery temperature, battery voltage, and gas concentration data are mapped to the corresponding nodes of the dynamic adjacency matrix to obtain a node feature sequence; A preset graph neural network is used to perform message passing between the node feature sequence and the dynamic adjacency matrix to obtain the embedding vector of each node; The gas concentration data is incorporated into the embedding vectors of each node and propagated directionally along the adjacent edges of the storage space topology to obtain updated node embedding vectors. Based on the updated node embedding vector, a stacked matrix is ​​constructed along the time dimension to obtain the spatiotemporal feature tensor; The spatiotemporal feature tensor is processed using a preset temporal prediction model to obtain the probability of abnormal states; When any node in the abnormal state probability exceeds the preset abnormal state probability threshold, the risk source node and potential propagation path are traced along the reverse propagation path to determine the thermal runaway risk area.

[0007] In a second aspect, the present invention provides a battery storage management system, comprising: The data acquisition module is used to acquire battery temperature, battery voltage, gas concentration data, raw signal strength and unique identifier, calculate the real-time position coordinates of the battery based on the raw signal strength, and construct a spatial topology structure based on the real-time position coordinates of the battery and the unique identifier to obtain the warehouse space topology structure. The matrix construction module is used to determine the set of adjacent batteries of a single battery based on the topology of the storage space, and obtain a dynamic adjacency matrix. The data mapping module is used to map the battery temperature, battery voltage and gas concentration data to the corresponding nodes of the dynamic adjacency matrix to obtain the node feature sequence; The data association module is used to perform message passing between the node feature sequence and the dynamic adjacency matrix using a preset graph neural network to obtain the embedding vector of each node; The vector update module is used to incorporate the gas concentration data into the embedding vectors of each node and propagate it directionally along the adjacent edges of the storage space topology to obtain the updated node embedding vectors. The spatiotemporal integration module is used to construct a stacked matrix along the time dimension based on the updated node embedding vector to obtain a spatiotemporal feature tensor; The feature analysis module is used to process the spatiotemporal feature tensor using a preset time series prediction model to obtain the probability of abnormal states; The risk location module is used to trace the risk source node and potential propagation path along the reverse propagation path and determine the thermal runaway risk area when any node in the abnormal state probability exceeds the preset abnormal state probability threshold.

[0008] Thirdly, the present invention also provides an electronic device including a processor, a memory, and a computer program stored in the memory and configured to be executed by the processor, wherein the processor executes the computer program to implement the pool storage management method described in any one of the above.

[0009] Fourthly, the present invention also provides a computer-readable storage medium comprising a stored computer program, wherein the computer program, when running, controls the device where the computer-readable storage medium is located to execute the battery storage management method described in any one of the above.

[0010] Compared with the prior art, the present invention has the following beneficial effects: (1) This invention solves the problem of difficulty in accurately locating risk sources due to the inability to dynamically perceive early gradual risks and spatial propagation paths caused by the inability to dynamically perceive the coupling of multiple physical fields in the battery.

[0011] (2) By constructing a dynamic adjacency matrix to map the battery adjacency relationship and integrating multi-dimensional features such as temperature, voltage, and gas concentration, this invention realizes closed-loop reasoning from spatial correlation to temporal evolution and traces the risk source and risk propagation path, thereby improving the identification accuracy and early warning timeliness of thermal runaway risk areas.

[0012] (3) This invention overcomes the lag and isolation of traditional fixed threshold monitoring by using dynamic topology modeling and multiphysics coupling analysis, and realizes spatial simulation and temporal prediction of heat conduction and gas diffusion effects between batteries. It can capture early gradual abnormal trends and accurately trace the risk source and propagation path when the abnormal probability exceeds the limit, thereby improving the timeliness of early warning, the accuracy of positioning and the ability to judge risks, and providing decision support for battery storage safety management. Attached Figure Description

[0013] Figure 1 This is a schematic diagram of a battery storage management method provided in the first embodiment of the present invention; Figure 2 This is a schematic diagram of a battery storage management system provided in the second embodiment of the present invention. Detailed Implementation

[0014] 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.

[0015] Reference Figure 1 The first embodiment of the present invention provides a battery storage management method, including the following steps: S11, acquire battery temperature, battery voltage, gas concentration data, original signal strength and unique identifier, calculate the real-time position coordinates of the battery based on the original signal strength, construct a spatial topology based on the real-time position coordinates of the battery and the unique identifier, and obtain the warehouse space topology. S12, determine the set of adjacent batteries for a single battery based on the topology of the storage space, and obtain a dynamic adjacency matrix; S13, map the battery temperature, battery voltage and gas concentration data to the corresponding nodes of the dynamic adjacency matrix to obtain the node feature sequence; S14, A preset graph neural network is used to perform message passing between the node feature sequence and the dynamic adjacency matrix to obtain the embedding vector of each node; S15, the gas concentration data is incorporated into the embedding vectors of each node, and directional propagation is performed along the adjacent edges of the storage space topology to obtain the updated node embedding vectors; S16, construct a stacked matrix along the time dimension based on the updated node embedding vector to obtain the spatiotemporal feature tensor; S17, The spatiotemporal feature tensor is processed using a preset time series prediction model to obtain the probability of abnormal states; S18, when any node in the abnormal state probability exceeds the preset abnormal state probability threshold, the risk source node and potential propagation path are traced along the reverse propagation path to determine the thermal runaway risk area.

[0016] In step S11, battery temperature, battery voltage, gas concentration data, original signal strength, and a unique identifier are acquired. The real-time battery location coordinates are calculated based on the original signal strength. A spatial topology is constructed based on the real-time battery location coordinates and the unique identifier to obtain the warehouse space topology, including: The battery temperature, battery voltage, and gas concentration data are acquired by sensors. The original signal strength and unique identifier are collected by a receiving and reading device. The real-time position coordinates of the battery are obtained by performing trilateration calculations based on the original signal strength. The real-time position coordinates of the battery are offset from the preset reference position to obtain corrected position data. An identifier correspondence is generated based on the corrected position data and the unique identifier. Topology nodes are generated based on the identification correspondence and the corrected location data, and the warehouse space topology structure is constructed based on the topology nodes.

[0017] It should be noted that the battery node identified as "BAT001" has real-time status data collected by sensors: temperature 45.2 degrees Celsius, voltage 3.75 volts, and ambient gas concentration 0.012 ppm, with real-time coordinates of (2.1, 3.5, 1.0) meters. These data are linked by a unique identifier, collectively forming a complete spatiotemporal snapshot of the battery node's state. The location coordinates determine the node's adjacency relationship in the warehouse spatial topology, while temperature, voltage, and gas concentration reflect its own operating status and local environment. When the temperature of a node abnormally rises to 50.5 degrees Celsius, the gas concentration of the adjacent node "BAT002" within 0.5 meters of its surrounding coordinates may increase to 0.018 ppm, while the voltage may drop to 3.70V. This spatial coupling change of multiple parameters is a key characteristic of the transmission of thermal runaway risk.

[0018] First, in the warehouse battery storage positioning system, the reading and writing devices continuously collect the raw signal strength (RSSI) values ​​emitted by the tags. For example, at a certain moment, a battery tag might be simultaneously captured by three fixed readers, yielding RSSI values ​​of -65dBm, -72dBm, and -58dBm, respectively. These raw signal strengths are then converted into distance estimates; generally, stronger signals indicate a closer distance. Specifically, a mapping table between RSSI and distance can be pre-established. A log-distance path loss model can be used to map the conversion relationship between RSSI values ​​and distance. For instance, the tag signal strength can be measured multiple times at 0.5-meter intervals, and the average RSSI value recorded, thus constructing a static mapping table between RSSI values ​​and actual distances. When a specific RSSI value is collected in real time, the closest RSSI entry is directly looked up in the table, and its corresponding distance value is used as the estimation result. If the RSSI falls between two values ​​in the table, linear interpolation is used to calculate the corresponding distance. The average RSSI is approximately -45 dBm for a distance of 1 meter, -52 dBm for 1.5 meters, -58 dBm for 2 meters, -64 dBm for 2.5 meters, and -70 dBm for 3 meters. These three values ​​are then converted to estimated distances of approximately 2.6 meters, 3.4 meters, and 2 meters, respectively.

[0019] Next, the trilateration method is used. Using the known coordinates of the three readers as centers and the estimated distance as radii, the intersection point is found to obtain the real-time coordinates of the battery tag. Centered on the ball, with the corresponding estimated distance Given a radius, the equation for a sphere can be written. Taking three readers as an example, the solution typically involves first calculating the difference between pairwise equations to obtain a linear relationship, then constructing a system of linear equations and using the least squares method to solve for the tag coordinates. Due to the error in distance estimation, the three spheres usually do not intersect precisely at one point. Therefore, the least squares solution can obtain the optimal approximate intersection point, which is the estimated position of the tag. For example, the final calculated position is (4.3, 5.8, 1.4) meters.

[0020] Subsequently, due to multipath reflections and occlusions in the actual environment, the coordinates may shift. For example, the preset reference position is an ideal coordinate (4.0, 6.0, 1.0) meters obtained through initial calibration. After comparing it with the real-time coordinates, the offset is found to be 0.3 meters horizontally, -0.2 meters vertically, and 0.4 meters vertically. Specifically, the offset vector (0.3, -0.2, 0.4) meters from the reference position to the real-time coordinates is calculated, and then the direction angle of this vector is obtained, including the horizontal azimuth angle θ, the elevation angle φ, and the offset distance L.

[0021] By analyzing historical data, a statistical relationship was established between signal distortion caused by environmental interference and the offset direction angle. For example, it was found that when the signal multipath mainly comes from a certain direction, the positioning error vector will stably deviate to a related angle. Calculate the implicit error direction angle in the current real-time positioning result. With environmental feature angle The difference is obtained by fitting historical error data based on a preset value. For example, the scaling factor k is 0.1 (angle-distance), and the angle difference is converted into a distance correction value. Finally, the scaling factor is corrected in the opposite direction of the offset vector. The original offset vector is scaled to obtain a correction vector. Subtracting this vector from the real-time coordinates yields the corrected accurate position. Finally, the battery identifier and the corrected coordinates are aggregated to obtain the identifier correspondence. Finally, the adjacency relationship is determined based on the Euclidean distance between nodes. Considering the influence range of thermal radiation and gas diffusion during thermal runaway, the adjacency determination threshold is expanded to 4 meters, which can effectively cover the adjacent battery group that may be affected by the single cell anomaly. Therefore, the threshold is set to 4 meters. If the distance between A and B is only 0.3 meters, which is less than the set strong adjacency threshold of 1 meter, then they are considered strong adjacency. If the distance between A and C is too far, then they are not adjacent. Thus, an adjacency matrix is ​​constructed. In the matrix, the corresponding position is set to 1 to indicate connectivity and 0 to indicate non-connectivity.

[0022] Specifically, once the adjacency matrix confirms that the entire graph is connected—meaning that any node can reach other nodes through adjacency—the mapping between the battery's unique identifier and the current corrected coordinates is formally established. For example, battery ID "BAT001" maps to (2.0, 3.0), and ID "BAT002" maps to (2.1, 3.2). The current warehouse topology can only be constructed when connectivity requirements are met, avoiding distortion of the overall structure due to abnormal signals from individual tags. For instance, in a real warehouse, if excessively high stacks in a certain area obstruct signals, causing some nodes to temporarily isolate, and an adjacency matrix disconnection is detected, the current topology construction is rejected, and the attempt is retried only after collecting more complete data.

[0023] In step S12, determining the set of adjacent batteries for a single battery based on the topology of the storage space to obtain a dynamic adjacency matrix includes: Obtain the coordinate data of all battery nodes in the warehouse space topology, calculate the Euclidean spatial distance based on the battery node coordinate data and construct a matrix to obtain the initial distance matrix; The spatial distance values ​​are obtained from the initial distance matrix, and the spatial distance values ​​are marked with blocking status in combination with the preset warehouse obstacle distribution map to obtain the marked distance data; Based on the marked distance data, filter connected node pairs that meet the dynamic neighborhood conditions, and generate a distance-oriented weighted list based on the connected node pairs; Based on the distance-oriented weighted list, construct the set of adjacent batteries for each battery to obtain a dynamic adjacency matrix.

[0024] Firstly, in the field of warehouse battery management, obtaining the coordinate data of all battery nodes in the warehouse spatial topology is fundamental to constructing spatial relationships. Consider a warehouse with 10 battery nodes. The coordinate data of each node is recorded; for example, node 1 is located at (2.0, 3.0, 1.0) meters, and node 2 is located at (2.5, 3.2, 1.0) meters. Using these coordinates, the Euclidean spatial distance between the nodes is calculated, forming an initial distance matrix. The calculated straight-line distance between node 1 and node 2 is approximately 0.5 meters. This matrix provides the initial data foundation for subsequent analysis.

[0025] Subsequently, marking the blocking status based on the spatial distance values ​​in the initial distance matrix and the preset warehouse obstacle distribution map becomes particularly important. In the warehouse, shelves or walls may block direct paths between nodes. The straight-line distance between node 1 and node 3 is 3.0 meters, but there is a tall shelf in between. This distance will be marked as a blocking state, indicating that the two nodes cannot be directly connected. In the initial distance matrix, the distance value of the corresponding node pair, such as node 1 and node 3, is replaced with a special blocking identifier, a negative number "-1". In an associated adjacency state matrix, the element value corresponding to this node pair is set to "0" to indicate no direct connection, distinguishing it from the actual distance weight of the normal connected state. In this way, the distance matrix records geometric distance, while the state matrix explicitly expresses connectivity. The preset warehouse obstacle distribution map is pre-set manually, and through on-site surveying, the outlines of fixed shelves, walls, and pillars are digitized into polygonal regions and stored as a set of coordinate boundaries. The intersection of node lines with these polygonal regions is calculated: if a line intersects with any obstacle polygon, it is determined to be "blocked". By manually labeling and connecting the key boundary points of each obstacle in a counter-clockwise order, a closed polygon is formed and represented using a sequence of vertex coordinates. For example, a rectangular shelf can be represented as the coordinates of its four corner points [(x1,y1),(x2,y2),(x3,y3),(x4,y4)]. For complex shapes, they can be decomposed into combinations of multiple simple polygons. Finally, the set of polygons representing all obstacles, along with their identifiers, is stored as structured data, forming a digital obstacle distribution map that can be queried by a computational program.

[0026] Secondly, based on the labeled distance data, connected node pairs that meet the dynamic neighborhood condition radius are selected and adjacency relationships are constructed. The neighborhood radius is set to 2.0 meters. Preliminary experiments show that during battery thermal runaway, the temperature or gas concentration within 2.0 meters of the battery significantly affects adjacent batteries, so 2.0 meters can be used as the neighborhood determination radius. Node 1 and Node 2 are 0.5 meters apart, meeting the condition, and are therefore confirmed as a connected node pair. Node 1 and Node 4 are 2.5 meters apart, exceeding the radius range, and are therefore not considered. Simultaneously, a distance-orientation weighted list containing distance and orientation weights is generated for the connected node pairs. For example, the connection weight of Node 1 and Node 2 is higher due to their close distance, while the orientation weight is further adjusted according to their relative direction. Specifically, the distance weight is usually inversely proportional to the distance and can be calculated using an exponential decay function, such as the weight... This results in a weight of 0.78. A directional weight can also be introduced; for example, if the line connecting two nodes aligns with the main ventilation direction of the warehouse, a higher weight can be assigned. For example, in 1.2, if it's perpendicular, a base weight of 1.0 is assigned; if it's at an angle, linear interpolation is used to obtain the weight. The final comprehensive weight... This results in a final overall weight of 0.94, which is then used to create a distance- and orientation-weighted list. This method provides a more detailed description of the spatial relationships between nodes.

[0027] Finally, based on the distance-orientation weighted list, the neighboring battery set for each battery is constructed, resulting in a dynamic adjacency matrix that reflects the spatial relationships at the current moment. The neighboring set of node 1 includes nodes 2 and 5. Traversing these sets, the connectivity relationships are expressed in matrix form. A valid connection threshold > 0.1 is set for the weighted list; this threshold is only used to determine matrix connectivity. When constructing the dynamic adjacency matrix, for any pair of nodes (a, b), the comprehensive weight value in its distance-orientation weighted list is queried. .like If a node exists and its value exceeds a preset threshold, then the node in the adjacency matrix is ​​marked as 1 in row a, column b and row b, column a, indicating that the two nodes are effectively connected in spatial relation at the current time; if If a node does not exist or is less than or equal to the threshold, it is marked as 0, indicating that it is not directly connected. This dynamic adjacency matrix can intuitively display the adjacency status between battery nodes, providing a reliable basis for subsequent inventory management and path optimization.

[0028] In step S13, mapping the battery temperature, battery voltage, and gas concentration data to the corresponding nodes of the dynamic adjacency matrix to obtain the node feature sequence includes: The battery temperature, voltage, and gas concentration data collected by the sensors are integrated to obtain the physical quantity values; The physical quantity values ​​are mapped to the warehouse space topology to generate a multi-dimensional original feature vector. The multidimensional original feature vectors are weighted and fused according to the dynamic adjacency matrix to obtain a weighted feature vector; The weighted feature vectors are stacked using a time-series sliding window to generate a sequence tensor, and the sequence tensor is filled into the corresponding nodes of the dynamic adjacency matrix to obtain the node feature sequence.

[0029] Firstly, in the field of warehouse battery management, sensors are deployed near each battery node to collect analog signals in real time and convert them into digital messages. These digital messages contain physical quantities such as battery temperature, voltage, and gas concentration. Specifically, the sensors use analog-to-digital converters to convert continuous analog waveforms into discrete digital formats, ensuring the accuracy and reliability of data transmission. The temperature analog signal collected by the sensors at a single battery node is a sinusoidal waveform, which, after conversion, generates a digital message, such as a temperature value encoded as 45.2, a voltage of 3.75, and a gas concentration of 0.012 ppm. Temperature, voltage, and gas concentration reflect its own operating status and local environment. When the temperature of a node abnormally rises to 50.5 degrees Celsius, the gas concentration of the adjacent node "BAT002" within 0.5 meters of its surrounding coordinates may increase to 0.018 ppm, while the voltage may drop to 3.70V. These values ​​directly reflect the real-time state of the battery, providing the raw data foundation for subsequent analysis.

[0030] Subsequently, after parsing the digital message, the obtained physical quantity values ​​need to be mapped to the storage space topology to generate a multi-dimensional original feature vector. Temperature (45.2 degrees Celsius), voltage (3.75 volts), and gas concentration (0.012 ppm) are extracted. Then, based on the node identifier "BAT001," matching is performed to associate the values ​​with nodes, forming a three-dimensional vector [45.2, 3.75, 0.012]. The message contains the identifier of the source battery and its measured physical quantity values. The corresponding node is found in the topology using this identifier, and the physical quantity values ​​are assigned to that node, generating its multi-dimensional original feature vector, for example, [45.2, 3.75, 0.012]. The coordinates (2.0, 3.0, 1.0 meters) represent the node's spatial attributes. This mapping method ensures a close connection between physical quantities and spatial location, facilitating the capture of local environmental differences within the batteries in the warehouse.

[0031] Secondly, based on the previously constructed dynamic adjacency matrix, the multidimensional original feature vectors are weighted and fused to obtain a weighted feature vector. Specifically, for node 1, its neighboring nodes include nodes 2 and 5, and the corresponding positions in the matrix are marked as 1. The comprehensive weight value is then retrieved from the distance-orientation weighted list based on the adjacency relationship. For example, if the distance weight between node 1 and node 2 is 0.8 and between node 1 and node 5 is 0.6, then the weighted average of the feature vectors of adjacent nodes needs to be normalized. The distance weight between node 1 and node 2 (0.8 / (0.8+0.6)) is approximately 0.57, and the distance weight between node 1 and node 5 (0.6 / (0.8+0.6)) is approximately 0.43. The original vector of node 2 is [44.8,3.80,0.010], and that of node 5 is [46.1,3.70,0.015]. After fusion, the weighted feature vector of node 1 may become [45.4,3.76,0.012]. This reflects the comprehensive state under the influence of the neighborhood and improves the robustness and context relevance of the features.

[0032] Finally, the weighted feature vectors are time-series sliding windowed and stacked to generate a sequence tensor. This sequence tensor is then associated with the corresponding nodes in the dynamic adjacency matrix, resulting in a node feature sequence. For example, with a sliding window length of 5 time steps, the weighted feature vectors from the most recent 5 time steps are collected, such as those from time t-4 to t, and stacked into a 5x3 sequence tensor. This tensor is then filled into node 1 of the adjacency matrix, forming enhanced graph structure data. This serialization process captures the temporal dynamics of battery state, which helps predict potential anomalies, such as a gradual temperature increase trend.

[0033] In step S14, the use of a preset graph neural network to perform message passing between the node feature sequence and the dynamic adjacency matrix to obtain the embedding vector of each node includes: Based on the dynamic adjacency matrix, conduction paths are established between nodes in the storage space topology, and energy gradients are determined. A message carrier is constructed based on the energy gradient, and the node feature sequence is mapped to a multidimensional space to obtain the initial message. A preset graph neural network is used to assign neighborhood weights to the initial message. When the diffusion rate of the energy gradient exceeds a preset diffusion rate threshold, the transmission path is updated to obtain the updated transmission path. Based on the updated propagation path, multiple rounds of message passing are performed to compress the state tensor into an embedded representation, resulting in the embedding vector of each node.

[0034] First, the transmission path is constructed based on the positional relationships and connection strengths of the battery nodes within the warehouse space. There are 10 battery nodes in the warehouse. A dynamic adjacency matrix is ​​generated based on spatial distance and connection weights. Paths may preferentially select connections between nodes that are closer and have higher weights. For example, if the weight from node A to node B is 0.9, but only 0.3 to node C, the path tends to pass through B. Regarding the step of determining the energy gradient through the dynamic adjacency matrix, the energy gradient can be seen as a quantitative indicator of the state differences between battery nodes. After Z-value normalization of the data for node (a,b), the energy gradient... ,in and These are the temperature or voltage characteristic values ​​of nodes b and a in a certain state, respectively. This is the corresponding connection weight in the adjacency matrix. This calculation result... That is, the quantized value of the energy gradient from node a to node b. A positive value indicates a tendency for state values ​​to flow from node B to node A; a negative value indicates the opposite direction. The absolute value reflects both the strength of the state difference and the weight of the spatial connection, thus providing a directed and weighted basis for message passing. For example, if node A has a temperature of 40.5 degrees Celsius, node B has a temperature of 42.8 degrees Celsius, and the weight is 0.9, then the temperature gradient is 2.07, quantifying the tendency and strength of heat flowing from B to A. This gradient provides a basis for subsequent message passing, ensuring that the information flow is consistent with changes in physical state.

[0035] Subsequently, the process of constructing a message carrier and mapping the node feature sequence to a high-dimensional mapping space aims to extract deeper evolutionary features. Specifically, the multi-dimensional feature sequence of node A over consecutive time steps is taken as an input sample and subjected to a nonlinear transformation by an encoder. This encoder learns to compress and map dynamic patterns in the time series, such as temperature trends and voltage fluctuations, into a fixed-length high-dimensional vector, i.e., the "initial message." For example, the original feature sequence might be a matrix of 5 time steps × 3 features (temperature, voltage, and gas concentration), which, after being mapped by the encoder, becomes a 128-dimensional vector, forming the initial message. The initial message contains the temperature change trend of node A over the past 5 time steps, enabling a better understanding of its state evolution. This mapping can be seen as an abstract expression of the node's state, facilitating the capture of hidden dynamic change patterns.

[0036] The nonlinear transformation can be achieved by flattening the 5×3 temporal feature matrix of node A or directly inputting it into a fully connected layer for dimensionality increase. The convolutional kernel size is set to 3, and the stride is set to 1, enabling time-step sliding scanning to ensure temporal continuity. "Same" padding is used to maintain the temporal dimension of the output sequence. The number of convolutional kernels is typically set to 32 or 64 in the first layer to extract basic local features, and can be multiplied layer by layer in subsequent layers to enhance representational capabilities. The pooling layer uses max pooling with a size of 2 and a stride of 2, halving the temporal dimension successively, achieving dimensionality reduction while preserving significant features. These parameters are typical empirical configurations for processing short temporal sequences, aiming to balance the granularity of feature extraction with model computational efficiency.

[0037] Secondly, the step of assigning neighborhood weights to the initial message using an aggregation function focuses on balancing the influence of neighborhoods. Node A has nodes B and C in its neighborhood. The message influence is assigned based on distance weights of 0.7 and 0.5 respectively. Let the initial message of node A be... The initial messages of its neighboring nodes B and C are respectively , The predefined distance weights in the dynamic adjacency matrix are: , The first step is to perform a linear transformation, applying a shared weight matrix W to the messages of each node to obtain the transformed features. , , The second step is to calculate the attention score. and ( The feature references of the neighboring nodes B and C of the central node A are concatenated and then multiplied by a learnable attention vector a. The original score is then obtained by applying the LeakyReLU activation function. and The third step introduces distance weights as priors, combining the original scores with the distance weights to calculate the adjusted scores. = , The distance weights between critical nodes A and B can be traced back in S12. The fourth step is normalization: the Softmax function is applied to the adjusted scores of node A and all its neighbors to obtain the normalized attention weights. , , The fifth step is weighted aggregation, which ultimately yields the new message for node A. , where σ is a nonlinear activation function . When the diffusion rate exceeds a preset diffusion rate threshold, which is determined by analyzing the statistical distribution of characteristic diffusion rates between nodes under historical normal operating conditions, the 99th percentile of this distribution is taken as the initial threshold setting, such as 0.2 units per second. This triggers a propagation path update, indicating that the information transmission efficiency of the current propagation path is insufficient to match the dynamic changes in the actual state. The propagation path update mechanism is then activated. The message diffusion rate on each active adjacent edge is continuously monitored. It is obtained by calculating the norm of the change in the eigenvector difference between nodes per unit time. For nodes i and j at two consecutive time points... and The feature vectors are respectively , and , First, calculate the difference vector of eigenvectors between the two time points. , Next, calculate the change in the difference vector. .at last, Defined as this change norm divided by time interval ,Right now And with a preset threshold Compare.

[0038] when When this happens, a weight update calculation is triggered. Based on the latest node state, it is quantized into three core inputs: the real-time temperature gradient, etc. Secondly, the real-time voltage gradient Third, the diffusion rate exceeds the standard. Weight updates are performed via the Adma optimizer to ensure the timeliness of information delivery.

[0039] Finally, the propagation path is updated and multiple rounds of message passing are performed to compress the state tensor into an embedded representation. Through three rounds of message passing, the state tensor of node A is compressed into an embedding vector, which contains the combined state of itself and its neighborhood, such as temperature influence and voltage fluctuation trends. Regarding the acquisition of each node's embedding vector, it can be understood that the embedding vector is not only a compressed representation of the state but also incorporates the influence of its neighbors. In each round of message passing, each node merges the current feature vectors of its neighboring nodes through an aggregation function, such as a weighted summation with attention, and updates its own features. After three iterations, each node's feature vector has undergone information integration from the local to a broader neighborhood, ultimately forming the node feature vector. This vector not only encodes the node's own original state sequence, such as temperature and voltage timing, but also encodes the state influence and topological relationships of its multi-hop neighborhood nodes using LeakyReLU. In node A's embedding vector, the temperature-related component is slightly enhanced by the high temperature influence of node B. This fusion allows the system to more accurately grasp the overall operating status of the batteries in the warehouse, providing a reliable basis for management decisions.

[0040] In step S15, the process of incorporating the gas concentration data into the embedding vectors of each node and propagating it directionally along the adjacent edges of the storage space topology to obtain updated node embedding vectors includes: Obtain the gas concentration values ​​and topological connections of each node, calculate the gas concentration difference between adjacent nodes, and obtain the gas concentration gradient vector. The concentration gradient vector is projected onto the adjacent edge direction of the topological connection, and a propagation weight characterizing the gas flow along the adjacent edge is generated to obtain the directional propagation weight. The features of adjacent nodes are weighted according to the directed propagation weights to obtain the diffusion message tensor; The spatial decay operation is performed on the diffusion message tensor based on the physical length of the adjacent edges to obtain the signal sequence to be aggregated; The signal sequence to be aggregated is aggregated and fused with the current node state to obtain the updated node embedding vector.

[0041] Firstly, in the field of warehouse battery management, the analysis of gas concentration monitoring and diffusion impact can explore how to optimize the system's perception of the battery operating environment from multiple perspectives. To obtain gas concentration values ​​and topological connections for each node, it is first necessary to collect gas concentration data around each battery node within the warehouse and construct a topological structure based on the spatial connections between nodes. In a warehouse with 8 battery nodes, the gas concentration at node A is 2.5 ppm, and at node B it is 3.8 ppm. A topological map is generated based on the physical distance and connection strength between nodes, laying the foundation for subsequent analysis. For example, in the step of calculating the concentration difference between adjacent nodes to determine the concentration gradient vector, the concentration differences between adjacent nodes are analyzed to form a vector characterizing the gas flow trend. For instance, if node A and node B are adjacent and the concentration difference is 1.3 ppm, it can be determined that the gas may flow from B to A, forming a directional vector. This represents a vector with a magnitude of 1.3 ppm, directed from A to B. This vector reflects the potential trend of gas diffusion, providing a basis for subsequent processing.

[0042] Subsequently, regarding the process of projecting the concentration gradient vector onto the adjacent edge direction defined by the topological connection relationship, it is necessary to determine whether the gas flow is consistent with the direction of the connection edge. If the inner product from node A to B is positive, such as 0.7, it indicates that the gas is more likely to flow along this edge, and a corresponding directional propagation weight is generated, which is the projection component value. This weight intuitively reflects the directionality of gas diffusion, which helps to accurately simulate the flow path.

[0043] Secondly, in the step of weighting the source node features using directed propagation weights to construct the diffusion message tensor, the influence of the node features needs to be adjusted according to the weights. Node B has a high concentration, so its weight is 0.8. A message tensor containing concentration and flow intensity is constructed to characterize the gas diffusion characteristics from B outwards. Its shape can be represented as [number of target nodes, maximum number of adjacent edges, message feature dimension]. The first dimension indexes each target node, the second dimension enumerates all directed adjacent edges pointing to the target node, and the third dimension stores the weighted message feature vector transmitted on each edge. This feature vector is obtained by multiplying the state features of the source node, such as concentration, temperature, and voltage, with the corresponding directed propagation weights, and can selectively concatenate the weight values ​​and the direction encoding of the edges. For example, for target node A, its message vector from node B is [B concentration 0.8]. ,Temperature B 0.8, Voltage B 0.8, Direction code BA], the direction code BA can be obtained through the gas flow direction 1 , get direction The message vector from node C is similar. This message tensor provides a carrier for subsequent information transmission.

[0044] Then, a spatial attenuation operation is performed on the diffusion message tensor based on the physical length of the adjacent edges, taking into account the effect of distance on gas diffusion. The physical distance between nodes A and B is 5 meters. The message tensor is attenuated and adjusted according to the distance to obtain a signal sequence reflecting transmission loss. Specifically, for each adjacent edge, let its physical length be... And calculate an attenuation coefficient based on the diffusion coefficient. ,in The attenuation rate parameter is determined by fitting the observed attenuation coefficient with distance, based on the actual law of gas concentration attenuation with distance as calibrated in historical data. For example, The value is 0.15 per meter. Then, the diffusion message vector corresponding to that edge is... That is, each element of the feature slice at the corresponding position in the tensor is multiplied by the attenuation coefficient to obtain the attenuated message vector. This operation is performed independently on each edge, ultimately updating all message vectors in the entire diffuse message tensor. This approach ensures that the influence of distant nodes is reasonably mitigated.

[0045] Finally, regarding the aggregation operation on the signal sequence and its fusion with the current node state, the received signal is combined with the node's own state to output an updated embedding vector. For example, after node A receives an attenuated signal from B, the gas concentration influence component in its embedding vector increases, reflecting the diffusion effect. For instance, node A's current embedding vector is [temperature 45.2, voltage 3.75, gas concentration: 0.012], and the attenuated diffusion message from node B, transmitted to A, results in a signal vector of [temperature influence 0.3, voltage influence 0.0, gas concentration influence 0.008]. After element-wise addition, node A's updated embedding vector becomes [45.5, 3.75, 0.020]. This updated vector provides more comprehensive data support for warehouse environment monitoring, helping to promptly identify potential risks and optimize management strategies.

[0046] In step S16, the step of constructing a stacked matrix along the time dimension based on the updated node embedding vector to obtain the spatiotemporal feature tensor includes: The updated node embedding vectors are arranged and aligned in chronological order to obtain the standardized evolution trajectory; The standardized evolution trajectories are stacked along the time axis to obtain a multidimensional stacked matrix; The stacked matrix is ​​scanned using a sliding convolution kernel group to extract local temporal patterns, thus obtaining the hidden layer states; The association weights are calculated based on the hidden state, and the hidden state is weighted and fused based on the association weights to obtain the spatiotemporal feature tensor.

[0047] First, the updated node embedding vectors generated in each sampling period are obtained. These vectors incorporate the latest state after the influence of gas diffusion. By arranging and aligning these vectors in chronological order, a standardized evolution trajectory can be formed, reflecting the trend of gas concentration changes in the warehouse over time. The sampling period is once every 10 minutes. In the first period, the gas component of the embedding vector of node A is 2.8, which rises to 3.2 in the second period and 3.5 in the third period. These values ​​are arranged sequentially, forming a clear upward trajectory, which facilitates the subsequent capture of the gas accumulation process.

[0048] Subsequently, the evolution trajectories are stacked along the time axis to construct a multi-dimensional stacked matrix. This operation integrates the node states from multiple cycles into a unified structure. For example, for a warehouse with eight battery nodes, the time-series vectors of each node are stacked into a matrix, where rows represent different time points and columns correspond to the node feature dimensions. This matrix intuitively preserves the spatiotemporal correlation and helps to comprehensively analyze the dynamic propagation of gas within the warehouse.

[0049] Secondly, a sliding scan of the multidimensional stacked matrix is ​​performed using a group of convolutional kernels to extract hidden state states that reflect local temporal patterns. This process is similar to detecting concentration change patterns within a time window. The convolutional kernel size is set to three time steps. When scanning the trajectory of node B, if the concentration rises from 3.8 to 4.1 and then to 4.4 in the previous three cycles, the local pattern of accelerated increase is extracted as the hidden state. This state accurately captures short-term gas leakage trends, thereby improving the sensitivity to anomalies.

[0050] Finally, association weights are calculated based on the hidden state, and these weights are used to perform weighted fusion of the hidden states to generate a spatiotemporal feature tensor that captures the temporal accumulation process. For example, for each node, its nearest neighbors are extracted. The sequence consists of hidden state vectors at each time step. Calculate the time-series characteristic indicators of the sequence, including the trend component. The slope was obtained by performing linear regression on the time index of the sequence. , The covariance between the time index vector t and the hidden state sequence H represents the variance of the sequence. and deviation components The Euclidean distance between the sequence mean and the global hidden state mean.

[0051] Next, these three scalar indicators are concatenated into a feature vector. The input is a two-layer fully connected neural network. The first layer undergoes a linear transformation and is activated by ReLU. The second layer undergoes a linear transformation and is normalized to the (0,1) interval by the Sigmoid function to obtain the final weights.

[0052] Specifically, the weight matrix With input feature vector After multiplying the transpose of the matrix and adding the bias vector The weight matrix The dimension is h rows and 3 columns. It maps the three-dimensional input features to an h-dimensional hidden space, with a bias vector. It is an h-dimensional vector used to perform a global translation of the linear combination result, and is the input feature vector. The transpose of the matrix transforms it from a row vector to a column vector to conform to the matrix multiplication rules. The second linear transformation converts the weight matrix... After matrix multiplication with the intermediate eigenvector u, a scalar bias is added. The weight matrix The dimension is 1 row and h columns, responsible for compressing the intermediate features of the h dimensions into a scalar value, with scalar bias. The benchmark used to adjust this scalar is the intermediate feature vector u, which is the output of the first-layer linear transformation after being processed by the ReLU nonlinear activation function. The output is then normalized by the Sigmoid function to generate the final association weights between 0 and 1.

[0053] If the hidden state near node A shows a continuous upward trend, its correlation weight is assigned a value of 0.9, while the weight of the stable region is 0.3. After weighting and fusion, a spatiotemporal feature tensor is generated. This tensor centrally reflects the long-term gas accumulation region, such as the location near a specific battery pack. This tensor provides high-dimensional support for subsequent risk assessment, effectively revealing the temporal dependence of gas diffusion within the warehouse, improving the accuracy and foresight of environmental perception, and ultimately supporting more timely battery management decisions.

[0054] In step S17, the step of processing the spatiotemporal feature tensor using a preset temporal prediction model to obtain the probability of abnormal states includes: The spatiotemporal feature tensor is input into a preset temporal prediction model to extract time dimension information, resulting in a deep temporal coding vector; Decode the deep temporal coding vector to obtain the feature sequence of future time periods; The future time period feature sequence is dimension-mapped to obtain a state category score matrix; The state category score matrix is ​​normalized to obtain a multi-category probability distribution matrix; Abnormal values ​​are extracted based on the multi-category probability distribution matrix to determine the probability of abnormal states.

[0055] First, the spatiotemporal feature tensor formed by stacking node embedding vectors can be used as input data and passed to a pre-defined temporal prediction model to generate a deep temporal encoding vector. Specifically, pattern information in the time dimension is extracted from the tensor to reflect the changing pattern of gas concentration over time. For example, if there are 5 battery nodes in a warehouse, the spatiotemporal feature tensors of each node over the past 6 hours are stacked into a multi-dimensional tensor. After processing by the pre-defined temporal prediction model, an encoding vector is output, which condenses the dynamic features of each node within the time window.

[0056] It should be noted that the preset time-series prediction model typically employs an encoder-decoder architecture, such as LSTM. Its training is based on a historical battery monitoring dataset, which contains long-term time-series node states, such as temperature, voltage, gas concentration, and corresponding environmental parameters and anomaly labels. The hidden state dimension of the encoder and decoder LSTM is set to 128, the number of stacked layers is set to 2, and the dropout rate of 0.2 is set to randomly discard 20% of neurons during forward propagation to mitigate overfitting. For example, in a training process, the initial learning rate is set to 0.001 and may decay with training, the sample batch size used for each parameter update is set to 32, and the number of iterations to fully traverse the training set is set to 100. The input time window length corresponds to 6 hours; if the sampling interval is 5 minutes, the time step is 72; and the output step size is 12 for predicting the next hour. During training, the encoder progressively encodes the input multi-node time-series data, and the final hidden state is the required deep time-series encoding vector with a dimension of 128. The decoder then uses this vector as the initial state to progressively generate the prediction sequence for the next 12 time steps.

[0057] Subsequently, when the deep temporal encoding vector is input into the decoder structure, the aim is to generate a feature representation sequence for future time periods. The decoder, through parsing the encoding vector, predicts the state trend of each node over a future period. For a specific warehouse node, the decoder predicts the gas concentration change characteristics over the next 3 hours, generating a sequence showing that the concentration may gradually increase from the current 2.5 to 2.9. This sequence provides the foundational data for subsequent analysis. Let the deep temporal encoding vector be... Decoder initial hidden state Initial input The current time node features. For each future prediction time step... For example, T=12 corresponds to 3 hours, and the hidden state is updated based on the previous hidden state and the current input. The DecoderCell is the core recursive computation unit within the decoder, responsible for integrating historical information and updating the internal state at each step of prediction. Its implementation can be any neural network structure conforming to this functional definition, such as GRU or LSTM. For example, LSTM is used here, where t-1 represents the previous time step, and the predicted features are calculated through the output layer. ,in , For learnable parameters, Activation function .Will As input for the next time step Final output sequence This refers to a characteristic sequence for future time periods, such as a gas concentration sequence [2.5, 2.55, ..., 2.9].

[0058] Secondly, the process of performing dimensionality mapping on the feature sequence representation of future time periods to obtain the state category score matrix mainly involves transforming the feature sequence into scores for different state categories. Key statistical features are then extracted from the feature sequence of future time periods, such as the maximum gas concentration within the future time period. ,average value The difference between the final time and the initial time and the slope of the linear fit of its upward trend. These statistical features are concatenated into a comprehensive feature vector. The vector is then input into a fully connected network for classification mapping, and the calculation formula is as follows: ,in and These are learnable parameters, automatically optimized on historical datasets using backpropagation and gradient descent algorithms. Before training, they are typically randomly initialized, such as sampling from a normal distribution with a mean of 0 and a standard deviation of 0.01. After training, they may converge to a set of values ​​that effectively distinguish states, for example... Possible values ​​are [[0.12, -0.05, 0.30, 1.20], [0.50, 0.80, 0.15, 0.90], [0.05, 0.10, 0.80, 2.50]]. The input feature vector f could be [0.1, 0.3, -0.2], such that the input feature vector f is calculated... The original score is then obtained. The resulting vector... This is the state category score matrix, whose three components represent whether the node is judged as normal, slightly abnormal, or severely abnormal, with scores of 0.2, 0.6, and 0.1 respectively.

[0059] Next, Softmax normalization is performed on the state category score matrix to generate a multi-class probability distribution matrix. This aims to convert the scores into probability values, ensuring that the sum is 1. After normalization, the probability distributions of the matrices with scores of 0.2, 0.6, and 0.1 are adjusted to 0.17, 0.5, and 0.33, making the results more consistent with probabilistic logic and facilitating subsequent decision-making.

[0060] Finally, extracting values ​​for the anomaly label dimension from the multi-category probability distribution matrix to determine the probability of anomalies in future time periods is a crucial step in assessing anomaly risk. From the matrix, a node is identified as having a minor anomaly probability of 0.5 and a severe anomaly probability of 0.33. Combining this with historical data and business rules, it is determined that this node may pose a future gas leak risk. For nodes in high-risk areas, the prediction frequency is increased from once per hour to once every 30 minutes to improve monitoring density. This logical design, from core to extension, ensures the comprehensiveness and flexibility of the solution.

[0061] In step S18, when any node in the abnormal state probability exceeds a preset abnormal state probability threshold, the risk source node and potential propagation path are traced along the reverse propagation path to determine the thermal runaway risk area, including: When the probability of the abnormal state exceeds the preset abnormal state probability threshold, the target node to be traced is determined. Based on the target node to be traced, a reverse propagation path is traced to obtain the upstream connection node. The feature contribution value of the upstream connection node is calculated to obtain the risk source node. A potential propagation path is constructed based on the risk source node and the target node to be traced, and a weighted summation is performed on the abnormal state probabilities of the nodes on the potential propagation path to obtain the comprehensive risk value of the path. By aggregating all potential propagation paths whose combined risk value exceeds the safety limit and the upstream connection nodes, the thermal runaway risk area is determined.

[0062] First, after obtaining the probability of abnormal states for each node, if the probability of an abnormal state for a node exceeds a preset threshold, which is initialized by statistically analyzing the historical abnormal state probabilities of each node, then the target node to be traced is identified. The threshold is set to 0.4. If the probability of a battery node experiencing a severe abnormality reaches 0.55 within the next 3 hours, then that node is marked as a target node to be traced. This judgment helps to quickly pinpoint potential thermal runaway risks and improves response efficiency.

[0063] Subsequently, based on the topology diagram and the energy transfer correlation weight matrix between nodes, the characteristic contribution values ​​of upstream connected nodes to the traceable abnormal target node are calculated along the backpropagation path to determine the risk source node. The topology diagram reflects the physical connections and energy flow relationships between battery nodes within the warehouse, while the weight matrix quantifies the transfer intensity. For the traceable abnormal target node A, its upstream node B has a weight of 0.7 and node C has a weight of 0.3. Through backpropagation calculation, the characteristic contribution of node B is 0.62, and that of node C is 0.28, thus identifying node B as the risk source node. This calculation method accurately traces the source of the anomaly, avoiding blind inspections.

[0064] Specifically, the implementation of backpropagation computation is based on the principle of backward gradient flow in message passing within graph neural networks. After model training is complete, once an abnormal target node to be traced is identified, the system fixes all model parameters, uses the abnormal probability of node A (0.55) as the initial gradient, and calculates the gradient backpropagation from the target node to its upstream neighbor nodes along the reverse direction of the topological connections using the chain rule. Specifically, it calculates the gradient of node A's embedding vector with respect to its abnormal probability. .

[0065] For example, the embedding vector of node A The input is fed into the subsequent fully connected layer and the softmax output layer to calculate its anomaly probability value. In the reverse tracing phase, all model parameters are fixed, and a model is constructed based on the probability output. Backtracking to the embedding vector The local computation graph. The probability... Starting from itself, the loss function is calculated according to the chain rule. right The partial derivative of the gradient is typically initialized to 1 or a scalar related to the error. This gradient is then backpropagated through the Jacobian matrix of the Softmax layer and the weight matrix of the fully connected layer, ultimately reaching the embedding vector. In the layer in question, the gradient vector obtained at this point is the desired result. .

[0066] Secondly, based on the inverse formula of the aggregation function during message passing, the gradient is decomposed into the contribution of each of its upstream neighbor nodes j. For the aggregation function using attention weights... ， Node j pairs The contribution gradient is ,in These are the forward propagation attention weights recorded during training, such as 0.7 and 0.3. This is the weight matrix of a linear transformation. The weights need to be declared here. These are the normalized attention weights from S14 mentioned earlier.

[0067] By Projecting onto the feature dimension and calculating its norm or dot product with the original features of node j yields the feature contribution values ​​of node j: 0.62 and 0.28. This process essentially utilizes the internal gradient flow path of a trained graph neural network to quantify the "influence" of upstream nodes on the abnormal state of the target node. Its correctness depends on the model successfully learning the causal relationships between nodes, and the tracing path being influenced by trained attention weights. and transformation matrix The constraints mean that the calculated contribution can reflect the causal relationship contained in the training data, thus achieving accurate tracing.

[0068] Secondly, a potential propagation path is constructed based on the risk source node and the target node to be traced for anomalies. The anomaly probabilities of the nodes along this potential propagation path are then weighted and summed to obtain a comprehensive path risk value. The path from risk source node B to target node A includes BDA, with node anomaly probabilities of 0.45, 0.30, and 0.55, respectively. After weighted summation, the comprehensive path risk value is 0.47. The weights are then allocated based on the percentage of anomaly probabilities, resulting in weight ratios of 0.35, 0.23, and 0.42. Summing these anomaly probabilities yields a final risk percentage of 0.46. This risk value reflects the overall threat level of the entire path, facilitating priority intervention on high-risk links.

[0069] Finally, all potential propagation paths whose combined risk value exceeds the safety limit and their coverage areas are aggregated to determine the thermal runaway risk area. The safety limit can be set to 0.45. By analyzing historical normal and abnormal data, the 95th percentile of the risk value of the normal path is taken as the benchmark reference for the safety limit. If the risk values ​​of the three paths are 0.47, 0.52, and 0.41, respectively, then the node groups covered by the first two paths are aggregated to form a high-risk area containing 8 battery nodes.

[0070] In summary, this invention discloses a battery storage management method that significantly improves the accuracy of risk identification and the timeliness of early warning through dynamic topology modeling and multi-feature time series fusion, thereby achieving safe management of battery storage.

[0071] Reference Figure 2 The second embodiment of the present invention provides a battery storage management system, including: The data acquisition module is used to acquire battery temperature, battery voltage, gas concentration data, raw signal strength and unique identifier, calculate the real-time position coordinates of the battery based on the raw signal strength, and construct a spatial topology structure based on the real-time position coordinates of the battery and the unique identifier to obtain the warehouse space topology structure. The matrix construction module is used to determine the set of adjacent batteries of a single battery based on the topology of the storage space, and obtain a dynamic adjacency matrix. The data mapping module is used to map the battery temperature, battery voltage and gas concentration data to the corresponding nodes of the dynamic adjacency matrix to obtain the node feature sequence; The data association module is used to perform message passing between the node feature sequence and the dynamic adjacency matrix using a preset graph neural network to obtain the embedding vector of each node; The vector update module is used to incorporate the gas concentration data into the embedding vectors of each node and propagate it directionally along the adjacent edges of the storage space topology to obtain the updated node embedding vectors. The spatiotemporal integration module is used to construct a stacked matrix along the time dimension based on the updated node embedding vector to obtain a spatiotemporal feature tensor; The feature analysis module is used to process the spatiotemporal feature tensor using a preset time series prediction model to obtain the probability of abnormal states; The risk location module is used to trace the risk source node and potential propagation path along the reverse propagation path and determine the thermal runaway risk area when any node in the abnormal state probability exceeds the preset abnormal state probability threshold.

[0072] It should be noted that the battery storage management system provided in this embodiment of the invention is used to execute all the process steps of the battery storage management 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.

[0073] 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.

[0074] 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 battery storage management method, characterized in that, include: Acquire battery temperature, battery voltage, gas concentration data, raw signal strength and unique identifier; calculate the real-time position coordinates of the battery based on the raw signal strength; construct a spatial topology based on the real-time position coordinates of the battery and the unique identifier to obtain the warehouse space topology. Based on the storage space topology, the set of adjacent batteries for a single battery is determined, and a dynamic adjacency matrix is ​​obtained. The battery temperature, battery voltage, and gas concentration data are mapped to the corresponding nodes of the dynamic adjacency matrix to obtain a node feature sequence; A preset graph neural network is used to perform message passing between the node feature sequence and the dynamic adjacency matrix to obtain the embedding vector of each node; The gas concentration data is incorporated into the embedding vectors of each node and propagated directionally along the adjacent edges of the storage space topology to obtain updated node embedding vectors. Based on the updated node embedding vector, a stacked matrix is ​​constructed along the time dimension to obtain the spatiotemporal feature tensor; The spatiotemporal feature tensor is processed using a preset temporal prediction model to obtain the probability of abnormal states; When any node in the abnormal state probability exceeds the preset abnormal state probability threshold, the risk source node and potential propagation path are traced along the reverse propagation path to determine the thermal runaway risk area.

2. The battery storage management method according to claim 1, characterized in that, The process involves acquiring battery temperature, battery voltage, gas concentration data, raw signal strength, and a unique identifier; calculating the real-time battery location coordinates based on the raw signal strength; and constructing a spatial topology based on the battery location coordinates and the unique identifier to obtain the warehouse space topology, including: The battery temperature, battery voltage, and gas concentration data are acquired by sensors. The original signal strength and unique identifier are collected by a receiving and reading device. The real-time position coordinates of the battery are obtained by performing trilateration calculations based on the original signal strength. The real-time position coordinates of the battery are offset from the preset reference position to obtain corrected position data. An identifier correspondence is generated based on the corrected position data and the unique identifier. Topology nodes are generated based on the identification correspondence and the corrected location data, and the warehouse space topology structure is constructed based on the topology nodes.

3. The battery storage management method according to claim 1, characterized in that, The step of determining the set of adjacent batteries for a single battery based on the topology of the storage space to obtain a dynamic adjacency matrix includes: Obtain the coordinate data of all battery nodes in the warehouse space topology, calculate the Euclidean spatial distance based on the battery node coordinate data and construct a matrix to obtain the initial distance matrix; The spatial distance values ​​are obtained from the initial distance matrix, and the spatial distance values ​​are marked with blocking status in combination with the preset warehouse obstacle distribution map to obtain the marked distance data; Based on the marked distance data, filter connected node pairs that meet the dynamic neighborhood conditions, and generate a distance-oriented weighted list based on the connected node pairs; Based on the distance-oriented weighted list, construct the set of adjacent batteries for each battery to obtain a dynamic adjacency matrix.

4. The battery storage management method according to claim 1, characterized in that, The step of mapping the battery temperature, battery voltage, and gas concentration data to the corresponding nodes of the dynamic adjacency matrix to obtain a node feature sequence includes: The battery temperature, battery voltage, and gas concentration data collected by the sensor are integrated to obtain the physical quantity values; The physical quantity values ​​are mapped to the warehouse space topology to generate a multi-dimensional original feature vector. The multidimensional original feature vectors are weighted and fused according to the dynamic adjacency matrix to obtain a weighted feature vector; The weighted feature vectors are stacked using a time-series sliding window to generate a sequence tensor, and the sequence tensor is filled into the corresponding nodes of the dynamic adjacency matrix to obtain the node feature sequence.

5. The battery storage management method according to claim 1, characterized in that, The step involves using a pre-defined graph neural network to perform message passing between the node feature sequence and the dynamic adjacency matrix to obtain the embedding vector of each node, including: Based on the dynamic adjacency matrix, conduction paths are established between nodes in the storage space topology, and energy gradients are determined. A message carrier is constructed based on the energy gradient, and the node feature sequence is mapped to a multidimensional space to obtain the initial message. A preset graph neural network is used to assign neighborhood weights to the initial message. When the diffusion rate of the energy gradient exceeds a preset diffusion rate threshold, the transmission path is updated to obtain the updated transmission path. Based on the updated propagation path, multiple rounds of message passing are performed to compress the state tensor into an embedded representation, resulting in the embedding vector of each node.

6. The battery storage management method according to claim 1, characterized in that, The process of incorporating the gas concentration data into the embedding vectors of each node and propagating it directionally along the adjacent edges of the storage space topology to obtain updated node embedding vectors includes: Obtain the gas concentration values ​​and topological connections of each node, calculate the gas concentration difference between adjacent nodes, and obtain the gas concentration gradient vector. The concentration gradient vector is projected onto the adjacent edge direction of the topological connection, and a propagation weight characterizing the gas flow along the adjacent edge is generated to obtain the directional propagation weight. The features of adjacent nodes are weighted according to the directional propagation weights to obtain the diffusion message tensor; The spatial decay operation is performed on the diffusion message tensor based on the physical length of the adjacent edges to obtain the signal sequence to be aggregated; The signal sequence to be aggregated is aggregated and fused with the current node state to obtain the updated node embedding vector.

7. The battery storage management method according to claim 1, characterized in that, The step of constructing a stacked matrix along the time dimension based on the updated node embedding vector to obtain the spatiotemporal feature tensor includes: The updated node embedding vectors are arranged and aligned in chronological order to obtain the standardized evolution trajectory; The standardized evolution trajectories are stacked along the time axis to obtain a multidimensional stacked matrix; The local temporal patterns are extracted by scanning the multidimensional stacked matrix using a sliding convolution kernel group to obtain the hidden layer states; The association weights are calculated based on the hidden state, and the hidden state is weighted and fused based on the association weights to obtain the spatiotemporal feature tensor.

8. The battery storage management method according to claim 1, characterized in that, The step of processing the spatiotemporal feature tensor using a preset temporal prediction model to obtain the probability of abnormal states includes: The spatiotemporal feature tensor is input into a preset temporal prediction model to extract time dimension information, resulting in a deep temporal coding vector; Decode the deep temporal coding vector to obtain the feature sequence of future time periods; The future time period feature sequence is dimension-mapped to obtain a state category score matrix; The state category score matrix is ​​normalized to obtain a multi-category probability distribution matrix; Abnormal values ​​are extracted based on the multi-category probability distribution matrix to determine the probability of abnormal states.

9. The battery storage management method according to claim 1, characterized in that, When any node in the abnormal state probability exceeds a preset abnormal state probability threshold, the risk source node and potential propagation path are traced along the reverse propagation path to determine the thermal runaway risk area, including: When the probability of the abnormal state exceeds the preset abnormal state probability threshold, the target node to be traced is determined. Based on the target node to be traced, a reverse propagation path is traced to obtain the upstream connection node. The feature contribution value of the upstream connection node is calculated to obtain the risk source node. A potential propagation path is constructed based on the risk source node and the target node to be traced, and a weighted summation is performed on the abnormal state probabilities of the nodes on the potential propagation path to obtain the comprehensive risk value of the path. By aggregating all potential propagation paths and upstream connection nodes whose combined risk value exceeds a preset safety limit, thermal runaway risk areas are determined.

10. A battery storage management system, characterized in that, include: The data acquisition module is used to acquire battery temperature, battery voltage, gas concentration data, raw signal strength and unique identifier, calculate the real-time position coordinates of the battery based on the raw signal strength, and construct a spatial topology structure based on the real-time position coordinates of the battery and the unique identifier to obtain the warehouse space topology structure. The matrix construction module is used to determine the set of adjacent batteries of a single battery based on the topology of the storage space, and obtain a dynamic adjacency matrix. The data mapping module is used to map the battery temperature, battery voltage and gas concentration data to the corresponding nodes of the dynamic adjacency matrix to obtain the node feature sequence; The data association module is used to perform message passing between the node feature sequence and the dynamic adjacency matrix using a preset graph neural network to obtain the embedding vector of each node; The vector update module is used to incorporate the gas concentration data into the embedding vectors of each node and propagate it directionally along the adjacent edges of the storage space topology to obtain the updated node embedding vectors. The spatiotemporal integration module is used to construct a stacked matrix along the time dimension based on the updated node embedding vector to obtain a spatiotemporal feature tensor; The feature analysis module is used to process the spatiotemporal feature tensor using a preset time series prediction model to obtain the probability of abnormal states; The risk location module is used to trace the risk source node and potential propagation path along the reverse propagation path and determine the thermal runaway risk area when any node in the abnormal state probability exceeds the preset abnormal state probability threshold.

Images