Energy storage power station fire prediction method based on spatio-temporal data

Abstract

The invention discloses an energy storage power station fire hazard prediction method based on spatio-temporal data, and the method comprises the steps: constructing a multi-dimensional spatio-temporal data set, constructing and training a Fireformer model, carrying out the real-time reasoning prediction of a fire hazard risk through the deployment of the trained model, and carrying out the continuous optimization and iteration upgrading of the Fireformer model through an actual prediction result. According to the method, environment data of detectors at all parts in an energy storage cabin are collected, spatial-temporal characteristics are effectively extracted by using a cube attention mechanism, a hierarchical encoder-decoder architecture is combined with a global vector mechanism, local and global dynamic information in detector data is captured, and accurate fire risk prediction is realized. According to the method, the fire risk can be evaluated in real time, and early warning is triggered when the risk reaches the preset threshold value.

Description

Technical Field

[0001] This invention relates to the field of fire early warning for energy storage power stations, and in particular to a multivariable spatiotemporal sequence fire early warning method and edge terminal. Background Technology

[0002] Existing fire early warning systems in electrochemical energy storage power plants generally rely on time-series data generated by various sensors (such as temperature, smoke, and VOC gas detectors) deployed inside the cabin. Traditional analysis methods and prediction models typically treat this sensor data as independent or flattened multivariate time series, focusing only on mining and modeling its time-domain characteristics.

[0003] However, this type of method has a fundamental flaw: it completely ignores the crucial information contained in the three-dimensional distribution of detectors in physical space. In a typical energy storage module design, a large number of detectors are not randomly deployed, but rather arranged in a regular, logical three-dimensional spatial array structure (e.g., a 2 (width) × 6 (length) × 8 (height) topological network) based on the physical layout of the battery pack. Each detector data point is not only correlated with its preceding and following states on the time axis, but also strongly correlated with its neighboring detectors in the three-dimensional grid on the spatial dimension.

[0004] The lack of this spatial topological relationship leads to several challenges for existing technologies: 1. Inability to accurately locate and trace the source of fires: Fires typically originate in a localized location (such as an overheated battery pack) and gradually spread outwards. Traditional time-series models cannot effectively utilize the spatial clustering of alarm signals, making it difficult to determine the precise origin of the fire and thus missing the optimal opportunity for early intervention. 2. Difficulty in modeling the spatial propagation patterns of fires: The diffusion of heat, smoke, and gases has a clear direction and continuity in physical space. Flattening the data disrupts this continuous spatial relationship, preventing the model from learning the dynamic patterns of fire propagation along specific paths (such as cable trays and ventilation ducts), thus limiting its predictive capabilities. 3. Insufficient information utilization and high noise sensitivity: False alarms from a single sensor are common, but a real fire usually triggers a coordinated response from multiple adjacent detectors in space. Existing methods, unable to perceive this spatial correlation, often treat isolated alarms as noise, leading to underreporting of real risks; or conversely, they cannot distinguish between spatially discrete false alarm signals. 4. Although some studies have attempted to introduce convolutional neural networks (CNNs) to capture spatial features, their fixed local receptive fields make it difficult to efficiently model long-range spatial dependencies and complex spatiotemporal dynamic interactions. Recurrent neural networks (RNNs), on the other hand, suffer from vanishing or exploding gradients when dealing with long-sequence time dependencies. Therefore, developing a novel model architecture capable of deeply fusing time series and three-dimensional spatial structure information has become a key technological breakthrough for improving the accuracy, timeliness, and reliability of fire early warning systems for electrochemical energy storage power plants. 5. Traditional fire early warning methods typically utilize detectors to detect relevant indicators in the target environment, such as smoke and temperature, and determine whether these indicators exceed set thresholds to achieve fire detection. However, traditional detector methods only consider a single environmental parameter with a fixed threshold, making it difficult to effectively cope with complex, variable, and interference-filled scenarios, leading to false alarms and missed alarms.

[0005] For example, invention application number 202211545345.6 discloses a method and system for early warning of fires in energy storage power stations based on multiple sensors. This application uses multiple sensors and a normal model obtained through extensive data analysis to identify abnormal situations. It employs a hierarchical cumulative decision-making approach to avoid false alarms. Furthermore, it uses different handling methods for different alarm levels, helping regulatory personnel quickly determine the stage of the fire and improving early warning efficiency. However, this solution does not consider the three-dimensional array structure of the detectors, which could lead to false alarms and missed alarms.

[0006] Therefore, a fire prediction method for energy storage power stations based on spatiotemporal data is needed to fully utilize the spatiotemporal data characteristics within the energy storage power station, effectively compensate for the shortcomings of traditional fire early warning methods, improve the accuracy and timeliness of fire early warning for energy storage power stations, and provide more reliable protection for the safe operation of energy storage power stations. Summary of the Invention

[0007] To address the aforementioned problems, the present invention aims to provide a fire prediction method for energy storage power stations based on spatiotemporal data. This method utilizes the spatiotemporal data characteristics of detectors within the energy storage power station to significantly reduce false alarm and false alarm rates and achieve early and accurate warnings.

[0008] This invention provides a method for predicting fires in energy storage power stations based on spatiotemporal data.

[0009] S1. Based on the spatially distributed three-dimensional structure, historical data of the energy storage power station sensor network are collected, preprocessed, and then a multi-dimensional spatiotemporal dataset is constructed.

[0010] S2. Build and configure the Fireformer model;

[0011] S3. Train and optimize the Fireformer model based on a multidimensional spatiotemporal dataset;

[0012] S4. Deploy the trained and optimized Fireformer model to perform real-time fire inference and prediction.

[0013] S5. Utilize real-time inference prediction results to continuously optimize and iteratively upgrade the Fireformer model.

[0014] In one embodiment of the present invention, the preprocessing in S1 includes:

[0015] The data is standardized to eliminate the differences in dimensions between different sensors; missing values ​​are repaired using an interpolation algorithm based on spatiotemporal nearest neighbors.

[0016] In one embodiment of the present invention, the construction of a multidimensional spatiotemporal dataset in S1 includes:

[0017] The energy storage power station sensor network is deployed in three dimensions according to the physical layout of the battery pack. Based on the sensors, temperature, smoke concentration and VOC concentration are collected to form historical data. The historical data is extracted using a sliding time window and encoded into a multidimensional spatiotemporal tensor. The multidimensional spatiotemporal tensor is used to construct a multidimensional spatiotemporal dataset.

[0018] In one embodiment of the present invention, the Fireformer model adopts a hierarchical encoder-decoder architecture based on a cube attention mechanism and introduces a global vector.

[0019] In one embodiment of the present invention, the following is stated:

[0020] The cube attention mechanism includes: decomposing the input spatiotemporal tensor into multiple non-overlapping cubes, performing self-attention computation within each cube, and capturing spatiotemporal features at different scales by setting different cube sizes and decomposition strategies.

[0021] The global vectors include: introducing a set of learnable global vectors as a summary or memory unit of the system state, and each cube pays attention not only to its internal elements but also to the global vectors when performing attention;

[0022] The hierarchical encoder-decoder consists of: the encoder progressively compresses spatial resolution and extracts multi-level features through a downsampling module and multi-layer cubic attention blocks; the decoder generates predictions in a coarse-to-fine manner through an upsampling module and cubic cross-attention blocks.

[0023] In one embodiment of the present invention, the encoder employs a two-level hierarchy.

[0024] The first-stage encoder and the second-stage encoder employ an axial attention mode within the stacked cubic attention block, including:

[0025] Extract global temporal dependencies along the time dimension, extract dependencies along the spatial height dimension, and extract dependencies along the spatial width dimension.

[0026] In one embodiment of the present invention, the decoder employs a cube cross-attention mechanism to fuse features from the decoding process with multi-level features from the encoder output.

[0027] In one embodiment of the present invention, when training and optimizing the Fireformer model in S3, a multi-class FocalLoss function is used as the training objective of the model to optimize the imbalanced sample, as expressed by the formula:

[0028]

[0029] in, For the number of categories, The category weights are set based on the reciprocal of the number of samples for each fire level in the training set. To focus parameters, For the model to class The predicted probability, The true label for the sample.

[0030] In one embodiment of the present invention, when training and optimizing the Fireformer model in S3, the AdamW optimizer is used, and a linear learning rate warm-up and cosine annealing scheduling strategy is employed.

[0031] In one embodiment of the present invention, when deploying the trained and optimized Fireformer model, the trained and optimized model is pruned and quantized, and the trained and optimized model is integrated into the inference engine of the edge terminal, including the following modules:

[0032] Data acquisition and preprocessing module: Receives sensor network data in real time and performs preprocessing and multidimensional spatiotemporal tensor reconstruction operations;

[0033] Model inference module: Loads the trained and optimized model, performs forward computation on the input multidimensional spatiotemporal tensor, and outputs the probability distribution prediction results;

[0034] Early warning decision module: Analyzes the probability distribution output by the model and triggers the corresponding level of fire early warning by combining the dynamic confidence threshold.

[0035] The beneficial effects of this invention are:

[0036] 1. The method of the present invention collects environmental data from each detector in the energy storage compartment and preprocesses it. It effectively extracts spatiotemporal features through the Cuboid Attention mechanism and adopts a hierarchical encoder-decoder architecture combined with a global vector mechanism to capture local and global dynamic information in the data, thereby making accurate fire risk prediction. It can assess the risk of fire occurrence in real time and trigger an early warning when the preset threshold is reached.

[0037] 2. The method of the present invention deeply integrates three-dimensional spatial and temporal series information and captures multi-scale spatiotemporal features through a cube attention mechanism, which solves the shortcomings of traditional methods that ignore spatial topological relationships, achieves accurate fire location and source tracing, and improves the timeliness of early warning.

[0038] 3. The method of this invention optimizes the sample imbalance problem by using a multi-class FocalLoss function and AdamW optimizer, reduces the risk of false alarms and false negatives from a single sensor, and significantly improves the accuracy and reliability of early warning by combining dynamic confidence thresholds.

[0039] 4. The method of the present invention adapts the model to edge terminals after pruning and quantization optimization. The time for a single inference is no more than 10 milliseconds, which meets the real-time requirements. It also supports continuous iterative upgrades, adapts to different energy storage power station environments, and has strong applicability. Attached Figure Description

[0040] Figure 1 This is a flowchart of the method of the present invention;

[0041] Figure 2 This is a flowchart of the cubic attention block structure of the method of the present invention;

[0042] Figure 3This is a structural diagram of the Fireformer model of the present invention; Figure 4 This is a schematic diagram of the electronic device structure of the present invention. Detailed Implementation

[0043] Embodiments of the present invention are described in detail below. Examples of these embodiments are illustrated in the accompanying drawings, wherein the same or similar symbols denote the same or similar elements or elements having the same or similar functions throughout. The embodiments described below with reference to the accompanying drawings are exemplary and are only used to explain the present invention, and should not be construed as limiting the present invention.

[0044] This invention provides a fire early warning method and system based on a spatiotemporal Transformer architecture, which can significantly reduce false alarm and false negative rates and achieve early and accurate warnings. This invention deeply fuses the multidimensional spatial structure information of the detector network with multivariate time series data and designs a dedicated deep learning model that can simultaneously capture spatiotemporal multi-scale features and overcome the sample imbalance problem.

[0045] Example 1:

[0046] like Figure 1 As shown, this embodiment discloses a fire prediction method for energy storage power stations based on spatiotemporal data, including:

[0047] S1. Based on the spatially distributed three-dimensional structure, historical data of the energy storage power station sensor network are collected, preprocessed, and then a multi-dimensional spatiotemporal dataset is constructed.

[0048] The energy storage power station's sensor network is deployed in three dimensions according to the physical layout of the battery packs. Based on the sensors, it collects data on temperature, smoke concentration, and VOC concentration to form historical data. A sliding time window is used to extract segments of this historical data, which are then encoded into a multidimensional spatiotemporal tensor. This multidimensional spatiotemporal tensor is then used to construct a multidimensional spatiotemporal dataset. Specifically:

[0049] First, data acquisition is performed by synchronously collecting historical data from the sensor network of the electrochemical energy storage power station's fire protection system. The sensor network should be deployed in three dimensions according to the physical layout of the battery pack (e.g., a 2 (width) × 6 (length) × 8 (height) topology), and the types of data collected should include at least temperature, smoke concentration, and VOC (volatile organic compound) concentration.

[0050] Next, data structuring is performed, using a sliding time window (preferably 160 pixels in size with a step size of 1) to truncate the original continuous data. Sensor readings at all spatial points (2 × 6 × 8 = 96 in total) within each time window are no longer considered independent time series, but are reconstructed into a multidimensional spatiotemporal tensor X ∈ R^(T × 6 × 2 × 8 × C), where T is the time step size and C is the number of sensor types. The warning level corresponding to the last time step of this tensor, labeled by expert rules, is the label for that sample. This step explicitly encodes the spatial topological relationships of the detector data into the input data structure.

[0051] The data is no longer a flat time series in the traditional sense, but a spatial feature map that naturally preserves the spatial positional relationships of the detectors. Its time series constitutes the sequence of changes of this spatial feature map over time, laying the foundation for subsequent spatiotemporal collaborative analysis.

[0052] Then, the dataset is partitioned and preprocessed, and the generated multidimensional spatiotemporal dataset samples are divided into training set and test set according to a ratio (e.g., 7:3).

[0053] Preprocessing includes data standardization (such as Z-score) to eliminate differences in dimensions between different sensors. For missing values, a spatiotemporal nearest neighbor-based interpolation algorithm is used for repair.

[0054] S2. Build and configure the Fireformer model.

[0055] like Figure 3 As shown, this embodiment uses Fireformer, which is based on the Cuboid Attention mechanism, as the core prediction model. The Fireformer model adopts a hierarchical encoder-decoder architecture and introduces a global vector.

[0056] like Figure 2 As shown, the cube attention mechanism includes: decomposing the input spatiotemporal tensor into multiple non-overlapping cubes, performing self-attention computation within each cube, and capturing spatiotemporal features at different scales by setting different cube sizes and decomposition strategies.

[0057] In the hierarchical encoder-decoder, the encoder progressively compresses the spatial resolution and extracts multi-level features through a downsampling module and multi-layer cubic attention blocks; the decoder generates predictions in a coarse-to-fine manner through an upsampling module and cubic cross-attention blocks.

[0058] During encoding, the input projection and downsampling are performed first. The input spatiotemporal tensor is first passed through an initial downsampling module consisting of multiple 3×3 2D convolutional layers, group normalization layers, and LeakyReLU activation functions to initially extract spatial features and reduce computational resolution. For example, a spatial size of 6×2×8 can be downsampled to a 32×32 or 16×16 grid through the PatchMerging operation, while increasing the number of channels.

[0059] The encoder design employs a hierarchical structure, comprising two levels.

[0060] First-level encoder: Stack D1 (e.g., 4) cube attention blocks at a relatively high spatial resolution. Each block employs an "axial attention" mode, performing three cube decompositions sequentially:

[0061] Cuboid Attention(cuboid_size=(T,1,1)): Extracts global temporal dependencies along the time dimension.

[0062] Cuboid Attention(cuboid_size=(1,H,1)): Extracts dependencies along the spatial height dimension.

[0063] Cuboid Attention(cuboid_size=(1,1,W)): Extracts dependencies along the spatial width dimension.

[0064] The second-level encoder further downsamples the features and stacks D2 (e.g., 4) of cube attention blocks with the same "axial attention" pattern to capture global features at a lower resolution and a higher level of abstraction.

[0065] The decoder and encoder have a symmetrical structure and employ a coarse-to-fine upsampling process. It uses a cuboid cross-attention mechanism to fuse features from the decoding process with multi-level features from the encoder output. Finally, the spatial resolution of the target prediction is gradually restored through an upsampling module (containing nearest neighbor interpolation and convolutional layers).

[0066] The global vector includes a set of learnable global vectors G∈R^(P×C) (P is preferably 8 in this embodiment), which serve as a summary or memory unit of the system state. When performing attention, each cube pays attention not only to its internal elements but also to the global vector. The global vector also aggregates information from all input elements through attention. This design solves the problem of information isolation between cubes with negligible computational overhead, enabling the model to perceive global dynamics.

[0067] S3. Train and optimize the Fireformer model based on a multidimensional spatiotemporal dataset.

[0068] To alleviate the imbalance between a large number of normal samples and a very small number of fire samples, this embodiment uses a multi-class FocalLoss function as the training target of the model.

[0069]

[0070] in, The category weights are set based on the reciprocal of the number of samples for each fire level in the training set. To focus the parameters (set to 2 in this embodiment), the model is forced to focus on difficult samples (i.e., early fire samples that are prone to misclassification), where k is the number of categories. Let be the predicted probability of the model for category i. The true label for the sample.

[0071] The FocalLoss function works through two mechanisms: first, it introduces class weights, giving higher loss weights to fire categories with fewer samples; second, it sets a focus parameter, causing the model training process to focus on those "difficult samples" that are easily misclassified, forcing the model to prioritize learning high-risk features, thereby greatly improving its sensitivity to real fires.

[0072] When training hyperparameters for the model, the AdamW optimizer was used (learning rate set to 0.001, weight decay to 0.00001), and the batch size was set to 64. Linear learning rate warm-up (warm-up steps accounted for 20% of the total steps) and cosine annealing scheduling strategies were adopted.

[0073] During model training, the model is trained for a total of 100 epochs. An early stopping mechanism is adopted: when the validation set loss no longer decreases within 20 consecutive epochs, training is terminated and the model is rolled back to the point where the optimal model parameters are saved.

[0074] The model ultimately outputs the probability distribution of the warning level for the next K time steps.

[0075] S4. Deploy the trained and optimized Fireformer model to perform real-time fire inference and prediction.

[0076] This embodiment provides a complete end-to-end solution, which performs optimization operations such as pruning and quantization on the trained optimal model, and converts it into a format suitable for inference on edge devices (such as high-performance industrial gateways and industrial control computers) (such as ONNX and TFLite).

[0077] The optimized model is integrated into the inference engine of the edge terminal. After deployment, this terminal includes the following modules:

[0078] Data acquisition and preprocessing module: Receives sensor network data in real time and performs the same standardization and 3D tensor reconstruction operations as in the training phase.

[0079] Model Inference Module: Loads the optimized model, performs forward computation on the input 3D spatiotemporal tensor, and outputs the prediction results.

[0080] Early warning decision module: Analyzes the probability distribution output by the model and triggers the corresponding level of fire early warning based on the dynamic confidence threshold (e.g., 0.7).

[0081] After optimization, using this model, a single inference task should be completed within 10 milliseconds, which can meet the low latency requirement of electrochemical energy storage power stations for real-time early warning.

[0082] S5. Utilize real-time inference prediction results to continuously optimize and iteratively upgrade the Fireformer model.

[0083] The system should record all prediction results and subsequent real-world fire verification information to form a closed loop. New data should be used periodically for incremental training or full-scale fine-tuning of the model to achieve continuous optimization and iterative upgrades, adapting to environmental differences and battery aging at different power plants.

[0084] The present invention also provides an electronic device, Figure 4 This is a schematic diagram of the structure of an electronic device provided in an embodiment of the present invention, such as... Figure 4 As shown, the electronic device may include a processor, a communications interface, memory, and a communication bus, wherein the processor, communications interface, and memory communicate with each other via the communication bus. The processor can invoke logical instructions from the memory, for example, to execute the following method:

[0085] S1. Based on the spatially distributed three-dimensional structure, historical data of the energy storage power station sensor network are collected, preprocessed, and then a multi-dimensional spatiotemporal dataset is constructed.

[0086] S2. Build and configure the Fireformer model;

[0087] S3. Train and optimize the Fireformer model based on a multidimensional spatiotemporal dataset;

[0088] S4. Deploy the trained and optimized Fireformer model to perform real-time fire inference and prediction.

[0089] S5. Utilize real-time inference prediction results to continuously optimize and iteratively upgrade the Fireformer model.

[0090] Furthermore, the logical instructions in the aforementioned memory can be implemented as software functional units and sold or used as independent products, and can be stored in a computer-readable storage medium. Based on this understanding, the technical solution of the present invention, or the part that contributes to the prior art, or a part of the technical solution, can be embodied in the form of a software product. This computer software product is stored in a storage medium and includes several instructions to cause a computer device (which may be a personal computer, server, or network device, etc.) to execute all or part of the steps of the methods described in the various embodiments of the present invention. The aforementioned storage medium includes various media capable of storing program code, such as USB flash drives, portable hard drives, read-only memory (ROM), random access memory (RAM), magnetic disks, or optical disks.

[0091] This invention also provides a non-transitory computer-readable storage medium storing a computer program thereon, which, when executed by a processor, is implemented to perform the methods provided in the above embodiments, including, for example:

[0092] S1. Based on the spatially distributed three-dimensional structure, historical data of the energy storage power station sensor network are collected, preprocessed, and then a multi-dimensional spatiotemporal dataset is constructed.

[0093] S2. Build and configure the Fireformer model;

[0094] S3. Train and optimize the Fireformer model based on a multidimensional spatiotemporal dataset;

[0095] S4. Deploy the trained and optimized Fireformer model to perform real-time fire inference and prediction.

[0096] S5. Utilize real-time inference prediction results to continuously optimize and iteratively upgrade the Fireformer model.

[0097] The device embodiments described above are merely illustrative. The units described as separate components may or may not be physically separate. 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. Those skilled in the art can understand and implement this without any creative effort.

[0098] Through the above description of the embodiments, those skilled in the art can clearly understand that each embodiment can be implemented by means of software plus necessary general-purpose hardware platforms, and of course, it can also be implemented by hardware. Based on this understanding, the above technical solutions, in essence or the part that contributes to the prior art, can be embodied in the form of a software product. This computer software product can be stored in a computer-readable storage medium, such as ROM / RAM, magnetic disk, optical disk, etc., and includes several instructions to cause a computer device (which may be a personal computer, server, or network device, etc.) to execute the methods described in the various embodiments or some parts of the embodiments.

[0099] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention, and not to limit them; although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some of the technical features; and these modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the spirit and scope of the technical solutions of the embodiments of the present invention.

Claims

1. A method for predicting a fire in an energy storage power station based on spatiotemporal data, characterized in that, The method comprises the following steps: S1, based on the spatial distribution of the three-dimensional structure of the energy storage power station sensor network historical data, after preprocessing, a multi-dimensional spatio-temporal data set is constructed; S2, constructing and configuring the Fireformer model; S3, training and optimizing the Fireformer model based on the multi-dimensional spatio-temporal data set; S4, deploying the trained and optimized Fireformer model to conduct real-time inference and prediction of fire; S5, using the real-time inference and prediction results, continuously optimizing and iteratively upgrading the Fireformer model.

2. The method of claim 1, wherein, The preprocessing in S1 includes: Standardizing the data to eliminate the dimensional differences of different sensors; for missing values, use the interpolation algorithm based on spatio-temporal neighbors to repair.

3. The method of claim 1, wherein, The construction of the multi-dimensional spatio-temporal data set in S1 includes: The sensor network of the energy storage power station is deployed in three dimensions according to the physical layout of the battery pack, and temperature, smoke concentration and VOC concentration are collected based on the sensors to form historical data. The historical data is intercepted using a sliding time window, and the data is encoded as a multi-dimensional spatio-temporal tensor. The multi-dimensional spatio-temporal data set is constructed using the multi-dimensional spatio-temporal tensor.

4. The method of claim 1, wherein, The Fireformer model adopts a hierarchical encoder-decoder architecture based on a cubic attention mechanism and introduces a global vector.

5. The method of claim 4, wherein, The cubic attention mechanism includes: Decompose the input spatio-temporal tensor into multiple non-overlapping cubes, and perform self-attention calculation in each cube; by setting different cube sizes and decomposition strategies, capture spatio-temporal features of different scales; The global vector includes: Introduce a set of learnable global vectors as a summary or memory unit of the system state. Each cube not only pays attention to internal elements, but also pays attention to global vectors when performing attention; 6. The method of claim 5, wherein, The hierarchical encoder-decoder includes: The encoder gradually compresses the spatial resolution and extracts multi-level features through the downsampling module and multiple layers of cubic attention blocks; the decoder generates predictions from coarse to fine through the upsampling module and cubic cross-attention blocks. The encoder adopts two levels of hierarchy, 7. The method of claim 5, wherein, The first and second level encoders use axial attention mode in the stacked cubic attention blocks, including:

8. The method of claim 1, wherein, Extracting global temporal dependencies along the time dimension, dependencies along the spatial height dimension, and dependencies along the spatial width dimension. where k is the number of classes, is the class weight set according to the inverse of the number of samples of each fire grade in the training set, is the focus parameter, is the predicted probability of class i by the model, is the true label of the sample.

9. The method of claim 1, wherein, The decoder uses a cubic cross-attention mechanism to fuse the features in the decoding process with the multi-level features output by the encoder.

10. The method of claim 1, wherein, In S3, when training and optimizing the Fireformer model, use the multi-class FocalLoss function as the training target of the model to optimize the sample imbalance, and the formula is: In S3, when training and optimizing the Fireformer model, use the AdamW optimizer, and use linear learning rate warm-up and cosine annealing scheduling strategies. In S4, when deploying the trained and optimized Fireformer model, pruning and quantizing the trained and optimized model, integrating the trained and optimized model into the inference engine of the edge terminal, including the following modules: Data acquisition and preprocessing module: real-time receive sensor network data, perform preprocessing and multidimensional spatiotemporal tensor reconstruction operation; Model inference module: load the model optimized by training, perform forward calculation on the input multidimensional spatiotemporal tensor, and output the probability distribution prediction result; Early warning decision module: analyze the probability distribution output by the model, and trigger the corresponding level of fire warning combined with the dynamic confidence threshold.

Images