Fire-fighting equipment fault prediction method based on deep learning

Abstract

The invention relates to the technical field of fire safety monitoring, and discloses a fire-fighting equipment fault prediction method based on deep learning. According to the method, multi-source operation data of a fire fighting system and an alarm / linkage system are collected, an equipment topological graph is constructed, and a two-channel neural network model fusing time sequence convolution and graph convolution is established. Adaptivity and information propagation efficiency of a topological structure are enhanced through spectrum gap optimization and a greedy topology reconnection mechanism, and flexible alignment and stable fusion of time sequence features and structural features are realized by introducing a multi-modal mixed comparison consistency fusion mechanism. The method can identify the potential fault of the fire-fighting equipment in advance under a complex working condition, realizes accurate prediction and graded alarm, improves the reliability, robustness and intelligent operation and maintenance capability of the system, and is suitable for the field of operation safety monitoring of building fire-fighting systems and industrial fire-fighting equipment.

Description

Technical Field

[0001] This invention relates to the field of fire safety monitoring technology, and in particular to a method for predicting fire equipment failures based on deep learning. Background Technology

[0002] With the increasing complexity of fire protection systems in urban public buildings and industrial facilities, the operational safety and stability of fire protection equipment have become key factors in ensuring the reliability of these systems. Core equipment such as fire pumps, valves, sprinkler / hydrant circuits, detectors, and controllers are susceptible to performance degradation, parameter drift, or potential failure during long-term operation due to various factors including pressure fluctuations, electrical shocks, mechanical vibrations, and changes in ambient temperature and humidity. Failure to identify and address faults early can lead to reduced fire extinguishing capacity, delayed response times, and even system-wide malfunction, posing serious safety risks. While existing intelligent monitoring systems have achieved data acquisition and status identification to some extent, they are mostly limited to anomaly detection based on single-point parameter thresholds or shallow learning algorithms, failing to comprehensively reflect the dynamic evolution of equipment operating status. Although some systems employ neural networks or time-series models for trend prediction, they lack sufficient modeling of the correlation between multi-source data and do not adequately consider the complex physical connections and control logic within the fire protection system, making it difficult to identify potential bottleneck nodes and related fault propagation paths in the topology. Furthermore, existing models exhibit weak robustness and interpretability under complex operating conditions, often resulting in misjudgments or omissions of abnormal signals. Summary of the Invention

[0003] This invention proposes a deep learning-based method for predicting fire equipment faults, aiming to address the shortcomings of existing intelligent systems in multi-source data fusion, structural dependency modeling, and prediction stability. This method introduces a dual-channel neural network structure that integrates temporal dynamic features and equipment topology features, achieving global modeling and dynamic prediction of the operating status of fire equipment. Unlike traditional shallow learning or single-modal models, this invention constructs a spectral gap-guided graph neural network on the complex physical and logical structure of the fire system. Through spectral optimization and a greedy topology reconnection mechanism, it dynamically adjusts the connectivity and information propagation path of the graph structure, significantly enhancing the model's ability to identify structural bottlenecks and abnormal nodes. Simultaneously, a multimodal hybrid contrastive consistency fusion mechanism is proposed, using bidirectional Mixup contrastive loss and many-to-many SoftClip consistency constraints to construct a cross-modal flexible alignment process. This improves the semantic consistency of temporal and structural features while suppressing overfitting under noise interference. Through this innovative system, the model can achieve adaptive feature enhancement and spectral feature optimization while maintaining the safety constraints of the fire system, thereby identifying potential fault risks in advance and outputting stable and interpretable hierarchical alarm results, providing high-precision prediction support for intelligent operation and maintenance and risk decision-making of fire equipment.

[0004] This invention provides a deep learning-based method for predicting fire equipment faults. This method is applied to a fire equipment fault prediction system, which specifically includes a central monitoring server, an alarm controller, fire equipment installed in a pump room, and an edge server. The method includes:

[0005] Step S1: Collect multi-source operation data of fire protection equipment and align them according to timestamps to obtain the aligned multi-source data;

[0006] Step S2: Perform preprocessing on the aligned multi-source data in the edge server to obtain windowed time-series data;

[0007] Step S3: Construct the device topology diagram using the built-in topology modeling engine of the central monitoring server;

[0008] Step S4: Deploy a temporal convolutional-graph convolutional fusion prediction model in the central monitoring server; input windowed temporal data and device topology graph into the temporal convolutional-graph convolutional fusion model; predict the failure probability of the target device;

[0009] Step S5: Based on the failure probability of the target equipment, the alarm controller outputs graded alarm signals to the corresponding fire-fighting equipment, and automatically triggers equipment maintenance, reset and standby equipment linkage operations according to the alarm level. At the same time, it generates disposal suggestions and uploads them to the fire monitoring center to realize predictive maintenance and closed-loop management of fire-fighting equipment.

[0010] Furthermore, step S4 specifically includes the following steps:

[0011] Step S41: Input the windowed time series data into the time series coding branch; The time series coding branch performs hierarchical convolution and dynamic gating operations on the windowed time series data to extract the short-term fluctuation features and long-term evolution trend of the equipment operating status, and obtain time-dependent feature data;

[0012] Step S42: Input the device topology graph into the graph convolution branch. The graph convolution branch processes the device topology graph through the spectral gap-guided graph neural network model to form structural context feature data containing spatial correlation and upstream and downstream dependencies. The construction process of the spectral gap-guided graph neural network model is as follows: Build a graph neural network model, introduce spectral gap optimization and greedy topology reconnection mechanism to optimize the structural adaptability and spectral feature expression ability of the graph neural network model, and obtain the spectral gap-guided graph neural network model.

[0013] Step S43: Employ a multimodal hybrid contrast consistency fusion mechanism to uniformly integrate time-dependent feature data and structural context feature data in the fusion layer of the temporal convolution-graph convolution fusion model to obtain fused state data;

[0014] Step S44: Input the fused state data into the prediction output layer of the temporal convolutional-graph convolutional fusion model, and generate the target device failure probability after fully connected mapping and normalization.

[0015] Furthermore, step S42 specifically includes the following steps:

[0016] Step S421: Based on the node set, edge set, and node feature matrix of the equipment topology graph, construct the adjacency matrix and degree matrix to form a symmetric normalized Laplace matrix; classify and label the edges in the equipment topology graph to form an engineering safety constraint set; the engineering safety constraint set is divided into three categories according to safety priority: edges that cannot be modified, edges that can be weakened and adjusted, and edges that can be logically added.

[0017] Step S422: Perform eigenvalue decomposition using the symmetric normalized Laplacian matrix to obtain eigenvalues ​​and eigenvectors, where the smallest eigenvalue is zero, and the first non-zero eigenvalue is used as the spectral gap; based on the distribution of eigenvalues ​​and the gradient changes of eigenvectors, obtain the spectral gap analysis results; on this basis, locate the regions where the eigenvector gradient changes abruptly and the local spectral gap shrinks significantly, and mark and locate the structural bottleneck region identifiers;

[0018] Step S423: Under the premise of satisfying the engineering safety constraint set, based on the spectral gap analysis results and the structural bottleneck region identification, generate two types of candidate edge sets: deleted candidate set and added candidate set;

[0019] Step S424: For each candidate edge in the two sets of candidate edges, calculate the approximate change in spectral gap one by one using the spectral gap proxy approximation formula, and perform greedy topology reconnection optimization using the principle of maximizing spectral gap gain;

[0020] Step S425: Introduce the Eldan criterion as a stopping condition to dynamically constrain and control the termination of the spectral optimization process, prevent greedy topology reconnection optimization from causing local overfitting, and output the reconnected topology structure and updated spectral feature results.

[0021] Step S426: Perform upstream and downstream control logic consistency verification and topology connectivity verification on the reconnected topology and updated spectral feature results to verify the consistency and effectiveness of the reconnected structure at the physical constraint and control logic level; when the verification is passed, input the verified topology and spectral features into the graph convolution part of the graph neural network model to generate structural context feature data.

[0022] Furthermore, step S43 specifically includes the following steps:

[0023] Step S431: Input the time-dependent feature data and structural context feature data into the modal encoder of the temporal convolution-graph convolution fusion model respectively, and map them to obtain two modal features; wherein, the time-dependent features are mapped to the first modal latent representation vector by the temporal feature encoder, and the structural context features are mapped to the second modal latent representation vector by the graph structure encoder; the two latent representations are unified to the same latent space dimension through the learnable linear mapping layer of the temporal convolution-graph convolution fusion model, and L2 normalization is performed to obtain a unified sample set;

[0024] Step S432: Randomly select an index in the unified sample set, and perform a convex combination operation on the two modal features respectively to obtain a cross-sample mixed sample set;

[0025] Step S433: Construct a bidirectional Mixup contrastive loss function. The bidirectional Mixup contrastive loss function performs forward and reverse alignment between two modalities in the cross-sample mixed sample set simultaneously. This captures the shared relationship between samples in the early stages of training, enhances the robustness of feature fusion, and forms highly consistent fused data.

[0026] Step S434: In the later stages of training, a many-to-many SoftClip consistency mechanism is introduced to replace the bidirectional Mixup contrastive loss function to avoid overfitting noise, thereby obtaining optimized temporal feature embeddings and graph structure embeddings;

[0027] Step S435: The optimized temporal feature embedding and graph structure embedding are concatenated in the fusion layer of the temporal convolution-graph convolution fusion model to obtain fused state data.

[0028] By adopting the above solution, the beneficial effects achieved by the present invention are as follows:

[0029] This invention achieves global modeling and refined fault prediction of fire equipment operation status by constructing a deep prediction model that integrates temporal dynamic features and topological structural features, realizing a transformation from single signal detection to multi-source, multi-dimensional intelligent prediction. By introducing a spectral gap-guided graph neural network structure, it can identify key nodes, bottleneck connections, and potential abnormal paths in the complex pipe network and linkage control topology of the fire protection system, improving the model's understanding of system structural dependencies and energy transfer laws. This mechanism enables efficient information propagation in multi-level control chains such as fire pumps, valves, and sprinkler branches, solving the problem that traditional models cannot reflect the upstream and downstream influences between equipment, enhancing the structural consistency and engineering interpretability of fault prediction results, and significantly improving the reliability and accuracy of fire protection system operation status monitoring.

[0030] This invention introduces a multimodal hybrid contrastive consistency fusion mechanism to achieve dynamic alignment of temporal signals and topological features in the semantic space, improving the robustness and representational accuracy of cross-modal feature fusion. By employing a bidirectional Mixup contrastive loss function in the early stages of training, the model learns a stable shared feature distribution from multidimensional operational samples of fire-fighting equipment, enhancing its generalization ability under imbalanced data and complex operating conditions. In the later stages of training, a many-to-many SoftClip consistency mechanism further suppresses feature drift caused by noise fluctuations, solving the problem of difficult alignment between temporal and structural features in complex environments. This mechanism enables the invention to stably output reliable fault probabilities even when facing typical fire-fighting scenarios such as sudden flow changes, motor overload, abnormal vibration, or multi-point linkage interference, enhancing the system's adaptability and predictive robustness to non-stationary signals.

[0031] Furthermore, this invention achieves adaptive topology reconnection while maintaining fire safety logic constraints by introducing a collaborative mechanism of maximizing spectral gap gain and engineering safety constraint sets in graph structure optimization. This mechanism effectively avoids interference from topology reconfiguration on safety paths such as the main water supply line and control closed loop, ensuring that the model's structural optimization process complies with fire regulations and physical constraints. Simultaneously, it utilizes the spectral gap proxy approximation formula to achieve rapid evaluation and greedy optimization, significantly improving the convergence efficiency and real-time prediction performance of the model training. Through this mechanism, the model not only possesses the ability to predict fault risks but also dynamically identifies vulnerable connection points in the topology and performs inference corrections, enhancing the overall stability and operational guidance value of the system. This realizes the intelligent evolution of fire equipment from "passive response" to "active defense," providing strong technical support for the safety decision-making of fire protection systems. Attached Figure Description

[0032] Figure 1 This is a flowchart illustrating a deep learning-based method for predicting fire equipment faults proposed in this invention.

[0033] Figure 2 This is a graph showing the evolution of the loss component proposed in Example 5. Detailed Implementation

[0034] The technical solutions of the present invention will be clearly and completely described below with reference to the accompanying drawings of the embodiments of the present invention. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. All other embodiments obtained by those skilled in the art based on the embodiments of the present invention without creative effort are within the scope of protection of the present invention.

[0035] Example 1, according to Figure 1This invention provides a deep learning-based method for predicting fire equipment failures. This method is applied to a fire equipment failure prediction system, specifically including fire equipment, an edge server in a pump room, a central monitoring server, and an alarm controller. The method includes:

[0036] Step S1: Data Acquisition and Alignment: Collect multi-source operational data of fire protection equipment in the fire protection system and alarm / linkage system through the data acquisition layer, and align them according to timestamps to obtain the aligned multi-source data;

[0037] Firefighting equipment includes main fire pumps, pressure-stabilizing pumps, solenoid valves, sprinkler / fire hydrant circuits, alarm controllers, gas extinguishing devices, and environmental monitoring units. Multi-source operational data includes: pipeline pressure and flow rate, motor current and voltage, pump speed and start / stop status, valve opening, ambient temperature and humidity, vibration and acoustic signals, alarm / feedback signals from smoke and temperature detectors and gas extinguishing devices, controller event logs, and maintenance records. The data acquisition layer is installed in key parts of the fire protection system to achieve real-time sensing of the operating status of firefighting equipment, environmental parameters, and control signals. This includes: 1. Pipeline monitoring unit: A pressure sensor (0–1.6 MPa range) and a flow sensor (electromagnetic type) are installed at the pump outlet, main pipeline, and branch node to collect pipeline pressure, flow rate, and transient fluctuation data. 2. Pump Unit Monitoring Unit: Three-phase current transformers and voltage detection modules are installed at the motor ends of the main fire pump and the pressure-stabilizing pump to monitor electrical load and power supply status; triaxial MEMS vibration sensors are attached to the pump bearings and base to collect vibration acceleration and spectral characteristics; acoustic pickup sensors (microphone arrays) are installed on the motor cover or pump body surface to monitor abnormal noise characteristics; temperature and humidity sensors are installed on the pump room roof and inside the control cabinet to record changes in ambient temperature and humidity. 3. Alarm and Control Unit: Alarm, feedback, and event log signals from smoke and temperature detectors, gas extinguishing devices, and alarm / linkage controllers are collected to record equipment status changes and control logic responses. Analog signals output by each monitoring unit are converted to digital signals and timestamped, then aggregated via an RS485 bus to an edge computing node and transmitted to an edge server within the pump room.

[0038] Step S2: Preprocessing and windowing: The data acquisition layer is connected to the edge server in the pump room via CAN bus. The edge server performs noise reduction, missing value imputation, outlier suppression and normalization on the aligned multi-source data, and divides the sequence into sample segments according to the sliding window to obtain windowed time series data.

[0039] Step S3: Equipment Topology Modeling: The central monitoring server is located in the control room of the computer room and communicates with the edge server in the pump room. The central monitoring server has a built-in topology modeling engine. In the topology modeling engine, the equipment topology diagram is constructed based on the physical connection relationship and logical control relationship between the fire protection system and the alarm / linkage system. The equipment topology diagram includes a set of nodes, a set of edges, and a node feature matrix. Among them, the set of nodes represents water pumps, valves, sprinkler / fire hydrant circuits, detectors and controllers, the set of edges represents pipeline connections, circuit affiliation and control / linkage relationships, and the node feature matrix records the operating parameters and status characteristics of each node.

[0040] Step S4: Deep prediction model inference: Deploy a temporal convolutional-graph convolutional fusion prediction model in the central monitoring server; input windowed temporal data and device topology map into the temporal convolutional-graph convolutional fusion model; extract time-dependent features from the windowed temporal data through a temporal coding branch, extract structural context features from the device topology map through a graph convolutional branch, fuse the features of the two branches, and predict the failure probability of the target device;

[0041] Step S5: Alarm and Handling Suggestions: The central monitoring server connects to the alarm controller, sets adaptive thresholds for equipment categories and operating conditions, compares the failure probability of the target equipment with the adaptive thresholds for equipment categories and operating conditions, and outputs graded alarm signals to the corresponding fire-fighting equipment through the alarm controller. Based on the alarm level, it automatically triggers equipment maintenance, reset, and standby equipment linkage operations, and generates handling suggestions and uploads them to the fire monitoring center to achieve predictive maintenance and closed-loop management of fire-fighting equipment.

[0042] Example 2, this example is based on Example 1. In this example, step S4 specifically includes the following steps:

[0043] Step S41: Input the windowed time series data into the time series coding branch; The time series coding branch performs hierarchical convolution and dynamic gating operations on the windowed time series data to extract the short-term fluctuation features and long-term evolution trend of the equipment operating status, and obtain time-dependent feature data;

[0044] Step S42: Input the device topology graph into the graph convolution branch. The graph convolution branch processes the device topology graph through the spectral gap-guided graph neural network model to form structural context feature data containing spatial correlation and upstream and downstream dependencies. The construction process of the spectral gap-guided graph neural network model is as follows: Build a graph neural network model, introduce spectral gap optimization and greedy topology reconnection mechanism to optimize the structural adaptability and spectral feature expression ability of the graph neural network model, and obtain the spectral gap-guided graph neural network model.

[0045] Step S43: A multimodal hybrid contrast consistency fusion mechanism is adopted to integrate time-dependent feature data and structural context feature data in the fusion layer of the temporal convolution-graph convolution fusion model to obtain fused state data. The construction method of the multimodal hybrid contrast consistency fusion mechanism is as follows: a multi-stage optimization strategy based on hybrid sample contrast and consistency constraints is introduced. By performing bidirectional hybrid sample contrast learning in the early stage of training, shared semantic relationships and feature mapping consistency are established between different modalities. In the later stage of training, a many-to-many SoftClip consistency constraint mechanism is introduced to calculate the soft weight distribution within the modality and perform weighted consistency alignment across modalities to form a multimodal hybrid contrast consistency fusion mechanism.

[0046] Step S44: Input the fused state data into the prediction output layer of the temporal convolutional-graph convolutional fusion model, and generate the target device failure probability after fully connected mapping and normalization.

[0047] Example 3, this example is based on Example 1. In this example, step S4 specifically includes the following steps:

[0048] Step E1: Input the windowed time series data into the time series coding branch; The time series coding branch performs hierarchical convolution and dynamic gating operations on the windowed time series data to extract the short-term fluctuation features and long-term evolution trends of the equipment operating status, and obtain time-dependent feature data;

[0049] Step E2: Input the device topology graph into the graph convolution branch. The graph convolution branch processes the device topology graph through the graph neural network model to form structural context feature data containing spatial relationships and upstream and downstream dependencies.

[0050] Step E3: Employ a multimodal hybrid contrast consistency fusion mechanism to uniformly integrate time-dependent feature data and structural context feature data at the fusion layer to obtain fused state data;

[0051] Step E4: Input the fused state data into the prediction output layer, and generate the target device failure probability after fully connected mapping and normalization.

[0052] Example 4, this example is based on Example 2. In this example, step S42 specifically includes the following steps:

[0053] Step S421: Based on the node set, edge set, and node feature matrix of the equipment topology graph, construct the adjacency matrix and degree matrix to form a symmetric normalized Laplace matrix; classify and label the edges (connection relationships) in the equipment topology graph to form an engineering safety constraint set. This constraint set is used to provide safety boundary conditions in the subsequent spectral optimization and edge reconnection process to ensure that structural adjustments do not damage the functional integrity, signal transmission stability, and compliance of the fire protection system; the engineering safety constraint set is divided into three categories according to safety priority: edges that cannot be modified (cannot be deleted or weakened), edges that can be weakened and adjusted (adjustable edges), and edges that can be logically added (candidate new edges).

[0054] Edges that cannot be modified (cannot be deleted or weakened): These include the main fire water supply line, the main water supply path from the fire pump to the main pipe, forced linkage loops (such as the control closed loop between the pump, valve, and detector), legally required inspection paths, and signal trunk lines directly connected to the central alarm control unit. These edges play a core role in the system in terms of transmission, control, and monitoring. Deleting or weakening them will lead to functional failure or safety risks. Therefore, they are defined as strongly constrained edges in the process of spectral gap optimization and graph reconnection. They are only allowed to participate in feature propagation calculations and their weights or topology cannot be adjusted.

[0055] Adjustable edges (adjustable edges): These refer to ordinary operational connections other than those that are prohibited from being modified, including auxiliary loops between similar functional units, redundant control paths, and connections between non-critical measurement points. These edges can be dynamically weighted, weakened, or reduced in the subsequent spectrum optimization stage based on the importance of edge weights, autocorrelation, or bottleneck influence factors, in order to optimize information flow paths and alleviate excessive aggregation effects.

[0056] Candidate new edges (logically new edges): These are virtual edges established between node pairs that are functionally compatible and within the limits of the physical structure. These edges are used to enhance topological connectivity and spectral gap width, and improve the effective coverage of information propagation, when the topology is sparse or there are structural bottlenecks. The generation of new edges is based on the similarity of node attributes, functional domain affiliation, and spatial association weights, and must meet the feasibility constraints of fire safety logic (such as electrical isolation, loop tolerance, and signal safety thresholds).

[0057] Step S422: Perform eigenvalue decomposition using the symmetric normalized Laplacian matrix to obtain eigenvalues ​​and eigenvectors, where the smallest eigenvalue is zero. Use the first non-zero eigenvalue as the spectral gap. Based on the eigenvalue distribution and eigenvector gradient changes, obtain the spectral gap analysis results. On this basis, locate regions with abrupt changes in eigenvector gradients and significant reductions in local spectral gaps, and mark and locate structural bottleneck regions. The spectral gap analysis results are used to characterize the connectivity and information flow capacity of the current topology. The smaller the spectral gap, the more likely there is communication blockage or energy transmission bottleneck in that region, making it a key subgraph for subsequent structural optimization.

[0058] Step S423: Under the premise of satisfying the engineering safety constraint set, based on the spectral gap analysis results and the structural bottleneck area identification, generate two types of candidate edge sets: deleted candidate set and new candidate set; among them, the deleted candidate set is only selected from the "weakeable and adjustable edge class", and the new candidate set is only constructed within the scope of "logically addable edge class", to ensure that the candidate generation process meets the safety constraints and physical connectivity requirements of the fire protection system;

[0059] Step S424: For each candidate edge in the two sets of candidate edges, calculate the approximate change in spectral gap (i.e., surrogate gain) one by one using the spectral gap surrogate approximation formula, and perform greedy topology reconnection optimization using the principle of maximizing spectral gap gain;

[0060] The spectral gap surrogate approximation formula is constructed based on matrix perturbation theory and is used to quickly evaluate the impact of adding or removing a single edge on the overall spectral structure. It is specifically expressed as follows:

[0061] ;

[0062] in, This represents the target eigenvalues ​​of the symmetric normalized Laplacian matrix before modifying the edges. This indicates that after flipping a candidate edge (adding / removing), the new eigenvalues ​​are approximately equal; , Representing the eigenvector At the node ,node The components on; Indicates the edge operation identifier;

[0063] Based on the above formula, the approximate change in spectral gap (i.e., surrogate gain) is calculated for each candidate edge, and a greedy reconnection optimization is performed according to the principle of maximizing spectral gap gain: when a candidate edge belongs to the deleted candidate set and the surrogate gain is positive and meets the engineering safety constraints, the edge deletion operation is performed; when a candidate edge belongs to the newly added candidate set and the surrogate gain is positive and the false aggregation threshold is not triggered, the edge addition operation is performed; after each edge operation, the system only performs a few steps of exponential iteration to update the main feature vector and spectral gap, achieving local fast convergence and entering the next round of iteration; this step effectively alleviates the problems of over-smoothing of the topology and information compression by "flipping a few edges and significantly improving the spectral gap" while ensuring real-time performance and scalability.

[0064] Greedy topology reconnection optimization: In the candidate edge set generated in step S423, the spectral gap proxy gain value of each candidate edge is calculated one by one. According to the principle of maximizing gain, the candidate edge with the highest gain value is iteratively selected to perform edge addition or deletion operations. After each edge operation, the spectral feature vector and spectral gap are updated quickly using exponential iteration to realize the dynamic reconnection of the topology and the enhancement of spectral features. This optimization process replaces global recalculation with local greedy updates, and achieves spectral gap improvement and structural bottleneck reduction with low computational complexity under the premise of ensuring engineering safety constraints.

[0065] Step S425: Introduce the Eldan criterion as a stopping condition to dynamically constrain and control the termination of the spectral optimization process, prevent greedy topology reconnection optimization from causing local overfitting, and output the reconnected topology structure and updated spectral feature results.

[0066] Specifically: when the discrimination function of a candidate edge satisfies the sufficient condition that "edge deletion can preserve gap increase", the corresponding edge deletion operation is allowed; when it is predicted that adding an operation will lead to a decrease in spectral gap, the operation is rejected, thereby ensuring that structural adjustment is carried out only along the direction of spectral gap gain and avoiding negative topology reconnection; at the same time, the model continuously monitors the convergence trend of the spectral gap proxy gain, and when the gain of several consecutive iterations is lower than the preset threshold, or the spectral gap has reached the preset upper limit, the stopping condition is automatically triggered to terminate the topology reconnection process; after the stopping condition is met, the system outputs the reconnected topology structure and the updated spectral feature results as inputs for subsequent graph convolution calculations and equipment state inference, providing a highly connected and stable structural foundation for the state assessment and prediction of the fire protection system.

[0067] Step S426: Perform upstream and downstream control logic consistency verification and topological connectivity verification on the reconnected topology and updated spectral feature results to verify the consistency and effectiveness of the reconnected structure at the physical constraint and control logic level. After the verification is passed, the qualified topology and spectral features are input into the graph convolution part of the graph neural network (GNN) model to perform multi-layer message passing and feature normalization update process. In the graph convolution part, a cross-layer skip connection mechanism and residual aggregation strategy are introduced to enhance the transmission stability of deep structural information and feature expression ability. Finally, the structural context feature data that integrates spatial correlation and upstream and downstream dependence is generated through pooling operation.

[0068] Example 5, according to Figure 2 This embodiment is based on Embodiment 4. In this embodiment, step S43 specifically includes the following steps:

[0069] Step S431: Input the time-dependent feature data and structural context feature data into the modal encoder of the temporal convolution-graph convolution fusion model respectively, and map them to obtain two modal features; wherein, the time-dependent features are mapped to the first modal latent representation vector by the temporal feature encoder, and the structural context features are mapped to the second modal latent representation vector by the graph structure encoder; the two latent representations are unified to the same latent space dimension through the learnable linear mapping layer of the temporal convolution-graph convolution fusion model, and L2 normalization is performed to obtain a unified sample set;

[0070] Step S432: Randomly select an index in the unified sample set, and perform convex combination operation on the features of the two modalities respectively to obtain a cross-sample mixed sample set; the convex combination operation performs linear interpolation between samples by setting a mixing factor, thereby simulating the potential correlation between modalities in the latent space and providing a mixed sample basis for subsequent contrastive fusion learning;

[0071] Step S433: Construct a bidirectional Mixup contrastive loss function. The bidirectional Mixup contrastive loss function performs forward and reverse alignment between two modalities in the cross-sample mixed sample set simultaneously. This captures the shared relationship between samples in the early stages of training, enhances the robustness of feature fusion, and forms highly consistent fused data.

[0072] A bidirectional Mixup contrastive loss function is constructed by simultaneously performing forward and reverse Mixup sample alignment between temporal feature embeddings and structural feature embeddings in a cross-sample mixed sample set.

[0073] Specifically, the Mixup contrast loss in the first modal direction is defined as:

[0074] ;

[0075] in, This indicates the first modality direction, using "Modality 1 (temporal feature embedding)" as the anchor point to align with the Mixup contrastive loss of "Modality 2 structural feature embedding"; This represents the number of multimodal sample pairs in a training batch. This represents the index of the current anchor sample. Representation and Sample Perform another sample indexing for the same modality (modality 1) in Mixup; Indicates sample Mixup coefficient, This represents the mixed sample embedding of mode 1 (first perform Mixup, then feed it into the mode 1 encoder). This represents the complete set of normalized embeddings for mode 2 in this batch; This represents the loss for unidirectional multimodal comparison (similarity).

[0076] And the second-mode orientation loss is defined symmetrically:

[0077] ;

[0078] in, This represents the mixed sample embedding of mode 2 (first perform Mixup, then feed it into the mode 2 encoder). This represents the complete set of normalized embeddings for mode 1 in this batch;

[0079] Averaging the losses in both directions yields the bidirectional Mixup contrastive loss function:

[0080] ;

[0081] in, This represents the total loss of the bidirectional Mixup contrastive loss function;

[0082] Through this bidirectional structure, the model can capture shared relationships between samples in the early stages of training, enhance the robustness of feature fusion, and reduce the dependence on "strict one-to-one sample matching", thereby forming a highly consistent fusion representation between temporal and topological features.

[0083] Step S434: In the later stages of training, a many-to-many SoftClip consistency mechanism is introduced to replace the bidirectional Mixup contrastive loss function to avoid overfitting noise, thereby obtaining optimized temporal feature embeddings and graph structure embeddings;

[0084] The many-to-many SoftClip consensus mechanism is based on intramodal soft weights:

[0085] Consistency weight distribution in the modality 1 direction (temporal feature side):

[0086] ;

[0087] in, This indicates the index of the candidate reference sample in mode 1. This indicates that for mode 1, the first... The sample relative to the first The soft weights of each sample. This represents the traversal index used for summing the denominators. In mode 1 (i.e., sequential feature embedding), the first... The embedding vector of each sample, In mode 1, the first The embedding vector of each sample, and They belong to the same mode (timing). In mode 1, the first Embedding of a sample; Represents an exponential function. Indicates the temperature coefficient;

[0088] Consistency weight distribution in mode 2 direction (structural feature side):

[0089] ;

[0090] in, This indicates that for mode 2, the first... The sample relative to the first The soft weights of each sample. In mode 2, the first The structural feature representation vector of each sample In mode 2, the first The embedding vector of each sample, and They belong to the same mode (structural side). In mode 2, the first Embedding of a sample;

[0091] The weights are then used for cross-modal similarity calculation to achieve flexible "many-to-many" matching, thereby maintaining the structural consistency and distribution stability of the fused features under complex working conditions.

[0092] The many-to-many SoftClip consistency mechanism is used in the later stages of model training. It introduces a soft-weight-based many-to-many SoftClip consistency mechanism, which calculates soft similarity distributions within each modality and performs weighted consistency constraints across modalities to achieve flexible many-to-many alignment between two modalities. This mechanism differs from traditional one-to-one or one-to-many contrastive learning methods. It establishes adaptive weight distributions within both temporal and structural modalities and forms a flexible consistency loss for cross-modal aggregation based on these distributions. This maintains the stability of feature distributions and semantic consistency between modalities under complex conditions and noisy environments.

[0093] During model training, to characterize the changes in the dominant relationships of different loss components during the training phase, such as... Figure 2 As shown, the evolution curve of the generated loss component is plotted. Figure 2 In the diagram, the horizontal axis represents the training epoch, and the vertical axis represents the loss value. The red curve represents the bidirectional Mixup contrastive loss L_mixup, and the blue curve represents the many-to-many SoftClip consistency loss L_softclip. In the early stages of training, L_mixup dominates, and the model establishes shared representations of temporal and structural features through contrastive learning. When training reaches around the 40th epoch, the model switches to a many-to-many consistency constraint stage dominated by L_softclip to maintain the stability and consistency of cross-modal embedding distribution.

[0094] Step S435: The optimized temporal feature embedding and graph structure embedding are concatenated in the fusion layer of the temporal convolution-graph convolution fusion model to obtain fused state data.

[0095] Example 6, this example is based on Example 4. In this example, step S43 specifically includes the following steps:

[0096] Step R1: Input the temporal dependent feature data and structural context feature data into the modal encoder of the temporal convolution-graph convolution fusion model respectively, and map them to obtain two modal features; wherein, the temporal dependent features are mapped to the first modal latent representation vector by the temporal feature encoder, and the structural context features are mapped to the second modal latent representation vector by the graph structure encoder; the two latent representations are unified to the same latent space dimension through the learnable linear mapping layer of the temporal convolution-graph convolution fusion model, and L2 normalization is performed to obtain a unified sample set;

[0097] Step R2: Randomly select indices from the unified sample set and perform convex combination operations on the features of the two modalities respectively to obtain a cross-sample mixed sample set; the convex combination operation performs linear interpolation between samples by setting a mixing factor, thereby simulating the potential correlation between modalities in the latent space and providing a mixed sample basis for subsequent contrastive fusion learning;

[0098] Step R3: Construct a contrastive loss function based on InfoNCE. By maximizing the similarity of positive samples and minimizing the similarity of negative samples in the cross-modal mixed sample set, feature alignment and consistency enhancement between cross-modal samples are achieved, generating highly consistent fused data.

[0099] Step R4: The highly consistent fused data is spliced ​​together at the fusion layer to obtain the fused state data.

[0100] Example 7, this example is based on Example 5, in this example:

[0101] Step S4: Input the windowed time series data and the device topology map into the temporal convolution-graph convolution fusion model; the windowed time series data is processed through the temporal coding branch to extract time-dependent features, and the device topology map is processed through the graph convolution branch to extract structural context features. The features of the two branches are fused to predict the failure probability of the target device.

[0102] Step S5: Set the adaptive threshold for equipment category and operating condition, compare the failure probability of the target equipment with the adaptive threshold for equipment category and operating condition, output graded alarms, and generate handling suggestions for maintenance, reset, and linkage verification.

[0103] In this embodiment:

[0104] Initial normalized Laplace spectral gap: 0.5657;

[0105] Spectral gap after reconnection: 0.6667;

[0106] Target equipment failure probability:

[0107] The fault probability field is fused: fault_prob (0~1);

[0108] Adaptive threshold (equipment category = fire pump + dynamic operating condition fluctuation): adaptive_th;

[0109] Alarm tiers: alarm_level (normal / Level I - low / Level II - medium / Level III - high);

[0110] The 3 most serious moments: 09:19

[0112] fault_prob=0.983, adaptive_th=0.670 → Level III - High;

[0113] Specifications: Pressure 6.813 bar (significantly declining), Flow rate 42.99 L / s (overflow), Current 157.8 A (overcurrent), Vibration 4.10 mm / s, Noise 69.1 dB, Controller events 2, smoke_alarm=1; 09:20:

[0115] fault_prob=0.983, adaptive_th=0.655 → Level III - High;

[0116] Specifications: Pressure 7.220 bar, Flow rate 44.58 L / s, Current 164.7 A, Vibration 4.32 mm / s, Noise 69.6 dB, Controller events 2, smoke_alarm=1; 09:18

[0118] fault_prob=0.979, adaptive_th=0.665 → Level III - High;

[0119] Specifications: Pressure 6.894 bar, Flow rate 44.20 L / s, Current 155.9 A, Vibration 3.99 mm / s, Noise 69.8 dB, Controller events 2, smoke_alarm=1;

[0120] The resonant combination of "low pressure + high flow + overcurrent + vibration / noise rise + short-term smoke temperature alarm" from 09:18 to 09:20 matches the actual signs of abnormal pump operation / local problems in the pipeline network. Therefore, the probability of failure is rapidly increased, exceeding the adaptive threshold and triggering Level III-High.

[0121] Alarms and handling suggestions:

[0122] Threshold strategy: The basic threshold for pump equipment is 0.62, which is dynamically adjusted by +0 to 0.06 based on current / pressure / vibration fluctuation.

[0123] suggestion:

[0124] Immediate inspection: Check the pump set bearings and couplings (for vibration and current rise);

[0125] Reset / Check: Reset the relevant smoke temperature detector;

[0126] Short-term retest: Retest the pressure / flow curve and current waveform within 10-15 minutes to confirm whether there are signs of leakage / partial blockage or valve jamming.

[0127] The present invention and its embodiments have been described above. This description is not restrictive. The accompanying drawings are only one embodiment of the present invention, and the actual structure is not limited thereto. In short, if a person skilled in the art is inspired by this description and designs a similar structure and embodiment without departing from the spirit of the present invention, such design should fall within the protection scope of the present invention.

Claims

1. A deep learning-based method for predicting fire equipment faults, applied to a fire equipment fault prediction system, the fire equipment fault prediction system specifically including a central monitoring server, an alarm controller, and fire equipment and edge servers installed in the pump room, characterized in that: The method includes: Step S1: Collect multi-source operation data of fire protection equipment and align them according to timestamps to obtain the aligned multi-source data; Step S2: Perform preprocessing on the aligned multi-source data in the edge server to obtain windowed time-series data; Step S3: Construct the device topology diagram using the built-in topology modeling engine of the central monitoring server; Step S4: Deploy a temporal convolutional-graph convolutional fusion prediction model in the central monitoring server; input windowed temporal data and device topology graph into the temporal convolutional-graph convolutional fusion model; predict the failure probability of the target device; Step S5: Based on the failure probability of the target equipment, the alarm controller outputs graded alarm signals to the corresponding fire-fighting equipment, and automatically triggers equipment maintenance, reset and standby equipment linkage operations according to the alarm level. At the same time, it generates disposal suggestions and uploads them to the fire monitoring center to realize predictive maintenance and closed-loop management of fire-fighting equipment.

2. The method for predicting fire equipment faults based on deep learning according to claim 1, characterized in that: The temporal convolution-graph convolution fusion model is constructed as follows: a temporal coding branch based on a gated temporal convolutional network and a graph convolution branch based on a spectral gap guided graph neural network model are established respectively, and the features of the two branches are fused to form a dual-channel neural network structure, thus obtaining the temporal convolution-graph convolution fusion model.

3. The fire equipment fault prediction method based on deep learning according to claim 2, characterized in that: Step S4 specifically includes the following steps: Step S41: Input the windowed time series data into the time series coding branch to obtain time-dependent feature data; Step S42: Input the device topology graph into the graph convolution branch. The graph convolution branch processes the device topology graph through the spectral gap-guided graph neural network model to form structural context feature data. Step S43: Employ a multimodal hybrid contrast consistency fusion mechanism to uniformly integrate time-dependent feature data and structural context feature data in the fusion layer of the temporal convolution-graph convolution fusion model to obtain fused state data; Step S44: Input the fused state data into the prediction output layer of the temporal convolutional-graph convolutional fusion model to obtain the target device failure probability.

4. The fire equipment fault prediction method based on deep learning according to claim 3, characterized in that: The construction process of the spectral gap-guided graph neural network model: A graph neural network model is established, and a spectral gap optimization and greedy topology reconnection mechanism are introduced to optimize the structural adaptability and spectral feature representation ability of the graph neural network model, thus obtaining the spectral gap-guided graph neural network model.

5. The fire equipment fault prediction method based on deep learning according to claim 4, characterized in that: Step S42 specifically includes the following steps: Step S421: Based on the node set, edge set, and node feature matrix of the equipment topology graph, construct the adjacency matrix and degree matrix to form a symmetric normalized Laplace matrix; classify and label the edges in the equipment topology graph to form an engineering safety constraint set; Step S422: Perform eigenvalue decomposition using the symmetric normalized Laplacian matrix to obtain eigenvalues ​​and eigenvectors, using the first non-zero eigenvalue as the spectral gap; based on the eigenvalue distribution and eigenvector gradient changes, obtain the spectral gap analysis results and mark the structural bottleneck region. Step S423: Under the premise of satisfying the engineering safety constraint set, based on the spectral gap analysis results and the structural bottleneck region identification, generate two types of candidate edge sets: deleted candidate set and added candidate set; Step S424: For each candidate edge in the two sets of candidate edges, calculate the approximate change in spectral gap one by one using the spectral gap proxy approximation formula, and perform greedy topology reconnection optimization using the principle of maximizing spectral gap gain; Step S425: Introduce the Eldan criterion as a stopping condition to dynamically constrain and control the termination of the spectral optimization process, prevent greedy topology reconnection optimization from causing local overfitting, and output the reconnected topology structure and updated spectral feature results. Step S426: Perform upstream and downstream control logic consistency verification and topology connectivity verification on the reconnected topology and updated spectral feature results; when the verification is passed, input the qualified topology and spectral features into the graph convolution part of the graph neural network model to obtain structural context feature data.

6. The fire equipment fault prediction method based on deep learning according to claim 5, characterized in that: The engineering safety constraint set is divided into three categories according to safety priority: prohibited modification of class edges, weakenable adjustment of class edges, and logical addition of class edges.

7. The fire equipment fault prediction method based on deep learning according to claim 3, characterized in that: Step S43 specifically includes the following steps: Step S431: Input the time-dependent feature data and structural context feature data into the modal encoder of the temporal convolution-graph convolution fusion model respectively, map them to obtain two modal features, and perform linear mapping layer processing to obtain a unified sample set; Step S432: Randomly select an index in the unified sample set, and perform a convex combination operation on the two modal features respectively to obtain a cross-sample mixed sample set; Step S433: Construct a bidirectional Mixup contrastive loss function. The bidirectional Mixup contrastive loss function performs forward and backward alignment between two modalities in the cross-sample mixed sample set simultaneously, capturing the shared relationship between samples in the early stage of training. Step S434: In the later stages of training, a many-to-many SoftClip consistency mechanism is introduced to replace the bidirectional Mixup contrastive loss function to avoid overfitting noise, thereby obtaining optimized temporal feature embeddings and graph structure embeddings; Step S435: The optimized temporal feature embedding and graph structure embedding are concatenated in the fusion layer of the temporal convolution-graph convolution fusion model to obtain fused state data.

Images