A fire-fighting equipment fault prediction method based on deep learning

Abstract

The application relates to the technical field of fire safety monitoring, and discloses a fire-fighting equipment fault prediction method based on deep learning. The method collects multi-source operation data of a fire-fighting system and an alarm / interlocking system, constructs an equipment topology graph, and establishes a dual-channel neural network model integrating time series convolution and graph convolution. The adaptability and information propagation efficiency of the topology structure are enhanced through spectral gap optimization and a greedy topology reconnection mechanism, and a multi-modal mixed contrast consistency fusion mechanism is introduced to realize flexible alignment and stable fusion of time series features and structure features. The method can identify potential faults of fire-fighting equipment in advance under complex working conditions, realize accurate prediction and graded alarm, improve the reliability, robustness and intelligent operation and maintenance capability of the system, and is suitable for operation safety monitoring of building fire-fighting systems and industrial fire-fighting equipment.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of fire safety monitoring, and particularly relates to a fire-fighting equipment fault prediction method based on deep learning. BACKGROUND

[0002] With the increasing complexity of fire-fighting systems of urban public buildings and industrial facilities, the operation safety and stability of fire-fighting equipment have become key factors to ensure the reliability of fire-fighting systems. Core equipment such as fire pumps, valves, sprinkler / hydrant circuits, detectors, and controllers are easily subject to performance degradation, parameter drift, or potential failure due to factors such as pressure fluctuations, electrical shocks, mechanical vibrations, and changes in environmental temperature and humidity during long-term operation. If the faults are not identified and addressed in the early stage, it may lead to a decrease in fire extinguishing capacity, a delay in linkage response, and even a failure of the overall system function, resulting in serious safety risks. Although existing intelligent monitoring systems have achieved data collection and state identification to some extent, they are mostly limited to abnormal detection based on single-point parameter threshold or shallow learning algorithms, and cannot fully reflect the dynamic evolution characteristics of the equipment operation state. Some systems use neural networks or time series models for trend prediction, but they lack correlation modeling of multi-source data and do not fully consider the complex physical connections and control logic within the fire-fighting system, making it difficult to identify potential bottleneck nodes and associated fault propagation paths in the topology structure. In addition, existing models have weak robustness and interpretability under complex working conditions, often resulting in false positives or false negatives of abnormal signals. SUMMARY

[0003] The present application proposes a fire-fighting equipment fault prediction method based on deep learning, aiming to solve the shortcomings of existing intelligent systems in multi-source data fusion, structure-dependent modeling, and prediction stability. The method realizes global modeling and dynamic prediction of the operation state of fire-fighting equipment by introducing a dual-channel neural network structure that integrates time series dynamic features and device topology features. Unlike traditional shallow learning or single-modal models, the present application constructs a spectral gap-oriented graph neural network on the complex physical and logical structure of the fire-fighting system, dynamically adjusts the connectivity and information propagation path of the graph structure through spectral optimization and greedy topology reconnection mechanisms, significantly enhancing the model's ability to identify structural bottlenecks and abnormal nodes. At the same time, a multi-modal mixed contrast consistency fusion mechanism is proposed to construct a cross-modal flexible alignment process with bidirectional Mixup contrast loss and many-to-many SoftClip consistency constraints, which not only improves the semantic consistency of time series and structural features, but also suppresses overfitting under noise interference. Through this innovative system, the model can adaptively enhance features and optimize spectral features while maintaining the safety constraints of the fire-fighting system, thereby identifying potential fault risks in advance and outputting stable and interpretable graded alarm results, providing high-precision prediction support for intelligent operation and risk decision-making of fire-fighting equipment.

[0004] The application provides a fire-fighting equipment fault prediction method based on deep learning, which is applied to a fire-fighting equipment fault prediction system, and the fire-fighting equipment fault prediction system specifically comprises a central monitoring server, an alarm controller, and fire-fighting equipment and an edge server arranged in a pump room; the method comprises the following steps:

[0005] Step S1: collecting multi-source operation data of the fire-fighting equipment and aligning the multi-source data according to timestamps to obtain aligned multi-source data;

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

[0007] Step S3: constructing a device topology graph through a topology modeling engine built in the central monitoring server;

[0008] Step S4: deploying a time series convolution-graph convolution fusion prediction model in the central monitoring server; inputting the windowed time series data and the device topology graph into the time series convolution-graph convolution fusion model; and obtaining a target device fault probability through prediction;

[0009] Step S5: based on the target device fault probability, outputting a hierarchical alarm signal to the corresponding fire-fighting equipment through the alarm controller, and automatically triggering device maintenance, reset and standby device linkage operation according to the alarm level, while generating a disposal suggestion and uploading it to a fire-fighting monitoring center, so as to realize predictive maintenance and closed-loop management of the fire-fighting equipment.

[0010] Further, step S4 specifically comprises the following steps:

[0011] Step S41: inputting the windowed time series data into a time series encoding branch; the time series encoding branch performs hierarchical convolution and dynamic gating operation on the windowed time series data, extracts short-term fluctuation features and long-term evolution trend of the device operation state, and obtains time-dependent feature data;

[0012] Step S42: inputting the device topology graph into a graph convolution branch; the graph convolution branch processes the device topology graph through a spectrum gap-oriented graph neural network model to form structure context feature data containing spatial correlation and upstream and downstream dependence; the construction process of the spectrum gap-oriented graph neural network model is as follows: establishing a graph neural network model, introducing a spectrum gap optimization and a greedy topology reconnection mechanism to optimize the structure adaptability and spectrum feature expression capability of the graph neural network model, and obtaining the spectrum gap-oriented graph neural network model;

[0013] Step S43: adopting a multi-modal mixed contrast consistency fusion mechanism to uniformly integrate the time-dependent feature data and the structure context feature data in a fusion layer of the time series convolution-graph convolution fusion model to obtain fusion state data;

[0014] Step S44: input the fusion state data into the prediction output layer of the time series convolution-graph convolution fusion model, perform full connection mapping and normalization processing, and generate the target device fault probability.

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

[0016] Step S421: according to the node set, edge set and node feature matrix of the device topology graph, construct an adjacency matrix and a degree matrix to form a symmetric normalized Laplacian matrix; classify and label the edges in the device topology graph to form an engineering safety constraint set; the engineering safety constraint set is divided into three categories according to the safety priority: forbidden modification class edge, weak adjustment class edge and logic addition class edge;

[0017] Step S422: perform eigenvalue decomposition on the symmetric normalized Laplacian matrix to obtain eigenvalues and eigenvectors, wherein the smallest eigenvalue is zero, and the first non-zero eigenvalue is taken as the spectral gap; according to the eigenvalue distribution and the gradient change of the eigenvector, the spectral gap analysis result is obtained; on this basis, the region where the gradient of the eigenvector suddenly changes and the local spectral gap significantly narrows is located, and the structural bottleneck region is marked and located;

[0018] Step S423: under the premise of meeting the engineering safety constraint set, according to the spectral gap analysis result and the structural bottleneck region, generate two types of candidate edge sets: deletion candidate set and addition candidate set;

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

[0020] Step S425: introduce Eldan criterion as a stopping condition to dynamically constrain and terminate the spectral optimization process to prevent local overfitting caused by greedy topology reconnection optimization, and output the reconnected topology structure and updated spectral feature result;

[0021] Step S426: perform upstream and downstream control logic consistency checking and topology connectivity checking on the reconnected topology structure and updated spectral feature result to verify the consistency and effectiveness of the reconnected structure at the physical constraint and control logic level; when the checking passes, input the checked and qualified topology structure and spectral feature to the graph convolution part of the graph neural network model to generate the structural context feature data.

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

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

[0025] Step S433: construct a bidirectional Mixup contrast loss function, simultaneously perform forward and reverse alignment between the two modalities in the cross-sample mixed sample set through the bidirectional Mixup contrast loss function, capture the shared relationship between samples in the early training stage, enhance the robustness of feature fusion, and form high-consistency fusion data;

[0026] Step S434: in the middle and late training stage, introduce a many-to-many SoftClip consistency mechanism to replace the bidirectional Mixup contrast loss function to avoid overfitting noise, and obtain optimized time feature embedding and graph structure embedding;

[0027] Step S435: splice the optimized time feature embedding and graph structure embedding in the fusion layer of the time convolution-graph convolution fusion model to obtain fusion state data.

[0028] The beneficial effects achieved by the above scheme are as follows:

[0029] The present application realizes global modeling and fine fault prediction of the running state of the fire-fighting equipment by constructing a deep prediction model that fuses time-dependent dynamic features and topological structure features, realizes the transition from single signal detection to multi-source and 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-fighting system, and improve the understanding ability of the model to the system structure dependence and energy transmission law. This mechanism realizes efficient propagation of information in multi-level control chains such as fire-fighting pump-valve-sprinkler branches, solves the problem that traditional models cannot reflect the upstream and downstream influence between devices, enhances the structural consistency and engineering interpretability of the fault prediction results, and significantly improves the reliability and accuracy of the fire-fighting system running state monitoring.

[0030] The application realizes dynamic alignment of time sequence signals and topological features in semantic space by introducing a multi-modal mixed contrast consistency fusion mechanism, and improves the robustness and expression accuracy of cross-modal feature fusion. By using a bidirectional Mixup contrast loss function in the early training stage, the model can learn a stable shared feature distribution from the multi-dimensional operation samples of the fire-fighting equipment, and the generalization ability of the model under data imbalance and complex working conditions is enhanced. In the middle and late training stage, a many-to-many SoftClip consistency mechanism is used to further suppress feature drift caused by noise fluctuations and solve the problem of difficulty in aligning time sequence features and structure features in complex environments. This mechanism enables the application to still output reliable fault probability when facing typical fire-fighting scenes such as sudden flow changes, motor overload, vibration abnormalities or multi-point linkage interference, thereby enhancing the adaptability and prediction robustness of the system to non-stationary signals.

[0031] In addition, the application realizes adaptive topology reconnection under the premise of maintaining fire safety logical constraints by introducing a spectrum gap gain maximization and engineering safety constraint set collaborative mechanism in graph structure optimization. This mechanism effectively avoids interference of topology reconstruction with safety paths such as main water supply trunk lines and control closed loops, ensuring that the structure optimization process of the model conforms to fire regulations and physical constraints; meanwhile, the spectrum gap proxy approximation formula is used to realize fast evaluation and greedy optimization, thereby significantly improving the convergence efficiency and real-time prediction performance of model training. Through this mechanism, the model not only has the ability to predict future fault risks, but also can dynamically identify weak connection points in the topology and make inference corrections, thereby improving the overall stability and operation and maintenance guidance value of the system, and realizing the intelligent evolution of fire-fighting equipment from passive response to active defense, thereby providing strong technical support for safety decision-making of the fire-fighting system. BRIEF DESCRIPTION OF DRAWINGS

[0032] Figure 1 A flowchart of a fire-fighting equipment fault prediction method based on deep learning is provided for the application.

[0033] Figure 2 A loss component evolution curve graph is provided for Example Five. DETAILED DESCRIPTION

[0034] The technical solutions in the embodiments of the application will be described clearly and completely below with reference to the drawings in the embodiments of the application. Obviously, the described embodiments are only part of the embodiments of the application, rather than all the embodiments of the application. Based on the embodiments in the application, all other embodiments obtained by those skilled in the art without creative labor fall within the scope of protection of the application.

[0035] Example One, according to Figure 1The application provides a fire-fighting equipment fault prediction method based on deep learning, which is applied to a fire-fighting equipment fault prediction system and specifically comprises fire-fighting equipment, an edge server in a pump room, a central monitoring server and an alarm controller.

[0036] Step S1: data acquisition and alignment: multi-source operation data of the fire-fighting equipment in the fire-fighting system and the alarm / interlocking system are collected through a data collection layer, and the multi-source data after alignment is obtained by time stamp alignment;

[0037] The fire-fighting equipment comprises a fire-fighting main pump, a pressure stabilizing pump, an electromagnetic valve, a spraying / fire hydrant circuit, an alarm controller, a gas fire extinguishing device and an environmental monitoring unit; the multi-source operation data comprises pipe network pressure and flow, motor current and voltage, pump rotating speed and start-stop state, valve opening degree, environmental temperature and humidity, vibration and acoustic signals, alarm / feedback signals of a smoke temperature detector and the gas fire extinguishing device, controller event logs and maintenance and repair records. The data collection layer is installed at key positions of the fire-fighting system and is used for realizing real-time sensing of the operation state of the fire-fighting equipment, environmental parameters and control signals, comprising: 1. a pipe network monitoring unit: a set of pressure sensors (0-1.6 MPa range) and flow sensors (electromagnetic type) are installed at the pump outlet, the main pipe and the branch node, and are used for collecting pipe network pressure, flow and transient fluctuation data. 2. A pump set monitoring unit: three-phase current transformers and voltage detection modules are installed at the motor end of the fire-fighting main pump and the pressure stabilizing pump, and are used for monitoring the electrical load and power supply state; three-axis MEMS vibration sensors are pasted at the pump bearing and the base, and are used for collecting vibration acceleration and frequency spectrum characteristics; acoustic pickup sensors (microphone arrays) are arranged on the surface of the motor cover or the pump body, and are used for monitoring abnormal noise characteristics; temperature and humidity sensors are installed on the top of the pump room and in the control cabinet, and are used for recording environmental temperature and humidity changes. 3. An alarm and control unit: alarm, feedback and event log signals of the smoke temperature detector, the gas fire extinguishing device and the alarm / interlocking controller are collected, and are used for recording equipment state changes and control logic responses. Analog signals output by each monitoring unit are converted into digital signals and are attached with time stamps, are gathered to the edge computing node through an RS485 bus and are transmitted to the edge server in the pump room.

[0038] Step S2: preprocessing and windowing: the data collection layer is communicatively connected with the edge server in the pump room through a CAN bus, the aligned multi-source data is subjected to denoising, missing value filling, abnormal value suppression and normalization processing in the edge server, and the sequence is cut into sample segments according to a sliding window, and windowed time series data are obtained;

[0039] Step S3: Device topology modeling: the central monitoring server is arranged in the machine room control room, the central monitoring server is in communication connection with the edge server in the pump room; a topology modeling engine is built in the central monitoring server, a device topology graph is constructed based on the physical connection relationship and the logical control relationship of the fire fighting system and the alarm / linkage system in the topology modeling engine, the device topology graph includes a node set, an edge set and a node characteristic matrix; wherein the node set represents the water pump, the valve, the spray / hydrant circuit, the detector and the controller, the edge set represents the pipeline communication, the circuit membership and the control / linkage relationship, and the node characteristic matrix records the operating parameters and the state characteristics of each node;

[0040] Step S4: Deep prediction model inference: a time series convolution-graph convolution fusion prediction model is deployed in the central monitoring server; the windowed time series data and the device topology graph are input into the time series convolution-graph convolution fusion model; the windowed time series data is extracted through the time series encoding branch to extract the time dependent features, the device topology graph is extracted through the graph convolution branch to extract the structural context features, the features of the two branches are fused, and the target device failure probability is predicted;

[0041] Step S5: Alarm and disposal suggestion: the central monitoring server interfaces with the alarm controller, sets the device category and the working condition adaptive threshold, compares the target device failure probability with the device category and the working condition adaptive threshold, outputs a hierarchical alarm signal to the corresponding fire fighting device through the alarm controller, automatically triggers the device repair, reset and standby device linkage operation according to the alarm level, generates a disposal suggestion and uploads it to the fire fighting monitoring center, realizes the predictive maintenance and closed-loop management of the fire fighting device.

[0042] Embodiment two, based on embodiment one, in this embodiment, step S4 specifically includes the following steps:

[0043] Step S41: input the windowed time series data into the time series encoding branch; the time series encoding branch performs hierarchical convolution and dynamic gating operation on the windowed time series data, extracts the short-term fluctuation features and the long-term evolution trend of the device operating state, and obtains 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 spectrum gap guided graph neural network model, forms structural context feature data containing spatial correlation and upstream / downstream dependence; the construction process of the spectrum gap guided graph neural network model: establish a graph neural network model, introduce a spectrum gap optimization and a greedy topology reconnection mechanism to optimize the structure adaptability and the spectrum feature expression ability of the graph neural network model, and obtain the spectrum gap guided graph neural network model;

[0045] Step S43: The time-dependent feature data and the structural context feature data are unified and integrated in the fusion layer of the time convolution-graph convolution fusion model by using a multi-modal mixed contrast consistency fusion mechanism to obtain fusion state data; the multi-modal mixed contrast consistency fusion mechanism is constructed in the following manner: a multi-stage optimization strategy based on mixed sample contrast and consistency constraint is introduced, bidirectional mixed sample contrast learning is performed in the early training stage to establish shared semantic relationship and feature mapping consistency between different modalities; and a multi-to-multi SoftClip consistency constraint mechanism is introduced in the middle and late training stages to calculate a soft weight distribution within the modalities and perform weighted consistency alignment between cross modalities, thereby forming the multi-modal mixed contrast consistency fusion mechanism.

[0046] Step S44: The fusion state data is input into the prediction output layer of the time convolution-graph convolution fusion model, and the target device failure probability is generated after full connection mapping and normalization processing.

[0047] Embodiment Three, based on Embodiment One, in this embodiment, step S4 specifically includes the following steps:

[0048] Step E1: The windowed time series data is input into the time encoding branch; the time encoding branch performs hierarchical convolution and dynamic gating operation on the windowed time series data to extract short-term fluctuation features and long-term evolution trend of the device running state, thereby obtaining time-dependent feature data;

[0049] Step E2: The device topology graph is input into the graph convolution branch, and the graph convolution branch processes the device topology graph through a graph neural network model to form structural context feature data containing spatial correlation and upstream and downstream dependence;

[0050] Step E3: The time-dependent feature data and the structural context feature data are unified and integrated in the fusion layer by using a multi-modal mixed contrast consistency fusion mechanism to obtain fusion state data;

[0051] Step E4: The fusion state data is input into the prediction output layer, and the target device failure probability is generated after full connection mapping and normalization processing.

[0052] Embodiment Four, based on Embodiment Two, in this embodiment, step S42 specifically includes the following steps:

[0053] Step S421: According to the node set, edge set and node feature matrix of the device topology graph, an adjacency matrix and a degree matrix are constructed to form a symmetric normalized Laplacian matrix; the edges (connection relationships) in the device topology graph are classified and labeled to form an engineering safety constraint set, which is used to provide safety boundary conditions in subsequent spectral optimization and edge reconnection processes to ensure that the functional integrity, signal transmission stability and compliance of the fire protection system are not damaged by structural adjustment; the engineering safety constraint set is divided into three categories according to the safety priority: prohibited modification class edge (not deletable, not weakenable), adjustable class edge (adjustable edge) and logical addition class edge (candidate addition edge);

[0054] Prohibited modification class edge (not deletable, not weakenable): including fire main water supply trunk, main water supply path from fire pump to main pipe, forced linkage loop (such as control closed loop between pump-valve-detector), mandatory inspection path and signal main trunk line directly connected with central alarm control unit; this kind of edge bears the core transportation, control and monitoring functions in the system, and if it is deleted or weakened, it will lead to functional failure or safety risk, so it is defined as a strong constraint edge in the spectral gap optimization and graph reconnection process, and only allowed to participate in feature propagation calculation without adjusting edge weight or topology.

[0055] Adjustable class edge (adjustable edge): refers to ordinary running connection except prohibited modification class edge, including auxiliary loop between same type functional units, redundant control path, connection between non-critical measuring points; this kind of edge allows dynamic weighting, edge weight attenuation or weakening operation according to edge weight importance, self-correlation degree or bottleneck influence factor in subsequent spectral optimization stage, for optimizing information flow path and relieving excessive aggregation effect.

[0056] Logical addition class edge (candidate addition edge): refers to the virtual edge established between node pairs within the allowable range of physical structure and compatible in function logic; this kind of edge is used to enhance the connectivity of topology and the width of spectral gap, and improve the effective coverage of information propagation when the topology is sparse or there is a structural bottleneck; the generation of the new edge is based on node attribute similarity, functional domain affiliation and spatial correlation weight, and must meet the feasibility constraints of fire safety logic (such as electrical isolation, loop tolerance, signal safety threshold).

[0057] Step S422: eigenvalues and eigenvectors are obtained by eigenvalue decomposition of the symmetric normalized Laplacian matrix, wherein the smallest eigenvalue is zero, and the first non-zero eigenvalue is taken as the spectral gap; spectral gap analysis results are obtained according to the eigenvalue distribution and eigenvector gradient change; on this basis, the region where the eigenvector gradient suddenly changes and the local spectral gap significantly narrows is located, and the structural bottleneck region is marked and located; the spectral gap analysis results are used to represent the connectivity and information flow capacity of the current topology structure; the smaller the spectral gap, the more communication blockage or energy transmission bottleneck exists in the region, which is the focus of subsequent structural optimization.

[0058] Step S423: Under the premise of meeting the set of engineering safety constraints, according to the spectral gap analysis result and the structure bottleneck region identification, two types of candidate edge set are generated: a deletion candidate set and an added candidate set; wherein the deletion candidate set is selected only from the "weakening adjustment type edge", and the added candidate set is constructed only in the range of "logical added type edge", so as 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 types of candidate edge set, the spectral gap proxy approximation formula is used to calculate the spectral gap approximation change (i.e. proxy gain) one by one, and the greedy topology reconnection optimization is performed according to the spectral gap gain maximization principle;

[0060] The spectral gap proxy approximation formula is constructed based on the matrix perturbation theory, which is used to quickly evaluate the influence of the addition and deletion of a single edge on the overall spectral structure, and is specifically expressed as follows:

[0061] ;

[0062] Wherein, represents the target eigenvalue of the symmetric normalized Laplacian matrix before modifying the edge, represents the new eigenvalue approximation after flipping a candidate edge (addition / deletion); 、 represents the components of the eigenvector on node and node ; represents the edge operation identifier;

[0063] According to the above formula, the spectral gap approximation change (i.e. proxy gain) of each candidate edge is calculated, and the greedy reconnection optimization is performed according to the spectral gap gain maximization principle: when the candidate edge belongs to the deletion candidate set and the proxy gain is positive, and the engineering safety constraints are met, the edge deletion operation is performed; when the candidate edge belongs to the added candidate set and the proxy 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 power iteration to update the main eigenvector and the spectral gap, realizing local fast convergence and entering the next iteration; this step effectively alleviates the problems of over-smoothing and information compression of the topology structure in the premise of ensuring real-time performance and scalability through "a small number of edge flips and significant improvement of spectral gap";

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

[0070] Step S432: randomly select an index in the unified sample set, and perform a convex combination operation on the two modal features 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 association relationship between modalities in the hidden space, and providing a mixed sample basis for subsequent contrastive fusion learning;

[0071] Step S433: construct a bidirectional Mixup contrastive loss function, and simultaneously perform forward and reverse alignment between the two modalities in the cross-sample mixed sample set through the bidirectional Mixup contrastive loss function, to capture the shared relationship between samples, enhance the robustness of feature fusion, and form high-consistency fusion data in the early stage of training;

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

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

[0074] ;

[0075] Wherein, represents the Mixup contrastive loss of the first modal direction, with "modality 1 (temporal feature embedding)" as the anchor point, and "modality 2 structural feature embedding" as the sample to be aligned; represents the number of multi-modal sample pairs in a training batch; represents the index of the current anchor sample, represents another sample index of the same modality (modality 1) that is mixed with the sample ; represents the Mixup mixing coefficient of the sample ; represents the mixed sample embedding of modality 1 (first Mixup, then input into the modality 1 encoder to obtain), represents the normalized embedding set of modality 2 in the current batch; represents a unidirectional multi-modal contrast (similarity) loss;

[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 time sequence feature embedding is spliced with the graph structure embedding at a fusion layer of the time sequence convolution-graph convolution fusion model to obtain fusion state data.

[0095] In this embodiment based on the fourth embodiment, step S43 specifically includes the following steps:

[0096] Step R1: The time-dependent feature data and the structural context feature data are respectively input into the modal encoders of the time sequence convolution-graph convolution fusion model to obtain two modal features through mapping; the time-dependent feature is mapped into a first modal latent representation vector through a time sequence feature encoder, and the structural context feature is mapped into a second modal latent representation vector through a graph structure encoder; the two types of latent representations are unified to the same hidden space dimension through a learnable linear mapping layer of the time sequence convolution-graph convolution fusion model, and L2 normalization processing is performed to obtain a unified sample set;

[0097] Step R2: An index is randomly selected in the unified sample set, and convex combination operations are respectively performed on the two modal features to obtain a cross-sample mixed sample set; the convex combination operation simulates the latent correlation relationship between the modalities in the hidden space by setting a mixing factor to perform linear interpolation between the samples, thereby providing a mixed sample basis for subsequent contrastive fusion learning;

[0098] Step R3: An InfoNCE-based contrastive loss function is constructed, and the feature alignment and consistency enhancement between the cross-modal samples are realized by maximizing the similarity of positive samples and minimizing the similarity of negative samples in the cross-sample mixed sample set, to generate high-consistency fusion data;

[0099] Step R4: The high-consistency fusion data is spliced at the fusion layer to obtain fusion state data.

[0100] In this embodiment based on the fifth embodiment, in this embodiment:

[0101] Step S4: The windowed time sequence data and the device topology graph are input into the time sequence convolution-graph convolution fusion model; the windowed time sequence data is extracted through a time sequence encoding branch to extract time-dependent features, and the device topology graph is extracted through a graph convolution branch to extract structural context features; the features of the two branches are fused to predict the target device failure probability;

[0102] Step S5: The device category and working condition adaptive threshold are set, the target device failure probability is compared with the device category and working condition adaptive threshold, a graded alarm is output, and disposal suggestions for maintenance, reset, and linkage verification are generated.

[0103] In this embodiment:

[0104] Initial normalized laplace gap: 0.5657;

[0105] Post-reconnection laplace gap: 0.6667;

[0106] Target device failure probability:

[0107] Fusion failure probability field: fault_prob (0~1);

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

[0109] Hierarchical alarm: alarm_level (normal / level-I-low / level-II-medium / level-III-high);

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

[0112] fault_prob = 0.983, adaptive_th = 0.670 → level-III-high;

[0113] Indicators: pressure 6.813 bar (significant drop), flow 42.99 L / s (uprush), current 157.8 A (overcurrent), vibration 4.10 mm / s, noise 69.1 dB, controller event 2, smoke_alarm = 1; 09:20:

[0115] fault_prob = 0.983, adaptive_th = 0.655 → level-III-high;

[0116] Indicators: pressure 7.220 bar, flow 44.58 L / s, current 164.7 A, vibration 4.32 mm / s, noise 69.6 dB, controller event 2, smoke_alarm = 1; 09:18:

[0118] fault_prob = 0.979, adaptive_th = 0.665 → level-III-high;

[0119] Indicators: pressure 6.894 bar, flow 44.20 L / s, current 155.9 A, vibration 3.99 mm / s, noise 69.8 dB, controller event 2, smoke_alarm = 1;

[0120] The resonant combination of "low pressure + high flow + overcurrent + vibration / noise lift + short-term smoke temperature alarm" from 09:18-09:20 matches the actual signs of pump operating condition abnormalities / local problems in the pipe network, so the failure probability is quickly pulled up, exceeding the adaptive threshold and triggering level-III-high.

[0121] Alarm and treatment suggestion:

[0122] Threshold strategy: pump equipment basic threshold 0.62, combined with current / voltage / vibration fluctuation +0~0.06 dynamic correction;

[0123] Suggestion:

[0124] Immediate repair: check the pump group bearing and coupling (for vibration and current rise);

[0125] Reset / verification: 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 is a leakage / local blockage or valve jamming sign.

[0127] The above describes the present application and its embodiments, which is not restrictive, and the drawings shown are only one of the embodiments of the present application, and the actual structure is not limited thereto; in general, if a person skilled in the art is inspired thereby, without departing from the purpose of the present application, without creative design, similar structure and embodiments to the technical solution, which should belong to the protection scope of the present application.

Claims

1. A deep learning-based fire-fighting equipment failure prediction method, the method being applied to a fire-fighting equipment failure prediction system, the fire-fighting equipment failure prediction system specifically comprising a central monitoring server, an alarm controller, and fire-fighting equipment and an edge server arranged in a pump room, characterized in that: The method comprises: Step S1: Collecting multi-source operation data of the fire-fighting equipment and aligning the multi-source data by timestamps to obtain aligned multi-source data; Step S2: Preprocessing the aligned multi-source data in the edge server to obtain windowed time series data; Step S3: Constructing a device topology graph through a topology modeling engine built in the central monitoring server; Step S4: Deploying a time series convolution-graph convolution fusion prediction model in the central monitoring server; inputting the windowed time series data and the device topology graph into the time series convolution-graph convolution fusion model; and predicting a target device failure probability; Step S5: Based on the target device failure probability, outputting a hierarchical alarm signal to the corresponding fire-fighting equipment through an alarm controller, and automatically triggering device maintenance, reset and standby device linkage operation according to the alarm level, while generating a disposal suggestion and uploading it to a fire-fighting monitoring center, to realize predictive maintenance and closed-loop management of the fire-fighting equipment; The construction method of the time series convolution-graph convolution fusion model is: a time series encoding branch based on a gated time series convolution 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 double-channel neural network structure, thereby obtaining the time series convolution-graph convolution fusion model; The spectral gap guided graph neural network model is used to process the device topology graph to form structural context feature data; the construction process of the spectral gap guided graph neural network model is: a graph neural network model is established, and a spectral gap optimization and a greedy topology reconnection mechanism are introduced to optimize the structural adaptability and spectral feature expression capability of the graph neural network model, thereby obtaining the spectral gap guided graph neural network model; The specific process of processing the device topology graph through the spectral gap guided graph neural network model is as follows: According to the node set, the edge set and the node feature matrix of the device topology graph, an adjacency matrix and a degree matrix are constructed to form a symmetric normalized Laplacian matrix; the edges in the device topology graph are classified and labeled to form an engineering safety constraint set; The symmetric normalized Laplacian matrix is subjected to eigenvalue decomposition to obtain eigenvalues and eigenvectors, and the first non-zero eigenvalue is taken as a spectral gap; according to the eigenvalue distribution and the gradient change of the eigenvectors, a spectral gap analysis result is obtained, and a structural bottleneck region identifier is marked; On the premise of meeting the engineering safety constraint set, according to the spectral gap analysis result and the structural bottleneck region identifier, two types of candidate edge sets, a deletion candidate set and an addition candidate set, are generated; For each candidate edge in the two types of candidate edge sets, the spectral gap proxy approximation formula is used to calculate the approximate change amount of the spectral gap one by one, and the spectral gap gain maximization principle is used to perform greedy topology reconnection optimization, and the formula used is as follows: ； wherein, denotes the target eigenvalue of the symmetric normalized Laplacian before modifying an edge, denotes the new eigenvalue approximation after flipping one candidate edge; 、 denotes the eigenvector at node , the component at node ; denotes the edge operation identifier; The Eldan criterion is introduced as a stopping condition to dynamically constrain and terminate the spectral optimization process, prevent local overfitting caused by the greedy topology reconnection optimization, and output the reconnected topology structure and the updated spectral feature result; Specifically, when the discriminant function of the candidate edge satisfies the sufficient condition of "edge deletion can ensure gap increase", the corresponding edge deletion operation is allowed to be performed; when the predicted new operation will lead to the decrease of spectral gap, the new operation is rejected, thereby ensuring that the structure adjustment is only along the direction of spectral gap gain, avoiding the generation of 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 continuous several rounds of iterations is lower than the preset threshold, or the spectral gap has reached the preset upper limit, the stop condition is automatically triggered to terminate the topology reconnection process; after meeting the stop condition, the system outputs the reconnected topology structure and the updated spectral characteristic results as the input of subsequent graph convolution calculation and device state reasoning, providing a high connectivity and stable structure basis for state evaluation and prediction of the fire fighting system; The reconnected topology structure and the updated spectral characteristic results are subjected to consistency check of upstream and downstream control logic and topology connectivity check; when the check passes, the topology structure and the spectral characteristics that pass the check are input into the graph convolution part of the graph neural network model to obtain structure context feature data.

2. The fire-fighting equipment failure prediction method based on deep learning according to claim 1, characterized in that: Step S4 specifically includes the following steps: Step S41: input the windowed time series data into the time series encoding branch to obtain time-dependent feature data; Step S42: input the device topology graph into the graph convolution branch, and the graph convolution branch processes the device topology graph through the spectral gap guided graph neural network model to form structure context feature data; Step S43: adopt a multi-modal mixed contrast consistency fusion mechanism to uniformly integrate the time-dependent feature data and the structure context feature data in the fusion layer of the time series convolution-graph convolution fusion model to obtain fusion state data; Step S44: input the fusion state data into the prediction output layer of the time series convolution-graph convolution fusion model to obtain the target device fault probability.

3. The fire-fighting equipment failure prediction method based on deep learning according to claim 1, characterized in that: The engineering safety constraint set is divided into three categories according to safety priority: prohibited modification class edge, adjustable adjustment class edge and logical new class edge.

4. The fire-fighting equipment failure prediction method based on deep learning according to claim 2, characterized in that: Step S43 specifically includes the following steps: Step S431: input the time-dependent feature data and the structure context feature data into the modal encoder of the time series convolution-graph convolution fusion model respectively, map to obtain two modal features, and perform linear mapping layer processing to obtain a uniform sample set; Step S432: randomly select an index in the uniform sample set, and perform convex combination operation on the two modal features to obtain a cross-sample mixed sample set; Step S433: construct a bidirectional Mixup contrast loss function, which simultaneously performs forward and reverse alignment between the two modalities in the cross-sample mixed sample set through the bidirectional Mixup contrast loss function to capture the shared relationship between samples in the early training stage; Step S434: in the middle and late training stage, introduce a many-to-many SoftClip consistency mechanism to replace the bidirectional Mixup contrast loss function to avoid overfitting noise, and obtain optimized time series feature embedding and graph structure embedding; Step S435: splice the optimized time series feature embedding and graph structure embedding in the fusion layer of the time series convolution-graph convolution fusion model to obtain fusion state data.

Images