Fault detection method based on fast-assembly type prefabricated footpath

Abstract

The invention discloses a fault detection method based on a fast-assembly type prefabricated footpath, and relates to the technical field of equipment fault diagnosis. According to the method, a sensor network deployed on each module of a prefabricated footpath is utilized to collect vibration signals; a graph structure is dynamically constructed based on the real-time similarity of the signals, nodes represent modules, and edges represent the function connection relation between the nodes; and inputting the dynamic graph sequence into the space-time graph convolutional network model for processing. The model captures the propagation and distribution of a fault in a spatial dimension through a graph attention network layer, and achieves the preliminary positioning of a fault region. Analyzing spatio-temporal characteristics through a time convolutional network layer, and identifying a fault type; and finally, accurate positioning and classification of faults are realized through spatio-temporal feature fusion. According to the method, the problems that a traditional method is weak in spatial perception capability and poor in model generalization are solved, the method is particularly suitable for a modularized and extensible fast-assembly type prefabricated footpath system, and early warning, efficient diagnosis and accurate positioning are achieved.

Description

Technical Field

[0001] This invention belongs to the field of equipment fault diagnosis technology, and in particular, a fault detection method based on quick-installation prefabricated walkways. Background Technology

[0002] To meet the demands of rapid construction and modular building in modern architecture, prefabricated walkways have emerged. These walkways are prefabricated in factories using standardized modular units, requiring only rapid assembly on-site, greatly improving construction efficiency. The accompanying prefabricated freight transport channels also employ the same modular design, installed parallel to the walkways for transporting goods.

[0003] Traditional fault diagnosis methods primarily rely on installing a limited number of vibration sensors on key components (such as the drive unit and critical truss points) and analyzing the time-domain or frequency-domain characteristics of individual sensor signals (such as amplitude exceeding thresholds or the presence of abnormal frequency components in the spectrum) to determine the condition. However, prefabricated walkways are flexible, expandable, and modular systems, making them a natural platform for deploying high-density, standardized sensor networks. Each module can be integrated into an independent sensing unit, forming a naturally distributed monitoring network. Therefore, using traditional fault diagnosis methods for prefabricated walkways has significant limitations. 1. Lack of spatial awareness: The walkway is a continuous rigid metal structure, and vibrations caused by local faults will propagate along the structure. Existing methods analyze sensor data in isolation, failing to utilize the propagation and attenuation patterns of vibration signals in space, leading to difficulties in fault location and susceptibility to interference.

[0004] 2. Poor model generalization ability: Diagnostic models are mostly designed for specific models and fixed structures of walkways, making it difficult to adapt to modular, quickly installable prefabricated walkway systems. When the walkway length or sensor layout changes, the model needs to be readjusted, lacking flexibility.

[0005] 3. Insufficient dynamic relationship modeling: Although recent research has introduced graph neural networks into fault diagnosis and modeled sensor networks as graph structures, the graph connections are usually based on static presets of physical locations. This fails to capture the dynamic changes in the "functional connections" between sensors caused by faults during equipment operation, which contain rich fault characteristic information and are crucial for early warning and accurate location. Summary of the Invention

[0006] The purpose of this invention is to overcome the shortcomings of the prior art and propose a fault detection method based on quick-installation prefabricated walkways, which can achieve early warning, accurate location and efficient diagnosis of faults.

[0007] The technical problem solved by this invention is achieved through the following technical solution: A fault detection method based on quick-installation prefabricated walkways includes the following steps: Step 1: Obtain vibration signals collected by sensors deployed on each module of the quick-installation prefabricated walkway; Step 2: Calculate the signal similarity between corresponding nodes of each module, and dynamically generate an adjacency matrix based on an adaptive threshold, thereby constructing a dynamic graph structure with modules as nodes and functional connections as edges; Step 3: Input the dynamic graph structure into the pre-trained spatiotemporal graph convolutional network model ST-GCN; the spatiotemporal graph convolutional network model performs spatial convolution on the dynamic graph structure through the graph attention network GAT layer to generate new features containing spatial context information; and then performs temporal convolution on the new feature sequence through the temporal convolutional network layer TCN to extract spatiotemporal features; Step 4: Based on the spatiotemporal features output by the spatiotemporal graph convolutional network model, perform fault classification and fault location.

[0008] Moreover, the specific implementation method of step 1 is as follows: each module deployed on the quick-installation prefabricated walkway is defined as a node, the vibration signals collected by the sensors on each module are extracted, and a node feature matrix is ​​formed.

[0009] Furthermore, the specific implementation method of step 2 is as follows: combining the feature vectors of each module into a feature matrix. Calculate the real-time similarity of vibration signals between all node pairs and apply an adaptive threshold. Dynamically generate an adjacency matrix that reflects the functional connection relationships between nodes at the current time. , node feature matrix and adjacency matrix Combined, they form a dynamic graph structure for the current time window. .

[0010] Furthermore, the graph attention network layer performs spatial dimension modeling, assigning different attention weights to the neighbors of each central node: in, Here, the original attention coefficient represents the unnormalized importance of neighbor node j to center node i, where... and Let be the input feature vectors of the two nodes, and W be a trainable shared weight matrix used to perform linear transformation on the node features. The original attention coefficients are normalized using the softmax function, resulting in the following formula: in, The attention weights are normalized and their values ​​range from 0 to 1, representing the set of all neighbors of node j at node i. The relative importance of each node is given by k, which is the index of a neighboring node. The denominator is summed with respect to the exponential coefficients of all neighbors of node i to ensure that the sum of the weights is 1. ∥ represents the vector concatenation operation, a is a trainable attention vector that maps the concatenated features to scalar scores, and LeakyReLU introduces a non-linear activation function to update node features through a weighted aggregation mechanism. in, It is a non-linear activation function. For the new feature, the transformation feature of each neighbor node j According to its attention weight The weighted results are summed, and then passed through a nonlinear activation function. The process ultimately generates a new feature representation for node i that incorporates neighborhood context information. ; For the new features output by the GAT layer that already contain rich spatial context information New features of it across all time slices Stack them in chronological order to form a new feature sequence, which is then input into the temporal convolutional network layer for temporal dimension modeling.

[0011] Furthermore, the TCN employs one-dimensional dilated causal convolution to process the new feature sequence of each node after GAT layer enhancement. At each independent time slice, GAT is used to aggregate neighbor information and generate new node features rich in spatial context. For each node, then its new features across all time slices are calculated. Stacked chronologically, they form a time series. Ultimately, TCN receives this sequence of node features output by GAT, arranged chronologically. It no longer cares about the graph structure, but focuses on learning temporal dynamic patterns from this sequence.

[0012] in, This represents the convolution operation. f It is a length of K The convolution kernel function, whose parameters are determined by... k index, These are the weight parameters of the convolution kernel at position k; z It is the input time-series signal; t Represents the current time step in the calculation. d The expansion factor is used to obtain the spatiotemporal characteristics.

[0013] Moreover, the specific implementation method of step 4 is as follows: obtain the fault level according to the spatiotemporal characteristics and classify it: minor abnormalities are recorded in the system log; general faults are issued with early warning notifications; serious faults trigger emergency alarms and recommend shutdown for maintenance.

[0014] The advantages and positive effects of this invention are: 1. This invention adopts an intelligent diagnostic method based on dynamic graph neural network (ST-GCN), combined with modular sensor network and spatiotemporal signal analysis technology, which has many significant advantages and implementation effects.

[0015] 2. The dynamic graph neural network model of this invention achieves collaborative capture and joint analysis of the spatial and temporal features of fault signals through a serial fusion architecture of Graph Attention Network (GAT) and Temporal Convolutional Network (TCN). The GAT layer assigns differentiated weights to each sensor node through an attention mechanism, enabling precise perception of the spatial propagation path of the fault in the modular structure and achieving preliminary localization of the fault area. The TCN layer, through a dilated causal convolutional structure, effectively captures the temporal evolution of fault features and identifies typical fault modes such as periodic impacts and trend changes. This "space-first, time-follower" processing mechanism allows the system to accurately distinguish between local faults and global interference, significantly improving diagnostic accuracy. Through the dynamic graph structure construction mechanism, the system can adaptively expand and reconstruct the modular structure of the walkway. Within each time window, the system dynamically generates an adjacency matrix based on the real-time similarity of vibration signals between nodes, without the need for pre-defined fixed sensor network topology. This dynamic adaptability allows this invention to be flexibly applied to prefabricated walkway systems of different lengths and module configurations, greatly enhancing the versatility and scalability of the method.

[0016] 3. This invention fully mines the multi-dimensional spatiotemporal representation information of faults through attention-weighted aggregation of GAT and long-term dependency capture of TCN. Compared with traditional single-point sensing analysis methods, it can identify weak fault signals earlier and achieve early warning of faults. At the same time, based on the attention weight matrix generated by the GAT layer, the system can accurately locate the physical module where the fault occurs, providing clear guidance for subsequent maintenance.

[0017] 4. This invention can be directly deployed on existing modular freight channels without modifying the walkway itself, resulting in low implementation costs and easy integration. Through an end-to-end intelligent diagnostic process, it achieves fully automated processing from raw vibration signals to fault diagnosis decisions, significantly improving detection efficiency and providing a complete technical solution for the intelligent operation and maintenance of quick-installation prefabricated walkway systems.

[0018] 5. This invention achieves significant progress in the accuracy, early detection, positioning precision, and system adaptability of fault diagnosis through the innovative combination of dynamic graph neural network and spatiotemporal feature fusion. It has important engineering application value and broad development prospects in quick-installation prefabricated walkway systems. Attached Figure Description

[0019] Figure 1 This is a flowchart of the present invention; Figure 2 This is a flowchart of the GAT-TCN algorithm of the present invention. Detailed Implementation

[0020] The present invention will be further described in detail below with reference to the accompanying drawings.

[0021] A fault detection method based on quick-installation prefabricated walkways is proposed. This method utilizes a spatiotemporal graph convolutional network (ST-GCN) model, which achieves collaborative capture and joint analysis of spatial and temporal features in fault signals through deep fusion of graph attention networks (GAT) and temporal convolutional networks (TCN). This "space-time" sequential processing mode is the key innovation of this invention for achieving high-precision diagnosis and localization.

[0022] Traditional fault diagnosis methods either analyze individual sensor signals in isolation (ignoring spatial correlation) or simply stitch together data from multiple sensors (failing to effectively model complex spatial relationships). This invention recognizes that fault information in trail systems is not presented solely in the spatial or temporal dimensions, but simultaneously in both spatial distribution and temporal evolution. Spatial dimension: Vibrations generated by local faults propagate through the structure, forming a "spatial field" with a specific attenuation pattern on the sensor network. The signal similarity of nodes near the fault source changes. Temporal dimension: Fault development (e.g., crack propagation) or the characteristics of the fault itself (e.g., the periodic impact of bearing failure) manifests as specific patterns in the signal over time. Therefore, neither single-dimensional model can fully understand the fault. This invention combines the spatial and temporal dimensions: GAT (Gas-Atlas) interprets the "spatial distribution map" of the fault, while TCN (Tracking Networking) interprets the "temporal evolution history" of each point; only by combining these two methods can the complete picture of the fault be reconstructed.

[0023] like Figure 1 As shown, the present invention includes the following steps: Step 1: Obtain vibration signals collected by sensors deployed on each module of the quick-installation prefabricated walkway.

[0024] Each module deployed on the prefabricated walkway is defined as a node. Vibration signals collected by sensors on each module are extracted and a node feature matrix is ​​formed. Therefore, each node is an "intelligent module" capable of sensing its own vibration state.

[0025] Step 2: Calculate the signal similarity between corresponding nodes of each module, and dynamically generate an adjacency matrix based on an adaptive threshold, thereby constructing a dynamic graph structure with modules as nodes and functional connections as edges.

[0026] The feature vectors of each module are combined into a feature matrix. Calculate the real-time similarity of vibration signals between all node pairs and apply an adaptive threshold. Dynamically generate an adjacency matrix that reflects the functional connection relationships between nodes at the current time. , node feature matrix and adjacency matrix Combined, they form a dynamic graph structure for the current time window. .

[0027] Generate dynamic adjacency matrix The key lies in the adaptive threshold The determination of the threshold needs to be based on the current actual operating status of the system, dynamically selecting the most significant functional connections to avoid using fixed thresholds that may not adapt to changes in operating conditions. First, the optimal connection threshold is dynamically determined by analyzing the overall distribution characteristics of the similarity between nodes within each time window. This is done from the similarity matrix of the current time window. After extracting all off-diagonal elements to form a set U containing N(N-1) / 2 similarity values, the p-quantile of this set is calculated based on the preset quantile parameter p and used as the adaptive threshold for the current window. .

[0028] This design allows the system to automatically retain the most similar and valid connections, filtering out accidental weak correlations. Then, based on a threshold... The generated adjacency matrix The rules for the values ​​of its elements are as follows: This adaptive mechanism can accurately capture the essential changes in the functional connections between nodes under different operating conditions, providing a reliable topological foundation for subsequent graph neural network analysis.

[0029] Step 3, as follows Figure 2As shown, the dynamic graph structure is input into a pre-trained spatiotemporal graph convolutional network model ST-GCN; the spatiotemporal graph convolutional network model performs spatial convolution on the dynamic graph structure through the graph attention network GAT layer to generate new features containing spatial context information; then, the new feature sequence is temporally convolved through the temporal convolutional network layer TCN to extract spatiotemporal features.

[0030] The graph attention network layer models the spatial dimension, assigning different attention weights to the neighbors of each central node: in, Here, the original attention coefficient represents the unnormalized importance of neighbor node j to center node i, where... and Let be the input feature vectors of the two nodes, and W be a trainable shared weight matrix used to perform linear transformation on the node features. The original attention coefficients are normalized using the softmax function, resulting in the following formula: in, The attention weights are normalized and their values ​​range from 0 to 1, representing the set of all neighbors of node j at node i. The relative importance of each node is given by k, which is the index of a neighboring node. The denominator is summed with respect to the exponential coefficients of all neighbors of node i to ensure that the sum of the weights is 1. ∥ represents the vector concatenation operation, a is a trainable attention vector that maps the concatenated features to scalar scores, and LeakyReLU introduces a non-linear activation function to update node features through a weighted aggregation mechanism. in, It is a non-linear activation function. For the new feature, the transformation feature of each neighbor node j According to its attention weight The weighted results are summed, and then passed through a nonlinear activation function. The process ultimately generates a new feature representation for node i that incorporates neighborhood context information. The entire process utilizes learnable parameters 'a' and 'W' to enable the model to adaptively capture important relationships between nodes. The GAT layer dynamically and selectively aggregates information by assigning different attention weights to neighboring nodes, thereby not only capturing spatial dependencies but also accurately locating the node regions that contribute the most to anomalies. Subsequently, the node feature sequences output by the GAT layer, which already contain rich spatial context information, are input into the temporal convolutional network (TCN) layer for temporal dimension modeling.

[0031] TCN uses one-dimensional dilated causal convolution to process the new feature sequence of each node after GAT layer enhancement: in, This represents the convolution operation. f It is a length of K The convolution kernel function, whose parameters are determined by... k index, These are the weight parameters of the convolution kernel at position k; z It is the input time-series signal; t Inflation factor represents the current time step of the calculation. d The interval span of the convolution kernel when upsampling the input sequence is controlled by introducing The offset allows the convolution kernel to expand its receptive field while maintaining a constant total number of parameters, thus effectively capturing long-term temporal dependencies. Summation operation This indicates a weighted summation across all positions of the convolution kernel, while This design ensures that each computation step relies solely on data from the current and past time steps, strictly adhering to causal constraints and preventing the leakage of future information. This approach enables the model to efficiently process long sequences and identify temporal patterns. Compared to traditional RNNs / LSTMs, TCN offers advantages such as high parallel computation efficiency, stable gradients, and strong pattern recognition capabilities.

[0032] Step 4: Based on the spatiotemporal features output by the spatiotemporal graph convolutional network model, perform fault classification and fault location.

[0033] The specific implementation method of step 4 is as follows: obtain the fault level based on the spatiotemporal characteristics and classify and process it: minor anomalies are recorded in the system log; general faults are issued with early warning notifications; serious faults trigger emergency alarms and recommend shutdown for maintenance.

[0034] In the spatiotemporal graph convolutional network constructed in this invention, the graph attention network (GAT) and the temporal convolutional network (TCN) work together in a hierarchical manner to support the two core output dimensions of fault diagnosis. The GAT layer acts as a spatial perceptron, dynamically calculating the attention weights between nodes. A spatial propagation map of fault signals in the sensor network was constructed. When a fault occurs at a specific location, its neighboring nodes form a high-weight focus of attention due to vibration propagation. This spatial dependency directly provides a theoretical basis for the fault localization branch. By identifying clusters of high-weight nodes, the physical area where the fault occurs can be accurately located. Subsequently, the TCN layer, acting as a temporal analyzer, performs temporal pattern mining on the feature sequences spatially enhanced by GAT. Its dilated causal convolution structure can effectively capture long-term dependency features related to the fault type in the vibration signal. These deeply extracted spatiotemporal features are finally transformed into a probability distribution of fault types through a fully connected layer, completing the fault classification task. The synergistic effect of the two branches reflects a complete diagnostic logic: GAT's solution of "where is the fault?" provides spatial context for TCN's analysis of "what is the fault?", while the fault type confirmed by TCN, in turn, verifies whether the area located by GAT exhibits typical temporal characteristics of that type of fault. This spatiotemporal cross-validation mechanism ensures that the diagnostic results have both localization accuracy and classification reliability, ultimately forming an actionable diagnostic conclusion of "a certain type of fault occurs in a certain area." This allows for the precise localization of the physical area where the fault occurs. Ultimately, the system triggers corresponding warnings or maintenance commands based on the diagnostic results.

[0035] Based on the above-mentioned fault detection method for quick-installation prefabricated walkways, the effectiveness of the present invention is verified by performing the following calculation tests.

[0036] like Figure 1 As shown, after system startup, the system first performs hardware self-tests and sensor network initialization. The triaxial vibration acceleration sensors deployed on each module of the quick-assembly prefabricated freight channel begin operation, and the required sampling frequency and range are set. The data acquisition module synchronizes data between modules via a CAN bus network to ensure timestamp consistency.

[0037] The raw vibration signals acquired by the sensors first enter the preprocessing stage. The signal preprocessing module performs quality checks on the raw signals and marks or interpolates abnormal signals. A 4th-order Butterworth bandpass filter preserves the vibration characteristic frequency bands within the normal range, followed by z-score normalization to eliminate the influence of dimensions. The preprocessed signal then enters the feature extraction stage, calculating time-domain features including root mean square value, peak value, and kurtosis, as well as frequency-domain features such as spectral centroid and bandwidth.

[0038] like Figure 2 As shown, the dynamic graph structure system employs a sliding time window mechanism, with a window length of 1.0 second and a sliding step of 0.5 seconds. For each time window, a node feature matrix is ​​first constructed, combining the feature vectors of each module into a feature matrix. The system then proceeds to the similarity calculation stage, offering two algorithm options: cosine similarity and dynamic time warping. The appropriate similarity measurement method can be selected based on the specific application scenario.

[0039] The choice of similarity measurement method is strictly based on the specific vibration signal characteristics and system operating conditions. The cosine similarity algorithm is mainly suitable for scenarios where the system is operating in a steady state. At this time, the statistical characteristics of the vibration signals collected by each sensor node (such as root mean square value, peak value, kurtosis, spectral centroid, etc.) remain stable. This method efficiently evaluates the overall distribution similarity of vibration modes between nodes by calculating the cosine value of the spatial angle between fixed-dimensional feature vectors, meeting the high computational efficiency requirements of real-time monitoring.

[0040] Dynamic Time Warping (DTW) is designed for non-steady-state conditions or scenarios requiring detailed analysis of waveform propagation characteristics. When the system experiences speed fluctuations, load changes, or fault impact signals that cause time delays and waveform distortions during propagation, DTW can directly compare the original vibration waveforms through nonlinear sequence alignment, accurately capturing similar patterns with phase differences between signals.

[0041] In practical applications, the system can automatically switch measurement methods based on the operating condition identification results: under steady-state conditions, cosine similarity is used first to achieve rapid screening, while under non-steady-state conditions, DTW is enabled for deep waveform matching, thereby optimizing the allocation of computing resources while ensuring diagnostic accuracy.

[0042] After similarity calculation, the system selects an adaptive threshold or a fixed threshold according to a preset strategy to generate a dynamic adjacency matrix. Finally, the node feature matrix will be... and adjacency matrix Combine them to form the graph structure of the current time window. The dynamic graph sequence is then incorporated into the spatiotemporal graph sequence. The completed dynamic graph sequence is input into the spatiotemporal graph neural network for analysis. In the GAT+TCN joint diagnostic mode, the graph attention network first performs spatial convolution on the graph structure of each time slice, calculates the attention weights between nodes using an 8-head attention mechanism, and generates node features containing spatial context by weighted aggregation of neighbor node information.

[0043] Given the current time window graph structure Two adjacent nodes i and j in the array have input feature vectors respectively. and First, the node features are linearly transformed using a shared weight matrix W. Then, the original attention coefficients are calculated. Finally, the normalized attention weights are obtained by normalizing all neighbors of node i using the softmax function. .

[0044] Spatially enhanced features are then fed into a temporal convolutional network (TCN). The TCN utilizes its unique dilated causal convolution structure, employing residual connections to ensure gradient stability and effectively capture long-term temporal dependencies in the sequence. In this collaborative mechanism, the GAT layer uses attention weight distribution to achieve initial fault region localization, providing the TCN layer with spatially semantic feature representations. The TCN layer then analyzes the temporal evolution patterns of these spatial features to further confirm the fault type and verify the reliability of the localization results. Finally, spatial and temporal features are integrated in a fusion layer to form a final spatiotemporal feature representation with clear physical meaning.

[0045] The spatiotemporal feature input fault diagnosis results are categorized and processed according to fault severity: minor anomalies are recorded in the system log, general faults trigger early warning notifications, and severe faults trigger emergency alarms and recommend shutdown for maintenance. Simultaneously, the fault detection system proposed in this invention can provide precise fault location information, offering navigation support for maintenance personnel. After each diagnosis, the system updates its status information and decides whether to continue monitoring or stop based on operational judgment. In continuous monitoring mode, the system returns to the sliding time window step and begins data processing for the next time window, forming a closed-loop intelligent monitoring cycle. The entire system adopts a modular design, supports hot-swapping and online upgrades, ensuring long-term stable operation.

[0046] It should be emphasized that the embodiments described in this invention are illustrative rather than limiting. Therefore, this invention includes, but is not limited to, the embodiments described in the specific implementation. Any other implementations derived by those skilled in the art based on the technical solutions of this invention are also within the scope of protection of this invention.

Claims

1. A fault detection method based on quick-installation prefabricated walkways, characterized in that, Includes the following steps: Step 1: Obtain vibration signals collected by sensors deployed on each module of the quick-installation prefabricated walkway; Step 2: Calculate the signal similarity between corresponding nodes of each module, and dynamically generate an adjacency matrix based on an adaptive threshold, thereby constructing a dynamic graph structure with modules as nodes and functional connections as edges; Step 3: Input the dynamic graph structure into the pre-trained spatiotemporal graph convolutional network model ST-GCN; the spatiotemporal graph convolutional network model performs spatial convolution on the dynamic graph structure through the graph attention network GAT layer to generate new features containing spatial context information; Then, the new feature sequence is temporally convolved through the Temporal Convolutional Network (TCN) layer to extract spatiotemporal features; Step 4: Based on the spatiotemporal features output by the spatiotemporal graph convolutional network model, perform fault classification and fault location.

2. The fault detection method based on quick-installation prefabricated walkways according to claim 1, characterized in that, The specific implementation method of step 1 is as follows: each module deployed on the quick-installation prefabricated walkway is defined as a node, the vibration signals collected by the sensors on each module are extracted, and a node feature matrix is ​​formed.

3. The fault detection method based on quick-installation prefabricated walkways according to claim 1, characterized in that, The specific implementation method of step 2 is as follows: combine the feature vectors of each module into a feature matrix. Calculate the real-time similarity of vibration signals between all node pairs and apply an adaptive threshold. Dynamically generate an adjacency matrix that reflects the functional connection relationships between nodes at the current time. , node feature matrix and adjacency matrix Combined, they form a dynamic graph structure for the current time window. .

4. The fault detection method based on quick-installation prefabricated walkways according to claim 1, characterized in that, In step 3, the graph attention network (GAT) layer performs spatial dimension modeling, assigning different attention weights to the neighbors of each central node: ； in, Here, the original attention coefficient represents the unnormalized importance of neighbor node j to center node i, where... and Let be the input feature vectors of the two nodes, and W be a trainable shared weight matrix used to perform linear transformation on the node features. The original attention coefficients are normalized using the softmax function, resulting in the following formula: ； in, The attention weights are normalized and their values ​​range from 0 to 1, representing the set of all neighbors of node j at node i. The relative importance of each node is given by k, which is the index of a neighboring node. The denominator is summed with respect to the exponential coefficients of all neighbors of node i to ensure that the sum of the weights is 1. ∥ represents the vector concatenation operation, a is a trainable attention vector that maps the concatenated features to scalar scores, and LeakyReLU introduces a non-linear activation function to update node features through a weighted aggregation mechanism. ； in, It is a non-linear activation function. For the new feature, the transformation feature of each neighbor node j According to its attention weight The weighted results are summed, and then passed through a nonlinear activation function. The process ultimately generates a new feature representation for node i that incorporates neighborhood context information. ; For the new features output by the GAT layer that already contain rich spatial context information New features of it across all time slices Stack them in chronological order to form a new feature sequence, which is then input into the temporal convolutional network layer for temporal dimension modeling.

5. The fault detection method based on quick-installation prefabricated walkways according to claim 1, characterized in that, In step 3, the temporal convolutional network layer TCN uses one-dimensional dilated causal convolution to process the new feature sequence of each node after enhancement by the GAT layer: ； in, This represents the convolution operation. f It is a length of K The convolution kernel function, whose parameters are determined by... k index, These are the weight parameters of the convolution kernel at position k; z It is the input time-series signal; t Represents the current time step in the calculation. d The expansion factor is used to obtain the spatiotemporal characteristics.

6. The fault detection method based on quick-installation prefabricated walkways according to claim 1, characterized in that, The specific implementation method of step 4 is as follows: obtain the fault level according to the spatiotemporal characteristics and classify it: minor abnormalities are recorded in the system log; general faults are issued with early warning notifications; serious faults trigger emergency alarms and recommend shutdown for maintenance.

Images