A catenary integrated intelligent detection system based on multi-sensor fusion

Abstract

The application relates to the technical field of rail transit detection, and particularly discloses an integrated intelligent detection system for a contact network based on multi-sensor fusion, which comprises a multi-source heterogeneous sensor array, an edge computing gateway, a data fusion and feature decoupling center, an interpretable fault diagnosis engine and a decision and visualization platform. Through a cross-modal attention fusion network and a variational autoencoder feature decoupling module, deep and interpretable data fusion is realized, fault features are decoupled into independent factors corresponding to different physical modes, and accurate classification, positioning and evaluation are carried out by the diagnosis engine. The application improves the transparency, accuracy and early warning capability of fault diagnosis.

Description

Technical Field

[0001] This invention belongs to the field of rail transit inspection technology, specifically relating to an integrated intelligent inspection system for overhead contact lines based on multi-sensor fusion. Background Technology

[0002] In the field of intelligent operation and maintenance of rail transit infrastructure, efficient and accurate condition monitoring and fault diagnosis of the overhead contact system are crucial to ensuring the safe and stable operation of trains. As a direct device for supplying electrical energy to electric locomotives, the overhead contact system has a complex structure and operates in a harsh environment. Its condition monitoring involves multi-dimensional information such as geometric parameters, electrical characteristics, and the integrity of mechanical components.

[0003] The intelligent contact network detection technology based on multi-sensor fusion aims to achieve comprehensive perception and evaluation of the contact network status by integrating multiple sensing devices such as vision, laser, infrared, and ultrasound to collaboratively collect images, point clouds, temperature, and defect feature data.

[0004] Simple data-layer or feature-layer fusion strategies are insufficient to effectively handle heterogeneous data from sensors based on different physical principles. These data differ significantly in dimensionality, scale, semantics, and noise characteristics, leading to redundancy and conflicts in the fused feature information, failing to form a consistent and in-depth representation of the overhead contact line's condition. Furthermore, traditional fusion methods often rely on pre-defined rules or shallow models. When faced with complex and coupled fault modes such as loose overhead contact line suspension components, insulator contamination, and conductor wear, their analysis lacks interpretability and struggles to accurately trace the root causes and evolution mechanisms of faults. This directly limits the detection system's ability to warn of early potential faults and the reliability of diagnostic results, making maintenance decisions still highly dependent on human experience and judgment. Summary of the Invention

[0005] The purpose of this invention is to provide an integrated intelligent detection system for overhead contact lines based on multi-sensor fusion, in order to solve the problems of insufficient fusion depth of multi-source heterogeneous sensor data, difficulty in coordinating feature conflicts, and the resulting poor interpretability of fault diagnosis and weak early warning capability in the prior art.

[0006] This invention provides an integrated intelligent detection system for overhead contact lines based on multi-sensor fusion. The system includes a multi-source heterogeneous sensor array, an edge computing gateway, a data fusion and feature decoupling center, an interpretable fault diagnosis engine, and a decision and visualization platform.

[0007] The multi-source heterogeneous sensor array is used to synchronously acquire multi-dimensional physical state raw data of the overhead contact system. This array consists of a high-resolution linear array camera module, a 3D laser scanning module, an infrared thermal imaging module, and an ultrasonic flaw detection module. The high-resolution linear array camera module continuously acquires visible light image sequences of the overhead contact system suspension components and conductors along the train's direction of travel at a sampling rate of 5000 lines per second. The 3D laser scanning module generates spatial 3D point cloud data of key areas of the overhead contact system with an angular resolution of 0.1 degrees and a ranging accuracy of 2 millimeters. The infrared thermal imaging module is equipped with an uncooled focal plane array detector, capturing real-time temperature field distribution images of the electrical connection points of the overhead contact system at a frame rate of 30 Hz. The ultrasonic flaw detection module uses a pulse-echo probe to perform online ultrasonic scanning of predefined key mechanical structural parts of the overhead contact system at a center frequency of 5 MHz to obtain echo signals of internal defects.

[0008] The edge computing gateway is used to perform localized preprocessing and temporal-spatial alignment of raw data acquired by a multi-source heterogeneous sensor array. This gateway has a built-in high-precision synchronization unit that provides a unified microsecond-level timestamp for all sensor modules. Its spatial alignment unit, based on a preset sensor extrinsic parameter calibration matrix, uniformly registers laser point cloud data, infrared thermal images, and visible light images to the same world coordinate system. The preprocessing includes adaptive histogram equalization and dehazing of the visible light image, outlier filtering and downsampling of the laser point cloud, non-uniformity correction of the infrared image, and bandpass filtering and gain compensation of the ultrasonic signal.

[0009] The data fusion and feature decoupling center is the core computing unit of the system, used to achieve deep fusion of multi-source data from the raw layer to the decision layer, and to decouple independent feature factors representing different physical fault modes. This center consists of a cascaded cross-modal attention fusion network and a variational autoencoder feature decoupling module. The cross-modal attention fusion network receives multimodal data preprocessed and aligned by the edge computing gateway as input. The network first extracts high-dimensional visual feature tensors and geometric feature tensors from image data and point cloud data, respectively, through parallel convolutional neural network branches and point cloud neural network branches.

[0010] Furthermore, the network introduces a cross-modal attention mechanism, which calculates the attention weight of each spatial location in the visual feature tensor to all points in the geometric feature tensor, while simultaneously calculating the symmetric attention weight of the geometric features to the visual features. Based on the calculated bidirectional attention weights, the network performs weighted fusion of visual and geometric features to generate a primary fused feature containing information about the relationship between appearance and structure. Subsequently, this primary fused feature, along with the infrared temperature distribution features extracted by the convolutional neural network and the ultrasonic time-frequency features extracted by the one-dimensional convolutional network, are input into a multilayer perceptron for secondary feature interaction and fusion, outputting a unified multi-source deep fused feature vector.

[0011] The variational autoencoder feature decoupling module is used to perform decoupling analysis on the unified multi-source deep fusion feature vector. The encoder of this module maps the fused feature vector to the mean vector and log-variance vector of a multivariate Gaussian distribution. The latent space dimension of this multivariate Gaussian distribution is pre-divided into four mutually orthogonal subspaces, corresponding to the four core physical fault modes of the overhead contact system: mechanical loosening, electrical overheating, geometric deformation, and surface contamination and wear. During training, by applying subspace orthogonality constraint loss and feature factor independence loss, the encoder is forced to learn to encode the representation information of different fault modes into different statistically independent dimensional subsets of the latent space. The decoder then reconstructs the input multi-source deep fusion features based on the feature factors sampled from the latent space.

[0012] After training, for any fused feature in the input, the magnitude of the value in a specific subspace dimension of the mean vector output by the encoder directly quantifies the significance of the corresponding type of fault mode, thereby achieving interpretability decoupling of fault features.

[0013] The explained fault diagnosis engine is used to accurately classify, locate, and assess the severity of faults based on decoupled feature factors output by the data fusion and feature decoupling center. The engine includes a fault classifier and a regression estimator. The fault classifier takes a decoupled 4-dimensional feature factor vector as input and outputs a multi-dimensional fault probability vector through a fully connected neural network. Each element of this vector corresponds to a specific fault type, such as a loose locator, broken dropper, overheated clamp, excessive conductor wear, or insulator flashover. The regression estimator receives the same feature factor vector and intermediate layer features from the fault classifier. Through another fully connected network, it regresses the precise 3D spatial coordinates of the fault point and a fault severity index between 0 and 1. The engine's diagnostic report not only includes the fault type and location but also explicitly lists the contribution values ​​of each decoupled feature factor, thus providing a basis for tracing the diagnostic decision.

[0014] The decision-making and visualization platform integrates diagnostic results, generates maintenance strategies, and provides a comprehensive display. The platform receives all diagnostic information output by the interpretable fault diagnosis engine, including fault type, probability, location coordinates, severity index, and feature factor contribution. The platform's built-in decision rule library automatically matches preset maintenance priorities and recommended handling measures based on the fault type and severity index. The visualization module uses a 3D point cloud model of the overhead contact line as the base map, overlays the infrared temperature field in pseudo-color onto the corresponding locations, and highlights the diagnosed fault points in the 3D scene with different colors and icons. It also displays a detailed text diagnostic report and a feature factor contribution bar chart.

[0015] As one embodiment of the present invention, the calculation process of bidirectional attention weights in the cross-modal attention fusion network is as follows: First, the visual feature tensor and the geometric feature tensor are projected onto the same feature dimension through independent linear transformation layers; then, the dot product similarity between each feature vector in the projected visual features and all feature vectors in the geometric features is calculated and normalized using the Softmax function to obtain the visual-to-geometric attention weight matrix; the geometric-to-visual attention weight matrix is ​​symmetrically calculated; then, the geometric features are weighted and summed using the visual-to-geometric attention weight matrix to obtain the context geometric features from the visual perspective; the visual features are weighted and summed using the geometric-to-visual attention weight matrix to obtain the context visual features from the geometric perspective; finally, the original features are concatenated with their corresponding context features and a primary fusion feature is generated through a nonlinear fusion layer.

[0016] In one embodiment of the present invention, the subspace orthogonality constraint loss applied during the training process of the variational autoencoder feature decoupling module is specifically achieved by calculating the cosine similarity matrix between feature vectors corresponding to different subspaces of the latent space and minimizing the Frobenius norm of this matrix with the identity matrix. The feature factor independence loss is applied by calculating and minimizing the mutual information estimate of the decoupled feature factors.

[0017] In one embodiment of the present invention, the training of the fault classifier in the interpretable fault diagnosis engine employs a focus loss function to alleviate the imbalance problem between normal contact network samples and various fault samples. In the regression evaluator, a smoothing L1 loss function is used for the regression of the three-dimensional coordinates of the fault point, and a mean squared error loss function is used for the regression of the fault severity index.

[0018] In one embodiment of the invention, the system operates in a hybrid architecture combining vehicle-mounted mobile detection mode and fixed-point monitoring mode. The vehicle-mounted mobile detection mode is executed while the inspection train is running, completing a periodic and rapid general survey of the entire overhead contact line. The fixed-point monitoring mode is deployed in key sections, such as phase separation zones and tunnel entrances, for continuous monitoring 24 / 7. Data collected by both modes is uploaded to a cloud data center, where it is uniformly processed and analyzed by a data fusion and feature decoupling center and an interpretable fault diagnosis engine. The analysis results are simultaneously updated to the decision-making and visualization platform.

[0019] In one embodiment of the present invention, the edge computing gateway and the data fusion and feature decoupling center transmit data via a dedicated 5G network or high-speed Ethernet. The data fusion and feature decoupling center and the interpretable fault diagnosis engine are deployed in a high-performance computing cluster in the cloud or a railway bureau-level private cloud server.

[0020] Compared with the prior art, the beneficial effects of the present invention are as follows: 1. This invention constructs a deep and interpretable multi-source data fusion core by designing a cascaded cross-modal attention fusion network and a variational autoencoder feature decoupling module. The cross-modal attention mechanism adaptively balances the correlation between image, point cloud, infrared, and ultrasonic data, achieving information complementarity and conflict coordination, and generating deep fusion features with low information redundancy and high consistency. The variational autoencoder decouples the fusion features into independent factors corresponding to different physical fault modes by applying subspace orthogonality and independence constraints. This allows subsequent diagnosis to not only provide results but also clearly reveal which physical state anomaly is dominant, greatly improving the transparency and credibility of the system diagnosis process and providing direct evidence for maintenance personnel to understand the root cause of faults.

[0021] 2. The interpretable fault diagnosis engine proposed in this invention directly utilizes decoupled, physically meaningful feature factors for fault classification and assessment. This changes the traditional approach's reliance on black-box model outputs. The quantified output of feature factor contributions allows the system to clearly identify the key evidence sources leading to the diagnostic decision, such as whether the geometric deformation factor is significantly high or the electrical overheating factor is abnormally prominent. This traceable diagnostic logic significantly reduces the false alarm rate and enables the system to identify the coupling effects of multiple fault modes. It is crucial for the accurate location and refined severity assessment of complex faults, thus supporting the advancement from preventative maintenance to predictive maintenance.

[0022] 3. This invention integrates a complete technology chain from edge preprocessing and cloud-based deep analysis to terminal decision visualization, constructing an integrated intelligent detection closed loop. A multi-source heterogeneous sensor array provides a comprehensive data foundation, the edge computing gateway ensures data quality and spatiotemporal consistency, the cloud center completes core intelligent analysis, and the decision platform enables intuitive application of the results. This systematic design not only improves the accuracy and depth of single detections but also achieves continuous three-dimensional perception of the catenary status in both time and space through a hybrid monitoring mode combining onboard and fixed-point monitoring. This lays a solid data and analytical foundation for building a digital twin model and realizing full lifecycle health management, comprehensively improving the intelligence level and safety reliability of rail transit catenary operation and maintenance. Attached Figure Description

[0023] Figure 1 This is a schematic diagram of the overall technical solution architecture of the present invention; Figure 2 This is a schematic diagram of the core principle framework of the data fusion and feature decoupling center in this invention; Figure 3 This is a cascaded logical framework diagram of the cross-modal attention fusion network and the variational autoencoder feature decoupling module in this invention; Figure 4 This is a logical flow diagram of the fault diagnosis engine in this invention that can be explained by performing diagnosis based on decoupled feature factors; Figure 5 This is a schematic diagram of the multi-level interaction and data flow of the hybrid architecture of vehicle-mounted mobile detection and fixed-point monitoring in this invention. Detailed Implementation

[0024] Example 1: Please refer to the appendix Figures 1 to 5 This invention provides an integrated intelligent detection system for overhead contact lines based on multi-sensor fusion. The system consists of five core functional units: a multi-source heterogeneous sensor array, an edge computing gateway, a data fusion and feature decoupling center, an interpretable fault diagnosis engine, and a decision-making and visualization platform. These units collaborate efficiently through standardized data interfaces and communication protocols, forming a complete closed loop from initial sensing to intelligent decision-making. The following detailed description of each component, in conjunction with the accompanying drawings, will illustrate the system.

[0025] A multi-source heterogeneous sensor array, serving as the system's front-end sensing layer, is deployed on a dedicated bracket on the top of the inspection train or at a fixed monitoring point. It is used to synchronously acquire multi-dimensional physical field data of the overhead contact system during operation. This sensor array consists of four sub-modules: a high-resolution linear array camera module, a 3D laser scanning module, an infrared thermal imaging module, and an ultrasonic flaw detection module. Each module achieves microsecond-level time synchronization under hardware trigger signal control. The high-resolution linear array camera module employs a global shutter CMOS image sensor with an effective pixel count of no less than 8192. It continuously acquires visible light image sequences of the overhead contact system suspension components (including positioners, droppers, clamps, insulators, etc.) and conductor surfaces along the train's direction of travel at a sampling rate of 5000 lines per second. Each frame has a 12-bit grayscale resolution and a dynamic range of no less than 70 dB, ensuring that critical structural details are preserved even under strong backlight or low-light conditions.

[0026] The 3D laser scanning module employs a galvanometer scanning mechanism with a horizontal angular resolution of 0.1 degrees and a vertical angular resolution of 0.05 degrees. Its ranging accuracy is better than ±2 mm, and its maximum effective ranging distance is 15 meters. It can generate high-density spatial 3D point cloud data for key support structures such as catenary supports, cantilever arms, and positioning tubes, with a point cloud density of no less than 40,000 points per square meter at a distance of 10 meters. The infrared thermal imaging module is equipped with an uncooled vanadium oxide focal plane detector with a pixel size of 12 micrometers, an array size of 640×512, an operating wavelength of 8 to 14 micrometers, a frame rate of 30 Hz, and a thermal sensitivity better than 50 milliklvin. It is used to capture real-time temperature field distribution images of electrical connection points in the catenary (such as electrical connection clamps and sectional insulator joints), with a temperature measurement range covering -20 degrees Celsius to 250 degrees Celsius.

[0027] The ultrasonic flaw detection module employs a pulse-reflection piezoelectric ceramic probe with a center frequency of 5 MHz, a bandwidth of 80%, and a repetition frequency of 1 kHz. It uses a non-contact electromagnetic transducer for coupling. Online ultrasonic scanning is performed on predefined key mechanical structural parts (such as the end pressing area of ​​the dropper and the root weld of the locator) to acquire A-scan echo signals. The sampling frequency is 100 MHz, and the dynamic range is 60 dB, effectively identifying internal cracks or porosity defects deeper than 0.5 mm. All sensing modules are equipped with independent temperature control and dustproof / waterproof housings, operating in an ambient temperature range of -40°C to +70°C, with a protection rating of at least IP66.

[0028] The edge computing gateway, acting as a local data processing hub, is directly connected to the multi-source heterogeneous sensor array, undertaking the preprocessing and spatiotemporal alignment of raw data. This gateway incorporates a high-precision synchronization unit, employing the IEEE 1588 precision time protocol or a BeiDou / GNSS dual-mode timing module to provide a unified microsecond-level timestamp for all sensor modules, ensuring strict alignment of data across modalities on the time axis with a time synchronization error of less than 5 microseconds. The spatial alignment unit performs coordinate transformation based on a pre-calibrated sensor extrinsic parameter matrix. This extrinsic parameter matrix is ​​obtained by jointly calibrating the relative poses of each sensor's optical axis and mounting base using a high-precision total station in a standard calibration field. It includes a rotation matrix R and a translation vector t, used to uniformly register laser point cloud data, infrared thermal images, and visible light images to a world coordinate system based on the contact wire centerline. Specifically, for each pixel (u,v) in a visible light image frame, it is back-projected to world coordinates using the camera intrinsic parameter matrix K and extrinsic parameter [R|t]; for each three-dimensional point in the laser point cloud... Through extrinsic parameter transformation Mapped to world coordinates; infrared images are then mapped using an affine transformation matrix. The pixel coordinates are mapped to the visible light image coordinate system, and then transformed to world coordinates via visible light extrinsic parameters.

[0029] The preprocessing steps are performed separately for different modal data characteristics: For visible light image sequences, adaptive histogram equalization is first performed to enhance local contrast, followed by a dehazing algorithm based on dark channel priors to eliminate scattering effects in tunnel or rain / fog environments, resulting in a corrected image with over 30% improved clarity; For laser point cloud data, a statistical outlier filter is applied to remove outliers caused by interference from birds, raindrops, etc., and then voxel grid downsampling is used to control the point cloud density to within 2000 points per cubic meter to reduce subsequent computational load; For infrared images, two-point non-uniformity correction is performed, using reference images acquired by a blackbody radiation source at both room temperature and high temperature to calculate the gain and bias parameters of each pixel, eliminating detector response inconsistencies; For ultrasonic echo signals, a bandpass filter with a center frequency of 5 MHz and a bandwidth of 2 MHz is used to suppress low-frequency vibration noise and high-frequency electronic interference, and time-varying gain compensation is applied to offset the attenuation effect of sound waves propagating in materials, keeping the defect echo amplitude stable. The preprocessed multimodal data is packaged into structured data frames, which include timestamps, coordinate system identifiers, data type labels, and checksums, and then transmitted to the next processing unit via a high-speed interface.

[0030] The data fusion and feature decoupling center is the core intelligent analysis unit of the system, and its internal logical framework is shown in the attached figure. Figure 2 With appendix Figure 3 As shown, the center receives preprocessed and aligned multimodal data streams from edge computing gateways, performs deep fusion from the raw layer to the semantic layer, and further decouples independent fault feature factors with clear physical meaning. The center consists of a cascaded cross-modal attention fusion network and a variational autoencoder feature decoupling module. The cross-modal attention fusion network first extracts modality-specific features from the input data. Visible light images and infrared thermal images are fed into parallel convolutional neural network branches, where the visible light branch uses a ResNet-50 backbone network and the infrared branch uses a lightweight MobileNetV3 structure. Both are fine-tuned through transfer learning on ImageNet and railway-specific image datasets, outputting a high-dimensional visual feature tensor of dimension 2048. The 3D laser point cloud data is then input into the PointNet++ network, which extracts geometric feature tensors through multi-scale grouping and feature propagation mechanisms. , where N is the number of point clouds.

[0031] The ultrasonic echo signal is processed by a one-dimensional convolutional network containing five convolutional layers, each with a kernel size of 31, a stride of 2, and a ReLU activation function. The final output is a 128-dimensional time-frequency feature vector. Cross-modal attention mechanisms build upon this foundation to establish a dynamic relationship between visual and geometric modalities. Specifically, and First, project the data onto the same feature dimension d=512 using independent 1×1 convolutional layers (or fully connected layers) to obtain the desired results. and Then, the attention weight matrix from vision to geometry is calculated. Its elements , representing the degree of attention given to geometric point j by visual feature location i. Symmetrically, the attention weight matrix from geometry to vision is calculated. Its elements .based on ,right We perform a weighted summation to obtain the contextual geometric features from the visual perspective. ;based on ,right We perform a weighted summation to obtain the contextual visual features from a geometric perspective. The original features are concatenated with their context features, and then nonlinearly fused using two MLP layers (512-dimensional hidden layer and 256-dimensional output layer) to generate primary fused features. This primary fusion feature and (Infrared characteristics) and The ultrasound features are input to a multilayer perceptron (MLP), which contains three fully connected layers with dimensions of 1024, 512, and 256, respectively. The activation function is GELU, and the final output is a 256-dimensional unified multi-source deep fusion feature vector. .

[0032] The variational autoencoder feature decoupling module receives the aforementioned deep fusion feature vector. As input, it performs a feature decoupling task. The encoder of this module consists of a three-layer fully connected network: a 256-dimensional input layer, a 512-dimensional hidden layer, and an output layer that is the mean vector of the latent variables. With log-variance vector The potential space dimension 64 is pre-divided into four mutually orthogonal 16-dimensional subspaces S1, S2, S3, and S4, corresponding to four core physical fault modes: mechanical loosening, electrical overheating, geometric deformation, and surface contamination and wear, respectively. During training, two types of constraint losses are applied to guide decoupling learning. Subspace orthogonality constraint loss. By calculating the cosine similarity matrix between feature vectors of different subspaces ,in and minimize ( (The matrix is ​​a 4th-order identity matrix), forcing the representations of each subspace to be mutually orthogonal. Eigenfactor independence loss. The mutual information between feature dimensions within each subspace is calculated using the Mutual Information Neural Estimator (MINE). And minimize its sum to ensure that features within the subspace remain statistically independent. The total loss function is ,in The reconstruction loss is calculated using mean squared error. and The balancing coefficients are set to 0.5 and 0.3 respectively. The decoder consists of a symmetrical three-layer fully connected network that samples from the latent space. Reconstructed into the original fused features ẑ. After training, for any input sample, the mean vector output by the encoder is... In this model, the L2 norm of subspace S1 quantifies the significance of mechanical loosening faults, S2 corresponds to electrical overheating, S3 to geometric deformation, and S4 to surface contamination and wear. This decoupling mechanism gives each feature factor a clear physical reference, laying the foundation for subsequent interpretable diagnosis.

[0033] The interpretable fault diagnosis engine is based on the aforementioned decoupled feature factor vector. The logic flow for fault diagnosis is as follows: Figure 4 As shown, the engine comprises two parallel sub-modules: a fault classifier and a regression estimator. The fault classifier employs a three-layer fully connected network, with an input of a 4-dimensional feature factor vector f, hidden layers of dimensions 64 and 32, and an output layer of dimension [missing information]. ( For the total number of predefined fault types, for example =8, corresponding to locator loosening, dropper wire breakage, clamp overheating, conductor wear exceeding limits, insulator flashover, bracket misalignment, segment insulator damage, and surge arrester deterioration), with the activation function being Softmax, outputting a multi-dimensional fault probability vector. ,satisfy The classifier uses a focus loss function during the training phase. ,in For category weights, =2, to alleviate the class imbalance problem where normal samples far outnumber various fault samples. The regression estimator receives the same feature factor vector f and the 32-dimensional intermediate features h from the second hidden layer of the fault classifier, and performs regression prediction through a four-layer fully connected network (dimensions 64, 128, 64, 4). The output contains the three-dimensional spatial coordinates of the fault point. With Fault Severity Index Coordinate regression uses a smoothed L1 loss function, while severity index regression uses mean squared error loss. The severity index s is defined based on historical maintenance records and expert knowledge: s < 0.3 indicates a minor anomaly, requiring observation; 0.3 ≤ s < 0.7 indicates a moderate fault, requiring planned maintenance; and s ≥ 0.7 indicates a severe fault, requiring immediate action. The diagnostic report not only includes the fault type, probability, location, and severity, but also clearly lists the values ​​of each decoupling characteristic factor and their contribution ranking in this diagnosis. For example, "This diagnosis is wire clamp overheating (probability 92%), mainly driven by the electrical overheating factor (value 0.85), with the geometric deformation factor (value 0.12) contributing less," thus providing a complete decision tracing chain.

[0034] The decision-making and visualization platform serves as the system's terminal interface, deployed at the railway bureau's dispatch center or on mobile terminal devices. This platform receives all diagnostic information from the interpretable fault diagnosis engine, including fault type, probability, three-dimensional coordinates, severity index, and feature factor contribution. The platform's built-in decision rule base is implemented using an if-then rule engine. For example, "If the fault type is 'clamp overheating' and s≥0.7, the maintenance priority is 'emergency,' and the recommended measure is 'immediately disconnect power and replace the clamp'"; "If the fault type is 'excessive conductor wear' and 0.3≤s<0.7, the priority is 'high,' and the recommended measure is 'arrange to replace the conductor section within 72 hours'."

[0035] The visualization module uses a 3D point cloud model of the overhead contact line as its base map, which is constructed from historical laser scanning data and updated regularly. The infrared temperature field is overlaid in pseudo-color (blue for low temperature, red for high temperature) onto the corresponding world coordinates. Diagnosed fault points are marked with different colored icons: red triangles indicate urgent faults, orange circles for high-priority faults, and yellow squares for medium-to-low-priority anomalies. Clicking any fault icon brings up a detailed diagnostic report window, displaying text descriptions, probability values, severity indices, and bar charts showing the contribution of characteristic factors. All data supports filtering and statistical analysis by line section, time range, and fault type, generating weekly, monthly, and trend warning charts.

[0036] The system operates in a hybrid architecture combining vehicle-mounted mobile detection and fixed-point monitoring, as shown in the attached diagram. Figure 5As shown. The vehicle-mounted mobile detection mode is activated when the inspection train is running at a speed of 80 km / h, completing a periodic rapid survey of the entire overhead contact line (typically 200 km long), with a survey cycle of once every 7 days. The fixed-point monitoring mode is deployed at key locations such as phase separation zones, tunnel entrances, and windy sections, with several monitoring stations set up. Each station is equipped with a complete multi-source heterogeneous sensor array and edge computing gateway, performing continuous monitoring 24 / 7 with a sampling interval of 5 minutes. Preprocessed data collected by both modes is uploaded to the cloud data center via a dedicated 5G network or gigabit Ethernet. The data fusion and feature decoupling center and the interpretable fault diagnosis engine are deployed on a railway bureau-level private cloud server cluster, using Kubernetes containerization management to support elastic scaling. Cloud analysis results are synchronized to the decision-making and visualization platform in real time and pushed to the mobile terminals of relevant maintenance personnel via message queues. This hybrid architecture achieves three-dimensional monitoring with full spatial coverage and high temporal density, ensuring both the breadth of the entire line survey and the depth of monitoring in key sections.

[0037] The entire system employs AES-256 encryption and two-way TLS authentication during data transmission to ensure data security. Communication between functional units uses the Protobuf serialization protocol, defining a strict data structure including fields such as timestamp (microsecond level), data source identifier, coordinate system version number, and checksum hash value. The system has a comprehensive fault self-diagnosis and degradation operation mechanism: if a sensor module fails, the edge computing gateway automatically marks the data of that modality as invalid, and the data fusion center switches to a three-modal fusion mode; if the network is interrupted, the edge gateway locally caches the data of the most recent 24 hours and automatically resumes transmission after recovery. The system software version is remotely updated via OTA, and the firmware upgrade process has rollback capabilities to ensure operational stability.

[0038] Example 2: Building upon Example 1, this example optimizes the training strategy for the data fusion and feature decoupling center to further enhance the robustness and generalization ability of feature decoupling. Specifically, an adversarial perturbation enhancement mechanism is introduced during the training process of the variational autoencoder feature decoupling module. In each iteration, a small-amplitude adversarial perturbation is applied to the input deep fusion feature vector z. This perturbation is generated using the Fast Gradient Sign Method (FGSM), i.e. ,in The perturbation strength is set to 0.01. Characteristics after perturbation. The data is fed into the encoder, which is required to output stable decoupling characteristic factors even under disturbances. This process is equivalent to expanding the decision boundaries of various fault modes in the potential space, enhancing the model's resistance to sensor noise and environmental interference.

[0039] Furthermore, this embodiment improves the regression estimator of the interpretable fault diagnosis engine by introducing an uncertainty quantification mechanism. The output layer of the regression estimator is expanded to predict the mean. With variance The system employs negative log-likelihood as the loss function. For any diagnostic result, the system not only outputs a point estimate of the fault severity index s, but also provides its 95% confidence interval. When the confidence interval width exceeds a preset threshold (e.g., 0.3), the system automatically marks the diagnostic result as "low confidence" and displays a flashing border in the visualization platform, suggesting manual review. This mechanism effectively avoids making overconfident judgments when data quality is poor or the failure mode is rare, thus improving the system's reliability.

[0040] Regarding the hybrid monitoring architecture, this embodiment adds an edge-cloud collaborative inference mechanism. The edge computing gateway has a built-in lightweight decoupled feature extraction model (with only 1 / 10 the number of parameters of the cloud model), which can perform local preliminary diagnosis when network latency is too high. The severity index of the local diagnosis result... When the value is ≥ 0.6, a local alarm is immediately triggered and the original data is uploaded first; otherwise, only the compressed feature vector is uploaded. This strategy ensures timely response to critical faults while significantly reducing 5G network bandwidth usage.

[0041] The above-mentioned improvement scheme, without changing the overall system architecture, further enhances the robustness, reliability and resource efficiency of the system through refined design at the algorithm level, making it suitable for complex and ever-changing railway operating environments.

Claims

1. An integrated intelligent detection system for overhead contact lines based on multi-sensor fusion, characterized in that, include: A multi-source heterogeneous sensor array is used to synchronously acquire multi-dimensional physical state raw data of the contact network. The multi-source heterogeneous sensor array consists of a high-resolution linear array camera module, a three-dimensional laser scanning module, an infrared thermal imaging module, and an ultrasonic flaw detection module. An edge computing gateway is used to perform localized preprocessing and temporal-spatial alignment on the raw data collected by the multi-source heterogeneous sensor array. The edge computing gateway has a built-in high-precision synchronization time unit and spatial alignment unit. The data fusion and feature decoupling center is used to perform deep fusion and feature decoupling on the multimodal data that has been preprocessed and aligned by the edge computing gateway. The data fusion and feature decoupling center consists of a cascaded cross-modal attention fusion network and a variational autoencoder feature decoupling module. An interpretable fault diagnosis engine is used to accurately classify, locate, and assess the severity of faults based on the decoupled feature factors output by the data fusion and feature decoupling center. The interpretable fault diagnosis engine includes a fault classifier and a regression estimator. The decision-making and visualization platform is used to integrate the diagnostic results output by the interpretable fault diagnosis engine, generate maintenance strategies, and display them comprehensively. The decision-making and visualization platform has a built-in decision rule base and visualization module.

2. The integrated intelligent detection system for overhead contact lines based on multi-sensor fusion according to claim 1, characterized in that, The cross-modal attention fusion network performs the following process: High-dimensional visual feature tensors and geometric feature tensors are extracted from image data and point cloud data respectively through parallel convolutional neural network branches and point cloud neural network branches. A cross-modal attention mechanism is introduced to calculate the attention weight of each spatial location in the visual feature tensor to all points in the geometric feature tensor, and at the same time calculate the symmetric attention weight of the geometric features to the visual features. Based on the calculated bidirectional attention weights, visual features and geometric features are weighted and fused to generate primary fused features; The primary fusion features, along with the infrared temperature distribution features extracted by the convolutional neural network and the ultrasonic time-frequency features extracted by the one-dimensional convolutional network, are input into a multilayer perceptron for secondary feature interaction and fusion, outputting a unified multi-source deep fusion feature vector.

3. The integrated intelligent detection system for overhead contact lines based on multi-sensor fusion according to claim 2, characterized in that, The calculation process for the bidirectional attention weights in the cross-modal attention mechanism is as follows: First, the visual feature tensor and the geometric feature tensor are projected onto the same feature dimension through independent linear transformation layers; Next, the dot product similarity between each feature vector in the projected visual features and all feature vectors in the geometric features is calculated, and then normalized using the Softmax function to obtain the attention weight matrix from vision to geometry. Symmetrically compute the geometry-to-vision attention weight matrix; Then, the geometric features are weighted and summed using a visual-to-geometric attention weight matrix to obtain the contextual geometric features from the visual perspective. The visual features are weighted and summed using a geometry-to-vision attention weight matrix to obtain the contextual visual features from a geometric perspective. Finally, the original features are concatenated with their corresponding context features, and the primary fused features are generated through a non-linear fusion layer.

4. The integrated intelligent detection system for overhead contact lines based on multi-sensor fusion according to claim 1, characterized in that, The variational autoencoder feature decoupling module performs the following process: Its encoder maps the unified multi-source deep fusion feature vector into the mean vector and log-variance vector of a multivariate Gaussian distribution. The potential spatial dimension of this multivariate Gaussian distribution is pre-divided into four mutually orthogonal subspaces, which correspond to the four core physical fault modes of the contact network: mechanical loosening, electrical overheating, geometric deformation, and surface contamination and wear. Its decoder reconstructs the multi-source deep fusion features of the input based on feature factors sampled from the latent space; During training, by applying subspace orthogonality constraint loss and feature factor independence loss, the encoder is forced to learn to encode the representation information of different fault modes into different and statistically independent dimensional subsets of the latent space.

5. The integrated intelligent detection system for overhead contact lines based on multi-sensor fusion according to claim 4, characterized in that, The subspace orthogonality constraint loss is achieved by calculating the cosine similarity matrix between the eigenvectors corresponding to different subspaces of the latent space and minimizing the Frobenius norm of this matrix and the identity matrix; the eigenfactor independence loss is applied by calculating and minimizing the mutual information estimate of the decoupled eigenfactors.

6. The integrated intelligent detection system for overhead contact lines based on multi-sensor fusion according to claim 1, characterized in that, The input to the fault classifier is a decoupled 4-dimensional feature factor vector, which outputs a multi-dimensional fault probability vector through a fully connected neural network. The regression evaluator receives the same feature factor vector and the intermediate layer features of the fault classifier, and regresses the precise three-dimensional spatial coordinates of the fault point and the fault severity index between 0 and 1 through another fully connected network.

7. The integrated intelligent detection system for overhead contact lines based on multi-sensor fusion according to claim 6, characterized in that, The fault classifier is trained using the focus loss function; the regression evaluator uses the smoothing L1 loss function for the regression of the three-dimensional coordinates of the fault point and the mean squared error loss function for the regression of the fault severity index.

8. The integrated intelligent detection system for overhead contact lines based on multi-sensor fusion according to claim 1, characterized in that, The preprocessing process of the edge computing gateway includes: adaptive histogram equalization and dehazing of visible light images; outlier filtering and downsampling of laser point clouds; non-uniformity correction of infrared images; and bandpass filtering and gain compensation of ultrasonic signals.

9. The integrated intelligent detection system for overhead contact lines based on multi-sensor fusion according to claim 1, characterized in that, The system operates in a hybrid architecture that combines vehicle-mounted mobile detection mode and fixed-point monitoring mode; The vehicle-mounted mobile inspection mode is executed while the inspection train is running, completing a periodic and rapid general survey of the entire overhead contact line. The fixed-point monitoring mode is deployed in key sections to conduct continuous monitoring 24 hours a day, 7 days a year. Data collected in both modes is uploaded to the cloud data center, where it is processed and analyzed in a unified manner by the data fusion and feature decoupling center and the interpretable fault diagnosis engine.

10. The integrated intelligent detection system for overhead contact lines based on multi-sensor fusion according to claim 1, characterized in that, The edge computing gateway and the data fusion and feature decoupling center transmit data via a dedicated 5G network or high-speed Ethernet. The data fusion and feature decoupling center and the interpretable fault diagnosis engine are deployed in a high-performance computing cluster in the cloud or a railway bureau-level private cloud server.

Images