Method and system for detecting running state of facility in track connection channel

Abstract

The invention relates to the technical field of point cloud recognition, in particular to a method and system for detecting the running state of facilities in a track contact channel. The method comprises the following steps: acquiring point cloud data and sensing data of facilities in a track contact channel; carrying out data preprocessing on the obtained point cloud data and sensing data; performing feature extraction and feature selection on the preprocessed data; constructing a generative adversarial network model based on deep learning; a generative adversarial network model based on deep learning is used to carry out simulation analysis on the extracted features; and outputting a detection result. According to the invention, by deploying the three-dimensional laser scanner and the portable laser range finder, the high-precision point cloud data of the track contact channel and facilities are obtained, and three-dimensional space structure information which is difficult to capture by a traditional sensor is supplemented.

Description

Technical Field

[0001] The present invention relates to the field of point cloud recognition technology, and in particular to a method and system for detecting the operating status of facilities in a track connection channel. Background Art

[0002] As a key facility connecting adjacent tunnels, rail access tunnels perform important functions such as ventilation, drainage, and personnel evacuation. Their structural integrity (such as cracks and deformation) and equipment operating status (such as fan blade wear and drainage pump displacement) directly impact the safety and reliability of rail transit systems. Traditional monitoring methods rely primarily on vibration, temperature, and pressure sensors to collect one- or two-dimensional data, which makes it difficult to fully reflect the three-dimensional spatial changes in the tunnel structure and subtle damage to the equipment's appearance. With the maturity of point cloud data acquisition technologies (such as 3D laser scanning), it has become possible to obtain high-precision, high-density three-dimensional spatial information, providing a new dimension for facility status monitoring.

[0003] There are many defects in the existing technology:

[0004] Limitations of traditional sensors: Vibration and temperature sensors can only capture local physical parameters of equipment (such as vibration amplitude and temperature changes) and cannot detect three-dimensional changes in the structure or equipment (such as crack expansion in channel walls or deviation in the installation angle of ventilation equipment). Existing technology relies on manual inspections to detect external damage to facilities (such as dents in drainage pump casings and warped tracks), which is inefficient and highly subjective, making real-time quantitative assessment difficult.

[0005] Insufficient multi-source data fusion: In existing systems, different types of sensor data (such as vibration signals and structural point clouds) are usually processed independently, lacking effective spatial and temporal correlation. This makes it impossible to comprehensively analyze the coupling relationship between "structural deformation and equipment vibration anomalies" (for example, channel settlement causes track deformation, which in turn leads to increased vibration during train operation).

[0006] Rough outlier and noise processing: Traditional point cloud cleaning methods (such as simple threshold filtering) are difficult to adapt to complex environments (such as random noise caused by dust and lighting changes in the channel), and are prone to residual outliers or mistaken deletion of valid points, affecting the accuracy of subsequent feature extraction. There is a lack of customized filtering strategies for rail-connected channel scenarios (such as ignoring interference points caused by dynamic pedestrians and temporary equipment occlusion in the channel).

[0007] Large point cloud registration errors: The registration of multi-period point cloud data relies on manually selected reference points or simple rigid transformations, which makes it difficult to handle nonlinear deformations of channel structures (such as thermal expansion and contraction caused by temperature changes). This leads to large deviations in the alignment of data at different time points and makes it impossible to accurately quantify dynamic structural changes.

[0008] Insufficient utilization of geometric and topological features: Existing technologies mostly focus on the basic geometric features of point clouds (such as coordinates and normal vectors), ignoring the sensitive indication of topological features (such as changes in connected areas and hole distribution) on structural damage (for example, changes in the topological connectivity of channel lining voids occur earlier than visible cracks).

[0009] Insufficient dimensionality reduction capabilities of nonlinear features: Traditional principal component analysis (PCA) is only applicable to linear data and cannot effectively handle the coupling relationship of nonlinear features such as point cloud curvature and sensor signal spectrum, resulting in the neglect of key patterns hidden in high-dimensional data.

[0010] Lack of point cloud data-driven models: Existing deep learning models (such as CNN) rely on regularized grid input and are unable to directly process the disorder and sparsity characteristics of point clouds. They must first be converted into voxels or grids, resulting in information loss and reduced computational efficiency.

[0011] The lack of a generative adversarial model for multimodal data (point cloud + time series signal) makes it impossible to detect anomalies by simulating normal state data, especially in scenarios with unbalanced samples (such as scarce fault data), where the performance drops significantly. Summary of the Invention

[0012] In order to solve the above-mentioned problems, the present invention provides a method and system for detecting the operating status of facilities in a track connection channel.

[0013] In a first aspect, the present invention provides a method for detecting the operating status of facilities in a rail connection channel, which adopts the following technical solution:

[0014] A method for detecting the operating status of facilities in a rail communication channel, comprising:

[0015] Obtain point cloud data and sensor data of facilities within the rail connection channel;

[0016] Perform data preprocessing on the acquired point cloud data and sensor data;

[0017] Perform feature extraction and feature selection on the preprocessed data;

[0018] Build a generative adversarial network model based on deep learning;

[0019] Use the generative adversarial network model based on deep learning to simulate and analyze the extracted features;

[0020] Output the detection results.

[0021] Furthermore, the acquired point cloud data and sensor data are preprocessed, including calculating the distance between each point in the point cloud data and its neighboring points based on statistical methods to remove outliers; using a radius-based filtering method to count the number of point clouds within the radius and identify isolated points, and removing noise points through point cloud deletion; performing point cloud registration on point cloud data collected at different times through an iterative nearest point algorithm; and associating and fusing the point cloud data and sensor data by setting a unified time base and using a sensor data association mapping method.

[0022] Furthermore, the pre-processed data is subjected to feature extraction and feature selection, including geometric feature extraction by calculating the curvature features of the point cloud data, and judging the surface integrity of the facility by analyzing the consistency of the normal vector; a triangular mesh model of the point cloud is constructed using the Delaunay triangulation algorithm of the point cloud, and topological feature extraction is performed based on the triangular mesh model. For the point cloud data of the channel structure, the damage and voids of the facility structure are judged by the number of voids and the number of connected areas.

[0023] Furthermore, the feature extraction and feature selection of the preprocessed data also includes combining the point cloud features and the sensor data features into a high-order feature matrix, mapping the original features to a high-dimensional space using a kernel principal component analysis algorithm, and extracting the main feature components in the high-dimensional space; constructing an objective function, using a Bayesian optimization algorithm to search for the optimal feature subset, and using acquisition function balance exploration to determine the feature combination that is most critical for judging the operating status of the facility, thereby achieving feature selection.

[0024] Furthermore, the construction of a generative adversarial network model based on deep learning includes constructing a generator and a discriminator, wherein constructing the generator includes adopting a point cloud-based deconvolution operation to convert the feature vector back into an upsampling operation in the point cloud format, splicing the feature map after deconvolution of the point cloud and the feature map after deconvolution of the sensor data according to the channel dimension at a specific layer, and outputting simulated data corresponding to the original data, including simulated point cloud data and sensor data; constructing the discriminator includes extracting and fusing features through a point cloud-based convolution layer, downsampling and further extracting features from the fused features, and passing a LeakyReLU activation function after each convolution layer to avoid the gradient disappearance problem.

[0025] Furthermore, the construction of a generative adversarial network model based on deep learning also includes forming complete multimodal training data by dividing the data set and enhancing the sensor data, and training and optimizing the generative adversarial network model based on the reconstruction loss function of the point cloud, wherein an optimization algorithm based on natural gradient is adopted, and the batch size and the learning rate are adjusted dynamically according to the computational characteristics of point cloud data processing, and a learning rate decay strategy is adopted to enable the model to converge quickly in the early stage of training, and fine-tune the parameters in the later stage to improve the training effect.

[0026] Furthermore, the construction of the generative adversarial network model based on deep learning also includes using a deep belief network to perform unsupervised pre-training on point cloud data and sensor data before GAN training, and learning the hierarchical feature representation of the data by training RBM layer by layer; the pre-trained DBN weights are used to initialize the corresponding part of the network in the GAN, so that the GAN has an initial state when the adversarial training starts, which helps to converge faster and generate more reasonable data. At the same time, the feature representation of multimodal data after DBN pre-training is used to improve the GAN model's processing ability for multimodal data.

[0027] Furthermore, the extracted features are simulated and analyzed using a generative adversarial network model based on deep learning, including using a particle swarm optimization algorithm to adjust the GAN training strategy. Based on the comparison results between the generated data and the real data in terms of point cloud features and other sensor data features, the fitness value is calculated for each particle, and the particles are guided to move towards a better parameter combination, so that the GAN can intelligently search for the optimal parameters during the training process. In particular, when processing multimodal data including point cloud data, it can balance the training of the generator and the discriminator, generate simulated data that conforms to the actual situation, and improve the accuracy of facility operation status detection.

[0028] Furthermore, the use of a deep learning-based generative adversarial network model to simulate and analyze the extracted features also includes regularly using test set data to comprehensively evaluate the system's detection performance, calculate the accuracy, recall rate, F1 value and root mean square error index. Among them, for the point cloud data part, the point cloud matching accuracy and point cloud reconstruction error are introduced, and the confusion matrix is ​​used to deeply analyze the model's misjudgment and missed judgment in the processing of point cloud data and other sensor data; if it is found that the model performance has declined, new facility operation status data is collected, the model is incrementally learned and retrained, and the model structure and parameters are adjusted according to various types of data in the new data.

[0029] In a second aspect, a system for detecting the operating status of facilities in a rail communication channel includes:

[0030] The data acquisition module is configured to acquire facility point cloud data and sensor data within the rail communication channel;

[0031] A preprocessing module is configured to perform data preprocessing on the acquired point cloud data and sensor data;

[0032] The feature module is configured to perform feature extraction and feature selection on the preprocessed data;

[0033] The model building module is configured to build a generative adversarial network model based on deep learning;

[0034] The simulation analysis module is configured to perform simulation analysis on the extracted features using a generative adversarial network model based on deep learning;

[0035] The output module is configured to output the detection result.

[0036] In a third aspect, the present invention provides a computer-readable storage medium storing a plurality of instructions, wherein the instructions are suitable for being loaded and executed by a processor of a terminal device, for a method for detecting the operating status of facilities in a rail connection channel.

[0037] In a fourth aspect, the present invention provides a terminal device comprising a processor and a computer-readable storage medium, wherein the processor is used to implement various instructions; and the computer-readable storage medium is used to store multiple instructions, wherein the instructions are suitable for being loaded and executed by the processor to implement a method for detecting the operating status of facilities in a rail connection channel.

[0038] In summary, the present invention has the following beneficial technical effects:

[0039] This invention deploys a 3D laser scanner and a portable laser rangefinder to acquire high-precision point cloud data of rail access corridors and facilities, supplementing the 3D spatial structure information that traditional sensors struggle to capture. The millimeter-level precision of the scanning equipment clearly reveals subtle deformations in the corridor structure and changes in the facility's appearance, such as deformation of ventilation blades and the expansion of cracks in the corridor walls. Furthermore, combined with traditional sensor data such as vibration and temperature, multimodal monitoring is achieved, comprehensively reflecting the operational status of the facilities and providing a rich data foundation for accurate assessment.

[0040] The present invention optimizes the design by constructing a deep learning network model - Generative Adversarial Network (GAN) in combination with the characteristics of point cloud data. The generator can generate simulated data, and the discriminator can effectively judge the operating status of the facility by comparing real data with simulated data, and analyzing the differences between point cloud features and other sensor data features. When the discrimination result shows an abnormality, the expert system and fault tree analysis method are combined to comprehensively consider the structural information reflected by the point cloud data and the abnormal characteristics of other sensor data, and the root cause of the fault can be traced more accurately. The multi-level early warning mechanism promptly notifies the operation and maintenance personnel according to the severity of the fault and takes corresponding measures, which greatly improves the accuracy of fault diagnosis and the timeliness of early warning.

[0041] During model training, multiple optimization strategies were employed. Data enhancement, particularly geometric transformations such as rotation, translation, and scaling of point cloud data, expanded data diversity and enhanced the model's ability to recognize point cloud data in different poses and positions. The introduction of a natural gradient-based optimization algorithm, deep belief network pre-training, and particle swarm optimization algorithms improved model performance, enabling it to better balance the training of the generator and discriminator when processing multimodal data containing point clouds, generating simulated data that is more consistent with actual conditions. Furthermore, regular performance evaluation and model update mechanisms can rationally adjust the model structure and parameters based on the characteristics of new data, ensuring that the system adapts to the dynamic changes in the operating status of rail liaison channel facilities and maintains long-term stable monitoring and diagnostic capabilities. BRIEF DESCRIPTION OF THE DRAWINGS

[0042] Figure 1 This is a schematic diagram of a method for detecting the operating status of facilities in a rail connection channel according to Example 1 of the present invention. DETAILED DESCRIPTION

[0043] The present invention will be further described in detail below with reference to the accompanying drawings.

[0044] Example 1

[0045] Reference Figure 1 , this embodiment

[0046] 1. Obtaining operating status parameters

[0047] (I) Sensor Selection and Deployment (New Point Cloud Related) Point Cloud Data Acquisition Equipment Deployment: Within the rail access corridor, 3D laser scanners are installed at both ends and at a certain distance (e.g., 50 meters) in the middle to acquire point cloud data of the corridor's overall structure and facilities. Scanner selection should consider its scanning accuracy, range, and speed. For example, a device with millimeter-level accuracy, a scanning range that covers the entire corridor cross-section, and a high scanning speed should be selected to meet efficient acquisition requirements.

[0048] For some key facilities, such as ventilation equipment, drainage pumps, etc., portable laser rangefinders are added at appropriate locations around them to supplement the collection of more detailed point cloud data of local facilities, so as to more accurately monitor the changes in the shape of the facilities and the offset of the installation position.

[0049] (2) Data collection (new point cloud collection content)

[0050] Point cloud data acquisition: The 3D laser scanner automatically scans at set time intervals (such as once a day) to obtain the latest point cloud data of the channel. During the scanning process, ensure that the equipment is installed stably to avoid vibration or displacement affecting the data quality. After the scan is completed, the original point cloud data is stored on the local server. The data format uses the common PLY or LAS format to facilitate subsequent processing. Portable laser rangefinders are used by inspection personnel during regular inspections to perform manual scans on key facilities. Before scanning, the rangefinder needs to be calibrated to ensure measurement accuracy. The collected data is imported into the local server via wireless transmission or memory card and integrated with the data obtained by the 3D laser scanner.

[0051] 2. Data preprocessing (new point cloud data preprocessing step)

[0052] (1) Point cloud data cleaning

[0053] Remove outliers: Use a statistical method to calculate the distance between each point in the point cloud data and its neighboring points. Points whose distance exceeds a certain standard deviation range (such as 3 times the standard deviation) are determined to be outliers and removed. For example, for a point P, calculate the average distance d between it and the k nearest neighboring points in the neighborhood. If d is greater than the average distance of all points plus 3 times the standard deviation, then point P is considered an outlier. Use a radius-based filtering method, set a radius threshold (such as 0.1 meters) with each point as the center, and count the number of points within the radius. If the number of points is lower than a certain set value (such as 5), the point is considered an isolated point and deleted, thereby removing noise points and improving the quality of the point cloud data. Point cloud registration: For point cloud data collected at different times, the iterative closest point (ICP) algorithm is used for registration. First, the point cloud data collected for the first time is selected as the reference point cloud, and the point cloud data collected subsequently is used as the point cloud to be registered. Through continuous iterative calculations, the correspondence between the points in the point cloud to be registered and the nearest points in the reference point cloud is found. Then, the rotation and translation transformation matrices are calculated based on the corresponding point pairs, and the point cloud to be registered is transformed into a coordinate system consistent with the reference point cloud, achieving precise alignment of multi-phase point cloud data, which is convenient for subsequent comparative analysis.

[0054] (2) Time synchronization of pre-processing of point cloud data fusion with other sensor data: Since the acquisition time of point cloud data may differ from the data acquisition time of other sensors (such as vibration and temperature sensors), time synchronization processing is required. By setting a unified time base in the data acquisition system, such as using a high-precision GPS timing module, ensure that the timestamps recorded by all sensors when collecting data are accurately synchronized. For the asynchronous data that has been collected, linear interpolation or spline interpolation and other methods are used according to the timestamp information to align the point cloud data with other sensor data in the time dimension, so that the subsequent fusion analysis has time consistency. Data association: Establish the association relationship between point cloud data and other sensor data. For facilities within the rail connection channel, according to the location information of the facility in the point cloud data and the installation location record of the sensor, the part of the point cloud data corresponding to the facility is associated with the sensor data related to the facility.

[0055] By determining the spatial position of the ventilation equipment in the point cloud model, a mapping relationship is established between it and the vibration, temperature, torque and other sensor data installed on the ventilation equipment, so as to facilitate subsequent comprehensive analysis of the operating status of the facility.

[0056] 3. Feature extraction and selection (integration of point cloud features)

[0057] (1) Point cloud feature extraction Geometric feature extraction:

[0058] Calculate the curvature characteristics of the point cloud data. For each point, fit the surface of its neighboring points to calculate the curvature value at that point. Curvature reflects the degree of curvature of the point cloud surface and is important for detecting deformation of channel structures and changes in the external appearance of facilities. For example, in point cloud data of a channel wall, areas with abnormally increased curvature may indicate cracks or localized deformation. Extract the normal vector characteristics of the point cloud. Normal vectors are perpendicular to the point cloud surface and reflect the orientation of the point cloud surface. By analyzing the consistency and variation of normal vectors, the flatness and integrity of the channel structure or facility surface can be determined. In track point cloud data, significant inconsistency in the normal vector direction of a certain track segment may indicate distortion or deformation. Extract topological features: Use the Delaunay triangulation algorithm to construct a triangular mesh model of the point cloud. Based on this, extract topological features, such as the number of holes and the number of connected areas. For point cloud data of channel structures, an increase in the number of holes may indicate damage or a void in the structure; a change in the connected area may reflect a change in the connection between the facilities within the channel and the structure.

[0059] (2) Feature selection of multi-source data fusion

[0060] Feature selection based on kernel principal component analysis (KPCA): Point cloud features are combined with other sensor data features (such as time-domain and frequency-domain features like vibration and temperature) into a high-dimensional feature matrix. Using the kernel principal component analysis algorithm, by selecting an appropriate kernel function (such as the Gaussian kernel function), the original features are mapped to a high-dimensional space. Principal component analysis is then performed in this high-dimensional space to extract the main characteristic components of the data. KPCA can handle nonlinear relationships in the data and identify hidden low-dimensional structures within the data. By setting a contribution rate threshold (such as 95%), it retains the principal components that explain most of the data variance, achieving feature dimensionality reduction and selection, removing redundant features, and improving the efficiency and accuracy of subsequent model training.

[0061] Feature selection based on Bayesian optimization: Construct an objective function that takes a feature subset as input and outputs the model's performance indicators (such as accuracy, F1 value) on the validation set. Use the Bayesian optimization algorithm to search for the optimal feature subset. Bayesian optimization approximates the objective function by constructing a proxy model (such as a Gaussian process model). Based on the existing sampling point information, the position of the next sampling point is calculated so that the maximum value of the objective function (i.e., the optimal feature subset) can be found as much as possible within a limited number of sampling times. During the search process, the proxy model is continuously updated, and the acquisition function (such as the expected improvement function) is used to balance exploration and utilization, gradually determining the most critical feature combination for judging the operating status of the facility, reducing the blindness of feature selection, and improving the efficiency and quality of feature selection.

[0062] 4. Building a Deep Learning Network Model - Generative Adversarial Network (GAN)

[0063] (1) Network structure design (integration of point cloud data processing module)

[0064] Generator: In the input layer, in addition to receiving a random noise vector, the point cloud data is first encoded into a low-dimensional feature vector. A point cloud-based encoder, such as PointNet or PointNet++, can be used to convert the point cloud data into a fixed-length feature vector, which is then concatenated with the random noise vector before being input to subsequent layers. Fully connected layers map the concatenated vector into an intermediate feature space, and a series of deconvolution layers (transposed convolution layers) gradually restore the data's dimensionality and structure. During the deconvolution process, point cloud-based deconvolution operations are performed on the point cloud data, such as upsampling to convert the feature vector back into a point cloud format. Feature fusion is also performed with other modal data (such as vibration and temperature data, recovered via deconvolution). For example, at a specific layer, the deconvolved feature map of the point cloud is concatenated with the deconvolved feature maps of the other modal data along the channel dimension. Finally, the output is simulated data corresponding to the original multimodal data, including simulated point cloud data and other simulated sensor data. The output layer uses a suitable activation function for point cloud data (such as Tanh, which maps the point cloud coordinate values ​​to a reasonable range), and uses corresponding activation functions (such as Tanh or Sigmoid) for other data.

[0065] Discriminator: The input layer receives real multimodal data (including real point cloud data and other sensor data) or simulated multimodal data generated by the generator.

[0066] For point cloud data, features are first extracted through point cloud-based convolutional layers (such as the convolution operations in PointNet or PointNet++). The point cloud data is converted into feature maps and fused with the feature maps extracted by the convolutional layers from other modalities (e.g., element-wise addition or channel concatenation). A series of convolutional layers are then used to downsample and further extract features from the fused features. Each convolution layer is activated with a LeakyReLU activation function with a slope of 0.2 to prevent the vanishing gradient problem. Following the convolutional layer, the extracted features are mapped to a scalar value through a fully connected layer. The Sigmoid activation function is used to output a probability value between 0 and 1, indicating the probability that the input data is real data.

[0067] (2) Model training (considering the characteristics of point cloud data)

[0068] Data division and enhancement (new point cloud data enhancement): For point cloud data, in terms of data enhancement, geometric transformation operations such as rotation, translation, and scaling are used to generate new point cloud data samples. For example, the point cloud data of the training set is randomly rotated (rotation angle within the range of ±30 degrees), translated (translation distance within the range of ±0.1 meters), and scaled (scaling ratio between 0.8-1.2) with a certain probability (such as 0.5) to expand the diversity of point cloud data and improve the model's ability to recognize point cloud data of different postures and positions. At the same time, other sensor data (such as time domain and frequency domain data) continue to be enhanced by translation, scaling, noise addition, etc., and are combined with the enhanced samples of point cloud data to form complete multimodal training data.

[0069] Loss function and optimization algorithm (adjusted in combination with point cloud data characteristics): In terms of the loss function, in addition to the original adversarial loss, a point cloud-based reconstruction loss is introduced for the point cloud data part.

[0070] The differences between the generated simulated point cloud data and the real point cloud data in terms of point positions, normal vectors, and other features are calculated, using the mean squared error (MSE) or Chamfer distance as the reconstruction loss metric. The reconstruction loss and adversarial loss are weighted and summed to obtain the final loss function. The weights can be adjusted based on experimental results (e.g., a weight of 0.8 for the adversarial loss and a weight of 0.2 for the reconstruction loss). The optimization algorithm uses a natural gradient-based optimization algorithm, such as Natural Gradient Descent.

[0071] The natural gradient takes into account the geometric structure of the parameter space, enabling more efficient parameter updates along the direction of fastest function descent. Compared to traditional gradient descent algorithms, this approach may achieve faster convergence and better performance when processing complex models (such as GAN models that involve point cloud data processing). During training, the batch size is appropriately adjusted (e.g., appropriately reducing the batch size to accommodate the memory requirements of point cloud data) based on the computationally intensive nature of point cloud data processing. The learning rate is dynamically adjusted, employing learning rate decay strategies such as exponential decay and cosine annealing. This allows for rapid model convergence in the early stages of training, allowing for fine-tuning of parameters later to improve training effectiveness. The learning rates for the generator and discriminator can be set separately: 0.0002 for the generator and 0.0001 for the discriminator, with the β1 parameter set to 0.5.

[0072] Combined with Deep Belief Network (DBN) for unsupervised pre-training (considering point cloud data):

[0073] Before GAN training, unsupervised pre-training is performed using a deep belief network for point cloud data and other sensor data. A deep belief network is composed of multiple stacked restricted Boltzmann machines (RBMs). By training the RBMs layer by layer, it can automatically learn the hierarchical feature representation of the data. For point cloud data, the point cloud data is used as input to train the RBMs of each layer of the DBN to learn the intrinsic structure and features of the point cloud data. For other sensor data, the corresponding data is also input into the DBN for pre-training. The pre-trained DBN weights can initialize the corresponding parts of the GAN network (such as the initial layers of the generator and discriminator), so that the GAN has a better initial state when starting adversarial training, which helps to converge faster and generate more reasonable data. At the same time, the feature representation of multimodal data after DBN pre-training is used to improve the GAN model's ability to process multimodal data.

[0074] Introducing particle swarm optimization (PSO) to adjust the GAN training strategy (combined with point cloud data):

[0075] The GAN training process is considered an optimization problem, where the parameters of the generator and discriminator are the optimization variables. The particle swarm optimization algorithm is used to adjust the GAN training strategy. The particle swarm optimization algorithm simulates the foraging behavior of a flock of birds. Each particle represents a combination of generator and discriminator parameters. The particles fly through the solution space, continuously updating their positions and velocities to find the optimal parameter combination to maximize the discriminator's discriminative ability and the quality of the generator's generated data.

[0076] During the optimization process, a fitness value is calculated for each particle based on the comparison of generated data with real data in terms of point cloud features (such as curvature, normal vectors, topological features, etc.) and other sensor data features, guiding the particles towards a more optimal parameter combination. This combination enables GAN to more intelligently search for optimal parameters during training, improving model performance. This is particularly true when processing multimodal data including point clouds, enabling better balance between generator and discriminator training, generating simulated data that better reflects actual conditions, and improving the accuracy of facility operation status detection.

[0077] 5. Data processing and test result generation.

[0078] (1) Real-time data processing (adding point cloud real-time processing flow) The operating status data of facilities in the rail communication channel collected in real time, including point cloud data and other sensor data, are processed in sequence through the steps of data cleaning, normalization, feature extraction and selection to form a format suitable for input into the generative adversarial network model.

[0079] For point cloud data, preprocessing operations such as outlier removal and point cloud registration are performed in real time to extract geometric and topological features. For other sensor data, corresponding time domain, frequency domain, and time-frequency domain feature extraction and selection are completed. During the feature extraction process, kernel principal component analysis and Bayesian optimization algorithms are used to select features from multi-source data and identify key features.

[0080] The processed data is fed into a trained GAN model. The generator generates simulated data (including simulated point cloud data and other simulated sensor data), and the discriminator distinguishes between real and simulated data. The operational status of the facility is determined by comparing the discriminator's output and the generated data with real data based on point cloud features (such as curvature, normal vectors, topology, etc.) and other sensor data features.

[0081] If the discriminator's probability of judging the real point cloud data is much higher than that of the generated point cloud data, and the generated point cloud data differs significantly from the real point cloud data in key geometric and topological features, and other sensor data also show anomalies, then the comprehensive judgment is that the facility's operating status is abnormal. During the model training and judgment process, an optimization algorithm based on natural gradient, deep belief network pre-training, and particle swarm optimization are used to improve model performance and detection accuracy.

[0082] (2) Fault diagnosis and early warning (fault analysis combined with point cloud data)

[0083] When the discriminator output results or the comparative analysis of generated data with real data shows that the facility may be in an abnormal state, the expert system and fault tree analysis (FTA) method are combined to trace the root cause of the fault not only based on the abnormal characteristics of different sensor data and the logical relationship between them, but also considering the information reflected by the point cloud data.

[0084] If the point cloud data shows a significant change in the ventilation equipment's shape (e.g., blade deformation causing a change in the blade shape in the point cloud model), and if the ventilation equipment's vibration, torque, and other sensor data also show anomalies, fault tree analysis can identify a series of faults likely caused by blade damage. During the fault diagnosis process, the geometric and topological features of the point cloud data, along with other sensor data features, are fully utilized, combined with the results of model training using multiple algorithms, to improve the accuracy of fault diagnosis.

[0085] A multi-tiered early warning mechanism is developed based on the severity and scope of the fault. Minor faults are notified to maintenance personnel via text messages and in-site messaging, and recorded in the equipment maintenance log. Moderate faults, in addition to notifying personnel, activate backup equipment for switching, and monitor the faulty equipment in real time. Severe faults immediately trigger audible and visual alarms, send emergency notifications to relevant departments, and initiate emergency response plans, such as evacuating personnel from corridors and shutting down power to relevant areas. During the early warning process, visualization models generated from point cloud data can be presented to maintenance personnel along with abnormal information from other sensor data, helping them to more intuitively understand the fault situation.

[0086] (3) Performance Evaluation and Model Update (Considering the Impact of Point Cloud Data on Performance): Regularly (e.g., weekly or monthly) use the test set data to comprehensively evaluate the system's detection performance, calculating metrics such as accuracy, recall, F1 score, and root mean square error (RMSE, used for regression tasks). Furthermore, for point cloud data, specialized evaluation metrics are introduced, including cloud matching accuracy (calculating the accuracy of key point matching between the generated point cloud and the ground truth) and point cloud reconstruction error (measuring the geometric differences between the generated point cloud and the ground truth). Confusion matrix analysis visually demonstrates the model's classification performance across different categories, allowing for in-depth analysis of the model's misclassifications and omissions in processing point cloud data and other sensor data.

[0087] In the performance evaluation, the impact of multiple algorithms on model performance is comprehensively considered, and the contribution of each algorithm in different data types and tasks is analyzed. If the model performance is found to have degraded, such as the accuracy rate is lower than the set threshold (such as 90%), the model update process is automatically triggered. New facility operation status data is collected, including point cloud data and other sensor data, and data preprocessing, feature extraction, model training and other steps are repeated to perform incremental learning or retraining of the model. During the model training process, the model structure and parameters are reasonably adjusted according to the characteristics of various types of data (point cloud data and other data) in the new data, such as adjusting the number and parameters of convolutional layers in the point cloud data processing module, as well as multimodal data fusion.

[0088] Specifically, the following steps are included:

[0089] 1. Obtaining operating status parameters

[0090] (1) Sensor selection and deployment

[0091] 1. 3D laser scanner deployment

[0092] Position planning: Install fixed 3D laser scanners at both ends and every 50 meters in the middle of the rail connection channel to ensure that the equipment is stable and vibration-free (if installed on a fixed bracket on the channel wall, the bracket should be made of shock-absorbing material).

[0093] Equipment selection: Select equipment with millimeter-level accuracy (such as ±1mm), a scanning range covering the channel cross-section (radius ≥5 meters), and a scanning speed ≥100,000 points / second (such as FARO Focus 3D).

[0094] 2. Deployment of portable laser rangefinders

[0095] Key facility coverage: Within a 0.5-meter radius around ventilation equipment, drainage pumps, and other facilities, a handheld or portable rangefinder (such as the Leica Disto X3) is used to ensure that the scanning angle covers the entire shape of the facility (e.g., a 360-degree scan around the facility).

[0096] (2) Data Collection

[0097] 1. Automatic scanning (3D laser scanner)

[0098] Scheduled task: Automatic scanning is triggered at 0:00 every day, and the single scanning time is ≤ 10 minutes. During the scanning process, mobile devices in the channel are prohibited to avoid vibration.

[0099] Data storage: The original point cloud data is stored in PLY format with the naming convention of "channel ID_scanner position_acquisition time.ply" and stored in a dedicated folder on the local server (automatically classified by timestamp).

[0100] 2. Manual scanning (portable rangefinder)

[0101] Inspection linkage: Inspection personnel carry equipment for inspection every 7 days. Before scanning, they use a standard cube (with known side length) to calibrate the rangefinder to ensure that the error is ≤0.5mm.

[0102] Data import: Import data to the server via WiFi (such as IEEE 802.11n) or TF card, and create an associated folder with the scanner data based on the facility ID (such as ventilation equipment number V-01).

[0103] 2. Data Preprocessing

[0104] (1) Point cloud data cleaning

[0105] 1. Remove outliers (statistical filtering + radius filtering)

[0106] Step 1: Statistical filtering

[0107] For each point P, calculate the Euclidean distance between it and its k=20 nearest neighbor points and obtain the average distance dp.

[0108] Calculate the average distance μ and standard deviation σ of all points, and delete points with dp>μ+3σ.

[0109] Step 2: Radius Filter

[0110] With each point as the center, set the radius r = 0.1 meter and count the number of points in the neighborhood.

[0111] Delete isolated points with less than 5 neighborhood points.

[0112] 2. Point cloud registration (ICP algorithm)

[0113] Step 1: Reference point cloud setting**: Select the first collected point cloud \(P_{ref}\) as the reference, and the subsequent point cloud \(P_{curr}\) is to be registered.

[0114] Step 2: Iterative alignment**:

[0115] For each point in Pcurr, find the nearest point in Pref to form a point pair (pi,qi).

[0116] 2. Calculate the center of mass Constructing the covariance matrix

[0117]

[0118] 3. Solve the rotation matrix R = VU T (where H=U∑V T is SVD decomposition), translation vector

[0119] 4. Apply R,t to Pcurr and repeat until the mean square error (MSE) converges (e.g., iterations ≤ 50 or MSE < 1e-4).

[0120] (2) Multi-source data fusion preprocessing

[0121] 1. Time synchronization

[0122] Hardware synchronization: All sensors are connected to the GPS timing module (accuracy ±1μs), and UTC timestamps are recorded during collection.

[0123] Interpolation alignment: For asynchronous data (such as point cloud acquisition every 24 hours and vibration sensor acquisition at 100 Hz per second), cubic spline interpolation is used to align the timestamps of other sensor data based on the point cloud acquisition time.

[0124] 2. Data Association

[0125] Spatial mapping: Establish a facility location ledger (including 3D coordinates and sensor installation coordinates), and bind the facility area in the point cloud (such as the point cloud clustering results of ventilation equipment) to the corresponding sensor (such as vibration sensor V-01-acc) ID through spatial coordinate matching (error ≤ 0.05 meters).

[0126] 3. Feature Extraction and Selection

[0127] (1) Point cloud feature extraction

[0128] 1. Geometric features

[0129] Curvature calculation:

[0130] For each point pi, take k=50 neighboring points to fit the quadratic surface ax 2 +by 2 +cz 2 +dxy+eyz+fxz+gx+hy+iz+j=0, calculate the Gaussian curvature K and the mean curvature H.

[0131] Normal vector calculation:

[0132] The covariance matrix is ​​constructed for k=30 neighborhood points, and the eigenvector corresponding to the minimum eigenvalue is the normal vector, which is normalized and used as the normal vector feature of point pi.

[0133] 2. Topological features

[0134] Delaunay triangulation: Project the point cloud onto the XY plane to construct a two-dimensional Delaunay triangulation network, which is then extended to three dimensions to generate a tetrahedral mesh.

[0135] Hole detection: Count the number of boundary edges (edges belonging to only one tetrahedron) in the grid and calculate the hole area (the area enclosed by the boundary edges is greater than 0.1㎡ and is considered a valid hole).

[0136] (2) Feature selection of multi-source data fusion

[0137] 1. KPCA feature dimensionality reduction

[0138] Step 1: Feature concatenation: Merge the point cloud curvature, normal vector, number of holes, root mean square (RMS) of the vibration signal, and mean value of the temperature signal into a high-dimensional vector X (dimension ≤ 200).

[0139] Step 2: Kernel Mapping: Select Gaussian Kernel

[0140] k(x i , x j )=exp(-γ||x i -x j || 2 )(γ=0.1), mapping X to a high-dimensional space.

[0141] Step 3: Principal component extraction: Calculate the eigenvalues ​​of the kernel matrix and retain the principal components (first 20 dimensions) with cumulative contribution rates ≥ 95%.

[0142] 2. Bayesian Optimization Feature Selection

[0143] Step 1: Objective function definition: The F1 value of the random forest classifier on the validation set is the target, and the feature subset is input

[0144] Step 2: Surrogate model construction: Use Gaussian process regression (GPR) to fit the objective function and initially sample 5 sets of random feature subsets.

[0145] Step 3: Iterative search: Select the next set of feature subsets through the expected improvement (EI) acquisition function, iterate 20 times, and retain the subset with the highest F1 value (such as containing 15 key features).

[0146] 4. Building a Generative Adversarial Network (GAN) Model

[0147] (1) Network structure design

[0148] 1. Generator (PointNet + Deconvolution)

[0149] Input layer: The point cloud data (N×3) is encoded into a 128-dimensional feature vector by PointNet and concatenated with 100-dimensional random noise to form a 228-dimensional vector.

[0150] Middle layer: Mapped to the intermediate space through three fully connected layers (256→512→1024), and then through three deconvolution layers (1024→512→256), combined with upsampling (such as farthest point sampling + linear interpolation) to restore the point cloud resolution.

[0151] Output layer: Generates point clouds (N×3, Tanh activation mapped to [-1, 1]) and other sensor data (such as vibration signals, Sigmoid activation mapped to [0, 1]).

[0152] 2. Discriminator (PointNet + fused convolution)

[0153] Input layer: The real / generated point cloud (N×3) is extracted with 128-dimensional features by PointNet and concatenated with other sensor data (10-dimensional temporal features) into a 138-dimensional vector.

[0154] Middle layer: 3 convolutional layers (128→256→512, LeakyReLU activation, slope 0.2) downsampling, and fully connected layers mapped to 1D probability values ​​(Sigmoid activation).

[0155] (2) Model training

[0156] 1. Data Augmentation

[0157] Point cloud enhancement: The training point cloud is randomly rotated (±30° around the Z axis), translated (±0.1 meters), and scaled (0.8-1.2 times) with a probability of 50%, keeping the point order unchanged.

[0158] Sensor enhancement: Gaussian noise (standard deviation 0.05) was added to the vibration data, and the temperature data was shifted by ±2°C.

[0159] 2. Loss Function and Optimization

[0160] Loss function: L = L adv +λL rec

[0161] Among them, adversarial loss:

[0162]

[0163] , reconstruction loss L rec =ChamferDistance(G(z),x), weight λ=0.2.

[0164] Optimization algorithm: Natural Gradient Descent (NGD), generator learning rate 0.0002, discriminator 0.0001, β1 = 0.5, batch size 32 (dynamically adjusted to 16-32 due to the large memory usage of point cloud).

[0165] 3. Pre-training and parameter optimization

[0166] DBN pre-training: The point cloud (N×3) is input into the first layer of RBM (784 units in the visible layer and 256 units in the hidden layer), and trained layer by layer to 3 layers, and the weights are initialized to the generator encoder.

[0167] PSO parameter adjustment**: The particle dimension is the key parameter of the generator / discriminator (such as the number of convolution kernels), the fitness function is the Chamfer distance on the validation set + the discrimination accuracy, and the parameters are updated after 50 iterations.

[0168] 5. Data processing and test result generation

[0169] (1) Real-time data processing

[0170] 1. Preprocessing pipeline

[0171] Point cloud: Real-time outlier removal (statistics + radius filtering) → ICP registration (aligned with the reference point cloud from the previous 24 hours) → extraction of curvature, normal vector, and number of holes.

[0172] Sensor: FFT of vibration signals is used to extract frequency domain features (such as the energy ratio between 10 and 50 Hz), and temperature signals are normalized (0-1).

[0173] 2. Model Inference

[0174] The preprocessed feature vector (10-dimensional point cloud features + 8-dimensional sensor features) is input, the generator outputs simulated data, and the discriminator outputs a probability value \p.

[0175] Anomaly determination: If \(p<0.3\) and the Chamfer distance between the generated point cloud and the real point cloud is greater than 0.05 meters, it is marked as a suspected anomaly.

[0176] (2) Fault diagnosis and early warning

[0177] 1. Multi-layer diagnosis

[0178] Preliminary judgment: Based on the discriminator probability and feature differences, the expert system rules are triggered (such as "the number of holes in the ventilation equipment point cloud is greater than 3 and the vibration RMS is greater than 0.8g → the blade is damaged").

[0179] Fault tree analysis: Trace back from abnormal features, such as "blade deformation → abnormal point cloud curvature → increase in high-frequency components of vibration signals", and match fault tree nodes.

[0180] 2. Early warning mechanism

[0181] Minor (Level 1): The inspection personnel will be notified via SMS and the error will be recorded in the log (e.g., equipment displacement of 0.5 mm).

[0182] Moderate (Level 2): ​​Switch to the backup device, start high-frequency scanning every 10 minutes, and push real-time point cloud comparison images to the app.

[0183] Serious (Level 3): Trigger channel audio and visual alarms, link the SCADA system to power off, and send an email to the emergency department (if the point cloud shows a structural crack ≥10cm).

[0184] (3) Performance evaluation and model update

[0185] 1. Evaluation Metrics

[0186] Common indicators: accuracy, recall, and F1 value (for fault classification).

[0187] Point cloud-specific indicators: point cloud matching accuracy (number of key point matches / total key points × 100%), reconstruction error (root mean square distance, RMSE < 0.03 meters is qualified).

[0188] 2. Model Update

[0189] Trigger condition: F1 value < 85% or RMSE > 0.05 meters for 3 consecutive days.

[0190] Update process: Collect new data from the past month → Redo point cloud cleaning / registration → Update feature subset using KPCA+Bayesian optimization → Incremental training of GAN (freeze the first 50% of network layers and fine-tune the last 50%), taking ≤ 4 hours.

[0191] Through the above steps, a complete process from point cloud acquisition to fault warning is formed, ensuring the efficiency and accuracy of the operation status monitoring of the rail connection channel facilities.

[0192] This embodiment provides a system for detecting the operating status of facilities in a rail communication channel, including:

[0193] The data acquisition module is configured as

[0194] A computer-readable storage medium stores a plurality of instructions, wherein the instructions are suitable for being loaded and executed by a processor of a terminal device for a method for detecting the operating status of facilities in a track connection channel.

[0195] A terminal device includes a processor and a computer-readable storage medium, wherein the processor is used to implement various instructions; the computer-readable storage medium is used to store multiple instructions, and the instructions are suitable for being loaded and executed by the processor to describe a method for detecting the operating status of facilities in a rail connection channel.

[0196] The above are all preferred embodiments of the present invention, and are not intended to limit the scope of protection of the present invention. Therefore, any equivalent changes made based on the structure, shape, and principle of the present invention should be included in the scope of protection of the present invention.

Claims

1. A method for detecting the operating status of facilities in a track communication channel, characterized in that: include: Obtain point cloud data and sensor data of facilities within the rail connection channel; Perform data preprocessing on the acquired point cloud data and sensor data; Perform feature extraction and feature selection on the preprocessed data; Build a generative adversarial network model based on deep learning; Use the generative adversarial network model based on deep learning to simulate and analyze the extracted features; Output the detection results.

2. A method for detecting the operating status of facilities in a track connection channel according to claim 1, characterized in that: The data preprocessing of the acquired point cloud data and sensor data includes calculating the distance between each point in the point cloud data and its neighboring points based on a statistical method to remove outliers; using a radius-based filtering method to count the number of point clouds within the radius and identify isolated points, and removing noise points through point cloud deletion; performing point cloud registration on point cloud data collected at different times through an iterative nearest point algorithm; and associating and fusing the point cloud data and sensor data by setting a unified time base and using a sensor data association mapping method.

3. A method for detecting the operating status of facilities in a track connection channel according to claim 2, characterized in that: The feature extraction and feature selection of the preprocessed data include extracting geometric features by calculating the curvature features of the point cloud data, and judging the surface integrity of the facility by analyzing the consistency of the normal vector; constructing a triangular mesh model of the point cloud using the Delaunay triangulation algorithm of the point cloud, and extracting topological features based on the triangular mesh model. For the point cloud data of the channel structure, the damage and voids of the facility structure are judged by the number of voids and the number of connected areas.

4. A method for detecting the operating status of facilities in a track connection channel according to claim 3, characterized in that: The feature extraction and feature selection of the preprocessed data also includes combining point cloud features and sensor data features into a high-order feature matrix, mapping the original features to a high-dimensional space using a kernel principal component analysis algorithm, and extracting main feature components in the high-dimensional space; constructing an objective function, using a Bayesian optimization algorithm to search for the optimal feature subset, and using acquisition function balance exploration to determine the feature combination that is most critical for judging the operating status of the facility, thereby achieving feature selection.

5. A method for detecting the operating status of facilities in a track connection channel according to claim 4, characterized in that: The construction of a generative adversarial network model based on deep learning includes constructing a generator and a discriminator, wherein constructing the generator includes adopting a point cloud-based deconvolution operation, converting the feature vector back into an upsampling operation in a point cloud format, splicing the feature map after deconvolution of the point cloud and the feature map after deconvolution of the sensor data according to the channel dimension at a specific layer, and outputting simulated data corresponding to the original data, including simulated point cloud data and sensor data; constructing the discriminator includes extracting and fusing features through a point cloud-based convolution layer, downsampling and further extracting features from the fused features, and using a LeakyReLU activation function after each convolution layer to avoid the gradient disappearance problem.

6. A method for detecting the operating status of facilities in a track connection channel according to claim 5, characterized in that: The construction of the generative adversarial network model based on deep learning also includes forming complete multimodal training data by dividing the data set and enhancing the sensor data, and training and optimizing the generative adversarial network model based on the reconstruction loss function of the point cloud. Specifically, an optimization algorithm based on natural gradient is adopted, and the batch size and the learning rate are dynamically adjusted according to the computational characteristics of point cloud data processing. A learning rate decay strategy is adopted to enable the model to converge quickly in the early stage of training, and the parameters are fine-tuned in the later stage to improve the training effect.

7. A method for detecting the operating status of facilities in a track connection channel according to claim 6, characterized in that: The construction of the generative adversarial network model based on deep learning also includes performing unsupervised pre-training on point cloud data and sensor data using a deep belief network before GAN training, and learning the hierarchical feature representation of the data by training the RBM layer by layer; the pre-trained DBN weights are used to initialize the corresponding part of the network in the GAN, so that the GAN has an initial state when the adversarial training begins, which helps to converge faster and generate more reasonable data. At the same time, the feature representation of multimodal data after DBN pre-training is used to improve the GAN model's processing ability for multimodal data.

8. A method for detecting the operating status of facilities in a track connection channel according to claim 7, characterized in that: The extracted features are simulated and analyzed using a deep learning-based generative adversarial network model, including using a particle swarm optimization algorithm to adjust the GAN training strategy. Based on the comparison results between the generated data and the real data in terms of point cloud features and other sensor data features, a fitness value is calculated for each particle, guiding the particles to move towards a more optimal parameter combination. This allows the GAN to intelligently search for optimal parameters during training. In particular, when processing multimodal data including point cloud data, the training of the generator and the discriminator can be balanced to generate simulated data that conforms to the actual situation, thereby improving the accuracy of facility operation status detection.

9. A method for detecting the operating status of facilities in a track connection channel according to claim 8, characterized in that: The method uses a generative adversarial network model based on deep learning to simulate and analyze the extracted features, and also includes regularly using test set data to comprehensively evaluate the detection performance of the system, calculate the accuracy, recall rate, F1 value and root mean square error index. Among them, for the point cloud data part, the point cloud matching accuracy and point cloud reconstruction error are introduced, and the confusion matrix is ​​used to deeply analyze the model's misjudgment and missed judgment in the processing of point cloud data and other sensor data; if it is found that the model performance has declined, new facility operation status data is collected, the model is incrementally learned and retrained, and the model structure and parameters are adjusted according to various types of data in the new data.

10. A system for detecting the operating status of facilities in a track connection channel, characterized in that: include: The data acquisition module is configured to acquire point cloud data and sensor data of facilities in the rail communication channel; A preprocessing module is configured to perform data preprocessing on the acquired point cloud data and sensor data; The feature module is configured to perform feature extraction and feature selection on the preprocessed data; The model building module is configured to build a generative adversarial network model based on deep learning; The simulation analysis module is configured to perform simulation analysis on the extracted features using a generative adversarial network model based on deep learning; The output module is configured to output the detection result.

Images