Railway vehicle operation abnormity judgment method, system and equipment based on tunnel section multi-source information fusion

Abstract

The invention relates to the technical field of railway vehicles, in particular to a railway vehicle operation abnormity judgment method, system and equipment based on tunnel section multi-source information fusion, and the method comprises the steps of data acquisition, data preprocessing, image data feature extraction, information fusion, abnormity judgment model establishment and alarm threshold setting. According to the railway vehicle operation abnormity judgment method based on tunnel section multi-source information fusion, a mathematical model for vehicle operation abnormity judgment is established, machine learning methods such as a support vector machine (SVM) and a neural network are adopted, and a large amount of normal and abnormal operation data are trained, so that the model can learn feature modes of normal and abnormal states; when new data is input, the model can judge whether the vehicle runs abnormally according to the learned mode, the comprehensive state of the vehicle in the running process can be comprehensively reflected in real time, and meanwhile abnormal information can be sent to a mobile terminal of a maintainer so that maintenance measures can be taken in time.

Description

Technical Field

[0001] The present invention relates to the technical field of rail vehicles, and in particular to a method, system and equipment for judging abnormal operation of rail vehicles based on multi-source information fusion of tunnel sections. Background Art

[0002] With the rapid development of rail transit, the safe operation of rail vehicles has become a critical issue. During the operation of rail vehicles, a variety of abnormal situations may occur, such as vehicle derailment, wheel scratches, vehicle structure deformation, etc. If these abnormal situations are not discovered and handled in time, they will seriously endanger the safety of passengers and the normal operation of the rail transit system.

[0003] At present, the detection methods of rail vehicle operation anomalies mainly include detection methods based on vehicle-mounted sensors and detection methods based on trackside detection equipment. Although the detection method based on vehicle-mounted sensors can directly obtain the vehicle's own operating status information, it has limited information on the interaction between the vehicle and the track and the geometric state of the track. The detection methods based on trackside detection equipment, such as track flaw detectors and track geometry state measuring instruments, can detect some parameters of the track, but it is difficult to fully and real-time reflect the comprehensive state of the vehicle during operation. Summary of the invention

[0004] The purpose of the present invention is to provide an intelligent monitoring device, method and means for urban rail transit management departments, and to provide a layout plan for roadside equipment in track tunnels, so as to assist operation and management personnel to quickly judge the vehicle status and ensure safety.

[0005] To achieve the above object, the present invention provides the following technical solutions:

[0006] The rail vehicle operation abnormality judgment system based on tunnel section multi-source information fusion includes a data acquisition and processing module, a feature extraction module, an information fusion module, a judgment module and an early warning module. The data acquisition and processing module includes a laser radar sensor, a high-definition camera, an acceleration sensor and a displacement sensor. The data acquisition and processing is used to obtain the three-dimensional contour information of the tunnel section, collect the visual image inside the tunnel, and the vibration acceleration situation;

[0007] The feature extraction module uses image recognition technology to extract the texture features of the tunnel wall surface, uses a grayscale co-occurrence matrix to describe the texture, characterizes the texture changes by calculating parameters such as energy, contrast, and correlation in the matrix, and extracts the shape features of the object in the image;

[0008] The information fusion module fuses the raw data of lidar, images and other sensors at an early stage, and fuses the features extracted from different sensor data;

[0009] The judgment module establishes a mathematical model for judging abnormal vehicle operation, and judges whether the vehicle operation is abnormal based on the learned pattern by training a large amount of normal and abnormal operation data;

[0010] The early warning module sets a reasonable alarm threshold based on the fused features and judgment results. If a certain threshold is exceeded, it is judged as abnormal and an alarm signal is issued.

[0011] The method for judging abnormal operation of rail vehicles based on multi-source information fusion of tunnel sections includes the following steps:

[0012] S1: Data collection:

[0013] Installing a laser radar sensor in the tunnel can obtain the three-dimensional profile information of the tunnel section with high precision. The laser radar scans the tunnel wall at a certain scanning frequency to obtain a series of point cloud data. These point cloud data can accurately represent the spatial position of the tunnel wall at a certain moment.

[0014] S2: Data preprocessing:

[0015] Preprocessing of LiDAR point cloud data: Since LiDAR will be disturbed by environmental noise during the acquisition process, a filtering algorithm is needed to remove these noise points. A statistical filtering algorithm can be used to remove outliers based on the statistical characteristics of the distance distribution of points in the point cloud.

[0016] S3: Image data feature extraction:

[0017] The texture features of the tunnel wall surface are extracted using image recognition technology, and the gray-level co-occurrence matrix (GLCM) is used to describe the texture. The energy, contrast, correlation and other parameters in the matrix are calculated to characterize the texture changes.

[0018] Extract the shape features of objects in the image. For possible foreign objects, use edge detection algorithms to extract the edge contours of the objects, and then calculate the shape parameters of the contours to determine the nature of the objects.

[0019] S4: Information Fusion:

[0020] Fuse the raw data of lidar, images and other sensors at an early stage, fuse the features extracted from different sensor data, and use multiple decision algorithms to make comprehensive judgments on the fused information;

[0021] S5: Establish an abnormality judgment model:

[0022] Establish a mathematical model for judging abnormal vehicle operation, and use machine learning methods such as support vector machine (SVM) and neural network. Through training on a large amount of normal and abnormal operation data, the model can learn the characteristic patterns of normal and abnormal states. When new data is input, the model can judge whether the vehicle operation is abnormal based on the learned patterns.

[0023] S6: Set the alarm threshold:

[0024] According to the fused features and judgment results, a reasonable alarm threshold is set. For the vehicle vibration acceleration amplitude, if it exceeds a certain threshold, it is judged as abnormal and an alarm signal is issued. The alarm signal can be displayed in the tunnel control center through the sound and light alarm. At the same time, the abnormal information can be sent to the maintenance personnel's mobile terminal so that maintenance measures can be taken in time.

[0025] Preferably, in S1, a high-definition camera is installed in the tunnel to collect visual images inside the tunnel. The camera is installed at a position that can ensure coverage of key tunnel section areas.

[0026] Preferably, in S1, an acceleration sensor and a displacement sensor are installed in the tunnel at the same time, and the acceleration sensor is installed on the track or the vehicle to detect the vibration acceleration of the vehicle.

[0027] Preferably, in data preprocessing in S2, when the vehicle moves, point cloud data collected at different times need to be registered to establish a unified coordinate system. Feature-based registration is to first extract the features of the point cloud and then achieve registration through feature matching.

[0028] Preferably, in S2, a median filtering method may be used during data preprocessing. The median filtering method has a good effect on removing salt and pepper noise, and it replaces the current pixel value by calculating the median value in the pixel neighborhood.

[0029] Preferably, the acceleration sensor data needs to be filtered to remove high-frequency noise interference, and a low-pass filter can be used to only allow signals below a certain frequency to pass.

[0030] The equipment for judging abnormal operation of railway vehicles based on multi-source information fusion of tunnel sections comprises a memory, a processor and a computer program stored in the memory and executable on the processor, and is characterized in that when the processor executes the computer program, the steps of the method for judging abnormal operation of railway vehicles based on multi-source information fusion of tunnel sections are implemented.

[0031] Compared with the prior art, the present invention has the following beneficial effects:

[0032] The rail vehicle operation abnormality judgment method, system and equipment based on tunnel section multi-source information fusion establishes a mathematical model for vehicle operation abnormality judgment, adopts machine learning methods such as support vector machine (SVM) and neural network, and trains a large amount of normal and abnormal operation data. The model can learn the characteristic patterns of normal and abnormal states. When new data is input, the model can judge whether the vehicle operation is abnormal based on the learned pattern, and can comprehensively and real-time reflect the comprehensive state of the vehicle during operation. BRIEF DESCRIPTION OF THE DRAWINGS

[0033] In order to more clearly illustrate the technical solutions in the embodiments of the present invention, the accompanying drawings required for use in the description of the embodiments will be briefly introduced below. Obviously, the accompanying drawings in the following description are only some embodiments of the present invention. For ordinary technicians in this field, other accompanying drawings can be obtained based on these accompanying drawings without paying creative labor.

[0034] Figure 1 This is a flow chart of the rail vehicle operation abnormality judgment system of the present invention. DETAILED DESCRIPTION

[0035] The following will be combined with the drawings in the embodiments of the present invention to clearly and completely describe the technical solutions in the embodiments of the present invention. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without creative work are within the scope of protection of the present invention.

[0036] See also Figure 1 The rail vehicle operation abnormality judgment system based on tunnel section multi-source information fusion includes a data acquisition and processing module, a feature extraction module, an information fusion module, a judgment module and an early warning module. The data acquisition and processing module includes a laser radar sensor, a high-definition camera, an acceleration sensor and a displacement sensor. The data acquisition and processing is used to obtain the three-dimensional contour information of the tunnel section, collect the visual image inside the tunnel, and the vibration acceleration situation;

[0037] The feature extraction module uses image recognition technology to extract the texture features of the tunnel wall surface, adopts gray-level co-occurrence matrix to describe the texture, characterizes the texture changes by calculating parameters such as energy, contrast, and correlation in the matrix, and extracts the shape features of objects in the image;

[0038] The information fusion module fuses the raw data of lidar, images and other sensors at an early stage and fuses the features extracted from different sensor data;

[0039] The judgment module establishes a mathematical model for judging abnormal vehicle operation. By training a large amount of normal and abnormal operation data, it judges whether the vehicle operation is abnormal based on the learned pattern.

[0040] The early warning module sets a reasonable alarm threshold based on the fused features and judgment results. If it exceeds a certain threshold, it is judged as abnormal and an alarm signal is issued.

[0041] The method for judging abnormal operation of rail vehicles based on multi-source information fusion of tunnel sections includes the following steps:

[0042] S1: Data collection:

[0043] Installing a laser radar sensor in the tunnel can obtain the three-dimensional profile information of the tunnel section with high precision. The laser radar scans the tunnel wall at a certain scanning frequency to obtain a series of point cloud data. These point cloud data can accurately represent the spatial position of the tunnel wall at a certain moment.

[0044] At the same time, high-definition cameras are installed in the tunnel to collect visual images inside the tunnel. The camera installation position must ensure that it can cover the key tunnel section area.

[0045] At the same time, acceleration sensors and displacement sensors are installed in the tunnel. The acceleration sensors are installed on the tracks or vehicles to detect the vibration acceleration of the vehicles.

[0046] S2: Data preprocessing:

[0047] Preprocessing of LiDAR point cloud data: Since LiDAR will be disturbed by environmental noise during the acquisition process, a filtering algorithm is needed to remove these noise points. A statistical filtering algorithm can be used to remove outliers based on the statistical characteristics of the distance distribution of points in the point cloud.

[0048] When the vehicle moves, the point cloud data collected at different times need to be registered to establish a unified coordinate system. Feature-based registration is to first extract the features of the point cloud and then achieve registration through feature matching.

[0049] You can use the median filter method. The median filter is effective in removing salt and pepper noise. It replaces the current pixel value by calculating the median value in the pixel neighborhood.

[0050] For the acceleration sensor data, filtering is required to remove high-frequency noise interference. A low-pass filter can be used to only allow signals below a certain frequency to pass.

[0051] Taking the digital Butterworth low-pass filter as an example, the Butterworth low-pass filter is a commonly used filter. Its characteristic is that it has the flattest frequency response in the passband, allowing signals below the cutoff frequency to pass through without attenuation as much as possible, while attenuating signals above the cutoff frequency, thereby achieving the purpose of removing high-frequency noise.

[0052] Collect initial data: When the displacement sensor is installed and in a stationary state with no displacement, collect data for a period of time to obtain a data sequence x0[n] (n represents the sampling point number).

[0053] Calculate the zero offset: Calculate the average value of the data series This average value is the zero point offset, which reflects the degree to which the sensor deviates from the zero point in the initial state.

[0054] Calibration data: For the actual measurement data x[n] of the displacement sensor collected later, the calibrated displacement data y[n] is obtained by the following formula:

[0055] In this way, the zero point of the displacement sensor is calibrated to the theoretical zero position, thereby improving the accuracy of the measurement.

[0056] The statistical filtering algorithm determines and removes outliers based on the statistical characteristics of the distance distribution of points in the point cloud. Assuming that the point cloud data is relatively evenly distributed in a local area, points that are far away from most points may be noise points.

[0057] Determine the local region: For each point in the point cloud, define a local region, usually a sphere or cube centered at the point. The radius of the sphere (or the side length of the cube) is an important parameter that determines the size of the local region. For example, the radius is set to 0.1 meters, depending on factors such as the density of the point cloud and the size of the object.

[0058] Calculate statistical parameters: In each local area, calculate the statistical parameters of the distance from a point to its neighboring points, such as the mean and standard deviation of the distance.

[0059] Set point P i There are n adjacent points in the local area of ​​P i The distances are d1, d2, …, d n , then the distance mean μ and standard deviation σ can be calculated by the following formula:

[0060]

[0061] Determine outliers: Set a threshold (usually determined based on experience or experiment), such as k times the standard deviation (k is usually 2-3). If the distance from a point to its adjacent point is greater than μ+kσ, the point is determined to be an outlier and removed.

[0062] Feature-based point cloud registration algorithm

[0063] Edge extraction: For point cloud data, edge detection algorithms can be used to extract edge features. A common method is to detect edges based on the change in normal vectors. The normal vector of each point in the point cloud is calculated. When the difference in the normal vectors of adjacent points exceeds a certain threshold, these points may be on the edge. For example, the normal vector is obtained by calculating the local surface fit of the point (such as fitting a plane using the least squares method), and then comparing the angles of the normal vectors of adjacent points.

[0064] Corner point extraction: Corner points are another important feature. Corner points can be extracted by calculating curvature. For a point in a point cloud, calculate the curvature of its local area. The curvature can be estimated by the second-order derivative (which can be approximated by differential calculation in discrete point clouds). When the curvature exceeds a certain threshold, the point may be a corner point.

[0065] Feature Matching

[0066] Feature descriptor calculation: For the extracted features such as edge points and corner points, feature descriptors need to be calculated for matching. For example, for edge points, the edge direction histogram can be used as a descriptor. The local neighborhood of the edge point is divided into several fan-shaped areas, and the direction distribution of the edge points in each area is counted to form a histogram as a descriptor. For corner points, descriptors based on curvature and normal vectors can be used.

[0067] Matching algorithm: Use feature matching algorithms, such as nearest neighbor matching or matching based on distance thresholds. Match between point cloud feature descriptors at two different times. For each feature point descriptor, find the most similar one in the feature descriptor set of another point cloud (such as using Euclidean distance to measure similarity). If the distance is less than the set threshold, the two feature points are considered to match.

[0068] Registration calculation: Calculate the transformation relationship between the two point clouds through the matching feature point pairs. The least squares method can be used to solve the transformation matrix (including rotation and translation) that minimizes the error between the matching feature point pairs. Assume that the matching feature point pairs are Where P 1i is the feature point in the first point cloud, P 2i is the corresponding matching feature point in the second point cloud. To find a transformation matrix T, By solving this least squares problem, we can obtain the transformation matrix required for registration.

[0069] S3: Image data feature extraction:

[0070] The texture features of the tunnel wall surface are extracted using image recognition technology, and the gray-level co-occurrence matrix (GLCM) is used to describe the texture. The energy, contrast, correlation and other parameters in the matrix are calculated to characterize the texture changes.

[0071] Determine the window size and displacement: First, for the acquired tunnel wall image, it is necessary to determine the window size (for example, a window of 3×3, 5×5, or 7×7 pixels) and displacement for calculating the gray-level co-occurrence matrix. The displacement usually includes horizontal, vertical, and diagonal directions. Generally, d is used to represent the displacement distance, and θ is used to represent the displacement direction (for example, 0 degrees represents the horizontal direction, 90 degrees represents the vertical direction, and 45 degrees and 135 degrees represent the diagonal direction).

[0072] Count the frequency of pixel pairs: In the selected window, for each pixel (x, y), find the corresponding pixel according to the set displacement (d, θ), (x+Δx, y+Δy) (where Δx and Δy are determined by d and θ), and count the frequency of the gray value combination of these two pixels. Assuming that the gray level of the image is L (usually 0-255), the constructed gray level co-occurrence matrix P is an L×L two-dimensional matrix, and P(i, j) represents the number of times the pixel with gray value i and the pixel with gray value j appear under a given displacement.

[0073] The energy is the sum of the squares of the gray-level co-occurrence matrix elements and is calculated as:

[0074]

[0075] Energy reflects the uniformity of the image grayscale distribution and the coarseness of the texture. When the image texture is regular and the grayscale distribution is uniform, the energy value is large; when the tunnel wall has stains, diseases, etc., resulting in texture changes and uneven grayscale distribution, the energy value will decrease.

[0076] The formula for calculating contrast is:

[0077] Contrast mainly reflects the contrast of local grayscale changes in the image. The rougher the texture and the more dramatic the grayscale changes, the greater the contrast value. For example, when cracks and other defects appear on the tunnel wall, the grayscale contrast between the crack area and the surrounding normal area is obvious, and the contrast will increase.

[0078] The correlation calculation formula is: in:

[0079]

[0080]

[0081] Correlation is a measure of the linear relationship between row or column elements in the gray-level co-occurrence matrix. Correlation changes when the image texture has directionality (e.g., streak-like defects on a tunnel wall). Correlation is high if the texture has a strong linear relationship in a certain direction and low when the texture is disordered.

[0082] By calculating these parameters, a texture feature vector (such as [E, C, R]) can be constructed for subsequent texture change analysis and abnormality judgment. In practical applications, these parameters can be calculated for normal tunnel wall textures and textures with stains or diseases, and the feature vectors can be compared to determine whether the texture has undergone abnormal changes.

[0083] Extract the shape features of objects in the image. For possible foreign objects, use edge detection algorithms to extract the edge contours of the objects, and then calculate the shape parameters of the contours to determine the nature of the objects.

[0084] If the image is a color image, it needs to be converted to a grayscale image first. The common grayscale conversion method is the weighted average method. For the RGB value of each pixel in the color image (R represents the red component, G represents the green component, and B represents the blue component), the formula G ray =0.299R+0.587G+0.114B converts it into grayscale value to obtain a grayscale image.

[0085] Use appropriate filtering methods to remove noise from the image, such as Gaussian filtering. Gaussian filtering is to perform convolution operations on the image through a two-dimensional Gaussian kernel. The Gaussian kernel function is as follows:

[0086]

[0087] Among them, (x, y) is the coordinate position of the pixel point in the kernel relative to the center, and σ is the standard deviation, which determines the width of the Gaussian distribution, that is, the degree of filtering. By selecting a suitable σ value (for example, σ = 1) to perform a convolution operation on the image, noise interference is removed, and the image is smoother, which is conducive to subsequent edge detection.

[0088] Calculate gradient magnitude and direction

[0089] Use the Sobel operator (or other gradient operators) to calculate the gradient of the image in the horizontal and vertical directions respectively. The horizontal template of the Sobel operator is:

[0090]

[0091] The vertical template is:

[0092]

[0093] By convolving the two templates with the image, we can get the horizontal gradient G x (i, j) and the vertical gradient G y (i,j) (where (i,j) represents the coordinate position of the pixel in the image).

[0094] Then the gradient magnitude M(i, j) and gradient direction θ(i, j) of each pixel are calculated according to the following formula:

[0095]

[0096] The gradient magnitude reflects the strength of the edge, and the gradient direction indicates the direction of the edge.

[0097] The calculated gradient amplitude image is subjected to non-maximum suppression processing to refine the edge and make the edge as single-pixel wide as possible. Along the gradient direction, the amplitude of each pixel is compared with the amplitude of its adjacent pixels in the gradient direction. If the amplitude of the pixel is not the local maximum, its amplitude is set to 0, otherwise its amplitude is retained, so as to obtain a more refined edge candidate point.

[0098] Circumference calculation:

[0099] Discrete contour: If the edge contour is a curve composed of discrete pixels, the chain code can be used to calculate the perimeter for the contour pixels connected in sequence. The chain code uses a series of numbers to represent the direction of the contour pixels. The commonly used 8-direction chain code uses 0-7 to represent 8 different directions. The contour perimeter is calculated by counting the number of occurrences of each direction code in the chain code and combining the distance between pixels (usually the distance between adjacent pixels is 1 unit length) according to the corresponding chain code perimeter calculation formula.

[0100] In the case of curve fitting: you can also first fit the contour to a curve (such as using the least squares method to fit a polynomial curve, etc.), and then use mathematical methods such as integration to calculate the length of the curve as the perimeter based on the mathematical expression of the fitted curve. For example, for a fitted curve y=f(x) on a two-dimensional plane, the length on the interval [a,b] can be calculated using the arc length formula to calculate.

[0101] Area calculation:

[0102] Method based on pixel statistics: Count the number of pixels in the area enclosed by the outline, and multiply the number by the actual area represented by a single pixel (if the correspondence between the physical size of the image and the pixel size is known), then the approximate area value can be obtained. If each pixel in the image represents an actual 1mm×1mm area, and the number of pixels in the outline is N, then the area is Nmm 2 .

[0103] Circularity calculation:

[0104] Common formulas and calculations: Circularity is an indicator that measures how close an object's shape is to a circle. There are many commonly used circularity calculation formulas. The circularity calculation formula based on the perimeter C and area S is: Its value varies between 0 and 1. The closer it is to 1, the closer the shape of the object is to a circle, and the closer it is to 0, the more irregular the shape is. Substituting the calculated perimeter and area values ​​into the formula, the circularity of the object can be calculated.

[0105] By calculating these shape parameters, we can preliminarily determine some basic properties of the object, such as whether the object is a nearly circular component or irregularly shaped debris, and then assist in determining whether there are any abnormalities in the operation of the rail vehicle.

[0106] Extract features from accelerometer data:

[0107] Vibration frequency extraction (based on FFT)

[0108] Data preprocessing: First, obtain the time domain signal a(t) collected by the acceleration sensor, where t represents time. Since the actual collected data may contain noise, the data needs to be filtered before performing the fast Fourier transform (FFT). For example, a Butterworth filter can be used to select the appropriate filter order and cutoff frequency based on the frequency range and noise characteristics of the signal. If it is known that the main frequency range of vehicle vibration is between f1 and f2, then a bandpass filter can be designed to only allow signals within this frequency range to pass.

[0109] FFT calculation: Perform FFT on the filtered acceleration signal a(t) to convert the time domain signal into the frequency domain signal A(f). In the discrete case, let the sampling frequency be f s , the number of sampling points is N, then the calculation formula of FFT is: Where: k = 0, 1, ..., N-1, a(n) is the discretized acceleration signal, and A(k) is the discrete value of the frequency domain signal.

[0110] Frequency component analysis: The frequency domain signal A(f) obtained by FFT has a frequency resolution of Δf=f s / N. The main frequency components of vibration can be determined by analyzing the amplitude distribution of the frequency domain signal. Find the frequencies corresponding to the peak values ​​of the frequency domain signal amplitude. These peak frequencies are the main frequencies of vibration. At the same time, the amplitudes of different frequency components can be observed to determine whether there are abnormal frequencies. If the main frequency of vehicle vibration is f0 under normal operation, and a frequency component significantly higher or lower than f0 appears at a certain moment, and its amplitude is large, this may indicate problems such as loose vehicle parts and uneven tracks.

[0111] Vibration amplitude extraction:

[0112] Directly obtain the amplitude: The vibration amplitude can be directly obtained from the time domain acceleration signal a(t). For simple cases, the peak amplitude of the acceleration signal can be calculated, that is, A=max(a(t)), where a(t) represents the absolute value of the acceleration signal. The root mean square amplitude (RMS) can also be calculated, and the calculation formula is:

[0113]

[0114] Where T is the time interval of analysis. In the discrete case, the calculation formula is:

[0115]

[0116] Where N is the number of sampling points. The root mean square amplitude can better reflect the energy of the vibration.

[0117] Extract features from displacement sensor data:

[0118] Extraction of the maximum wheel-rail relative displacement:

[0119] Data acquisition and processing: Obtain the wheel-rail relative displacement signal d(t) measured by the displacement sensor. Similarly, in order to improve data quality, it may be necessary to filter the data to remove high-frequency noise and outliers. For example, using a moving average filter, for a discrete displacement signal d(n) (n is the sampling point number), its moving average filtered signal d1(n) can be calculated by the following formula:

[0120]

[0121] Among them, M is the window size of the moving average filter, and a suitable value needs to be selected according to the characteristics of the signal and the noise situation.

[0122] Maximum value calculation: In the processed displacement signal d1(t) (or d1(n)), calculate the maximum displacement value Dmax=max(d1(t)). This maximum value can reflect the maximum relative displacement between the wheel and the rail at a certain moment. When the maximum value exceeds a certain safety threshold, it may indicate that the wheel has a tendency to derail or an abnormal collision has occurred between the wheel and the rail.

[0123] Extraction of wheel-rail relative displacement change rate

[0124] Differential calculation (discrete case): For the discrete wheel-rail relative displacement signal d(n), the displacement change rate can be calculated by differential operation. The first-order difference can approximate the displacement change rate, and the calculation formula is:

[0125]

[0126] Wherein, t(n) is the time corresponding to the nth sampling point. Usually, in the case of equal-interval sampling, t(n)=t(n-1)=Δt (sampling interval) is a constant.

[0127] Continuous case (derivative calculation): If the displacement signal is a continuous function d(t), then the rate of change of displacement is the derivative of the displacement function In practical applications, derivatives can be calculated by numerical differentiation methods, such as the central difference method, which is calculated as follows:

[0128] Statistical analysis of the rate of change: After calculating the displacement rate of change r(t), you can perform statistical analysis on it. For example, calculate the maximum, minimum, average, and standard deviation of the rate of change. These statistics can help determine whether the interaction between the wheel and the rail is stable. If the maximum value of the rate of change is too large or the standard deviation is too large, it may indicate that the interaction between the wheel and the rail is abnormal, the wheel may be at risk of derailment, or the track is in poor condition.

[0129] S4: Information Fusion:

[0130] Data layer fusion

[0131] The raw data from LiDAR, images and other sensors are fused at an early stage. For example, the LiDAR point cloud data and the corresponding image data are spatially registered so that each point in the point cloud data can correspond to the pixel position in the image. This can comprehensively utilize the geometric information of the point cloud and the texture information of the image to more comprehensively describe the tunnel section and vehicle operation status.

[0132] Feature layer fusion

[0133] The features extracted from different sensor data are fused. A weighted fusion method can be used to assign different weights to each feature according to its importance. For example, for determining whether a vehicle is at risk of derailment, the weight of the wheel-rail displacement feature may be high, while the weight of the tunnel wall roughness feature may be relatively low. Through feature fusion, a comprehensive feature vector can be obtained to more accurately determine abnormal vehicle operation.

[0134] Tandem Fusion

[0135] Principle: The feature vectors from different sensors are directly concatenated together to form a new feature vector with higher dimension. This method is simple and direct, and can retain all the original feature information, but it may cause the feature vector dimension to be too high, increasing the computational complexity of subsequent processing.

[0136] Assume that the feature vector extracted from the lidar data is F1 = [f 11 ,f 12 ,…,f 1n ], the feature vector extracted from the image data is F2 = [f 21 ,f 22 ,…,f 2m ], the feature vector extracted from other sensors (such as acceleration and displacement sensors) is F3 = [f 31 ,f 32 ,…,f 3k ], the fused feature vector F f =[f 11 ,f 12 ,…,f 1n ,f 21 ,f 22 ,…,f 2m ,f 31 ,f 32 ,…,f 3k ].

[0137] Decision-making level integration

[0138] Use a variety of decision-making algorithms to make a comprehensive judgment on the fused information. For example, methods such as Bayesian network and DS evidence theory can be used. Bayesian network can infer the possibility of abnormal vehicle operation based on prior knowledge and fused feature probability; DS evidence theory can process uncertain information and combine evidence from different sensors to obtain more reliable judgment results.

[0139] Principle: The decision results of multiple classifiers (or decision makers) are voted. Each classifier makes a judgment on the vehicle's operating status (such as normal, abnormal), and then the votes for each judgment are counted, and the result of the majority vote is used as the final fusion decision. This method is simple and intuitive, and is suitable for multiple relatively independent classifiers with similar performance.

[0140] Steps: Assume that there are n classifiers, and their judgment results on the vehicle running status are d1, d2, …, d n (d i The value is normal or abnormal). Count the normal and abnormal votes. For example, the normal votes are c1 and the abnormal votes are c2. If c1>c2, the final fusion decision is normal; otherwise, it is abnormal.

[0141] S5: Establish an abnormality judgment model:

[0142] Establish a mathematical model for judging abnormal vehicle operation, and use machine learning methods such as support vector machine (SVM) and neural network. Through training on a large amount of normal and abnormal operation data, the model can learn the characteristic patterns of normal and abnormal states. When new data is input, the model can judge whether the vehicle operation is abnormal based on the learned patterns.

[0143] Each collected data record corresponds to the relevant feature information of the vehicle at a certain moment, forming a feature vector. For example, the feature vector can be [tunnel wall roughness, image texture feature parameters, acceleration amplitude, maximum wheel-rail relative displacement], etc.

[0144] The collected data is manually labeled, and the data of normal operation is marked as one category (for example, marked as +1), and the data of abnormal operation is marked as another category (for example, marked as -1).

[0145] In order to eliminate the influence of different feature dimensions and improve the model training effect, each feature in the feature vector is normalized. Common normalization methods include minimum-maximum normalization, and the formula is:

[0146]

[0147] Where: x is the original eigenvalue, x min and x max are the minimum and maximum values ​​of the feature in all data, respectively, new is the normalized eigenvalue.

[0148] Select kernel function: SVM uses kernel function to map low-dimensional data to high-dimensional space in order to find a better classification hyperplane. Commonly used kernel functions include linear kernel function Polynomial kernel function ( where c is a constant and d is the degree of the polynomial), Gaussian kernel function (also called radial basis function, σ is the kernel width parameter). Select an appropriate kernel function based on the distribution characteristics of the data and the complexity of the problem. For example, if the data is linearly separable or approximately linearly separable in the original space, the linear kernel function can be tried first; if the data is more complex and has obvious nonlinear characteristics, the Gaussian kernel function can be considered.

[0149] Determine model parameters: Use cross-validation and other methods to determine the key parameters of the SVM model. For example, for the Gaussian kernel function, you need to determine the value of σ and the penalty parameter C (which controls the degree of penalty for misclassification. The larger C is, the heavier the penalty for misclassification is). Cross-validation can divide the data set into training set, validation set, and test set. For example, the commonly used K-fold cross-validation method is used (divide the data set into K equal parts, take K-1 parts as training sets each time, and the remaining 1 part as validation set, repeat K times, and take the average performance index to select the optimal parameters).

[0150] Training model: Using labeled training data, based on the selected kernel function and determined parameters, use the optimization algorithm to solve the objective function of SVM and find the optimal classification hyperplane, that is, determine the support vector and the corresponding weight coefficient, so that the two types of data (normal and abnormal) can be separated as accurately as possible.

[0151] For a given training data set {(x1,y1),(x2,y2),…,(xn,yn)}, where x i is the input feature vector (for example, in the case of abnormal vehicle operation, it is a vector containing various sensor features), y i ∈{+1,-1} is the corresponding category label (indicating normal or abnormal), and the objective function of the soft margin linear separable SVM can be expressed as:

[0152] sty i (ω T x i +b)≥1-ζ i ,ζ i ≥0,i=1,2,…,n;

[0153] Where: ω is the weight vector, b is the bias term, ζ i is a slack variable (used to allow a certain degree of misclassification), and C is a penalty parameter used to control the degree of penalty for misclassification.

[0154] Model evaluation: Use the test set data to evaluate the trained SVM model. Common evaluation indicators include accuracy. TP is the true positive, that is, the number of cases correctly predicted as abnormal; TN is the true negative, that is, the number of cases correctly predicted as normal; FP is the false positive, that is, the number of cases that are misjudged as abnormal; FN is the false negative, that is, the number of cases that are misjudged as normal. Recall F1 value Determine whether the model meets the requirements based on the evaluation results. If not, adjust parameters, change kernel functions, or further optimize data to retrain the model.

[0155] Abnormal judgment application: When new data is input, it is normalized in the same way and then substituted into the trained SVM model to judge whether the vehicle operation is abnormal based on the output of the model. If the output value is greater than 0, it is judged as normal operation (corresponding to the mark +1); if the output value is less than or equal to 0, it is judged as abnormal operation (corresponding to the mark -1).

[0156] S6: Set the alarm threshold:

[0157] According to the fused features and judgment results, a reasonable alarm threshold is set. For the vehicle vibration acceleration amplitude, if it exceeds a certain threshold, it is judged as abnormal and an alarm signal is issued. The alarm signal can be displayed in the tunnel control center through the sound and light alarm. At the same time, the abnormal information can be sent to the maintenance personnel's mobile terminal so that maintenance measures can be taken in time.

[0158] The invention discloses a railway vehicle operation abnormality judgment device based on tunnel section multi-source information fusion, comprising a memory, a processor and a computer program stored in the memory and executable on the processor. When the processor executes the computer program, the steps of the railway vehicle operation abnormality judgment method based on tunnel section multi-source information fusion are implemented.

[0159] Although embodiments of the present invention have been shown and described, it will be appreciated by those skilled in the art that various changes, modifications, substitutions and variations may be made to the embodiments without departing from the principles and spirit of the present invention, and that the scope of the present invention is defined by the appended claims and their equivalents.

Claims

1. A rail vehicle operation abnormality judgment system based on tunnel section multi-source information fusion, including a data acquisition and processing module, a feature extraction module, an information fusion module, a judgment module and an early warning module, characterized in that: The data acquisition and processing module includes a laser radar sensor, a high-definition camera, an acceleration sensor and a displacement sensor. The data acquisition and processing is used to obtain the three-dimensional profile information of the tunnel section, collect the visual image inside the tunnel, and the vibration acceleration situation; The feature extraction module uses image recognition technology to extract the texture features of the tunnel wall surface, uses a grayscale co-occurrence matrix to describe the texture, characterizes the texture changes by calculating parameters such as energy, contrast, and correlation in the matrix, and extracts the shape features of the object in the image; The information fusion module fuses the raw data of lidar, images and other sensors at an early stage, and fuses the features extracted from different sensor data; The judgment module establishes a mathematical model for judging abnormal vehicle operation, and judges whether the vehicle operation is abnormal based on the learned pattern by training a large amount of normal and abnormal operation data; The early warning module sets a reasonable alarm threshold based on the fused features and judgment results. If a certain threshold is exceeded, it is judged as abnormal and an alarm signal is issued.

2. A method for determining abnormal operation of rail vehicles based on multi-source information fusion of tunnel sections, characterized in that: The following steps are involved: S1: Data collection: Installing a laser radar sensor in the tunnel can obtain the three-dimensional profile information of the tunnel section with high precision. The laser radar scans the tunnel wall at a certain scanning frequency to obtain a series of point cloud data. These point cloud data can accurately represent the spatial position of the tunnel wall at a certain moment. S2: Data preprocessing: Preprocessing of LiDAR point cloud data: Since LiDAR will be disturbed by environmental noise during the acquisition process, a filtering algorithm is needed to remove these noise points. A statistical filtering algorithm can be used to remove outliers based on the statistical characteristics of the distance distribution of points in the point cloud. S3: Image data feature extraction: The texture features of the tunnel wall surface are extracted using image recognition technology, and the gray-level co-occurrence matrix (GLCM) is used to describe the texture. The energy, contrast, correlation and other parameters in the matrix are calculated to characterize the texture changes. Extract the shape features of objects in the image. For possible foreign objects, use edge detection algorithms to extract the edge contours of the objects, and then calculate the shape parameters of the contours to determine the nature of the objects. S4: Information Fusion: Fuse the raw data of lidar, images and other sensors at an early stage, fuse the features extracted from different sensor data, and use multiple decision algorithms to make comprehensive judgments on the fused information; S5: Establish an abnormality judgment model: Establish a mathematical model for judging abnormal vehicle operation, and use machine learning methods such as support vector machine (SVM) and neural network. Through training on a large amount of normal and abnormal operation data, the model can learn the characteristic patterns of normal and abnormal states. When new data is input, the model can judge whether the vehicle operation is abnormal based on the learned patterns. S6: Set the alarm threshold: According to the fused features and judgment results, a reasonable alarm threshold is set. For the vehicle vibration acceleration amplitude, if it exceeds a certain threshold, it is judged as abnormal and an alarm signal is issued. The alarm signal can be displayed in the tunnel control center through the sound and light alarm. At the same time, the abnormal information can be sent to the maintenance personnel's mobile terminal so that maintenance measures can be taken in time.

3. The method for determining abnormal operation of a railway vehicle based on multi-source information fusion of a tunnel section according to claim 1 is characterized in that: In the S1, a high-definition camera is installed in the tunnel to collect visual images inside the tunnel. The camera is installed at a position that can ensure coverage of key tunnel section areas.

4. The method for determining abnormal operation of a railway vehicle based on multi-source information fusion of a tunnel section according to claim 1 is characterized in that: In the S1, an acceleration sensor and a displacement sensor are installed in the tunnel at the same time. The acceleration sensor is installed on the track or the vehicle to detect the vibration acceleration of the vehicle.

5. The method for determining abnormal operation of a railway vehicle based on multi-source information fusion of a tunnel section according to claim 1 is characterized in that: In the data preprocessing in S2, when the vehicle moves, the point cloud data collected at different times need to be registered to establish a unified coordinate system. The feature-based registration is to first extract the features of the point cloud and then achieve registration through feature matching.

6. The method for determining abnormal operation of a railway vehicle based on multi-source information fusion of a tunnel section according to claim 1 is characterized in that: In the data preprocessing in S2, a median filtering method may be used. Median filtering is effective in removing salt and pepper noise. It replaces the current pixel value by calculating the median value in the pixel neighborhood.

7. The method for determining abnormal operation of a railway vehicle based on tunnel section multi-source information fusion according to claim 4 is characterized in that: For the acceleration sensor data, filtering is required to remove high-frequency noise interference. A low-pass filter can be used to only allow signals below a certain frequency to pass.

8. A rail vehicle operation abnormality judgment device based on tunnel section multi-source information fusion, comprising a memory, a processor, and a computer program stored in the memory and executable on the processor, characterized in that: When the processor executes the computer program, the steps of the method for determining abnormal operation of a rail vehicle based on multi-source information fusion of a tunnel section as described in any one of claims 1 to 7 are implemented.

Images