A multi-dimensional train derailment monitoring method based on vision and vibration

Abstract

The present invention discloses a multi-dimensional train derailment monitoring method based on vision and vibration, which belongs to the field of rail transportation. The method comprises: respectively collecting wheel-rail contact area images and vibration data of each wheel in different directions, and aligning the wheel-rail contact area images with the vibration data in time and space; extracting visual features from the wheel-rail contact area images by combining edge detection algorithm and Hough transform technology, and extracting vibration features from the vibration data, correlating the visual features with the vibration features to generate a wheel rim lateral position-vibration feature mapping table; establishing a D-S evidence theory fusion model based on the wheel rim lateral position-vibration feature mapping table, and utilizing the D-S evidence theory fusion model to fuse the visual features with the vibration features, and determining the risk of train derailment according to the feature fusion result. The present invention realizes deep coupling analysis of visual images and vibration data, and solves the defects of the existing technology that it cannot adapt to changes in working conditions, is very likely to generate false alarms and affect operations.

Description

Technical Field

[0001] The present invention relates to the field of rail transportation, and in particular to a multi-dimensional train derailment monitoring method based on vision and vibration. Background Art

[0002] With the development of driverless rail transit technology, derailment detection has become a research hotspot. Currently, a contact-type obstacle and derailment detection device suitable for driverless subways has been developed. This is of great significance for improving the safety of driverless subways, enabling emergency braking in the event of a derailment, reducing accidental injuries. Furthermore, building on the existing online running gear monitoring system, the vehicle derailment monitoring function has been further expanded. Through the use of relevant derailment monitoring algorithms, the ability to predict vehicle derailments based on vibration impact has been realized.

[0003] Derailment detection: The working principle of derailment detection usually involves monitoring the vehicle's operating status, including the vehicle's running parts, stability, and comfort. These monitoring systems can assess whether the vehicle is at risk of derailment through different sensors and analysis models.

[0004] Existing technologies primarily rely on contact detection and vibration and impact detection, both of which are widely used. However, contact detection has a high false alarm rate, while vibration and impact detection has a high missed alarm rate, resulting in less than ideal operational performance. Contact obstacle detection can only generate an alarm after a derailment has occurred, providing no early warning. Vibration and impact detection is affected by the vibrations of the train after it is running, resulting in a high probability of false alarms and ineffective prevention, impacting actual operations.

[0005] For example, Chinese invention patent CN114368411B discloses a train derailment safety monitoring and early warning method. This method uses a visual behavior-based monitoring and early warning method to analyze images of the wheel-rail contact area, calculate the relative lateral displacement of the wheel and rail, and assess the derailment safety level based on the relative lateral displacement of the wheel and rail in the world coordinate system. However, the operating environment under the train is relatively harsh, which affects the accuracy of visual detection. The presence of debris under the train and the complex visual environment (lighting, background, etc.) lead to a relatively high false alarm rate.

[0006] Chinese invention patent CN107521511B discloses a wheel derailment alarm method and system. After obtaining a vibration signal, the vibration signal is processed to obtain a sample to be analyzed; time domain identification is performed on the sample to be analyzed to obtain a time domain derailment identification result; fast Fourier transform is performed on the sample to be analyzed to obtain a corresponding spectrum, and frequency domain identification is performed on the spectrum to obtain a frequency domain derailment identification result. However, derailment vibration information can generally only occur when derailment has already occurred. Although there is an early warning function, it involves vehicle control, resulting in a relatively conservative threshold setting, which may result in missed alarms. In addition, the algorithm is not mature and complex enough and is prone to false alarms. Once a false alarm is reported, emergency braking will be triggered, seriously affecting actual operations.

[0007] Currently, no effective solutions have been proposed for the problems in related technologies. Summary of the Invention

[0008] (1) Technical problems solved

[0009] In response to the shortcomings of the existing technology, the present invention provides a multi-dimensional train derailment monitoring method based on vision and vibration, which has the advantage of deep coupling analysis of visual images and vibration data, thereby solving the problem that the existing technology cannot adapt to changes in working conditions, is prone to false alarms, and affects operations.

[0010] (2) Technical solution

[0011] To achieve the advantages of deep coupled analysis of visual images and vibration data, the specific technical solutions adopted by the present invention are as follows:

[0012] A multi-dimensional train derailment monitoring method based on vision and vibration, the method comprising:

[0013] S1. Collect wheel-rail contact area images and vibration data of each wheel in different directions, and align the wheel-rail contact area images and vibration data in time and space;

[0014] S2. Combining edge detection algorithm and Hough transform technology to extract visual features from the wheel-rail contact area image, and extract vibration features from the vibration data, and then correlating the visual features with the vibration features to generate a wheel flange lateral position-vibration feature mapping table;

[0015] S3. A DS evidence theory fusion model is established based on the wheel flange lateral position-vibration feature mapping table. The visual features and vibration features are fused using the DS evidence theory fusion model, and the train derailment risk is determined based on the feature fusion results.

[0016] Preferably, respectively collecting the wheel-rail contact area image and the vibration data of each wheel in different directions, and performing spatiotemporal alignment of the wheel-rail contact area image and the vibration data includes:

[0017] S11. Using a video device to capture an image of the wheel-rail contact area, and using an acceleration sensor to capture the three-axis acceleration of the wheel to obtain vibration data, and performing time alignment between the wheel-rail contact area image and the vibration data using a clock protocol;

[0018] S12. Establishing a visual coordinate system based on the track plane with the track center reference point as the origin, and establishing a vibration coordinate system with the bogie center as the origin;

[0019] S13. Mapping the vibration data of the vibration coordinate system to the visual coordinate system through the rigid body transformation matrix for spatial alignment.

[0020] Preferably, mapping the vibration data of the vibration coordinate system to the visual coordinate system through the rigid body transformation matrix for spatial alignment includes:

[0021] S131. Using a laser rangefinder to collect the installation status of the acceleration sensor and measure the Euler angle of the vibration coordinate system relative to the visual coordinate system;

[0022] S132. Calculate the rotation matrix based on the Euler angles and the rotation matrix synthesis formula, measure the coordinates of the origin of the vibration coordinate system in the visual coordinate system, and determine the translation vector;

[0023] S133, converting the vibration data of the vibration coordinate system into components of the visual coordinate system through the rotation matrix and the translation vector, and spatially correlating them with the wheel rim position in the visual coordinate system;

[0024] S134. Compare the wheel flange position of the actual wheel-rail contact area image with the spatial distribution of the vibration data, and optimize the rotation matrix and translation vector.

[0025] Preferably, the edge detection algorithm and Hough transform technology are combined to extract visual features from the wheel-rail contact area image, and vibration features are extracted from the vibration data. The visual features and vibration features are associated to generate a wheel flange lateral position-vibration feature mapping table, which includes:

[0026] S21. Divide the visual area in the visual coordinate system based on predefined division conditions, where the visual area includes a safety area, a warning area, and an alarm area;

[0027] S22, combining the Canny edge detection algorithm and Hough transform technology to extract the wheel-rail profile features in the wheel-rail contact area image in real time, and calculate the lateral offset of the wheel tread centerline;

[0028] S23, calculating the time domain characteristics of the acceleration signal in the vibration data, and extracting the frequency domain characteristics of the acceleration signal using an adaptive time-frequency fusion fast Fourier transform algorithm;

[0029] S24. Based on the time-domain and frequency-domain characteristics of the acceleration signal, the adaptive empirical mode decomposition algorithm is used to generate the transformed marginal spectrum and analyze the time-frequency characteristics of the derailment precursor;

[0030] S25. Combining the wheel-rail profile characteristics and the time-frequency characteristics, a wheel flange lateral position-vibration characteristic mapping table is generated.

[0031] Preferably, extracting the frequency domain features of the acceleration signal using the adaptive time-frequency fusion fast Fourier transform algorithm includes:

[0032] Preprocess the acceleration signal, retain the wheel-rail coupling related frequency band, and use the Hanning window function to perform windowing processing on the wheel-rail coupling related frequency band;

[0033] The Fourier transform parameters of the windowed acceleration signals in different visual areas are dynamically adjusted, and the visual evidence weight is introduced into the calculation of the Fourier transform parameters to obtain the amplitude spectrum.

[0034] The amplitude spectrum is converted into a single-sided spectrum, and the peak value and amplitude within the preset range are searched in the single-sided spectrum. The wheel-rail coupling characteristic frequency is obtained according to the peak value and amplitude of each frequency point.

[0035] Preferably, based on the time domain characteristics and frequency domain characteristics of the acceleration signal, an adaptive empirical mode decomposition algorithm is used to generate a transform marginal spectrum, and the time-frequency characteristics of the derailment precursor are analyzed, including:

[0036] S241. Dynamically adjusting the empirical mode decomposition parameters of the visual area using an empirical mode decomposition algorithm to obtain a plurality of intrinsic mode function components;

[0037] S242. Performing a Hilbert transform on each intrinsic mode function component to obtain an analytical signal, and calculating the instantaneous amplitude and instantaneous frequency of the analytical signal;

[0038] S243, superimposing the temporal distributions of the instantaneous amplitudes and instantaneous frequencies of all intrinsic mode function components to form a Hilbert spectrum;

[0039] S244. Integrate the Hilbert spectrum on the time axis to obtain a marginal spectrum, and identify abnormal frequency components of derailment precursors based on the marginal spectrum.

[0040] Preferably, the empirical mode decomposition algorithm is used to dynamically adjust the empirical mode decomposition parameters of the visual area to obtain several intrinsic mode function components including:

[0041] S2411. Selecting a decomposition method of the empirical mode decomposition algorithm based on different visual areas;

[0042] S2412. Obtain local maximum and minimum values of the acceleration signal, connect the maximum points to form an upper envelope, connect the minimum points to form a lower envelope, and fit the upper and lower envelopes by cubic spline interpolation;

[0043] S2413, calculating a preliminary intrinsic mode function based on the fitted upper and lower envelopes, repeating steps S2411 to S2412 for the preliminary intrinsic mode function until the standard deviation criterion is met, and obtaining the first intrinsic mode function component;

[0044] S2414. Subtract the first intrinsic mode function component from the original acceleration signal to obtain a residual signal, and repeatedly decompose the residual signal until the residual signal is a monotonic function to obtain a plurality of intrinsic mode function components.

[0045] Preferably, the decomposition method of the empirical mode decomposition algorithm based on different visual areas includes:

[0046] S24111. In the safe zone, the energy distribution of the low-frequency fundamental frequency and its harmonics is extracted using a coarse-grained decomposition method;

[0047] S24112. In the warning area and alarm area, a fine-grained decomposition method is used and visual constraints are introduced to decompose high-frequency impact modes.

[0048] Preferably, a DS evidence theory fusion model is established based on the wheel flange lateral position-vibration feature mapping table, and the visual features and vibration features are fused using the DS evidence theory fusion model. Determining the train derailment risk based on the feature fusion result includes:

[0049] S31. defining a derailment risk state set based on the divided visual area;

[0050] S32. Calculate the confidence level of each risk state in the derailment risk state set based on the area where the wheel tread centerline is located and use it as visual evidence;

[0051] S33. Calculate the confidence level based on the degree of acceleration anomaly and the proportion of high-frequency components and use it as vibration evidence;

[0052] S34. Based on the predefined Dempster synthesis rule, the visual evidence and the vibration evidence are fused, and the train derailment risk is determined based on the fused features.

[0053] Preferably, the rotation matrix synthesis formula is:

[0054] ;

[0055] Where, represents the rotation matrix; Indicates winding The basic rotation matrix of the axis; Indicates winding The basic rotation matrix of the axis; Indicates winding The basic rotation matrix of the axis; Indicates the pitch angle; represents the yaw angle; Indicates the roll angle.

[0056] (3) Beneficial effects

[0057] Compared with the existing technology, the present invention provides a multi-dimensional train derailment monitoring method based on vision and vibration, which has the following beneficial effects:

[0058] 1. The present invention breaks through the limitations of single vibration data through a multi-dimensional architecture that combines visual area quantification and grading with multi-axis analysis of vibration acceleration. It also realizes deep coupling analysis of visual images and vibration data through a multi-sensor fusion algorithm with a three-layer architecture of spatiotemporal alignment, feature-level fusion, and decision-level fusion, thus solving the defects of existing technologies that cannot adapt to changes in working conditions, are prone to false alarms, and affect operations.

[0059] 2. The present invention adopts a bogie-integrated sensor layout, eliminating the need for track facility modification. It implements risk probability calculation through an evidence theory fusion algorithm, rather than simple signal triggering logic. This represents a substantial improvement in detection logic and control strategy, forming an intelligent closed loop of "multi-dimensional perception - cross-analysis - accurate decision-making," significantly enhancing the technical solution's non-obviousness and practical application value. BRIEF DESCRIPTION OF THE DRAWINGS

[0060] In order to more clearly illustrate the configuration method of the present invention or the technical solutions in the prior art, the following briefly introduces the drawings required for use in the embodiments. Obviously, the drawings described below are only some embodiments of the present invention. For ordinary technicians in this field, other drawings can be obtained based on these drawings without paying any creative work.

[0061] Figure 1 is a flow chart of a multi-dimensional train derailment monitoring method based on vision and vibration according to an embodiment of the present invention;

[0062] Figure 2 is a flow chart of an empirical mode decomposition algorithm in a multi-dimensional train derailment monitoring method based on vision and vibration according to an embodiment of the present invention;

[0063] Figure 3 is a principle block diagram of a multi-dimensional train derailment monitoring method based on vision and vibration according to an embodiment of the present invention;

[0064] Figure 42. A schematic diagram of the position of the wheel flange centerline in the multi-dimensional train derailment monitoring method based on vision and vibration according to an embodiment of the present invention;

[0065] Figure 5 This is a pre-alarm area identification diagram in a multi-dimensional train derailment monitoring method based on vision and vibration according to an embodiment of the present invention. DETAILED DESCRIPTION

[0066] To further illustrate each embodiment, the present invention provides drawings, which are part of the disclosure of the present invention. They are mainly used to illustrate the embodiments and can be used in conjunction with the relevant descriptions in the specification to explain the operating principles of the embodiments. By referring to these contents, ordinary technicians in this field should be able to understand other possible implementation methods and advantages of the present invention. The components in the figures are not drawn to scale, and similar component symbols are generally used to represent similar components.

[0067] According to an embodiment of the present invention, a multi-dimensional train derailment monitoring method based on vision and vibration is provided.

[0068] The present invention will now be further described with reference to the accompanying drawings and specific embodiments. Figure 1 As shown, according to a multi-dimensional train derailment monitoring method based on vision and vibration according to an embodiment of the present invention, the method includes:

[0069] S1. Collect wheel-rail contact area images and vibration data of each wheel in different directions respectively, and align the wheel-rail contact area images and vibration data in time and space.

[0070] The wheel-rail contact area image and vibration data of each wheel in different directions are collected separately, and the wheel-rail contact area image and vibration data are aligned in time and space, including:

[0071] S11. Using a video device to capture an image of the wheel-rail contact area, and using an acceleration sensor to capture the three-axis acceleration of the wheel to obtain vibration data, and performing time alignment between the wheel-rail contact area image and the vibration data using a clock protocol;

[0072] S12. Establishing a visual coordinate system based on the track plane with the track center reference point as the origin, and establishing a vibration coordinate system with the bogie center as the origin;

[0073] S13. Mapping the vibration data of the vibration coordinate system to the visual coordinate system through the rigid body transformation matrix for spatial alignment.

[0074] Mapping the vibration data of the vibration coordinate system to the visual coordinate system through the rigid body transformation matrix for spatial alignment includes:

[0075] S131. Using a laser rangefinder to collect the installation status of the acceleration sensor and measure the Euler angle of the vibration coordinate system relative to the visual coordinate system;

[0076] S132. Calculate the rotation matrix based on the Euler angles and the rotation matrix synthesis formula, measure the coordinates of the origin of the vibration coordinate system in the visual coordinate system, and determine the translation vector;

[0077] S133, converting the vibration data of the vibration coordinate system into components of the visual coordinate system through the rotation matrix and the translation vector, and spatially correlating them with the wheel rim position in the visual coordinate system;

[0078] S134. Compare the wheel flange position of the actual wheel-rail contact area image with the spatial distribution of the vibration data, and optimize the rotation matrix and translation vector.

[0079] It should be noted that the following further describes the process of separately collecting the wheel-rail contact area image and the vibration data of each wheel in different directions, and performing spatiotemporal alignment on the wheel-rail contact area image and the vibration data, in conjunction with specific implementation methods.

[0080] 1. Spatiotemporal registration (establishing a unified data benchmark):

[0081] 1. Hardware synchronization mechanism

[0082] To meet the high-speed train operation requirements, it is planned to use high-refresh, high-resolution, and anti-interference video acquisition equipment. The hardware visual acquisition equipment is installed between the first and second axles of each bogie, with two sets of equipment installed on each bogie.

[0083] Visual data: The laser binocular camera captures images of the wheel-rail contact area at a refresh rate of 500fps, with a built-in hardware clock accuracy of ±1μs;

[0084] Vibration data: The three-axis accelerometer collects X (longitudinal), Y (transverse), and Z (vertical) accelerations simultaneously at a 10kHz sampling rate, enabling hardware trigger synchronization.

[0085] Spatiotemporal alignment: Using the IEEE 1588 precision clock protocol, the visual frame timestamp and vibration sampling point timestamp are unified to the nanosecond level, ensuring alignment of multi-source data at the same physical moment.

[0086] 2. Spatial coordinate system mapping

[0087] Visual coordinate system: Take the track plane as the XY plane and define the lateral position of the wheel rim centerline O (zone I / II / III). The position of the wheel rim centerline O is as follows: Figure 4 As shown ( Figure 4 A in the figure indicates the rim centerline O);

[0088] like Figure 5 As shown, the I / II / III zone division is based on:

[0089] (1) According to Chinese railway standards, the lateral displacement of the wheelset shall not exceed 70% of the rail head width.

[0090] (2) In the wheel-rail contact simulation, when using tools such as ANSYS to simulate the lateral displacement of the wheelset, when the lateral displacement reaches 67% of the rail head width (i.e., 46.67 mm), the contact stress at the root of the wheel flange reaches 80% of the material yield strength, and an early warning needs to be triggered; when it exceeds the rail head width, the contact stress exceeds the yield strength.

[0091] (3) Field test data: In railway tests, when the wheelset lateral displacement exceeds 67% of the rail head width, the wheel flange wear depth increases by 1.2 mm per month, which is 2.5 times the normal wear rate. Based on the above three points, the classification is: Zone I is safe, Zone II is warning, and Zone III is alarm.

[0092] Zone I width a = 0.67 (L / 2); Zone II width b = 0.33 (L / 2); Zone III widths greater than L. For example, when L = 70 mm, the wheelset can operate safely within a range of ±23.45 mm.

[0093] Zone II (warning zone): indicates that a warning is required when the lateral displacement of the wheelset exceeds 23.45mm but does not reach 35mm.

[0094] Zone III (alarm zone): If the width exceeds L (i.e., lateral displacement > 35 mm), emergency braking is triggered to prevent the wheel flange from completely separating from the rail head and causing derailment.

[0095] Vibration coordinate system: With the bogie center as the origin, the vibration coordinate system and the visual coordinate system are mapped through the rigid body transformation matrix (rotation matrix R, translation vector T) to ensure that the spatial physical meaning of the vibration data is consistent with the visual area judgment.

[0096] It should be noted that the rigid body transformation matrix contains the rotation matrix and the translation vector, which is used to describe the spatial transformation relationship between the two coordinate systems:

[0097] Rotation matrix: is a 3×3 coordinate system with respect to the visual coordinate system (such as the rotation of axis rotation angle), satisfying (identity matrix).

[0098] Translation vector: A 3×1 vector that represents the translation distance of the origin of the vibration coordinate system relative to the origin of the visual coordinate system (offset along the x / y / z axis).

[0099] Transformation expression (vector form): Assume the visual coordinate system is , the vibration coordinate system is , a point at The coordinates in are:

[0100] ;

[0101] The coordinates in are:

[0102] ;

[0103] Then the transformation relationship is:

[0104] ;

[0105] The homogeneous coordinate form (for easy processing by the preprocessor or host) is:

[0106] ;

[0107] Where, represents a 1×3 zero vector; Represents the rotation matrix, which describes the rotation posture of the vibration coordinate system relative to the visual coordinate system; Represents the translation vector, which describes the position displacement of the origin of the vibration coordinate system in the visual coordinate system; Represents the point coordinate vector in the visual coordinate system; Represents the point coordinate vector in the visual coordinate system; Indicates the three-dimensional coordinates of a point in the vibration coordinate system; Indicates the three-dimensional coordinates of a point in the visual coordinate system.

[0108] 2. Specific implementation steps of the rigid body transformation matrix:

[0109] 1. Coordinate system definition and hardware installation

[0110] Visual coordinate system: The track plane is the XY plane. The origin can be set to a reference point in the center of the track. The X-axis is along the longitudinal direction of the track (the direction of train operation), the Y-axis is along the transverse direction of the track (the direction of wheelset lateral movement), and the Z-axis is perpendicular to the track plane (vertical).

[0111] Vibration coordinate system: With the center of the bogie as the origin, the sensor is installed at the axle box position, and its coordinate axis direction is aligned with the bogie structure (for example, the X axis corresponds to the longitudinal acceleration , the Y axis corresponds to the lateral acceleration , the Z axis corresponds to the vertical acceleration ).

[0112] 2. Determine the rotation matrix

[0113] Step 1: Measure the sensor installation posture. Use a laser rangefinder or inertial navigation equipment to measure the Euler angle (pitch angle) of the vibration coordinate system relative to the visual coordinate system. , yaw angle , roll angle ).

[0114] Step 2: Construct the rotation matrix according to the Euler angle order (such as zyx rotation) and calculate it through the rotation matrix synthesis formula :

[0115] ;

[0116] Where, Respectively around The basic rotation matrix for an axis, for example:

[0117] ;

[0118] It should be noted that the order of Euler angles must be consistent with the actual installation posture to avoid the universal joint lock problem.

[0119] 3. Determine the translation vector

[0120] Measuring the origin coordinates of the sensor Using high-precision positioning tools, measure the coordinates of the origin of the vibration coordinate system in the visual coordinate system ,Right now:

[0121] ;

[0122] For example, the sensor is mounted on the bogie axle box, and its lateral offset relative to the track center is , the vertical height is , the vertical position is .

[0123] 4. Data transformation and verification

[0124] Real-time coordinate conversion of acceleration data collected by vibration sensors Pass first and Converted into components in the visual coordinate system, and then compared with the wheel rim position in the visual coordinate system (such as ) for spatial association.

[0125] Calibration and error correction are carried out through field tests (such as simulating wheelset lateral movement), comparing the wheel rim position detected by visual inspection with the spatial distribution of vibration data, and fine-tuning and Parameters are used to ensure that the transformation accuracy reaches nanosecond-level time synchronization and millimeter-level spatial alignment.

[0126] S2. Combine the edge detection algorithm and Hough transform technology to extract visual features from the wheel-rail contact area image, and extract vibration features from the vibration data. Associate the visual features with the vibration features to generate a wheel flange lateral position-vibration feature mapping table.

[0127] Among them, the edge detection algorithm and Hough transform technology are combined to extract visual features from the wheel-rail contact area image, and vibration features are extracted from the vibration data. The visual features and vibration features are associated to generate a wheel flange lateral position-vibration feature mapping table including:

[0128] S21. Divide the visual area in the visual coordinate system based on predefined division conditions, where the visual area includes a safety area, a warning area, and an alarm area;

[0129] S22, combining the Canny edge detection algorithm and Hough transform technology to extract the wheel-rail profile features in the wheel-rail contact area image in real time, and calculate the lateral offset of the wheel tread centerline;

[0130] S23. Calculate the time domain features of the acceleration signal in the vibration data, and extract the frequency domain features of the acceleration signal using an adaptive time-frequency fusion fast Fourier transform algorithm.

[0131] Among them, the frequency domain features of the acceleration signal extracted using the adaptive time-frequency fusion fast Fourier transform algorithm include:

[0132] Preprocess the acceleration signal, retain the wheel-rail coupling related frequency band, and use the Hanning window function to perform windowing processing on the wheel-rail coupling related frequency band;

[0133] The Fourier transform parameters of the windowed acceleration signals in different visual areas are dynamically adjusted, and the visual evidence weight is introduced into the calculation of the Fourier transform parameters to obtain the amplitude spectrum.

[0134] The amplitude spectrum is converted into a single-sided spectrum, and the peak value and amplitude within the preset range are searched in the single-sided spectrum. The wheel-rail coupling characteristic frequency is obtained according to the peak value and amplitude of each frequency point.

[0135] S24. Based on the time-domain and frequency-domain characteristics of the acceleration signal, the adaptive empirical mode decomposition algorithm is used to generate the transformed marginal spectrum and analyze the time-frequency characteristics of the derailment precursor.

[0136] Among them, based on the time domain characteristics and frequency domain characteristics of the acceleration signal, the adaptive empirical mode decomposition algorithm is used to generate the transformation marginal spectrum, and the time-frequency characteristics of the derailment precursor are analyzed, including:

[0137] S241. Dynamically adjust the empirical mode decomposition parameters of the visual area using an empirical mode decomposition algorithm to obtain several intrinsic mode function components.

[0138] like Figure 2 As shown in FIG, the empirical mode decomposition algorithm is used to dynamically adjust the empirical mode decomposition parameters of the visual area to obtain several intrinsic mode function components including:

[0139] S2411. Select a decomposition method of the empirical mode decomposition algorithm based on different visual areas.

[0140] Among them, the decomposition methods of the empirical mode decomposition algorithm based on different visual areas include:

[0141] S24111. In the safe zone, the energy distribution of the low-frequency fundamental frequency and its harmonics is extracted using a coarse-grained decomposition method;

[0142] S24112. In the warning area and alarm area, a fine-grained decomposition method is used and visual constraints are introduced to decompose high-frequency impact modes.

[0143] S2412. Obtain local maximum and minimum values of the acceleration signal, connect the maximum points to form an upper envelope, connect the minimum points to form a lower envelope, and fit the upper and lower envelopes by cubic spline interpolation;

[0144] S2413, calculating a preliminary intrinsic mode function based on the fitted upper and lower envelopes, repeating steps S2411 to S2412 for the preliminary intrinsic mode function until the standard deviation criterion is met, and obtaining the first intrinsic mode function component;

[0145] S2414. Subtract the first intrinsic mode function component from the original acceleration signal to obtain a residual signal, and repeatedly decompose the residual signal until the residual signal is a monotonic function to obtain a plurality of intrinsic mode function components.

[0146] S242. Performing a Hilbert transform on each intrinsic mode function component to obtain an analytical signal, and calculating the instantaneous amplitude and instantaneous frequency of the analytical signal;

[0147] S243, superimposing the temporal distributions of the instantaneous amplitudes and instantaneous frequencies of all intrinsic mode function components to form a Hilbert spectrum;

[0148] S244. Integrate the Hilbert spectrum on the time axis to obtain a marginal spectrum, and identify abnormal frequency components of derailment precursors based on the marginal spectrum.

[0149] S25. Combining the wheel-rail profile characteristics and the time-frequency characteristics, a wheel flange lateral position-vibration characteristic mapping table is generated.

[0150] It should be noted that the following will further illustrate, in conjunction with specific implementation methods, the extraction of visual features from the wheel-rail contact area image by combining the edge detection algorithm and the Hough transform technology, the extraction of vibration features from the vibration data, and the association of the visual features with the vibration features to generate a wheel rim lateral position-vibration feature mapping table.

[0151] Feature-level fusion (feature extraction and association of multi-source data):

[0152] 1. Visual feature extraction

[0153] Track surface partitioning: The track interface is divided into Zone I (safety zone, width 0.67L / 2), Zone II (warning zone, width 0.33L / 2), and Zone III (alarm zone, width > L / 2). The wheel-rail profile is extracted in real time through Canny edge detection + Hough transform, and the lateral offset Δ of the wheel tread centerline O is calculated. x ;

[0154] It should be noted that Canny edge detection + Hough transform is a step for real-time extraction of wheel and rail contours.

[0155] (1) Image preprocessing

[0156] Gaussian filtering: adds a "blur filter" to the image to reduce random noise interference;

[0157] Adaptive Threshold: Dynamically adjusts the black and white dividing line based on the brightness of the local area to make the wheel and track edges clearer.

[0158] (2) Canny edge detection

[0159] Compute gradients: Find areas in the image where brightness changes dramatically (e.g., wheel / track boundaries);

[0160] Non-maximum suppression: only retain the brightest edge points and "cut" the wide edges into thin lines;

[0161] Double threshold screening: Use two threshold values to filter edges. Strong edges are retained directly, and weak edges are retained only when they are connected to strong edges.

[0162] (3) Hough transform

[0163] Line detection: finding the most likely location (track) of a line among the points in an image;

[0164] Circle detection: Similar to line detection, but it is used to find circles (wheels).

[0165] (4) Contour fitting

[0166] Least squares method: find a straight line or curve so that the sum of the distances from all data points to this line is minimized;

[0167] RANSAC algorithm: Repeated random sampling to find the most "reliable" model (such as a straight track).

[0168] 2. Vibration feature extraction

[0169] Time domain characteristics: Calculate the root mean square value (RMS), crest factor, and kurtosis of the acceleration signal to characterize the vibration impact intensity;

[0170] (1) The formula for calculating the root mean square value (RMS) is:

[0171] ;

[0172] Where, The acceleration signal Sampling point value (unit: , acceleration due to gravity), Indicates the number of sampling points, RMS value Used to characterize the energy intensity of vibration signals and reflect the average energy level of vibration.

[0173] (2) The calculation formula of Peak Factor is:

[0174] Crest Factor ;

[0175] Where: Indicates the absolute peak value of the acceleration signal (unit: ); Represents the root mean square value; the crest factor is used to measure the relative intensity of the impact peak in the vibration signal. The larger the crest factor, the more severe the impact (such as the moment of wheel-rail collision).

[0176] (3) The calculation formula of Kurtosis is:

[0177] Kurtosis ;

[0178] Where, Indicates the mean value of the acceleration signal (unit: ), Indicates the standard deviation of the acceleration signal (unit: ), kurtosis is used to describe the amplitude distribution of vibration signals. The kurtosis of normal vibration signals is close to 3 (normal distribution). Sudden impacts will cause the kurtosis to increase significantly, which is used to identify abnormal impact events (such as high-frequency micro-shocks that are precursors to derailment).

[0179] Frequency domain features: Adaptive time-frequency fusion FFT algorithm is applied to the acceleration signal to extract the wheel-rail coupling characteristic frequency.

[0180] It should be noted that the FFT feature extraction implementation steps include:

[0181] (1) Acceleration signal preprocessing

[0182] Sampling parameters: According to the documentation, the vibration sensor collects triaxial acceleration signals at a sampling rate of 10kHz ( ), each analysis takes a data segment of length N (e.g. N = 1024 points, corresponding to a duration of 102.4ms).

[0183] Denoising: Use mean filtering to remove DC drift (such as , represents the signal mean);

[0184] The wheel-rail coupling related frequency band (fundamental frequency 25 Hz and its harmonics, high frequency impulses > 500 Hz are processed by wavelet packets) is retained by bandpass filtering (e.g. 5-500 Hz).

[0185] (2) Windowing

[0186] Window function selection: Use Hanning Window to reduce spectrum leakage. The window function expression is:

[0187] ;

[0188] Where, Indicates the length of the window function, usually a power of 2 (such as 512, 1024); Indicates the discrete sampling point number of the time domain signal ( ); Windowing is used to reduce the spectrum diffusion caused by signal truncation and improve the calculation accuracy of frequency component amplitude.

[0189] (3) Adaptive time-frequency fusion FFT (Fast Fourier Transform) algorithm

[0190] The windowed signal Perform N-point FFT to obtain a frequency domain complex sequence X ( k );

[0191] Spatiotemporal linkage preprocessing: Dynamically adjust the FFT parameters of vibration signals in different areas according to the I / II / III zones divided by the visual coordinate system:

[0192] Zone I (safe zone): Use long-window FFT (e.g., N=4096) to improve the resolution of the low-frequency band (25 Hz fundamental frequency) and monitor long-term vibration trends;

[0193] Zone II / III (warning / alarm zone): Switch to short-window FFT (e.g., N=256) + overlap sampling (50% overlap rate) to capture instantaneous frequency changes of high-frequency shocks (>500Hz).

[0194] Cross-modal weighting: Introducing visual evidence weights into FFT magnitude calculations (such as Zone II =0.8, area I =0.2), enhancing the frequency characteristic response of the abnormal area:

[0195] ;

[0196] Where, Indicates the visual area trigger factor (1 for area II / III, 0 for area I).

[0197] Compute the amplitude spectrum:

[0198] ;

[0199] Where, Represents the frequency domain complex sequence after visual evidence weight adjustment; Re represents the real part of the complex number; It represents the imaginary part of a complex number.

[0200] Convert to a single-sided spectrum: Take the first N / 2 points and multiply the amplitude by 2 / N (keeping the DC component amplitude unchanged).

[0201] Frequency resolution:

[0202] ;

[0203] Where, Indicates frequency resolution; Indicates the sampling rate of the acceleration signal; Indicates the number of sample points when performing fast Fourier transform.

[0204] (4) Wheel-rail coupling characteristic frequency extraction

[0205] Fundamental Frequency Identification: Searching in the Spectrum The peak value within the range confirms the fundamental frequency of the wheelset passing through the track joint (such as 25Hz mentioned in the document);

[0206] Harmonic detection: Check the amplitude of the integer multiple frequency points such as 50Hz, 75Hz, etc. in turn to build a harmonic sequence ,in, Indicates the fundamental vibration frequency generated by the wheelset through the track joint, etc. Indicates the harmonic order, which is a positive integer ( =1,2,3…), used to represent the integer multiple frequencies of the fundamental frequency.

[0207] Characteristic parameters: extract the amplitude of each frequency point and phase , for subsequent spatiotemporal association with visual data.

[0208] Time-frequency characteristics: Adaptive EMD (Empirical Mode Decomposition) decomposition based on the visual area is used to generate the Hilbert-Huang transform (HHT) marginal spectrum to analyze the abnormal time-frequency energy distribution of derailment precursors.

[0209] It should be noted that the Hilbert-Huang transform implementation steps include:

[0210] (1) Adaptive EMD decomposition based on visual area

[0211] Input signal: Single-axis acceleration signal collected by vibration sensor (such as longitudinal acceleration ), sampling rate 10kHz, data length (Corresponding to 10,000 points).

[0212] Adaptive decomposition: Dynamically adjust the EMD decomposition parameters based on the visual division of zones I / II / III (rim displacement) in the document:

[0213] Zone I (safe zone): Coarse-grained decomposition (limiting the number of IMF layers to ≤3) is used to focus on extracting the energy distribution of the low-frequency fundamental frequency (25 Hz) and its harmonics;

[0214] Zone II / III (early warning / alarm zone): trigger fine-grained decomposition (IMF layer number ≥ 5) and introduce visual constraints ( Indicates the lateral offset of the wheel rim, represents the weight coefficient), and the high-frequency impact modes are forced to be decomposed (for example, the micro-impact in the early stage of track climbing corresponds to IMF4-IMF5).

[0215] Decomposition process:

[0216] Determine the extreme points: Find All local maxima and minima of .

[0217] Construct envelope: Fit the upper and lower envelopes by cubic spline interpolation to ensure that the two envelopes are tangent everywhere in the data segment and the envelope mean is .

[0218] Extract IMF: Calculate preliminary intrinsic mode functions (IMFs):

[0219] ;

[0220] Screening criteria: Preliminary eigenmode functions Repeat steps 1-3 until the standard deviation criterion is met ( ), get the first IMF component .

[0221] Recursive decomposition: Subtract from the original signal , and obtain the residual signal ,right Repeat the decomposition until the residual signal is a monotonic function, and we get IMF components and the residual .

[0222] (2) Hilbert transform and marginal spectrum generation

[0223] Hilbert transform: For each IMF component Perform Hilbert transform to obtain the analytical signal:

[0224] ;

[0225] Where, Represents the Hilbert transform operator, the amplitude of the analytical signal is the instantaneous amplitude , Indicates the phase, Indicates the IMF components, the instantaneous frequency calculation formula is:

[0226] ;

[0227] Where, represents the instantaneous frequency; Indicates the change in phase over time (radians); Indicates the time unit (seconds); Indicates the rate of change of phase (radians per second).

[0228] Hilbert spectrum construction: superimpose the instantaneous amplitude and instantaneous frequency of all IMFs over time to form the Hilbert spectrum , characterizes the time-frequency energy distribution of the signal.

[0229] Marginal spectrum calculation: Integrate the Hilbert spectrum on the time axis to obtain the marginal spectrum :

[0230] ;

[0231] The marginal spectrum reflects the total energy distribution of each frequency component in the signal and is used to identify abnormal frequency components that are precursors to derailment.

[0232] 3. Feature association modeling: Establish a wheel rim lateral position-vibration feature mapping table, as shown in Table 1.

[0233] Table 1: Wheel rim lateral position-vibration characteristic mapping table

[0234]

[0235] S3. A DS evidence theory fusion model is established based on the wheel flange lateral position-vibration feature mapping table. The visual features and vibration features are fused using the DS evidence theory fusion model, and the train derailment risk is determined based on the feature fusion results.

[0236] Among them, a DS evidence theory fusion model is established based on the wheel flange lateral position-vibration feature mapping table. The DS evidence theory fusion model is used to fuse visual features with vibration features. The train derailment risk is determined based on the feature fusion results, including:

[0237] S31. defining a derailment risk state set based on the divided visual area;

[0238] S32. Calculate the confidence level of each risk state in the derailment risk state set based on the area where the wheel tread centerline is located and use it as visual evidence;

[0239] S33. Calculate the confidence level based on the degree of acceleration anomaly and the proportion of high-frequency components and use it as vibration evidence;

[0240] S34. Based on the predefined Dempster synthesis rule, the visual evidence and the vibration evidence are fused, and the train derailment risk is determined based on the fused features.

[0241] It should be noted that the following will further explain, in conjunction with specific implementation methods, how to establish a DS evidence theory fusion model based on the wheel flange lateral position-vibration feature mapping table, how to use the DS evidence theory fusion model to fuse visual features with vibration features, and how to determine the risk of train derailment based on the feature fusion results.

[0242] Decision-level fusion (multi-dimensional evidence synthesis and risk assessment):

[0243] 1.DS Evidence Theory Fusion Framework

[0244] Identification framework: defining a set of derailment risk states ={safety, warning, alarm};

[0245] Basic Probability Allocation (BPA):

[0246] Visual evidence: Calculate the confidence level of each risk state based on the area where the center line O is located:

[0247] Zone I: ;

[0248] Zone II:

[0249] Zone III: ;

[0250] Vibration evidence: Calculate the confidence level based on the degree of acceleration abnormality (ax>60g or ay>30g) and the proportion of high-frequency components:

[0251] No exceptions: ;

[0252] Longitudinal abnormalities: ;

[0253] Horizontal and vertical double abnormalities: ;

[0254] Calculating the conflict coefficient (The conflict coefficient is the sum of the products of the probabilities of all combinations whose intersections are empty):

[0255] ;

[0256] Take "visual area II + vibration longitudinal abnormality" as an example:

[0257] Visual evidence: ;

[0258] Vibration evidence: ;

[0259] Calculating the conflict coefficient :

[0260] Calculate the sum of the products of the probability of all combinations whose intersections are empty,

[0261] ;

[0262] Substituting the values into the equation, we get: ;

[0263] Calculate the basic probability distribution value after fusion:

[0264] Warning status:

[0265] ;

[0266] Substituting the values:

[0267] ;

[0268] Alarm status:

[0269] ;

[0270] Substituting the values:

[0271] ;

[0272] Security Status:

[0273] ;

[0274] By integrating the Dempster synthesis rule, the system fuses the evidence from two different sources, visual and vibration, to produce a more comprehensive and reliable risk assessment. In the above example, the confidence level of the warning status after fusion significantly increased to 0.89, exceeding the confidence level of 0.8 for visual warnings alone and 0.7 for vibration warnings alone. This indicates that by combining information from both sensors, the system is more confident in determining that the train is in a warning state. Compared to the judgment of a single sensor, this reduces the risk of misjudgment, provides a more accurate basis for subsequent warning and braking decisions, and enhances the reliability and accuracy of the train derailment monitoring system.

[0275] Risk Assessment:

[0276] Three-coordinate sensors are installed at each axlebox position on the bogie to collect acceleration in three directions (primarily longitudinal and lateral) at each wheel position. When derailment is imminent, the wheel flange will inevitably climb the rail, which will inevitably cause abnormal longitudinal acceleration (ax) and even lateral acceleration (ay). However, longitudinal acceleration is not necessarily a precursor to derailment and may also be radial runout of the wheel flange. Therefore, it is necessary to combine the vision system to determine the relative position of the wheel and rail.

[0277] like Figure 3 As shown, the overall multi-dimensional derailment detection method logic includes:

[0278] The entire multi-dimensional derailment system based on vision and vibration uses a bogie as a basic judgment unit, in which each wheel has a visual and vibration monitoring unit (visual equipment and vibration sensor). The sensor ax>60g and ay>30g are judged as abnormal acceleration.

[0279] a. If the visual judgment center line O is in zone I:

[0280] There is no abnormality in the longitudinal acceleration of the vibration unit, no early warning alarm, and the train does not move;

[0281] The abnormal longitudinal acceleration of the vibration unit ax>60g, and the warning confidence of the composite judgment is greater than the single warning confidence, a level 1 warning occurs, and the vehicle does not slow down.

[0282] b. If the visual judgment center line O is in Zone II:

[0283] If the longitudinal acceleration of the vibration unit is normal and the composite warning confidence is greater than the single warning confidence, a Level 1 warning is issued and the vehicle does not decelerate.

[0284] The abnormal longitudinal acceleration of the vibration unit ax>60g, and the warning confidence of the composite judgment is greater than the single warning confidence, a level 2 warning is issued and the vehicle slows down.

[0285] c. If the visual judgment unit is in Zone III:

[0286] In case of any data condition of the vibration unit, and the alarm confidence of the composite judgment is greater than the single alarm confidence, an alarm will be issued and the vehicle will be braked urgently.

[0287] d. If ax>60g and ay>30g, and the composite alarm confidence is greater than the single alarm confidence, both are directly alarmed and the vehicle is braked urgently.

[0288] It should be noted that the present invention can be applied to a new rail transit vehicle derailment monitoring system, which mainly consists of a bogie equipped with four sets of binocular cameras and four three-coordinate sensors, and equipped with a pre-processor (mounted on the bogie or vehicle chassis) to pre-process the signal and then transmit it to the host, thereby accurately issuing early warning and alarm at the derailment site, effectively avoiding the hazards caused by train derailment.

[0289] The present invention uses an adaptive time-frequency fusion FFT algorithm and an adaptive visual area EMD decomposition to achieve an adaptive decomposition design driven by the visual area, deeply couples the vibration time-frequency analysis with the spatial position information, and forms a technical path of "visual guidance-vibration precision judgment".

[0290] Through the three-level architecture of "physical layer synchronization - feature layer association - decision-making layer intelligence", the present invention realizes the optimization of the entire chain from data collection to risk decision-making, significantly improves the reliability and intelligence level of train derailment monitoring, and reduces the false alarm rate of a single sensor through multi-source data cross-validation.

[0291] The above description is only a preferred embodiment of the present invention and is not intended to limit the present invention. Any modifications, equivalent substitutions, improvements, etc. made within the spirit and principles of the present invention should be included in the scope of protection of the present invention.

Claims

1. A multi-dimensional train derailment monitoring method based on vision and vibration, characterized in that: The method includes: Collect wheel-rail contact area images and vibration data of each wheel in different directions, and align the wheel-rail contact area images and vibration data in time and space; Combining edge detection algorithms and Hough transform technology, visual features are extracted from the wheel-rail contact area image, and vibration features are extracted from the vibration data. The visual features and vibration features are then correlated to generate a wheel flange lateral position-vibration feature mapping table. A DS evidence theory fusion model is established based on the wheel flange lateral position-vibration feature mapping table. The visual features and vibration features are fused using the DS evidence theory fusion model, and the train derailment risk is determined based on the feature fusion results. The method of extracting visual features from the wheel-rail contact area image by combining the edge detection algorithm and the Hough transform technology, and extracting vibration features from the vibration data includes: Combined with edge detection algorithm and Hough transform technology, wheel-rail profile features in the wheel-rail contact area image are extracted in real time, and the lateral offset of the wheel tread centerline is calculated; Calculate the time domain characteristics of the acceleration signal in the vibration data, and use the adaptive time-frequency fusion fast Fourier transform algorithm to extract the frequency domain characteristics of the acceleration signal; Based on the time-domain and frequency-domain characteristics of the acceleration signal, the adaptive empirical mode decomposition algorithm is used to generate the transform marginal spectrum and analyze the time-frequency characteristics of the derailment precursor. The method of combining the edge detection algorithm and the Hough transform technology to extract the wheel-rail profile features in the wheel-rail contact area image in real time and calculate the lateral offset of the wheel tread centerline also includes: Dividing a visual area in a visual coordinate system based on predefined division conditions, wherein the visual area includes a safety area, a warning area, and an alarm area; The method further includes: generating a transform marginal spectrum based on the time domain characteristics and frequency domain characteristics of the acceleration signal by using an adaptive empirical mode decomposition algorithm to analyze the time-frequency characteristics of the derailment precursor; Combining wheel-rail profile features with time-frequency features, a wheel flange lateral position-vibration feature mapping table is generated; The method of extracting the frequency domain features of the acceleration signal by using the adaptive time-frequency fusion fast Fourier transform algorithm includes: Preprocess the acceleration signal, retain the wheel-rail coupling related frequency band, and use the Hanning window function to perform windowing processing on the wheel-rail coupling related frequency band; The Fourier transform parameters of the windowed acceleration signals in different visual areas are dynamically adjusted, and the visual evidence weight is introduced into the calculation of the Fourier transform parameters to obtain the amplitude spectrum. The amplitude spectrum is converted into a single-sided spectrum, and the peak value and amplitude within the preset range are searched in the single-sided spectrum. The wheel-rail coupling characteristic frequency is obtained according to the peak value and amplitude of each frequency point.

2. A multi-dimensional train derailment monitoring method based on vision and vibration according to claim 1, characterized in that, The collecting of the wheel-rail contact area image and the vibration data of each wheel in different directions respectively, and performing spatiotemporal alignment of the wheel-rail contact area image and the vibration data comprises: Video equipment is used to capture images of the wheel-rail contact area, and acceleration sensors are used to collect triaxial acceleration of the wheel to obtain vibration data. The wheel-rail contact area images and vibration data are then time-aligned using a clock protocol. A visual coordinate system is established based on the track plane with the track center reference point as the origin, and a vibration coordinate system is established with the bogie center as the origin; The vibration data of the vibration coordinate system is mapped to the visual coordinate system through the rigid body transformation matrix for spatial alignment.

3. A multi-dimensional train derailment monitoring method based on vision and vibration according to claim 2, characterized in that, Mapping the vibration data of the vibration coordinate system to the visual coordinate system through the rigid body transformation matrix for spatial alignment includes: Use a laser rangefinder to collect the installation status of the acceleration sensor and measure the Euler angle of the vibration coordinate system relative to the visual coordinate system; Calculate the rotation matrix based on the Euler angle and the rotation matrix synthesis formula, measure the coordinates of the origin of the vibration coordinate system in the visual coordinate system and determine the translation vector; The vibration data of the vibration coordinate system are converted into components of the visual coordinate system through the rotation matrix and translation vector, and spatially associated with the wheel rim position in the visual coordinate system; The wheel flange position in the actual wheel-rail contact area image is compared with the spatial distribution of vibration data, and the rotation matrix and translation vector are optimized.

4. A multi-dimensional train derailment monitoring method based on vision and vibration according to claim 1, characterized in that, The time-domain characteristics and frequency-domain characteristics of the acceleration signal are based on the adaptive empirical mode decomposition algorithm to generate the transformation marginal spectrum and analyze the time-frequency characteristics of the derailment precursor, including: The empirical mode decomposition algorithm is used to dynamically adjust the empirical mode decomposition parameters of the visual area to obtain several intrinsic mode function components; Perform Hilbert transform on each intrinsic mode function component to obtain an analytical signal, and calculate the instantaneous amplitude and instantaneous frequency of the analytical signal; The instantaneous amplitudes and instantaneous frequencies of all eigenmode function components are superimposed over time to form the Hilbert spectrum. The Hilbert spectrum is integrated on the time axis to obtain the marginal spectrum, and the abnormal frequency components of the derailment precursor are identified based on the marginal spectrum.

5. A multi-dimensional train derailment monitoring method based on vision and vibration according to claim 4, characterized in that, The empirical mode decomposition algorithm is used to dynamically adjust the empirical mode decomposition parameters of the visual area to obtain several intrinsic mode function components including: Select the decomposition method of the empirical mode decomposition algorithm based on different visual areas; Obtain the local maximum and minimum values of the acceleration signal, connect the maximum points to form the upper envelope, connect the minimum points to form the lower envelope, and fit the upper and lower envelopes through cubic spline interpolation; Calculate the preliminary intrinsic mode function based on the fitted upper and lower envelopes, iterate the preliminary intrinsic mode function repeatedly until the standard deviation criterion is met, and obtain the first intrinsic mode function component; The first intrinsic mode function component is subtracted from the original acceleration signal to obtain a residual signal, and the residual signal is repeatedly decomposed until the residual signal is a monotonic function to obtain several intrinsic mode function components.

6. A multi-dimensional train derailment monitoring method based on vision and vibration according to claim 5, characterized in that: The decomposition methods of the empirical mode decomposition algorithm based on different visual areas include: In the safe zone, the energy distribution of the low-frequency fundamental frequency and its harmonics is extracted using a coarse-grained decomposition method; In the early warning area and alarm area, fine-grained decomposition is used and visual constraints are introduced to decompose high-frequency impact modes.

7. The multi-dimensional train derailment monitoring method based on vision and vibration according to claim 5 is characterized in that: The DS evidence theory fusion model is established based on the wheel rim lateral position-vibration feature mapping table, and the visual features and vibration features are combined using the DS evidence theory fusion model. Fusion, based on the feature fusion results, determines the risk of train derailment, including: Determine the derailment risk state set based on the divided visual area; According to the area where the wheel tread centerline is located, the confidence level of each risk state in the derailment risk state set is calculated and used as visual evidence; Calculate the confidence level based on the degree of acceleration anomaly and the proportion of high-frequency components and use it as vibration evidence; Based on predefined synthesis rules, visual evidence and vibration evidence are fused, and the risk of train derailment is determined based on the fused features.

8. The multi-dimensional train derailment monitoring method based on vision and vibration according to claim 3 is characterized in that: The rotation matrix synthesis formula is: ； Where, represents the rotation matrix; Indicates winding The basic rotation matrix of the axis; Indicates winding The basic rotation matrix of the axis; Indicates winding The basic rotation matrix of the axis; Indicates the pitch angle; represents the yaw angle; Indicates the roll angle.

Images