A self-diagnostic method for abnormalities in highway traffic overload detection equipment

Abstract

This invention provides a method for self-diagnosing anomalies in highway traffic overload control detection equipment. The method includes: collecting time-domain and frequency-domain features of sensor data samples from a laboratory environment as the source domain; collecting relevant data from dynamic truck scales at overload control stations and calculating corresponding time-domain and frequency-domain features as the target domain; fusing the time-domain and frequency-domain features to obtain corresponding source domain samples and unlabeled target domain samples; mapping these samples to the feature space again using kernel function processing and the Transfer Component Analysis (TCA) method to obtain source domain sample features and target domain sample features; training a pre-defined WKNN classifier using the TCA-transferred source and target domain labeled features to output a target domain pseudo-label; retraining the WKNN classifier using the target domain pseudo-label samples and the source domain to obtain a trained WKNN classifier; and using the trained WKNN classifier to diagnose anomalies in dynamic truck scales. This method enables real-time self-diagnosis of anomalies in overload control equipment.

Description

Technical Field

[0001] This invention relates to the field of abnormal detection technology for overload control station equipment, and in particular to a self-diagnosis method for abnormalities in highway traffic overload control detection equipment. Background Technology

[0002] Overload control devices are crucial for ensuring road safety, maintaining traffic order, and reducing traffic accidents. They can promptly detect overloaded vehicles, preventing damage to bridges, roads, and other infrastructure, thereby reducing maintenance and repair costs. With the continuous advancement of overload control efforts, more and more off-site detection points are being built. These points are equipped with various detection devices, such as lidar for contour detection, and weighing instruments, load cells, and loop detectors for overload detection. However, these detection devices are prone to malfunctions when exposed to harsh environments or improper installation or use, resulting in the inability to properly weigh passing vehicles. Therefore, regular calibration of measuring instruments such as dynamic weighbridges and contour measuring devices used in road transport is extremely important for supervising and enforcing traffic regulations, preventing illegal overloading, improving the efficiency of traffic facilities, and ensuring the safety and reliability of the road network. However, due to the wide distribution and large number of measuring instruments such as dynamic truck scales and contour measuring devices in road transportation, traditional inspection methods require manual inspection one by one, which is inefficient and increases the difficulty and complexity of the inspection work. Currently, some methods use fuzzy algorithms to calculate abnormal situations, but these algorithms are not very accurate and have low practicality in actual use.

[0003] Therefore, a more efficient and accurate automatic detection method for faults in overload control station equipment is needed to better ensure road safety, maintain traffic order, and reduce traffic accidents. Summary of the Invention

[0004] The technical problem to be solved by the present invention is to provide a self-diagnosis method for abnormalities in highway traffic overload detection equipment, thereby solving the problems of low detection efficiency and low detection accuracy of existing equipment abnormalities.

[0005] This invention provides a self-diagnostic method for abnormalities in highway traffic overload detection equipment, the method comprising the following steps:

[0006] Step S1: Collect the time-domain and frequency-domain characteristics of weighing sensor data samples under various fault types in the laboratory, and use them as the source domain;

[0007] Step S2: Collect relevant data of dynamic truck scales at the overload control stations, calculate the output value of each weighing sensor through the correlation model, and calculate the time domain characteristics and frequency domain characteristics based on the output values ​​of each weighing sensor as the target domain;

[0008] Step S3: Fuse the time-domain features and frequency-domain features of the source domain and the target domain respectively to obtain the processed source domain sample and the unlabeled target domain sample.

[0009] Step S4: Process the source domain samples and target domain samples using kernel functions, and then map them to the feature space using the transfer component analysis (TCA) method to obtain the features of the source domain samples and the features of the target domain samples.

[0010] Step S5: Use the labeled features of the source and target domains after TCA feature transfer to train the preset WKNN classifier and output the target domain pseudo label;

[0011] Step S6: Merge the target domain pseudo-label samples with the source domain to form an expanded labeled dataset. Retrain the WKNN classifier using the expanded dataset to obtain the trained WKNN classifier.

[0012] Step S7: Collect relevant data from any dynamic truck scale at the overload control station in real time, preprocess it, and input it into the trained WKNN classifier to output the predicted fault type.

[0013] Furthermore, the fault types include substantial faults and performance degradation faults, with substantial faults including load cell faults, foundation and weighing platform faults, and junction box and wiring faults.

[0014] Furthermore, the relevant data of the dynamic truck scale at the overload control station includes environmental monitoring data, weighing instrument data, multiple sensor data used for weighing each dynamic truck scale, lidar data, camera and coil vehicle detector.

[0015] Furthermore, step S2 specifically includes:

[0016] Step S21: Collect the values ​​of multiple weighing sensors corresponding to the dynamic truck scale at the overload control station, and use a Butterworth filter to filter out dynamic signal interference.

[0017] Step S22: Calculate the output value of each weighing sensor using an association model; specifically, the association model is: ;

[0018] in, Let j be the output value of the j-th weighing sensor on the dynamic truck scale. Let M be the value of the nth weighing sensor on the dynamic truck scale, and M be the number of weighing sensors corresponding to the current dynamic truck scale. For the corresponding number The coefficient between the first load cell and the remaining M-1 load cells. , , and These are the vehicle speed factor, load type factor, and load factor under dynamic conditions, respectively. The correlation weights for these three influencing factors;

[0019] Step S23: Calculate the time-domain and frequency-domain characteristics based on the output values ​​of each weighing sensor, using this as the target domain. The time-domain characteristics include the root mean square value, impulse index, kurtosis index, margin index, peak value, average amplitude, and waveform index. The frequency-domain characteristics include the standard deviation frequency, centroid frequency, root mean square frequency, and mean frequency. The calculation expressions for each characteristic are as follows:

[0020] Root mean square value: ;

[0021] Pulse Indicators: ;

[0022] Kurtosis index: ;

[0023] Margin indicators: ;

[0024] Peak metrics: ;

[0025] Average amplitude: ;

[0026] Waveform Indicators: ;

[0027] Standard deviation frequency: ;

[0028] Center of gravity frequency: ;

[0029] Root mean square frequency: ;

[0030] Mean frequency: ;

[0031] in, This represents the amplitude of the i-th point of the output signal from a sensor, where N represents the number of signal points. This represents the frequency value of the i-th spectral line in the power spectrum. This indicates a reference output value.

[0032] Furthermore, before step S3, the method further includes: performing PCA or KPCA feature dimensionality reduction analysis on the source domain and target domain respectively to obtain the variance explanation rate of each principal component, and then fusing the time domain features and frequency domain features to obtain labeled weighing sensor samples and unlabeled truck scale samples respectively. In this process, a reference model of the feature signals under different fault types is established to determine the threshold range under different faults.

[0033] Furthermore, in step S4, "by mapping TCA features to the feature space, we obtain the source domain sample features and the target domain sample features" specifically means:

[0034] TCA feature transfer is used to map source and target domain samples to the regenerating kernel Hilbert space. By minimizing the maximum mean difference between source and target domain samples in the feature subspace, a dimensionality-reduced feature subspace is obtained, which reduces the distribution difference between the source and target domains in this subspace. This results in source and target domain sample features with a certain degree of similarity, thereby realizing cross-domain transfer of fault knowledge.

[0035] Furthermore, the preprocessing in step S7 includes using a Butterworth filter to filter out dynamic signal interference from the acquired weighing sensors, obtaining the output values ​​of each weighing sensor of the dynamic truck scale through an association model, calculating time-domain and frequency-domain features based on the output values ​​of each sensor, and then fusing the time-domain and frequency-domain features through dimensionality reduction processing.

[0036] Furthermore, the method also includes simultaneously judging abnormal behavior using the collected data from the dynamic truck scale, and further screening the output of step S7. If no abnormal behavior is found, it is determined to be the corresponding fault type; otherwise, a vehicle abnormal behavior warning is issued directly. The abnormal vehicle behaviors include: crossing lanes, driving against traffic, pressing against the edge, S-shaped behavior, parking, jumping the scale, dragging the scale, grinding the scale, queuing, and jacking. The judgment of these abnormal behaviors is achieved by deeply fusing the sensing data of the roadside end equipment and using a combination of multi-channel data from the weighing system and optical measurement methods to judge the vehicle weight and behavior.

[0037] Furthermore, the specific process and handling methods for judging abnormal behavior include:

[0038] The lane-crossing behavior is determined based on its characteristics: the vehicle crosses the lane centerline, part of its weight is obtained from the adjacent lane, and the weight measured by the current lane's weighing equipment is inaccurate. The determination method for lane-crossing behavior is as follows: the vehicle's position and status are estimated based on multi-channel sensor signals from all weighing equipment within the station; the driving position and direction are determined by coil signals; the vehicle's size and position are measured by lidar and compared with the weighing signal at the same moment to determine whether the vehicle is in a lane-crossing state. The lane-crossing behavior is processed as follows: the system displays an abnormal status message, and sensor signals belonging to the vehicle's total weight are matched based on the location and accumulated and uploaded.

[0039] The reverse driving behavior is determined based on its characteristics, which are: the vehicle crosses the lane centerline, its entire weight is acquired by the weighing equipment in the adjacent oncoming lane, there is no weighing data in the current lane, and an error occurs in the weighing process of the oncoming lane. The method for determining reverse driving behavior is as follows: the vehicle's position and status are estimated based on the multi-channel sensor signals of all weighing equipment in the station; the driving position and direction are determined by the coil signal; the vehicle size, position, and driving direction measured by the lidar are compared with the weighing signal at the same time to determine whether the vehicle is in a reverse driving state; the reverse driving behavior is processed as follows: the abnormal status is indicated in the system, and the sensor signal of the oncoming lane that belongs to the total weight of the vehicle is matched according to the position and accumulated and uploaded; the rear camera is used as the front camera to identify the vehicle license plate number.

[0040] The edge-pressing behavior is determined based on its characteristics, which are: the weighing equipment does not cover the entire width of the lane, and the vehicle partially presses on the road surface without a weighing platform. The determination method for edge-pressing behavior is as follows: based on the multi-channel sensor signals of the weighing equipment, it is determined whether the balance of the left and right weights exceeds a threshold. Combined with the vehicle size, position, and driving direction measured by the lidar, it is compared with the weighing signal at the same moment to determine whether the vehicle is in an edge-pressing state. The edge-pressing behavior is processed as follows: the abnormal state is indicated in the system, and the left half weight of the vehicle is matched according to the position as a reference half weight. Twice of this weight is uploaded as temporary data for subsequent verification.

[0041] The S-shaped behavior is determined based on its characteristics, which are: the vehicle sways left and right as it passes through the weighing area, ensuring that the two wheels on the same axle do not pass through the weighing area in the same lane simultaneously. The determination method for S-shaped behavior is as follows: based on the multi-channel sensor signals of all weighing equipment at the station, the characteristics of peak axle load and wheel load are identified. This is combined with the vehicle's trajectory measured by lidar and compared with the weighing signals from the same period to determine if the vehicle is in an S-shaped driving state. The following processing is applied to S-shaped behavior: an abnormal status is indicated in the system, and all weights measured by the weighing equipment during the time the vehicle passes through the weighing area are uploaded as temporary data for subsequent verification.

[0042] The parking behavior is determined based on its characteristics: prolonged parking in the weighing area causes errors in the inductive loop vehicle detector signal, leading to errors in the dynamic truck scale program that calculates weight by axle, and thus preventing the correct calculation of the total weight. The method for determining parking behavior is as follows: the vehicle status is estimated based on the multi-channel sensor signals of the weighing equipment; the inductive loop signal is used to determine if the vehicle is in motion; the vehicle trajectory measured by lidar is used for confirmation; and the result is compared with the weighing signal at the same time to determine if the vehicle is parked. The parking behavior is processed as follows: an abnormal status is displayed in the system, and all weights measured by the weighing equipment during the period when all axles of the vehicle pass through the weighing area are uploaded as data.

[0043] The aforementioned "jumping the scale" behavior is determined based on its characteristics, which are: the vehicle accelerates from a standstill, causing the center of gravity to shift rearward, thereby reducing the measured front axle weight. The determination method for the "jumping the scale" behavior is: judging its acceleration behavior based on the vehicle speed-time curve obtained from LiDAR measurement. The "jumping the scale" behavior is handled as follows: an abnormal status is displayed in the system.

[0044] The towing behavior is determined based on its characteristics, which are: the vehicle drags on the weighing equipment by locking the trailer axle, increasing the lateral force on the load cell and causing the load cell to jam, thus reducing the measured trailer weight; the towing behavior is determined by judging its acceleration and deceleration behavior based on the vehicle speed-time curve obtained by lidar measurement; the towing behavior is handled as follows: the abnormal status is displayed in the system.

[0045] The scale-grinding behavior is judged based on its characteristics, which are: the vehicle repeatedly moves forward and backward in the weighing area, causing errors in the cumulative axle count and total weight. The judgment method for scale-grinding behavior is: based on the weight information of all weighing equipment in the station, the vehicle's forward and backward state is estimated, as well as the axle weight that is repeatedly weighed. Combined with the speed-time curve obtained by lidar measurement, it is compared with the weighing signal at the same moment to determine the weight of the vehicle that is repeatedly weighed. The scale-grinding behavior is processed as follows: the abnormal state is indicated in the system, and the weight of repeated weighing and the number of axles repeatedly calculated are removed.

[0046] The queuing behavior is judged based on its characteristics, which are: multiple vehicles queuing closely to reduce the distance between vehicles, intending to interfere with the vehicle separator and cause errors in the measured weight, and preventing the camera from recognizing the license plate; the queuing behavior is judged by: estimating the vehicle separation status based on the weight information of all weighing equipment in the station, combined with the vehicle separation information obtained by LiDAR; the queuing behavior is processed as follows: the abnormal status is indicated in the system, and the weight of each vehicle in the queue is matched according to the spatiotemporal information;

[0047] The jack behavior is determined based on its characteristics, which are: by installing a jack device at the bottom of the vehicle, the weight is transferred to a non-weighing area, reducing the measured axle load; the determination method for the jack behavior is: based on the weight of each axle obtained by the weighing equipment, it is estimated whether the axle load imbalance is caused by the jack, and at the same time, combined with the speed-time curve obtained by the lidar, it is determined whether the jack was operated while the vehicle was stopped to evade inspection; the jack behavior is handled as follows: the abnormal status is indicated in the system.

[0048] Furthermore, the method also includes step 8: after detecting a fault, using a dynamic weighing system intelligent fault-tolerant method based on an improved particle swarm optimization generalized regression neural network under a multi-association model to measure and predict passing vehicles, as detailed below:

[0049] Remove data that deviates due to faults;

[0050] The Lévy-Flight Improved Particle Swarm Optimization (LPSO-GRNN) algorithm was used to search for the optimal solution to optimize the generalized regression neural network, resulting in an LPSO-GRNN network model that matches the truck scale data samples. This model was then used to establish the output prediction network.

[0051] A multi-association model is used to determine the expected characteristic values ​​under different faults, and then the output prediction network is used to obtain the expected output value of the fault sensor, thereby realizing the measurement prediction value of passing vehicles.

[0052] The advantages of this invention are: it can realize real-time self-diagnosis of abnormalities in overload control equipment, improve the efficiency and accuracy of diagnosing equipment abnormalities, so as to maintain faulty equipment in a timely manner, ensure stable operation of equipment, and is of great significance for ensuring road safety, maintaining traffic order and reducing traffic accidents. Attached Figure Description

[0053] The present invention will be further described below with reference to the accompanying drawings and embodiments.

[0054] Figure 1 This is a flowchart illustrating the execution of an abnormal self-diagnosis method for a highway traffic overload detection device according to the present invention.

[0055] Figure 2 This is a schematic diagram illustrating the principle of an abnormal self-diagnosis method for a highway traffic overload detection device according to the present invention.

[0056] Figure 3 This is a block diagram illustrating the model building principle of the abnormal self-diagnosis method for highway traffic overload detection equipment according to the present invention. Detailed Implementation

[0057] This application provides a self-diagnosis method for abnormalities in highway traffic overload detection equipment, which can realize automatic diagnosis and analysis of abnormal data and quickly locate the cause of abnormalities.

[0058] The overall approach of the technical solution in this application is as follows: The sensor output values ​​of a dynamic truck scale are calculated using an association model to determine its time-domain and frequency-domain features, serving as the target domain. Simultaneously, sensor data under various fault conditions simulated in the laboratory are used to obtain the source domain. Addressing the challenge of online detection of dynamic faults in truck scales during service, a fault detection method for dynamic weighing systems based on kernel principal component analysis (KPCA) under a multi-association model is proposed. A Butterworth filter is used to remove dynamic signal interference, and KPCA and PCA are used to reduce the dimensionality and filter features, determining feature indicators and corresponding control limits to construct a comprehensive fault evaluation method. Furthermore, addressing the problem of insufficient on-site fault data collection leading to decreased diagnostic performance, a cross-sensor fault diagnosis method based on transfer component analysis (TCA) combined with weighted k-nearest neighbor (WKNN) algorithm is adopted under multi-feature fusion. TCA maps sensor samples and truck scale samples to the feature space, thereby reducing distribution differences. WKNN is used as a classifier to classify and match the data. This method achieves a diagnostic accuracy of 93.33% with a support set sample size of 3 for each fault state, effectively avoiding the problem of imbalanced sample distribution. The invention also diagnoses abnormal behavior in the data, eliminating false fault outputs caused by abnormal behavior, thus improving the accuracy of fault analysis. Furthermore, by predicting the expected output of faulty sensors, it improves the accuracy of predicting oncoming vehicles under fault conditions.

[0059] To better understand the above technical solutions, the following will provide a detailed explanation of the technical solutions in conjunction with the accompanying drawings and specific implementation methods.

[0060] like Figures 1 to 3 As shown in the figure, this embodiment provides a self-diagnosis method for abnormalities in highway traffic overload detection equipment, the method including the following steps:

[0061] Step S1: Collect the time-domain and frequency-domain characteristics of weighing sensor data samples under various fault types in the laboratory, and use them as the source domain;

[0062] Step S2: Collect relevant data of dynamic truck scales at the overload control stations, calculate the output value of each weighing sensor through the correlation model, and calculate the time domain characteristics and frequency domain characteristics based on the output values ​​of each weighing sensor as the target domain;

[0063] Step S3: Fuse the time-domain features and frequency-domain features of the source domain and the target domain respectively to obtain the processed source domain sample and the unlabeled target domain sample.

[0064] Step S4: Process the source domain samples and target domain samples respectively using kernel functions, and then map them to the feature space using TCA feature transfer to obtain the source domain sample features and target domain sample features. Preferably, TCA feature transfer is used to map the source domain and target domain samples to the regenerating kernel Hilbert space. By minimizing the maximum mean difference between the source domain and target domain samples in the feature subspace, a dimensionality-reduced feature subspace is obtained. The distribution difference between the source domain and target domain is reduced in this subspace, thereby realizing the cross-domain transfer of fault knowledge.

[0065] Specifically, for vibration-type load cells, the vibration signal of the load cell is first acquired. The vibration signal of the load cell with a known fault state is used as the source domain sample, and the vibration signal of the dynamic weighing system with an unknown fault state is used as the target domain sample. After processing the differences between the source domain samples and the target domain samples using the TCA algorithm, the similarity between the source domain data samples and the target domain samples is increased. The source domain data to be migrated is determined by the maximum mean difference, which solves the problem of insufficient labeled data supply in the target domain, i.e., a small number of actual fault samples, so as to form a cross-equipment fault diagnosis model for dynamic weighing systems based on TCA-WKNN.

[0066] Step S5: Use the labeled features of the source and target domains after TCA feature transfer to train the preset WKNN classifier and output the target domain pseudo-label; WKNN realizes cross-domain transfer of fault knowledge. After the previous series of operations, the WKNN algorithm can effectively capture the potential feature differences between the source and target domains, allowing some target domains to also be labeled. Semi-supervised learning is performed again, and finally the class with the highest prediction probability in the target domain that is not labeled is labeled with pseudo-labels, so that all target domains are labeled.

[0067] Step S6: Filter the target domain pseudo-labels to obtain the filtered target domain pseudo-label samples. Use the filtered target domain pseudo-label samples to merge with the source domain to form an expanded labeled dataset. Use the expanded dataset to retrain the WKNN classifier to obtain the trained WKNN classifier.

[0068] Step S7: Collect relevant data from any dynamic truck scale at the overload control station in real time, preprocess it, and input it into the trained WKNN classifier to output the predicted fault type.

[0069] The steps S1 and S2 described above can be performed in any order.

[0070] Preferably, the fault types include substantial faults and performance degradation faults, wherein substantial faults include load cell faults, foundation and platform faults, and junction box and wiring faults.

[0071] Preferably, the relevant data of the dynamic truck scale at the overload control station includes environmental monitoring data, weighing instrument data, multiple sensor data used for weighing each dynamic truck scale, lidar data, camera and coil vehicle detector.

[0072] Preferably, step S2 specifically comprises:

[0073] Step S21: Collect the values ​​of multiple weighing sensors corresponding to the dynamic truck scale at the overload control station, and use a Butterworth filter to filter out dynamic signal interference.

[0074] Step S22: Calculate the output value of each weighing sensor using an association model; specifically, the association model is: ;

[0075] The above formula can be expanded as follows: ;

[0076] in, Let j be the output value of the j-th weighing sensor on the dynamic truck scale. Let M be the value of the nth weighing sensor on the dynamic truck scale, and M be the number of weighing sensors corresponding to the current dynamic truck scale. For the corresponding number The coefficient between the first load cell and the remaining M-1 load cells. , , and These are the vehicle speed factor, load type factor, and load factor under dynamic conditions, respectively. The correlation weights for these three influencing factors;

[0077] Step S23: Calculate the time-domain and frequency-domain characteristics based on the output values ​​of each weighing sensor, using this as the target domain. The time-domain characteristics include the root mean square value, impulse index, kurtosis index, margin index, peak value, average amplitude, and waveform index. The frequency-domain characteristics include the standard deviation frequency, centroid frequency, root mean square frequency, and mean frequency. The calculation expressions for each characteristic are as follows:

[0078] Root mean square value: ;

[0079] Pulse Indicators: ;

[0080] Kurtosis index: ;

[0081] Margin indicators: ;

[0082] Peak metrics: ;

[0083] Average amplitude: ;

[0084] Waveform Indicators: ;

[0085] Standard deviation frequency: ;

[0086] Center of gravity frequency: ;

[0087] Root mean square frequency: ;

[0088] Mean frequency: ;

[0089] in, This represents the amplitude of the i-th point of the output signal from a sensor, where N represents the number of signal points. This represents the frequency value of the i-th spectral line in the power spectrum. This indicates a reference output value.

[0090] Preferably, before step S3, the method further includes: performing PCA or KPCA feature dimensionality reduction analysis on the source domain and the target domain respectively to obtain the variance explanation rate of each principal component, and then fusing the time domain features and frequency domain features to obtain labeled weighing sensor samples and unlabeled truck scale samples respectively. In this way, by establishing reference models of feature signals under different fault types, the threshold range under different faults is determined.

[0091] Through the above steps S1-S3, the time domain and frequency domain features of the source domain and target domain samples are extracted, and then fused using PCA feature fusion technology to obtain a labeled source domain sample feature set and an unlabeled target domain sample feature set, thus realizing feature extraction and fusion.

[0092] Preferably, the preprocessing in step S7 includes using a Butterworth filter to filter out dynamic signal interference from the acquired weighing sensors, obtaining the output values ​​of each weighing sensor of the dynamic truck scale through an association model, calculating time-domain and frequency-domain features based on the output values ​​of each sensor, and then fusing the time-domain and frequency-domain features through dimensionality reduction processing.

[0093] Preferably, the method further includes simultaneously judging abnormal behavior using the relevant data collected from the dynamic truck scale, and further screening the output of step S7. If no abnormal behavior is found, it is determined to be the corresponding fault type; otherwise, a vehicle abnormal behavior warning is directly issued. The abnormal vehicle behaviors include: crossing lanes, driving against traffic, pressing against the edge, S-shaped behavior, parking, jumping the scale, dragging the scale, grinding the scale, queuing, and jacking. The judgment of these abnormal behaviors is achieved by deeply fusing the sensing data of the roadside end equipment and using a combination of multi-channel data from the weighing system and optical measurement methods to judge the vehicle weight and behavior.

[0094] Preferably, the specific process and handling methods for judging abnormal behavior include:

[0095] The lane-crossing behavior is determined based on its characteristics: the vehicle crosses the lane centerline, part of its weight is obtained from the adjacent lane, and the weight measured by the current lane's weighing equipment is inaccurate. The determination method for lane-crossing behavior is as follows: the vehicle's position and status are estimated based on multi-channel sensor signals from all weighing equipment within the station; the driving position and direction are determined by coil signals; the vehicle's size and position are measured by lidar and compared with the weighing signal at the same moment to determine whether the vehicle is in a lane-crossing state. The lane-crossing behavior is processed as follows: the system displays an abnormal status message, and sensor signals belonging to the vehicle's total weight are matched based on the location and accumulated and uploaded.

[0096] The reverse driving behavior is determined based on its characteristics, which are: the vehicle crosses the lane centerline, its entire weight is acquired by the weighing equipment in the adjacent oncoming lane, there is no weighing data in the current lane, and an error occurs in the weighing process of the oncoming lane. The method for determining reverse driving behavior is as follows: the vehicle's position and status are estimated based on the multi-channel sensor signals of all weighing equipment in the station; the driving position and direction are determined by the coil signal; the vehicle size, position, and driving direction measured by the lidar are compared with the weighing signal at the same time to determine whether the vehicle is in a reverse driving state; the reverse driving behavior is processed as follows: the abnormal status is indicated in the system, and the sensor signal of the oncoming lane that belongs to the total weight of the vehicle is matched according to the position and accumulated and uploaded; the rear camera is used as the front camera to identify the vehicle license plate number.

[0097] The edge-pressing behavior is determined based on its characteristics, which are: the weighing equipment does not cover the entire width of the lane, and the vehicle partially presses on the road surface without a weighing platform. The determination method for edge-pressing behavior is as follows: based on the multi-channel sensor signals of the weighing equipment, it is determined whether the balance of the left and right weights exceeds a threshold. Combined with the vehicle size, position, and driving direction measured by the lidar, it is compared with the weighing signal at the same moment to determine whether the vehicle is in an edge-pressing state. The edge-pressing behavior is processed as follows: the abnormal state is indicated in the system, and the left half weight of the vehicle is matched according to the position as a reference half weight. Twice of this weight is uploaded as temporary data for subsequent verification.

[0098] The S-shaped behavior is determined based on its characteristics, which are: the vehicle sways left and right as it passes through the weighing area, ensuring that the two wheels on the same axle do not pass through the weighing area in the same lane simultaneously. The determination method for S-shaped behavior is as follows: based on the multi-channel sensor signals of all weighing equipment at the station, the characteristics of peak axle load and wheel load are identified. This is combined with the vehicle's trajectory measured by lidar and compared with the weighing signals from the same period to determine if the vehicle is in an S-shaped driving state. The following processing is applied to S-shaped behavior: an abnormal status is indicated in the system, and all weights measured by the weighing equipment during the time the vehicle passes through the weighing area are uploaded as temporary data for subsequent verification.

[0099] The parking behavior is determined based on its characteristics: prolonged parking in the weighing area causes errors in the inductive loop vehicle detector signal, leading to errors in the dynamic truck scale program that calculates weight by axle, and thus preventing the correct calculation of the total weight. The method for determining parking behavior is as follows: the vehicle status is estimated based on the multi-channel sensor signals of the weighing equipment; the inductive loop signal is used to determine if the vehicle is in motion; the vehicle trajectory measured by lidar is used for confirmation; and the result is compared with the weighing signal at the same time to determine if the vehicle is parked. The parking behavior is processed as follows: an abnormal status is displayed in the system, and all weights measured by the weighing equipment during the period when all axles of the vehicle pass through the weighing area are uploaded as data.

[0100] The aforementioned "jumping the scale" behavior is determined based on its characteristics, which are: the vehicle accelerates from a standstill, causing the center of gravity to shift rearward, thereby reducing the measured front axle weight. The determination method for the "jumping the scale" behavior is: judging its acceleration behavior based on the vehicle speed-time curve obtained from LiDAR measurement. The "jumping the scale" behavior is handled as follows: an abnormal status is displayed in the system.

[0101] The towing behavior is determined based on its characteristics, which are: the vehicle drags on the weighing equipment by locking the trailer axle, increasing the lateral force on the load cell and causing the load cell to jam, thus reducing the measured trailer weight; the towing behavior is determined by judging its acceleration and deceleration behavior based on the vehicle speed-time curve obtained by lidar measurement; the towing behavior is handled as follows: the abnormal status is displayed in the system.

[0102] The scale-grinding behavior is judged based on its characteristics, which are: the vehicle repeatedly moves forward and backward in the weighing area, causing errors in the cumulative axle count and total weight. The judgment method for scale-grinding behavior is: based on the weight information of all weighing equipment in the station, the vehicle's forward and backward state is estimated, as well as the axle weight that is repeatedly weighed. Combined with the speed-time curve obtained by lidar measurement, it is compared with the weighing signal at the same moment to determine the weight of the vehicle that is repeatedly weighed. The scale-grinding behavior is processed as follows: the abnormal state is indicated in the system, and the weight of repeated weighing and the number of axles repeatedly calculated are removed.

[0103] The queuing behavior is judged based on its characteristics, which are: multiple vehicles queuing closely to reduce the distance between vehicles, intending to interfere with the vehicle separator and cause errors in the measured weight, and preventing the camera from recognizing the license plate; the queuing behavior is judged by: estimating the vehicle separation status based on the weight information of all weighing equipment in the station, combined with the vehicle separation information obtained by LiDAR; the queuing behavior is processed as follows: the abnormal status is indicated in the system, and the weight of each vehicle in the queue is matched according to the spatiotemporal information;

[0104] The jack behavior is determined based on its characteristics, which are: by installing a jack device at the bottom of the vehicle, the weight is transferred to a non-weighing area, reducing the measured axle load; the determination method for the jack behavior is: based on the weight of each axle obtained by the weighing equipment, it is estimated whether the axle load imbalance is caused by the jack, and at the same time, combined with the speed-time curve obtained by the lidar, it is determined whether the jack was operated while the vehicle was stopped to evade inspection; the jack behavior is handled as follows: the abnormal status is indicated in the system.

[0105] Preferably, the method further includes step 8: after detecting a fault, using a dynamic weighing system intelligent fault-tolerant method based on an improved particle swarm optimization generalized regression neural network under a multi-association model to measure and predict passing vehicles, as follows:

[0106] Remove data that deviates due to faults;

[0107] The Lévy-Improved Particle Swarm Optimization (LPSO) algorithm is used to search for the optimal solution to optimize the generalized regressive neural network (GRNN), and an LPSO-GRNN network model matching the truck scale data sample is obtained, thereby establishing the output prediction network;

[0108] A multi-association model is used to determine the expected characteristic values ​​under different fault conditions. Then, an output prediction network is used to obtain the expected output value of the faulty sensor, thereby realizing the measurement prediction value of passing vehicles. The expected characteristic values ​​refer to the signal mean frequency, standard deviation frequency, impulse index, and root mean square frequency that should occur due to sensor failure. The expected output value refers to the measurement value that the sensor should output if it is fault-free when subjected to the current load.

[0109] The technical solution provided in this application embodiment has at least the following technical effects or advantages: The present invention utilizes the TCA algorithm to process the differences between source domain samples and target domain samples, thereby increasing the similarity between source domain data samples and target domain samples. It identifies the source domain data to be migrated by the maximum mean difference and also labels the target domain features, thereby solving the problem of insufficient supply of labeled data in the target domain, i.e., a small number of actual fault samples. The distribution difference between the source domain and the target domain is reduced in this subspace, thereby realizing the cross-domain transfer of fault knowledge and forming a cross-equipment fault diagnosis model for dynamic weighing systems based on TCA-WKNN. At the same time, combined with the abnormal behavior analysis algorithm, it can realize the automatic diagnosis and analysis of abnormal conditions of equipment at the overload control station, while improving the diagnostic accuracy, improving the efficiency of equipment fault self-diagnosis, quickly locating the fault point and the fault cause, realizing efficient maintenance, improving the stability of equipment operation, and improving the working efficiency of equipment at the overload control station.

[0110] While specific embodiments of the present invention have been described above, those skilled in the art should understand that the specific embodiments described are merely illustrative and not intended to limit the scope of the present invention. Equivalent modifications and variations made by those skilled in the art in accordance with the spirit of the present invention should be covered within the scope of protection of the claims of the present invention.

Claims

1. A self-diagnostic method for abnormalities in highway traffic overload detection equipment, characterized in that: The method includes the following steps: Step S1: Collect the time-domain and frequency-domain characteristics of weighing sensor data samples under various fault types in the laboratory, and use them as the source domain; Step S2: Collect relevant data from the dynamic truck scale at the overload control station, and calculate the output value of each weighing sensor through an association model. Calculate the time-domain and frequency-domain characteristics based on the output values ​​of each weighing sensor, using these as the target domain. Specifically, step S2 involves: Step S21: Collect the values ​​of multiple weighing sensors corresponding to the dynamic truck scale at the overload control station, and use a Butterworth filter to filter out dynamic signal interference. Step S22: Calculate the output value of each weighing sensor using an association model; specifically, the association model is: ; in, Let j be the output value of the j-th weighing sensor on the dynamic truck scale. Let M be the value of the nth weighing sensor on the dynamic truck scale, and M be the number of weighing sensors corresponding to the current dynamic truck scale. The coefficient between the j-th load cell and the remaining M-1 load cells is given. , and These are the vehicle speed factor, load type factor, and load factor under dynamic conditions, respectively. The correlation weights for these three influencing factors; Step S23: Calculate the time-domain and frequency-domain characteristics based on the output values ​​of each weighing sensor, and use them as the target domain; Step S3: Fuse the time-domain features and frequency-domain features of the source domain and the target domain respectively to obtain the processed source domain sample and the unlabeled target domain sample. Step S4: Process the source domain samples and target domain samples using kernel functions, and then map them to the feature space using the transfer component analysis (TCA) method to obtain the features of the source domain samples and the features of the target domain samples. Step S5: Use the labeled features of the source and target domains after TCA feature transfer to train the preset WKNN classifier and output the target domain pseudo label; Step S6: Merge the target domain pseudo-label samples with the source domain to form an expanded labeled dataset. Retrain the WKNN classifier using the expanded dataset to obtain the trained WKNN classifier. Step S7: Collect relevant data from any dynamic truck scale at the overload control station in real time, preprocess it, and input it into the trained WKNN classifier to output the predicted fault type. The method also includes using the collected data from the dynamic truck scale to simultaneously judge abnormal behavior, and further screening the output of step S7. If there is no abnormal behavior, it is determined to be the corresponding fault type; otherwise, a vehicle abnormal behavior warning is issued directly. The method further includes step 8: after detecting a fault, a dynamic weighing system intelligent fault-tolerant method based on an improved particle swarm optimization generalized regression neural network under a multi-association model is used to measure and predict passing vehicles, as detailed below: Remove data that deviates due to faults; The Lévy-Flight Improved Particle Swarm Optimization (LPSO-GRNN) algorithm was used to search for the optimal solution to optimize the generalized regression neural network, resulting in an LPSO-GRNN network model that matches the truck scale data samples. This model was then used to establish the output prediction network. A multi-association model is used to determine the expected characteristic values ​​under different faults, and then the output prediction network is used to obtain the expected output value of the faulty sensor, thereby realizing the measurement prediction value of passing vehicles. The expected characteristic values ​​refer to the signal mean frequency, standard deviation frequency, impulse index and root mean square frequency that occur due to sensor faults.

2. The self-diagnosis method for abnormalities in a highway traffic overload detection device according to claim 1, characterized in that: The fault types include substantive faults and performance degradation faults. Substantive faults include load cell faults, foundation and weighing platform faults, and junction box and wiring faults.

3. The method for self-diagnosing abnormalities in a highway traffic overload detection device according to claim 1, characterized in that: The relevant data of the dynamic truck scales at the overload control stations include environmental monitoring data, weighing instrument data, data from multiple sensors used for weighing each dynamic truck scale, lidar data, cameras, and coil vehicle detectors.

4. The method for self-diagnosing abnormalities in a highway traffic overload detection device according to claim 1, characterized in that: The time-domain features include root mean square value, impulse index, kurtosis index, margin index, peak value, average amplitude, and waveform index; the frequency-domain features include standard deviation frequency, centroid frequency, root mean square frequency, and mean frequency. The calculation expressions for each feature are as follows: Root mean square value: ; Pulse Indicators: ; Kurtosis index: ; Margin indicators: ; Peak metrics: ; Average amplitude: ; Waveform Indicators: ; Standard deviation frequency: ; Center of gravity frequency: ; Root mean square frequency: ; Mean frequency: ; in, This represents the amplitude of the i-th point of the output signal from a sensor, where N represents the number of signal points. This represents the frequency value of the i-th spectral line in the power spectrum. This indicates a reference output value.

5. The method for self-diagnosing abnormalities in a highway traffic overload detection device according to claim 1, characterized in that: Before step S3, the method further includes: performing PCA or KPCA feature dimensionality reduction analysis on the source domain and target domain respectively to obtain the variance explanation rate of each principal component, and then fusing the time domain features and frequency domain features to obtain labeled weighing sensor samples and unlabeled truck scale samples respectively. In this process, a reference model of the feature signals under different fault types is established to determine the threshold range under different faults.

6. The method for self-diagnosing abnormalities in a highway traffic overload detection device according to claim 1, characterized in that: In step S4, "the source domain sample features and target domain sample features are obtained by mapping TCA feature transfer to the feature space" specifically means: TCA feature transfer is used to map source and target domain samples to the regenerating kernel Hilbert space. By minimizing the maximum mean difference between source and target domain samples in the feature subspace, a dimensionality-reduced feature subspace is obtained, which reduces the distribution difference between the source and target domains in this subspace. This results in source and target domain sample features with a certain degree of similarity, thereby realizing cross-domain transfer of fault knowledge.

7. The self-diagnosis method for abnormalities in a highway traffic overload detection device according to claim 1, characterized in that: The preprocessing in step S7 includes using a Butterworth filter to filter out dynamic signal interference from the acquired weighing sensors, obtaining the output values ​​of each weighing sensor of the dynamic truck scale through an association model, calculating time-domain and frequency-domain features based on the output values ​​of each sensor, and then fusing the time-domain and frequency-domain features through dimensionality reduction processing.

8. The method for self-diagnosing abnormalities in a highway traffic overload detection device according to claim 1, characterized in that: The abnormal vehicle behaviors include: crossing lanes, driving against traffic, driving along the edge, S-shaped behavior, parking, jumping on the weighbridge, dragging the weighbridge, grinding the weighbridge, queuing, and using a jack. These abnormal behaviors are determined by deeply fusing the sensing data from the roadside equipment and using a combination of multi-channel data from the weighing system and optical measurement methods to judge the vehicle's weight and behavior.

9. The method for self-diagnosing abnormalities in a highway traffic overload detection device according to claim 8, characterized in that: The specific process and handling methods for judging abnormal behavior include: The lane-crossing behavior is determined based on its characteristics: the vehicle crosses the lane centerline, part of its weight is obtained from the adjacent lane, and the weight measured by the current lane's weighing equipment is inaccurate. The determination method for lane-crossing behavior is as follows: the vehicle's position and status are estimated based on multi-channel sensor signals from all weighing equipment within the station; the driving position and direction are determined by coil signals; the vehicle's size and position are measured by lidar and compared with the weighing signal at the same moment to determine whether the vehicle is in a lane-crossing state. The lane-crossing behavior is processed as follows: the system displays an abnormal status message, and sensor signals belonging to the vehicle's total weight are matched based on the location and accumulated and uploaded. The reverse driving behavior is determined based on its characteristics, which are: the vehicle crosses the lane centerline, its entire weight is acquired by the weighing equipment in the adjacent oncoming lane, there is no weighing data in the current lane, and an error occurs in the weighing process of the oncoming lane. The method for determining reverse driving behavior is as follows: the vehicle's position and status are estimated based on the multi-channel sensor signals of all weighing equipment in the station; the driving position and direction are determined by the coil signal; the vehicle size, position, and driving direction measured by the lidar are compared with the weighing signal at the same time to determine whether the vehicle is in a reverse driving state; the reverse driving behavior is processed as follows: the abnormal status is indicated in the system, and the sensor signal of the oncoming lane that belongs to the total weight of the vehicle is matched according to the position and accumulated and uploaded; the rear camera is used as the front camera to identify the vehicle license plate number. The edge-pressing behavior is determined based on its characteristics, which are: the weighing equipment does not cover the entire width of the lane, and the vehicle partially presses on the road surface without a weighing platform. The determination method for edge-pressing behavior is as follows: based on the multi-channel sensor signals of the weighing equipment, it is determined whether the balance of the left and right weights exceeds a threshold. Combined with the vehicle size, position, and driving direction measured by the lidar, it is compared with the weighing signal at the same moment to determine whether the vehicle is in an edge-pressing state. The edge-pressing behavior is processed as follows: the abnormal state is indicated in the system, and the left half weight of the vehicle is matched according to the position as a reference half weight. Twice of this weight is uploaded as temporary data for subsequent verification. The S-shaped behavior is determined based on its characteristics, which are: the vehicle sways left and right as it passes through the weighing area, ensuring that the two wheels on the same axle do not pass through the weighing area in the same lane simultaneously. The determination method for S-shaped behavior is as follows: based on the multi-channel sensor signals of all weighing equipment at the station, the characteristics of peak axle load and wheel load are identified. This is combined with the vehicle's trajectory measured by lidar and compared with the weighing signals from the same period to determine if the vehicle is in an S-shaped driving state. The following processing is applied to S-shaped behavior: an abnormal status is indicated in the system, and all weights measured by the weighing equipment during the time the vehicle passes through the weighing area are uploaded as temporary data for subsequent verification. The parking behavior is determined based on its characteristics: prolonged parking in the weighing area causes errors in the inductive loop vehicle detector signal, leading to errors in the dynamic truck scale program that calculates weight by axle, and thus preventing the correct calculation of the total weight. The method for determining parking behavior is as follows: the vehicle status is estimated based on the multi-channel sensor signals of the weighing equipment; the inductive loop signal is used to determine if the vehicle is in motion; the vehicle trajectory measured by lidar is used for confirmation; and the result is compared with the weighing signal at the same time to determine if the vehicle is parked. The parking behavior is processed as follows: an abnormal status is displayed in the system, and all weights measured by the weighing equipment during the period when all axles of the vehicle pass through the weighing area are uploaded as data. The aforementioned "jumping the scale" behavior is determined based on its characteristics, which are: the vehicle accelerates from a standstill, causing the center of gravity to shift rearward, thereby reducing the measured front axle weight. The determination method for the "jumping the scale" behavior is: judging its acceleration behavior based on the vehicle speed-time curve obtained from LiDAR measurement. The "jumping the scale" behavior is handled as follows: an abnormal status is displayed in the system. The towing behavior is determined based on its characteristics, which are: the vehicle drags on the weighing equipment by locking the trailer axle, increasing the lateral force on the load cell and causing the load cell to jam, thus reducing the measured trailer weight; the towing behavior is determined by judging its acceleration and deceleration behavior based on the vehicle speed-time curve obtained by lidar measurement; the towing behavior is handled as follows: the abnormal status is displayed in the system. The scale-grinding behavior is judged based on its characteristics, which are: the vehicle repeatedly moves forward and backward in the weighing area, causing errors in the cumulative axle count and total weight. The judgment method for scale-grinding behavior is: based on the weight information of all weighing equipment in the station, the vehicle's forward and backward state is estimated, as well as the axle weight that is repeatedly weighed. Combined with the speed-time curve obtained by lidar measurement, it is compared with the weighing signal at the same moment to determine the weight of the vehicle that is repeatedly weighed. The scale-grinding behavior is processed as follows: the abnormal state is indicated in the system, and the weight of repeated weighing and the number of axles repeatedly calculated are removed. The queuing behavior is judged based on its characteristics, which are: multiple vehicles queuing closely to reduce the distance between vehicles, intending to interfere with the vehicle separator and cause errors in the measured weight, and preventing the camera from recognizing the license plate; the queuing behavior is judged by: estimating the vehicle separation status based on the weight information of all weighing equipment in the station, combined with the vehicle separation information obtained by LiDAR; the queuing behavior is processed as follows: the abnormal status is indicated in the system, and the weight of each vehicle in the queue is matched according to the spatiotemporal information; The jack behavior is determined based on its characteristics, which are: by installing a jack device at the bottom of the vehicle, the weight is transferred to a non-weighing area, reducing the measured axle load; the determination method for the jack behavior is: based on the weight of each axle obtained by the weighing equipment, it is estimated whether the axle load imbalance is caused by the jack, and at the same time, combined with the speed-time curve obtained by the lidar, it is determined whether the jack was operated while the vehicle was stopped to evade inspection; the jack behavior is handled as follows: the abnormal status is indicated in the system.

Images