A driving data analysis system and method based on target detection

Abstract

The application discloses a kind of driving data analysis system and method based on target detection, obtain vibration signal waterfall chart by modulating and demodulating to the original vibration signal, label file is made according to vibration signal waterfall chart, and waterfall chart and label file are randomly divided into training set, verification set and test set;Improved Yolov5s model is used to construct the neural network for target detection of waterfall chart, the neural network calls the data in training set, verification set and test set to train, obtains driving data analysis model;The model is used to carry out target detection of abnormal vibration signal to vibration signal waterfall chart, and the measurement area and duration period where abnormal vibration signal in driving data is located are obtained.The improved Yolov5s model is used to identify demodulated vibration signal in the application, and the improved Yolov5s model is more accurate and faster in recognition speed;When abnormal condition occurs in track, it can timely give early warning.

Description

Technical Field

[0001] This invention belongs to the field of image recognition technology, specifically relating to a vehicle data analysis system and method based on target detection. Background Technology

[0002] With the diversified development of train and railway technology in China, rail transit has taken on increasingly diverse types and occupies a significant proportion of transportation. However, during train operation, wheel-rail malfunctions may occur, and external intrusion incidents can disrupt the train's operating environment, severely impacting safety and causing economic and personal losses. Therefore, accurately and promptly identifying abnormal situations is crucial for ensuring rail transit safety. Traditional safety detection technologies cannot achieve distributed dynamic detection (real-time data collection from multiple points), cannot perform all-time, all-area detection of the track, cannot perform detailed detection of wheel-rail conditions (such as detecting wear on subway train wheels and rails), and cannot provide real-time categorized alarms for external construction, wheel-rail anomalies, and personnel intrusion.

[0003] In recent years, image recognition technology has made some progress in the application of track safety inspection. Existing track safety inspection methods collect and store images of abnormal situations using handheld fault recorders, establishing anomaly datasets. Traditional machine learning methods are then used to extract features from specific anomalies, followed by classification using algorithms such as SVM. Therefore, the recognition accuracy largely depends on manually designed features. However, the representation of anomalies in rail transit is complex and diverse; specific anomaly features cannot fully reflect track safety anomalies, resulting in low recognition accuracy and poor generalization. Furthermore, image recognition technology also suffers from problems such as large model parameters and low detection efficiency. Summary of the Invention

[0004] To improve the efficiency and accuracy of track safety inspection, this invention proposes a method and system for analyzing train operation data based on target detection.

[0005] A vehicle data analysis system based on target detection, which achieves one of the objectives of this invention, includes a vibration signal acquisition module, a dataset establishment module, and a vehicle data analysis model construction module.

[0006] The vibration signal acquisition module is used to acquire the original vibration signal of the track through a distributed optical fiber vibration sensor network arranged on the track, and to modulate and demodulate the original vibration signal to obtain a vibration signal waterfall plot.

[0007] The dataset creation module is used to obtain label files for identifying abnormal vibration signals based on the vibration signal waterfall plot, and randomly divides the vibration signal waterfall plot and the corresponding label files into training set, validation set and test set according to a set ratio;

[0008] The driving data analysis model construction module is used to construct a neural network for target detection of vibration signal waterfall plots using an improved Yolov5s model. This neural network is trained by calling the training set, validation set, and test set to obtain the driving data analysis model. This model is used to detect abnormal vibration signals in the vibration signal waterfall plots to obtain the measurement area and duration segment of the abnormal vibration signal in the driving data.

[0009] In the aforementioned system, the improved Yolov5s model adds an SPD-Conv module before each C3 module of the existing Yolov5s model. The SPD-Conv module is used to cut an intermediate feature map of arbitrary size into multiple sub-feature maps and perform non-staggered convolution to extract abnormal features from the waterfall plot, which can effectively improve the abnormal target detection effect of the train data analysis model. The non-staggered convolution modifies the number of channels in the sub-feature map. The abnormal features include: comparing the waterfall plot of each train vibration signal measurement area at each time period with the standard waterfall plot of the corresponding train vibration signal measurement area. If there is a disappearance of vibration signal in a certain measurement area, the duration of vibration signal is longer or shorter than the duration in the corresponding standard waterfall plot, or the vibration intensity of vibration signal is different from the vibration intensity at the corresponding position in the standard waterfall plot, then it is considered that there are abnormal features in the waterfall plot. The standard waterfall plot of the train vibration signal measurement area is obtained in the early stage under the condition that there are no track anomalies.

[0010] The SPD-Conv module consists of an SPD module and a Conv module. The SPD module is used to cut an intermediate feature map of arbitrary size into multiple sub-feature maps to extract richer features. The Conv module is used to perform non-stepping convolution on the sub-feature maps segmented by the SPD module.

[0011] The Conv module uses non-staggered convolutional layers with a C2 filter, where the C2 filter changes the number of channels in the sub-feature map from C1 to C2 during convolution, and C2 = scale. 2 ×C1; where scale is the cutting step size of the SPD module, and C1 is the number of channels in the intermediate feature map or sub-feature map, with the intermediate feature map and sub-feature map having the same number of channels.

[0012] The SPD module sets the cutting step size (scale) to 2 when cutting the intermediate feature map. The cutting step size (scale) means that the sub-feature map is evenly divided into scale*scale sub-feature maps. If scale is set to 2, the sub-feature map is cut into 4 (2*2) sub-maps, and the length and width of each sub-map are 1 / 2 of the atomic feature map. If scale is set to 3, the sub-feature map is cut into 9 (3*3) sub-maps, and the length and width of each sub-map are 1 / 3 of the atomic feature map.

[0013] Furthermore, if the vibration signal waterfall plot contains abnormal vibration signals, then the vibration signal waterfall plot after target detection includes the measurement area, time period, anomaly type, and confidence level of the abnormal vibration signal. The abnormal vibration signal is determined by comparing the waterfall plot of each time period in each train vibration signal measurement area with the standard waterfall plot of the corresponding train vibration signal measurement area (obtained from previous experiments under conditions of no track anomalies). If, in a certain measurement area, the vibration signal disappears, or the duration of the vibration signal is too long or too short, or the vibration intensity of the vibration signal differs from the vibration intensity at the corresponding position in the standard waterfall plot, then an abnormal vibration signal is considered to exist.

[0014] The system also includes a tag file creation module, which is used to obtain tag files for identifying abnormal signals based on the vibration signal waterfall plot; the tag files and their corresponding vibration waterfall plots together serve as the original dataset for training the driving data analysis model.

[0015] The prefix of the filename of the label file is the same as the prefix of the filename of the corresponding waterfall plot. Each line of data in the label file corresponds to a selection box in the image. The selection box is used to select the abnormal part of the image. The first number in each line is the type of abnormality, and the four values ​​after each type of abnormality correspond to the coordinate parameters of the selection box in the image.

[0016] Furthermore, methods for obtaining vibration signal waterfall plots include:

[0017] S101. A tunable narrow-linewidth laser emits continuous light with a set optical power as the light source for a distributed fiber optic vibration sensing network; an acousto-optic modulator pulses the continuous light to obtain pulsed light; and the pulsed light is amplified and then enters the distributed fiber optic vibration sensing network.

[0018] S102. The sensing channel in the distributed optical fiber vibration sensing network reflects back a pulse light sequence. The reflected pulse light sequence enters the unequal-arm Mach-Zehnder interferometer. The interferometer converts the optical phase information caused by the vibration at the corresponding sensing position into optical intensity information. The signal is then output to a multi-channel photodetector through a multi-channel coupler to obtain multiple electrical signals. The vibration sensing information is obtained by using a phase demodulation algorithm on the multiple electrical signals.

[0019] S103. The vibration sensing information is integrated into a waterfall plot. The horizontal axis of the waterfall plot represents different test areas, the vertical axis represents time, and the color and the color-intensity metric table on the side together represent the intensity of the vibration sensing information in the test area during the time period.

[0020] A driving data analysis method based on target detection, which achieves the second objective of this invention, includes the following steps:

[0021] S1. Obtain the original vibration signal of the track through a distributed optical fiber vibration sensor network arranged at the track, and modulate and demodulate the original vibration signal to obtain a vibration signal waterfall diagram.

[0022] S2. Create a corresponding tag file for each vibration signal waterfall plot, and randomly divide the waterfall plot and the corresponding tag file into a training set, a validation set and a test set according to a set ratio; the tag file is used to record the types of abnormal parts existing in the corresponding waterfall plot and their position parameters in the waterfall plot;

[0023] S3. Construct a neural network for target detection of vibration signal waterfall plots using an improved Yolov5s model. This neural network is trained by calling the waterfall plots and corresponding label files in the training set, validation set, and test set to obtain a driving data analysis model. This model is used to detect abnormal vibration signals in the vibration signal waterfall plot to obtain the measurement area and duration segment of the abnormal vibration signal in the driving data.

[0024] Beneficial effects:

[0025] This invention employs an improved YOLOv5s model to identify demodulated vibration signals, enabling faster identification of anomalous features in the corresponding vibration signal waterfall plot. This allows for quicker analysis of the measurement area and duration of abnormal vibration signals on the track. The improved YOLOv5s model integrates the SPD-Conv module, which further extracts anomalous features from the input waterfall plot by segmenting the feature maps in the network structure. This makes the learned anomalous features more detailed, resulting in higher accuracy and faster identification speed for the improved YOLOv5s model. Using the system and method described in this invention, early warnings can be issued more quickly when abnormal conditions occur on the track. Attached Figure Description

[0026] Figure 1 This is a general flowchart of an embodiment of the method described in this invention;

[0027] Figure 2 This is a schematic diagram of a distributed fiber optic vibration sensing network based on Mach-Zehnder interferometric OTDR technology in the embodiment.

[0028] 1. Narrow bandwidth laser; 2. Modulation device; 3. First circulator; 6. Second circulator; 4. Long-distance transmission fiber; 5. Multiple chirped Bragg fiber grating array; 7. 3×3 coupler; 8. Delay fiber; 9. First Faraday rotator mirror; 11. Second Faraday rotator mirror; 10. Unbalanced Michelson interferometer; 12. First photodetector; 13. Second photodetector; 14. Third photodetector; 15. Host computer.

[0029] Figure 3 This is a structural diagram of the improved Yolov5s model in the embodiment;

[0030] Figure 4 This is a schematic diagram of the test area and situation dataset in the embodiment;

[0031] Figure 5 yes Figure 4 A schematic diagram of the contents of the tag file corresponding to the waterfall chart shown;

[0032] Figure 6 This is a diagram showing the recognition effect of the track safety detection model in the embodiment. Detailed Implementation

[0033] The following detailed embodiments are provided to explain the technical solutions of the claims of this invention, so that those skilled in the art can understand the claims. The scope of protection of this invention is not limited to the following specific embodiments. Any modifications made by those skilled in the art that incorporate the technical solutions of the claims but differ from the following detailed embodiments are also within the scope of protection of this invention.

[0034] like Figure 1 As shown in the figure, this application provides a vehicle data analysis method based on target detection, including the following steps:

[0035] Step 1: Obtain the original vibration signal of the track through a distributed optical fiber vibration sensor network deployed at the track, and modulate and demodulate the original vibration signal to obtain a vibration signal waterfall plot; compared with other existing images, the vibration signal waterfall plot shows abnormal signals more obviously, is easier to process, and does not require too much consideration of the difference between the measurement areas.

[0036] Step 2: Create a corresponding tag file for each vibration signal waterfall plot. Randomly divide the waterfall plot and its corresponding tag file into a training set, a validation set, and a test set in a ratio of 8:1:1, and place them into the training set folder, the validation set folder, and the test set folder in sequence. The tag file is used to record the anomaly type of the abnormal vibration signal present in the corresponding waterfall plot and its position parameters in the waterfall plot.

[0037] Step 3: Run the improved Yolov5s model by calling the waterfall plots and corresponding label files in the training set folder, validation set folder, and test set folder to obtain the driving data analysis model; the input variables of this model are as follows: Figure 4 The waterfall plot in the image outputs a waterfall plot after object detection, such as... Figure 6 The waterfall plot is labeled with the anomaly type of the abnormal vibration signal and its location parameters in the waterfall plot. Using the aforementioned driving data analysis model, a large number of waterfall plots can be processed efficiently, and abnormal parts on the waterfall plot can be detected in real time. Based on the abnormal parts, the measurement area where the abnormal signal is located and the time period in which it occurred can be found.

[0038] Step 4: Run the driving data analysis model to perform abnormal target detection on the waterfall plot, obtain test results, and observe the correctness.

[0039] In the above technical solution, the schematic diagram of the distributed optical fiber vibration sensing network described in step 1 is as follows: Figure 2 As shown, an interferometric grating array vibration sensing system is adopted, which mainly consists of a UWFBG array and an unbalanced Michelson interferometer structure.

[0040] As a sensing network, the UWFBG array consists of two adjacent UWFBGs and the optical fiber within them forming a sensing area, and the reflectivity, bandwidth, center wavelength and other parameters of each UWFBG are basically the same.

[0041] The unbalanced Michelson interferometer structure mainly consists of a 3×3 coupler, two delay fibers of unequal length, and two Faraday rotators. The optical path difference between the two delay fibers is the same as the spacing between adjacent UWFBGs, used to compensate for the optical path difference between adjacent UWFBGs; both Faraday rotators are rotated at a 45° angle to eliminate polarization effects in the interferometer.

[0042] By using an equal-arm Mach-Zehnder interferometer to improve the frequency response range of vibration signal sensing, and by using an ultra-weakly reflective fiber grating with a reflectivity of only 0.01% to form a sensing array, spatial location information is converted into time-domain delay, thereby realizing multi-point vibration localization in distributed vibration sensing.

[0043] The structure of the aforementioned interferometric grating array vibration sensing system includes:

[0044] Narrow-bandwidth laser 1 is connected to first circulator 3 via modulation device 2. One port of first circulator 3 is connected to second circulator 6 via long-distance transmission fiber 4 and multiple chirped Bragg fiber grating arrays 5.

[0045] The first port of the 3×3 coupler 7 is connected to the first Faraday rotator mirror 9 via the delay fiber 8, the second port is connected to the unbalanced Michelson interferometer 10, the third port is connected to the second Faraday rotator mirror 11, the fourth port is connected to the first photodetector 12 via the second circulator 6, the fifth port is connected to the second photodetector 13, and the sixth port is connected to the third photodetector 14.

[0046] The first photodetector 12, the second photodetector 13 and the third photodetector 14 are all connected to the host computer 15;

[0047] The working principle of the above-mentioned interferometric grating array vibration sensing system is as follows:

[0048] The continuous narrow-bandwidth laser output from the narrow-bandwidth laser 1 is modulated by the modulation device 2 and then passes through the first circulator 3. It is then transmitted through the long-distance transmission fiber 4 to the chirped Bragg fiber grating array 5. The chirped Bragg fiber grating array 5 at different distances generates reflected light pulse signals. When the environment around the chirped Bragg fiber grating array 5 changes, the phase of the reflected light pulse shifts for the same chirped Bragg fiber grating array 5. The resulting phase-shifted reflected light pulses enter the 3×3 coupler 7 through the first circulator 3 and the second circulator 6. Then, they enter the first Faraday rotator mirror 9 and the second Faraday rotator mirror 11 with the delay fiber 8, respectively. After passing through the unbalanced Michelson interferometer 10, interference light pulse signals are generated and return to the 3×3 coupler 7. After passing through the first photodetector 12, the second photodetector 13, and the third photodetector 14, the three interference light pulse signals are converted into electrical signals and input into the host computer 15 to form a waterfall diagram.

[0049] In this embodiment, the method for obtaining the vibration signal waterfall diagram in step 1 includes:

[0050] S101, a tunable narrow-linewidth laser emits continuous light with a power of 3mW as the light source for a distributed fiber optic vibration sensing network. Its wavelength is tunable from 1500nm to 1630nm, making it suitable for fiber optic grating sensing networks with various center wavelengths. The acousto-optic modulator in the modulation device pulses the continuous light, achieving a high extinction ratio of 63dB. The modulated light pulse width is 20ns, and the frequency is 100kHz. After the pulsed light is amplified by an EDFA, it enters the sensing network, which consists of a weakly reflective fiber optic grating with a reflectivity of 0.01% and a sensing fiber with a length of 3m.

[0051] S102. The pulse light sequence reflected back from the sensing channel enters the unequal-arm Mach-Zehnder interferometer with an arm length difference of 6m. The interferometric technique is used to convert the light phase information caused by the vibration at the corresponding sensing position into light intensity information. The light is then output to three identical photodetectors through a 3×3 coupler for photoelectric conversion. Finally, the electrical signal is acquired and the vibration sensing information is obtained through a phase demodulation algorithm.

[0052] The formulas for calculating the light intensity at the three output terminals of the 3×3 symmetrical coupler under ideal splitting ratio conditions are as follows:

[0053]

[0054] In the formula:

[0055] I n n = 1, 2, 3: representing the first, second, and third outputs of the 3×3 symmetrical coupler, respectively;

[0056] Phase changes caused by external factors.

[0057] S103. The vibration sensing information is integrated into a waterfall chart by the host computer. The horizontal axis of the waterfall chart represents different test areas, the vertical axis represents time, and the color and the color-intensity measurement table on the side together represent the intensity of the vibration sensing information in the test area during the time period.

[0058] In the above technical solution, the method for creating the label file in step 2 includes:

[0059] Use a tool that can create label file datasets to create label files. When using this tool, manually select the abnormal images. The tool will automatically output the label file corresponding to each waterfall chart based on the manually selected abnormal images; for example... Figure 4 The image shown is a waterfall plot of the original vibration signal. Figure 5 As shown Figure 4 The corresponding tag file is saved in .txt format in this embodiment. The prefix of the tag file name is the same as the prefix of the corresponding waterfall chart file name. Each line of data in the tag file corresponds to a selection box in the image. The selection box is a box to select the abnormal part of the image. The first number of each line is the type of abnormality, such as 2, 3, 4 and 5 in the figure, which are four types of abnormalities. The four floating-point numbers after each type of abnormality are the coordinate parameters of the selection box in the figure (the image has been normalized first, and the calculation and output of this parameter is automatically completed by the tool). The first floating-point parameter is the x-coordinate of the center of the box, the second is the y-coordinate of the center of the box, the third is the width of the box, and the fourth is the width and height.

[0060] In the above technical solution, the random partitioning in step 2 involves treating the waterfall plots and the label files with the same filename prefix as a group, randomly selecting from many groups, and assigning them to the training set, the validation set, and the test set, ensuring that the waterfall plots and the label files with the same filename prefix are in the same set.

[0061] In the above technical solution, the training set folder, validation set folder, and test set folder in step 2 are used to train the driving data analysis model. During the iterative process of model training, the contents of the training set folder are used to provide a feature library that the model can learn. After the model has learned all the images in the training set, the contents of the validation set folder and test set folder are automatically called to test abnormal target detection. The test results are automatically recorded, and the iteration parameters are automatically modified according to the test results to continue iterative training until the predetermined number of training rounds is reached, at which point the training ends and the trained driving data analysis model is obtained.

[0062] In the above technical solution, the total number of datasets used for training in step 2 is 100 sets, and the predetermined number of training rounds is 50 rounds, which can obtain a high-precision driving data analysis model, which is saved in .pt format.

[0063] In the above technical solution, the improved Yolov5s mentioned in step 3 is an algorithm that incorporates the SPD-Conv module. The SPD-Conv module includes an SPD module and a Conv module. The SPD-Conv module can better learn the features of the detected target, and the output results can be directly displayed in the image or output as text, which is highly flexible and provides intuitive results.

[0064] Specifically, in the improved YOLOv5s model, an SPD-Conv module is added before each C3 module in the backbone of the existing YOLOv5s neural network model, for a total of four SPD-Conv modules. All SPD-Conv modules have the same settings. The C3 module is a module that learns residual features and contains three standard convolutional layers and multiple Bottleneck modules (the number is determined by the product of the n and depth_multiple parameters in the configuration file .yaml). The C3 module can learn the features obtained from the previous convolutions. Inserting the SPD-Conv module before the C3 module can extract more detailed features from the image to provide for the C3 module to learn features. Figure 3 This is a structural diagram of the improved Yolov5s model in an embodiment of the present invention, wherein the SPD-Conv module is inserted into the positions of layers 2 and 3, 6 and 7, 10 and 11, and 14 and 15 to form the improved Yolov5s model.

[0065] The specific method for constructing an improved Yolov5s neural network model is as follows:

[0066] An existing Yolov5s neural network is constructed. The input waterfall plot is automatically augmented and fed into the backbone of the neural network. The target features of the input waterfall plot are extracted through four Conv layers and four C3 modules in the backbone. The features are then fed into the Neck part of the network to generate a feature pyramid from the features extracted by the backbone and perform multi-scale feature fusion. Finally, these features are passed to the head layer of the prediction layer for final regression prediction.

[0067] Preferably, the SPD module is used to cut an intermediate feature map X of arbitrary size S×S×C1 (length, width, number of channels) into a series of sub-feature maps. The cutting formula is as follows:

[0068] f 0,0 =X[0:S:scale,0:S:scale]

[0069] f 1,0 =X[1:S:scale,0:S:scale],…,

[0070] f scale-1,0 =X[scale-1:S:scale,0:S:scale];

[0071] f 0,1 =X[0:S:scale,1:S:scale],

[0072] f 1,1 =X[1:S:scale,1:S:scale],…,

[0073] f scale-1,1 =X[scale-1:S:scale,1:S:scale];

[0074] f 0,scale-1 =X[0:S:scale,scale-1:S:scale],

[0075] f 1,scale-1 =X[1:S:scale,scale-1:S:scale],…,

[0076] f scale-1,scale-1 =X[scale-1:S:scale,scale-1:S:scale].

[0077] In the formula:

[0078] f x,yIn the model, x = 0, 1, 2, ..., scale-1, y = 0, 1, 2, ..., scale-1, which represent all the sub-feature maps, X is the feature map, S is the image length, and scale is the cutting step size;

[0079] f 1,0 =X[1:S:scale,0:S:scale] means that during the cutting process, the horizontal axis starts from the second pixel and ends at the Sth pixel, with a step size of scale; the vertical axis starts from the first pixel and ends at the Sth pixel, with a step size of scale. Then, all the cut pixels are concatenated in sequence to obtain a sub-feature map f. 1,0 .

[0080] Preferably, the scale value in the SPD module is set to 2, which can significantly improve the confidence of target detection while ensuring high speed. The resulting four sub-images f 0,0 f 1,0 f 0,1 f 1,1 Each sub-image has a size of (S / 2×S / 2×C1), which is 1 / 4 the size of the original feature map X, i.e., X is downsampled by 2 times.

[0081] Preferably, the Conv module uses a non-staggered convolutional layer with a C2 filter, where the C2 filter means that the number of channels is changed from C1 to C2 during convolution, where C2 = scale. 2 ×C1, after non-stepping convolution processing by the Conv module in the SPD-Conv module, the size of each subgraph is (S / 2×S / 2×C2).

[0082] Preferably, the non-stepping refers to setting the step size to 1. Setting the step size of the Conv module to 1 can avoid the non-discriminatory loss of information when the step size is greater than 1, that is, avoid the asymmetric sampling process.

[0083] In the above technical solution, the test process described in step 4 is to use the driving data analysis model to test the waterfall plot. The test result is to mark the predicted boxes of abnormal vibration signals on the test waterfall plot and mark the prediction labels and prediction confidence levels. The prediction labels are used to indicate the type of abnormality of the abnormal signal, and the prediction confidence level is used to indicate the credibility of the type of abnormality to which the abnormal signal belongs.

[0084] In the above technical solution, the observation accuracy step 4 is based on comparing the horizontal and vertical coordinates and predicted labels of the partial images within the predicted box with the actual anomalies to distinguish different anomalies. Figure 6This is a waterfall chart showing the recognition results of a track safety detection model. The image contains four boxes, each enclosing a portion of the image at a different location. The first number above each box represents the anomaly type predicted by the vehicle data analysis model in this waterfall chart. The chart shows four types: 003, 004, 005, and 006, indicating that the model identified four different types of targets, or anomalies. The second number above each box represents the confidence level of the model's prediction for these four anomaly types in this waterfall chart.

[0085] It should be understood that the sequence number of each step in the above embodiments does not imply the order of execution. The execution order of each process should be determined by its function and internal logic, and should not constitute any limitation on the implementation process of the embodiments of this application.

[0086] The contents not described in detail in this specification are existing technologies known to those skilled in the art.

Claims

1. A vehicle data analysis system based on target detection, characterized in that, It includes a vibration signal acquisition module, a dataset creation module, and a vehicle data analysis model construction module; The vibration signal acquisition module is used to acquire the original vibration signal of the track through a distributed optical fiber vibration sensor network arranged on the track, and to modulate and demodulate the original vibration signal to obtain a vibration signal waterfall plot. The dataset creation module is used to obtain label files for identifying abnormal vibration signals based on the vibration signal waterfall plot, and randomly divides the vibration signal waterfall plot and the corresponding label files into training set, validation set and test set according to a set ratio; The driving data analysis model construction module is used to construct a neural network for target detection of vibration signal waterfall plots using an improved Yolov5s model. This neural network is trained by calling waterfall plots and corresponding label files from the training set, validation set, and test set to obtain the driving data analysis model. This model is used to detect abnormal vibration signals from the vibration signal waterfall plots to obtain the measurement area and duration segment of the abnormal vibration signal in the driving data. The improved Yolov5s model adds an SPD-Conv module before each C3 module of the existing Yolov5s model. The SPD-Conv module is used to cut an intermediate feature map of arbitrary size into multiple sub-feature maps and perform non-stride convolution on the sub-feature maps in order to extract anomalous features in the waterfall plot. The SPD-Conv module includes the SPD module, which is used to cut an intermediate feature map of arbitrary size into multiple sub-feature maps. The SPD-Conv module includes a Conv module; the Conv module is used to perform non-stepping convolution on multiple sub-feature maps segmented by the SPD module.

2. The vehicle data analysis system based on target detection as described in claim 1, characterized in that, The SPD module sets the cutting step size (scale) to 2 when cutting the intermediate feature map.

3. The vehicle data analysis system based on target detection as described in claim 1, characterized in that, If the vibration signal waterfall plot contains abnormal vibration signals, then the vibration signal waterfall plot after target detection includes the measurement area, time period, anomaly type, and confidence level of the abnormal vibration signals.

4. The vehicle data analysis system based on target detection as described in claim 1, characterized in that, It also includes a tag file creation module, which is used to obtain tag files for identifying abnormal signals based on the vibration signal waterfall plot; the tag files and their corresponding vibration waterfall plots together serve as the original dataset for training the driving data analysis model.

5. The vehicle data analysis system based on target detection as described in claim 4, characterized in that, The prefix of the filename of the label file is the same as the prefix of the filename of the corresponding waterfall plot. Each line of data in the label file corresponds to a selection box in the image. The selection box is used to select the position of the abnormal signal in the waterfall plot. The first number of each line of data in the label file indicates the type of abnormality. Multiple numbers after each number indicating the type of abnormality are used to indicate the coordinate parameters of the selection box in the image.

6. The vehicle data analysis system based on target detection as described in claim 1, characterized in that, Methods for obtaining vibration signal waterfall plots include: S101. A tunable narrow-linewidth laser emits continuous light with a set optical power as the light source for a distributed fiber optic vibration sensing network; an acousto-optic modulator pulses the continuous light to obtain pulsed light; and the pulsed light is amplified and then enters the distributed fiber optic vibration sensing network. S102. The sensing channel in the distributed optical fiber vibration sensing network reflects back a pulse light sequence. The reflected pulse light sequence enters the unequal-arm Mach-Zehnder interferometer. The interferometer converts the optical phase information caused by the vibration at the corresponding sensing position into optical intensity information. The signal is then output to a multi-channel photodetector through a multi-channel coupler to obtain multiple electrical signals. The vibration sensing information is obtained by using a phase demodulation algorithm on the multiple electrical signals. S103. The vibration sensing information is integrated into a waterfall plot. The horizontal axis of the waterfall plot represents different test areas, the vertical axis represents time, and the color and the color-intensity metric table on the side together represent the intensity of the vibration sensing information in the test area during the time period.

7. A vehicle data analysis method based on target detection according to the system of claim 1, characterized in that, Includes the following steps: S1. Obtain the original vibration signal of the track through a distributed optical fiber vibration sensor network arranged at the track, and modulate and demodulate the original vibration signal to obtain a vibration signal waterfall diagram. S2. Create a corresponding tag file for each vibration signal waterfall plot, and randomly divide the waterfall plot and the corresponding tag file into a training set, a validation set and a test set according to a set ratio. The tag file is used to record the types of abnormal parts existing in the corresponding waterfall plot and their position parameters in the waterfall plot. S3. Construct a neural network for target detection of vibration signal waterfall plots using an improved Yolov5s model. This neural network is trained by calling the waterfall plots and corresponding label files in the training set, validation set, and test set to obtain a driving data analysis model. This model is used to detect abnormal vibration signals in the vibration signal waterfall plot to obtain the measurement area and duration segment of the abnormal vibration signal in the driving data.

Images