Railway Safety Detection System and Method Based on Improved AlexNet

Abstract

The present invention discloses an orbit safety detection system based on an improved AlexNet. Its abnormal vibration signal judgment module determines whether the train operation vibration signals in each selected historical period of each train operation vibration signal measurement area are normal train operation vibration signals or abnormal train operation vibration signals; the vibration signal data set establishment module forms a model data set; the data enhancement module performs data enhancement on the model data set, and divides the vibration signal data in the model data set after sample expansion into a training set, a validation set and a test set; the convolutional neural network model construction module constructs a convolutional neural network; the model training module uses the training set to train the AlexNet model; the signal testing module determines the orbit safety status of the train operation vibration signal measurement area to be measured. The present invention can detect the orbit in real time and across the whole area, and can give early warnings for problems such as external construction, orbit anomalies, and personnel intrusion.

Description

Technical Field

[0001] The present invention relates to the technical field of track safety detection, and specifically refers to a track safety detection system and method based on an improved AlexNet. Background Art

[0002] With the diversified development of trains and railway technologies in China, there are more and more types of rail transit, which occupies a large proportion in transportation. However, during the operation of trains, faults may occur in the wheel-rail system, and external personnel intrusion events will also damage the train operation environment, thus seriously affecting the safety of train operation and causing economic and personnel losses. Therefore, accurately and timely identifying the types of abnormal situations is particularly important for ensuring track traffic safety.

[0003] Traditional safety detection technologies cannot achieve distributed dynamic detection, cannot detect the state of the wheel-rail system in detail, and cannot alarm against intrusion behaviors.

[0004] In recent years, the application of image recognition-based technologies in track safety detection has made some progress. Existing track safety detection methods collect and store abnormal situation images through hand-held disease recorders, establish an abnormal situation dataset, and after extracting features of specific abnormal situation features through traditional machine learning methods, use machine learning algorithms such as SVM to classify the extracted features. Therefore, their recognition effect depends to a large extent on manually designed features. The characteristics of abnormal situations in rail transit are complex and diverse, and specific abnormal situation features cannot fully reflect the abnormal situations of track safety, resulting in low recognition accuracy and poor generalization effect. In addition, image recognition technologies also have problems such as large model parameters and low detection efficiency. Summary of the Invention

[0005] The purpose of the present invention is to provide a track safety detection system and method based on an improved AlexNet. The present invention can detect the track in real time and across the entire area, and can give early warnings about problems such as external construction, track anomalies, and personnel intrusion.

[0006] To achieve this purpose, a track safety detection system based on an improved AlexNet designed by the present invention is characterized in that it includes an abnormal vibration signal judgment module, a vibration signal dataset establishment module, a data enhancement module, a convolutional neural network model construction module, a model training module, and a signal testing module;

[0007] Each driving vibration signal measurement area in the track distributed acoustic wave sensing system can sense the driving vibration signal of the corresponding measurement area;

[0008] The abnormal vibration signal judgment module is used to compare the driving vibration signal of each selected historical period in each driving vibration signal measurement area with the standard driving vibration signal of the corresponding driving vibration signal measurement area in terms of amplitude and frequency, and determine whether the driving vibration signal of each selected historical period in each driving vibration signal measurement area is a normal driving vibration signal or an abnormal driving vibration signal;

[0009] The vibration signal data set establishment module is used to set a corresponding folder for each driving vibration signal measurement area, and put the normal driving vibration signal or abnormal driving vibration signal of each historical time period selected in the corresponding driving vibration signal measurement area into each folder to form a model data set;

[0010] The data enhancement module is used to perform data enhancement operations on the model data set to obtain a model data set after sample expansion, and divide the vibration signal data in the model data set after sample expansion into a training set, a verification set and a test set;

[0011] The convolutional neural network model building module is used to build a convolutional neural network using the AlexNet method;

[0012] The model training module is used to set the hyperparameters of the AlexNet model training in the convolutional neural network according to the accuracy required when training the model, and use the above hyperparameters to train the AlexNet model using the training set;

[0013] The signal testing module is used to test the test set using the trained AlexNet model, generate a track vehicle passing condition diagram based on the test results, and determine the track safety condition of the vehicle vibration signal measurement area to be tested through the track vehicle passing condition diagram.

[0014] Beneficial effects of the present invention:

[0015] 1. The optical fiber laid in the track covers the entire track and has a wide range. It can continuously collect vibration information of the entire track and realize full-time and full-domain detection of track status, which solves the problem that traditional track detection systems cannot achieve full-time and full-domain detection, improves detection efficiency, and gives early warning of abnormal track conditions;

[0016] 2. The present invention adopts an improved AlexNet model to identify the demodulated vibration signal. Compared with the original model, the improved AlexNet model does not overfit, has higher accuracy and faster recognition speed. When an abnormal condition occurs on the track, a timely warning can be given. BRIEF DESCRIPTION OF THE DRAWINGS

[0017] Figure 1 It is a structural schematic diagram of the present invention;

[0018] Figure 2Schematic diagram of the distributed fiber optic vibration sensing network based on Mach-Zehnder interference OTDR technology in the present invention;

[0019] Figure 3 Structural diagram of the AlexNet model in the embodiment of the invention;

[0020] Figure 4 Recognition effect diagram of the track safety detection model in the embodiment of the present invention;

[0021] Figure 5 Schematic diagram of the measurement area and situation data set in the embodiment of the present invention;

[0022] Figure 6 Time-domain waveform diagram of abnormal passing vehicle vibration signal;

[0023] Figure 7 Time-domain waveform diagram of normal passing vehicle vibration signal;

[0024] Figure 3 In it, conv means convolution, maxpooling means max pooling, the number after @ represents the number of convolution kernels, s represents the stride, the Flatten layer is used to "flatten" the input, that is, to make the multi-dimensional input one-dimensional, which is commonly used in the transition from the convolutional layer to the fully connected layer, FC represents the fully connected layer, and softmax represents that the softmax activation function is used in the last fully connected layer. Specific implementation manners

[0025] The following further elaborates the present invention in detail in conjunction with the accompanying drawings and specific embodiments:

[0026] Such as Figure 1 The track safety detection system based on the improved AlexNet shown, which is characterized in that it includes an abnormal vibration signal judgment module, a vibration signal data set establishment module, a data enhancement module, a convolutional neural network model construction module, a model training module and a signal testing module;

[0027] Each train vibration signal measurement area in the track distributed acoustic sensing system can sense the train vibration signal corresponding to the measurement area;

[0028] The abnormal vibration signal judgment module is used to compare the train vibration signals in each selected historical period in each train vibration signal measurement area with the standard train vibration signal of the corresponding train vibration signal measurement area (obtained through experiments in the early stage when there is no abnormality in the track) in terms of amplitude and frequency, and determine whether the train vibration signals in each selected historical period in each train vibration signal measurement area are normal train vibration signals or abnormal train vibration signals. The abnormal signal judgment module is used to assist in making the data set;

[0029] The vibration signal dataset establishment module is used to set a corresponding folder for each train running vibration signal measurement area, and put the normal train running vibration signals or abnormal train running vibration signals of the selected historical time periods in each folder to form a model dataset;

[0030] The data augmentation module is used to perform data augmentation operations on the model dataset to obtain a model dataset with augmented samples, and divide the vibration signal data in the model dataset with augmented samples into a training set, a validation set, and a test set according to a ratio of 3:1:1;

[0031] The convolutional neural network model construction module is used to construct a convolutional neural network using the AlexNet method;

[0032] The model training module is used to set the hyperparameters for training the AlexNet model in the convolutional neural network according to the accuracy required for training the model (the parameters can be modified in the program). Using the above hyperparameters, the training set is used to train the AlexNet model. The AlexNet model is used to process and analyze a large amount of data, reducing the workload and improving efficiency and accuracy;

[0033] The signal testing module is used to test the test set using the trained AlexNet model, generate a track passing condition diagram based on the test results, and determine the track safety condition of the train running vibration signal measurement area to be measured. The specific testing process is as follows: put the image to be tested in a folder, run the above-trained model to read the image in the file, and then perform recognition and output. The monitoring situation of each picture is output, and the output is concentrated in a TXT document. The specific output is as attached Figure 4 as shown.

[0034] In the above technical solution, two adjacent ultra-weak fiber Bragg gratings and the optical fiber between them in the rail transit distributed acoustic sensing system form a train running vibration signal measurement area, and each train running vibration signal measurement area can sense the corresponding train running vibration signal in the measurement area;

[0035] The phase change caused by changes in the diameter, refractive index, length, etc. of the optical fiber due to vibration sensed is resolved into an optical signal through a 3*3 coupler. The optical signal is transmitted to the upper computer by the PD. The PD is a photodetector. Then the optical signal is converted into an electrical signal through the PD, and further the vibration signal is restored.

[0036] In the above technical solution, the abnormal vibration signal judgment module is used to compare the train vibration signals in each time period of each train vibration signal measurement area with the standard train vibration signal of the corresponding train vibration signal measurement area (obtained through experiments when there is no abnormality in the track in the early stage). When the difference between the amplitude of the train vibration signal at a certain moment in a certain train vibration signal measurement area and the amplitude of the standard train vibration signal of the corresponding train vibration signal measurement area > the preset amplitude threshold, or when the difference between the frequency of the train vibration signal at a certain moment in this train vibration signal measurement area and the frequency of the standard train vibration signal of the corresponding train vibration signal measurement area > the preset frequency threshold, it is determined that the train vibration signal at this moment in this train vibration signal measurement area is an abnormal vibration signal;

[0037] When the difference between the amplitude of the train vibration signal at a certain moment in a certain train vibration signal measurement area and the amplitude of the standard train vibration signal of the corresponding train vibration signal measurement area ≤ the preset amplitude threshold, and at the same time, the difference between the frequency of the train vibration signal at this moment in this train vibration signal measurement area and the frequency of the standard train vibration signal of the corresponding train vibration signal measurement area ≤ the preset frequency threshold, it is determined that the train vibration signal at this moment in this train vibration signal measurement area is a normal vibration signal.

[0038] In the above technical solution, the data augmentation operation is to call the OpenCV module under the TensorFlow unit to perform sample augmentation on the model data set. The sample augmentation methods include: random angle rotation, horizontal flipping, and vertical flipping. The OpenCV module is a software development toolkit that integrates many image processing modules. Here, the OpenCV module is used to perform sample augmentation processing on the images in the data set, and data augmentation is performed on the collected data set. After augmentation, a more abundant data set is formed. The pictures in this data set are more abundant, and the features of the monitoring target are more prominent. The purpose of augmentation is to enrich the data set so that the algorithm can more easily extract image features. The data augmentation method adopted in this embodiment is to enlarge the pictures.

[0039] In the above technical solution, the specific method for the convolutional neural network model construction module to construct a convolutional neural network using the AlexNet method is as follows:

[0040] Perform feature extraction operations on the AlexNet model in the existing convolutional neural network structure. The feature extraction operations include setting 3 convolutional layers in the AlexNet model and unifying the convolutional kernel sizes of each convolutional layer of the AlexNet model;

[0041] Then, the scale of the convolutional kernels is uniformly set to 3×3 (the 3×3 convolutional size is relatively stable, and the receptive field is suitable for the images we train). The numbers of convolutional kernels are 32 for the first convolutional layer, 64 for the second convolutional layer, and 128 for the third convolutional layer. The convolutional layers are used to extract image features from the input information. The image features are reflected by each pixel in the image in a combined or independent manner, such as the texture features and color features of the picture. Here, the convolution operation is to perform cross-correlation operations on the matrices of each channel from left to right and from top to bottom by the convolutional kernels (first from left to right, and then from top to bottom, so the convolution operation also preserves the position information). It is like a small window that slides step by step from the upper left corner to the lower right corner. The sliding step size is a hyperparameter. The meaning of cross-correlation operation is to multiply and add the corresponding positions, and finally add the values of the three channels correspondingly to get a value.

[0042] Then, improve the 5 convolutional layers in the AlexNet model, replace these 5 convolutional layers with 3 convolutional layers, and add a layer of max pooling layer after each convolutional layer, that is, perform pooling operation after the convolutional calculation of the convolutional layer. After the third pooling layer, perform the calculation of the fully connected layer and connect 2 fully connected layers.

[0043] Then, connect a softmax classifier in the AlexNet model, that is, select the softmax activation function for the second fully connected layer.

[0044] Then, send the vibration signal data in the model dataset after sample expansion to the first convolutional layer of the AlexNet model for convolutional kernel shift convolution calculation. The calculated result is pooled and then sent to the second convolutional layer for convolutional kernel shift convolution calculation. The calculated result is pooled and then sent to the third convolutional layer for convolutional kernel shift convolution calculation. The calculated result is pooled, and then connected to two fully connected layers to obtain a convolutional feature map, that is, the construction of the convolutional neural network model is completed. The dimension of the convolutional feature values in this convolutional feature map is equal to the dimension of the vibration signal data in the model dataset after expansion.

[0045] In the above technical solution, the original five convolutional layers in the AlexNet model are as follows: the first convolutional layer uses 96 convolutional kernels of 11×11×3, with a stride of 4; the second convolutional layer uses 128 convolutional kernels of 5×5×48, with a stride of 1; the third convolutional layer uses 384 convolutional kernels of 3×3×256, with a stride of 1; the fourth convolutional layer uses 384 convolutional kernels of 3×3×192, with a stride of 1; the fifth convolutional layer uses 256 convolutional kernels of 3×3×192, with a stride of 1. After improvement, the three convolutional layers adopted in the present invention are as follows: the first convolutional layer is 32 convolutional kernels of 3×3×3, with a stride of 1; the second convolutional layer is 3×3×32 convolutional kernels, with a stride of 1; the third convolutional layer is 3×3×64 convolutional kernels, with a stride of 1. Replacing these five convolutional layers with three convolutional layers can prevent overfitting and improve the model accuracy, as Figure 3 shown.

[0046] In the above technical solution, the main role of the pooling layer in the convolutional neural network structure is to downsample the feature map. In the image classification task, the model does not need to learn the position of the target object, but it must ensure that the position where the object is located has no impact on the final recognition result of the model. The characteristic of the pooling layer is that it can ensure the invariance of transformations such as translation and rotation, enhancing the robustness of the model. At the same time, the downsampling process of pooling reduces the size of the feature map, reducing the computational amount of the model, but it can better maintain high-resolution features. Usually, the pooling layer in the convolutional neural network model is cascaded after the convolutional layer, and the commonly used pooling types are average pooling (Averagepooling) and max pooling (Maxpooling). Average pooling means taking the average of all values of the feature map on the pooling window, and max pooling means taking the maximum value of the feature values in the area where the pooling window is located as the feature value of the next layer of the feature map. In this paper, max pooling is adopted to avoid the blurring effect of average pooling and enhance the richness of features. The functions of the max pooling layer are, first, to further reduce the dimension of the information extracted by the convolutional layer and reduce the computational amount; second, to strengthen the invariance of image features, making it more robust to image offsets, rotations, etc.

[0047] In the above technical solution, the fully connected layer (Fully Connected Layer) is generally located at the end of the entire convolutional neural network and is responsible for converting the two-dimensional feature map output by the convolution into a one-dimensional vector, thereby realizing an end-to-end learning process (that is: input an image or a piece of speech, and output a vector or information). Each node in the fully connected layer is connected to all nodes in the previous layer, so it is called a fully connected layer. Due to its fully connected characteristics, generally, the fully connected layer has the most parameters.

[0048] The main function of the fully connected layer is to map the feature space samples obtained from the previous layers (such as convolutional and pooling layers) to the sample label space. Simply put, it integrates the feature representations into a single value. Its advantage lies in reducing the impact of feature positions on the classification results and improving the robustness of the entire network.

[0049] In the above technical solution, the Softmax function is an activation function used for multi-class classification problems. In multi-class classification problems, when there are more than two class labels, class membership is required. It can compress a K-dimensional vector z containing arbitrary real numbers into another K-dimensional real vector σ(z), such that the range of each element is between (0, 1), and the sum of all elements is 1. This function is mostly used in multi-class classification problems.

[0050] The softmax function is first and foremost a function that transforms a vector containing K real values into a K-dimensional real vector with a sum of 1. The K input values of the vector can be positive, negative, zero, or greater than 1, but softmax can transform them into values between 0 and 1, so they can be interpreted as probabilities. If one of the inputs is small or negative, softmax makes it a small probability, and if the input is large, it makes it a large probability, but always keeps it between 0 and 1.

[0051] In the above technical solution, the formula for the convolution kernel is as follows:

[0052]

[0053] Where: f[x, y] represents the vibration signal data in the augmented model dataset, f[n1, n2] is the vibration signal data at the n1-th row and n2-th column in the augmented model dataset, g[n1, n2] is the convolution matrix, which is the weight value at the n1-th row and n2-th column of the convolution kernel, used to retain the feature information in the image and remove useless pixels. The feature value at the x-th row and y-th column in the output feature map after convolution operation, n1 and n2 are the convolution pixel positions in the convolution region, "×" represents the multiplication of each pixel point, and a represents the set number of summation times.

[0054] In the above technical solution, the specific method for setting the hyperparameters of the AlexNet model training in the convolutional neural network by the model training module is as follows: The resolution of the vibration signal data in the model dataset input to the AlexNet model is 224×224. The loss function of the AlexNet model is set to cross-entropy. The cross-entropy loss function is located in the training part of the model. In the last layer (fully connected layer) of the model, SGD (stochastic gradient descent) is used to optimize and update the hyperparameters. The hyperparameters include the weights and biases in the convolutional layer and the fully connected layer, and the network improves the recognition effect through the weight parameters.

[0055] In the above technical solution, the AlexNet model is iterated for 20 epochs, and the batch training method is used to train the vibration signal data after sample augmentation in the model data set. The final result is shown in Figure 4 , batch training is a training method in deep learning. It divides the training data set into several batches, randomly extracts a batch of data from the data set for training each time, and the parameter update during training is based on the gradient calculated from this batch of data, rather than the gradient of the entire data set;

[0056] When training the AlexNet model with three designed convolutional layers, after each epoch iteration, the accuracy of the validation set is tested once. When training the AlexNet model, the data set is divided into three parts: training set, validation set, and test set. During the training process, the performance of the model can be monitored by testing the accuracy of the validation set after each epoch iteration. Specifically, after each epoch iteration, the validation set can be used to test the model, and the accuracy or error of the model on the validation set can be calculated. This accuracy or error can be used as an indicator of the model's performance to determine whether the model is overfitting or underfitting, and to adjust the hyperparameters of the model, etc.

[0057] In machine learning, an epoch refers to the number of times the entire training set data is passed through completely. When training a neural network, the training data set is usually divided into multiple mini-batches, and a mini-batch of data is randomly selected from the data set for training each time. Such an iteration is called an iteration, and an epoch refers to the situation where all the training set data has undergone one iteration. Usually, one epoch of training will contain multiple iterations. After one epoch ends, all the data in the data set will be used once, and at this time, the parameters of the model will be updated once.

[0058] In the above technical solution, when the signal test module uses the trained AlexNet model to test the test set, first, the trained AlexNet model is saved in the SavedModel format, then, the test set is stored in the saved AlexNet model, and then the inference operation is performed.

[0059] In the above technical solution, an orbital vehicle passing condition diagram is generated according to the test results (all the test results of testing the test set with the saved AlexNet model are saved in a txt file, and then the content in the txt file is processed with Excle to generate an orbital vehicle passing condition diagram). The specific method for determining the orbital safety condition of the measured area of the vehicle vibration signal to be measured through the orbital vehicle passing condition diagram is as follows:

[0060] When the time-domain waveforms in the same measurement area in the track passing vehicle condition diagram are regular and the passing vehicle signals are unchanged for multiple times, it indicates that the measurement area is normal. When the time-domain waveforms in the same measurement area in the track passing vehicle condition diagram are not regular and / or the amplitudes and / or frequencies of the waveforms corresponding to different passing vehicles exceed the corresponding thresholds, it indicates that the measurement area is abnormal. As shown in the appendix Figure 6 and the appendix Figure 7 It can be seen that there are obvious differences in signal amplitude and waveform between abnormal signals and normal signals;

[0061] When A consecutive measurement areas are abnormal, it indicates that the train wheel-rail condition is abnormal; when B discontinuous measurement areas are abnormal, it indicates that there are problems with part of the roadbed; when signals are not detected in C measurement areas and signals are detected in D measurement areas, it indicates that there is an external intrusion. In this embodiment, A is taken as 120, B is taken as 2, C is taken as 117, and D is taken as 3.

[0062] An orbit safety detection method based on an improved AlexNet, as Figures 1 to 7 shown, it includes the following steps:

[0063] Step 1: Each measurement area of the vehicle running vibration signal in the track distributed acoustic wave sensing system can sense the vehicle running vibration signal of the corresponding measurement area;

[0064] Step 2: Compare the vehicle running vibration signals in each selected historical period in each measurement area of the vehicle running vibration signal with the standard vehicle running vibration signal of the corresponding measurement area of the vehicle running vibration signal to determine whether the vehicle running vibration signals in each selected historical period in each measurement area of the vehicle running vibration signal are normal vehicle running vibration signals or abnormal vehicle running vibration signals;

[0065] Step 3: Set corresponding folders for each measurement area of the vehicle running vibration signal, and put the normal vehicle running vibration signals or abnormal vehicle running vibration signals in each selected historical time period of the corresponding measurement area of the vehicle running vibration signal in each folder to form a model data set;

[0066] Step 4: Perform data enhancement operations on the model data set to obtain a model data set with expanded samples, and divide the vibration signal data in the model data set with expanded samples into a training set, a validation set, and a test set;

[0067] Step 5: Use the AlexNet method to construct a convolutional neural network;

[0068] Step 6: According to the accuracy required when training the model, set the hyperparameters for training the AlexNet model in the convolutional neural network, and use the above hyperparameters to train the AlexNet model using the training set;

[0069] Step 7: Use the trained AlexNet model to test the test set, generate an orbital vehicle passing condition diagram according to the test results, and determine the track safety condition of the measured vehicle vibration signal measurement area through the orbital vehicle passing condition diagram.

[0070] The said Step 1 includes:

[0071] Step 101: The interference grating array vibration sensing system adopted in the present invention mainly consists of a UWFBG array and an unbalanced Michelson interference structure. The UWFBG array serves as a sensing network. Every two adjacent UWFBGs and the optical fiber therein form a sensing measurement area, and parameters such as the reflectivity, bandwidth, and central wavelength of each UWFBG are basically the same. The unbalanced Michelson interference structure mainly includes a 3×3 coupler, two optical fibers with different lengths for delaying light, and two Faraday mirrors. The optical path difference between the two optical fibers for delaying light is the same as the spacing between adjacent UWFBGs, which is used to make up for the optical path difference between adjacent UWFBGs; both of the two Faraday mirrors adopt a rotation angle of 45°, which is used to eliminate the polarization effect in the interferometer. The frequency response range of vibration signal sensing is improved through an equal-arm Mach-Zehnder interferometer. A sensing array is composed of ultra-weak reflection fiber gratings with a reflectivity of only 0.01%, and the spatial position information is converted into a time-domain delay to realize the multi-point vibration signal positioning in distributed vibration sensing, as Figure 2 shown.

[0072] Step 102: A tunable narrow linewidth laser emits continuous light with an optical power of 3 mW, which serves as the light source of the distributed fiber vibration sensing network. Its wavelength tunable range is 1500 nm to 1630 nm, and it can be applicable to various fiber grating sensing networks with different central wavelengths; an acousto-optic modulator performs pulse modulation on the continuous light. It has a high extinction ratio, reaching 63 dB, and the modulated optical pulse width is 20 ns, and the frequency is 100 kHz; after the pulsed light is amplified in power by an EDFA, it enters the sensing network. The sensing network is composed of weak reflection fiber gratings with a reflectivity of 0.01% and a sensing optical fiber with a length of 3 m.

[0073] The pulsed light sequence reflected back by the sensing channel enters an unequal-arm Mach-Zehnder interferometer with an arm length difference of 6 m. The optical phase information corresponding to the vibration at the sensing position is converted into optical intensity information by using interference technology, output through a 3×3 coupler, and connected to 3 identical photodetectors for photoelectric conversion. Finally, the electrical signals are collected and the vibration sensing information is obtained through a phase demodulation algorithm. Among them, in the case of an ideal splitting ratio, the optical intensity formulas of the three output ends of the said 3×3 symmetric coupler are as follows:

[0074]

[0075] where, I n, where n = 1, 2, 3, representing the first, second, and third output channels of a 3×3 symmetric coupler respectively. The interfering light is the input signal of the 3×3 coupler. When output from the 3×3 coupler, it is evenly divided into three optical signals. I1 and I2 are the intensities of the two optical pulses that generate the interfering light. is the phase change caused by external factors.

[0076] The two signals interfere at the 3×3 coupler after being reflected by two non-equidistant delay optical fibers and a Faraday rotator mirror. Since the length difference between the delay optical fibers is the distance of the sensing area, the time delay between the two interfering signals is equal to the time difference between the probe light reaching two UWFBGs in the sensing area, that is, each interfering pulse is obtained by the interference of the reflected lights of two UWFBGs in the corresponding area. Finally, three photodetectors (PD1, PD2, PD3) are used to convert the three output optical signals of the 3×3 coupler into electrical signals, and then the vibration signal is restored according to the phase change caused by vibration. The role of the above formula is to obtain the output signal of the 3×3 coupler, that is, the input signals of the three photodetectors. Through the photoelastic effect, changes in the external environment (vibration is most obvious when a subway train passes by) will cause changes in the refractive index, diameter, and length of the optical fiber, and these indicators have a certain linear relationship with the phase of the reflected light in the optical fiber. Changes in these indicators will cause corresponding changes in the phase of the reflected light, and the subsequent demodulation device reads the vibration information by demodulating the phase change. That is to say, through the photoelastic effect, the vibration information is modulated into a phase change, and then the phase information is calculated through demodulation means such as a 3*3 data coupler, and the vibration information can be restored through demodulation.

[0077] The present invention uses a weak grating fiber array to collect vibration signals and a neural network to process and identify the demodulated signals. The optical fiber can achieve full-time and full-domain detection of the track, and the neural network can improve the accuracy and efficiency of identification and give early warnings of abnormal conditions of the track.

[0078] The content not described in detail in this specification belongs to the prior art well-known to those skilled in the art.

Claims

1. An orbit safety detection system based on improved AlexNet, characterized in that: It includes an abnormal vibration signal judgment module, a vibration signal data set establishment module, a data enhancement module, a convolutional neural network model construction module, a model training module and a signal testing module; Each vehicle vibration signal measuring area in the track distributed acoustic wave sensing system can sense the vehicle vibration signal of the corresponding measuring area; The abnormal vibration signal judgment module is used to compare the driving vibration signal of each selected historical period in each driving vibration signal measurement area with the standard driving vibration signal of the corresponding driving vibration signal measurement area in terms of amplitude and frequency, and determine whether the driving vibration signal of each selected historical period in each driving vibration signal measurement area is a normal driving vibration signal or an abnormal driving vibration signal; The vibration signal data set establishment module is used to set a corresponding folder for each driving vibration signal measurement area, and put the normal driving vibration signal or abnormal driving vibration signal of each historical time period selected in the corresponding driving vibration signal measurement area into each folder to form a model data set; The data enhancement module is used to perform data enhancement operations on the model data set to obtain a model data set after sample expansion, and divide the vibration signal data in the model data set after sample expansion into a training set, a verification set and a test set; The convolutional neural network model building module is used to build a convolutional neural network using the AlexNet method; The model training module is used to set the hyperparameters of the AlexNet model training in the convolutional neural network according to the accuracy required when training the model, and use the above hyperparameters to train the AlexNet model using the training set; The signal test module is used to test the test set using the trained AlexNet model, generate a track vehicle passing condition diagram based on the test results, and determine the track safety status of the vehicle vibration signal measurement area to be tested through the track vehicle passing condition diagram; The track traffic condition diagram is generated based on the test results. The specific method of determining the track safety condition of the vehicle vibration signal measurement area to be tested through the track traffic condition diagram is as follows: When the time domain waveform of the same measurement area in the track vehicle passing condition diagram has regularity and the multiple vehicle passing signals remain unchanged, it indicates that the measurement area is normal. When the time domain waveform of the same measurement area in the track vehicle passing condition diagram does not have regularity and / or the amplitude and / or frequency of the waveforms corresponding to different vehicle passings exceeds the corresponding threshold, it indicates that the measurement area is abnormal. When A consecutive measurement areas are abnormal, it means that the train wheel and rail conditions are abnormal; When the B intermittent measurement areas are abnormal, it means that there is a problem with part of the roadbed; when no signal is detected in the C measurement area and a signal is detected in the D measurement area, it indicates that external intrusion has occurred.

2. The track safety detection system based on the improved AlexNet according to claim 1, characterized in that: The abnormal vibration signal judgment module is used to compare the driving vibration signal of each driving vibration signal measurement area in each time period with the standard driving vibration signal of the corresponding driving vibration signal measurement area. When the difference between the driving vibration signal amplitude of a certain driving vibration signal measurement area at a certain moment and the standard driving vibration signal amplitude of the corresponding driving vibration signal measurement area is greater than a preset amplitude threshold, or the difference between the driving vibration signal frequency of the driving vibration signal measurement area at that moment and the standard driving vibration signal frequency of the corresponding driving vibration signal measurement area is greater than a preset frequency threshold, the driving vibration signal of the driving vibration signal measurement area at that moment is determined to be an abnormal vibration signal; When the difference between the amplitude of the train vibration signal at a certain moment in a certain train vibration signal measurement area and the amplitude of the standard train vibration signal in the corresponding train vibration signal measurement area ≤ the preset amplitude threshold, and at the same time, the difference between the frequency of the train vibration signal at this moment in this train vibration signal measurement area and the frequency of the standard train vibration signal in the corresponding train vibration signal measurement area ≤ the preset frequency threshold, it is determined that the train vibration signal at this moment in this train vibration signal measurement area is a normal vibration signal.

3. The rail safety detection system based on the improved AlexNet according to claim 1, characterized in that: The data augmentation operation is to call the OpenCV module under the TensorFlow unit to perform sample augmentation of the model dataset. The ways of sample augmentation include: random angle rotation, horizontal flipping, and vertical flipping.

4. The rail safety detection system based on the improved AlexNet according to claim 1, characterized in that: The specific method for the convolutional neural network model construction module to construct a convolutional neural network using the AlexNet method is as follows: Perform feature extraction operations on the AlexNet model in the existing convolutional neural network structure. The feature extraction operations include setting 3 convolutional layers in the AlexNet model and unifying the convolutional kernel sizes of each convolutional layer of the AlexNet model; Then, uniformly set the convolutional kernel scale to 3×3; Then, improve the 5 convolutional layers in the AlexNet model, replace these 5 convolutional layers with 3 convolutional layers, and add a max-pooling layer after each convolutional layer, that is, perform pooling operation after the convolutional calculation of the convolutional layer. After the third pooling layer, perform the calculation of the fully connected layer and connect 2 fully connected layers; Then, connect a softmax classifier in the AlexNet model; Then, send the vibration signal data in the model dataset after sample augmentation to the first convolutional layer of the AlexNet model for convolutional kernel shift convolutional calculation. The calculated result is pooled and then sent to the second convolutional layer for convolutional kernel shift convolutional calculation. The calculated result is pooled and then sent to the third convolutional layer for convolutional kernel shift convolutional calculation. The calculated result is pooled and then connected to two fully connected layers to obtain a convolutional feature map, that is, the construction of the convolutional neural network model is completed. The dimension of the convolutional feature values in this convolutional feature map is equal to the dimension of the vibration signal data in the model dataset after augmentation.

5. The track safety detection system based on the improved AlexNet according to claim 4, characterized in that: The formula for the convolutional kernel is as follows: Where: f[x, y] represents the vibration signal data in the model dataset after augmentation, f[n1, n2] is the vibration signal data in the n1-th row and n2-th column in the model dataset after augmentation, g[n1, n2] is the convolutional matrix, which is the weight value of the n1-th row and n2-th column of the convolutional kernel, used to retain the feature information in the image, remove useless pixels, and the eigenvalue of the x-th row and y-th column in the output feature map after convolutional operation. n1 and n2 are the positions of the convolutional pixels in the convolutional region, "×" represents the multiplication of each pixel point, and a represents the set number of summations.

6. The rail safety detection system based on the improved AlexNet according to claim 1, characterized in that: The specific method for the model training module to set the hyperparameters for AlexNet model training in the convolutional neural network is as follows: the resolution of the vibration signal data in the model data set input to the AlexNet model is 224×224, the loss function of the AlexNet model is set to cross entropy, and SGD optimization is used to update the hyperparameters, which include weights and biases in the convolutional layer and the fully connected layer.

7. The rail safety detection system based on the improved AlexNet according to claim 6, characterized in that: The AlexNet model iterates for 20 epochs, and uses a batch training method to train the vibration signal data after sample expansion in the model data set; When training the AlexNet model with three convolution layers, the accuracy of the validation set is tested once after each epoch.

8. The rail safety detection system based on the improved AlexNet according to claim 1, characterized in that: When the signal testing module uses the trained AlexNet model to test the test set, the trained AlexNet model is first saved in a SavedModel format, and then the test set is stored in the saved AlexNet model, and then an inference operation is performed.

9. An orbit safety detection method based on improved AlexNet, characterized in that, It includes the following steps: Step 1: Each vehicle vibration signal measuring area in the track distributed acoustic wave sensing system can sense the vehicle vibration signal of the corresponding measuring area; Step 2: Compare the amplitude and frequency of the driving vibration signal of each selected historical period in each driving vibration signal measurement area with the standard driving vibration signal of the corresponding driving vibration signal measurement area, and determine whether the driving vibration signal of each selected historical period in each driving vibration signal measurement area is a normal driving vibration signal or an abnormal driving vibration signal; Step 3: Set a corresponding folder for each driving vibration signal measurement area, and put the normal driving vibration signal or abnormal driving vibration signal of each historical time period selected in the corresponding driving vibration signal measurement area into each folder to form a model data set; Step 4: Perform data enhancement operation on the model data set to obtain a model data set after sample expansion, and divide the vibration signal data in the model data set after sample expansion into a training set, a validation set, and a test set; Step 5: Use AlexNet method to build a convolutional neural network; Step 6: According to the accuracy required when training the model, set the hyperparameters of the AlexNet model training in the convolutional neural network, and use the above hyperparameters to train the AlexNet model using the training set; Step 7: Use the trained AlexNet model to test the test set, generate a track traffic condition map based on the test results, and determine the track safety status of the vehicle vibration signal measurement area to be tested through the track traffic condition map; The track traffic condition diagram is generated based on the test results. The specific method of determining the track safety condition of the vehicle vibration signal measurement area to be tested through the track traffic condition diagram is as follows: When the time domain waveform of the same measurement area in the track vehicle passing condition diagram has regularity and the multiple vehicle passing signals remain unchanged, it indicates that the measurement area is normal. When the time domain waveform of the same measurement area in the track vehicle passing condition diagram does not have regularity and / or the amplitude and / or frequency of the waveforms corresponding to different vehicle passings exceeds the corresponding threshold, it indicates that the measurement area is abnormal. When A consecutive measurement areas are abnormal, it means that the train wheel and rail conditions are abnormal; When there are abnormalities in B discontinuous measurement areas, it indicates that there are problems with part of the roadbed; when signals are not detected in C measurement areas and signals are detected in D measurement areas, it indicates an external intrusion.

Images