阅读说明：本技术 基于Faster R-CNN的铁路入侵行为检测方法 (Railway intrusion behavior detection method based on fast R-CNN ) 是由 惠一龙 马鑫蕊 肖潇 李长乐 段江华 于 2021-07-12 设计创作，主要内容包括：本发明提出了一种基于Faster R-CNN的铁路异常入侵行为检测方法,用于解决现有技术中存在的检测准确率较低的技术问题,实现步骤为：构建DAS数据处理系统；获取训练样本集和测试样本集；构建铁路入侵行为检测网络模型Faster R-CNN；对铁路入侵行为检测网络模型Faster R-CNN进行迭代训练；获取铁路异常入侵行为检测结果。本发明所构建的网络模型Faster R-CNN使用归一化时空信号图像作为训练样本集,充分结合信号的时空特征,区分背景噪声信号的干扰,减少误报,同时候选区域生成网络精确预测特征图的区域候选框位置,一定程度上提高了检测准确率,可用于保护铁路列车的安全运行。(The invention provides a railway abnormal intrusion behavior detection method based on Faster R-CNN, which is used for solving the technical problem of lower detection accuracy in the prior art and comprises the following steps: constructing a DAS data processing system; acquiring a training sample set and a test sample set; building a railway intrusion behavior detection network model Faster R-CNN; performing iterative training on a railway intrusion behavior detection network model Faster R-CNN; and obtaining a detection result of the abnormal intrusion behavior of the railway. The network model Faster R-CNN constructed by the invention uses the normalized space-time signal image as a training sample set, fully combines the space-time characteristics of signals, distinguishes the interference of background noise signals, reduces false alarm, and simultaneously generates the area candidate frame position of the network accurate prediction characteristic diagram in the candidate area, thereby improving the detection accuracy to a certain extent and being used for protecting the safe operation of railway trains.)

1. A railway abnormal intrusion behavior detection method based on FasterR-CNN is characterized by comprising the following steps:

(1) constructing a DAS data processing system:

constructing a DAS data processing system comprising a cascaded optical fiber distributed acoustic sensing DAS sub-module and a data processing sub-module, wherein the optical fiber distributed acoustic sensing DAS sub-module comprises a DAS vibration detection optical cable, an optical signal demodulation host and a monitoring terminal analysis host which are cascaded in sequence, the DAS vibration detection optical cable is laid along a railway fence and comprises N sampling points distributed at equal intervals; the output end of the monitoring terminal analysis host is connected with the data processing submodule, wherein N is more than or equal to 2;

(2) acquiring a training sample set V and a testing sample set F:

(2a) the optical fiber distributed acoustic sensing DAS sub-module acquires a preprocessed vibration optical signal:

(2a1) n sampling points of DAS vibration detection optical cables collect vibration optical signals C ═ C { C ═ generated by disturbance of K external invasion behavior categories of T time railway fences_n,tN is more than 0 and less than N, T is more than 0 and less than T, and C is transmitted to the optical signal demodulation host, wherein T is more than or equal to 6000000, K is more than or equal to 1, and C is more than or equal to T_n,tRepresenting the vibration optical signal of the t moment acquired by the nth sampling point;

(2a2) the optical signal demodulation host is used for each vibration lightSignal of science C_n,tModulating and amplifying the signal, and amplifying the amplified vibration optical signal C '({ C'_n,tL 0 < N < N,0 < T < T } is transmitted to the analysis host computer of the monitoring terminal, wherein C'_n,tIs represented by C_n,tA corresponding amplified vibration optical signal;

(2a3) the monitoring terminal analysis host machine filters the received vibration optical signal C' and sets the vibration optical signal S after filtering to { S ═ S_n,tL 0 < N < N,0 < T < T) is drawn into a railway abnormal intrusion behavior signal data table with dimension of T multiplied by N, and then the table is sent to a data processing submodule, wherein S_n,tIs C'_n,tFiltering results;

(2b) the data processing submodule converts each filtered vibration optical signal S in the table_n,tNormalization is performed simultaneously along the dimension T and the dimension N to obtain a normalized vibration optical signal S '═ S'_n,tThe method comprises the steps of obtaining a railway abnormal intrusion behavior signal data table with |0 < N < N,0 < T < T), uniformly dividing S' into Q groups according to the dimension T, then drawing a normalized space-time signal image of each group of vibration optical signals by taking the time T of each group of normalized vibration optical signals as a vertical coordinate, taking a sampling point N as a horizontal coordinate, and expressing the magnitude of the normalized vibration optical signal value by different colors to obtain a data set containing Q normalized space-time signal images, wherein Q is more than or equal to 2000;

(2c) labeling abnormal intrusion behavior signal regions in each normalized space-time signal image, and randomly selecting M of Q normalized space-time signal images containing abnormal intrusion behavior signal region labeling frames as training sample sets V ═ V₁,V₂,...,V_m,...,V_MV is Z ═ Z₁',Z'₂,...,Z'_m,...,Z'_MAnd taking the rest Q-M normalized spatio-temporal signal images containing the abnormal intrusion behavior signal region labeling frame as a test sample set F ═ F₁,F₂,...,F_w,...,F_WAnd (c) the step of (c) in which,V_mrepresenting the m-th training sample with a label box, Z'_mRepresents V_mCorresponding authentic label, F_wDenotes the W-th test sample with label box, W-Q-M;

(3) constructing a railway intrusion behavior detection network model FasterR-CNN:

(3a) constructing a railway intrusion behavior detection network model FasterR-CNN comprising a cascaded feature extraction network, a candidate region generation network RPN, an ROI pooling layer and a classification regression network, wherein the feature extraction network adopts a VGG16 convolutional neural network comprising 13 convolutional layers, 13 activation function layers and 4 pooling layers; the RPN comprises a cascade shared convolution layer, an activation function layer and a candidate region recommendation layer, wherein the activation function layer and the candidate region recommendation layer are directly loaded with SoftMax classification branches B which are arranged in parallel₁And bounding Box regression convolution layer B₂(ii) a The classification regression network comprises two full connection layers F₁And with F₁Parallel connected SoftMax sorted fully connected layer F₂And regression full-junction layer F₃；

(3b) Constructing a classification loss function L_clsAnd a regression loss function L_regLoss function L of composed candidate region generation network RPN_RPN：

Wherein p is_iIndicates the i-th preset-box prediction label,represents the real label, u, corresponding to the ith preset box_i＝{u_x,u_y,u_w,u_hDenotes the predicted bounding box regression parameter vector for the ith preset box,representing the regression parameter vector, N, of the true bounding box corresponding to the ith preset box_clsRepresenting the total number of samples, N, in a minimum batch value_regExpressing the number of pixel points of each characteristic graph, and expressing a balance weight parameter by lambda;

(4) performing iterative training on a railway intrusion behavior detection network model FasterR-CNN:

(4a) initializing the iteration number as K, the maximum iteration number as K, and setting the K-th iteration railway intrusion behavior detection network model FasterR-CNN as H^kAnd let k equal to 1, H^k＝H；

(4b) Taking the training sample set V as a railway intrusion behavior detection network model H^kFor each training sample V, the feature extraction network_mExtracting the abnormal behavior signal characteristics to obtain M characteristic graphs; predicting the accurate position of the candidate frame of the abnormal intrusion behavior region of each feature map by the candidate region generation network RPN to obtain M feature maps containing the accurate position prediction of the candidate frame of the abnormal intrusion behavior region;

(4c) the ROI pooling layer performs pooling operation on each feature map output by the network RPN generated by the candidate region to obtain M feature maps with the same size;

(4d) two fully-connected layers F in a classification regression network₁Performing convolution operation on each characteristic graph, and classifying the full-connected layer F by SoftMax₂Classifying the convolution result of each feature map to obtain a prediction label set Z ═ Z corresponding to the training sample set V₁,Z₂,...,Z_m,...,Z_MGet back to the full connection layer F at the same time₃For each special frameRegression is carried out on the convolution result of the figure graph to obtain a predicted position offset vector set A which corresponds to the training sample set V and is { A }₁,A₂,...,A_m,...,A_MIn which Z is_mAnd A_mRespectively represent V_mA corresponding prediction tag and prediction position offset vector;

(4e) using a classification loss function L_clsCalculating each predictive label Z_mAnd its corresponding real tag Z'_mThen carrying out mean value processing on the M loss values to obtain H after the kth iteration^kClassification loss value ofAnd using a back propagation algorithm based onClassifying fully-connected layers F for SoftMax in RPN (resilient packet network)₂Parameter (d) ofUpdating is carried out;

(4f) using a regression loss function L_regCalculating each predicted position offset vector A_mTrue position vector A 'corresponding thereto'_mThen carrying out mean value processing on the M loss values to obtain H after the kth iteration^kValue of regression loss ofAnd using a back propagation algorithm based onGenerating regression full connectivity layer F in network RPN for region₃Parameter (d) ofUpdating to obtain H after the kth iteration^k；

(4g) Judging that K is KIf not, obtaining a trained railway intrusion behavior detection network model H^*Otherwise, let k equal to k +1, and execute step (4 b);

(5) obtaining a detection result of the abnormal intrusion behavior of the railway:

(5a) taking the test sample set F as a trained railway intrusion behavior detection network model H^*Performing abnormal behavior signal feature extraction on each test sample by a feature extraction network to obtain a feature map of W images;

(5b) predicting the accurate position of the candidate frame of the abnormal intrusion behavior region of each feature map by the candidate region generation network RPN to obtain W feature maps containing the accurate position prediction of the candidate frame of the abnormal intrusion behavior region;

(5c) the ROI pooling layer performs pooling operation on each feature map output by the candidate area generation network to obtain W feature maps with the same size;

(5d) two fully-connected layers F in a classification regression network₁Performing convolution operation on each characteristic graph, and classifying the full-connected layer F by SoftMax₂Classifying the convolution result of each characteristic graph to obtain a prediction label set corresponding to the test sample set FSimultaneous regression full junction layer F₃Regression is carried out on the convolution result of each characteristic graph to obtain a prediction position offset vector set corresponding to the test sample set FWherein the content of the first and second substances,andrespectively represent F_wA corresponding prediction tag and a predicted position offset vector.

2. The method for detecting abnormal intrusion behavior of railway based on Faster R-CNN as claimed in claim 1, wherein the candidate region generating network RPN in step (4b) predicts the accurate position of the candidate frame of the abnormal intrusion behavior region of each feature map by the following steps:

(4b1) the shared convolution layer extracts a plurality of pixel feature points of each feature map and plots each pixel feature point on the feature map with a size of 128²、256²、512²Nine preset frames with the length-width ratio of 2:1, 1:1 and 1:2, and outputting M characteristic diagrams each comprising a plurality of preset frames;

(4b2) the activation function layer carries out nonlinear mapping on pixel feature points in each feature graph output by the shared convolution layer to obtain M feature graphs after the nonlinear mapping, wherein the formula for carrying out the nonlinear mapping on each pixel feature point is as follows:

RELU(c)＝max(0,c)

wherein c is the pixel characteristic point in each characteristic graph output by the shared convolutional layer;

(4b3) SoftMax sorting leg B₁After convolution is carried out on each feature graph output by the activation function layer, the convolution results are classified, and the prediction label p of the ith preset frame in each feature graph is output_iWhile bounding box regresses convolutional layer B₂Regression is carried out on the convolution result of each feature graph output by the activation function layer, and a border regression parameter vector u of the ith preset frame in each feature graph is output_i＝{u_x,u_y,u_w,u_hIn which u_x、u_y、u_w、u_hRespectively representing the central abscissa, the central ordinate, the width and the height of the ith preset frame;

(4b4) selecting a branch B between the real label set Z and SoftMax classification in the candidate area recommendation layer₁Outputting a predictive label p of the ith preset box in each feature map_iThe same preset frame is used as a region candidate frame of abnormal intrusion behavior, and then the convolution layer B is regressed through a boundary frame₂Outputting a regression parameter vector u of the prediction boundary box of the ith preset box in each feature map_iRefinement to region candidate boxesAnd (4) performing regression on the determined position to obtain a feature map containing the accurate position prediction of the candidate frame of the abnormal intrusion behavior region.

3. The FasterR-CNN-based railway abnormal intrusion behavior detection method according to claim 1, wherein the step (4e) is based onClassifying fully-connected layers F for SoftMax in RPN (resilient packet network)₂Parameter (d) ofPerforming an update, and according to step (4f)Generating regression full connectivity layer F in network RPN for region₃Parameter (d) ofUpdating, wherein the updating formulas are respectively as follows:

wherein eta represents the learning step length, 0 < eta is less than or equal to 0.1,andrespectively representAs a result of the update, the result of the update,representing a derivative operation.

Technical Field

The invention belongs to the technical field of intelligent traffic, and relates to a method for detecting railway abnormal intrusion behaviors, in particular to a method for detecting railway abnormal intrusion behaviors based on Faster R-CNN. The method can be used for safety monitoring of railway perimeter and protecting the safe operation of railway trains.

Background

Along with the development of railways and intelligent traffic systems, people bring rapidness and convenience in traffic, and security work along railways is more and more important. The illegal damage or the illegal crossing of the rail fence, the theft of the rail fence cable and other dangerous or malicious intrusions can cause serious accident potential hazards to the train operation safety, bring economic loss to people, possibly cause traffic jam in partial areas, and even more seriously cause the casualties of related personnel. The railway abnormal intrusion behavior is a behavior which threatens the running safety of a railway line and a train, and particularly refers to a behavior that a solid object with certain volume and mass outside the railway line rushes into the railway perimeter due to the action of some external force. How to monitor the safety of the railway perimeter and detect the abnormal intrusion behavior of the railway is an important and urgent problem to be concerned and solved.

Most railway abnormal intrusion behavior detection methods take the improvement of the detection accuracy as a first criterion, so the improvement of the detection accuracy of the abnormal intrusion behavior method is the key of research. The existing railway abnormal intrusion behavior detection method comprises an infrared laser correlation monitoring method, a video camera monitoring method, a vibration cable monitoring method, a microwave correlation monitoring method, an electronic fence monitoring method, an optical fiber distributed sound sensing monitoring method and the like.

With the wide application of big data and machine learning, some researchers use a method combining deep learning technology and distributed optical fiber acoustic sensing to detect railway intrusion behaviors so as to improve the detection accuracy rate of abnormal intrusion behaviors. For example, in the paper "Fiber distributed adaptive sensing using a resonant short-term memory network," optical express 283 (2020):2925 and 2938 published by Li, Zhongqi and Jiangwei Zhang in 2020, a field test on high-speed-gained railroads intrusion detection method based on a convolutional long-short term memory artificial neural network (ConvLSTM) and Fiber distributed acoustic sensing is disclosed. The method comprises the steps of firstly, periodically scanning collected vibration optical signals, constructing frame images required by a detection model, then inputting the frame images into the detection model based on ConvLSTM, and finally detecting each frame image to obtain a final railway intrusion behavior detection result. The method improves the detection accuracy, but false alarm is caused because the convolutional neural network in the ConvLSTM detection model is insufficient in extracting the characteristic information of the frame image, only the spatial characteristic of the signal can be extracted, the time dependence of the signal cannot be analyzed, and the interference caused by the background noise signal cannot be distinguished.

Disclosure of Invention

The invention aims to overcome the defects in the prior art, and provides a railway abnormal intrusion behavior detection method based on fast R-CNN, which is used for solving the technical problem of low detection accuracy in the prior art.

In order to achieve the purpose, the technical scheme adopted by the invention comprises the following steps:

(1) constructing a DAS data processing system:

constructing a DAS data processing system comprising a cascaded optical fiber distributed acoustic sensing DAS sub-module and a data processing sub-module, wherein the optical fiber distributed acoustic sensing DAS sub-module comprises a DAS vibration detection optical cable, an optical signal demodulation host and a monitoring terminal analysis host which are cascaded in sequence, the DAS vibration detection optical cable is laid along a railway fence and comprises N sampling points distributed at equal intervals; the output end of the monitoring terminal analysis host is connected with the data processing submodule, wherein N is more than or equal to 2;

(2) acquiring a training sample set V and a testing sample set F:

(2a) the optical fiber distributed acoustic sensing DAS sub-module acquires a preprocessed vibration optical signal:

(2a1) n sampling points of DAS vibration detection optical cables collect vibration optical signals C ═ C { C ═ generated by disturbance of K external invasion behavior categories of T time railway fences_n,tN is more than 0 and less than N, T is more than 0 and less than T, and C is transmitted to the optical signal demodulation host, wherein T is more than or equal to 6000000, K is more than or equal to 1, and C is more than or equal to T_n,tRepresenting the vibration optical signal of the t moment acquired by the nth sampling point;

(2a2) the optical signal demodulation host computer processes each vibration optical signal C_n,tModulating and amplifying the signal, and amplifying the amplified vibration optical signal C '({ C'_n,tL 0 < N < N,0 < T < T } is transmitted to the analysis host computer of the monitoring terminal, wherein C'_n,tIs represented by C_n,tA corresponding amplified vibration optical signal;

(2a3) the monitoring terminal analysis host machine filters the received vibration optical signal C' and sets the vibration optical signal S after filtering to { S ═ S_n,tL 0 < N < N,0 < T < T) is drawn into a railway abnormal intrusion behavior signal data table with dimension of T multiplied by N, and then the table is sent to a data processing submodule, wherein S_n,tIs C'_n,tFiltering results;

(2b) the data processing submodule converts each filtered vibration optical signal S in the table_n,tNormalization is performed simultaneously along the dimension T and the dimension N to obtain a normalized vibration optical signal S '═ S'_n,tThe method comprises the steps of obtaining a railway abnormal intrusion behavior signal data table with |0 < N < N,0 < T < T), uniformly dividing S' into Q groups according to the dimension T, then drawing a normalized space-time signal image of each group of vibration optical signals by taking the time T of each group of normalized vibration optical signals as a vertical coordinate, taking a sampling point N as a horizontal coordinate, and expressing the magnitude of the normalized vibration optical signal value by different colors to obtain a data set containing Q normalized space-time signal images, wherein Q is more than or equal to 2000;

(2c) labeling abnormal intrusion behavior signal regions in each normalized space-time signal image, and randomly selecting Q regional labels containing the abnormal intrusion behavior signal regionsM frames in the framed normalized spatio-temporal signal image are used as a training sample set V ═ V₁,V₂,...,V_m,...,V_MV is Z ═ Z'₁,Z'₂,...,Z'_m,...,Z'_MAnd taking the rest Q-M normalized spatio-temporal signal images containing the abnormal intrusion behavior signal region labeling frame as a test sample set F ═ F₁,F₂,...,F_w,...,F_WAnd (c) the step of (c) in which,V_mrepresenting the m-th training sample with a label box, Z'_mRepresents V_mCorresponding authentic label, F_wDenotes the W-th test sample with label box, W-Q-M;

(3) constructing a railway intrusion behavior detection network model, namely, fast R-CNN:

(3a) constructing a railway intrusion behavior detection network model, namely, a fast R-CNN, which comprises a cascaded feature extraction network, a candidate region generation network RPN, an ROI pooling layer and a classification regression network, wherein the feature extraction network adopts a VGG16 convolutional neural network comprising 13 convolutional layers, 13 activation function layers and 4 pooling layers; the RPN comprises a cascade shared convolution layer, an activation function layer and a candidate region recommendation layer, wherein the activation function layer and the candidate region recommendation layer are directly loaded with SoftMax classification branches B which are arranged in parallel₁And bounding Box regression convolution layer B₂(ii) a The classification regression network comprises two full connection layers F₁And with F₁Parallel connected SoftMax sorted fully connected layer F₂And regression full-junction layer F₃；

(3b) Constructing a classification loss function L_clsAnd a regression loss function L_regLoss function L of composed candidate region generation network RPN_RPN：

Wherein p is_iIndicates the i-th preset-box prediction label,represents the real label, u, corresponding to the ith preset box_i＝{u_x,u_y,u_w,u_hDenotes the predicted bounding box regression parameter vector for the ith preset box,representing the regression parameter vector, N, of the true bounding box corresponding to the ith preset box_clsRepresenting the total number of samples, N, in a minimum batch value_regExpressing the number of pixel points of each characteristic graph, and expressing a balance weight parameter by lambda;

(4) performing iterative training on a railway intrusion behavior detection network model, namely, fast R-CNN:

(4a) initializing the iteration number as K, the maximum iteration number as K, and setting the kth iterative railway intrusion behavior detection network model Faster R-CNN as H^kAnd let k equal to 1, H^k＝H；

(4b) Taking the training sample set V as a railway intrusion behavior detection network model H^kFor each training sample V, the feature extraction network_mExtracting the abnormal behavior signal characteristics to obtain M characteristic graphs; predicting the accurate position of the candidate frame of the abnormal intrusion behavior region of each feature map by the candidate region generation network RPN to obtain M feature maps containing the accurate position prediction of the candidate frame of the abnormal intrusion behavior region;

(4c) the ROI pooling layer performs pooling operation on each feature map output by the network RPN generated by the candidate region to obtain M feature maps with the same size;

(4d) two fully-connected layers F in a classification regression network₁Performing convolution operation on each characteristic graph, and classifying the full-connected layer F by SoftMax₂Classifying the convolution result of each feature map to obtain a prediction label set Z ═ Z corresponding to the training sample set V₁,Z₂,...,Z_m,...,Z_MGet back to the full connection layer F at the same time₃Regression is carried out on the convolution result of each feature map, and a prediction position offset vector set A which corresponds to the training sample set V is obtained₁,A₂,...,A_m,...,A_MIn which Z is_mAnd A_mRespectively represent V_mA corresponding prediction tag and prediction position offset vector;

(4e) using a classification loss function L_clsCalculating each predictive label Z_mAnd its corresponding real tag Z'_mThen carrying out mean value processing on the M loss values to obtain H after the kth iteration^kClassification loss value ofAnd using a back propagation algorithm based onClassifying fully-connected layers F for SoftMax in RPN (resilient packet network)₂Parameter (d) ofUpdating is carried out;

(4f) using a regression loss function L_regCalculating each predicted position offset vector A_mTrue position vector A 'corresponding thereto'_mThen carrying out mean value processing on the M loss values to obtain H after the kth iteration^kValue of regression loss ofAnd using a back propagation algorithm based onGenerating regression full connectivity layer F in network RPN for region₃Parameter (d) ofUpdating to obtain H after the kth iteration^k；

(4g) Judging whether K is true or not, if so, obtaining a trained railway intrusion behavior detection network model H^*Otherwise, let k equal to k +1, and execute step (4 b);

(5) obtaining a detection result of the abnormal intrusion behavior of the railway:

(5a) taking the test sample set F as a trained railway intrusion behavior detection network model H^*Performing abnormal behavior signal feature extraction on each test sample by a feature extraction network to obtain a feature map of W images;

(5b) predicting the accurate position of the candidate frame of the abnormal intrusion behavior region of each feature map by the candidate region generation network RPN to obtain W feature maps containing the accurate position prediction of the candidate frame of the abnormal intrusion behavior region;

(5c) the ROI pooling layer performs pooling operation on each feature map output by the candidate area generation network to obtain W feature maps with the same size;

(5d) two fully-connected layers F in a classification regression network₁Performing convolution operation on each characteristic graph, and classifying the full-connected layer F by SoftMax₂Classifying the convolution result of each characteristic graph to obtain a prediction label set corresponding to the test sample set FSimultaneous regression full junction layer F₃Regression is carried out on the convolution result of each characteristic graph to obtain a prediction position offset vector set corresponding to the test sample set FWherein the content of the first and second substances,andrespectively represent F_wA corresponding prediction tag and a predicted position offset vector.

Compared with the prior art, the invention has the following advantages:

1. in the process of training a railway intrusion behavior detection network model Faster R-CNN and acquiring a railway abnormal intrusion behavior detection result, the candidate region generation network RPN accurately predicts the position of a region candidate frame of a feature map, and the classification regression network convolves each feature map, so that the defect that the feature information of a frame image extracted by a convolutional neural network in the prior art is insufficient is overcome, and the detection accuracy of the railway abnormal intrusion behavior is effectively improved.

2. According to the invention, through a data processing submodule contained in a constructed DAS data processing system, a railway abnormal intrusion behavior signal data table is normalized along a T dimension and an N dimension at the same time, the time T of each group of normalized vibration optical signals is taken as a vertical coordinate, a sampling point N is taken as a horizontal coordinate, different colors represent the magnitude of the normalized vibration optical signal value, and a normalized spatio-temporal signal image of the group of vibration optical signals is drawn to obtain a data set containing Q normalized spatio-temporal signal images.

Drawings

FIG. 1 is a flow chart of an implementation of the present invention.

Fig. 2 is a spatio-temporal image of five railway abnormal intrusion behavior signals according to an embodiment of the present invention.

FIG. 3 is a schematic structural diagram of a railway intrusion detection network model Faster R-CNN according to the present invention.

Detailed Description

The invention is described in further detail below with reference to the figures and specific examples.

Referring to fig. 1, the present invention includes the steps of:

step 1) constructing a DAS data processing system:

constructing a DAS data processing system comprising a cascaded optical fiber distributed acoustic sensing DAS sub-module and a data processing sub-module, wherein the optical fiber distributed acoustic sensing DAS sub-module comprises a DAS vibration detection optical cable, an optical signal demodulation host and a monitoring terminal analysis host which are cascaded in sequence, the DAS vibration detection optical cable is laid along a railway fence and comprises N sampling points distributed at equal intervals; the output end of the monitoring terminal analysis host is connected with the data processing sub-module, wherein N is more than or equal to 2, in the embodiment, the optical fiber distributed acoustic sensing DAS sub-module is deployed on a railway system for field data acquisition, a fence of an experimental field is 5 meters away from the railway horizontal distance, the vibration detection optical cable is fixed on the fence, a standard G652 single-mode optical fiber is adopted, the maximum sensing range can reach 40 kilometers, the signal sampling frequency is 488Hz, and N is 70;

step 2) obtaining a training sample set V and a testing sample set F:

(2a) the optical fiber distributed acoustic sensing DAS sub-module acquires a preprocessed vibration optical signal:

(2a1) n sampling points of DAS vibration detection optical cables collect vibration optical signals C ═ C { C ═ generated by disturbance of K external invasion behavior categories of T time railway fences_n,tN is more than 0 and less than N, T is more than 0 and less than T, and C is transmitted to the optical signal demodulation host, wherein T is more than or equal to 6000000, K is more than or equal to 1, and C is more than or equal to T_n,tRepresenting the vibration optical signal collected at the T-th moment of the nth sampling point, wherein in the embodiment, T is 18000000, K is 5, and the five collected railway abnormal intrusion behaviors are train background noise interference, fence shaking and cage pricking, fence climbing, wall chiseling and excavation respectively;

(2a2) the optical signal demodulation host computer processes each vibration optical signal C_n,tModulating and amplifying the signal, and amplifying the amplified vibration optical signal C '({ C'_n,tL 0 < N < N,0 < T < T } is transmitted to the analysis host computer of the monitoring terminal, wherein C'_n,tIs represented by C_n,tA corresponding amplified vibration optical signal;

(2a3) the monitoring terminal analyzes the received signal by the host computerAnd filtering the filtered oscillatory optical signal S ═ S_n,tL 0 < N < N,0 < T < T) is drawn into a railway abnormal intrusion behavior signal data table with dimension of T multiplied by N, and then the table is sent to a data processing submodule, wherein S_n,tIs C'_n,tFiltering results;

(2b) the data processing submodule converts each filtered vibration optical signal S in the table_n,tNormalization is performed simultaneously along the dimension T and the dimension N to obtain a normalized vibration optical signal S '═ S'_n,tAnd (2) a railway abnormal intrusion behavior signal data table with the value of I0 < N < N, and 0 < T < T), uniformly dividing S' into Q groups according to the dimension of T, then drawing a normalized space-time signal image of each group of vibration optical signals by taking the time T of each group of normalized vibration optical signals as a vertical coordinate, taking a sampling point N as a horizontal coordinate, and expressing the size of the normalized vibration optical signal value by different colors to obtain a data set containing Q normalized space-time signal images, wherein Q is more than or equal to 2000.

In this example, the signals S' are grouped into 6000 time instances to obtain a data set containing 3000 normalized spatio-temporal signal images, Q3000. The space-time image of five railway abnormal intrusion behavior signals is shown in fig. 2, fig. 2(a) shows the space-time image of a train background noise interference intrusion behavior signal, fig. 2(b) shows the space-time image of a shaking fence cage intrusion behavior signal, fig. 2(c) shows the space-time image of a digging intrusion behavior signal, fig. 2(d) shows the space-time image of a climbing fence intrusion behavior signal, and fig. 2(e) shows the space-time image of a wall-cutting intrusion behavior signal. The image fully combines the time information and the space information of the signals, the ordinate represents the time of each group of normalized vibration optical signals S', and the time dependence of the signals can be analyzed; the abscissa represents the sampling point of each group of normalized vibration optical signals S', which is beneficial to extracting the spatial characteristics of the signals; each point in the image represents the magnitude of the normalized vibro-optical signal value. The space-time images of the five abnormal intrusion behaviors are obviously different, and the interference of background noise signals can be distinguished;

different evaluation indexes are different in dimension anddimension units, thereby affecting the final detection result, and in order to eliminate the dimension influence among the indexes, normalization processing is required to solve the comparability among the data indexes. In this example, a linear normalization process is used, and each filtered vibro-optical signal S in the table is processed_n,tLinear normalization is carried out along the dimension T and the dimension N simultaneously, so that the convergence speed of the model can be further improved;

(2c) labeling abnormal intrusion behavior signal regions in each normalized space-time signal image, and randomly selecting M of Q normalized space-time signal images containing abnormal intrusion behavior signal region labeling frames as training sample sets V ═ V₁,V₂,...,V_m,...,V_MV is Z ═ Z'₁,Z'₂,...,Z'_m,...,Z'_MAnd taking the rest Q-M normalized spatio-temporal signal images containing the abnormal intrusion behavior signal region labeling frame as a test sample set F ═ F₁,F₂,...,F_w,...,F_WAnd (c) the step of (c) in which,V_mrepresenting the m-th training sample with a label box, Z'_mRepresents V_mCorresponding authentic label, F_wDenotes the W-th test sample with label box, W-Q-M, in this example, M-2000;

step 3) constructing a railway intrusion behavior detection network model fast R-CNN, wherein the structure diagram is shown in FIG. 3:

(3a) constructing a railway intrusion behavior detection network model, namely, a fast R-CNN, which comprises a cascaded feature extraction network, a candidate region generation network RPN, an ROI pooling layer and a classification regression network, wherein the feature extraction network adopts a VGG16 convolutional neural network comprising 13 convolutional layers, 13 activation function layers and 4 pooling layers; the RPN comprises a cascade shared convolution layer, an activation function layer and a candidate region recommendation layer, wherein the activation function layer and the candidate region recommendation layer are directly loaded with SoftMax classification branches B which are arranged in parallel₁And bounding Box regression convolution layer B₂(ii) a Classification regression networkComprising two fully-connected layers F₁And with F₁Parallel connected SoftMax sorted fully connected layer F₂And regression full-junction layer F₃；

In this embodiment, the RPN generates the network RPN by drawing nine preset frames for each pixel feature point on the feature map, accurately predicts the position of the candidate frame in the region of the feature map, and performs convolution on each feature map including the accurate position of the candidate frame in the region by using the classification regression network, thereby effectively extracting the feature information of the image and ensuring the detection accuracy.

(3b) Constructing a classification loss function L_clsAnd a regression loss function L_regLoss function L of composed candidate region generation network RPN_RPN：

Wherein p is_iIndicates the i-th preset-box prediction label,represents the real label, u, corresponding to the ith preset box_i＝{u_x,u_y,u_w,u_hDenotes the predicted bounding box regression parameter vector for the ith preset box,representing the real edge corresponding to the ith preset frameBounding box regression parameter vector, N_clsRepresenting the total number of samples, N, in a minimum batch value_regThe number of pixels of each feature map is represented, λ represents a balance weight parameter, and in this embodiment, N_cls＝256，N_reg＝2400，λ＝10；

In the present embodiment, the regression loss function L_regUse ofThe function can not cause unstable training due to the gradient of the predicted value, so that the gradient of the loss function near the true value is smoother, the predicted value is easier to stably converge, and the training speed and performance of the model are improved.

(4) Performing iterative training on a railway intrusion behavior detection network model, namely, fast R-CNN:

(4a) initializing the iteration number as K, the maximum iteration number as K, and setting the kth iterative railway intrusion behavior detection network model Faster R-CNN as H^kAnd let k equal to 1, H^kH, K in this example is 70000;

(4b) taking the training sample set V as a railway intrusion behavior detection network model H^kFor each training sample V, the feature extraction network_mPerforming abnormal behavior signal feature extraction to obtain M feature maps, where M is 2000 in this embodiment;

(4c) the candidate region generation network RPN predicts the accurate position of the abnormal intrusion behavior region candidate frame of each feature map, and the steps are as follows:

(4c1) the shared convolution layer extracts a plurality of pixel feature points of each feature map and plots each pixel feature point on the feature map with a size of 128²、256²、512²Nine preset frames with the length-width ratio of 2:1, 1:1 and 1:2, and outputting M characteristic diagrams each comprising a plurality of preset frames;

(4c2) the activation function layer carries out nonlinear mapping on pixel feature points in each feature graph output by the shared convolution layer to obtain M feature graphs after the nonlinear mapping, wherein the formula for carrying out the nonlinear mapping on each pixel feature point is as follows:

RELU(c)＝max(0,c)

wherein c is the pixel characteristic point in each characteristic graph output by the shared convolutional layer;

(4c3) SoftMax sorting leg B₁After convolution is carried out on each feature graph output by the activation function layer, the convolution results are classified, and the prediction label p of the ith preset frame in each feature graph is output_iWhile bounding box regresses convolutional layer B₂Regression is carried out on the convolution result of each feature graph output by the activation function layer, and a border regression parameter vector u of the ith preset frame in each feature graph is output_i＝{u_x,u_y,u_w,u_hIn which u_x、u_y、u_w、u_hRespectively representing the central abscissa, the central ordinate, the width and the height of the ith preset frame;

(4c4) selecting a branch B between the real label set Z and SoftMax classification in the candidate area recommendation layer₁Outputting a predictive label p of the ith preset box in each feature map_iThe same preset frame is used as a region candidate frame of abnormal intrusion behavior, and then the convolution layer B is regressed through a boundary frame₂Outputting a regression parameter vector u of the prediction boundary box of the ith preset box in each feature map_iAnd (4) performing regression on the accurate position of the area candidate frame to obtain a characteristic diagram containing the prediction of the accurate position of the area candidate frame of the abnormal intrusion behavior.

(4d) The ROI pooling layer performs pooling operation on each feature map output by the candidate region generation network RPN to obtain M feature maps with the same size, in the embodiment, the ROI pooling layer adopts a maximum pooling processing method, and the sizes of the output feature maps are 7 multiplied by 7;

(4e) two fully-connected layers F in a classification regression network₁Performing convolution operation on each characteristic graph, and classifying the full-connected layer F by SoftMax₂Classifying the convolution result of each feature map to obtain a prediction label set Z ═ Z corresponding to the training sample set V₁,Z₂,...,Z_m,...,Z_MGet back to the full connection layer F at the same time₃Convolution of each feature mapAnd (4) regression is carried out on the result to obtain a predicted position offset vector set A ═ A corresponding to the training sample set V₁,A₂,...,A_m,...,A_MIn which Z is_mAnd A_mRespectively represent V_mA corresponding prediction tag and prediction position offset vector;

(4f) using a classification loss function L_clsCalculating each predictive label Z_mAnd its corresponding real tag Z'_mThen carrying out mean value processing on the M loss values to obtain H after the kth iteration^kClassification loss value ofAnd using a back propagation algorithm based onClassifying fully-connected layers F for SoftMax in RPN (resilient packet network)₂Parameter (d) ofUpdating, wherein the updating formula is as follows:

wherein eta represents the learning step length, 0 < eta is less than or equal to 0.1,to representAs a result of the update, the result of the update,it is indicated that the operation of derivation is performed,the partial derivative calculation is shown, and in this example, η is 0.001.

(4g) Using a regression loss function L_regCalculating each predicted position offset vector A_mTrue position vector A 'corresponding thereto'_mThen carrying out mean value processing on the M loss values to obtain H after the kth iteration^kValue of regression loss ofAnd using a back propagation algorithm based onGenerating regression full connectivity layer F in network RPN for region₃Parameter (d) ofUpdating to obtain H after the kth iteration^kThe update formula is:

wherein eta represents the learning step length, 0 < eta is less than or equal to 0.1,respectively representAs a result of the update, the result of the update,it is indicated that the operation of derivation is performed,the partial derivative calculation is shown, and in this example, η is 0.001.

(4h) Judging whether K is true or not, if so, obtaining a trained railway intrusion behavior detection network model H^*Otherwise, let k equal to k +1, and execute step (4 b);

step 5) obtaining a detection result of the abnormal intrusion behavior of the railway:

(5a) set of test samples F asTrained railway intrusion behavior detection network model H^*Performing abnormal behavior signal feature extraction on each test sample by a feature extraction network to obtain a feature map of W images;

(5b) predicting the accurate position of the candidate frame of the abnormal intrusion behavior region of each feature map by the candidate region generation network RPN to obtain W feature maps containing the accurate position prediction of the candidate frame of the abnormal intrusion behavior region;

(5c) the ROI pooling layer performs pooling operation on each feature map output by the candidate area generation network to obtain W feature maps with the same size;

(5d) two fully-connected layers F in a classification regression network₁Performing convolution operation on each characteristic graph, and classifying the full-connected layer F by SoftMax₂Classifying the convolution result of each characteristic graph to obtain a prediction label set corresponding to the test sample set FSimultaneous regression full junction layer F₃Regression is carried out on the convolution result of each characteristic graph to obtain a prediction position offset vector set corresponding to the test sample set FWherein the content of the first and second substances,andrespectively represent F_wA corresponding prediction tag and a predicted position offset vector.

The effect of the present invention is further explained by combining the simulation experiment as follows:

1. simulation conditions and contents:

the hardware test platform adopted by the simulation experiment is as follows: the GPU is NVIDIA GeForce GTX 1060, the main frequency is 3.00GHz, and the memory is 16 GB; the software platform is as follows: ubuntu18.04 operating system and python 3.5.0.

The railway abnormal intrusion behavior image data sets used in the simulation experiment are real data sets, field acquisition is carried out on a railway system, and signal change near a fence can be detected by monitoring vibration detection optical cables laid on the railway fence, so that the behavior of threatening railway operation is detected. The total data of the abnormal intrusion behavior signal images is five kinds of abnormal intrusion behavior signal image data, including noise interference of train passing background, fence shaking and cage pricking, fence climbing, digging and wall chiseling. The size of the dataset image is 875 × 656, and the image format is jpg. And randomly selecting 2000 normalized space-time signal images in a normalized space-time signal image data set containing 3000 abnormal intrusion behavior signal area labeling boxes as a training set, and taking the other 1000 normalized space-time signal images as a test set.

The false alarm rate FAR and the detection precision TDR, F of the invention and the existing railway intrusion behavior detection method based on convolution long-short term memory artificial neural network (ConvLSTM) and optical fiber distributed acoustic sensing₁The scores were compared and simulated, and the results are shown in table 1.

2. And (3) simulation result analysis:

in order to quantify the detection results, the following 3 evaluation indexes were used in the simulation experiment.

(1) False Alarm rate far (false Alarm rate): the ratio of prediction error in all results of which the prediction is a positive sample is shown, and the value is between 0 and 100 percent, and the smaller the value is, the better the detection effect is.

(2) Detection accuracy tdr (thread Detect rate): the ratio of the samples which are truly predicted to be the positive samples in all the positive samples is represented, the value is between 0% and 100%, and the larger the value is, the better the detection effect is.

(3)F₁And (3) fractional: the detection accuracy and the recall rate are high as the index is higher, the value is between 0 and 100 percent, and the detection effect is better when the value is higher.

TABLE 1

In combination with Table 1As can be seen, the detection results of the present invention are in FAR, TDR, or F₁The method is obviously superior to the prior art in score, and meanwhile, the method can distinguish train background noise interference behaviors, reduce the false alarm rate and further solve the technical problem of low detection accuracy rate.

The foregoing description is only an example of the present invention and is not intended to limit the invention, so that it will be apparent to those skilled in the art that various changes and modifications in form and detail may be made therein without departing from the spirit and scope of the invention.