CN117609885A - High-speed rail intrusion monitoring method based on distributed optical fiber sensing and fused neural network - Google Patents

High-speed rail intrusion monitoring method based on distributed optical fiber sensing and fused neural network Download PDF

Info

Publication number
CN117609885A
CN117609885A CN202311740490.4A CN202311740490A CN117609885A CN 117609885 A CN117609885 A CN 117609885A CN 202311740490 A CN202311740490 A CN 202311740490A CN 117609885 A CN117609885 A CN 117609885A
Authority
CN
China
Prior art keywords
layer
disturbance
speed rail
neural network
data
Prior art date
Legal status (The legal status is an assumption and is not a legal conclusion. Google has not performed a legal analysis and makes no representation as to the accuracy of the status listed.)
Pending
Application number
CN202311740490.4A
Other languages
Chinese (zh)
Inventor
张益昕
姚媛媛
贾瑞铎
潘昱琦
张旭苹
邹宁睦
王峰
Current Assignee (The listed assignees may be inaccurate. Google has not performed a legal analysis and makes no representation or warranty as to the accuracy of the list.)
Nanjing University
Original Assignee
Nanjing University
Priority date (The priority date is an assumption and is not a legal conclusion. Google has not performed a legal analysis and makes no representation as to the accuracy of the date listed.)
Filing date
Publication date
Application filed by Nanjing University filed Critical Nanjing University
Priority to CN202311740490.4A priority Critical patent/CN117609885A/en
Publication of CN117609885A publication Critical patent/CN117609885A/en
Pending legal-status Critical Current

Links

Classifications

    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06F—ELECTRIC DIGITAL DATA PROCESSING
    • G06F18/00—Pattern recognition
    • G06F18/20—Analysing
    • G06F18/24—Classification techniques
    • G06F18/241—Classification techniques relating to the classification model, e.g. parametric or non-parametric approaches
    • G06F18/2415—Classification techniques relating to the classification model, e.g. parametric or non-parametric approaches based on parametric or probabilistic models, e.g. based on likelihood ratio or false acceptance rate versus a false rejection rate
    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06F—ELECTRIC DIGITAL DATA PROCESSING
    • G06F18/00—Pattern recognition
    • G06F18/10—Pre-processing; Data cleansing
    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06F—ELECTRIC DIGITAL DATA PROCESSING
    • G06F18/00—Pattern recognition
    • G06F18/20—Analysing
    • G06F18/21—Design or setup of recognition systems or techniques; Extraction of features in feature space; Blind source separation
    • G06F18/213—Feature extraction, e.g. by transforming the feature space; Summarisation; Mappings, e.g. subspace methods
    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06F—ELECTRIC DIGITAL DATA PROCESSING
    • G06F18/00—Pattern recognition
    • G06F18/20—Analysing
    • G06F18/21—Design or setup of recognition systems or techniques; Extraction of features in feature space; Blind source separation
    • G06F18/214—Generating training patterns; Bootstrap methods, e.g. bagging or boosting
    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06F—ELECTRIC DIGITAL DATA PROCESSING
    • G06F18/00—Pattern recognition
    • G06F18/20—Analysing
    • G06F18/25—Fusion techniques
    • G06F18/253—Fusion techniques of extracted features
    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
    • G06N3/00—Computing arrangements based on biological models
    • G06N3/02—Neural networks
    • G06N3/04—Architecture, e.g. interconnection topology
    • G06N3/044—Recurrent networks, e.g. Hopfield networks
    • G06N3/0442—Recurrent networks, e.g. Hopfield networks characterised by memory or gating, e.g. long short-term memory [LSTM] or gated recurrent units [GRU]
    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
    • G06N3/00—Computing arrangements based on biological models
    • G06N3/02—Neural networks
    • G06N3/04—Architecture, e.g. interconnection topology
    • G06N3/045—Combinations of networks
    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
    • G06N3/00—Computing arrangements based on biological models
    • G06N3/02—Neural networks
    • G06N3/04—Architecture, e.g. interconnection topology
    • G06N3/0464—Convolutional networks [CNN, ConvNet]
    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
    • G06N3/00—Computing arrangements based on biological models
    • G06N3/02—Neural networks
    • G06N3/08—Learning methods
    • G06N3/082—Learning methods modifying the architecture, e.g. adding, deleting or silencing nodes or connections
    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06F—ELECTRIC DIGITAL DATA PROCESSING
    • G06F2218/00—Aspects of pattern recognition specially adapted for signal processing
    • G06F2218/02—Preprocessing
    • G06F2218/04—Denoising
    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06F—ELECTRIC DIGITAL DATA PROCESSING
    • G06F2218/00—Aspects of pattern recognition specially adapted for signal processing
    • G06F2218/08—Feature extraction
    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06F—ELECTRIC DIGITAL DATA PROCESSING
    • G06F2218/00—Aspects of pattern recognition specially adapted for signal processing
    • G06F2218/12—Classification; Matching

Landscapes

  • Engineering & Computer Science (AREA)
  • Theoretical Computer Science (AREA)
  • Physics & Mathematics (AREA)
  • Data Mining & Analysis (AREA)
  • Evolutionary Computation (AREA)
  • Life Sciences & Earth Sciences (AREA)
  • Artificial Intelligence (AREA)
  • General Physics & Mathematics (AREA)
  • General Engineering & Computer Science (AREA)
  • Computer Vision & Pattern Recognition (AREA)
  • Bioinformatics & Cheminformatics (AREA)
  • Evolutionary Biology (AREA)
  • Bioinformatics & Computational Biology (AREA)
  • Computational Linguistics (AREA)
  • Health & Medical Sciences (AREA)
  • Biomedical Technology (AREA)
  • Biophysics (AREA)
  • General Health & Medical Sciences (AREA)
  • Molecular Biology (AREA)
  • Computing Systems (AREA)
  • Mathematical Physics (AREA)
  • Software Systems (AREA)
  • Probability & Statistics with Applications (AREA)
  • Image Analysis (AREA)

Abstract

The invention discloses a high-speed rail intrusion monitoring method based on distributed optical fiber sensing and a fused neural network, which comprises the following steps: collecting disturbance signals of the perimeter of the high-speed rail by using a distributed optical fiber acoustic wave sensing system, and dividing the collected disturbance signals to construct a sample data set containing various disturbance type labels; removing background noise contained in the sample data; the method comprises the steps of constructing a high-speed rail intrusion detection model, wherein the high-speed rail intrusion detection model comprises a convolutional neural network, a two-way long-short-term memory neural network, a splicing module, a RepMLP module and a softmax layer, and predicting the occurrence probability of each disturbance category corresponding to a disturbance signal. The invention realizes the long-distance, high-precision and low-time-consuming automatic abnormal disturbance monitoring of the high-speed rail perimeter.

Description

High-speed rail intrusion monitoring method based on distributed optical fiber sensing and fused neural network
Technical Field
The invention belongs to the technical field of perimeter monitoring along a high-speed railway, and particularly relates to a high-speed railway intrusion monitoring method based on distributed optical fiber sensing and a fused neural network.
Background
The high-speed railway generally has the characteristics of long mileage along the line, complex geology and the like, and the characteristics can cause sudden disasters such as falling rocks, debris flows and the like, thereby seriously threatening the running safety of the high-speed railway. Therefore, the safety of the high-speed railway is important, a real-time monitoring system for the perimeter safety of the track along the track is established, and the emergency is warned.
As an important branch of a distributed optical fiber sensing technology (DOFS), the DAS can realize long-distance, distributed and real-time quantitative monitoring of dynamic vibration of an optical fiber along a line, so that the DAS is very suitable for perimeter security monitoring of the high-speed rail along the line. The optical fiber sensing monitoring data not only comprises time sequence data of disturbance, but also comprises space sequence data of disturbance positions, and rich information is provided for subsequent analysis. In order to extract effective information from mass data, students at home and abroad try to combine a machine learning algorithm with a distributed optical fiber sensing technology, and after some characteristics are artificially defined, an appropriate machine learning algorithm, such as a related vector machine, a decision tree, a random forest and the like, is selected to obtain a final classification result. Although the methods can achieve higher recognition rate by manually extracting features, the methods have the defects of stronger pertinence and weaker generalization capability.
With the rapid development of the deep learning algorithm, the neural network structure is more complex and various, and the deep features of the signals can be automatically extracted, so that the manual intervention is reduced, and the method is more suitable for wider and changeable environments. CNN is a type of feedforward neural network consisting of a data input layer, a convolution layer, an excitation function layer, a pooling layer and a full-connection layer, and is one of the representative algorithms of deep learning. The CNN does not need to manually design a feature extractor, and the convolution kernel parameter sharing and the sparsity of interlayer connection in the hidden layer enable the convolution neural network to perform latticed features with smaller calculated amount, and the model performance can be improved by adjusting the layer number and parameters. However, CNN cannot extract time series features, so that an overfitting phenomenon is easy to generate, and the CNN is relatively sensitive to signal noise, and an additional technology is required to improve the robustness of the model. Compared with CNN, LSTM is a special cyclic neural network, which is designed to solve the long-term dependence problem of the common cyclic neural network, wherein a single cyclic unit has multiple gates, including an input gate, an output gate, a forgetting gate, a hidden feature and the like. Due to the unique design structure, LSTM is suitable for processing and analyzing correlations in significant events that are relatively long-spaced and delayed in time series. To fully take into account the forward and backward dependency information at a particular moment, we use the bi-directional LSTM (BiLSTM) model. However, biLSTM ignores spatial information in the data, and when processing longer sequence data, problems such as gradient extinction or gradient explosion may be encountered, and some techniques need to be adopted to deal with these problems. Therefore, when the DAS is used to monitor the perimeter of the high-speed rail, a single neural network model cannot effectively process the monitored data, and a new method is required to be provided to integrate the advantages of different neural networks.
In other fields, researchers have adopted a way to sub-classify the temporal and spatial features extracted by CNN and LSTM respectively. However, the established model cannot be directly migrated to the high-speed railway intrusion monitoring problem, because the high-speed railway intrusion data has the characteristics of fewer disturbance samples, shorter duration of disturbance events, smaller action range, multiple disturbance types and the like, if no network structure is further optimized, the defects of poor robustness, unsatisfactory precision, even indistinguishability and the like can be caused because the established model cannot be used for training the high-speed railway intrusion data aiming at the local characteristics of different disturbance types.
Therefore, the end-to-end fusion network structure suitable for processing the high-speed rail signal data is continuously designed at present, so that the feature extraction capacity and the anti-interference capacity of a model can be effectively improved, and the system can maintain higher accuracy and second-level reaction speed under a relatively complex environment.
Disclosure of Invention
The technical problems to be solved are as follows: the invention discloses a high-speed rail intrusion monitoring method based on a distributed optical fiber sensing and fusion neural network, which utilizes an existing optical cable of a railway and a distributed optical fiber acoustic wave sensing system (DAS) to realize real-time monitoring of sudden accidents along the high-speed rail and illegal intrusion related signals, utilizes neural network classification and identification technologies such as a front Convolutional Neural Network (CNN), a bidirectional long-short-term memory neural network (BiLSTM), a re-parameterized multi-layer perceptron (RepMLP) and the like, identifies specific intrusion types aiming at common intrusion events of the high-speed rail, and sends information such as accident positions and the like to related monitoring staff, thereby realizing long-distance, high-precision and low-time-consuming automatic monitoring of the perimeter of the high-speed rail.
The technical scheme is as follows:
a high-speed rail intrusion monitoring method based on distributed optical fiber sensing and fusion neural network comprises the following steps:
s1, collecting disturbance signals of the perimeter of a high-speed rail by using a distributed optical fiber acoustic wave sensing system, dividing the collected disturbance signals into a plurality of sampling points within a 1S adjacent space range, dividing the data into one sample data, and constructing a sample data set containing a plurality of disturbance type labels;
s2, aiming at background noise characteristics, removing background noise contained in sample data by using a heuristic threshold wavelet noise reduction method;
s3, constructing a high-speed rail intrusion detection model, wherein the high-speed rail intrusion detection model comprises a convolutional neural network, a two-way long-short-term memory neural network, a splicing module, a RepMLP module and a softmax layer, and training and verifying the high-speed rail intrusion detection model by adopting a sample data set; the convolutional neural network and the two-way long-short-term memory neural network are arranged in parallel, and feature extraction is carried out on sample data with background noise removed respectively to obtain spatial features and temporal features of the sample data: the spatial characteristics of the sample data are used for representing the change and the connection of disturbance signals in a window between adjacent sampling points, and the temporal characteristics of the sample data are used for representing the change behavior of the disturbance signals in the same window along with time; the splicing module splices the spatial characteristics and the temporal characteristics of the sample data, and inputs the space-time characteristic vector of the obtained disturbance signal to the RepMLP module; the RepMLP module utilizes the local capturing characteristic of the convolution layer to represent the geometric detail information of the space-time feature vector, and extracts the instantaneous disturbance features of disturbance signals of different categories; the softmax layer classifies the extracted instantaneous disturbance characteristics of disturbance signals of different categories, and predicts the occurrence probability of each disturbance category corresponding to the disturbance signal;
s4, acquiring acoustic wave signals of the perimeter of the high-speed rail by using a distributed optical fiber acoustic wave sensing system, carrying out segmentation processing and denoising processing on the acquired acoustic wave signals, then introducing the acoustic wave signals into a high-speed rail intrusion detection model, extracting instantaneous disturbance characteristics of the acoustic wave signals at different sampling point positions, outputting the occurrence probability of each disturbance category corresponding to the acoustic wave signals, and outputting the type with the highest occurrence probability as a classification result of the sampling point at the current moment.
Further, in step S1, the acquired disturbance signal is subjected to segmentation processing, and data of 5 sampling points in the adjacent spatial range within 1S are divided into one sample data.
Further, the disturbance type tag comprises seven types of climbing, knocking, digging, electric drill, foreign matter invasion, illegal intrusion and abnormal weather influence.
Further, in step S2, the process of removing the background noise included in the sample data by using the heuristic threshold wavelet noise reduction method includes the following steps:
s21, performing three-layer wavelet decomposition calculation on the disturbance signal by adopting a coif5 wavelet function;
s22, setting a fixed threshold value and an unbiased threshold value;
s23, quantifying a threshold value of the wavelet decomposition high-frequency coefficient, and adjusting according to the noise level estimation of the first-layer wavelet decomposition to remove the wavelet coefficient belonging to the noise; when the signal-to-noise ratio is larger than a preset signal-to-noise ratio threshold, adopting a fixed threshold to perform noise reduction processing, otherwise adopting an unbiased threshold to perform noise reduction processing
S24, reconstructing through wavelet inverse transformation to obtain a denoised disturbance signal.
Further, in step S2, the convolutional neural network includes a first two-dimensional convolutional layer, a second two-dimensional convolutional layer, a third two-dimensional convolutional layer, a pooling layer, a flattening layer, a first full-connection layer, a second full-connection layer, and a third full-connection layer that are sequentially connected;
the convolution kernel sizes of the first two-dimensional convolution layer, the second two-dimensional convolution layer and the third two-dimensional convolution layer are 3 multiplied by 3, the step length is 1, and the padding is 1; for input data with the size of 5×100, three two-dimensional convolution layers increase the number of channels from 1 to 64, and maintain the two-dimensional size of the data unchanged, and still be 5×100; the Pooling layer is used for carrying out two-dimensional Max Pooling operation with the size of 2 multiplied by 2 on the output of the third two-dimensional convolution layer, and reducing the data size of each channel to 2 multiplied by 50; the flattening layer is used for flattening the output of the pooling layer to obtain a one-dimensional vector with the length of 6400; the first full connection layer, the second full connection layer and the third full connection layer sequentially process the one-dimensional vector and map the one-dimensional vector to a 64-dimensional space feature vector.
Further, a two-dimensional regularization layer and a ReLu activation function layer are respectively connected behind the two-dimensional convolution layer; the two-dimensional regularization layer is used for carrying out batch normalization operation on the output to reduce adverse effects caused by data distribution migration, and the ReLu activation function layer is used for introducing nonlinear fitting capacity;
and a dropout layer is connected between the first full-connection layer and the second full-connection layer, and the dropout layer is used for improving generalization capability so as to avoid fitting input data.
Further, the two-way long-short-term memory neural network comprises two layers of LSTM networks stacked with each other, wherein the input size of each layer of LSTM network is 5, and the hidden layer size is 16; the second layer LSTM network takes the output of the first layer LSTM network as input, and each layer LSTM network inputs the sequence data according to the forward direction and the reverse direction and obtains the output of the two directions;
input data with the size of 5 multiplied by 100 is input into a first layer LSTM network, and finally two layers of bidirectional hidden layer outputs with the size of 16 are obtained, and the time feature vectors with the length of 64 are obtained by splicing.
Further, the RepMLP module comprises a fourth full connection layer, a convolution layer, a regularization layer and a fifth full connection layer which are sequentially connected;
the RepMLP module processes the space-time characteristic vector of the disturbance signal, and inputs the obtained logic vector with the length of 7 into the softmax layer.
Further, in step S4, the occurrence probability of each disturbance category corresponding to the output acoustic wave signal is determined, and if the occurrence probability of all the types is lower than a preset probability threshold, the non-destructive disturbance is determined; if the occurrence probability of at least one type is higher than the preset probability threshold, outputting the type with the highest occurrence probability as a classification result of the sampling point at the current moment, and determining the position of the sampling point as a disturbance point.
The beneficial effects are that:
first, the high-speed rail intrusion monitoring method based on the distributed optical fiber sensing and the fusion neural network directly utilizes the existing optical cable of the railway to realize the real-time monitoring of the high-speed rail along the line, does not need to additionally increase equipment, and reduces the monitoring cost and subsequent maintenance cost.
Secondly, according to the high-speed rail intrusion monitoring method based on the distributed optical fiber sensing and the fusion neural network, according to the characteristics of the high-speed rail environment, the heuristic threshold wavelet method is selected to reduce the noise of the data, the self-adaptive noise reduction processing can be carried out according to the characteristics of the signals, the influence of the background noise on the classification result in the complex environment is eliminated, and the overall signal-to-noise ratio of the data is improved.
Thirdly, the high-speed rail intrusion monitoring method based on the distributed optical fiber sensing and the fused neural network utilizes the fused neural network model to automatically extract and analyze the DAS signals after segmentation, fully utilizes the global characteristics of the signals, enhances the local capturing characteristics of the model by adding the RepMLP module to replace a single full-connection layer, is beneficial to accurately distinguishing different disturbance types, improves the classification precision, is more reasonable than a single CNN or LSTM model, has the advantages of high efficiency, comprehensiveness, strong robustness and the like, and is more suitable for practical high-speed rail perimeter security and protection application.
Drawings
FIG. 1 is a flow chart of a method for monitoring high-speed rail intrusion based on distributed optical fiber sensing and a fused neural network;
FIG. 2 is a graph showing the comparison of noise reduction effects; (a) Is an original signal diagram, and (b) is a denoised signal diagram.
Detailed Description
The following examples will provide those skilled in the art with a more complete understanding of the invention, but are not intended to limit the invention in any way.
The embodiment of the invention discloses a high-speed rail intrusion monitoring method based on distributed optical fiber sensing and a fused neural network, which comprises the following steps:
s1, collecting disturbance signals of the perimeter of a high-speed rail by using a distributed optical fiber acoustic wave sensing system, dividing the collected disturbance signals into a plurality of sampling points within a 1S adjacent space range, dividing the data into one sample data, and constructing a sample data set containing a plurality of disturbance type labels;
s2, aiming at background noise characteristics, removing background noise contained in sample data by using a heuristic threshold wavelet noise reduction method;
s3, constructing a high-speed rail intrusion detection model, wherein the high-speed rail intrusion detection model comprises a convolutional neural network, a two-way long-short-term memory neural network, a splicing module, a RepMLP module and a softmax layer, and training and verifying the high-speed rail intrusion detection model by adopting a sample data set; the convolutional neural network and the two-way long-short-term memory neural network are arranged in parallel, and feature extraction is respectively carried out on sample data with background noise removed, so that spatial features and temporal features of the sample data are obtained; the spatial characteristics of the sample data are used for representing the change and the connection of disturbance signals in a window between adjacent sampling points, and the temporal characteristics of the sample data are used for representing the change behavior of the disturbance signals in the same window along with time; the splicing module splices the spatial characteristics and the temporal characteristics of the sample data, and inputs the space-time characteristic vector of the obtained disturbance signal to the RepMLP module; the RepMLP module utilizes the local capturing characteristic of the convolution layer to represent the geometric detail information of the space-time feature vector, and extracts the instantaneous disturbance features of disturbance signals of different categories; the softmax layer classifies the extracted instantaneous disturbance characteristics of disturbance signals of different categories, and predicts the occurrence probability of each disturbance category corresponding to the disturbance signal;
s4, acquiring acoustic wave signals of the perimeter of the high-speed rail by using a distributed optical fiber acoustic wave sensing system, carrying out segmentation processing and denoising processing on the acquired acoustic wave signals, then introducing the acoustic wave signals into a high-speed rail intrusion detection model, extracting instantaneous disturbance characteristics of the acoustic wave signals at different sampling point positions, outputting the occurrence probability of each disturbance category corresponding to the acoustic wave signals, and outputting the type with the highest occurrence probability as a classification result of the sampling point at the current moment.
In the embodiment, firstly, a DAS system is utilized to collect disturbance signals of the perimeter of the high-speed rail, a data set containing seven classification labels is constructed, and the data set is divided into a test set and a verification set; secondly, aiming at the background noise characteristics, a heuristic threshold wavelet denoising method is utilized to denoise the background noise characteristics; and thirdly, fusing the CNN and the BiLSTM to establish a deep neural network, extracting data features and association with front and rear time periods respectively by utilizing the CNN and the BiLSTM in parallel, splicing the results of the CNN and the BiLSTM to obtain a fused feature vector, outputting occurrence probability of each type by utilizing the RepMLP finally, and determining the occurrence position of disturbance according to the classification result of each sampling point. Based on the model, the actual collected data is used for testing, network parameters are adjusted according to a random small-batch gradient descent method, a residual network model is added, the parameter setting of RepMLP is adjusted, model parameters are optimized until the model converges, and finally the high-speed rail intrusion detection model with high accuracy and low misjudgment rate is obtained.
Referring to fig. 1, the method for monitoring the intrusion of the high-speed rail comprises the following specific steps:
1. model training stage:
step 1, sample data can be acquired from an actual scene or an experimental scene. The invention uses the existing optical cable of the railway as a sensing medium, uses DAS to realize long-distance, distributed and real-time quantitative monitoring of perimeter disturbance, does not need to additionally increase equipment, and reduces the monitoring cost and subsequent maintenance cost.
In consideration of the fact that the disturbance data are less in the running process of the actual high-speed rail, the disturbance data in a part of experimental scenes are introduced in the embodiment. Under an experimental scene, seven disturbance types common to the perimeter of the high-speed rail are fully simulated, and data are acquired in real time by utilizing a DAS system. Under the simulation environment, the optical fibers are distributed on the perimeter guard rail of the experimental environment in an S shape and buried underground near the perimeter, so that the full coverage of the perimeter range is realized. The DAS system is connected with one end of the optical fiber, so that the normal operation of the signal collecting and transmitting parts of the DAS system is ensured.
And (3) manually performing a plurality of disturbance experiments around the laid optical fibers, wherein the disturbance experiments comprise: the method comprises the steps of (1) climbing, (2) knocking, (3) excavating, (4) electric drill, (5) foreign matter invasion, (6) illegal intrusion, (7) noise of abnormal weather and the like, and adding the seven classification labels according to specific types of signals collected by the optical fiber. The data of 5 sampling points in the adjacent space range within 1s are divided into a basic event window, namely one sample, a signal data set containing seven disturbance categories is constructed, a training set and a verification set are divided, and samples of each category are contained as much as possible. The corresponding two-dimensional data size is 5 x 100, 100 samples are collected for each type, and the training set and the verification set are divided according to the ratio of 7:3.
In an actual high-speed rail monitoring scenario, the disturbance event is a non-transient event, but the duration is short, usually in the range of seconds, so that besides the position where the disturbance event occurs, a time dimension feature is also required to be combined. Based on the principle, the embodiment performs segmentation processing on the acquired samples, divides data of 5 sampling points in a 1s adjacent spatial range into sample data, and processes the sample data through a fusion type neural network so as to accurately express the spatial and temporal characteristics of high-speed rail disturbance.
And 2, carrying out wavelet denoising on the constructed data set. In the specific implementation process of noise reduction, wavelet noise reduction adopts heuristic threshold values to carry out three-level wavelet transformation, and adjustment is carried out according to noise level estimation of first-layer wavelet decomposition. In a high-speed rail intrusion monitoring scene, a large amount of background noise exists in data due to the influence of coupling effect and other factors, namely, tiny disturbance irrelevant to the overall characteristics of signals is commonly existing in a time dimension, the amplitude of the disturbance is random and changeable, and the frequency domain characteristics of the disturbance are mainly expressed as 1-10Hz, so that the disturbance is submerged in the noise, and the recognition accuracy is influenced. Aiming at the noise problem, a wavelet change noise reduction method based on a heuristic threshold is adopted to preprocess the data set, the threshold is adjusted according to the signal characteristics, and the negative influence of background noise is reduced in a self-adaptive mode. And carrying out multiple lower-scale decomposition on the data by adopting a wavelet function, estimating a self-adaptive adjustment threshold according to the noise level of the first-layer wavelet decomposition, and then reconstructing by using estimated wavelet coefficients on each scale to obtain a denoised signal, thereby improving the overall signal-to-noise ratio of the data and eliminating the influence of background noise on a classification result in a complex environment.
The specific steps of wavelet noise reduction are as follows: 1. wavelet decomposition of the signal, adopting a coif5 wavelet function to perform three-layer wavelet decomposition calculation on the signal; 2. and selecting a proper threshold, wherein the formula is shown in (2), when the signal-to-noise ratio is large, a fixed threshold is selected, otherwise, an unbiased threshold is selected, so that the noise reduction method is more suitable for noise reduction of actual data. 3. Performing threshold quantization of the wavelet decomposition high-frequency coefficient, and adjusting according to the noise level estimation of the first-layer wavelet decomposition to remove the wavelet coefficient belonging to noise; 4. and reconstructing through wavelet inverse transformation to obtain a denoised signal. Fig. 2 shows the effect of noise reduction before and after noise reduction, where (a) is an original signal diagram and (b) is a denoised signal diagram. Let W be the vector, the square set of the element wavelet decomposition coefficients is arranged according to the order from small to large, as shown in formula (1):
W＝[ω 1 ，ω 2 ，...，ω n ] (1)
where S represents the sum of squares of wavelet coefficients,let->The heuristic threshold T is expressed as:
where σ is the mean square error of the noise signal and N is the noise signal length.
Step 3, constructing a CNN network, which comprises three convolution layers, a pooling layer and three full connection layers, and outputting a characteristic vector with the length of 64 for 5×100 input samples. Each convolution layer is followed by a Batch Normalization layer and a ReLu activation function layer, so that data is regularized, and the generalization capability of the system is improved. Considering that the change of the high-speed rail signal is complex and the amplitude change is large, a Dropout module is added in the full-connection layer, so that the model is prevented from being trained excessively and the over-fitting phenomenon is prevented.
And 4, constructing a BiLSTM network, analyzing the change of the signal along with time by using two-time bidirectional LSTM and extracting the time characteristics. Specifically, two layers of LSTM networks are stacked, each layer has an input size of 5 and an output length of 16, and two-way total four data results are spliced in a model to finally obtain a feature vector with a length of 64.
And 5, replacing the full-connection layer with a RepMLP module. After the two feature vectors are spliced to obtain a mixed feature vector, inputting the result into a RepMLP module, combining the advantages of a full-connection layer and a convolution layer, taking global features and local information into consideration, further obtaining the occurrence probability of each of the seven categories, and outputting the type with the highest occurrence probability as a classification result.
And 6, testing the fused neural network by using the data set after the noise reduction treatment, adjusting network parameters according to a random small batch gradient descent method, and optimizing model parameters and network structures until the model converges to obtain a high-speed rail intrusion detection model with high accuracy and low misjudgment rate.
In the actually collected high-speed rail intrusion monitoring data, a large amount of data are undisturbed data or data only containing background noise, and the actual disturbance data are fewer, so that a data set sample cannot distinguish multiple disturbance types under the condition of only extracting a single characteristic. In order to fully extract sample characteristics and reduce manual dependence, a parallel CNN and BiLSTM network structure is adopted to obtain space-time characteristics of the sample, and a RepMLP is used for replacing a traditional full-connection layer. The specific idea is as follows:
because of various disturbance types in an actual high-speed rail monitoring scene, the method for manually extracting the characteristics is excessively dependent on subjective experience, and meanwhile, the problems of low efficiency, poor robustness and the like exist. It is therefore contemplated that the data features are automatically extracted using a deep learning network model. Because the high-speed railway intrusion data has the characteristics of fewer disturbance samples, shorter duration of disturbance events, smaller action range, various disturbance types and the like, the spatial structure information and time characteristics of the high-speed railway intrusion data should be fully considered when the characteristics are extracted. In addition, since the high-speed railway disturbance signal generally has a short-time strong vibration characteristic, the transient increase of the amplitude is represented in the time domain, but the characteristic is not enough to distinguish a plurality of disturbance types. Therefore, the present embodiment adopts CNN and BiLSTM networks to perform feature extraction, and mainly aims to extract time domain features, frequency domain features and shallow morphology features in the acquired signals, and obtain classification results by stitching the features.
Because the CNN can automatically learn and extract spatial features of input data, and consider that similarity exists in distribution of signals of adjacent sampling points in a window on a frequency domain, the embodiment utilizes the CNN to process and characterize changes and relations between adjacent sampling points of high-speed rail signals, and considers the requirement of high-speed rail on real-time performance, and network structures need to be optimized as much as possible under the condition of realizing feature extraction functions.
For input data with the size of 5 multiplied by 100, in order to fully extract the morphological characteristics of the shallow layer, three convolution layers are utilized to respectively convolve the activation function, so that the nonlinear effect of the network is enhanced. Simultaneously, three full connection layers are arranged, the output of the convolution layers are linked, the time domain, the frequency domain and the morphological characteristics are integrated, and finally, a characteristic vector with the length of 64 is established, and specific parameters are shown in table 1.
TABLE 1
Firstly, through three two-dimensional convolution layers, the convolution kernel size is 3 multiplied by 3, the step length is 1, the padding is 1, the number of channels can be increased from 1 to 64 according to the above arrangement, and the two-dimensional size after data is maintained unchanged, and is still 5 multiplied by 100; then carrying out two-dimensional Max Pooling operation with the size of 2 multiplied by 2 on the output of the three convolution layers, and reducing the data size of each channel to 2 multiplied by 50; finally, flattening to obtain a one-dimensional vector with the length of 64 multiplied by 2 multiplied by 50=6400, passing through three full connection layers, outputting feature vectors with the sizes of 1024, 256 and 64 respectively, and mapping the feature vectors to 64 dimensions.
To improve the accuracy of the model, each convolution layer is followed by a two-dimensional Batch Normalization layer and a ReLu activation function layer. The former performs a batch normalization operation on the output to reduce the adverse effects of data distribution migration, and the latter introduces non-linear fitting capability. Meanwhile, a Dropout operation of 0.5 is used in the full connection layer, so that overfitting of input data is avoided, and generalization capability of the CNN module is improved.
Because the disturbance event is not instantaneous in nature and has continuity in the actual high-speed rail scene, the time dimension change behavior of the signal is taken as an important basis when the signal is classified, and the forward and backward long-term dependence information at a specific moment is also fully considered. Since BiLSTM can consider the context information of the input sequence and analyze long-term dependency, process and predict time sequence data, applicant intends to analyze the change of the signal in the input window with time by using two layers of BiLSTM, extract its features in the time dimension and build a feature vector with length of 64. The number of layers of the BiLSTM is set to two, the input size of each layer is set to 5, and the hidden layer size is set to 16. I.e. two layers of LSTM networks are stacked, the second layer takes the output of the first layer as input, while each layer of network inputs sequence data in forward and reverse directions and gets output in both directions. The input data is finally output by two layers of hidden layers with the total size of 16, wherein the hidden layers are obtained through the network, and the feature vectors with the length of 4 multiplied by 16=64 can be obtained through splicing.
Research shows that compared with a single network model, the network structure adopting parallel CNN and BiLSTM can obtain more accurate and reasonable classification results by comprehensively considering the characteristics extracted by CNN and BiLSTM, and meanwhile, the operation efficiency is improved. The applicant therefore intends to splice the two eigenvectors to obtain a fused eigenvector of length 128.
Finally, a feature classification result is formed. Considering that various disturbance types common to high-speed rails are relatively similar in time domain and frequency domain, the method has the characteristics of short duration, small action range and the like, and the identification accuracy can not meet the actual requirements under the condition of only extracting the space-time characteristics of the high-speed rails. Meanwhile, in the same sample, the actual disturbance data accounts for only 10% or even less of the whole data, and the traditional single full-connection layer realizes the classification function according to the efficient modeling of long-distance dependence and position mode, so that the performance in the aspect of small-area feature recognition is not satisfactory. Therefore, the network structure needs to be further optimized, and local information perception is considered while global features are extracted. The RepMLP module is an effective combination of a full-connection layer and a convolution layer, and can fully utilize the global expression and the position sensing capability of the full-connection layer and the local structure extraction capability of convolution, enhance the local capturing characteristic and improve the classification performance. Thus, applicants intend to replace the full-ligation layer for sorting with the RepMLP module, the specific parameters of which are shown in Table 2.
TABLE 2
Finally, the logic vector with the length of 7 is subjected to a softmax layer to obtain the classification probability that the sample belongs to each category, and the category with the highest probability is the classification result, so that whether the input signal is destructive disturbance and the disturbance category to which the input signal belongs can be determined.
For each sampling point, inputting the data in 5 adjacent sampling points in 1s into the fusion network, obtaining the classification result of the sampling point at the moment, and taking the sampling point classified as destructive disturbance as the potential disturbance occurrence position at the current moment and giving an alarm according to the potential disturbance occurrence position.
2. The actual application stage:
step 1, arranging a DAS system on a high-speed railway site to be monitored, connecting one end of an existing optical cable with the DAS system, and performing actual test to ensure that signal collecting and transmitting parts of the DAS system all operate normally and keep a lower packet loss rate and a second-level transmission speed.
And 2, inputting the real-time data of the perimeter of the high-speed rail collected by the DAS system into a modulated wavelet noise reduction module for noise reduction.
And step 3, inputting real-time data in a space-time window after noise reduction into a trained neural network identification module. And extracting feature vectors by using CNN and BiLSTM in parallel, splicing the CNN and BiLSTM to obtain a mixed feature vector, inputting the mixed feature vector into the RepMLP to predict the occurrence probability of each category, and outputting the type with the highest occurrence probability as the classification result of the sampling point at the moment. And after the abnormal disturbance is determined, the middle point of the event window is set as a disturbance position point, and a warning signal is automatically sent to remind relevant personnel of what type of disturbance occurs at what position. The noise reduction threshold value is further adjusted according to the data actually collected on site, and the model parameters are optimized, so that the model parameters have high robustness, and finally the high-speed rail intrusion detection model with high accuracy and low misjudgment rate is obtained.
In a high-speed railway invasion scene, the duration of single disturbance is short, and disturbance data in each data sample occupy less data, so that data segmentation is needed, and features are extracted from part of samples in sequence. Meanwhile, because the characteristic points of different network structures for the data characteristics are different, a model structure of the parallel neural network is selected for simultaneously acquiring the time characteristics and the space characteristics. For each data sample, dividing the data into a plurality of data with the size of 5 multiplied by 100, inputting the data into CNNs comprising three convolution layers, one pooling layer and three full connection layers, extracting the characteristics of the window, establishing a characteristic vector with the length of 64, and representing the change and the connection of the high-speed rail signals in the window between adjacent sampling points. Simultaneously, data with the size of 5 multiplied by 100 in the window are input to BiLSTM of the two layers in parallel to obtain a characteristic vector with the length of 64, and the time-dependent change behavior of the high-speed rail signals in the same window is represented.
The traditional network structure adopts a full connection layer as the output of the result, and local information cannot be captured depending on distance and position modes. In the high-speed railway intrusion monitoring, the action range of disturbance types is short, the expressions of different disturbance types on the frequency domain are close, the recognition accuracy cannot be further improved by simply splicing feature vectors, and the structure is required to be further optimized. Therefore, after the feature vectors of the CNN and the BiLSTM are spliced to obtain the space-time feature vector with the length of 128, the result is input to the RepMLP module to replace a single full-connection layer, the local capturing characteristic of the convolution layer is fully utilized, the geometric detail information of the data is represented, and the instantaneous disturbance feature is extracted by emphasis. Finally, the classification probability of the sample belonging to each category is obtained, and the category with the highest probability is the classification result, so that whether the input signal is destructive disturbance and the disturbance category to which the input signal belongs can be determined, and when the input signal is abnormal disturbance, an alarm function is started and the disturbance position is positioned.
In specific implementation, the fused neural network model is trained by using the training set data after noise reduction until convergence, and the verification set data is used for determining super parameters and evaluating training results so as to determine proper model parameters. In a real application scene, inputting 5×100 high-speed railway perimeter real-time data collected by each sampling point into a noise reduction module, inputting a trained network model, obtaining a predicted classification result, namely whether the classification result is destructive disturbance and disturbance category, and determining the position where disturbance occurs according to the classification result of each sampling point, namely the layout position of the sampling points with the classification result being destructive disturbance. For several consecutive outlier samples, the system will choose the geometric distance center between them to predict the site of the intrusion.
Through tests, the classification speed of the CNN and BiLSTM-based high-speed railway perimeter intrusion detection system on real-time data is in the second level, and the CNN and BiLSTM-based high-speed railway perimeter intrusion detection system can effectively respond to an emergency and give an alarm under an actual application scene, so that accidents are prevented. Because the network layer number is not deep, the parameters are relatively less, and the CNN module and the BiLSTM module can be calculated in a distributed parallel manner, the system has the advantage of high processing speed. The noise reduction module is added to further reduce the influence of background noise, is less influenced by actual environments such as larger people flow or bad weather, and can maintain ideal classification precision which can reach more than 95%.
The above is only a preferred embodiment of the present invention, and the protection scope of the present invention is not limited to the above examples, and all technical solutions belonging to the concept of the present invention belong to the protection scope of the present invention. It should be noted that modifications and adaptations to the invention without departing from the principles thereof are intended to be within the scope of the invention as set forth in the following claims.

Claims (9)

1. The high-speed rail intrusion monitoring method based on the distributed optical fiber sensing and fusion neural network is characterized by comprising the following steps of:
s1, collecting disturbance signals of the perimeter of a high-speed rail by using a distributed optical fiber acoustic wave sensing system, dividing the collected disturbance signals into a plurality of sampling points within a 1S adjacent space range, dividing the data into one sample data, and constructing a sample data set containing a plurality of disturbance type labels;
s2, aiming at background noise characteristics, removing background noise contained in sample data by using a heuristic threshold wavelet noise reduction method;
s3, constructing a high-speed rail intrusion detection model, wherein the high-speed rail intrusion detection model comprises a convolutional neural network, a two-way long-short-term memory neural network, a splicing module, a RepMLP module and a softmax layer, and training and verifying the high-speed rail intrusion detection model by adopting a sample data set; the convolutional neural network and the two-way long-short-term memory neural network are arranged in parallel, and feature extraction is respectively carried out on sample data with background noise removed, so that spatial features and temporal features of the sample data are obtained; the spatial characteristics of the sample data are used for representing the change and the connection of disturbance signals in a window between adjacent sampling points, and the temporal characteristics of the sample data are used for representing the change behavior of the disturbance signals in the same window along with time; the splicing module splices the spatial characteristics and the temporal characteristics of the sample data, and inputs the space-time characteristic vector of the obtained disturbance signal to the RepMLP module; the RepMLP module utilizes the local capturing characteristic of the convolution layer to represent the geometric detail information of the space-time feature vector, and extracts the instantaneous disturbance features of disturbance signals of different categories; the softmax layer classifies the extracted instantaneous disturbance characteristics of disturbance signals of different categories, and predicts the occurrence probability of each disturbance category corresponding to the disturbance signal;
s4, acquiring acoustic wave signals of the perimeter of the high-speed rail by using a distributed optical fiber acoustic wave sensing system, carrying out segmentation processing and denoising processing on the acquired acoustic wave signals, then introducing the acoustic wave signals into a high-speed rail intrusion detection model, extracting instantaneous disturbance characteristics of the acoustic wave signals at different sampling point positions, outputting the occurrence probability of each disturbance category corresponding to the acoustic wave signals, and outputting the type with the highest occurrence probability as a classification result of the sampling point at the current moment.
2. The method for monitoring the intrusion of the high-speed rail based on the distributed optical fiber sensing and fusion neural network according to claim 1, wherein in the step S1, the acquired disturbance signals are subjected to segmentation processing, and data of 5 sampling points in an adjacent space range within 1S are divided into one sample of data.
3. The method for monitoring the intrusion of the high-speed rail based on the distributed optical fiber sensing and fusion neural network according to claim 1, wherein the disturbance type labels comprise seven types of climbing, knocking, digging, electric drill, foreign matter intrusion, illegal intrusion and abnormal weather influence.
4. The method for monitoring the intrusion of the high-speed rail based on the distributed optical fiber sensing and fusion neural network according to claim 1, wherein in the step S2, the process of removing the background noise contained in the sample data by using a heuristic threshold wavelet noise reduction method according to the background noise characteristics comprises the following steps:
s21, performing three-layer wavelet decomposition calculation on the disturbance signal by adopting a coif5 wavelet function;
s22, setting a fixed threshold value and an unbiased threshold value;
s23, quantifying a threshold value of the wavelet decomposition high-frequency coefficient, and adjusting according to the noise level estimation of the first-layer wavelet decomposition to remove the wavelet coefficient belonging to the noise; when the signal-to-noise ratio is larger than a preset signal-to-noise ratio threshold, adopting a fixed threshold to perform noise reduction processing, otherwise adopting an unbiased threshold to perform noise reduction processing
S24, reconstructing through wavelet inverse transformation to obtain a denoised disturbance signal.
5. The method for monitoring the intrusion of the high-speed rail based on the distributed optical fiber sensing and fusion neural network according to claim 1, wherein in the step S2, the convolutional neural network comprises a first two-dimensional convolutional layer, a second two-dimensional convolutional layer, a third two-dimensional convolutional layer, a pooling layer, a flattening layer, a first full-connection layer, a second full-connection layer and a third full-connection layer which are sequentially connected;
the convolution kernel sizes of the first two-dimensional convolution layer, the second two-dimensional convolution layer and the third two-dimensional convolution layer are 3 multiplied by 3, the step length is 1, and the padding is 1; for input data with the size of 5×100, three two-dimensional convolution layers increase the number of channels from 1 to 64, and maintain the two-dimensional size of the data unchanged, and still be 5×100; the Pooling layer is used for carrying out two-dimensional Max Pooling operation with the size of 2 multiplied by 2 on the output of the third two-dimensional convolution layer, and reducing the data size of each channel to 2 multiplied by 50; the flattening layer is used for flattening the output of the pooling layer to obtain a one-dimensional vector with the length of 6400; the first full connection layer, the second full connection layer and the third full connection layer sequentially process the one-dimensional vector and map the one-dimensional vector to a 64-dimensional space feature vector.
6. The method for monitoring the intrusion of the high-speed rail based on the distributed optical fiber sensing and fusion neural network according to claim 5, wherein a two-dimensional regularization layer and a ReLu activation function layer are respectively connected behind the two-dimensional convolution layer; the two-dimensional regularization layer is used for carrying out batch normalization operation on the output to reduce adverse effects caused by data distribution migration, and the ReLu activation function layer is used for introducing nonlinear fitting capacity;
and a dropout layer is connected between the first full-connection layer and the second full-connection layer, and the dropout layer is used for improving generalization capability so as to avoid fitting input data.
7. The method for monitoring the intrusion of the high-speed rail based on the distributed optical fiber sensing and fusion neural network according to claim 1, wherein the two-way long-short-term memory neural network comprises two LSTM networks stacked with each other, wherein the input size of each LSTM network is 5, and the hidden size is 16; the second layer LSTM network takes the output of the first layer LSTM network as input, and each layer LSTM network inputs the sequence data according to the forward direction and the reverse direction and obtains the output of the two directions;
input data with the size of 5 multiplied by 100 is input into a first layer LSTM network, and finally two layers of bidirectional hidden layer outputs with the size of 16 are obtained, and the time feature vectors with the length of 64 are obtained by splicing.
8. The method for monitoring high-speed rail intrusion based on the distributed optical fiber sensing and fusion neural network according to claim 1, wherein the RepMLP module comprises a fourth full-connection layer, a convolution layer, a regularization layer and a fifth full-connection layer which are sequentially connected;
the RepMLP module processes the space-time characteristic vector of the disturbance signal, and inputs the obtained logic vector with the length of 7 into the softmax layer.
9. The method for monitoring the intrusion of the high-speed rail based on the distributed optical fiber sensing and fusion neural network according to claim 1, wherein in the step S4, the occurrence probability of each disturbance category corresponding to the output acoustic wave signal is judged, and if the occurrence probability of all the types is lower than a preset probability threshold, the non-destructive disturbance is judged; if the occurrence probability of at least one type is higher than the preset probability threshold, outputting the type with the highest occurrence probability as a classification result of the sampling point at the current moment, and determining the position of the sampling point as a disturbance point.
CN202311740490.4A 2023-12-15 2023-12-15 High-speed rail intrusion monitoring method based on distributed optical fiber sensing and fused neural network Pending CN117609885A (en)

Priority Applications (1)

Application Number Priority Date Filing Date Title
CN202311740490.4A CN117609885A (en) 2023-12-15 2023-12-15 High-speed rail intrusion monitoring method based on distributed optical fiber sensing and fused neural network

Applications Claiming Priority (1)

Application Number Priority Date Filing Date Title
CN202311740490.4A CN117609885A (en) 2023-12-15 2023-12-15 High-speed rail intrusion monitoring method based on distributed optical fiber sensing and fused neural network

Publications (1)

Publication Number Publication Date
CN117609885A true CN117609885A (en) 2024-02-27

Family

ID=89957931

Family Applications (1)

Application Number Title Priority Date Filing Date
CN202311740490.4A Pending CN117609885A (en) 2023-12-15 2023-12-15 High-speed rail intrusion monitoring method based on distributed optical fiber sensing and fused neural network

Country Status (1)

Country Link
CN (1) CN117609885A (en)

Cited By (2)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN118051830A (en) * 2024-04-16 2024-05-17 齐鲁工业大学(山东省科学院) Perimeter security intrusion event identification method
CN118180161A (en) * 2024-05-17 2024-06-14 太原理工大学 Method and device for constructing multidimensional coupling vibration identification model of strip mill roller system

Cited By (2)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN118051830A (en) * 2024-04-16 2024-05-17 齐鲁工业大学(山东省科学院) Perimeter security intrusion event identification method
CN118180161A (en) * 2024-05-17 2024-06-14 太原理工大学 Method and device for constructing multidimensional coupling vibration identification model of strip mill roller system

Similar Documents

Publication Publication Date Title
CN110097755B (en) Highway traffic flow state identification method based on deep neural network
CN117609885A (en) High-speed rail intrusion monitoring method based on distributed optical fiber sensing and fused neural network
CN113449660B (en) Abnormal event detection method of space-time variation self-coding network based on self-attention enhancement
CN111814699B (en) Deep learning earthquake prediction method for SWARM electromagnetic satellite data
CN106778841A (en) The method for building up of abnormal electricity consumption detection model
CN114692681B (en) SCNN-based distributed optical fiber vibration and acoustic wave sensing signal identification method
CN117421684B (en) Abnormal data monitoring and analyzing method based on data mining and neural network
CN113240199B (en) Port ship track prediction method based on DILATE _ TLSTM
CN113848417A (en) Rail transit power supply equipment fault prediction method and device
CN114926825A (en) Vehicle driving behavior detection method based on space-time feature fusion
CN112541480B (en) Online identification method and system for tunnel foreign matter invasion event
CN116596151B (en) Traffic flow prediction method and computing device based on time-space diagram attention
CN110263944A (en) A kind of multivariable failure prediction method and device
CN115060184B (en) Optical fiber perimeter intrusion detection method and system based on recursion diagram
CN116838955A (en) Two-stage oil and gas pipeline line interference identification method
CN117272102A (en) Transformer fault diagnosis method based on double-attention mechanism
CN115700542A (en) Optical fiber pipeline safety early warning algorithm based on deep learning
CN115017828A (en) Power cable fault identification method and system based on bidirectional long-short-time memory network
CN112836591B (en) Method for extracting optical fiber early warning signal characteristics of oil gas long-distance pipeline
Di et al. An automatic and integrated self-diagnosing system for the silting disease of drainage pipelines based on SSAE-TSNE and MS-LSTM
CN113762007B (en) Abnormal behavior detection method based on appearance and action feature double prediction
CN117648652A (en) Abnormal traffic state detection method based on dynamic hypergraph convolutional neural network
CN118015562A (en) Method and system for extracting key frames of traffic accident monitoring video in severe weather
CN116504005B (en) Perimeter security intrusion signal identification method based on improved CDIL-Bi-LSTM
Chaganti et al. Predicting Landslides and Floods with Deep Learning

Legal Events

Date Code Title Description
PB01 Publication
PB01 Publication
SE01 Entry into force of request for substantive examination
SE01 Entry into force of request for substantive examination