CN109583323B

CN109583323B - Subway vibration signal identification method based on door control circulation unit

Info

Publication number: CN109583323B
Application number: CN201811336175.4A
Authority: CN
Inventors: 江颉; 张继萍; 杨高级; 沈敖
Original assignee: Zhejiang University of Technology ZJUT
Current assignee: Zhejiang University of Technology ZJUT
Priority date: 2018-11-12
Filing date: 2018-11-12
Publication date: 2021-05-18
Anticipated expiration: 2038-11-12
Also published as: CN109583323A

Abstract

A subway vibration signal identification method based on a gate control cycle unit (GRU) network model is disclosed. The method comprises the following steps: sample data construction is carried out on the time tracking signal data of the sensor in two modes of program automatic marking and manual marking; carrying out dimensionality reduction on the sample data by adopting a self-coding network model; constructing a GRU subway vibration signal identification model; and testing the sample classification and identification effect. The invention effectively solves the problems that the non-professional personnel can not judge and identify and the working strength is overlarge in the past, and obviously improves the identification accuracy and efficiency.

Description

Subway vibration signal identification method based on door control circulation unit

Technical Field

The invention relates to the fields of deep learning, feature extraction, noise signal identification and the like, in particular to the field of subway vibration signal identification by applying a gate control cycle Unit (GRU).

Background

In densely populated urban areas, the metro network is the most efficient large-scale passenger transport system. Although subway systems have many advantages, there are various disadvantages that cannot be avoided. One of them is that vibration and radiation noise generated in the running process of the subway can have unpleasant influence on environments such as buildings along the subway. Because the subway network is generally under shallow ground, the vibration of the tunnel caused by the subway passing through the vehicle is difficult to avoid being transmitted to the adjacent buildings. Although the common buildings are generally not damaged due to the fact that buffering distance is considered during design of vibration caused by the subway or vibration reduction measures are taken for the railway, damage caused by subway vibration can be caused when cultural relics with high protection requirements exist nearby the subway. In addition, for some specific devices or instruments, attention needs to be paid; for example, in a nanotechnology laboratory or factory, vibrations caused by a subway may affect the normal operation of the equipment, thereby causing unnecessary losses.

In order to research the influence of subway vibration on the environment, a vibration instrument and a sensor are used to carry out field measurement on site, and a laboratory is brought back to extract a subway vibration signal from the whole vibration measurement signal, so that the method is a very critical means. Since the subway passing in the measurement site is intermittent, whether the subway passes through the measured vibration signal is generally determined in a manual mode. However, under the condition that a vibration measurement signal is formed by passing through a train in large-scale monitoring, the efficiency of a manual discrimination mode is too low.

Disclosure of Invention

In order to solve the problems that the subway passing vibration signals are distinguished and picked up manually from continuous vibration signals, the exposure efficiency is low, the technical personnel experience is seriously depended on, and the like, the subway train vibration signal identification problem is converted into a classification problem in deep learning, and a subway vibration signal identification method based on a gate control cycle Unit (GRU) with high identification accuracy and efficiency is provided.

The technical scheme adopted by the invention for solving the technical problems is as follows:

a subway vibration signal identification method based on a door control circulating unit comprises the following steps:

step 1, constructing a training sample, firstly preprocessing time tracking signal data recorded by a sensor, then constructing a sample from the preprocessed signal data, taking a signal intensity value contained in a time period T with subway passage as a positive sample and marking the signal intensity value as 1, and taking a vibration signal value without subway in other time periods except the time period T as a negative sample and marking the signal value as 0;

step 2, adopting a self-coding network model to perform dimensionality reduction processing on the training sample output in the step 1 to generate sample format data required by the GRU neural network model;

step 3, constructing a GRU subway vibration signal identification model, simulating the subway vibration signal identification problem to a multi-classification problem in a neural network, inputting the training set sample after dimensionality reduction into a deep learning model, training to obtain a classifier for subway vibration signal identification, and adopting a GRU algorithm in a recurrent neural network as a learning model; defining 2 placeholders, taking X as an input sequence and Y as a corresponding label, and setting the maximum time length of the sample sequence to be Max _ length; since Batch processing operation is adopted and the original data of X is input into the network one by one, converting X into Max _ length arrays according to the maximum time length, wherein each array comprises Batch _ size elements; inputting an input sequence X into a recurrent neural network; then, accessing the obtained result to a full connection layer to obtain a prediction sequence pred (0,1), and finishing the forward propagation of the network; then the model adopts Softmax (cross entropy loss function) to measure the probability distribution distance between the predicted pred value and the true value Y value; continuously adjusting the value of the weight parameter in the recurrent neural network by a back propagation algorithm and a gradient descent algorithm along with time until the stopping criterion is met;

step 4, subway vibration signal identification

Inputting the sample data of the test set into a GRU signal identification model, and judging the category of the sample; the type of the GRU signal identification model is 0 or 1, a 0 label represents that no subway passes through the time period contained in the sample, and a 1 label represents that the subway passes through the time period contained in the sample.

Further, in step 1, the sensor time tracking data set is preprocessed by the method comprising: smoothing the signal of the instantaneous jerky jitter of the positive sample, and replacing the signal with an adjacent conventional signal value; and aiming at the construction of the negative sample, two modes of program automatic marking and manual marking are combined.

Still further, the process of automatically marking the negative sample by the program comprises the following steps: after extracting the positive sample from the data set, the program takes a T time period as a sliding time window from the start time 0s of the data set, moves the T time (unit is second) to the right every time, and a signal value sequence generated in the sliding time window is a negative sample and is labeled as 0; the processing procedure of the manual marking mode is as follows: and respectively moving the sliding time window T where the positive sample is located for K seconds to two sides to serve as starting points of the sliding window, wherein K is T/3, the sliding time window T is moved once every 1 second until the sliding time window is completely extracted into a positive sample time period range, and at the moment, the signal value sequence values in the sliding time window T are all negative samples.

In step 3, the gate control circulation unit is a circulation neural network which shares weights in a time dimension through a neuron with self feedback, and achieves the accumulation speed of the control dependence information through selectively forgetting the accumulated information and selectively adding new information. For the construction of a GRU subway vibration signal identification network structure, a deep network model is utilized to improve the feature extraction function of a GRU network, a multi-layer GRU circulation network and a single-layer full-connection network structure are constructed, a Softplus function is selected from the GRU network as an activation function of the model, the Softplus function is smoother than a ReLU function, and meanwhile, an output value smaller than 0 is relatively reserved; in the aspect of hyper-parameter optimization, an optimization algorithm Adam with self-adaptive learning rate is adopted, the Adam is an extended random gradient descent algorithm, and each parameter is dynamically adjusted by utilizing first moment estimation and second moment estimation of a gradient.

The technical conception of the invention is as follows: and (3) judging the sensor time tracking signal data by applying a GRU network model, and identifying the time period when the subway vibration signal exists. Firstly, constructing sample data for sensor time tracking signal data in two modes of program automatic marking or manual marking; then, using an Auto-Encoder (AE) to perform dimensionality reduction processing on the sample data; and inputting the sample data subjected to dimensionality reduction into a GRU neural network model to finally obtain the subway vibration signal classifier. The performance of the protocol was evaluated by a series of experiments using conventional recall, accuracy and F1 values. The experimental result shows that the F1 value can reach 84% in solving the problem, the obstacle that non-professional personnel cannot judge and identify in the past is effectively solved, and the identification efficiency is remarkably improved.

The invention has the following beneficial effects: the method comprises the steps of constructing a training sample by using sensor time tracking signal data through two modes of program automatic marking and manual marking, adopting a self-coding network model to perform dimensionality reduction on the training sample, replacing an original high-dimensional characteristic model with a generated low-dimensional characteristic model, successfully constructing a GRU subway vibration signal recognition model, and improving the target recognition accuracy.

Drawings

Fig. 1 is a flowchart of a subway vibration signal identification method based on a door control cycle unit according to the present invention.

FIG. 2 is a flow chart of a dimension reduction process of a self-coding network.

Fig. 3 is a subway vibration signal recognition model architecture diagram.

Detailed Description

The invention is further described below with reference to the accompanying drawings.

Referring to fig. 1 to 3, a subway vibration signal identification method based on a door control cycle unit includes two stages of constructing a training sample and constructing a training model. Firstly, preprocessing the sensor time tracking signal data, and then generating a training sample by a process sequence automatic marking mode and a process sequence manual marking mode. Then, a self-encoding network (AE) model is used for compressing and reducing dimensions of the training sample characteristics, and a characteristic representation which is better than that of the original sample is generated. And finally, training the training sample as an input signal of a GRU subway vibration signal recognition model to obtain a classifier. The subway vibration signal identification method comprises the following steps:

step 1, constructing a training sample. The sensor time tracking signal data is first preprocessed and then a sample is constructed of the preprocessed sensor time tracking signal data. The signal intensity value included in the time period T when the subway train passes is taken as a positive sample (label is 1), and the signal value without subway vibration in other time periods T is taken as a negative sample (label is 0).

And 2, performing dimensionality reduction on the training sample output in the step one by adopting an Auto-Encoder (AE) model to generate sample format data required by the GRU neural network model. The basic model of the self-coding network is divided into three layers of neural networks, namely an input layer, a hidden layer and an output layer. The input is compressed in a mode that the number of the neurons in the next layer is less than that of the neurons in the previous layer in the hidden layer, and feature description better than original data is trained and learned. The training samples after the dimension reduction processing are split in a 3:1:1 mode, wherein 60% of the data samples are used for training the model, 20% of the samples are used as verification data of the model, and the remaining 20% are used as test samples.

And 3, constructing a GRU subway vibration signal identification model. The basic idea of the GRU subway vibration signal identification model is to analogize the subway vibration signal identification problem to the multi-classification problem in the neural network. And inputting the training set samples subjected to dimensionality reduction into a deep learning model, and training to construct a classifier for recognizing subway vibration signals. The invention adopts a GRU algorithm in a recurrent neural network as a learning model and is mainly divided into two parts of defining a network structure and training a model. X is used as an input sequence, Y is a corresponding label, and the maximum time length of the sample sequence is set to Max _ length. Since a Batch (Batch _ size) operation is employed and the input sequence X is input into the network one by one, the input sequence X is converted into Max _ length arrays each containing Batch _ size elements by the maximum time length Max _ length. Inputting an input sequence X into a recurrent neural network; and then accessing the result to a full connection layer to obtain a prediction sequence pred (0,1), and finishing the forward propagation of the network. Then the model adopts a Softmax cross entropy loss function to measure the probability distribution distance between the predicted value pred and the true value Y. The weight parameter values in the recurrent neural network are continuously adjusted by a back propagation algorithm and a gradient descent algorithm over time until a stopping criterion is met.

And 4, identifying subway vibration signals. And inputting the sample data of the test set into a GRU signal identification model, and judging the type of the sample. The GRU signal recognition model class is 0 or 1. The 0 label represents that no subway passes in the time period contained in the sample, and the 1 label represents that the subway passes in the time period contained in the sample. And three measures of Recall (Recall), Precision (Precision) and F1 values were used to evaluate the results.

Further, in step 1, preprocessing the sensor time tracking data set includes: the jerky signal is smoothed in a short time (which is much less than the subway running time) and replaced with a neighboring regular signal value. And aiming at the construction of the negative sample, two modes of program automatic marking and manual marking are combined.

Wherein the program automatically marks negative examples as: after extracting the positive samples from the data set, after removing the positive samples from the data set, the program takes a time period T (the time period which is the same as the time period of the positive samples) as a sliding time window from the beginning time 0s of the data set, moves for T seconds to the right every time, and generates a signal intensity sequence as a negative sample with a label of 0 in the sliding time window. This method cannot cover all types, such as a sample having a part overlapping with the subway signal in the T period, and the sample should be treated as a negative sample because there is a large deviation between the start time and the end time and the positive sample (including the complete subway vibration signal strength period). The method adopts a manual marking mode to make up for the problem that the situation that all sample types cannot be fully covered due to the existence of a program automatic marking mode. The manual marking mode comprises the following processing processes: and respectively moving the sliding time window T in which the positive samples are located for K seconds (K is T/3) towards two sides as starting points of the sliding window, moving the sliding time window T once every 1 second until the sliding time window completely removes the positive sample time period range, wherein the signal sequence values in the sliding time window T are all negative samples. The value of K needs to be balanced according to the actual situation, and if the value of K is too large, the strategy is degraded into a program automatic marking mode.

Furthermore, in the step 3, for constructing the GRU subway vibration signal identification network structure, the deep network model is used to improve the feature extraction function of the GRU network, and a multi-layer GRU circulation network and a one-layer fully-connected network structure are constructed. Selecting a Softplus function as an activation function of the model in the GRU network, wherein the Softplus function is smoother than a ReLU function and relatively reserves an output value smaller than 0. In the aspect of super-parameter optimization, a learning rate adaptive optimization algorithm adam (adaptive motion estimation) is adopted. Adam is an extended stochastic gradient descent algorithm, essentially combines the advantages of Adagrad processing sparse gradients and the characteristic that RMSprop is good at processing non-stationary targets, and dynamically adjusts each parameter by using first-order moment estimation and second-order moment estimation of the gradients. The method can ensure that the learning rate has a definite range in each iteration process after bias correction, so that the overall fluctuation of parameters is not large, and the memory has no extra requirement.

The subway vibration signal identification method based on GRU of the embodiment comprises the following steps:

step 1, constructing a training sample

Firstly, after sensor measurement data are visualized by a Gnuplot tool, all subway vibration signals are marked in a manual mode to serve as positive samples. And measuring the sensor near the subway to obtain a certain section of subway vibration signal. It can be observed that when two time periods have a clearly convex peak, but other times tend to be calm, and the time periods are both 9 seconds, i.e. T-9 s. In the actual measurement process, background noise generally exists in the sampling value of the time period without subway.

Then, the negative sample is combined and constructed by two ways of program automatic marking and manual marking. The program automatically marks: the program intercepts the signal strength sequence as a negative sample, labeled 0, for a 9s period of time (the same period of time as the positive sample) from the start time 0s of the data set. Moving the sliding time window backwards every 9 seconds automatically generates negative sample data. Manual marking: as shown in the case of fig. 3, the sliding time window (9s) in which the positive sample is located is shifted to both sides for 3(9/3 ═ 3) seconds, and then sliding time window 1 and sliding time window 2 are obtained. Then, from the starting point, the sliding time window 1 and the sliding time window 2 are respectively left and right at intervals of 1 second, and each sliding time window generates negative sample data until the sliding time window completely eliminates the positive sample time period range.

And 2, adopting an Auto-Encoder (AE) dimension reduction process.

As shown in fig. 2, the most important features in the subway vibration signal sample data are extracted by constructing a two-layer self-coding network. The initial sample data (sampling frequency 50HZ, 9 seconds total, 50 x 9 is 450 seconds) is reduced from 450 dimensions to 243 dimensions, and then from 243 dimensions to 36 dimensions again.

Step 3, constructing a GRU subway vibration signal identification model

As shown in fig. 3, a network model is constructed in which a GRU-loop network with 2 hidden layers (the hidden layers contain 100, 50 BN _ GRU units/definitions, respectively) is combined with a single-layer fully-connected network. And activating functions in the GRU network model, wherein the optimization algorithms are a Softplus function and an Adam algorithm respectively. 2 placeholders are defined, X as the input sequence, Y as the corresponding label, and the sample sequence maximum time length is set to Max _ length-36. The Batch size is set to Batch _ size 32. Operation and the raw data for X is input into the network one by one, so X is translated by the maximum length of time into Max _ length arrays, each containing Batch _ size elements. Firstly, inputting an input sequence X into a GRU function model; then, the result is accessed to a full connection layer to obtain a prediction sequence pred; the probability distribution distance between the predicted value pred and the true value Y is then evaluated using a Softmax cross entropy loss function. The weight parameter values in the recurrent neural network are continuously adjusted through a back propagation algorithm and a gradient descent algorithm over time until a stopping criterion is met.

Step 4, subway vibration signal identification

Inputting the test set into the GRU signal recognition model in the step 4 for classification and judgment, and evaluating an experimental result by adopting three evaluation indexes of Recall rate (Recall), Precision rate (Precision) and F1 value.

In order to verify the recognition effect of the method, the recognition rate obtained by the GRU signal recognition model is compared with the recognition rates of three traditional Machine learning classification models, namely logistic regression (logistic regression), Naive Bayes (Naive Bayes) and Support Vector Machine (Support Vector Machine), and the result is shown in table 1.

TABLE 1

As can be seen from table 1, the average recognition rate of the subway vibration signal recognition method based on GRU results test is greatly improved compared with the recognition rate of the traditional machine learning classification model, so that the recognition effect of the method of the present invention is superior to the recognition effect of the traditional machine learning classification model.

Claims

1. A subway vibration signal identification method based on a door control circulating unit is characterized in that: the subway vibration signal identification method comprises the following steps:

step 1, constructing a training sample, namely preprocessing time tracking signal data acquired by a vibration sensor and passed by a plurality of trains, constructing input sample data by using the preprocessed data, taking a signal intensity value contained in a time period T passed by a subway as a positive sample and marking the signal intensity value as 1, and treating a non-subway vibration signal background value outside the time period T as a negative sample and marking the signal intensity value as 0;

step 2, adopting a self-coding network model to perform dimensionality reduction on the training sample output in the step 1, and constructing sample format data required by a neural network model of a gate control circulation unit;

step 3, applying GRU to construct a subway vibration signal identification model, simulating the subway vibration signal identification problem of a vehicle passing through a neural network to a multi-classification problem in the neural network, inputting a training set sample after dimensionality reduction into a deep learning model, training to form a classifier for identifying subway vibration signals, and constructing the learning model by adopting a GRU algorithm in a cyclic neural network; defining 2 placeholders, taking X as an input sequence and Y as a corresponding label, and setting the maximum time length of the sample sequence with subway passing as Max _ length; because Batch processing operation is adopted and the original data of X is input into the network one by one, X is converted into Max _ length arrays according to the maximum time length, and each array comprises Batch _ size elements; inputting an input sequence X into a recurrent neural network; then, the result is accessed to a full connection layer to obtain a prediction sequence pred (0,1), and the forward propagation of the network is finished; then measuring the probability distribution distance between the model predicted value pred and the true value Y by adopting a Softmax cross entropy loss function; continuously adjusting the value of the weight parameter in the recurrent neural network through a back propagation algorithm and a gradient descent algorithm along with time until the stopping criterion is met;

step 4, subway vibration signal identification

Inputting the sample data of the test set into a GRU signal identification model, and judging the category of the sample; the type of the GRU signal identification model is 0 or 1, a 0 label represents that no subway passes through the time period contained in the sample, and a 1 label represents that the subway passes through the time period contained in the sample;

in step 1, the sensor time tracking data set is preprocessed, and the method includes: smoothing the rapidly jittered signal in a short time, and replacing the signal with the adjacent conventional signal intensity; aiming at the structure of the negative sample, two modes of program automatic marking and manual marking are combined;

the processing procedure of automatically marking the negative sample by the program is as follows: after the positive samples are intercepted from the data set, the program takes a T time period as a sliding time window from the beginning time 0s of the data set, moves for T seconds to the right every time, a signal intensity sequence generated in the sliding time window is a negative sample, and the label is 0; the processing process of the manual marking mode comprises the following steps: and respectively moving the sliding time window T where the positive sample is located for K seconds to two sides to serve as starting points of the sliding window, wherein K is T/3, the sliding time window T is moved once every 1 second until the sliding time window completely removes the time period range of the positive sample, and the signal intensity sequence values in the sliding time window T are all negative samples at the moment.

2. A subway vibration signal identification method as claimed in claim 1, which is based on door control circulation unit, characterized in that: in the step 3, for constructing a network structure for identifying the subway vibration signals of the GRUs, a deep network model is utilized to improve the function of extracting features of the GRU network, a multi-layer GRU cycle network and a one-layer fully-connected network structure are constructed, a Softplus function is selected from the GRU network as an activation function of the model, the Softplus function is smoother than a ReLU function, and an output value smaller than 0 is properly reserved; in the aspect of hyper-parameter optimization, an optimization algorithm Adam with self-adaptive learning rate is adopted, the Adam is an extended random gradient descent algorithm, and each parameter is dynamically adjusted by utilizing first moment estimation and second moment estimation of a gradient.