CN111623868B - Convolutional neural network construction method for rail corrugation identification - Google Patents
Convolutional neural network construction method for rail corrugation identification Download PDFInfo
- Publication number
- CN111623868B CN111623868B CN202010397122.4A CN202010397122A CN111623868B CN 111623868 B CN111623868 B CN 111623868B CN 202010397122 A CN202010397122 A CN 202010397122A CN 111623868 B CN111623868 B CN 111623868B
- Authority
- CN
- China
- Prior art keywords
- vibration
- domain information
- information
- time domain
- corrugation
- 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.)
- Active
Links
Images
Classifications
-
- G—PHYSICS
- G01—MEASURING; TESTING
- G01H—MEASUREMENT OF MECHANICAL VIBRATIONS OR ULTRASONIC, SONIC OR INFRASONIC WAVES
- G01H17/00—Measuring mechanical vibrations or ultrasonic, sonic or infrasonic waves, not provided for in the preceding groups
-
- G—PHYSICS
- G01—MEASURING; TESTING
- G01B—MEASURING LENGTH, THICKNESS OR SIMILAR LINEAR DIMENSIONS; MEASURING ANGLES; MEASURING AREAS; MEASURING IRREGULARITIES OF SURFACES OR CONTOURS
- G01B11/00—Measuring arrangements characterised by the use of optical techniques
- G01B11/02—Measuring arrangements characterised by the use of optical techniques for measuring length, width or thickness
-
- 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
- G06N3/08—Learning methods
Landscapes
- Physics & Mathematics (AREA)
- Engineering & Computer Science (AREA)
- Theoretical Computer Science (AREA)
- General Physics & Mathematics (AREA)
- Data Mining & Analysis (AREA)
- Molecular Biology (AREA)
- Biophysics (AREA)
- Computational Linguistics (AREA)
- Artificial Intelligence (AREA)
- Evolutionary Computation (AREA)
- General Health & Medical Sciences (AREA)
- Biomedical Technology (AREA)
- Computing Systems (AREA)
- General Engineering & Computer Science (AREA)
- Life Sciences & Earth Sciences (AREA)
- Mathematical Physics (AREA)
- Software Systems (AREA)
- Health & Medical Sciences (AREA)
- Measurement Of Mechanical Vibrations Or Ultrasonic Waves (AREA)
Abstract
The invention relates to monitoring of rail transit wheel-track short wave irregularity, in particular to a convolutional neural network construction method for rail track corrugation identification, which comprises the method for collecting and processing specific vibration information of a rail transit carrier, frequency domain data in a frequency range which does not influence the specific vibration detection characteristic representation in each frequency domain information is removed so as to respectively obtain optimized frequency domain information, each optimized frequency domain information is expressed as a frequency spectrum with the same frequency coordinate axis length, and the frequency spectrum data of the frequency spectrum is used as a training sample of a one-dimensional convolutional neural network. Compared with the traditional method for detecting and identifying by adopting the corrugation tester CAT, the method has the advantages that after the convolutional neural network training, vibration information and vehicle displacement information are acquired, and the signal is input into a computer containing the specially trained convolutional neural network for processing, so that the rail corrugation can be identified quickly and efficiently.
Description
Technical Field
The invention relates to monitoring of rail transit wheel-rail short wave irregularity, in particular to a convolutional neural network construction method for rail corrugation identification.
Background
Rail corrugation, i.e. rail wave wear, is shown in figure 1. Because the subway has high load capacity, large running density, complex line conditions (such as small curve radius and diversified track structure), and frequent vehicle starting and braking, the interaction between the wheel and the track is intensified, and the wavy abrasion (referred to as steel track corrugation) of the steel rail is serious.
Rail corrugation can cause a number of problems, such as abnormal vibration and noise pollution of the vehicle rails, and reduced fatigue reliability of the vehicle and rail components. According to literature reports, the noise difference of a subway line with a rail corrugation in a cab before and after grinding can reach nearly 10dBA, and the vibration acceleration of a subway rail fastener with a remarkable rail corrugation can even reach 195 g. In addition, fatigue fracture failure of some vehicle rail components is directly related to rail corrugation, such as vehicle axle box lug fracture, series steel spring fatigue fracture, rail fastener spring strip fracture, and the like. Rail corrugation has severely impacted vehicle safety operations and greatly increased maintenance costs. At present, rail grinding is one of the most main means for controlling rail corrugation, and when grinding strategies (such as grinding period and grinding amount) are determined, it is necessary to know the rail corrugation state in advance. Therefore, the method has important significance for monitoring the state and identifying the fault of the rail corrugation.
The traditional measuring method of the rail corrugation can be summarized into 3 types: chord measurement, inertial reference, and machine vision. The chord measuring method uses the steel rail as a moving reference system, so that the reference of the measurement reference is in a changing state along with the change of the height irregularity of the steel rail, the transmission function ratio (the ratio of the measured value to the actual value) is not constant to 1, and the chord measuring method cannot really and reliably test the corrugation of the steel rail. The inertia reference method usually represents the corrugation value through the quadratic integration of axle box acceleration, and has the defects of being easily interfered by wheel abrasion, and having larger measurement error under the low-speed working condition due to the influence of a high-pass filter, and generally only being used for large rail inspection vehicles. Machine vision methods often require precise photoelectric camera equipment and complex image processing means, employ a bulky and complicated pattern recognition technology to perform specific back-end processing, and have high practical application difficulty and high cost. For example, testing a subway line using a corrugation tester CAT based on one of the conventional methods often takes as long as a month, and is time-consuming, and the conventional detection methods have insufficient reliability and versatility and high cost.
Disclosure of Invention
The invention provides a convolutional neural network construction method, and the convolutional neural network constructed according to the collected specific mechanical motion information has accurate and efficient identification capability.
In order to achieve the above object, the present application adopts a technical solution of a method for collecting and processing specific vibration information generated by a mechanism moving along a set path, where the specific vibration information is used to characterize a specific vibration detection characteristic generated by the mechanism moving along the set path, and the method includes: the method comprises the steps that a vibration sensor arranged on a mechanism is used for collecting specific vibration information generated when the mechanism moves along a set path in real time, so that vibration time domain information is obtained; dividing the vibration time domain information into segmented vibration time domain information which sequentially corresponds to different time periods, and ensuring that the mechanism motion path corresponding to each segmented vibration time domain information is equal in length; and respectively converting the segmented vibration time domain information into frequency domain information.
When acquired data information needs to be input into a Convolutional Neural Network (CNN) for training, the CNN requires that an input sample is translated and unchanged, but the motion speed of a mechanism is often unstable in practice, so that the vibration time domain information is converted into a displacement space domain and then is segmented.
By adopting the method, when the mechanism moves along a set path, the problems of inconsistent time domain length and unchanged input sample translation required by CNN caused by continuous change of the mechanism speed can be adaptively solved, and the length of a space window is set by dividing the vibration time domain information into the segmented vibration time domain information sequentially corresponding to different time periods and ensuring that the mechanism movement path corresponding to each segmented vibration time domain information is equal in length, so that the positioning resolution of the result obtained by inputting CNN can be customized.
Furthermore, the motion is non-uniform motion, that is, in reality, the motion of most objects and mechanisms is non-uniform motion, and the method can provide higher accuracy.
Further, the converting the time domain information of each segment of vibration into the frequency domain information respectively comprises: and carrying out Fourier transform on the segmented vibration time domain information to obtain frequency domain information.
Further, the method includes removing frequency domain data of frequency ranges which do not affect the specific vibration detection feature characterization from each frequency domain information to obtain optimized frequency domain information respectively.
Further, each of the optimized frequency domain information may be expressed as a frequency spectrum having the same length of the frequency axis. The spectral data in the spectrum may then be input to a convolutional neural network for training.
According to the method, the invention provides a method for collecting and processing specific vibration information of a rail transit vehicle, wherein the specific vibration information is used for representing specific vibration detection characteristics generated when the vehicle moves, and the specific vibration information generated when the vehicle moves along a rail is collected in real time through a vibration sensor arranged on the vehicle so as to obtain vibration time domain information; dividing the vibration time domain information into segmented vibration time domain information which sequentially corresponds to different time periods, and ensuring that the motion path of a carrier corresponding to each segmented vibration time domain information is equal in length; and respectively converting the segmented vibration time domain information into frequency domain information.
Unlike other mechanical structures, they tend to have a fixed rotational frequency, i.e. a constant speed, during a certain period of operation. The speed of the rail vehicle usually shows a non-steady characteristic in actual operation, and the speed is in oscillation change at most moments. If the vibration signal is still sliced and divided by a common fixed time window, the displacement of the train passing through the steel rail in the time window is different, which brings more challenge to accurate positioning of subsequent rail corrugation. Therefore, the technology proposes to convert the vibration time domain signal into a displacement space domain and then divide the displacement space domain.
The invention provides a convolutional neural network construction method for rail corrugation identification, which comprises the steps of collecting and processing specific vibration information of a rail transit carrier, removing frequency domain data in a frequency range which does not influence the specific vibration detection characteristic representation in each frequency domain information to respectively obtain optimized frequency domain information, expressing each optimized frequency domain information as a frequency spectrum with the same frequency coordinate axis length, and using the frequency spectrum data in the frequency spectrum as a training sample of a one-dimensional convolutional neural network.
Compared with the traditional method for detecting and identifying by adopting the corrugation tester CAT, the method has the advantages that after the convolutional neural network training, vibration information and vehicle displacement information are acquired, and the signal is input into a computer containing the specially trained convolutional neural network for processing, so that the rail corrugation can be identified quickly and efficiently.
Further, classifying and marking the training samples according to a set corrugation threshold value to obtain a steel rail normal state label and a steel rail corrugation state label.
Specifically, the classifying and marking of the training samples according to the set corrugation threshold value comprises the following steps:
and acquiring a rail corrugation signal and an axle box vibration signal, setting a corrugation threshold value to obtain a nonlinear mapping relation of the rail corrugation signal and the axle box vibration signal, and taking the nonlinear mapping relation as a training label of the one-dimensional convolutional neural network.
Further, a corrugation tester CAT is used for collecting corrugation signals of the steel rail at the same interval position, an acceleration sensor is used for collecting vibration signals of the axle box at the corresponding position, and the nonlinear mapping relation between the two signals is obtained through mathematical conversion.
The present invention further provides a method for obtaining a detection result according to specific vibration information generated by a vehicle moving on a track, the specific vibration information being used for characterizing a specific vibration detection characteristic generated when the vehicle moves on the track, including:
acquiring specific vibration information generated when the carrier moves along the track in real time through a vibration sensor arranged on the carrier so as to obtain vibration time domain information; dividing the vibration time domain information into segmented vibration time domain information which sequentially corresponds to different time periods, and ensuring that the motion path of a carrier corresponding to each segmented vibration time domain information is equal in length; respectively converting the segmented vibration time domain information into frequency domain information; and inputting the frequency domain information into the trained convolutional neural network to obtain a detection result. Thus, a more accurate inspection result can be obtained.
The invention also provides a method for obtaining a detection result according to the specific vibration information generated by the movement of the carrier on the track, which comprises the following steps:
and removing frequency domain data of a frequency range which does not influence the specific vibration detection characteristic representation in each frequency domain information to respectively obtain optimized frequency domain information, expressing each optimized frequency domain information as a frequency spectrum with the same frequency coordinate axis length, and processing the frequency spectrum data in the frequency spectrum by a trained convolutional neural network to obtain a detection result. Just as mentioned above, the characteristics of rail transit are the entering and leaving station, the variable speed of vehicle etc. and the non-uniform motion in the whole circuit, adopt this method especially suitable for to the collection of rail transit vibration time domain signal, handle, can be better input in the convolutional neural network, obtain higher degree of accuracy.
Therefore, the invention also provides a rail corrugation identification method, which comprises the steps of acquiring specific vibration information generated when the vehicle moves along the rail in real time through a vibration sensor arranged on the vehicle to obtain vibration time domain information; dividing the vibration time domain information into segmented vibration time domain information which sequentially corresponds to different time segments, and ensuring that the vehicle motion path corresponding to each segmented vibration time domain information is equal in length; respectively converting the segmented vibration time domain information into frequency domain information; and inputting the frequency domain information into a convolutional neural network model for rail corrugation identification to obtain a classification result of the convolutional neural network on the rail corrugation state. The method can still keep higher identification precision and is stable at 99.2% under complex operation conditions and speed time-varying working conditions of the vehicle.
Further, the construction of the convolutional neural network model for rail corrugation identification comprises: acquiring specific vibration information generated when a vehicle moves along a track in real time through a vibration sensor arranged on the vehicle to obtain vibration time domain information; dividing the vibration time domain information into segmented vibration time domain information which sequentially corresponds to different time segments, and ensuring that the vehicle motion path corresponding to each segmented vibration time domain information is equal in length; respectively converting the segmented vibration time domain information into frequency domain information;
removing frequency domain data in a frequency range which does not influence the specific vibration detection characteristic representation in each frequency domain information to respectively obtain optimized frequency domain information, expressing each optimized frequency domain information as a frequency spectrum with the same frequency coordinate axis length, and using the frequency spectrum data in the frequency spectrum as a training sample of a convolutional neural network; and obtaining the convolutional neural network model for rail corrugation recognition after training.
In the method for recognizing the subway rail corrugation based on the convolutional neural network from end to end, the method provides a method for cutting a virtual space paragraph with a certain length such as a space domain to manufacture a sample set of the convolutional neural network, and the length of the sample set can be set to arbitrarily adjust the positioning resolution during rail corrugation recognition, so that the method can still maintain higher recognition accuracy and is stable at 99.2% under complex operation conditions and speed time-varying working conditions of vehicles.
Further, the convolutional neural network model is a one-dimensional convolutional neural network model. The required requirements can be met by adopting a one-dimensional convolution neural network model, and the complexity is reduced as much as possible.
Further, the rail corrugation identification method comprises the following steps:
the vibration sensor is arranged on the vehicle and used for acquiring vehicle vibration information;
the displacement sensor is arranged on the vehicle and used for acquiring vehicle displacement information;
the data acquisition system is respectively connected with the vibration sensor and the displacement sensor and is respectively used for collecting vehicle vibration information acquired by the vibration sensor and vehicle displacement information acquired by the displacement sensor;
and the data processing device comprises a processing program of a convolutional neural network model for rail corrugation recognition, is connected with the data acquisition system, is used for processing data received by the data acquisition system, and is used for recognizing the rail corrugation state after the vehicle vibration information and the vehicle displacement information are processed by the convolutional neural network model.
Specifically, when the acceleration sensor collects signals, the acceleration sensor is arranged on an axle box of the vehicle to collect vehicle vibration information, so that information collection is facilitated.
Further, the displacement sensor is a photoelectric rotation speed sensor, and the photoelectric rotation speed sensor is arranged on the vehicle frame.
And further, a corrugation tester for acquiring the corrugation signal of the steel rail performs corrugation detection on the steel rail to be identified, and the corrugation detection result is compared with the classification result of the convolution neural network on the corrugation state of the steel rail.
Further, the corrugation tester is a CAT corrugation tester.
The invention is further described with reference to the following figures and detailed description. Additional aspects and advantages of the invention will be set forth in part in the description which follows and, in part, will be obvious from the description. Or may be learned by practice of the invention.
Drawings
The accompanying drawings, which are incorporated in and constitute a part of this specification, are included to assist in understanding the invention, and are included to explain the invention and their equivalents and not limit it unduly. In the drawings:
FIG. 1 is a schematic diagram for explaining corrugation of a subway rail;
FIG. 2 is a schematic diagram for explaining vibration testing of an axle box of a subway;
FIG. 3 is a schematic flow chart for illustrating a 1-DCNN-based rail transit rail corrugation identification method according to an embodiment;
FIG. 4 is a diagram illustrating a 1-DCNN model established in an embodiment;
FIG. 5 is a bar graph of test results of test times, precision, and elapsed time in the embodiments;
FIG. 6 is a line graph of training round number versus accuracy in an embodiment;
FIG. 7 is a line graph of training round number and error in an embodiment;
fig. 8 is a diagram for explaining the positioning time point in the spatial domain in the present embodiment;
FIG. 9 is a diagram illustrating the division of a vibration time domain signal at a time point obtained by the present method;
fig. 10 is a schematic diagram for explaining the present system for rail transit rail corrugation identification.
Detailed Description
The invention will be described more fully hereinafter with reference to the accompanying drawings. Those skilled in the art will be able to implement the invention based on these teachings. Before the present invention is described in detail with reference to the accompanying drawings, it is to be noted that:
the technical solutions and features provided in the present invention in the respective sections including the following description may be combined with each other without conflict.
Moreover, the embodiments of the present invention described in the following description are generally only examples of a part of the present invention, and not all examples. Therefore, all other embodiments obtained by a person of ordinary skill in the art based on the embodiments of the present invention without any creative effort shall fall within the protection scope of the present invention.
With respect to terms and units in the present invention. The term "comprises" and any variations thereof in the description and claims of this invention and the related sections are intended to cover non-exclusive inclusions.
As shown in fig. 1 to 10, in the present embodiment, the method of identifying a rail corrugation of a rail transit is taken as an example to describe the present invention, and firstly, a convolutional neural network model is constructed by using the convolutional neural network construction method for identifying a rail corrugation of the present invention, and the method is divided into five steps when performing state identification on a rail corrugation based on a one-dimensional convolutional neural network, i.e., 1-DCNN: (1) signal acquisition and 'space domain' cutting; (2) nonlinear mapping of vibration-corrugation signals; (3) establishing a sample set; (4)1-DCNN structure design and training; (5) and identifying the corrugation state of the steel rail.
The method adopts equipment, and comprises the following steps:
the vibration signal acquisition device is arranged on the vehicle and used for acquiring vehicle vibration information;
the displacement signal collector 2 is arranged on the vehicle and used for obtaining vehicle displacement information;
the data acquisition system 3 is respectively connected with the vibration signal collector 1 and the displacement signal collector 2, and is used for collecting the vehicle vibration information collected by the vibration signal collector 1 and the vehicle displacement information collected by the displacement signal collector 2; the data acquisition system 3 is a multi-channel data acquisition system;
and the data processing device 4 is connected with the data acquisition system 3, is used for processing data received by the data acquisition system 3, and is used for identifying the rail corrugation state after vehicle vibration information and vehicle displacement information are input as one-dimensional convolutional neural network training samples.
The system for identifying rail corrugation of rail transit further comprises a corrugation tester for collecting corrugation signals of the rail transit, and the corrugation tester is connected with the data processing device 4.
The vibration signal collector 1 is an acceleration sensor. The acceleration sensor is arranged on the axle box of the train. The axle box all sets up about the train acceleration sensor to all detect out the rail corrugation on both sides like this, guarantee vibration information is more perfect.
The displacement sensor is a photoelectric rotating speed sensor arranged on a vehicle frame. The photoelectric rotating speed sensor is a single-head reflection type photoelectric sensor, and an emitting end of the single-head reflection type photoelectric sensor faces one side of a vehicle wheel.
Firstly, signal acquisition and space domain cutting are carried out. (spatial domain, the space consisting of image elements in which the pixel values are directly processed with length (distance) as an argument is called spatial domain processing.)
The vibration signal is obtained by installing an acceleration sensor at the axle box of the train, as shown in fig. 2. Fig. 3 (a) shows the collected time domain signals of the vertical vibration of the axle box. Unlike other mechanical structures, they tend to have a fixed rotational frequency, i.e. a constant speed, during a certain period of operation. The speed of the rail vehicle usually shows a non-steady characteristic in actual operation, and the speed is in oscillation change at most moments. If the vibration signal is still sliced and divided by a common fixed time window, the displacement of the train passing through the steel rail in the time window is different, which brings more challenge to accurate positioning of subsequent rail corrugation. Therefore, the application proposes to convert the vibration time domain signal into the displacement space domain and then divide the displacement space domain.
First, the train speed is integrated to obtain a displacement time-varying graph shown in (b) of fig. 3, and a vibration time domain signal is introduced into a displacement space domain signal. Here, a coordinate system may be established, and the displacement space domain means that the abscissa is time in unit s and the ordinate is displacement in unit m. Often, when a vibration signal is processed, the change rule of displacement along with time cannot be analyzed, the invention provides that a virtual fixed space paragraph is applied to introduce a time point obtained on a space domain into a vibration time domain signal, and the vibration time domain signal is cut and divided to finally obtain a training sample set.
Next, referring to fig. 8, a "spatial domain" window X is set as a displacement paragraph. Sliding on the displacement space domain by the length of each Xm to locate the time, namely finding the time T corresponding to the 1 Xm, the 2 Xm, the1、T2、...、TM-1、TMWherein S is the total displacement,
is rounded up.
Again, referring to FIG. 9, the time series of nodes T derived from the "spatial domain1、T2、...、TM-1、TMAs a cut point, the original time domain signal is divided, and (c) in fig. 3 is a sample time domain signal cut by the spatial domain node.
Finally, as the train speed per hour is constantly changed, the length of each time domain signal after cutting is uneven, even differs by several orders of magnitude, which can cause great discount on the quality of sample data, and is very unfavorable for training of a deep learning network model. Meanwhile, the translation characteristic of the mechanical vibration time domain signal is one of the difficulties of feature extraction and identification classification, and the translation invariance is also a common obstacle in the research fields of sound source identification, image tracking and the like, so that the method has great significance in solving the problems. The Fourier transform can reduce the influence caused by translation after converting the time domain signal into the frequency domain, and is a simple and effective method. The performance of the convolution network under three different data types of the time domain, the frequency domain and the time-frequency combination domain is contrastively analyzed, and the test precision can be greatly improved when the frequency domain data is taken as a sample set. Therefore, fourier transform is performed on the divided time domain signal, and the frequency domain signal below a certain cutoff frequency N is focused on in a unified manner to obtain the frequency spectrum shown in (d) in fig. 3, so that the changes caused by stretching, translation and distortion of the original time domain signal are greatly reduced, and the original time domain signal is used as a sample input of the 1-DCNN for training.
In conclusion, the problems that the length of a time domain signal is inconsistent due to the fact that the train speed per hour is changed constantly and the input sample translation required by the CNN is unchanged are solved adaptively, and the resolution of rail corrugation state identification and positioning can be adjusted at will by setting the length of a space window.
Two, vibration-corrugation signal nonlinear mapping
Different from mechanical systems such as bearings and gear boxes, the mechanical systems have a fault diagnosis standard data set which is accepted by academia, such as bearing vibration data disclosed by the university of Keiss West reservoir (CWRU), gear box data provided by PHM2009challenge and rolling bearing accelerated life test data which is first published for the whole world in China, and at present, no accepted standard vibration data set under rail corrugation excitation exists in the world. In addition, due to the wavy wear characteristic of rail corrugation, the traditional method of introducing specified fault defects through the electric spark machining technology is not applicable any more, and then the method of acquiring corresponding vibration signals is not applicable any more. Therefore, the mapping relation between the rail corrugation and the vibration signal needs to be characterized through other ways, and the input sample must be matched with the corresponding label type by taking 1-DCNN as a supervised model. The search for an accurate mapping relation between the vibration-corrugation signals is crucial to the training of the 1-DCNN.
As shown in fig. 2, CAT is used to collect rail corrugation signals for the rails at the same interval position, an acceleration sensor is used to collect the axle box vibration signals at the corresponding position, and the mapping relationship between the two signals can be obtained through mathematical conversion. Fig. 3 (f) shows the resulting corrugation test signal in the time domain, and fig. 3 (e) shows the 1/3 octave spectrum of the corrugation. The root mean square value calculation is carried out on the 2.1 sections of vibration time domain signals subjected to space domain cutting, so that waveforms shown in a space domain (f) in the figure 3 are obtained, and the nonlinear mapping relation of vibration-corrugation signals can be obtained by setting a proper corrugation threshold value, so that the rail corrugation state can be distinguished.
The CAT rail corrugation measuring instrument adopts a manual measuring mode, is specially used for measuring the corrugation and the roughness of the rail. The measurement can be done on a track of infinite length, walking at a speed of 1m/s (3.6 km/h).
Thirdly, establishing a sample set:
the sample input data for CNN is essentially a computer recognizable number matrix. For the 1-DCNN proposed by the present invention, the sample set is a matrix of M × N, that is, all the sample frequency domain data after the space domain division. In deep learning, a sample set is generally divided into a training set and a test set for verifying the generalization ability of a model, and if p% in the sample set is the training set, the rest (1-p)% is used as the test set.
Design and training of four, 1-DCNN structure
The 1-DCNN framework proposed by the present invention is shown in (h) of FIG. 3. From top to bottom: the input layer is a sample set matrix; the first four layers have the same structure, are changed into a group of characteristic graphs through a ReLU activation function after being convolved by a small convolution kernel, and are sent to the maximum pooling layer for down-sampling; the fifth layer is large convolution kernel convolution, and aims to automatically learn local features facing diagnosis for the previous layer of input feature map; and flattening all characteristic diagrams of the last pooling layer to form a full connection layer, and transmitting the full connection layer to a last sigmoid classification layer after the full connection layer is processed by a suppression fitting technology Dropout. And when the large convolution kernel is applied to a vibration time domain signal, the large convolution kernel of the first layer is beneficial to filtering high-frequency noise pollution so as to capture the related characteristic information of the middle and low frequency bands. Different from the first layer of large convolution kernels, the input samples of the 1-DCNN model are frequency domain data, and if the first layer of convolution kernels is large, the coherence relation of each frequency band in the frequency domain is damaged. The first four layers of small convolution kernel structures deepen the network under fewer parameters, and meanwhile overfitting is restrained.
The fifth layer of large convolution kernel increases the feature expression capability of convolution kernel filtering extraction, restricts the internal covariate transfer of the network to a certain extent, and improves the identification precision and generalization capability of the network. In the CNN model, appropriate model hyper-parameters are selected to ensure that the model training speed is considered on the premise of higher identification precision. Aiming at the 1-DCNN model constructed by the method, the optimal setting of each hyper-parameter is searched by using a random search method, and finally the 1-DCNN parameter configuration shown in the table 1 is obtained. FIG. 4 shows the details of the 1-DCNN model structure under the parameter configuration of Table 1.
The first four small convolution kernels have a size of 3 x 1, the fifth large convolution kernel has a size of 64 x 1, and the step sizes are all 1 x 1. The number of convolution kernels in the five-layer convolution pooling structure is respectively 2, 4 and 8, and the sizes and the step sizes of the convolution kernels of the pooling layers are both 2 x 1. The convolutional layer uses the "same" zero complement command to make the convolutional layer input and output length equal. The network was trained using an "Adam" optimizer, the learning rate was set to 0.001 and the loss function was "binary _ cross". To avoid gradient dispersion and gradient explosion, training was performed using batch samples, with a batch size of 128. The Dropout and EarlyStopping technology can effectively improve the overfitting problem in the training process. In the present invention, Dropout is set to 0.5, and probability in the early-stop mechanism is set to 20, i.e., training is stopped when the accuracy of the test set is no longer improved within 20 rounds. Due to the one-dimensional nature, the input layer has only length components in three directions, with width and depth being 1. With the alternation of convolution pooling, the length of the feature map is gradually reduced, the depth is deepened, and the topological structure features of the input signals are mined layer by the network and self-learned.
TABLE 11 DCNN parameter configuration
The 1-DCNN model provided by the invention is established in a Keras deep learning library based on Python language. The PC hardware configuration of the data processing apparatus 4 employs the i7-8700 processor 16GB memory Windows10 system.
Fifth, rail corrugation state identification
And aiming at the identification of the corrugation state of the steel rail, defining the label of the normal steel rail as 0 and the label of the corrugation steel rail as 1. And if the output of the sigmoid activation layer is less than 0.5, judging that the sample label is 0, otherwise, judging that the sample label is 1. Through the operation, the rail corrugation state can be accurately identified and positioned.
The following examples of the embodiments are given
Description of (A) data
The method comprises the steps of carrying out field test on a subway line of a certain city in China, testing the irregularity of steel rails in a plurality of intervals of the line by using a corrugation tester CAT, and acquiring axle box vibration signals by using a B & K3560D multichannel data acquisition system. It should be noted that the test train wheel is in an initial turning state, and the influence sources of abnormal vibration of vehicle parts caused by wheel out-of-round, abrasion, flat scars and the like are recovered to be good, and the axle box vibration signal is mainly from the unsmooth excitation of the track.
(II) data set establishment
The vibration data obtained by field test can be directly used for training 1-DCNN after proper division. The vibration signal is divided into a space domain, a space domain window X is set to be 10m, a time domain signal when the train passes through a steel rail and moves every 10m is cut, then Fourier transform is carried out on the time domain signal, a frequency spectrum with the cutoff frequency N below 1024Hz is focused, and the frequency spectrum is converted into a digital matrix which can be identified by a computer. Statistical analysis is carried out on a large number of vibration signals and corrugation signals obtained by field test, the corrugation threshold value is set to 35 (unit is m/s2), namely the label is 0 when the RMS value of the time domain signal is less than 35, otherwise, the label is 1. Thereby obtaining the non-linear mapping relation between the two, and editing the label for the sample set data.
In summary, the label type of the sample set is determined by the time domain signal after "space domain" cutting, and the frequency domain signal is used as the sample input of the 1-DCNN. The field test data is processed by the method to obtain 13460 samples and corresponding labels thereof, wherein 90% is used as a training set, and 10% is used as a testing set.
(III) analysis of results
In order to reduce random errors in the training process, performance of the 1-DCNN model under 10 tests is observed. As shown in FIG. 5, the recognition accuracy of the test set and the testing time consumption of each sample are used as criteria to verify the model recognition efficiency and timeliness. It can be seen that the diagnostic rate is not lower than 99% under 10 tests, the time consumed by a single test sample is less than 0.2ms, and the 1-DCNN model can effectively, quickly and stably carry out intelligent identification and classification on rail corrugation under complex field operation conditions and train speed time-varying working conditions. Meanwhile, the prediction of the rail corrugation occurrence position is well matched with the actual situation on site, and the theory of a space domain is verified, namely the rail corrugation is intelligently identified and classified and the space position of the rail corrugation is accurately positioned.
Statistics of FIG. 5 shows that the precision variation range of 10 tests is 99.03% -99.33%, the average precision is 99.2%, and the standard deviation is 0.1; the time-consuming change only slightly fluctuates in millisecond units, and the requirement of online monitoring timeliness of rail corrugation is met. To further understand the performance of the 1-DCNN network, the 2 nd test and the 7 th test with the highest and the lowest precision are taken as examples respectively, and the evolution rule of the test precision and the error curve in the model training process is analyzed, as shown in fig. 6 and 7. The curve is analyzed, the accuracy and the error of the two tests are slightly different in value, the error of the 2 nd test with higher accuracy is lower, but the general trend of the curve change of the two tests is the same. Due to the introduction of an EarlySopping mechanism, the precision is optimal and is not improved any more after the number of training rounds reaches 44 times, and errors are converged to be below 0.06. The phenomena all accord with self-learning and cognition rules in the convolutional network training process, and the robust characteristic of the 1-DCNN method provided by the invention is further demonstrated by analyzing the precision and error curve change of boundary examples under 10 tests.
Taking a certain urban subway line from which experimental data is sourced as an example, the corrugation time of the steel rail in each section of the line is tested by using a corrugation tester CAT and is up to one month, but the total time required by adopting the method is not more than 3 hours, wherein the total time comprises the acquisition of vibration data and the prediction of a model, and a train in normal operation can be arranged for on-line monitoring without influencing the daily operation plan of the train. Aiming at the problem of online monitoring of the rail corrugation, which is urgently to be solved, the technical means provided by the practioner can greatly improve the production efficiency, save a large amount of manpower and material resource expenditure and provide a new solution for online monitoring of the rail corrugation.
The contents of the present invention have been explained above. Those skilled in the art will be able to implement the invention based on these teachings. All other embodiments, which can be derived by a person skilled in the art from the above description without inventive step, shall fall within the scope of protection of the present invention.
Claims (10)
1. The method for collecting and processing the specific vibration information generated by the mechanism moving along the set path, wherein the specific vibration information is used for representing the specific vibration detection characteristics generated when the mechanism moves along the set path, and the method is characterized by comprising the following steps of: the method comprises the steps that a vibration sensor arranged on a mechanism is used for collecting specific vibration information generated when the mechanism moves along a set path in real time, so that vibration time domain information is obtained; dividing the vibration time domain information into segmented vibration time domain information which sequentially corresponds to different time periods, and ensuring that the mechanism motion path corresponding to each segmented vibration time domain information is equal in length; respectively converting the segmented vibration time domain information into frequency domain information, and inputting the frequency domain information into a convolutional neural network for data processing;
the method comprises the steps of dividing vibration time domain information into segmented vibration time domain information which sequentially corresponds to different time segments, ensuring that a mechanism motion path corresponding to each segmented vibration time domain information is equal in length, integrating mechanism speed to obtain a displacement variation graph along with time, introducing vibration time domain signals into displacement space domain signals, sliding on the displacement space domain by the length of each Xm, and dividing original time domain signals by taking positioning time sequence nodes as cutting points.
2. The method of collecting and processing specific vibration information generated by a mechanism moving along a set path as recited in claim 1,
the motion is non-uniform motion.
3. The method of collecting and processing specific vibration information generated by a mechanism moving along a set path as recited in claim 1,
converting the segmented vibration time domain information into frequency domain information respectively comprises:
and carrying out Fourier transform on the segmented vibration time domain information to obtain frequency domain information.
4. The method of claim 1, including removing frequency domain data from each frequency domain that does not affect the frequency range of the specific vibration detection signature to obtain optimized frequency domain information.
5. The method of claim 4 wherein each of said optimized frequency domain information is expressed as a spectrum having the same length of the frequency axis.
6. The method for collecting and processing the specific vibration information of the rail transit vehicle, wherein the specific vibration information is used for representing the specific vibration detection characteristics generated when the vehicle moves, and the method is characterized by comprising the following steps: acquiring specific vibration information generated when the carrier moves along the track in real time through a vibration sensor arranged on the carrier so as to obtain vibration time domain information; dividing the vibration time domain information into segmented vibration time domain information which sequentially corresponds to different time periods, and ensuring that the motion path of a carrier corresponding to each segmented vibration time domain information is equal in length; respectively converting the segmented vibration time domain information into frequency domain information, and inputting the frequency domain information into a convolutional neural network for data processing;
the method comprises the steps of dividing vibration time domain information into segmented vibration time domain information which sequentially corresponds to different time segments, ensuring that the motion path of a carrier corresponding to each segmented vibration time domain information is equal in length, integrating the speed of the carrier to obtain a displacement variation graph along with time, introducing a vibration time domain signal into a displacement space domain signal, sliding on the displacement space domain by the length of each Xm, and dividing an original time domain signal by taking a positioning time sequence node as a cutting point.
7. The convolutional neural network construction method for rail corrugation identification, characterized by comprising the method for acquiring and processing specific vibration information of a rail transit vehicle as claimed in claim 6, removing frequency domain data in a frequency range which does not affect the specific vibration detection characteristic representation in each frequency domain information to obtain optimized frequency domain information respectively, expressing each optimized frequency domain information as a frequency spectrum with the same frequency coordinate axis length, and using the frequency spectrum data in the frequency spectrum as a training sample of a one-dimensional convolutional neural network.
8. The method for constructing a convolutional neural network for rail corrugation identification according to claim 7, wherein the training samples are classified and labeled according to a set corrugation threshold value to obtain a rail normal state label and a rail corrugation state label.
9. The convolutional neural network construction method for rail corrugation identification as claimed in claim 8, wherein the classifying and marking of the training samples according to the set corrugation threshold comprises the following steps:
and acquiring a rail corrugation signal and an axle box vibration signal, setting a corrugation threshold value to obtain a nonlinear mapping relation of the rail corrugation signal and the axle box vibration signal, and taking the nonlinear mapping relation as a training label of the one-dimensional convolutional neural network.
10. The method for constructing a convolutional neural network for rail corrugation identification according to claim 9, wherein a corrugation tester CAT is used for collecting rail corrugation signals for rails at the same interval position, an acceleration sensor is used for collecting axle box vibration signals at corresponding positions, and a nonlinear mapping relation between the rail corrugation signals and the axle box vibration signals is obtained through mathematical conversion.
Priority Applications (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN202010397122.4A CN111623868B (en) | 2020-05-12 | 2020-05-12 | Convolutional neural network construction method for rail corrugation identification |
Applications Claiming Priority (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN202010397122.4A CN111623868B (en) | 2020-05-12 | 2020-05-12 | Convolutional neural network construction method for rail corrugation identification |
Publications (2)
| Publication Number | Publication Date |
|---|---|
| CN111623868A CN111623868A (en) | 2020-09-04 |
| CN111623868B true CN111623868B (en) | 2021-08-20 |
Family
ID=72270968
Family Applications (1)
| Application Number | Title | Priority Date | Filing Date |
|---|---|---|---|
| CN202010397122.4A Active CN111623868B (en) | 2020-05-12 | 2020-05-12 | Convolutional neural network construction method for rail corrugation identification |
Country Status (1)
| Country | Link |
|---|---|
| CN (1) | CN111623868B (en) |
Families Citing this family (11)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN112113655B (en) * | 2020-09-21 | 2021-06-01 | 西南交通大学 | Vibration signal detection device and health assessment method for subway track fastener |
| CN112215263B (en) * | 2020-09-23 | 2022-08-02 | 西南交通大学 | Rail corrugation detection method based on deep convolutional neural network |
| CN112836272B (en) * | 2021-01-20 | 2023-09-01 | 北京交通大学 | High-speed railway steel rail profile optimization design method based on neural network model |
| CN113420367B (en) * | 2021-04-27 | 2024-01-19 | 北京九州一轨环境科技股份有限公司 | Subway steel rail wave grinding detection method and device based on vibration and noise response |
| CN113947130B (en) * | 2021-08-03 | 2022-06-14 | 西南交通大学 | Training and using method and device for regression prediction AI model of polygonal wear waveform of wheel |
| CN113947129B (en) * | 2021-08-03 | 2022-06-14 | 西南交通大学 | Method, equipment and medium for training and using AI model for intelligently identifying wheel out-of-roundness state |
| CN113486874B (en) * | 2021-09-08 | 2021-11-05 | 西南交通大学 | Rail corrugation feature identification method based on wheel-rail noise wavelet packet decomposition |
| CN114881076B (en) * | 2022-04-29 | 2023-04-18 | 西南交通大学 | Rail corrugation identification method, device, equipment and medium based on support vector machine |
| CN115600086B (en) * | 2022-11-15 | 2023-04-07 | 西南交通大学 | Vehicle-mounted quantitative detection method for rail corrugation roughness based on convolution regression |
| CN115758289B (en) * | 2023-01-10 | 2023-04-25 | 西南交通大学 | Rail wave mill identification method based on multitask learning neural network |
| CN117235507B (en) * | 2023-11-13 | 2024-01-23 | 成都西交轨道交通技术服务有限公司 | Method for identifying abnormal state of grinding stone of steel rail grinding wagon |
Citations (3)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN110175434A (en) * | 2019-06-19 | 2019-08-27 | 西南交通大学 | A kind of rail fastener system injury detection method based on convolutional neural networks |
| CN110658005A (en) * | 2019-10-08 | 2020-01-07 | 西南交通大学 | Method for identifying rail corrugation diseases based on vehicle body acceleration |
| CN110806324A (en) * | 2019-11-11 | 2020-02-18 | 成都西交智众科技有限公司 | Wheel polygon abrasion detection method based on rail displacement and data acquisition equipment |
Family Cites Families (10)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| JP2000136988A (en) * | 1998-10-30 | 2000-05-16 | East Japan Railway Co | Detection method for wave-shaped abrasion of rail |
| CN101850772A (en) * | 2010-05-17 | 2010-10-06 | 唐德尧 | Vehicular monitoring device and monitoring method thereof for rail corrugation |
| JP5433516B2 (en) * | 2010-07-12 | 2014-03-05 | 公益財団法人鉄道総合技術研究所 | Rail wavy wear detection method and rail wavy wear detection system |
| CN103264712B (en) * | 2013-06-03 | 2016-03-30 | 北京绿创声学工程股份有限公司 | A kind of Rail corrugation detection method |
| CN105608312B (en) * | 2015-12-16 | 2018-11-16 | 广州地铁集团有限公司 | A kind of wave mill evaluation method instructing wave mill orbital maintenance |
| CN106326929B (en) * | 2016-08-24 | 2019-08-02 | 北京工商大学 | The diagnostic method of track train EEF bogie rolling bearing fault |
| CN107192763B (en) * | 2017-04-17 | 2019-06-28 | 北京交通大学 | Utilize the test method for the rail vibration attenuation rate that running train is excitation |
| CN108830127B (en) * | 2018-03-22 | 2021-08-10 | 南京航空航天大学 | Rotary machine fault feature intelligent diagnosis method based on deep convolutional neural network structure |
| CN111071291B (en) * | 2018-10-19 | 2022-03-18 | 株洲中车时代电气股份有限公司 | Train wheel set monitoring system and train wheel set monitoring method |
| CN110930357B (en) * | 2019-10-17 | 2023-11-14 | 中国矿业大学 | In-service steel wire rope surface defect detection method and system based on deep learning |
-
2020
- 2020-05-12 CN CN202010397122.4A patent/CN111623868B/en active Active
Patent Citations (3)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN110175434A (en) * | 2019-06-19 | 2019-08-27 | 西南交通大学 | A kind of rail fastener system injury detection method based on convolutional neural networks |
| CN110658005A (en) * | 2019-10-08 | 2020-01-07 | 西南交通大学 | Method for identifying rail corrugation diseases based on vehicle body acceleration |
| CN110806324A (en) * | 2019-11-11 | 2020-02-18 | 成都西交智众科技有限公司 | Wheel polygon abrasion detection method based on rail displacement and data acquisition equipment |
Also Published As
| Publication number | Publication date |
|---|---|
| CN111623868A (en) | 2020-09-04 |
Similar Documents
| Publication | Publication Date | Title |
|---|---|---|
| CN111623868B (en) | Convolutional neural network construction method for rail corrugation identification | |
| CN111637964B (en) | Rail corrugation identification method | |
| Islam et al. | Automated bearing fault diagnosis scheme using 2D representation of wavelet packet transform and deep convolutional neural network | |
| Cao et al. | Unsupervised domain-share CNN for machine fault transfer diagnosis from steady speeds to time-varying speeds | |
| US11761930B2 (en) | Prediction method of part surface roughness and tool wear based on multi-task learning | |
| CN111947928B (en) | Multi-source information fusion bearing fault prediction system and method | |
| CN111896625B (en) | Rail damage real-time monitoring method and monitoring system thereof | |
| CN113469060A (en) | Multi-sensor fusion convolution neural network aeroengine bearing fault diagnosis method | |
| Dou et al. | A rule-based intelligent method for fault diagnosis of rotating machinery | |
| CN110370649A (en) | On-Line Monitor Device, the system of 3D printing equipment | |
| Li et al. | Joint attention feature transfer network for gearbox fault diagnosis with imbalanced data | |
| CN107473036A (en) | Elevator traction machine remote online detects and diagnostic system and its checkout and diagnosis method | |
| CN107844067A (en) | A kind of gate of hydropower station on-line condition monitoring control method and monitoring system | |
| CN112364706A (en) | Small sample bearing fault diagnosis method based on class imbalance | |
| CN110631832A (en) | Subway vehicle bearing fault on-line detection method | |
| CN113988326A (en) | Subway equipment maintenance optimization method and system | |
| CN115238785A (en) | Rotary machine fault diagnosis method and system based on image fusion and integrated network | |
| CN115563553A (en) | Aviation bearing fault diagnosis method of deep migration map convolution network under fluctuation working condition | |
| CN114700587B (en) | Missing welding defect real-time detection method and system based on fuzzy inference and edge calculation | |
| CN114755017B (en) | Variable-speed bearing fault diagnosis method of cross-domain data driving unsupervised field shared network | |
| CN116861369A (en) | Air compressor health assessment method and system for industrial internet heterogeneous data fusion | |
| CN111597948A (en) | Fault detection and classification method for bearing vibration signals | |
| CN111401661A (en) | Predictive maintenance method and maintenance system for mechanical equipment | |
| CN114714145A (en) | Method for enhancing, comparing, learning and monitoring tool wear state by using Gelam angular field | |
| CN116644278A (en) | Sample screening and feature extraction-based dual-model bearing anomaly detection algorithm |
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 | ||
| GR01 | Patent grant | ||
| GR01 | Patent grant |