Train wheel state evaluation method

Abstract

The invention discloses a train wheel state evaluation method. The method comprises the steps that S1, dynamic strain data are collected through fiber grating sensors arranged on the bottom surface of a steel rail at equal intervals; s2, inputting the collected dynamic strain data into the trained nonlinear autoregressive network model to obtain prediction time domain data; s3, judging whether the nonlinear autoregressive network model predicts the time domain data to deviate or not; and S4, after the abnormal value is extracted, the abnormal part of the wheel is deduced according to the occurrence position of the abnormal value. According to the invention, flaw detection is carried out on the train wheel set through the position structure design of the fiber grating sensor, and whether the dynamic response signal has an abnormal value or not is evaluated through the nonlinear autoregression network model, so that the state of the train wheel is efficiently identified, and accurate evaluation is made.

Description

TECHNICAL FIELD

[0001] The present application belongs to the technical field of train wheel health state evaluation, and particularly relates to a train wheel state evaluation method. BACKGROUND

[0002] As the core carrier and operation basis of the railway system, the health condition of the train is directly related to the safety and comfort of the railway system. However, during the long-term service of the train wheelset, various disease problems will occur due to the wheel-rail interaction and environmental factors. In order to prevent train wheelset damage from causing railway safety accidents and property losses, the railway operation and maintenance department invests a large amount of manpower and material resources in offline detection and turning repair of the vehicle wheelset every year, and detailed technical specifications and implementation rules are formulated. However, railway safety accidents caused by wheelset damage still occur from time to time.

[0003] The current industry generally implements the "planned repair" strategy, that is, the train is periodically sent to the car plant for offline repair. The "planned repair" has the disadvantages of low detection efficiency, time-consuming and laborious, and inability to reflect the health condition and damage development of the wheelset in real time, and therefore an online monitoring means is needed to realize the "state repair" of the train wheelset, that is, the real-time state of the wheelset is used to guide the offline maintenance and turning repair. The problems of local non-circularity (flat scar or scratch), polygon (periodic non-circularity), abnormal vibration and the like of the wheel are solved. SUMMARY

[0004] In view of the problems of low detection efficiency, time-consuming and laborious, and inability to reflect the health condition and damage development of the wheelset in real time existing in the prior art train wheel which is periodically sent to the car plant for offline repair, the present application provides a train wheel state evaluation method.

[0005] To achieve the above technical purposes, the technical scheme adopted by the present application is as follows:

[0006] A train wheel state evaluation method, comprising the steps of:

[0007] S1, collecting dynamic strain data through the optical fiber grating sensor arranged at equal intervals on the rail bottom surface;

[0008] S2, inputting the collected dynamic strain data into the trained nonlinear autoregressive network model to obtain predicted time domain data;

[0009] S3, judging whether the predicted time domain data of the nonlinear autoregressive network model deviates;

[0010] S4, after extracting the abnormal value, deducing the abnormal part of the wheel according to the position of the abnormal value.

[0011] Further, the dynamic strain data is the bending stress generated by the rail under the action of the wheel-rail force; the detailed steps of collecting the dynamic strain data include:

[0012] The strain monitoring sensor array based on fiber grating is configured in series;

[0013] When the wheel acts on the monitoring section, the rail is subjected to bending stress under the wheel-rail force;

[0014] The above bending stress is collected by the fiber grating sensors arranged at equal intervals on the rail bottom surface.

[0015] Further, the length of the fiber grating sensor array is greater than 3 meters and greater than the circumference of the vehicle wheel set, and the distance between the single sensors is 0.15 meters. The spatial resolution of the measurement is ensured.

[0016] Further, the training step of the nonlinear autoregressive network model:

[0017] According to the time domain data corresponding to each fiber grating sensor;

[0018] The time domain data collected by two fiber grating sensors on the same position and different rails are used to train the nonlinear autoregressive network;

[0019] One of the sensor signals is used as the original signal for autoregressive prediction, and the other sensor signal is used as the external input to assist the autoregressive prediction of the original signal.

[0020] Further, the nonlinear autoregressive network model calculation formula is:

[0021]

[0022] Where F is a nonlinear function, y and are the original and predicted time domain signals, x is the time domain signal of the external input, d x and d y are the time delay steps of the external input and the original signal, W h and b h are the weights and bias terms in the network parameters, f is the activation function, N is the neuron depth / number, and t is the time.

[0023] Further, the detailed steps for judging whether the nonlinear autoregressive network model prediction time domain data deviates or not:

[0024] According to the network residual e b should conform to a certain normal distribution G(e b ), and the residual e u of the collected data after inputting the network should conform to another normal distribution G(e u );

[0025] By comparing the two distribution conditions, it is determined whether the signal is abnormal under the unknown working condition;

[0026] When the signal is abnormal, the prediction result of the nonlinear autoregressive network deviates, that is, G(e u ) deviates from G(e b ).

[0027] Further, the formula for determining whether the prediction time domain data deviates is:

[0028]

[0029] Where f b and f u are the peak values of the network residual probability density function, and σ b and σ u are the standard deviations of the two distributions.

[0030] Generally, the 95% confidence interval of the index can be obtained by multiple network training, and the index value outside the interval can be regarded as abnormal. The following explains the process of obtaining the confidence interval: under the prior condition that the train wheelset is healthy, n groups of dynamic strain data are collected and the index distribution is calculated. Assuming that is the average value of the index calculated n times, and σ is the standard deviation, then the upper and lower limit calculation formula of the index for determining whether the signal is abnormal is:

[0031]

[0032] Each peak value corresponds to a specific wheel of the train. Further, after the signal is abnormal, the wheel is determined according to the abnormal position, and the determination formula of whether the train wheelset is damaged is:

[0033]

[0034] Compared with the prior art, the present application has the following beneficial effects:

[0035] By designing the position structure of the fiber grating sensor to detect the train wheelset, and by using the nonlinear autoregressive network model to evaluate whether the dynamic response signal is abnormal, the state of the train wheel can be efficiently identified and accurately evaluated. BRIEF DESCRIPTION OF DRAWINGS

[0036] Figure 1 is a whole flow chart of a train wheel state evaluation method in an embodiment of the present application;

[0037] Figure 2 is a network prediction residual identification diagram in an embodiment of the present application;

[0038] Figure 3 Fig. 1 is a schematic diagram of the principle of monitoring the strain of a steel rail by using the fiber grating technology in an embodiment of the present application;

[0039] Figure 4 Fig. 2 is a side view and a sectional view of a rail load monitoring arrangement in an embodiment of the present application. DETAILED DESCRIPTION

[0040] For the convenience of those skilled in the art, the present application will be further described below in conjunction with the embodiments and the accompanying drawings. The content mentioned in the embodiments is not a limitation of the present application.

[0041] As shown in Fig. 1, the present embodiment provides a train wheel state evaluation method, which comprises the following steps: Figure 1

[0042] S1, collecting dynamic strain data by using fiber grating sensors which are equidistantly arranged on the rail bottom surface;

[0043] S2, inputting the collected dynamic strain data into a trained nonlinear autoregressive network model to obtain predicted time domain data;

[0044] S3, judging whether the predicted time domain data of the nonlinear autoregressive network model is deviated or not;

[0045] S4, after extracting the abnormal value, deducing the abnormal part of the wheel according to the position of the abnormal value.

[0046] As shown in Fig. 1, the dynamic strain data is the bending stress generated by the action of wheel-rail force on the rail; the detailed steps of collecting the dynamic strain data include: Figure 3

[0047] The fiber grating-based strain monitoring sensor array is constructed in a series mode;

[0048] When the wheel acts on the monitoring section, the rail will generate bending stress under the action of wheel-rail force;

[0049] The above bending stress is collected by the fiber grating sensors which are equidistantly arranged on the rail bottom surface.

[0050] As shown in Fig. 1, the length of the fiber grating sensor array is greater than 3 meters and greater than the circumference of the wheelset of the vehicle, and the distance between the single sensors is 0.15 meters. The spatial resolution of the measurement is ensured. Figure 4 The training steps of the nonlinear autoregressive network model include:

[0051] According to the time domain data corresponding to each fiber grating sensor;

[0052]

[0053] ​​​The time domain data collected by two fiber grating sensors on the same location but different steel rails are used to train the nonlinear autoregressive network;

[0054] One of the sensor signals is used as the original signal for autoregressive prediction, and the other sensor signal is used as an external input to assist the autoregressive prediction of the original signal.

[0055] The calculation formula of the nonlinear autoregressive network model is:

[0056]

[0057] Where F is a nonlinear function, y and are the original and predicted time domain signals, x is the time domain signal of the external input, d x and d y are the time delay steps of the external input and the original signal, W h and b h are the weights and bias terms in the network parameters, f is the activation function, N is the depth / number of neurons, and t is the time.

[0058] The detailed steps for determining whether the predicted time domain data of the nonlinear autoregressive network model deviates are as follows:

[0059] According to the trained network residual e b should conform to a certain normal distribution G(e b ), and the residual e u of the collected data after inputting the network should conform to another normal distribution G(e u );

[0060] By comparing the two distribution conditions, it is determined whether the signal under the unknown working condition is abnormal;

[0061] When the signal is abnormal, the prediction result of the nonlinear autoregressive network will deviate, i.e., G(e u ) deviates from G(e b ).

[0062] As shown in FIG. Figure 2 , the formula for determining whether the predicted time domain data deviates is:

[0063]

[0064] Where f b and f u are the peak values of the network residual probability density functions, σ b and σ u are the standard deviations of the two distributions.

[0065] The 95% confidence interval of the index can be obtained by multiple network training, and the index value outside the interval can be regarded as an anomaly. The process of obtaining the confidence interval is explained as follows: under the prior condition that the train wheelset is known to be healthy, n groups of dynamic strain data are collected and the index distribution is calculated. Assuming is the average value of the index calculated n times, and sigma is the standard deviation, then the upper and lower limits of the index for judging whether the signal is abnormal are calculated as follows:

[0066]

[0067] Each peak corresponds to a specific wheel of the train. After the signal appears abnormal, the wheel is determined according to the position where the abnormality occurs, and the formula for determining whether the train wheelset is damaged is as follows:

[0068]

[0069] Compared with the prior art, the present application has the following beneficial effects:

[0070] The fiber grating sensor position structure is designed to detect the train wheelset, and the nonlinear autoregressive network model is used to evaluate whether the dynamic response signal is abnormal, so as to realize efficient identification of the train wheel state and accurate evaluation.

[0071] The above describes in detail a train wheel state evaluation method provided by the present application. The description of the specific embodiments is only used to help understand the method and its core idea. It should be pointed out that for those skilled in the art, without departing from the principles of the present application, the present application can be improved and modified in several ways, and these improvements and modifications also fall within the protection scope of the claims of the present application.

Claims

1. A method of evaluating a state of a train wheel, characterized by, The method comprises the steps of: S1, collecting dynamic strain data through fiber grating sensors arranged at equal intervals on the rail bottom surface; S2, inputting the collected dynamic strain data into a trained nonlinear autoregressive network model to obtain predicted time domain data; S3, judging whether the predicted time domain data of the nonlinear autoregressive network model deviates; S4, after extracting the abnormal value, deducing the abnormal position of the wheel according to the position of the abnormal value.

2. The method of claim 1, wherein The dynamic strain data is the bending stress generated by the action of wheel-rail force on the rail; The detailed steps of collecting dynamic strain data include: A fiber grating-based strain monitoring sensor array is constructed in a series mode; When the wheel acts on the monitoring section, the rail will generate bending stress under the action of wheel-rail force; The above bending stress is collected through fiber grating sensors arranged at equal intervals on the rail bottom surface.

3. The method of claim 2, wherein The length of the fiber grating sensor array is greater than 3 meters and greater than the circumference of the vehicle wheelset, and the distance between single sensors is 0.15 meters.

4. The method of claim 3, wherein The training steps of the nonlinear autoregressive network model include: According to a section of time domain data corresponding to each fiber grating sensor; The time domain data collected by two fiber grating sensors at the same position on different rails are used to train the nonlinear autoregressive network; One of the sensor signals is used as the original signal for autoregressive prediction, and the other sensor signal is used as the external source input to assist the autoregressive prediction of the original signal.

5. The method of claim 4, wherein The calculation formula of the nonlinear autoregressive network model is: where F is a non-linear function, y and are the original and predicted time-domain signals, respectively, x is an exogenous input time-domain signal, d x and d y are the time-lagged steps of the exogenous input and original signals, respectively, W h and b h are the weight and bias terms in the network parameters, respectively, f is an activation function, N is the depth / number of neurons, and t is time.

6. The method of claim 5, wherein The detailed steps of judging whether the predicted time domain data of the nonlinear autoregressive network model deviates include: According to the trained network, the residual error e b should be subject to a certain normal distribution G(e b ), and the residual error e u of the collected data after inputting the network should be subject to another normal distribution G(e u ). By comparing the two distribution conditions, it is judged whether the signal under the unknown working condition is abnormal; When the signal is abnormal, the prediction result of the nonlinear autoregressive network deviates, that is, G(e u ) deviates from G(e b ).

7. The method of claim 6, wherein The formula for judging whether the predicted time domain data deviates is: where f b and f u are the peak values of the network residual probability density functions, σ b and σ u are the standard deviations of the two distributions, respectively.

8. The method of claim 7, wherein After the signal appears abnormal, the judgment formula for determining whether the wheel and train wheelset are damaged according to the position of the abnormality is:

Images