Bogie abnormality detection system and bogie abnormality detection method

Abstract

In order to provide a bogie abnormality detection system capable of simply detecting on the vehicle side, without depending on an acoustic sensor installed on the ground, an abnormality of a vehicle which is in a noisy traveling state, the bogie abnormality detection system that detects an abnormality of a bogie of the vehicle traveling on a track comprises: a vibration sensor or an acoustic sensor that senses the vibration of the bogie and outputs vibration data of the bogie; and a vehicle information control device that uses the vibration data and the traveling state information of the vehicle as inputs. The vehicle information control device has: an extraction unit that extracts, from the vibration data, on the basis of the traveling state information, the vibration data at least either during linear zone traveling or during coasting operation as vibration information; and a detection unit that detects, using a predictive model calculated by machine learning in which the vibration information and abnormality information of the bogie are used as training data, an abnormality of the bogie from the vibration information extracted by the extraction unit.

Description

Bogie abnormality detection system and bogie abnormality detection method

[0001] The present invention relates to a bogie abnormality detection system and a bogie abnormality detection method when an abnormality occurs in a bogie of a railway or track vehicle.

[0002] Traditionally, visual inspections and hammering tests have been used to check for abnormalities such as cracks, chips, breaks, and peeling in railway vehicle bogies. However, these inspections, which are performed while the vehicle is stationary, involve striking areas that are different from the vibrating points (places where loads are applied) when the vehicle is in motion, so they do not allow for observation of the condition of the vehicle while it is in motion. Furthermore, because the inspections are performed by humans, there is inevitably variation in accuracy.

[0003] In order to solve this problem, Non-Patent Document 1 and Patent Document 1 disclose a technique for detecting abnormalities in bogies using acoustic sensors installed at the side of the tracks.

[0004] JP 2020-73366 A JP 2019-167079 A

[0005] https: / / www.nikkei.com / articleDGXZQOHD225BY0S0A221C2000000 / Nikkei Newspaper JR West uses AI to distinguish between Shinkansen train running sounds and prevent accidents

[0006] The method disclosed in Patent Document 1 is limited to detecting abnormalities in the bogie of a vehicle that passes near the acoustic sensor, and detection using an acoustic sensor can only be applied to lines with relatively little noise, such as elevated Shinkansen tracks, making it difficult to apply to local lines in rural areas.

[0007] On the other hand, in the method disclosed in Patent Document 2, the thrust of the actuator is changed in the straight section / curve section by vibrations at the resonant frequency of the bogie, but simply combining the methods disclosed in these two patent documents does not result in removing the vibration data in the curved section, which is noisy, and it is difficult to detect abnormalities in the bogie.

[0008] The present invention has been made in consideration of the above circumstances, and aims to provide a technology for detecting abnormalities in a vehicle while it is running without relying on acoustic sensors installed on the ground.

[0009] In order to solve the above problems, one representative bogie abnormality detection system according to the present invention is a system that detects abnormalities in the bogie of a vehicle traveling on a track, and includes a vibration sensor or acoustic sensor that senses the vibrations of the bogie and outputs vibration data of the bogie, and a vehicle information control device that receives as input the vibration data and vehicle running state information.The vehicle information control device has an extraction unit that extracts, as vibration information, vibration data from the vibration data at least one of vibration data when traveling on a straight section or when coasting, based on the running state information, and a detection unit that detects abnormalities in the bogie from the vibration information extracted by the extraction unit, using a prediction model calculated by machine learning using the vibration information and bogie abnormality information as training data.

[0010] According to the present invention, it is possible to improve the accuracy of detecting abnormalities in the bogie by distinguishing between coasting and running on a straight section and filtering vibration frequencies, and easily separating only characteristic vibration frequencies using a prediction model calculated by machine learning. Problems, configurations, and effects other than those described above will become clear from the description of the following embodiments of the invention.

[0011] Fig. 1 is a diagram showing an example of the overall configuration of a bogie anomaly detection system according to an embodiment of the present invention. Fig. 2 is a diagram showing an example of the configuration of a vehicle information control device and its relationship with other devices. Fig. 3 is a diagram showing an example of the configuration of a prediction data calculation device and its relationship with other devices. Fig. 4 is a diagram showing an example of the configuration of machine learning by Deep Learning using a sigmoid function. Fig. 5 is a diagram showing a calculation graph of the multiplication part. Fig. 6 is a diagram showing a calculation graph of the addition part. Fig. 7 is a diagram showing a calculation graph of the sigmoid part. Fig. 8 is a diagram showing a spectrogram of sound data collected inside a vehicle when it is actually running between stations.

[0012] Hereinafter, embodiments for carrying out the present invention will be described as examples with reference to the drawings. Note that the present invention is not limited to these examples. In addition, in the description of the drawings, the same parts are designated by the same reference numerals.

[0013] 1 is a diagram showing an example of the overall configuration of a bogie abnormality detection system according to an embodiment of the present invention. First, the processing mode up to the transmission of information from the on-board side to the ground side will be described.

[0014] A vibration sensor or acoustic sensor S2 mounted on the vehicle S1 senses the vibration of the bogie, acquires vibration data, and outputs (transmits) the data to the vehicle information control device S4.

[0015] The driving information transmitting device S3 acquires information on the vehicle's driving state, such as information during driving, speed, wheel rotation speed, GPS information, signals from ground coils, position information that can identify the vehicle's driving position such as distance in kilometers, the state of the main controller (powering, coasting, braking, etc.), driving state information of the vehicle motor drive device, and braking state information of the brakes, and outputs (transmits) this information to the vehicle information control device S4.

[0016] The vehicle information control device S4 filters the vibration data input from the vibration sensor or acoustic sensor S2 based on the driving state information received from the driving information transmission device S3.

[0017] During this filtering, if the driving condition information includes information on at least one of the time period and section in which vehicles pass each other, the vibration data acquired during that time period and section will include vibration data caused by the vehicles passing each other, and by removing this data through filtering, the effects of vibrations caused by the vehicles passing each other can be eliminated.

[0018] The vibration information obtained by filtering is transmitted together with the driving state information to the ground-side communication device S6 via the vehicle-side communication device S5.

[0019] Next, a description will be given of the processing performed by the ground side until the information received is processed and transmitted to the on-board side. The ground side communication device S6 transmits the information received from the vehicle side communication device S5 to the prediction data calculation device S7.

[0020] The prediction data calculation device S7 manages information for each vehicle, and if a defect or the like is discovered in a bogie as a result of maintenance performed at the vehicle depot S8, it stores information relating to the defect and performs machine learning with information received from the ground-side communication device S6. The machine learning information performed by the prediction data calculation device S7 is transmitted to the vehicle-side communication device S5 via the ground-side communication device S6.

[0021] Next, the on-board processing based on the information received from the ground side will be described. The vehicle-side communication device S5 transmits the machine learning information received from the ground-side communication device S6 to the vehicle information control device S4.

[0022] The vehicle information control device S4 filters the vibration data received from the vibration sensor or acoustic sensor S2 using the driving status information received from the driving information transmission device S3, and performs calculations using the information from the machine learning performed on the ground side. If an abnormality is found as a result of the calculations, the vehicle information control device S4 notifies (announces) the abnormality to the crew S10 in the vehicle S1 or the maintenance worker S11 on the ground side.

[0023] Next, the operation of the onboard vehicle information control device S4 will be described. Fig. 2 is a diagram showing an example of the configuration of the vehicle information control device S4 and its relationship with other devices. The driving state calculation unit S44 determines whether the vehicle is coasting or traveling on a straight section based on the driving state information received from the driving information transmission device S3, and outputs determination information.

[0024] Regarding the determination of coasting, not only coasting but also a section where the driving state information of the drive device of the vehicle motor is equal to or less than a predetermined threshold may be determined as coasting. Furthermore, regarding the determination of running on a straight section, even if the section is not completely straight, a threshold may be set for the curvature to determine the running state.

[0025] This determination information is stored in the recording unit S45 together with information on the running state such as speed and acceleration, and is also sent to the vibration extraction unit S41.

[0026] The vibration extraction unit S41 filters the vibration data using the vibration data from the vibration sensor or acoustic sensor S2 and the discrimination information from the running state calculation unit S44. Through this filtering, vibration data from at least one of coasting and running on a straight section is extracted as vibration information, and transmitted to the prediction model confirmation unit S42 and the recording unit S45. This vibration information is stored in the recording unit S45.

[0027] Meanwhile, the prediction model storage unit S46 stores the prediction model received from the ground-side prediction data calculation device S7 via the ground-side communication device S6 and the vehicle-side communication device S5 in the recording unit S45. The prediction model storage unit S46 also reads out the running state information and vibration information from the recording unit S45 and transmits them to the ground-side communication device S6 via the vehicle-side communication device S5.

[0028] The prediction model confirmation unit S42 reads out the prediction model and running state information stored in the recording unit S45, and determines whether the vibration state of the bogie is abnormal by combining this with the vibration information received from the vibration extraction unit S41, and detects an abnormality. If an abnormality is detected in the bogie, the abnormality detection is notified to the crew etc. via the abnormality state display unit S43.

[0029] Here, the criterion for detecting an abnormality in the bogie may be whether the bogie has been determined to be abnormal for a specified number of times or whether the bogie has been determined to be abnormal for a certain period of time. An example of calculation in the prediction model confirmation unit S42 will be described in <Example of calculation using Deep Learning> below.

[0030] Next, the operation of the ground-side predicted data calculation device S7 will be described. Fig. 3 is a diagram showing an example of the configuration of the predicted data calculation device S7 and its relationship with other devices. The running state information and vibration information received from the ground-side communication device S6 are stored for each vehicle in a recording unit S75 via a received data storage unit S74.

[0031] Furthermore, inspection information such as which vehicle's bogie has an abnormality such as a crack, fissure, or break is stored in the recording unit S75 from the vehicle depot S8 via the storage data input unit S72.

[0032] The prediction model calculation unit S73 constructs a prediction model using the inspection information, running state information, and vibration information for each vehicle stored in the recording unit S75. An example of constructing a prediction model will be described later in <Example of calculation using Deep Learning>. The constructed prediction model is stored in the recording unit S75.

[0033] The prediction model transmission unit S71 transmits the prediction model stored in the recording unit S75 to the vehicle-side communication device S5 via the ground-side communication device S6.

[0034] <Calculation Example Using Deep Learning> As an example of a determination method using machine learning, a calculation example using Deep Learning will be described. Here, only a portion of the various information listed above may be used, or information limited to driving on a straight section may be used. Furthermore, the method is not limited to the Deep Learning method, and methods such as Bayesian statistics using conjugate distribution or MCMC methods, or SVM or t-tests that can statistically distinguish between normal and other data may also be used. Furthermore, in this embodiment, a sigmoid function is used as a calculation example using Deep Learning, but the method is not limited to this and other methods such as ReLU may also be used.

[0035] 4 is a diagram showing an example of the configuration of machine learning by deep learning using a sigmoid function. First, in an input data layer S91, information that serves as a reference when generating vibration information is input.

[0036] An example of input information is shown in Table 1. In the example shown in Table 1, for frequency X (X=1, 2, 3, ...), the vibration information is Fourier transformed, and the amplitude of the vibration divided into certain frequency intervals is used as input data.

[0037] In Table 1, acceleration, speed, curvature of the travel section, and gradient data are input as additional data. If vibration information is included as base data, only part of Table 1 may be added.

[0038] In addition to the information shown in Table 1, other information that may be useful when vibration occurs, such as other driving state information, occupancy rate, wind speed, wind direction (including vector decomposition to the front and sides of the vehicle), humidity, temperature, rainfall, etc. Unrelated data may be added as long as it does not affect the calculation results.

[0039] Next, the normalization layer S92 normalizes the input data. During normalization, the largest data in the input data (hereinafter referred to as "input vector") is set to 1, and normalization may be performed based on the relative magnitude or the maximum possible value for each data. In this embodiment, the normalized input data matrix is ​​defined as X = (x1, x2, x3, ... xn).

[0040] Next, in the Affine layer S93, the matrix W of the weighted signal is Then, let the bias matrix B be B = (b1, b2, b3, ..., bm) and the elements of the result matrix A be A = (a1, a2, a3, ..., am), then the result matrix A can be calculated by the following formula: A = XW + B

[0041] For example, if we take one of the elements of the result matrix A, we get a1=w11*x1+w12*x2+ . . . +w1n*xn+b1.

[0042] The values ​​of the matrices W and B that make up the above equations may be preset as initial values, or random values ​​may be generated. In the above example, a1 is still a linear expression and cannot express complex shapes.

[0043] Next, in the Sigmoid layer S94, the values ​​are converted into nonlinear values ​​using a Sigmoid function, where the Sigmoid function is as follows: y = 1 / (1 + exp(-x)). Values ​​that have passed through the Sigmoid function are normalized to a range between -1 and 1, and can therefore be used as input data to the next Affine layer S95.

[0044] The above process is repeated (Affine layer S95, Sigmoid layer S96, ...), and the repeatedly processed data becomes output data from the output data layer S97.

[0045] Next, as processing of the prediction model calculation unit S73, a teacher data matrix L is generated for the previous output data. Here, the teacher data matrix L is set as follows: L = (l1, l2, l3, ..., lm), where l1 is a parameter representing no bogie abnormality, l2 is an abnormality in the bogie part α, l3 is an abnormality in the bogie part β, ...

[0046] Here, if each parameter is true, it is set to 1, and if it is not true, it is set to 0. For example, if there is no abnormality in the cart, the teacher data matrix L is L = (1, 0, 0, ..., 0), and if an abnormality occurs only in the cart part α, the teacher data matrix L is L = (0, 1, 0, 0, ..., 0) (however, all "..." are 0).

[0047] The training data may be the results of a regular inspection or theoretical values. When the results of a regular inspection are used, the training data may be applied to vibration information obtained after the previous regular inspection, for example.

[0048] When using theoretical values, for example, a method may be used in which the range of vibration frequencies generated by the cart is determined in advance, data on the range of allowable vibration frequencies and frequencies that deviate from that range are input as input data, and the output data that serves as training data is trained as 0 for allowable frequencies and 1 for deviated frequencies.

[0049] The training data and the output data are compared, and any differences are reflected in the weighting signals and biases of the preceding Affine layer. Known methods for this include the gradient method and the backpropagation method.

[0050] In this embodiment, an example of changing weighting signals and biases using backpropagation will be described. Backpropagation uses the chain rule of differentiation. For example, if the multiplication, addition, and sigmoid parts of a calculation formula are expressed using a notation called a calculation graph, they will be expressed as shown in FIGS. 5, 6, and 7, respectively. In each figure, the notation above the arrow in the calculation graph is when output data is calculated, and the notation below the arrow in the calculation graph is when backpropagation is used from the output data. As shown in FIGS. 5 to 7, weighting signals and bias values ​​can be learned so that correct output data is output.

[0051] Furthermore, the training data may be generated based on faulty parts found during vehicle inspections such as monthly inspections, important part inspections, general inspections, and special inspections, or may be generated using simulation data.

[0052] As described above, the prediction model calculation unit S73 transmits the matrix of the prediction model learned using the input data layer S91 to the output data layer S97 to the vehicle-side communication unit S5 via the ground-side communication unit S6. This matrix of the prediction model is used by the on-board prediction model confirmation unit S42 to identify abnormalities.

[0053] According to the present invention, data with less noise can be easily selected using the information held by the vehicle information control device S4, thereby improving the accuracy of detecting bogie abnormalities. Furthermore, because data with less noise can be selected, sounds with less noise can be extracted using a vibration sensor or acoustic sensor additionally installed on the vehicle, without having to attach a vibrometer directly to the bogie.

[0054] Figure 8 shows a spectrogram of sound data collected inside a vehicle while it was actually traveling between stations. Between 130 s and 250 s while the train was coasting and traveling on a straight section, distinctive vibration frequencies were collected, with strong frequencies distributed in a striped pattern. However, in the spectrograms during other times, such as traveling on a curve or powering, no distinctive vibration frequencies were observed.

[0055] Although the embodiments of the present invention have been described above, the present invention is not limited to the above-described embodiments, and various modifications are possible within the scope of the gist of the present invention.

[0056] S1: vehicle, S2: vibration sensor or acoustic sensor, S3: driving information transmission device, S4: vehicle information control device, S5: vehicle-side communication device, S6: ground-side communication device, S7: prediction data calculation device, S8: vehicle depot, S10: crew member, S11: maintenance staff, S41: vibration extraction unit, S42: prediction model confirmation unit, S43: abnormal state display unit, S44: driving state calculation unit, S45: recording unit, S46: prediction model storage unit, S71: prediction model transmission unit, S72: maintenance data input unit, S73: prediction model calculation unit, S74: received data storage unit, S75: recording unit, S91: input data layer, S92: normalization layer, S93, S95: affine layer, S94, S96: sigmoid layer, S97: output data layer

Claims

1. A bogie abnormality detection system for detecting an abnormality of a bogie of a vehicle traveling on a track, comprising: a vibration sensor or an acoustic sensor that senses the vibration of the bogie and outputs vibration data of the bogie; a vehicle information control device that takes as input the vibration data and the running state information of the vehicle including the driving state information of the motor drive device of the vehicle; and the vehicle information control device includes an extraction unit that extracts, as vibration information, at least any one of the vibration data during straight-line section running or during the coasting operation from the vibration data based on the running state information; and a detection unit that detects an abnormality of the bogie from the vibration information extracted by the extraction unit using a prediction model calculated by machine learning with the vibration information and the abnormality information of the bogie as teacher data. A bogie abnormality detection system characterized by the above.

2. The bogie abnormality detection system according to claim 1, wherein the extraction unit extracts the vibration data in a section where the driving state information is equal to or less than a predetermined threshold value as the vibration information. A bogie abnormality detection system characterized by the above.

3. The bogie abnormality detection system according to claim 1 or 2, wherein the running state information includes at least any one of information on a time zone and a section where the vehicle passes by another vehicle, and the extraction unit removes at least any one of the vibration data in the time zone and the section from at least any one of the information on the time zone and the section. A bogie abnormality detection system characterized by the above.

4. The bogie abnormality detection system according to any one of claims 1 to 3, wherein the detection unit detects an abnormality of the bogie when the abnormality of the bogie is determined for a certain time or a certain number of times. A bogie abnormality detection system characterized by the above.

5. Sensing the vibration of a bogie of a vehicle traveling on a track to obtain vibration data of the bogie, extracting, as vibration information, at least any one of the vibration data during straight-line section running or during coasting operation from the vibration data based on the running state information of the vehicle including the driving state information of the motor drive device of the vehicle, calculating a prediction model by machine learning using the vibration information and the abnormality information of the bogie as teacher data, and detecting an abnormality of the bogie from the vibration information using the prediction model. A bogie abnormality detection method characterized by the above.

6. The bogie abnormality detection method according to claim 5, Extract the vibration data in the section where the drive state information is equal to or less than a predetermined threshold value as the vibration information A bogie abnormality detection method characterized by the above.

7. The bogie abnormality detection method according to claim 5 or 6, wherein the traveling state information includes at least one of information on a time zone and a section in which the vehicle passes by another vehicle, remove the vibration data of at least one of the time zone and the section from at least one of the information on the time zone and the section A bogie abnormality detection method characterized by the above.

8. The bogie abnormality detection method according to any one of claims 5 to 7, wherein when the abnormality of the bogie is determined as the abnormality for a certain time or a certain number of times, detect the abnormality of the bogie A bogie abnormality detection method characterized by the above.