An online detection method for train traction motor faults

Abstract

The present invention relates to an on-line detection method for train traction motor faults, comprising the following steps: dividing traction motor current data into blocks according to a specified unit time and preprocessing; using an inter-class difference network structure for multi-class fault prediction.

Description

Technical Field

[0001] The invention belongs to the field of intelligent operation and maintenance of rail transit, and relates to an online detection method for train traction motor faults. Background Art

[0002] The train traction motor is the driving power source of high-speed trains, subway trains and electric locomotives. It directly determines the running speed and driving stability of the train or train, and is the core key equipment of high-speed trains, subway trains and electric locomotives. The train traction motor is located in the train bogie, and is subjected to strong impact vibration between the wheel and rail. The working environment is harsh and it is a high-incidence point of the train traction system. Therefore, the online fault diagnosis of the train traction motor is of great significance to improving the operating reliability of high-speed trains, subway trains and electric locomotives.

[0003] At present, most of the fault diagnosis methods for train traction motors do not meet the rapidity requirements of online monitoring, and the false alarm and missed alarm rates of monitoring are far from meeting the requirements of operating units. Summary of the invention

[0004] Based on this, and in view of the deficiencies in the prior art, the present invention proposes a method for diagnosing a train traction motor fault.

[0005] Furthermore, the method comprises the following steps:

[0006] S1, traction motor current data is divided into blocks according to the specified unit time and pre-processed;

[0007] S2. Use the inter-class difference network structure for reasoning and make multi-classification fault prediction on the reasoning results.

[0008] Furthermore, the step S1 includes the following steps:

[0009] S11, assuming that the data acquisition frequency is f, obtain the data from the current time t0 to the previous cycle time t -1 Data set x at a fixed acquisition frequency f i (i=0,1,...,f-1);

[0010] S12. Use the following preprocessing function to obtain the feature vector The formula is as follows:

[0011]

[0012] Among them, min and max are x i The minimum and maximum values ​​in (i=0,1,...,f-1); a is the scale factor.

[0013] Furthermore, the step S2 comprises the following steps:

[0014] S21. The above is a feature vector of length f×1, and is briefly denoted as X1; construct a 1×4 zero kernel C0=(0,0,0,0) and a 1×4 difference kernel C1=(-3,-1,1,3), and calculate the feature vector X of length f / 4 21 , and the formula is as follows:

[0015]

[0016] where C 211 is a diagonal matrix with C1 on the diagonal, and the matrix size is f / 4×f.

[0017] S22. Calculate the sliding maximum and minimum difference of X1 with a step size of 4 and a length of 4, and obtain the feature vector X of length f / 4 22 , and the formula is as follows:

[0018]

[0019] S23. Calculate the sliding average value of X1 with a step size of 4 and a length of 4, and obtain the feature vector X of length f / 4 23 , and the formula is as follows:

[0020]

[0021] where mean represents the average value.

[0022] S24. Construct a 1×3 convolution kernel C2, and perform convolution calculation on the feature vector combination X3 = (X 21 X 22 X 23 ) of size f / 4×3 and activate it with the ReLU function to obtain the feature vector X4 of length f / 4, and the formula is as follows:

[0023]

[0024] where C 43 is a matrix of size f / 4×1.

[0025] S25. Perform multi-class prediction on the feature vector X4, as shown in the following formula:

[0026] Y i = WX4 + B

[0027] where W is a matrix of size b×f / 4, B is a matrix of size b×1, and Y i is a vector of size b×1.

[0028] S26. Calculate the fault category y, as shown in the following formula:

[0029]

[0030] Among them, y represents the category of the fault, and the range is 0 to b-1, and the function p represents the normalized exponential function. Description of the Drawings

[0031] Figure 1 is the detection flow block diagram of the present invention. Detailed Embodiments

[0032] The following combines the drawings in the embodiments of the present invention to clearly and completely describe the technical solutions in the embodiments of the present invention. Obviously, the described embodiments are only a part of the embodiments of the present invention, rather than all the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those of ordinary skill in the art without creative work shall fall within the protection scope of the present invention.

[0033] For the convenience of understanding by those skilled in the art, an embodiment is now used to illustrate the block division and preprocessing of the traction motor current data in a specified unit time: the data acquisition frequency is 256Hz, and the data set x from the current moment to the previous cycle moment, that is, from the 1st second to the 2nd second, is obtained i (i = 0, 1,..., 255), and the feature vector is obtained after the preprocessing function Select the scale factor 255 so that the range of the vector numerical value is 0 to 255.

[0034] For the convenience of understanding by those skilled in the art, an embodiment is now used to illustrate the inference using the inter-class difference network structure and the multi-class fault prediction of the inference result: X1 is a feature vector with a length of 256×1, and a 1×4 zero kernel C0=(0, 0, 0, 0) and a 1×4 difference kernel C1=(-3, -1, 1, 3) are constructed; a feature vector X with a length of 64 is calculated according to the S21 step 21 , where C 211 has a size of 64×256; a feature vector X with a length of 64 is calculated according to the S22 step 22 , where C 43 is a size of 64×1; a feature vector X with a length of 64 is calculated according to the S23 step 23 ; a feature vector X4 with a length of 64 is calculated according to the S24 step; a vector Y with a length of 5 is calculated according to the S25 step, where the size of W is 5×64 and the size of B is 5×1; according to the S26 step, the fault category y = 1 is calculated, where there are a total of 5 types of fault categories: broken bar fault, inter-turn short circuit, air-gap eccentricity, end-ring fracture, and bearing fault.

Claims

1. An online detection method for train traction motor fault, characterized in that: The following steps are involved: S1, dividing the traction motor current data into blocks according to a specified unit time and preprocessing, including steps S11 and S12; S11, assuming that the data collection frequency is f , get the current time t 0 to the previous cycle time t -1 At fixed frequency f The following dataset x i ( i =0,1,…, f -1); S12. Use the preprocessing function to obtain the feature vector ( i =0,1,…, f -1), the formula is as follows: ; Among them, min and max are x i ( i =0,1,…, f -1) and a is the scale factor; S2, using an inter-class difference network structure to perform multi-classification fault prediction, including steps S21 to S26; S21. Order , construct a 1×4 zero core C 0=(0,0,0,0), 1×4 difference kernel C 1=(-3,-1,1,3), the calculated length is f Eigenvector of / 4 X 21 , the formula is as follows: ;in, C 211 The diagonal line is C 1 diagonal matrix, the matrix size is f / 4× f ; S22, with a step size of 4 and a length of 4, calculate X The maximum and minimum difference of the sliding of 1, the length is f Eigenvector of / 4 X 22 , the formula is as follows: ; S23, with a step size of 4 and a length of 4, calculate X 1, the length is f Eigenvector of / 4 X 23 , the formula is as follows: ; Among them, mean represents the average value; S24, construct a 1×3 convolution kernel C 2, for size f / 4×3 feature vector combination X 3=( X 21 X 22 X 23 ) performs convolution calculation and uses ReLU function activation to obtain a length of f Eigenvector of / 4 X 4. The formula is as follows: ;in, C 43 Is the size of f / 4×1 matrix; S25. For the feature vector X 4. Perform multi-classification prediction, the formula is as follows: Y i = WX 4+ B ;in, W Is the size of b × f / 4 matrix, B Is the size of b ×1 matrix, Y i Is the size of b ×1 vector; S26, Calculation fault category y , the formula is as follows: ;in, y Indicates the type of fault; b is 5; y The range of is 0 to 4, corresponding to broken bar fault, inter-turn short circuit, air gap eccentricity, end ring fracture, and bearing fault; function p represents the normalized exponential function.

Images