Encoder signal processing method based on time domain averaging and depth unwinding

Abstract

The invention relates to an encoder signal processing method based on time domain averaging and depth unwinding. The method comprises the following steps: firstly, acquiring an angular position signal by using a rotary encoder, and converting the angular position signal into an instantaneous angular acceleration signal by using a second-order difference method; then, a time domain synchronous averaging method is used for dividing the whole instantaneous angular acceleration signal into segments containing five cycle lengths, averaging is carried out, and noise components are removed; setting N layers of neural networks, and for each layer of network, sequentially initializing a weight matrix, taking related kurtosis as a target function, utilizing an L-BFGS algorithm, and using the target function and a weight coefficient of previous iteration to update the weight matrix; and finally, performing dimension reduction processing on a plurality of signals output by the neural network, and reserving the signal with the maximum envelope spectrum kurtosis value as a final calculation result. According to the method, the correlation kurtosis is used as a target function, and robustness is achieved for interference of impulse noise; a Hanning window is adopted to initialize the filter bank, and the efficiency of the filtering process can be effectively improved.

Description

Technical Field

[0001] The present invention relates to the technical field of fault signal diagnosis of rotating mechanical equipment, and in particular to an encoder signal processing method based on time domain averaging and deep deconvolution. Background Art

[0002] The integration and intelligence level of rotating machinery is constantly improving, and the importance of its stable operation is becoming increasingly prominent. In mechanical systems, the failure of key transmission components such as gears and bearings is the main cause of mechanical equipment failure. Therefore, it is necessary to monitor its condition and diagnose its faults. With the development of complex equipment, traditional diagnostic methods based on vibration analysis face problems such as high sensor installation costs and fault signals being submerged by strong noise. In comparison, rotary encoders, as built-in devices for monitoring the operating status of mechanical equipment, are installed near transmission components, and their signals usually contain rich diagnostic information. Using encoder signals to diagnose rotating machinery faults has the following advantages:

[0003] 1) The rotary encoder is usually connected to the key components of the transmission, with a short transmission path and high signal fidelity;

[0004] 2) The encoder signal represents the torsional state of the rotating machinery. The stiffness change caused by initial faults such as gear cracks will lead to abnormal torsional vibration behavior, and the encoder is more sensitive to this.

[0005] 3) Rotary encoders are built into many mechanical systems, and signal acquisition is low cost and easy to obtain.

[0006] However, encoder signals are also subject to noise, and analysis of their signals requires specific conversion, processing, and denoising techniques. Summary of the invention

[0007] In order to analyze the encoder signal, the present invention provides an encoder signal processing method based on time domain averaging and deep deconvolution, which removes the noise component in the signal and separates the fault signal component, thereby performing fault diagnosis on the mechanical equipment.

[0008] The technical solution adopted by the present invention is:

[0009] The maximum correlation kurtosis deep deconvolution method for analyzing encoder signals comprises the following steps:

[0010] Step 1: Get the rotary encoder signal x e , and preprocess it. Using the second-order difference method, the incremental angular position signal collected by the encoder is converted into an instantaneous angular acceleration signal x a ;

[0011] Step 2: Use the time domain synchronous averaging method to divide the entire instantaneous angular acceleration signal into segments containing 5 full cycles, and take the average value of each segment signal to remove the noise component and enhance the periodic fault component in the signal to obtain the signal x;

[0012] Step 3: Set the number of network layers to N, the number of nodes to K, and the filter length to L;

[0013] Step 4: Initialize the weight matrix for the i-th layer of the neural network

[0014] W i =[w1; w2; ...; w k ] (1)

[0015] Where W i It can be regarded as a filter bank. The filter bank is initialized using the Hanning window of the same bandwidth. The center frequencies of the filters in the group are evenly distributed throughout the frequency band, and the upper and lower cutoff frequencies of each narrowband are f l and f u satisfy

[0016]

[0017] Step 5: Calculate the gradient of the objective function with respect to the weight coefficient The correlation kurtosis of signal y is used as the objective function to guide the iteration direction. Respectively expressed as

[0018]

[0019] in, T represents the number of sampling points corresponding to the fault characteristic cycle;

[0020] Step 6: Use the L-BFGS algorithm to update the weight matrix of the i-th layer neural network

[0021]

[0022] in, is the weight coefficient updated in the previous iteration, is the inverse of the Hessian matrix estimated by L-BFGS;

[0023] Step 7: Repeat steps 5 and 6 until the maximum number of iterations of the L-BFGS algorithm is reached or the stopping condition is met;

[0024] Step 8: Get the output of the i-th layer neural network through feedforward: X i =W i ·X i-1 , let i=i+1;

[0025] Step 9: Repeat steps 3 to 8 until all the set layers are trained;

[0026] Step 10: Perform dimensionality reduction processing on the output of the neural network. Among the multiple signals in the output matrix, select the signal with the largest envelope spectrum kurtosis as the final calculation result.

[0027] Compared with the traditional technology, the present invention has the following advantages:

[0028] 1) The present invention is robust to the interference of pulse noise and can effectively reduce non-periodic noise without any prior knowledge.

[0029] 2) The present invention adopts the Hanning window to initialize the filter group, which can effectively improve the efficiency of the filtering process.

[0030] 3) The present invention has higher feature extraction accuracy under heavy noise and pulse interference conditions. BRIEF DESCRIPTION OF THE DRAWINGS

[0031] Figure 1 is a flow chart of the present invention;

[0032] Figure 2 It is a planetary gearbox test bench in the embodiment;

[0033] Figure 3 The original signal collected by the encoder in the embodiment;

[0034] Figure 4 The instantaneous angular acceleration signal is obtained by converting the original signal in the embodiment;

[0035] Figure 5 The signal obtained by processing by the time domain synchronous averaging method in the embodiment;

[0036] Figure 6 It is the final result obtained after neural network processing and dimensionality reduction in the embodiment. DETAILED DESCRIPTION

[0037] The present invention is described in detail below with reference to the accompanying drawings and examples.

[0038] Taking the planetary gearbox test bench as an example, the test bench consists of a motor, a planetary gearbox, a brake, etc. Figure 1 As shown. The rotary encoder is installed at the input and output ends of the planetary gearbox. For the prefabricated pitting fault of the planetary gear, the signal is collected and processed by the present invention. The specific method includes the following steps:

[0039] Step 1: Get the rotary encoder signal x e ,like Figure 2As shown. Preprocess it. Use the second-order difference method to convert the incremental angular position signal collected by the encoder into an instantaneous angular acceleration signal x a ,like Figure 3 As shown;

[0040] Step 2: Use the time domain synchronous averaging method to divide the entire instantaneous angular acceleration signal into segments containing 5 full cycles, and take the average value of each segment signal to remove the noise component and enhance the periodic fault component in the signal to obtain the signal x, such as Figure 4 As shown;

[0041] Step 3: Set the number of network layers to 3, the number of nodes to 40, and the filter length to 40;

[0042] Step 4: Initialize the weight matrix for the i-th layer of the neural network

[0043] W i =[w1; w2; ...; w k ] (6)

[0044] Where W i It can be regarded as a filter bank. The filter bank is initialized using the Hanning window of the same bandwidth. The center frequencies of the filters in the group are evenly distributed throughout the frequency band, and the upper and lower cutoff frequencies of each narrowband are f l and f u satisfy

[0045]

[0046] Step 5: Calculate the gradient of the objective function with respect to the weight coefficient The correlation kurtosis of signal y is used as the objective function to guide the iteration direction. Respectively expressed as

[0047]

[0048] in, T represents the number of sampling points corresponding to the fault characteristic cycle;

[0049] Step 6: Use the L-BFGS algorithm to update the weight matrix of the i-th layer neural network

[0050]

[0051] in, is the weight coefficient updated in the previous iteration, is the inverse of the Hessian matrix estimated by L-BFGS;

[0052] Step 7: Repeat steps 5 and 6 until the maximum number of iterations of the L-BFGS algorithm is reached or the stopping condition is met;

[0053] Step 8: Get the output of the i-th layer neural network through feedforward: X i =W i ·X i-1 , let i=i+1;

[0054] Step 9: Repeat steps 3 to 8 until all the set layers are trained;

[0055] Step 10: Perform dimensionality reduction processing on the output of the neural network. Among the multiple signals in the output matrix, select the signal with the largest envelope spectrum kurtosis as the final calculation result, such as Figure 5 shown.

[0056] Depend on Figure 5 It can be seen that in the processed angular acceleration signal, periodic pulses appear every 31 teeth, so the fault can be discovered and located on the planetary gear with 31 teeth.

Claims

1. A method for processing encoder signals based on time domain averaging and deep deconvolution, characterized in that: The following steps are involved: Step 1: Get the rotary encoder signal x e The second-order difference method is used to convert the incremental angular position signal collected by the encoder into an instantaneous angular acceleration signal x a ; Step 2: Use the time domain synchronous averaging method to divide the entire instantaneous angular acceleration signal into segments containing 5 full cycles, and take the average value of each segment signal to remove the noise component and enhance the periodic fault component in the signal to obtain the signal x; Step 3: Set the number of network layers to N, the number of nodes to K, and the filter length to L; Step 4: Initialize the weight matrix for the i-th layer of the neural network IN i =[w1;w2;...;w k ] (1) Where W i It can be regarded as a filter bank. The filter bank is initialized using the Hanning window of the same bandwidth. The center frequencies of the filters in the group are evenly distributed throughout the frequency band, and the upper and lower cutoff frequencies of each narrowband are f l and f u satisfy Step 5: Calculate the gradient g of the objective function with respect to the weight coefficient Wi The relevant kurtosis of signal y is used as the objective function to guide the iteration direction. The relevant kurtosis CK and g Wi Respectively expressed as in, T represents the number of sampling points corresponding to the fault characteristic cycle; Step 6: Use the L-BFGS algorithm to update the weight matrix of the i-th layer neural network in, is the weight coefficient updated in the previous iteration, is the inverse of the Hessian matrix estimated by L-BFGS; Step 7: Repeat steps 5 and 6 until the maximum number of iterations of the L-BFGS algorithm is reached or the stopping condition is met; Step 8: Get the output of the i-th layer neural network through feedforward: X i =W i ·X i-1 , let i=i+1; Step 9: Repeat steps 3 to 8 until all the set layers are trained; Step 10: Perform dimensionality reduction processing on the output of the neural network. Among the multiple signals in the output matrix, select the signal with the largest envelope spectrum kurtosis as the final calculation result.

Images