A Fault Diagnosis Method for Elevator Bearing Components Based on an Improved Empirical Wavelet and Wavelet Packet Convolutional Neural Network

Abstract

This invention discloses a fault diagnosis method for elevator bearing components based on an improved empirical wavelet and wavelet packet convolutional neural network. The improved empirical wavelet method decomposes vibration signals into feature information of fixed frequency bands using a fixed bandpass filter design criterion. High-frequency filters in the filter bank are used as noise bands, and the corresponding sub-signals are filtered out and reconstructed as input samples for the network model. A multi-level wavelet decomposition full-component fusion structure is constructed using wavelet packet transform, resulting in a multi-level wavelet packet full-component residual module. This module is then embedded into a deep one-dimensional convolutional neural network. The elevator bearing component fault diagnosis model based on the multi-level full-component convolutional neural network is trained to diagnose faults in the sample signals, thereby achieving condition monitoring of the elevator bearing components.

Description

Technical Field

[0001] This invention relates to the field of component fault monitoring during the elevator manufacturing and operation stages, specifically to a fault diagnosis method for elevator bearing components based on a combination of an improved empirical wavelet and a wavelet packet convolutional neural network. Background Technology

[0002] With the development of digital and intelligent technologies, fault diagnosis methods based on wavelet feature extraction have seen significant performance improvements and are widely applied in many fields. Meanwhile, with the modernization of cities, elevators have become indispensable transportation equipment in people's daily lives, and their operational reliability is of great concern to passengers, manufacturers, and research institutions. Because the elevator is a complex integrated system, fault diagnosis methods based on whole-machine mechanism modeling suffer from significant lag. Therefore, designing fault diagnosis methods based on vibration signals from elevator bearings and other components is of great significance for elevator equipment condition monitoring, reliable operation, and operational safety. Against this backdrop, model frameworks combining data frequency feature extraction and deep learning have become mainstream methods and are widely studied. Currently, mainstream vibration signal time-frequency feature extraction methods include wavelet packet decomposition, empirical mode decomposition, and empirical wavelet transform, all of which decompose the target signal into sub-signals of different frequency modes as input features for the network. However, existing methods have the following significant drawbacks in signal decomposition design: a) Wavelet packet transform decomposes signals in a progressive manner, failing to extract features within a single, accurate frequency band; b) Empirical mode decomposition lacks theoretical interpretability, the resulting sub-signals have no practical physical meaning, and their number is uncontrollable; c) The determination of empirical wavelet transform sub-band filters does not follow the single-variable principle and is influenced by the extreme points of the signal's fast Fourier transform frequency domain. Therefore, designing a novel and effective method for extracting time-frequency features of vibration signals is of great significance for improving the accuracy of fault diagnosis in elevator bearing components. Summary of the Invention

[0003] This invention addresses the need for monitoring the potential fault conditions of elevator bearing components by proposing a fault diagnosis method for elevator bearing components based on an improved empirical wavelet and wavelet packet convolutional neural network. To address the shortcomings of existing technologies in signal feature extraction, an improved empirical wavelet method is designed to extract feature information of vibration signals in a fixed frequency band. Then, a full-component wavelet convolutional neural network is constructed to achieve condition monitoring of elevator bearing components.

[0004] A fault diagnosis method for elevator bearing components based on an improved empirical wavelet and wavelet packet convolutional neural network includes the following steps:

[0005] Step 1: Use sensors to collect vibration signals from elevator bearing components and perform data preprocessing to obtain sample data in a uniform format;

[0006] Step 2: Design a fixed multi-channel bandpass filter using empirical wavelet transform, and then decompose the uniform format sample data obtained in Step 1 into multiple sub-signals of different frequencies. The sub-signal with the highest frequency among the multiple sub-signals of different frequencies is taken as the noise signal.

[0007] Step 3: Filter the multiple sub-signals of different frequencies obtained in Step 2 to remove noise signals and obtain the sub-signals to be reconstructed. Reconstruct the sub-signals to be reconstructed by empirical wavelet transform and use the reconstructed sub-signals as a sample set.

[0008] Step 4: Design a multi-level residual deep convolutional neural network based on wavelet packet decomposition technology;

[0009] Step 5: Based on the multi-level residual deep convolutional neural network in Step 4, the sample set obtained in Step 3 is used as the input of the multi-level residual module of the deep convolutional neural network, and the corresponding fault state label is used as the target output of the multi-level residual deep convolutional neural network. The model is trained to obtain the fault diagnosis model of elevator bearing components.

[0010] Step 6: Collect vibration signals from elevator bearing components in real time and input them into the elevator bearing component fault diagnosis model obtained in Step 5 to diagnose the current health status of the elevator.

[0011] In step 1, the data preprocessing includes: standardization and normalization of sample length.

[0012] In step 2, a fixed multi-channel bandpass filter is designed using empirical wavelet transform. Specifically, the bandwidth of the filter is given by the following formula:

[0013]

[0014] The filter boundary is given by the following formula:

[0015] ;

[0016] Where B represents bandwidth, N represents the number of bandpass filters, w0 represents the boundary of filter 0, w1 represents the boundary of filter 1, and w n w represents the boundary of filter n. N This represents the boundary of filter N.

[0017] In step 3, the multiple sub-signals of different frequencies obtained in step 2 are filtered to remove noise signals, specifically including:

[0018] The multiple sub-signals S of different frequencies obtained in step 2 are decomposed into N-1 sub-signals and noise signals through convolution operation, as shown in the following formula:

[0019]

[0020] Where S(t) is the original signal at time t, s0 is sub-signal 0, s1 is sub-signal 1, and s N-1 It is sub-signal N-1, s n Subsignals n and s noise It is a noise signal; W S (0, t) represents the wavelet coefficients of sub-signal 0, φ1(t) represents the scaling function, and W S (n, t) represents the wavelet coefficients corresponding to the subsignal n, Ψ n (t) represents the wavelet function corresponding to the subsignal n, W S (N, t) represents the wavelet coefficients corresponding to the subsignal N, Ψ N (t) represents the wavelet function corresponding to the sub-signal N.

[0021] In step 4, a full-component multi-level residual module of a deep convolutional neural network is designed based on wavelet packet decomposition technology, specifically including:

[0022] The reconstructed sub-signals are decomposed into multiple sub-signals of different frequencies using wavelet packet technology. All sub-signals are then combined using a backpropagation neural network algorithm to obtain a first-level residual module.

[0023] The reconstructed sub-signals are decomposed into multiple sub-signals of different frequencies using wavelet packet technology. All sub-signals are then combined using a backpropagation neural network algorithm to obtain a two-level residual module.

[0024] The reconstructed sub-signals are decomposed into multiple levels using wavelet packet technology to obtain multiple sub-signals of different frequencies. All sub-signals are then combined using a backpropagation neural network algorithm to obtain a multi-level residual module.

[0025] The sub-signals obtained after wavelet decomposition are regarded as one component, and the full component is the input of the residual module using all components.

[0026] To achieve the above objectives, the technical solution designed in this invention includes a data feature reorganization stage and a deep learning stage, as detailed below:

[0027] Data feature reorganization stage:

[0028] Step 1: Collect vibration signals of elevator bearing components using sensors with the same sampling frequency, and preprocess the data.

[0029] Step 2: Design a fixed multi-channel bandpass filter using an empirical wavelet architecture to decompose the vibration signal into multiple sub-signals of different frequencies, and set the last high-frequency channel as a noise filter.

[0030] Step 3: Filter out the sub-signals from the high-frequency noise channels and reconstruct the sub-signals obtained from other bandpass filters as the sample set for neural network input;

[0031] Deep learning stages:

[0032] Step 4: Design a multi-level residual module for a deep convolutional neural network based on full-component wavelet packet decomposition technology;

[0033] Step 5: Construct a deep convolutional neural network model for elevator bearing components based on the multi-level residual module in Step 4. Use the sample set obtained in Step 3 as the input of the deep convolutional neural network model for elevator bearing components, and use the corresponding fault state label as the target output of the network model. Train the deep convolutional neural network for elevator bearing components to obtain the final model.

[0034] Step 6: Real-time acquisition of vibration signals from elevator bearing components. Preprocessing of the signals is performed using Step 1, followed by fixed multi-channel signal decomposition in Step 2. Then, signal filtering and feature reconstruction are performed using Step 3. Finally, the reconstructed features are input into the deep convolutional neural network model of elevator bearing components trained in Step 5 to diagnose the current health status of the elevator.

[0035] Compared with existing methods, the advantages of this invention are mainly reflected in:

[0036] This invention proposes a fault diagnosis method for elevator bearing components based on an improved empirical wavelet and a full-component convolutional neural network. The method involves collecting vibration signals from bearing components in the same location within an elevator and performing relevant data preprocessing to construct a preliminary sample set. Using a fixed-band empirical wavelet filter bank, sub-signals of different frequencies are extracted, denoised, and reconstructed to construct a direct training set. The reconstructed samples are then input into a wavelet packet full-component deep convolutional neural network model for fault diagnosis of elevator bearings and other components. This invention utilizes vibration signal fault modeling, enabling fault diagnosis of elevator components with various accelerometer sensor deployments. Furthermore, this invention proposes a multi-level wavelet packet full-component residual module, reducing information loss during signal downsampling. Attached Figure Description

[0037] Figure 1 This is a flowchart illustrating the specific process framework of the method of the present invention;

[0038] Figure 2 This refers to the multi-bandpass filter bank proposed in the embodiments of the present invention;

[0039] Figure 3 The empirical wavelet decomposition sub-signal proposed in this invention;

[0040] Figure 4 This invention relates to a multi-level wavelet packet full component residual module and a deep convolutional neural network proposed in an embodiment of the present invention. Detailed Implementation

[0041] To make the objectives, design concepts, and technical solutions of the embodiments of the present invention clearer, the present invention will be further described below with reference to the accompanying drawings.

[0042] like Figure 1 As shown, this invention proposes a fault diagnosis method for elevator bearing components based on a combination of improved empirical wavelet and wavelet packet convolutional neural network. The method includes the following steps:

[0043] In step 1, a sliding window is used to crop the vibration signal to obtain samples of the same length, thereby standardizing the sample length in the dataset, and then a normalization strategy is used to preprocess the sample data.

[0044] In the bandpass filter construction in step 2, the entire frequency band [0, π] is segmented using empirical wavelet transform. First, based on the ideas of Littlewood-Paley and Meyer wavelets, the following scaling function and wavelet function are constructed respectively:

[0045]

[0046]

[0047] Where β(x) is defined as x 4 (35-84x+70x 2 -20x 3 ), where w represents the frequency domain parameter, w n w n+1 This represents the boundary of a bandpass filter, where n represents the filter's ordinal number and the coefficients. When n = N, ω n+1 =n;

[0048] To distinguish it from the traditional method ω n The frequency domain extreme points determined by the fast Fourier transform of the vibration signal lead to excessive differences in the extracted features between different signals. This invention proposes a method for fixing filter boundaries and dividing the entire frequency band according to the following convention: the width of each bandpass filter is:

[0049]

[0050] The boundary of a bandpass filter is represented as:

[0051]

[0052] like Figure 2 As shown, this is a bandpass filter bank on [0, π].

[0053] In the wavelet decomposition and reconstruction of vibration signals from elevator bearing components, wavelet coefficients are calculated using the inner product operation, as shown in the following formula:

[0054] W S (n, t) = <S,Ψ n >=∫S(τ)Ψ n (τ-t)dτ

[0055] Among them, W S (n, t) represents the wavelet coefficients corresponding to the subsignal n, Ψ n Let represent the wavelet function corresponding to sub-signal n, t represent the translation parameter, and S represent the time-domain signal.

[0056] The signal S is then decomposed into N-1 sub-signals and noise signals through convolution operations, as shown in the following formula:

[0057]

[0058] Where S(t) is the original signal at time t, s0 is the sub-signal 0, sn is the sub-signal n, and s noise It is a noise signal, W S (0, t) represents the wavelet coefficients of sub-signal 0, and φ1(t) represents the scaling function.

[0059] In step 3, the sample recombination follows the formula below:

[0060] [s0, ..., s n , ..., s N-1 ]

[0061] Where s1 is sub-signal 1, s N-1 It is sub-signal N-1.

[0062] The improved empirical wavelet strategy is as follows:

[0063] Set all ω n This is a fixed value to ensure that the sub-signals extracted from the dataset are within the same frequency band;

[0064] High frequency band [ω N [π] serves as the passband for the noise filter, discarding sub-signals that pass through this passband;

[0065] Reconstructed eigensignals are used as input to a neural network model to replace the feature hiding after empirical wavelet reconstruction.

[0066] like Figure 3As shown, these are the individual sub-signals obtained after decomposing the vibration signal of the elevator bearing.

[0067] In step 4, wavelet packet transform is used to decompose the target signal to obtain its high-frequency and low-frequency components. For example... Figure 4 As shown, all components of the signal participate in the next level of decomposition, and the decomposed component signals are used as inputs to the neural network to calculate the weight of each component.

[0068] The specific calculation method is as follows:

[0069]

[0070] This represents the first-order wavelet packet decomposition of the signal S(x), where A -1 S and D -1 S represents the low-frequency and high-frequency components of the first-order transform of signal S(x), respectively; lp and hp represent the low-pass and high-pass filters, respectively; j is the scaling parameter; and n is the translation parameter.

[0071] Approximate coefficients A after decomposition -1 S and detail factor D -1 The length of S is always half of S(x), satisfying the requirements of first-level downsampling data; in the downsampling stage of the residual module, a weighted fusion method of high-frequency components and low-frequency components is used to retain more signal frequency information. The formula is:

[0072]

[0073] The backpropagation neural network was calculated using gradient descent to approximate the optimal w1 and w2, thus obtaining the first-order wavelet packet full air volume downsampling results. This is a retained term of the original residual signal.

[0074] Similarly, repeat the process for A. -1 S and D -1 S is decomposed into wavelet packets to obtain second-order wavelet packet components A of different frequencies. -2 SH, A -2 SL, D -2 SH, D -2 SL;

[0075] Next, a backpropagation neural network is constructed to calculate the fusion weights w1, w2, w3, and w4 for each second-level component, thereby obtaining the second-level downsampling result.

[0076]

[0077] Construct the three-level wavelet packet full component residual module in the same way.

[0078] Step 5 of the downsampling module for full component signal fusion based on wavelet packet decomposition constructs a multi-level wavelet packet full component residual convolutional neural network, the specific structure of which is as follows: Figure 4 As shown, the feature extraction module of this network structure includes one third-level wavelet packet residual module, two second-level wavelet packet residual modules, and four first-level wavelet packet residual modules. Each second-level wavelet packet residual module consists of two consecutive one-dimensional convolution operators. Then, an unfolding layer and a fully connected layer are added to the output decision module of the multi-level wavelet packet residual convolutional neural network.

[0079] The reconstructed elevator bearing vibration signals from the collected database were divided into a training set and a test set.

[0080] Through the above operations and optimization training, a fault diagnosis model for elevator bearing components based on the combination of improved empirical wavelet and full component convolutional neural network is obtained.

[0081] In step 6, the present invention utilizes an accelerometer deployed on the elevator bearing components to collect the vibration signal of the component in real time; the vibration signal is standardized and normalized using the data preprocessing method in step 1; based on the data preprocessed by the improved empirical wavelet decomposition, the decomposed noise sub-signal s is... noise Filter out and reconstruct the remaining sub-signals to generate input samples for model matching:

[0082] [s0, ..., s n , ..., s N-1 ]

[0083] Furthermore, the multi-level wavelet packet full-component residual convolutional neural network model trained in step 5 is used to diagnose faults in the samples.

[0084] Finally, the fault diagnosis results of the above model are used to achieve real-time status monitoring of elevator bearing components.

[0085] like Figure 3 As shown, the wavelet decomposition of the vibration signal of elevator bearing components is set with the number of sub-signals N=8, and the N=8 sub-signals s noise =s8 is treated as noise signal and removed.

Claims

1. A fault diagnosis method for elevator bearing components based on a combination of improved empirical wavelet and wavelet packet convolutional neural network, characterized in that, Includes the following steps: Step 1: Use sensors to collect vibration signals from elevator bearing components and perform data preprocessing to obtain sample data in a uniform format; Step 2: Design a fixed multi-channel bandpass filter using empirical wavelet transform, and then decompose the uniform format sample data obtained in Step 1 into multiple sub-signals of different frequencies. The sub-signal with the highest frequency among the multiple sub-signals of different frequencies is taken as the noise signal. Step 3: Filter the multiple sub-signals of different frequencies obtained in Step 2 to remove noise signals and obtain the sub-signals to be reconstructed. Reconstruct the sub-signals to be reconstructed by empirical wavelet transform and use the reconstructed sub-signals as a sample set. Step 4: Design a multi-level residual deep convolutional neural network based on wavelet packet decomposition technology, specifically including: The reconstructed sub-signals are decomposed into multiple sub-signals of different frequencies using wavelet packet technology. All sub-signals are then combined using a backpropagation neural network algorithm to obtain a first-level residual module. The reconstructed sub-signals are decomposed into multiple sub-signals of different frequencies using wavelet packet technology. All sub-signals are then combined using a backpropagation neural network algorithm to obtain a two-level residual module. The reconstructed sub-signals are decomposed into multiple levels using wavelet packet technology to obtain multiple sub-signals of different frequencies. All sub-signals are then combined using a backpropagation neural network algorithm to obtain a multi-level residual module. Step 5: Based on the multi-level residual deep convolutional neural network in Step 4, the sample set obtained in Step 3 is used as the input of the multi-level residual module of the deep convolutional neural network, and the corresponding fault state label is used as the target output of the multi-level residual deep convolutional neural network. The model is trained to obtain the fault diagnosis model of elevator bearing components. Step 6: Collect vibration signals from elevator bearing components in real time and input them into the elevator bearing component fault diagnosis model obtained in Step 5 to diagnose the current health status of the elevator.

2. The elevator bearing component fault diagnosis method based on the combination of improved empirical wavelet and wavelet packet convolutional neural network according to claim 1, characterized in that, In step 1, the data preprocessing includes: standardization and normalization of sample length.

3. The elevator bearing component fault diagnosis method based on the combination of improved empirical wavelet and wavelet packet convolutional neural network according to claim 1, characterized in that, In step 2, a fixed multi-channel bandpass filter is designed using empirical wavelet transform, specifically including: The bandwidth of the filter is given by the following formula: ； The filter boundary is given by the following formula: ； in Indicates bandwidth, and N represents the number of bandpass filters. Indicates the boundary of filter 0. Indicates the boundary of filter 1. Denotes the boundary of filter n. This represents the boundary of filter N.

4. The elevator bearing component fault diagnosis method based on the combination of improved empirical wavelet and wavelet packet convolutional neural network according to claim 1, characterized in that, In step 3, the multiple sub-signals of different frequencies obtained in step 2 are filtered to remove noise signals, specifically including: The multiple sub-signals S of different frequencies obtained in step 2 are decomposed into N-1 sub-signals and noise signals through convolution operation, as shown in the following formula: ； Where S(t) is the original signal at time t, N represents the number of bandpass filters, s0 is sub-signal 0, s1 is sub-signal 1, and s N-1 It is sub-signal N-1, s n Subsignals n and s noise It's a noise signal. The wavelet coefficients representing sub-signal 0, Represents the scaling function. This represents the wavelet coefficients corresponding to the sub-signal n. This represents the wavelet function corresponding to the sub-signal n. This represents the wavelet coefficients corresponding to the sub-signal N. This represents the wavelet function corresponding to the sub-signal N.

Images