Escalator fault diagnosis method

Abstract

The application discloses an escalator fault diagnosis method and system based on Markov transition field (MTF) and a light neural network, and belongs to the technical field of rotating machinery fault diagnosis. In view of the technical problems that the fault diagnosis precision of the escalator is low under variable load conditions, the traditional deep learning model has large parameters, which leads to slow training speed and difficulty in deployment, and it is difficult to extract signal features in a strong noise environment, the application proposes: a one-dimensional vibration signal is converted into a two-dimensional feature image by using a Markov transition field, and the dynamic evolution law of time series data is represented by a state transition probability matrix; a SE-ISHUFFLENETV2 light network model is constructed, an SENet attention mechanism is introduced to weight features of different channels, and a Hardswish activation function is used to solve the neuron necrosis problem; an ICEEMDAN joint correlation coefficient-adaptive wavelet threshold method is used for signal denoising, effective IMF components are screened through a Pearson correlation coefficient, and signal reconstruction is realized by combining adaptive wavelet threshold denoising. The application realizes high-precision and rapid diagnosis of bearing, gear and sprocket faults, and has excellent generalization performance and robustness in variable load and strong noise environments.

Description

Technical Field

[0001] This invention belongs to the field of rotating machinery fault diagnosis technology, specifically relating to a fault diagnosis method and system for escalators based on Markov Transition Field (MTF) and lightweight neural network. Background Technology

[0002] Escalators, as an important component of modern urban public transportation, are widely used in public places such as subway stations, shopping malls, and train stations. Statistics show that as of 2020, there were 1.37 million escalators in use worldwide, of which China had 625,000, accounting for 45.6% of the global total. The efficient operation of escalators not only affects the travel experience of passengers but also directly impacts public safety.

[0003] The main drive end of an escalator includes key transmission components such as bearings, gearboxes, and sprockets. These components are subjected to complex load impacts and wear during long-term operation, making them prone to failure. Traditional fault diagnosis methods mainly rely on regular manual maintenance, which suffers from high subjectivity, low efficiency, and difficulty in detecting early-stage faults. With the development of sensor technology and artificial intelligence, fault diagnosis methods based on vibration signal analysis have gradually become a research hotspot.

[0004] However, existing technologies still have the following technical shortcomings:

[0005] Variable load adaptability problem: In actual operation, the load of escalators fluctuates with changes in passenger flow, resulting in complex and variable vibration signal characteristics. Traditional deep learning models have insufficient generalization ability and significantly reduced diagnostic accuracy.

[0006] The contradiction between lightweight models and performance: While increasing the number of layers in deep neural networks can improve feature extraction capabilities, it leads to slow training speed and a large number of parameters, making it difficult to deploy on resource-constrained embedded devices; on the other hand, while lightweight networks train quickly, their diagnostic accuracy often fails to meet engineering requirements.

[0007] Strong noise interference problem: Escalators are installed in densely populated areas with complex background noise. The vibration signals collected by the sensors have a low signal-to-noise ratio, and the fault characteristics are masked by the noise, which seriously affects the accuracy of diagnosis.

[0008] Therefore, there is an urgent need for a fault diagnosis method for escalators that can maintain high diagnostic accuracy under variable load and high noise environments, while also being lightweight and easy to deploy in engineering. Summary of the Invention

[0009] The purpose of this invention is to overcome the shortcomings of the prior art and provide a method and system for escalator fault diagnosis based on Markov transition fields and lightweight neural networks, which solves technical problems such as low diagnostic accuracy, difficulty in model training and deployment, and strong noise interference under variable load conditions.

[0010] To achieve the above objectives, the present invention adopts the following technical solution:

[0011] In a first aspect, the present invention provides a method for escalator fault diagnosis based on Markov transfer fields and lightweight neural networks, comprising the following steps:

[0012] S1: Vibration signals of key components of the main drive end of the escalator are collected by vibration sensors. The key components include bearings, gears and sprockets.

[0013] S2: The Markov Transfer Field (MTF) method is used to convert one-dimensional time-series vibration signals into two-dimensional feature images;

[0014] S3: Input the two-dimensional feature image into the SE-IShufflenetV2 lightweight neural network model for fault diagnosis;

[0015] S4: Output fault diagnosis results.

[0016] Preferably, the specific implementation of the MTF method in step S2 includes: dividing the time series signal into Q statistical quantile intervals according to the amplitude, establishing a mapping relationship between the observed values ​​and the quantile levels, using a Markov chain model based on the time dimension to model and analyze the state transition probabilities of each quantile interval, and finally generating a Markov transition matrix with state transition characteristics, wherein the matrix elements correspond to the probability of transitioning from one quantile region to another. To overcome the shortcomings of traditional Markov state transition matrices, such as lack of sensitivity to time step parameters and neglect of signal spatial correlation, this invention adopts an improved MTF encoding method, which effectively captures the nonlinear time-varying characteristics in the data by constructing a multidimensional mapping relationship in the state space of the time series signal.

[0017] Preferably, the SE-IShufflenetV2 model described in step S3 includes several targeted improvements. At the infrastructure level, based on the IShufflenetV2 network, an additional set of 3×3 convolutional kernels is added to the basic units, achieving the receptive field effect of a 5×5 convolutional kernel without increasing too many parameters, thus expanding the feature extraction range. At the attention mechanism level, the network embeds SENet modules, which learn the correlation between channels by compressing and activating operations, and adaptively weight the feature maps of different channels to enhance the ability to express key features. At the activation function level, the Hardswish activation function is used to replace the traditional ReLU to solve the problem of neuron death and improve gradient propagation characteristics.

[0018] Preferably, a signal preprocessing step is provided after step S1. The vibration signal is denoised using the ICEEMDAN combined correlation coefficient-adaptive wavelet thresholding method. First, the noisy signal is decomposed into several intrinsic mode function components by an improved adaptive noise complete set empirical mode decomposition, overcoming the mode mixing and endpoint effects problems of traditional EMD. Then, the Pearson correlation coefficient between each IMF component and the original signal is calculated, and a threshold range of 0.3-0.5 is set to remove false components and retain the effective components containing fault information. Then, the effective IMF components after screening are denoised using the adaptive wavelet thresholding method. The threshold can be dynamically adjusted according to the noise level of each IMF component. Finally, the denoised IMF components are reconstructed to obtain a high-quality vibration signal.

[0019] This invention provides an escalator fault diagnosis system based on Markov transfer fields and lightweight neural networks, comprising a data acquisition module, a data transmission and storage module, a signal processing module, and an interactive display module. The data acquisition module consists of an IEPE piezoelectric accelerometer array, an NI PXIe-4492 dynamic signal acquisition card, an NI PXIe-1078 measurement and control platform, and a PXIe-8840 embedded controller. The sensor sensitivity is 100 mV / g, and the frequency range covers 0.5-10000 Hz. The acquisition card supports 8-channel parallel sampling, with a single-channel sampling rate of up to 204.8 kS / s, meeting the high-precision synchronous acquisition requirements of vibration signals from multiple parts of the escalator's main drive end. The signal processing module adopts a joint programming architecture of LabVIEW and Python. LabVIEW is responsible for data acquisition, storage, and interface interaction, while Python is responsible for MTF image encoding, deep learning model inference, and noise reduction algorithm execution. The two achieve efficient data interaction through the TCP / IP protocol. The SE-IShufflenetV2 model can be further compressed in size using knowledge distillation or model quantization techniques to meet the real-time inference requirements of embedded devices.

[0020] Compared with existing technologies, this invention has the following significant advantages: First, it possesses excellent adaptability to varying loads. MTF encoding characterizes the dynamic evolution of time-series data through the state transition probability matrix, effectively capturing nonlinear time-varying features. Combined with an overlapping sampling strategy, it enhances data features, enabling the model to maintain high diagnostic accuracy under varying load conditions. The diagnostic accuracy for bearing, gear, and sprocket datasets can all reach 100%. Second, it achieves a balance between lightweight design and high performance. The SE-IShufflenetV2 model significantly improves feature extraction capabilities while maintaining model lightweight design through the replacement of 3×3 convolutional kernels, the introduction of the SENet attention mechanism, and optimization of the Hardswish activation function. Training speed is 3-5 times faster than traditional deep networks, and the number of parameters is reduced by more than 60%, making it easier to deploy on embedded devices. Furthermore, it exhibits outstanding noise robustness. The ICEEMDAN combined with the correlation coefficient-adaptive wavelet threshold denoising method uses the Pearson correlation coefficient to select effective IMF components, combined with adaptive wavelet threshold denoising, to effectively extract fault features even in low signal-to-noise ratio environments, significantly improving the model's noise resistance. Finally, it possesses strong engineering practicality. The fault diagnosis system, developed based on LabVIEW and Python, can achieve real-time monitoring of vibration signals in the time-frequency domain, remote data access, and visual interaction, fully meeting the actual needs of escalator on-site operation and maintenance. Attached Figure Description

[0021] Figure 1 This is a flowchart of the fault diagnosis method of the present invention;

[0022] Figure 2 A schematic diagram of the Markov transition field coding principle;

[0023] Figure 3 This is a diagram of the SE-IShufflenetV2 network structure.

[0024] Figure 4 Flowchart of noise reduction processing for ICEEMDAN;

[0025] Figure 5 This is a diagram of the escalator fault diagnosis system architecture. Detailed Implementation

[0026] To make the objectives, technical solutions, and advantages of this invention clearer, the invention will be further described in detail below with reference to specific embodiments.

[0027] Example 1: Optimization of MTF Feature Encoding Parameters

[0028] For bearing fault diagnosis, the CWRU bearing dataset was used to conduct MTF parameter optimization experiments. The quantile interval number Q was set to a range of {4, 8, 16, 32, 64}, the image size was {64×64, 128×128, 224×224}, and the overlap sampling window step size was {50, 100, 200}.

[0029] Experimental results show that when Q=8, image size 224×224, and window stride 100, the SE-IShufflenetV2 model achieves optimal diagnostic accuracy of 99.8% on the test set while maintaining a relatively fast training speed. The two-dimensional feature image encoded by MTF can clearly distinguish four types: rolling element fault, inner race fault, outer race fault, and normal state.

[0030] Example 2: Ablation Experiment of SE-IShufflenetV2 Model

[0031] To verify the effectiveness of each improved module, an ablation experiment was designed:

[0032] Baseline model: Original IShufflenetV2;

[0033] Improved Model A: Replaced with IShufflenetV2+ 3×3 convolutional kernels;

[0034] Improved Model B: IShufflenetV2 + SENet attention mechanism;

[0035] Improved model C: IShufflenetV2 + Hardswish activation function;

[0036] The model of this invention (SE-IShufflenetV2) integrates all the above improvements.

[0037] Experimental results on the gear dataset show that the diagnostic accuracy of the model in this invention is improved by 4.2 percentage points compared with the benchmark model, the number of parameters increases by only 8.5%, and the single inference time is reduced by 35%.

[0038] Example 3: Verification of Generalization Performance under Variable Load Conditions

[0039] On the sprocket test platform, three load conditions of 0Nm, 50Nm, and 100Nm were set, and four state data were collected: pitting, tooth breakage, wear, and normal operation. A migration test setting was adopted with a source domain of 0Nm load and target domains of 50Nm and 100Nm load.

[0040] Comparative experiments show that the diagnostic accuracy of the method of the present invention reaches 96.5% under variable load conditions, which is significantly better than traditional CNN (78.3%), LSTM (82.1%) and the original ShufflenetV2 (89.4%), demonstrating excellent cross-condition adaptability.

[0041] Example 4: Verification of noise reduction effect in strong noise environment

[0042] Gaussian white noise of varying intensities was added to the bearing signal, with signal-to-noise ratios (SNR) set to -4dB, -2dB, 0dB, 2dB, and 4dB. Compared to denoising methods such as CEEMDAN, VMD, and wavelet thresholding, the ICEEMDAN combined with correlation coefficient-adaptive wavelet thresholding method of this invention achieved the best denoising effect at all SNR levels, and the diagnostic accuracy of the reconstructed signal was improved by 15-30 percentage points compared to the undenoised signal.

[0043] Example 5: Field Application of Fault Diagnosis System

[0044] The system of this invention was deployed on an escalator in a subway station, equipped with 4-channel IEPE sensors to monitor the main drive bearing, gearbox input shaft, gearbox output shaft, and sprocket respectively. After three months of continuous operation, the system successfully identified two cases of early bearing wear and one case of gear pitting, achieving a 100% early warning accuracy rate and an average diagnostic response time of less than 200ms, meeting the requirements for real-time monitoring.

Claims

1. A method for escalator fault diagnosis based on Markov transfer fields and lightweight neural networks, characterized in that, Includes the following steps: S1: Vibration signals of key components of the main drive end of the escalator are collected by vibration sensors. The key components include bearings, gears and sprockets. S2: The Markov Transition Field (MTF) method is used to convert one-dimensional time-series vibration signals into two-dimensional feature images. Specifically, this includes: dividing the time-series signal into Q statistical quantile intervals based on amplitude; establishing a mapping relationship between observed values ​​and quantile levels; and constructing a state transition probability matrix. , of which elements Representing the quantile region Towards The probability of transition; generating a Markov transition matrix with state transition characteristics as a two-dimensional feature image; S3: Input the two-dimensional feature image into the SE-IShufflenetV2 lightweight neural network model for fault diagnosis. The model includes: Based on the IShufflenetV2 network architecture, a 3×3 convolutional kernel is used instead of a 5×5 convolutional kernel; the SENet (Squeeze-and-Excitation Network) attention mechanism module is introduced to adaptively weight the feature maps of different channels; and the Hardswish activation function is used instead of the traditional ReLU activation function. S4: Output fault diagnosis results, including fault type and confidence level.

2. The method according to claim 1, characterized in that, The acquisition parameters of the vibration signal in step S1 are: sampling frequency 5.12kHz-12kHz, using IEPE piezoelectric accelerometer, sensitivity 100mv / g, frequency range 0.5-10000Hz.

3. The method according to claim 1, characterized in that, The MTF parameters mentioned in step S2 are determined through experimental optimization and include: the number of quantile intervals Q, the image size, and the size of the overlapping sampling window.

4. The method according to claim 1, characterized in that, The SE-IShufflenetV2 model described in step S3 also includes: adding an extra set of 3×3 convolutional kernels to the basic unit to achieve the receptive field effect of 5×5 convolutional kernels; using channel shuffle operation to enhance feature fusion; and connecting a fully connected layer and a Softmax classifier after the global average pooling layer.

5. The method according to claim 1, characterized in that, After step S1 and before step S2, a signal preprocessing step is also included: noise reduction of the vibration signal is performed using the ICEEMDAN combined correlation coefficient-adaptive wavelet thresholding method, specifically including: S1.1: The improved adaptive noise complete set empirical mode decomposition (ICEEMDAN) is used to decompose the noisy signal into several intrinsic mode function (IMF) components; S1.2: Calculate the Pearson correlation coefficient between each IMF component and the original signal, and remove spurious components with correlation coefficients below a set threshold; S1.3: The effective IMF components after screening are denoised using an adaptive wavelet thresholding method; S1.4: Reconstruct the denoised IMF components to obtain the denoised vibration signal.

6. The method according to claim 5, characterized in that, The formula for calculating the Pearson correlation coefficient in step S1.2 is as follows: in, The original signal, For IMF components, The arithmetic mean of the original signal. For the first The arithmetic mean of the IMF components. This represents the total number of signal sampling points. For the first The Pearson correlation coefficient between each IMF component and the original signal is used; a threshold range of 0.3-0.5 is set, and IMF components below this threshold are identified as noise-dominant components and are removed.

7. The method according to claim 5, characterized in that, The adaptive wavelet thresholding method described in step S1.3 uses a soft thresholding function: In the formula, For wavelet decomposition of the 1st The first layer Wavelet coefficients, These are the wavelet coefficients after thresholding. For the corresponding number The adaptive threshold of each IMF component is dynamically adjusted according to the noise level of each IMF component.

8. An escalator fault diagnosis system based on Markov transfer fields and lightweight neural networks, characterized in that, include: Data acquisition module: including IEPE piezoelectric accelerometer array, NI PXIe-4492 dynamic signal acquisition card, NI PXIe-1078 measurement and control platform and PXIe-8840 embedded controller, used to acquire vibration signals from the main drive end of the escalator in real time; Data transmission and storage module: Enables cloud migration of feature matrices via TCP / IP data channel, and builds a fault database that supports remote access; Signal processing module: integrated into the host computer monitoring platform, used to execute the fault diagnosis method according to any one of claims 1-7, including MTF feature encoding, ICEEMDAN noise reduction processing and SE-IShufflenetV2 model inference; Interactive display module: Built on the LabVIEW development environment, it realizes real-time monitoring and display of vibration signals in the time and frequency domain, visualization of fault diagnosis results, and query of historical data.

9. The system according to claim 8, characterized in that, The signal processing module adopts a joint programming architecture of LabVIEW and Python. LabVIEW is responsible for data acquisition, storage and interface interaction, while Python is responsible for MTF image encoding, deep learning model inference and noise reduction algorithm execution. The two achieve data interaction through TCP / IP protocol or shared memory.

10. The system according to claim 8, characterized in that, The SE-IShufflenetV2 model is deployed by further compressing the model size using model quantization or pruning techniques to meet the real-time inference requirements of the embedded controller PXIe-8840.

Images