| With the advent of Io T technology and Industry 4.0,the industrial sector is undergoing a significant upgrade from traditional to intelligent.As a result,the mechanical equipment used in industry is increasingly digital and complex.Rolling bearings,the most common components in machinery and equipment,are susceptible to failure under harsh working conditions such as prolonged exposure to high temperatures and heavy loads.Such failures can lead to machinery and equipment performance degradation,resulting in unplanned downtime,economic losses,and safety hazards.Consequently,precise and intelligent fault diagnosis of rolling bearings is crucial for efficient industrial digital production.Rolling bearings often operate in the presence of strong background noise and are prone to failure,which makes it challenging to extract fault features.To address this issue,this paper proposes a dual path fused shrinkage denoising temporal convolution network intelligent fault diagnosis algorithm.(1)A novel approach has been proposed to address the issue of Temporal Convolutional Network(TCN)models being unable to extract impulse fault features effectively from highly noisy vibration signals in real industrial environments.This approach involves incorporating an attention mechanism into a systolic noise-reducing TCN model.The TCN residual unit is modified by embedding a soft threshold,which enables the nonlinear mapping of features.Additionally,a one-dimensional convolutional attention mechanism is introduced to adaptively learn the soft threshold,thus improving the relationship between channels and achieving the abstract mapping of features.By doing so,the TCN can focus more on the relevant fault information in the signal while eliminating irrelevant noisy redundant features.Furthermore,a batch normalization is introduced into the TCN module to enhance the generalization of the model.(2)In order to overcome the issue of limited diagnostic efficiency caused by the inability of single time-domain information to fully reflect the differential characteristics of fault vibration signals,a hybrid domain TCN fault diagnosis algorithm is proposed.This algorithm leverages parallel two-way convolution to effectively learn fault information from both the time domain and time-frequency domain.However,general time-frequency domain feature extraction is often tedious and reliant on human experience,resulting in complexity and uncertainty.To address this issue,a Laplace wavelet convolution method is proposed to replace the traditional time-frequency domain signal acquisition approach,enabling end-to-end learning of time-frequency domain features.In the first layer of convolution,a wider convolution kernel is utilized for time-domain feature learning to reduce noise interference.After the initial extraction of both time-domain and time-frequency domain features using two-way convolution,the TCN units are stacked separately to capture the temporal long-term dependence of the signal.Subsequently,the extracted mixed-domain features are converged,and global average pooling is applied to output the global mapping and achieve fault classification.The ablation experiments and noise-resistance comparison experiments are conducted to evaluate the effectiveness of a dual path fused shrinkage denoising TCN intelligent fault diagnosis algorithm using three bearing datasets from Case Western Reserve University(CWRU),Jiangnan University(JNU),and PT300.The signals are intentionally added with Gaussian white noise to validate the algorithm’s robustness and stability.The results of the experimental comparison demonstrated that the proposed algorithm significantly improve the accuracy of fault diagnosis in noisy environments. |
