Research On Machine Health Detection Technology Based On Time-frequency Domain Analysis And Deep Learning

In recent years,machine health monitoring and fault diagnosis have become increasingly important in industrial automation and intelligent production.Rolling bearings are crucial for rotating machinery and are extensively utilized in mechanical equipment,serving a vital function in its operation.The safety of rotating mechanical systems is heavily influenced by the performance of their bearings,and the failure of bearings can result in the collapse of the entire machinery system.Therefore,this paper focuses on rolling bearings as the research subject and explores both traditional signal processing-based diagnostic methods and intelligent fault diagnosis based on convolutional neural networks.1.Resonance demodulation is a commonly used method for extracting fault features,but the difficulty lies in the selection of the parameters for the bandpass filter.The Fast Spectral Kurtosis algorithm can automatically select bandpass filter parameters,but its effectiveness in directly applying to signals is not obvious due to noise interference.The LMD algorithm can decompose signals,and by combining kurtosis and cross-correlation coefficient indicators to reconstruct the component signals,it can effectively denoise the original signal,increase the signal’s kurtosis value,and highlight the fault features.Therefore,a fault diagnosis method based on combining LMD denoising and fast spectral kurtosis is proposed.The method first uses the LMD algorithm based on kurtosis and cross-correlation criteria to reconstruct the original signal for the purpose of noise reduction.Then,fast spectral kurtosis is used to determine the optimal parameters for a bandpass filter,which is then applied.Ultimately,the filtered signal is squared to facilitate fault diagnosis.In Chapter 3,the proposed method was validated using both simulated signals and actual engineering signals.2.This paper proposes a dual-stream CNN model based on wavelet time-frequency spectrograms and FFT to address the problem of strong background noise and poor performance of intelligent diagnostic methods under variable operating conditions in bearing vibration data.The model acquires fault features from both FFT frequency domain data and wavelet time-frequency spectrograms through dual channels to improve feature extraction ability and noise resistance.To improve feature extraction,a larger convolution kernel can be used in the first layer of the lower-level onedimensional convolution,resulting in a larger receptive field.The two-dimensional feature data extracted by the upper-level convolution is flattened and concatenated with the lower-level one-dimensional data,which is then input into a fully connected layer.Finally,the concatenated one-dimensional feature data is classified using a Softmax classifier for fault diagnosis.The average accuracy of this method in variable operating condition experiments can reach 97.1%,and it also performs well in noise experiments,indicating good fault diagnosis performance.