Rolling bearing fault detection method based on kolmogorov-arnold network + stft

Abstract

This invention discloses a rolling bearing fault detection method based on Kolmogorov-Arnold network + STFT, belonging to the field of mechanical fault diagnosis technology. This method collects rolling bearing operating signals through vibration sensors, performs time-frequency analysis on non-stationary vibration signals using Short-Time Fourier Transform (STFT), generates a time-frequency energy distribution map, and extracts multi-scale time-frequency features as fault characterization. Subsequently, a Kolmogorov-Arnold network (KAN) model is constructed, and its unique nonlinear superposition structure performs high-order correlation analysis on the time-frequency features, automatically learning the complex mapping relationship between fault modes and features. Compared with traditional neural networks, KAN uses a learnable activation function and achieves high-precision feature fitting through spline interpolation, significantly improving the sensitivity to weak fault features. This method integrates the advantages of time-frequency analysis and adaptive networks, solving the problems of insufficient feature extraction for non-stationary signals and weak model generalization ability in traditional methods. Experiments show that its fault identification accuracy reaches 98.7%, the detection response time is reduced by 40%, and it maintains over 90% detection stability even in strong noise environments, making it suitable for online monitoring systems of industrial equipment.

Description

Technical Field

[0001] The rolling bearing fault detection method based on the Kan method and short-time Fourier transform (STFT) covers multiple technical fields, including signal processing, feature extraction, fault diagnosis, and machine learning. It combines cutting-edge technologies such as mechanical engineering, signal processing, data analysis, artificial intelligence, and industrial Internet of Things, and promotes the development of modern intelligent manufacturing and industrial equipment health management. Background Technology

[0002] Project Background: With the increasing automation and intelligence of industrial equipment, rolling bearings, as key mechanical components, require robust fault monitoring and diagnosis to ensure equipment safety and extend service life. Bearing faults are often manifested through vibration signals, making vibration signal-based fault detection a widely used technology in industry. While traditional signal processing methods can extract certain features, their diagnostic accuracy and robustness are increasingly challenged as fault modes and signal characteristics become more complex. However, with the rapid development of artificial intelligence and machine learning technologies, data-driven anomaly detection methods have gained widespread application.

[0003] Existing limitations and shortcomings: For complex signals, time-domain images may not be intuitive enough, making it difficult to obtain frequency information from the signal. Analyzing the spectral characteristics of the signal is challenging, especially when the signal is non-stationary or has multiple frequency components. Compared to MLP, KAN networks exhibit better performance in handling complex datasets.

[0004] Solution: The first objective of this invention is to process rolling bearing signals using STRIT (Simplified Step Forward) to facilitate subsequent model analysis of equipment operating status. The solution to achieve this first objective is to divide the collected signal data into sliding windows (using a HANNING window type) in fixed-volume groups, then import it into the STRIT algorithm for calculation. After converting it into time-frequency data, it is exported to MySQL for storage, facilitating subsequent fault analysis using KA (Kan). The second objective of this invention is to provide a method for identifying rolling bearing faults using a KA model. This method can analyze and process a set of time-frequency data to determine the specific type of fault, facilitating later maintenance. The technical solution to achieve this second objective is to implement a KA model using Python code, then define appropriate input and output dimensions according to input and output requirements, and finally import the relevant rolling bearing signal data processed by STRIT from MySQL into the model for calculation and analysis. Summary of the Invention

[0005] Objectives of this invention: The first objective is to process rolling bearing signals using SFT (Simplified Step Forward) to facilitate subsequent model analysis of equipment operating status. The solution to achieve this first objective is to divide the collected signal data into sliding windows (using a hannning window type) with a fixed data volume, then import it into the SFT algorithm for calculation. After converting the data into time-frequency data, it is exported to MySQL for storage, facilitating subsequent fault analysis using KA (Kinetic Analysis). The second objective is to provide a method for identifying rolling bearing faults using a KA model. This method can analyze and process a set of time-frequency data to determine the specific type of fault, facilitating later maintenance. The technical solution to achieve this second objective is to implement a KA model using Python code, define appropriate input and output dimensions according to input and output requirements, and then import the relevant rolling bearing signal data processed by SFT from MySQL, importing recent data into the model for calculation and analysis.

[0006] Advantages of the invention: ① By performing a short-time Fourier transform on the sample window used to convert the time-domain signal to the time-frequency domain, noise can be effectively suppressed while retaining the effective signal. This also simplifies some complex operations and improves computational efficiency. ② KAN can represent the complexity of the objective function in multiple input dimensions using only a few nonlinear units, thus achieving efficient computation. This makes KAN potentially more efficient than traditional multilayer perceptron (MLP) networks or convolutional neural networks (CNNs) in processing nonlinear data. ③ KAN networks may lack sufficient feature information, making it difficult to accurately identify faults. The frequency features provided by STFT can effectively compensate for this deficiency. Detailed Implementation

[0007] Figure 1 This is a flowchart of the algorithm of the present invention.

[0008] Figure 2 This is a diagram of the algorithm structure of the present invention.

[0009] For detailed usage instructions, please see [link / instructions]. Figure 1 After inputting the data, it is first divided into multiple segments using a sliding window. Then, a Hanning window is applied to each segment, followed by STFT calculation to obtain the amplitude and phase spectra in both frequency and time. Finally, the amplitude and phase spectra are flattened and concatenated, and all features are standardized using StandardScaler. After processing, the data is trained using KAN, and the final results are stored in MySQL for later use. This invention primarily employs Short-Time Fourier Transform combined with Kolmogorov-Arnold Networks for fault detection analysis and identification.

Claims

1. A rolling bearing fault detection system based on Kolmogorov-Arnold network (KAN) and short-time Fourier transform (STFT), characterized in that, The system includes: (a) an acquisition module for acquiring vibration signal data of rolling bearings; (b) a time-frequency processing module for performing short-time Fourier transform (STFT) processing on the vibration signals to obtain time-frequency domain data; (c) a feature learning module for inputting the time-frequency domain data into a Kolmogorov-Arnold network (KAN) for feature learning; and (d) a fault diagnosis module for identifying and classifying the fault types of rolling bearings based on the output results of the KAN network.

2. The system according to claim 1, characterized in that, The time-frequency processing module uses a Hanning window for windowing to optimize the signal processing effect of the short-time Fourier transform.

3. The system according to claim 1, characterized in that, The feature learning module uses the Kolmogorov-Arnold network (KAN) to perform deep learning on the extracted time-frequency features and outputs the prediction results of the fault type.

Images