CN-122023297-A - Infrared bearing fault diagnosis method and equipment based on edge-guided mutual information bottleneck

CN122023297ACN 122023297 ACN122023297 ACN 122023297ACN-122023297-A

Abstract

The invention belongs to the technical field related to mechanical fault diagnosis, and discloses an infrared bearing fault diagnosis method and equipment based on an edge-guided mutual information bottleneck, wherein the method comprises the steps of (1) constructing a data set based on infrared thermal images of bearing fault categories so as to train a fault diagnosis network; the fault diagnosis network comprises a main network, an edge perception branch, an edge guiding mutual information bottleneck module, an edge perception decoder and a region perception classification head, wherein the main network extracts multi-scale features based on infrared images of a bearing to be diagnosed, the edge perception branch predicts an edge probability map based on shallow features, the edge guiding mutual information bottleneck module sequentially carries out channel weighting and multi-scale bottleneck compression on the multi-scale features to obtain pure features, the edge perception decoder generates a segmentation mask based on the pure features and the edge probability map, and the region perception classification head carries out weighted pooling on the pure features by using the segmentation mask to obtain fault types. The invention improves the fault diagnosis precision.

Inventors

SHEN WEIMING
ZHONG SHUN
SONG WENBIN
YANG XIONGFEI

Assignees

华中科技大学

Dates

Publication Date: 20260512
Application Date: 20260120

Claims (10)

1. An infrared bearing fault diagnosis method based on an edge-guided mutual information bottleneck is characterized by comprising the following steps: (1) Constructing a data set containing a segmentation mask, an edge map and a classification label based on the infrared thermal image of each bearing fault type, and training a fault diagnosis network by adopting the data set, wherein the fault diagnosis network comprises a main network, an edge perception branch, an edge guidance mutual information bottleneck module, an edge perception decoder and a region perception classification head; (2) The method comprises the steps of extracting multi-scale features based on an infrared image of a bearing to be diagnosed, predicting an edge probability map by an edge perception branch based on shallow features in the multi-scale features, sequentially carrying out channel weighting and multi-scale bottleneck compression on the multi-scale features by an edge guiding mutual information bottleneck module to obtain pure features, generating a segmentation mask of a fault region by an edge perception decoder based on the obtained pure features and the edge probability map, and carrying out weighted pooling on the pure features by a region perception classification head by using the segmentation mask to obtain fault categories.
2. The infrared bearing fault diagnosis method based on the edge-guided mutual information bottleneck as set forth in claim 1, wherein the fault diagnosis network is trained end-to-end based on a total loss function.
3. The method for diagnosing infrared bearing faults based on the edge-guided mutual information bottleneck as claimed in claim 2, wherein the total loss function comprises a segmentation loss, an edge loss, a classification loss and a mutual information loss, and the formula is as follows: Wherein, the Is edge loss; loss for segmentation; Is a classification loss; Is the mutual information loss; The weight coefficient of the segmentation loss; The weight coefficient for edge loss; the weight coefficient of the classification loss; Is a weight coefficient of mutual information loss.
4. The method for diagnosing an infrared bearing failure based on an edge-directed mutual information bottleneck of claim 1, wherein the edge-directed mutual information bottleneck module calculates a mutual information loss between the purity feature and the edge probability map.
5. The infrared bearing fault diagnosis method based on edge-guided mutual information bottleneck as recited in claim 1, wherein the edge-aware decoder directs the diagnosis network to focus on physical boundaries by explicitly injecting edge position offsets in the self-attention mechanism.
6. The infrared bearing fault diagnosis method based on the edge-guided mutual information bottleneck as set forth in claim 5, characterized in that the up-sampled edge probability map is directly superimposed on self-attention Thereby directing the edge-aware decoder to focus on the physical boundary, generating a final segmentation mask 。
7. The infrared bearing fault diagnosis method based on the edge-guided mutual information bottleneck as set forth in claim 6, wherein the region-aware classification head utilizes a segmentation mask And the edge perception branches adopt gradient weighting loss functions to calculate gradient amplitude values of input images, and the calculated gradient amplitude values of the input images are used for dynamically adjusting pixel weights of cross entropy loss.
8. The infrared bearing fault diagnosis method based on the edge-guided mutual information bottleneck as set forth in any one of claims 1-7, wherein the edge-guided mutual information bottleneck module comprises an edge-guided channel attention unit, and a calculation formula of the edge-guided channel attention unit is as follows: Wherein, the Is an edge probability map The value at coordinates (h, w), To input the feature map of the EMIB module, in the first place Values at channel, coordinates (h, w), For the global foreground feature vector based on edge weighting, Global background feature vectors based on background weighting; Is a tiny constant that prevents the denominator from being zero; , For the channel weight to be a function of the channel weight, As the weight of the c-th channel, , Is a full-connection layer weight matrix of the multi-layer perceptron, Is a function of the activation of the ReLU, Is a Sigmoid activation function that is activated by, Is a characteristic diagram after channel recalibration; The edge-guided mutual information bottleneck module further comprises a mutual information projection unit, and the mutual information projection unit calculates mutual information loss between the pure feature and the edge probability map by utilizing InfoNCE loss functions.
9. An infrared bearing fault diagnosis system based on an edge-guided mutual information bottleneck is characterized by comprising a memory and a processor, wherein the memory stores a computer program, and the processor executes the infrared bearing fault diagnosis method based on the edge-guided mutual information bottleneck according to any one of claims 1-8 when executing the computer program.
10. A computer readable storage medium, characterized in that it stores machine executable instructions that, when invoked and executed by a processor, cause the processor to implement the infrared bearing fault diagnosis method based on edge-guided mutual information bottleneck of any one of claims 1-8.

Description

Infrared bearing fault diagnosis method and equipment based on edge-guided mutual information bottleneck Technical Field The invention belongs to the technical field related to mechanical fault diagnosis, and particularly relates to an infrared bearing fault diagnosis method and equipment based on an edge guiding mutual information bottleneck. Background In modern industrial production systems, rotating machinery (e.g., wind generators, numerically controlled machine tools, gas turbines, etc.) is a central source of power. The rolling bearing is used as a joint of the rotary machine, and the health state directly determines the operation safety and efficiency of the whole equipment. Statistics show that about 40% -50% of mechanical failures result from bearing failures. Once the bearing fails and is not found in time, the light weight causes the performance of the equipment to be reduced and the production to be stopped, and the heavy weight causes catastrophic safety accidents and huge economic losses. Therefore, the development of efficient and accurate bearing fault diagnosis and health management (PHM) research has important practical significance for guaranteeing industrial safety production. At present, a method based on vibration signal analysis is the main stream means which is most widely applied in the field of mechanical fault diagnosis and has the most mature technology. According to the method, vibration data are collected through a piezoelectric acceleration sensor arranged on the surface of equipment, and fault characteristics are extracted through time-frequency analysis or a deep learning algorithm. However, as industrial equipment moves toward high speed, heavy duty, and complex, conventional vibration diagnostics face the increasingly severe limitations of, first, inherent drawbacks of contact measurement. The vibration sensor must be closely attached to the surface of the apparatus by means of a magnetic mount, screws or glue, which is difficult to implement at high temperature, high pressure, high rotational speed or moving parts such as the rotating shaft itself, and the installation process may require a stop, affecting the production efficiency. Second, attenuation and distortion of the signal transmission path. Impact signals generated by fault sources (such as bearing inner rings) can be received by the sensor through transmission of multi-layer mechanical structures such as bearing blocks, shells and the like. The complex transmission path can cause serious attenuation of high-frequency impact components, is extremely easy to be interfered by background signals such as gear meshing, environmental noise and the like, and has extremely low signal to noise ratio, so that the early weak fault characteristics are difficult to extract. In addition, the transmission and storage of large amounts of data is stressed. In order to capture the high-frequency fault characteristics, the vibration signals generally need extremely high sampling rate, mass data are generated, and great challenges are brought to real-time transmission and edge calculation of the industrial Internet of things. In contrast, infrared thermography technology (Infrared Thermography, IRT) has shown unique advantages in the field of fault diagnosis as an emerging non-contact detection means. According to thermodynamic principles, mechanical failures (e.g., wear, fatigue, poor lubrication) are often accompanied by energy dissipation, manifesting as abnormal temperature rise. Infrared thermography has the significant advantages of first, non-contact remote detection. The thermal infrared imager does not need to contact the surface of equipment, can monitor rotating, electrified or high-temperature equipment in real time outside a safe distance, does not interfere with the normal operation of the equipment completely, and really realizes 'measuring and walking at once'. Second, intuitive two-dimensional field distribution information. Unlike a one-dimensional time series of vibration signals, infrared images provide a two-dimensional temperature field distribution. The fault is not only represented by numerical value change, but also by specific hot spot morphology and texture structure, namely 'what you see is what you get', and the spatial position and severity of the fault can be intuitively reflected. Thirdly, early warning ability is strong. In the early stage of mechanical property degradation, abnormal changes in the temperature field often occur prior to severe vibration or noise, which helps to achieve early detection and prevention of faults. Although infrared diagnosis has obvious advantages, the combination of infrared diagnosis and deep learning for automatic diagnosis still has technical bottlenecks. Due to the inherent thermal diffusion effect of infrared imaging, the boundary between the fault region (hot spot) and the background usually presents a 'diffuse shape', and the lack of a clear texture gradient as