Mechanical fault judgment method and system based on feature extraction

Abstract

The invention relates to the technical field of machine learning, in particular to a mechanical fault judgment method and system based on feature extraction. The method comprises the steps of obtaining a video image; preprocessing the acquired video image to obtain an image frame; performing feature extraction on the obtained image frame, and calculating step information of the image for the obtained image frame by using a Sobel operator; calculating a step matrix unit according to the calculated image frame step information; utilizing the step matrix unit to obtain a feature vector of the whole image; and performing feature analysis on the extracted feature vectors through a machine learning algorithm to obtain a fault judgment result. The principle of analyzing the mechanical fault through the video is mainly based on the morphological characteristics of the equipment and the change of the step direction distribution. Abnormity of mechanical parts can be quickly found, and then whether equipment breaks down or not is judged. And in combination with a machine learning classifier or a threshold method, efficient mechanical fault automatic detection and monitoring can be realized.

Description

Technical Field

[0001] The present invention relates to the technical field of machine learning, and in particular to a mechanical fault determination method and system based on feature extraction. Background Art

[0002] Mechanical fault determination based on feature extraction is a technology that combines signal processing, computer vision and machine learning. It is widely used in the condition monitoring and predictive maintenance of industrial equipment. The core idea is to extract features from sensors or video data, identify the difference between normal and abnormal operating states, and thus determine the health status of mechanical equipment. The following are some of the more popular technical means and methods of this technology.

[0003] Existing image feature extraction methods mainly focus on edge detection, local texture analysis, shape description, and key point detection. Although these methods can still judge the status of the device through feature extraction, they often have limitations when dealing with problems such as illumination changes, scale changes, and rotation. For example, the following are some common image feature extraction methods and their principles: 1. Edge Detection Method Edge detection is one of the earliest techniques used in image processing and feature extraction. It aims to extract edge information of an image by detecting areas where brightness or color changes significantly. Common edge detection operators include: 2. Local texture analysis method The extraction of local texture features is usually used for texture recognition, classification and segmentation of images. This type of method extracts local features of images by analyzing the relationship between pixel neighborhoods.

[0004] 3. Shape Description Method Early shape description methods were mainly used to identify the contours and geometric features of objects in images. Common methods include: 4. Key point detection and description Key point detection and description are also core technologies in current image matching and recognition. They detect salient points (key points) in the image and use feature descriptors to describe the local information of these points. Common methods include: 5. Global feature extraction method Global feature extraction methods extract features by analyzing the overall properties of the entire image and are often used for scene classification and overall image description.

[0005] In existing image feature extraction methods, pixel intensity values ​​are directly used to describe image features, but these intensity values ​​are very sensitive to changes in lighting and posture, resulting in unstable features and difficulty in performing well in practical applications. For example, simple edge detection algorithms are prone to failure when lighting changes, image rotation, scale changes, etc. Summary of the invention

[0006] In order to solve the above-mentioned problems, the present invention provides a mechanical fault determination method and system based on feature extraction.

[0007] In a first aspect, the present invention provides a mechanical fault determination method based on feature extraction, which adopts the following technical solution: A mechanical fault determination method based on feature extraction, comprising: Acquire video images; Preprocessing the acquired video image to obtain an image frame; Performing feature extraction on the obtained image frame, wherein the step information of the image is calculated using the Sobel operator on the obtained image frame; Calculate the step matrix unit according to the calculated image frame step information; The feature vector of the entire image is obtained using the ladder matrix unit; The extracted feature vectors are analyzed using a machine learning algorithm to obtain fault judgment results.

[0008] Furthermore, the preprocessing of the acquired video image to obtain the image frame includes extracting the image frame from the video image at a fixed frame rate, and performing grayscale and noise reduction processing on each frame of the image.

[0009] Furthermore, the step information of the image frame obtained by using the Sobel operator is calculated, including using the Sobel operator to calculate the brightness changes of each pixel in the image frame in the horizontal and vertical directions, that is, the horizontal steps and the vertical steps, and calculating the step amplitude and step direction at each pixel. Further, the step matrix unit is calculated according to the calculated image frame step information, including dividing the image frame into grid blocks, counting the step directions of all pixels in each grid block according to a predefined direction interval to form a step matrix unit, and the value of each direction interval is the accumulation of all pixel step amplitudes in the corresponding direction, expressed as: Where: M(x,y) is the step amplitude of the pixel; θ(x,y) is the step direction of the pixel; θbin is the direction interval of the matrix unit; w(θ(x,y),θbin) is a weight function used to assign the step direction θ(x,y) to adjacent direction intervals.

[0010] Furthermore, the step of calculating the step matrix unit according to the calculated image frame step information also includes normalizing a plurality of adjacent grid blocks, combining the plurality of grid blocks into a block, and normalizing the matrix unit in each block.

[0011] Furthermore, the step matrix unit is used to obtain the feature vector of the entire image, including connecting the normalized step matrix unit vectors of all blocks in series to obtain the feature vector of the entire image. The feature vector of the entire image is expressed as: , where the image is divided into N blocks, each block contains n grid block cells, and the feature vector of each block is the normalized result of all grid block cell matrix units in the block.

[0012] Furthermore, the feature analysis of the extracted feature vectors by a machine learning algorithm is performed to obtain a fault judgment result, including training a support vector machine (SVM) classifier using existing normal and fault samples, using the classifier to predict faults, and outputting the result.

[0013] In a second aspect, a mechanical fault determination system based on feature extraction includes: The data acquisition module is configured to acquire video images; The preprocessing module is configured to preprocess the acquired video image to obtain an image frame; The feature extraction module is configured to perform feature extraction on the obtained image frame, wherein the step information of the image is calculated using the Sobel operator for the obtained image frame; the step matrix unit is calculated according to the calculated step information of the image frame; and the feature vector of the entire image is obtained using the step matrix unit; The fault judgment module is configured to perform feature analysis on the extracted feature vector through a machine learning algorithm to obtain a fault judgment result.

[0014] In a third aspect, the present invention provides a computer-readable storage medium storing a plurality of instructions, wherein the instructions are suitable for being loaded and executed by a processor of a terminal device for the mechanical fault determination method based on feature extraction.

[0015] In a fourth aspect, the present invention provides a terminal device including a processor and a computer-readable storage medium, wherein the processor is used to implement various instructions; the computer-readable storage medium is used to store multiple instructions, wherein the instructions are suitable for being loaded by the processor and executing the method for determining a mechanical fault based on feature extraction.

[0016] In summary, the present invention has the following beneficial technical effects: 1. The principle of analyzing mechanical failures through video in the present invention is mainly based on the changes in the morphological characteristics of the equipment and the distribution of the step direction. When a mechanical device fails, the characteristics will significantly deviate from the normal operating state. By analyzing the changes in these characteristics, the abnormality of the mechanical parts can be quickly discovered, and then it can be determined whether the equipment has failed. Combined with machine learning classifiers or threshold methods, efficient automatic detection and monitoring of mechanical failures can be achieved.

[0017] 2. The present invention calculates the distribution of step directions in each part of the image and uses direction information to describe local structural features instead of relying directly on pixel intensity. Therefore, it is more robust to illumination changes and local deformations. Even if the illumination intensity changes, the perimeter and step directions of the parts in the image usually remain stable.

[0018] 3. The present invention divides the image into multiple small "cells" and "blocks" and calculates the step matrix unit in each cell. This method can capture the local structure of the object in the image and enhance the robustness to changes in illumination and contrast by normalizing within the block. BRIEF DESCRIPTION OF THE DRAWINGS

[0019] Figure 1 It is a schematic diagram of a mechanical fault determination method based on feature extraction according to Example 1 of the present invention. DETAILED DESCRIPTION

[0020] The present invention is further described in detail below in conjunction with the accompanying drawings.

[0021] Example 1 Reference Figure 1 , a mechanical fault determination method based on feature extraction in this embodiment includes: Acquire video images; Preprocessing the acquired video image to obtain an image frame; Performing feature extraction on the obtained image frame, wherein the step information of the image is calculated using the Sobel operator on the obtained image frame; Calculate the step matrix unit according to the calculated image frame step information; The feature vector of the entire image is obtained using the ladder matrix unit; The extracted feature vectors are analyzed using a machine learning algorithm to obtain fault judgment results.

[0022] Among them, the method of mechanical fault judgment based on feature extraction is an algorithm that extracts local features through the distribution of image steps. It is mainly used to capture the edges, shapes and contours of objects, and has good robustness to changes in illumination and scale. Using this method to analyze mechanical faults is to monitor the changes in the appearance of mechanical equipment, especially the changes in shape and contour caused by faults such as wear, deformation, and cracks. The following is the detailed process of using this method to determine mechanical faults: S1. Video frame acquisition and processing First, the video is decomposed into a series of image frames. Each frame is a static image and can be regarded as an independent picture. The application of this method in video analysis is based on frame-by-frame processing. By analyzing continuous image frames at multiple time points, the changing trend of mechanical failures over time can be observed.

[0023] step: Extract image frames from the video at a fixed frame rate (such as 10 or 30 frames per second).

[0024] Each frame is preprocessed, such as grayscale and noise reduction, to simplify the subsequent feature extraction process.

[0025] S2. Ladder calculation For each frame of the image, this method first calculates the image steps. Steps refer to the spatial rate of change of image brightness (grayscale value), which is used to describe the changes in edges or shapes in the image. Edge changes of mechanical equipment (such as cracks, wear, bending, etc.) will be reflected in the step information.

[0026] step: Calculate the brightness change (step) of each pixel in the horizontal direction (x-axis) and vertical direction (y-axis).

[0027] The Sobel operator is used to calculate the step magnitude and direction of each pixel.

[0028] Applications for fault detection: When the machine is running normally, the step distribution of its shape edge should be stable and regular (such as rotating shaft, gear edge, etc.).

[0029] When a device fails (such as surface damage or abnormal vibration), the step changes become irregular, such as the edge is no longer a smooth curve, but is broken or deformed.

[0030] S3. Ladder Matrix Unit Calculation After obtaining the step information, this method divides the image into several small cells and calculates the matrix unit of the step direction in each cell. The matrix unit is used to describe the distribution of the directional steps in the area, thereby capturing the local shape and edge information.

[0031] step: Split the image into several small cells (usually 8x8 or 16x16 pixels in size).

[0032] The step direction of the pixels in each unit cell is quantified (eg, divided into 9 directions), and the corresponding step amplitudes are accumulated in the corresponding directions to form a step matrix unit.

[0033] Applications for fault detection: The surface of normal equipment or moving parts should have a consistent step direction distribution, for example, rotating parts should have a regular circular or arc-shaped step direction distribution.

[0034] When a device fails, this regular step direction distribution may become unbalanced or chaotic. For example, cracks and worn areas of mechanical parts will show significant step anomalies in certain directions.

[0035] The principle of step calculation is to calculate the step directions of multiple blocks inside the image and represent these step information with matrix units. The following is the algorithm logic formula of step calculation in this method.

[0036] 1) Image ladder calculation The first step of feature extraction is to calculate the horizontal and vertical steps of the image at each pixel. In order to obtain the steps, a simple difference operator is usually used to approximate the steps, such as the common Sobel operator.

[0037] 1.1) Horizontal step Gx Perform a horizontal difference operation on the image I(x,y) to approximate the horizontal steps: Among them, I(x+1,y) is the pixel value on the right, and I(x−1,y) is the pixel value on the left.

[0038] 1.2) Vertical step Gy Perform a vertical difference operation on the image I(x,y) to approximate the vertical steps: Among them, I(x,y+1) is the pixel value below, and I(x,y−1) is the pixel value above.

[0039] 2) Calculation of step amplitude and direction Calculate the step magnitude and step direction at each pixel.

[0040] 2.1) Step amplitude M(x,y) The step amplitude represents the edge strength, which is the composite step of the horizontal step Gx(x,y) and the vertical step Gy(x,y), and can be calculated using the Euclidean norm: 2.2) Step direction θ(x,y) The step direction indicates the direction of the edge, usually defined as an angle relative to the horizontal axis: The angle of the step direction is usually constrained to be between 0∘ and 180∘, ignoring the sign of the step since the edge is identical in two opposite directions.

[0041] S4. Feature vector generation In order to obtain the features of the entire image, this method combines the ladder matrix units of each unit cell to form a feature vector. In order to enhance robustness, this method also normalizes multiple units to reduce the influence of factors such as lighting and contrast.

[0042] step: By grouping adjacent cells into a block, the matrix cells within the block are normalized.

[0043] The entire image is traversed in a sliding window manner to generate a global feature vector of the image.

[0044] Applications for fault detection: A normal mechanical device should have a smooth and regular distribution of feature vectors, describing the consistency of its edges, shapes, and contours.

[0045] When a fault occurs, the local features will change significantly. For example, cracks or mechanical wear will cause abnormal local features, which are manifested as significant deviations in the step direction in some areas, or the features of some blocks are inconsistent with the surrounding environment.

[0046] 1) Ladder Matrix Unit The core of this method is to aggregate the statistical information of the step direction into matrix units. The image is divided into multiple small grid blocks (cells), and the distribution of the pixel step direction is counted in each cell, and divided into several direction intervals (for example, 9 directions, each interval is 20∘).

[0047] 1.1) Matrix unit construction For each cell, the step directions of all pixels in the cell are counted according to the predefined direction interval to form a matrix unit. The value of each direction interval can be the accumulation of the step amplitudes of all pixels in the corresponding direction: in: M(x,y) is the step amplitude of the pixel.

[0048] θ(x,y) is the step direction of the pixel.

[0049] θbin is the orientation bin of the matrix cells.

[0050] w(θ(x,y),θbin) is a weight function used to assign step directions θ(x,y) to adjacent direction bins (e.g., bilinear interpolation).

[0051] 2) Block Normalization In order to enhance the robustness to changes in illumination and contrast, features are usually normalized for multiple adjacent cells after calculation. This step is done by grouping multiple cells into a block and then normalizing the matrix units in each block.

[0052] The normalization method usually uses L2 norm normalization: in: H is the echelon matrix unit vector within the block.

[0053] ϵ is a small constant used to avoid division by zero.

[0054] 3) Feature vector The last step of feature vector generation is to concatenate the normalized matrix unit vectors of all blocks to obtain the feature vector of the entire image.

[0055] Assuming that the image is divided into N blocks and each block contains n cells, the feature vector of the entire image can be expressed as: The eigenvector Hblocki of each block is the normalized result of all cell matrix units in the block.

[0056] S5. Fault Detection and Classification After extracting the features of each frame, these features can be analyzed through machine learning or deep learning algorithms to determine whether the device is faulty.

[0057] Failure mode identification: Classifier training: Through supervised learning methods (such as SVM, decision tree or deep learning network), a classifier can be trained using video data of normal and faulty states. The classifier learns the feature distribution of normal equipment and can identify feature distributions that deviate greatly from normal conditions, thereby determining faults.

[0058] Threshold method: In some simple applications, a feature change threshold can be set. When the feature change exceeds the threshold, it is considered that a mechanical failure has occurred.

[0059] Real-time monitoring: For real-time video monitoring, the features of each frame can be analyzed and combined with time series information to monitor the operating status of the equipment and identify the time point when the fault occurs.

[0060] Among them, example scenario: conveyor belt anomaly detection The size of the image IROI is 8×8 pixels: 1. Ladder calculation Horizontal direction: Calculation results: Vertical direction: Calculation results: Gradient increase formula: Substitute the data: The pixel enhancement value is: The pixel directions are: 2. Matrix unit calculation The staircase direction is divided into 9 bins (each bin is 20°). The pixel direction 45∘ corresponds to the 3rd bin.

[0061] Construct matrix cells (assuming each cell is 4×4): Among them, 224.64=16*14.14 is the accumulation of the step amplitude in this direction.

[0062] 3. Block Normalization Assuming that a block contains 2×2 cells, the normalization formula of the block feature vector is: Substitute the data: 4. Classifier Training The existing normal and fault samples are: Train a Support Vector Machine (SVM) classifier: 5. Fault Detection For new image feature extraction: Use the classifier to predict: If y^==0, it means normal; if y^=1, it means fault.

[0063] S6. Fault trend analysis By analyzing features in multiple frames of video, it is possible to identify trends in the development of mechanical failures. For example, increasingly drastic step changes in certain areas over time may indicate that mechanical wear is gradually increasing. This trend analysis can help predict failures and be used for preventive maintenance.

[0064] application: Equipment surface damage: By analyzing changes in surface edges, early signs of cracks or damage to mechanical surfaces can be captured.

[0065] Unbalanced Rotating Components: For rotating equipment, this method can detect abnormal vibration or deformation of rotating shafts or gears, which are often early warning signs of equipment failure.

[0066] Example 2 This embodiment provides a mechanical fault determination system based on feature extraction, including: The data acquisition module is configured to acquire video images; The preprocessing module is configured to preprocess the acquired video image to obtain an image frame; The feature extraction module is configured to perform feature extraction on the obtained image frame, wherein the step information of the image is calculated using the Sobel operator for the obtained image frame; the step matrix unit is calculated according to the calculated step information of the image frame; and the feature vector of the entire image is obtained using the step matrix unit; The fault judgment module is configured to perform feature analysis on the extracted feature vector through a machine learning algorithm to obtain a fault judgment result.

[0067] A computer-readable storage medium stores a plurality of instructions, wherein the instructions are suitable for being loaded and executed by a processor of a terminal device, for a mechanical fault determination method based on feature extraction.

[0068] A terminal device includes a processor and a computer-readable storage medium, wherein the processor is used to implement various instructions; the computer-readable storage medium is used to store multiple instructions, wherein the instructions are suitable for being loaded by the processor and executing the mechanical fault determination method based on feature extraction.

[0069] The above are all preferred embodiments of the present invention, and are not intended to limit the protection scope of the present invention. Therefore, any equivalent changes made based on the structure, shape, and principle of the present invention should be included in the protection scope of the present invention.

Claims

1. A mechanical fault determination method based on feature extraction, characterized in that: include: Acquire video images; Preprocessing the acquired video image to obtain an image frame; Performing feature extraction on the obtained image frame, wherein the step information of the image is calculated using the Sobel operator on the obtained image frame; Calculate the step matrix unit according to the calculated image frame step information; The feature vector of the entire image is obtained using the ladder matrix unit; The extracted feature vectors are analyzed using a machine learning algorithm to obtain fault judgment results.

2. A mechanical fault determination method based on feature extraction according to claim 1, characterized in that: The preprocessing of the acquired video image to obtain the image frame includes extracting the image frame from the video image at a fixed frame rate, and performing grayscale and noise reduction processing on each frame of the image.

3. A mechanical fault determination method based on feature extraction according to claim 2, characterized in that: The step information of the image frame obtained by using the Sobel operator is calculated, including using the Sobel operator to calculate the brightness change of each pixel in the image frame in the horizontal and vertical directions, that is, the horizontal step and the vertical step, and calculating the step amplitude and step direction at each pixel.

4. A mechanical fault determination method based on feature extraction according to claim 3, characterized in that: The step matrix unit is calculated according to the calculated image frame step information, including dividing the image frame into grid blocks, counting the step directions of all pixels in each grid block according to a predefined direction interval to form a step matrix unit, and the value of each direction interval is the accumulation of all pixel step amplitudes in the corresponding direction, expressed as: Where: M(x,y) is the step amplitude of the pixel; θ(x,y) is the step direction of the pixel; θbin is the direction interval of the matrix unit; w(θ(x,y),θbin) is a weight function used to assign the step direction θ(x,y) to adjacent direction intervals.

5. A mechanical fault determination method based on feature extraction according to claim 4, characterized in that: The step of calculating the step matrix unit according to the calculated image frame step information also includes normalizing a plurality of adjacent grid blocks, combining the plurality of grid blocks into a block, and normalizing the matrix unit in each block.

6. A mechanical fault determination method based on feature extraction according to claim 5, characterized in that: The step matrix unit is used to obtain the feature vector of the entire image, including connecting the normalized step matrix unit vectors of all blocks in series to obtain the feature vector of the entire image. The feature vector of the entire image is expressed as: , where the image is divided into N blocks, each block contains n grid block cells, and the feature vector of each block is the normalized result of all grid block cell matrix units in the block.

7. A mechanical fault determination method based on feature extraction according to claim 6, characterized in that: The method of performing feature analysis on the extracted feature vectors through a machine learning algorithm to obtain a fault judgment result includes training a support vector machine (SVM) classifier using existing normal and fault samples, using the classifier to predict faults, and outputting the result.

8. A mechanical fault determination system based on feature extraction, characterized in that: include: The data acquisition module is configured to acquire video images; The preprocessing module is configured to preprocess the acquired video image to obtain an image frame; The feature extraction module is configured to perform feature extraction on the obtained image frame, wherein the step information of the image is calculated using the Sobel operator for the obtained image frame; the step matrix unit is calculated according to the calculated step information of the image frame; and the feature vector of the entire image is obtained using the step matrix unit; The fault judgment module is configured to perform feature analysis on the extracted feature vector through a machine learning algorithm to obtain a fault judgment result.

9. A computer-readable storage medium storing a plurality of instructions, characterized in that: The instructions are suitable for being loaded by a processor of a terminal device and executing the method according to claim 1 .

10. A terminal device, comprising a processor and a computer-readable storage medium, wherein the processor is used to implement each instruction; and the computer-readable storage medium is used to store multiple instructions, characterized in that: The instructions are suitable for being loaded by a processor and executing the method as claimed in claim 1 .

Images