State monitoring and fault diagnosis method and system based on artificial intelligence

Abstract

The invention relates to the technical field of image analysis and mode recognition, in particular to a state monitoring and fault diagnosis method and system based on artificial intelligence. The invention discloses a state monitoring and fault diagnosis method based on artificial intelligence. The method comprises the following steps: S1, acquiring oil sample type data, performing background purification processing, and then executing a high-precision panoramic imaging process; the X-direction shooting number and the Y-direction shooting number are obtained based on the two-axis motion platform; s2, generating a grid coordinate sequence covering the deposition area according to the X-direction shooting number and the Y-direction shooting number, and forming a line-by-line scanning path; and synthesizing a full-focal-plane image based on the scanning path, and outputting a fused panoramic depth panorama according to the synthesized full-focal-plane image. According to the invention, through background purification, panoramic imaging, improved Mask-R-CNN mode recognition model segmentation and multivariate index fusion, precise recognition of wear particles and intelligent early warning and grading of wear states are realized.

Description

Technical Field

[0001] This invention relates to the field of image analysis and pattern recognition technology, and in particular to a method and system for state monitoring and fault diagnosis based on artificial intelligence. Background Technology

[0002] Artificial intelligence (AI) technology, as a core driving force for industrial intelligence, demonstrates significant advantages in equipment condition monitoring and fault diagnosis by simulating human cognitive mechanisms to achieve feature learning and pattern recognition of complex data. Existing methods, based on deep learning models, perform feature extraction and anomaly detection on multi-source data such as vibration signals and temperature sequences. These methods have achieved good results in identifying typical faults such as bearing wear and gear fracture under single operating conditions, providing technical support for predictive maintenance of equipment.

[0003] However, existing technologies have significant limitations in analyzing oil wear particles. Traditional pattern recognition methods rely on manually designed features and static thresholds, making them ill-suited for complex scenarios such as dark oil background interference, variable wear particle morphology, and overlapping particles. This often leads to incomplete feature separation and blurred classification boundaries. Furthermore, existing methods lack the ability to collaboratively model the semantic features of multi-scale wear particles and dynamic wear trends. In comprehensive condition evaluation, they are easily limited by single indicators, thus restricting the accuracy and reliability of fault prediction.

[0004] Therefore, it is necessary to construct an intelligent diagnostic method that integrates multi-scale feature learning and dynamic evaluation mechanisms. By enhancing the extraction accuracy and classification adaptability of wear particle features under complex backgrounds, the intelligent level and engineering applicability of wear condition monitoring can be improved. Summary of the Invention

[0005] To overcome the inaccuracy in identifying wear particle characteristics under complex oil samples, this invention provides a condition monitoring and fault diagnosis method and system based on artificial intelligence.

[0006] The technical implementation scheme of the present invention is as follows: a condition monitoring and fault diagnosis method based on artificial intelligence, comprising the following steps: S1: After acquiring oil sample type data and performing background purification processing, execute a high-precision panoramic imaging process; obtain the number of images in the X direction and the number of images in the Y direction based on a two-axis motion platform; S2: Generate a grid coordinate sequence covering the deposition area based on the number of images taken in the X direction and the number of images taken in the Y direction, forming a line-by-line scanning path; synthesize a full-focal-area image based on the scanning path, and output a fused panoramic depth-of-view image based on the synthesized full-focal-area image; S3: Construct an improved Mask-R-CNN pattern recognition model and use the improved Mask-R-CNN pattern recognition model to obtain multi-scale primary feature maps; use the multi-scale feature fusion module of the improved Mask-R-CNN pattern recognition model to process the multi-scale primary feature maps, and determine the wear particle mask based on the multi-scale primary feature maps. S4: Execute the Canny edge detection algorithm on each wear particle mask to extract sub-pixel precision contours, determine feature vectors based on the sub-pixel precision contours and output the final category; obtain the number of wear particles of each type according to the final category, and judge the wear status by combining the total content of wear particles, the number of large wear particles and the size of the largest wear particle, and automatically generate a detection report.

[0007] Preferably, the high-precision panoramic imaging process following the acquisition of oil sample type data and background purification includes: Connect the oil sample bottle to the inlet and the transparent auxiliary oil bottle to the auxiliary oil inlet. Obtain the oil sample type data. If the oil sample type data is not a dark oil sample, execute the high-precision panoramic imaging process. If the oil sample type data is a dark oil sample, keep the excitation coil energized, close the oil sample passage of the inlet through the three-way valve, open the auxiliary oil passage of the auxiliary oil inlet through the three-way valve, and start the peristaltic pump to pump the transparent auxiliary oil into the detection channel at a preset flow rate. The transparent auxiliary oil flows through the surface of the deposition glass slide in the detection channel. When the volume of the pumped transparent auxiliary oil reaches a preset threshold, stop the peristaltic pump, complete the background purification, and execute the high-precision panoramic imaging process.

[0008] Preferably, obtaining the number of images in the X direction and the number of images in the Y direction based on the two-axis motion platform includes: The two-axis motion platform is controlled to move to the mechanical origin position to establish a motion reference; the ring-shaped LED surface light source is turned on to provide illumination for the deposition slide; the length and width values ​​of the deposition area on the deposition slide are read; the length and width values ​​of the single-shot field of view of the camera are read; the length value of the deposition area is divided by the length value of the single-shot field of view of the camera, and the result is rounded up to obtain the number of shots in the X direction; the width value of the deposition area is divided by the width value of the single-shot field of view of the camera, and the result is rounded up to obtain the number of shots in the Y direction.

[0009] Preferably, the step of generating a grid coordinate sequence covering the deposition area based on the number of images taken in the X direction and the number of images taken in the Y direction, forming a line-by-line scanning path, includes: Using the mechanical origin as the starting point for scanning, the step spacing in the X direction is the length of the field of view captured by the camera in a single shot, and the step spacing in the Y direction is the width of the field of view captured by the camera in a single shot. Based on the number of shots captured in the X and Y directions, a grid coordinate sequence covering the deposition area is generated using the grid point coordinate calculation formulas in the X and Y directions, forming a line-by-line scanning path. The formula for calculating the coordinates of grid points in the X direction is: ;in, For the first The X coordinates of the grid points The X-coordinate of the starting point of the scan, i.e., the X-coordinate of the machine origin position. This represents the length of the field of view captured by the camera in a single shot. Index the grid points, starting from 1 and counting the number of images taken in the X direction; The formula for calculating the coordinates of grid points in the Y direction is: ;in, For the first The Y coordinate of the grid point The Y-coordinate of the starting point of the scan, i.e., the Y-coordinate of the machine origin position. This represents the width of the field of view in a single shot taken by the camera. Index the grid points, starting from 1 and counting the number of images taken in the Y direction; Traverse all ( , Combine them to generate a grid coordinate sequence.

[0010] Preferably, the step of synthesizing a full-focal-plane image based on the scanning path and outputting a fused panoramic depth-of-view image based on the synthesized full-focal-plane image includes: For each grid coordinate point in the scanning path, the following steps are performed: control the two-axis motion platform to move to the corresponding position of the grid coordinate point, control the two-axis motion platform to move step by step in the Z-axis direction according to the preset step size, trigger the camera to acquire images at each Z-axis position, construct a multi-focal image sequence according to the Z-axis position order, and select the pixel value with the highest contrast from all multi-focal images for each pixel position in the multi-focal image sequence to synthesize a full-focal image. The SIFT feature point matching algorithm is used to calculate the overlapping region and transformation matrix between adjacent full-focal plane images; all full-focal plane images are projected and aligned according to the transformation matrix; a weighted average fusion algorithm is performed on the pixels of the aligned overlapping region; and the fused panoramic depth image is output.

[0011] Preferably, the construction of the improved Mask-R-CNN pattern recognition model and the acquisition of multi-scale primary feature maps using the improved Mask-R-CNN pattern recognition model include: The improved Mask-R-CNN pattern recognition model includes a backbone network, a multi-scale feature fusion module, a region proposal network, a RoIAlign layer, a classification and regression head, and a mask head; The improved Mask-R-CNN pattern recognition model was trained using a large number of sample images containing labeled wear particle outlines and categories. A contrast stretching algorithm is applied to the panoramic depth image to obtain preprocessed image one; a noise reduction filtering algorithm is applied to preprocessed image one to obtain preprocessed image two; preprocessed image two is input into the trained improved Mask-R-CNN pattern recognition model; and multi-scale primary feature maps of preprocessed image two are extracted using the backbone network of the improved Mask-R-CNN pattern recognition model.

[0012] Preferably, the multi-scale feature fusion module utilizing the improved Mask-R-CNN pattern recognition model processes the multi-scale primary feature map and determines the wear particle mask based on the multi-scale primary feature map, including: The multi-scale feature fusion module first runs the top-down path of the feature pyramid network, and then runs the bottom-up enhancement path. Through convolution and upsampling operations, it fuses the low-level high-resolution features with the high-level high-semantic features step by step, and outputs a set of enhanced multi-scale feature maps. A region proposal network using an improved Mask-R-CNN pattern recognition model generates candidate regions on an enhanced multi-scale feature map; the region proposal network uses a set of anchor box sizes and aspect ratios specifically pre-defined to cover the range of wear particle sizes; The improved Mask-R-CNN pattern recognition model extracts fixed-size features from each candidate region using the RoIAlign layer. The improved Mask-R-CNN pattern recognition model's classification and regression heads filter wear particle candidate regions identified as wear particles based on the fixed-size features, and output precise bounding boxes. The improved Mask-R-CNN pattern recognition model's mask head outputs a pixel-level binary mask for each selected wear particle candidate region, denoted as the wear particle mask.

[0013] Preferably, the step of performing the Canny edge detection algorithm on each wear particle mask to extract sub-pixel precision contours, determining feature vectors based on the sub-pixel precision contours, and outputting the final category includes: Multiply the total number of pixels within the subpixel precision contour by the physical area of ​​a single pixel to obtain the area feature of each wear particle; fit the subpixel precision contour to an ellipse, and use the length of the major axis of the ellipse as the major axis dimension feature of each wear particle; multiply the area feature by 4π and then divide by the square of the perimeter of the subpixel precision contour to obtain the roundness feature of each wear particle. Based on the original image within the sub-pixel precision contour region, a gray-level co-occurrence matrix is ​​calculated; three statistics—contrast, entropy, and energy—are extracted from the gray-level co-occurrence matrix; the three statistics—contrast, entropy, and energy—are arranged in order into a multi-dimensional vector to obtain the texture features of each wear particle. The area features, major axis size features, roundness features, and texture features of each wear particle are combined into a feature vector; the feature vector is input into a post-processing classifier; the post-processing classifier outputs the category of each wear particle, including normal wear particles, cutting wear particles, severe sliding wear particles, spherical wear particles, and other wear particles; If the major axis dimension of the wear particle is smaller than a preset size threshold, the final category of the wear particle is determined as a normal wear particle; otherwise, the initial category output by the post-processing classifier is used as the final category of the wear particle.

[0014] Preferably, the step of obtaining the number of various characteristic wear particles according to the final category, judging the wear state by combining the total content of wear particles, the number of large wear particles, and the size of the largest wear particle, and automatically generating a detection report includes: Based on the sum of the area characteristics of all wear particles and the effective area of ​​the deposition region, the total content of wear particles is calculated using the formula: Total content of wear particles = (Sum of the area characteristics of all wear particles / Effective area of ​​the deposition region) * Calibration coefficient; the number of wear particles with a major axis dimension greater than or equal to the preset large wear particle size threshold is counted to obtain the number of large wear particles; the maximum value of the major axis dimension among all wear particles is selected to obtain the maximum wear particle size; the number of wear particles in the categories of cutting wear particles, severe sliding wear particles, and spherical wear particles is counted separately to obtain the number of wear particles of each type. The system evaluation database is constructed in the following way: historical detection data is continuously collected and stored, including component identification, wear particle data and equipment running time; when the historical detection data for a specific component accumulates to a preset amount, warning line thresholds and danger line thresholds are calculated based on the three-line value method for the total content of wear particles, the number of large wear particles, the maximum wear particle size and the number of characteristic wear particles, and the warning line thresholds and danger line thresholds are stored in the system database; Read the warning line threshold and danger line threshold corresponding to the currently tested component from the system database; compare the total wear particle content, number of large wear particles, maximum wear particle size, and number of characteristic wear particles calculated in this test with the read warning line threshold and danger line threshold respectively; If any indicator is greater than or equal to the danger threshold, the wear condition of the component is determined to be abnormal; if no indicator reaches the danger threshold, but any indicator is greater than or equal to the warning threshold, the wear condition of the component is determined to be cautious; if all indicators are less than the warning threshold, the wear condition of the component is determined to be normal. Automatically generate a PDF oil testing report; the oil testing report includes the component wear status conclusion, total wear particle content, number of large wear particles, maximum wear particle size, number of characteristic wear particles, wear particle distribution heat map, typical wear particle images, and historical test trend curves; output the oil testing report to the specified storage path and display it on the interface.

[0015] Preferably, an artificial intelligence-based condition monitoring and fault diagnosis system includes: The sample preparation and imaging control module is used to perform oil sample type determination and background purification, and to control the biaxial motion platform to complete the scanning path planning of the sedimentation area and the acquisition of multi-focal images. The panoramic image generation module is used to perform depth fusion and panoramic stitching processing on multi-focal plane image sequences to generate panoramic deep panoramic images. The wear particle intelligent recognition module is used to extract multi-scale primary feature maps through an improved Mask-R-CNN pattern recognition model, and output a wear particle mask after feature fusion and mask head processing. The status assessment and report generation module is used to calculate and classify the geometric and texture features of wear particles, and combine them with historical database thresholds to complete the wear status assessment and output of the detection report.

[0016] Beneficial Effects: This invention effectively eliminates background interference from dark oil samples and acquires panoramic and deep panoramic images through background purification processing and high-precision panoramic imaging, providing a high-quality image foundation for subsequent analysis. An improved Mask-R-CNN pattern recognition model is employed, using multi-scale feature fusion and mask generation to achieve accurate instance segmentation and target extraction of complex-shaped, adhered, and multi-scale wear particles. Based on this, geometric and texture features are extracted using sub-pixel contours, and a post-processing classifier is used to intelligently distinguish wear particle types, demonstrating the core role of pattern recognition methods in feature learning and classification decisions. Furthermore, by integrating multiple indicators such as total wear particle content and the number of characteristic wear particles, combined with dynamic threshold construction and a multi-dimensional evaluation mechanism, intelligent diagnosis from wear trend warning to abnormal state classification is achieved. Ultimately, a complete intelligent analysis system covering image acquisition, feature recognition, and state assessment is formed, significantly improving the accuracy, automation, and engineering practicality of wear state monitoring. Attached Figure Description

[0017] Figure 1 This is a flowchart of the artificial intelligence-based condition monitoring and fault diagnosis method of the present invention; Figure 2 This is a structural diagram of the artificial intelligence-based condition monitoring and fault diagnosis system of the present invention. Detailed Implementation

[0018] The present invention will be further described below with reference to specific embodiments. The illustrative embodiments and descriptions herein are used to explain the present invention, but are not intended to limit the present invention.

[0019] Example 1: A condition monitoring and fault diagnosis method based on artificial intelligence, such as Figure 1 As shown, it includes the following steps: S1-1: After acquiring oil sample type data and performing background purification processing, execute a high-precision panoramic imaging workflow, including: Connect the oil sample bottle to the inlet and the transparent auxiliary oil bottle to the auxiliary oil inlet. Obtain the oil sample type data. If the oil sample type data is not a dark oil sample, execute the high-precision panoramic imaging process. If the oil sample type data is a dark oil sample, keep the excitation coil energized, close the oil sample passage of the inlet through the three-way valve, open the auxiliary oil passage of the auxiliary oil inlet through the three-way valve, and start the peristaltic pump to pump the transparent auxiliary oil into the detection channel at a preset flow rate. The transparent auxiliary oil flows through the surface of the deposition glass slide in the detection channel. When the volume of the pumped transparent auxiliary oil reaches a preset threshold, stop the peristaltic pump, complete the background purification, and execute the high-precision panoramic imaging process.

[0020] It should be noted that in the analysis of wear particles in oil, dark-colored oil samples, due to their deep color and low transmittance, can severely interfere with image quality, leading to a decrease in the accuracy of subsequent feature extraction and recognition. Traditional ferrography often suffers from blurred wear particle outlines and loss of texture features due to background interference, affecting pattern recognition performance. This step, through background purification, aims to eliminate the masking effect of dark-colored oil samples on imaging and improve the image signal-to-noise ratio. The transmittance of the oil sample at a typical detection wavelength (e.g., 550 nm) is measured using an optical sensor. If the value is lower than a preset dark-color determination threshold (e.g., 30%), it is identified as a dark-colored oil sample, and the background purification process is then executed; otherwise, subsequent imaging proceeds directly. If the sample is determined to be dark oil, the oil path is automatically switched to transparent auxiliary oil while maintaining the magnetic field to adsorb wear particles. A peristaltic pump injects this auxiliary oil into the detection channel at a set flow rate (the preset flow rate is set based on the cross-sectional area of ​​the detection channel to ensure laminar flow across the deposition slide surface, preventing the adsorbed wear particles from being washed away) to replace the original dark oil sample. The preset threshold is greater than the total internal volume of the detection channel and connected flow paths to ensure complete replacement of the original dark oil sample. This process is completed in a closed flow path, avoiding displacement or loss of wear particles. Ultimately, an imaging environment with a uniform background and clear wear particle outlines is obtained, providing a high-quality image input foundation for high-precision feature learning and classification in subsequent pattern recognition.

[0021] S1-2: Obtain the number of images captured in the X and Y directions based on a two-axis motion platform, including: The two-axis motion platform is controlled to move to the mechanical origin position to establish a motion reference; the ring-shaped LED surface light source is turned on to provide illumination for the deposition slide; the length and width values ​​of the deposition area on the deposition slide are read; the length and width values ​​of the single-shot field of view of the camera are read; the length value of the deposition area is divided by the length value of the single-shot field of view of the camera, and the result is rounded up to obtain the number of shots in the X direction; the width value of the deposition area is divided by the width value of the single-shot field of view of the camera, and the result is rounded up to obtain the number of shots in the Y direction.

[0022] It should be noted that during the acquisition of wear particle images, the two-axis motion platform systematically scans the entire deposition area by controlling the camera's movement in the horizontal and vertical directions. First, the platform is reset to its mechanical origin position, which serves as the reference point for the entire coordinate system, ensuring a fixed starting position for each scan and thus guaranteeing spatial consistency in image acquisition. The deposition slide carries a sample of wear particles arranged in an orderly manner by a magnetic field; the length and width of the deposition area determine the required image range. A ring-shaped LED surface light source provides uniform illumination, effectively reducing shadow interference and creating favorable conditions for subsequent image processing. By dividing the dimensions (length and width) of the deposition area by the range (length and width) of the camera's single-shot field of view and rounding up, the number of shooting points required to cover the entire area can be accurately calculated, avoiding image omissions due to insufficient field of view. This quantitative calculation method overcomes the subjectivity and incompleteness of field-of-view selection in traditional manual observation, establishing a reliable and standardized digital spatial foundation for subsequent panoramic image stitching and pattern recognition algorithm feature extraction.

[0023] S2-1: Generate a grid coordinate sequence covering the deposition area based on the number of images taken in the X direction and the number of images taken in the Y direction, forming a line-by-line scanning path, including: Using the mechanical origin as the starting point for scanning, the step spacing in the X direction is the length of the field of view captured by the camera in a single shot, and the step spacing in the Y direction is the width of the field of view captured by the camera in a single shot. Based on the number of shots captured in the X and Y directions, a grid coordinate sequence covering the deposition area is generated using the grid point coordinate calculation formulas in the X and Y directions, forming a line-by-line scanning path. The formula for calculating the coordinates of grid points in the X direction is: ;in, For the first The X coordinates of the grid points The X-coordinate of the starting point of the scan, i.e., the X-coordinate of the machine origin position. This represents the length of the field of view captured by the camera in a single shot. Index the grid points, starting from 1 and counting the number of images taken in the X direction; The formula for calculating the coordinates of grid points in the Y direction is: ;in, For the first The Y coordinate of the grid point The Y-coordinate of the starting point of the scan, i.e., the Y-coordinate of the machine origin position. This represents the width of the field of view in a single shot taken by the camera. Index the grid points, starting from 1 and counting the number of images taken in the Y direction; Traverse all ( , Combine them to generate a grid coordinate sequence.

[0024] It should be noted that, to overcome the problems of incomplete coverage and positional deviation caused by manual field-of-view selection in traditional wear particle monitoring, this step achieves systematic image acquisition by establishing a standardized grid coordinate system. Using the machine origin position as the spatial reference, the dimensions (i.e., length and width values) of the camera's single-shot field of view are set as the motion step size to ensure seamless connection between fields of view. The X-axis coordinate follows... The linear growth law; the Y direction through The linear growth law. These two formulas together define a two-dimensional grid coordinate system covering the sedimentary region, where grid points are uniformly distributed along both the X and Y axes. In the formulas and The design aims to center each grid coordinate point within its corresponding single-field-of-view imaging region, ensuring sufficient overlap between adjacent full-focal-plane images and laying the geometric foundation for subsequent image stitching. By traversing all index combinations to generate a coordinate sequence, for example, when the number of points in the X direction is 3 and the number in the Y direction is 2, six core points are formed: (X1,Y1), (X2,Y1), (X3,Y1), (X1,Y2), (X2,Y2), and (X3,Y2). The scanning path constructed by these coordinate points in row-priority order allows the camera to move serpentinely along the deposition area. This avoids image omissions caused by traditional random sampling and ensures the accuracy of subsequent image stitching through geometric constraints, providing a complete spatial data foundation for pattern recognition.

[0025] S2-2: Synthesize a full-focal-plane image based on the scan path, and output a fused panoramic depth-of-view image based on the synthesized full-focal-plane image, including: For each grid coordinate point in the scanning path, the following steps are performed: control the two-axis motion platform to move to the corresponding position of the grid coordinate point, control the two-axis motion platform to move step by step in the Z-axis direction according to the preset step size, trigger the camera to acquire images at each Z-axis position, construct a multi-focal image sequence according to the Z-axis position order, and select the pixel value with the highest contrast from all multi-focal images for each pixel position in the multi-focal image sequence to synthesize a full-focal image. The SIFT feature point matching algorithm is used to calculate the overlapping region and transformation matrix between adjacent full-focal plane images; all full-focal plane images are projected and aligned according to the transformation matrix; a weighted average fusion algorithm is performed on the pixels of the aligned overlapping region; and the fused panoramic depth image is output.

[0026] It should be noted that, to address the core defect in traditional wear particle imaging where some particles are out of focus due to a fixed focal length, this invention employs precise Z-axis multi-plane sampling and fusion technology to acquire clear image data across the entire field. The preset step size must ensure coverage of the possible height distribution of wear particles, typically set to 1 / 2 to 1 / 3 of the theoretical depth of field value based on optical depth-of-field theory, to achieve sufficient sampling across the entire depth-of-field range.

[0027] At each grid coordinate point, the two-axis motion platform is moved to the corresponding position to fix the horizontal field of view. Subsequently, the platform moves incrementally along the Z-axis at the preset step size to systematically change the imaging focal plane, capturing clear images of wear particles at different heights. Camera acquisition is triggered at each discrete Z-axis position, resulting in a series of images with the same field of view but different focal points—a multi-focal-plane image sequence. This sequence carries the depth information of all objects in the scene.

[0028] By combining the highest contrast pixel values ​​from all images at each pixel location in a multi-focal plane sequence, the image contrast peaks at the actual object plane location of that pixel. This method automatically selects the clearest image information to generate a full-focal plane image of that local area. Full-focal plane images achieve an extended depth of field that cannot be achieved in a single shot.

[0029] To stitch together local full-focal-plane images into a complete panoramic image, it is necessary to handle image overlap and displacement. The SIFT (Scale Invariant Feature Transform) algorithm, due to its stability against rotation, scaling, and brightness changes, is used to automatically detect and match feature points between adjacent images. Successfully matched feature point pairs allow for robust calculation of the transformation matrix describing the geometric relationships between images. Projecting and aligning all images based on this transformation matrix eliminates geometric misalignments caused by camera movement and optical distortion, a prerequisite for seamless stitching. Subsequently, a weighted average fusion algorithm is performed on the overlapping areas of the aligned images. This algorithm smoothly transitions the boundaries between different images, eliminating brightness differences and stitching seams. The final output, a fused panoramic depth image, is a digital spectral map that completely covers the deposition area and clearly shows all wear particle details. Its purpose is to provide a high-fidelity, lossless, standardized input image for subsequent intelligent feature extraction and pattern recognition, directly overcoming the shortcomings of incomplete information in traditional single-focal-plane imaging.

[0030] S3-1: Construct an improved Mask-R-CNN pattern recognition model and use it to obtain multi-scale primary feature maps, including: The improved Mask-R-CNN pattern recognition model includes a backbone network, a multi-scale feature fusion module, a region proposal network, a RoIAlign layer, a classification and regression head, and a mask head; The improved Mask-R-CNN pattern recognition model was trained using a large number of sample images containing labeled wear particle outlines and categories. A contrast stretching algorithm is applied to the panoramic depth image to obtain preprocessed image one; a noise reduction filtering algorithm is applied to preprocessed image one to obtain preprocessed image two; preprocessed image two is input into the trained improved Mask-R-CNN pattern recognition model; and multi-scale primary feature maps of preprocessed image two are extracted using the backbone network of the improved Mask-R-CNN pattern recognition model.

[0031] It should be noted that this step employs an improved Mask-R-CNN pattern recognition model, aiming to overcome the shortcomings of traditional methods in recognizing complex wear particle morphologies and multi-scale targets. The improved Mask-R-CNN pattern recognition model comprises several core components working collaboratively. The backbone network, acting as a feature extractor, constructs multi-level features with different spatial resolutions and semantic levels from the input image through a series of convolution and downsampling operations. The multi-scale feature fusion module is a key improvement, based on a feature pyramid network architecture and innovatively embedding a channel attention mechanism. (The channel attention mechanism refers to the multi-scale feature fusion module connecting a channel attention submodule after each level of fusion in the top-down path of the feature pyramid network. This submodule first performs global average pooling on the input features, then generates channel weight vectors through two fully connected layers and activation functions, and finally multiplies the weight vectors with the original input features channel by channel to enhance the channel response for identifying key features of wear particles.) This mechanism can dynamically evaluate and weight the importance of different feature channels, thereby adaptively strengthening the feature responses that are effective in distinguishing small wear particles and adhesion areas when fusing high-level semantic features and low-level detail features from top to bottom, significantly improving the model's multi-scale perception capability.

[0032] The region proposal network slides pre-defined anchor boxes onto the fused multi-scale feature map to quickly generate candidate regions containing wear particles. The RoIAlign layer is responsible for accurately extracting fixed-dimensional feature representations from each candidate region of varying shapes in the feature map, solving the spatial misalignment problem between the feature map and the original image. Based on these aligned features, the classification and regression heads output the wear particle category to which each candidate region belongs and its precise bounding box coordinates in parallel. The masking head generates a fine-grained segmentation mask for each confirmed wear particle in a pixel-level prediction manner, achieving instance-level separation.

[0033] Before inference in the improved Mask-R-CNN pattern recognition model, the image needs to be preprocessed to optimize the input. First, contrast stretching is applied to the panoramic image to expand its dynamic range through a non-linear mapping based on the grayscale histogram, enhancing the distinction between the target and the background. This preprocessed image (Image 1) primarily addresses the issue of flat contrast in the original image. Subsequently, Gaussian filtering is applied to smooth Image 1, and the convolution operation of the Gaussian kernel effectively suppresses noise introduced by the image sensor, generating a cleaner preprocessed image (Image 2), providing a high-quality analytical foundation for the model. Finally, the backbone network extracts multi-scale primary feature maps from Image 2, forming a hierarchical feature space from local texture to global structure. This provides a unified and robust core feature representation for subsequent pattern recognition-based tasks such as wear particle target detection, instance segmentation, and classification.

[0034] S3-2: Processing multi-scale primary feature maps using the multi-scale feature fusion module of the improved Mask-R-CNN pattern recognition model, and determining wear particle masks based on the multi-scale primary feature maps, including: The multi-scale feature fusion module first runs the top-down path of the feature pyramid network, and then runs the bottom-up enhancement path. Through convolution and upsampling operations, it fuses the low-level high-resolution features with the high-level high-semantic features step by step, and outputs a set of enhanced multi-scale feature maps. A region proposal network using an improved Mask-R-CNN pattern recognition model generates candidate regions on an enhanced multi-scale feature map; the region proposal network uses a set of anchor box sizes and aspect ratios specifically pre-defined to cover the range of wear particle sizes; The improved Mask-R-CNN pattern recognition model extracts fixed-size features from each candidate region using the RoIAlign layer. The improved Mask-R-CNN pattern recognition model's classification and regression heads filter wear particle candidate regions identified as wear particles based on the fixed-size features, and output precise bounding boxes. The improved Mask-R-CNN pattern recognition model's mask head outputs a pixel-level binary mask for each selected wear particle candidate region, denoted as the wear particle mask.

[0035] It should be noted that, to address the problems of incomplete feature extraction and target omissions caused by the large differences in wear particle size, varied morphology, and mutual adhesion among wear particles in traditional wear particle identification methods, this step utilizes the multi-scale feature fusion module in the improved Mask-R-CNN model to enhance the primary features extracted by the backbone network. The multi-scale feature fusion module first runs the top-down path of the feature pyramid network, upsampling deep, high-semantic features and fusing them step-by-step with shallow, high-resolution features. Its function is to pass abstract category information to the detail layer, thereby improving the semantic perception ability for small-sized or blurred wear particles. Subsequently, a bottom-up enhancement path is run, using convolution operations to pass the fused features upwards again. This further optimizes the feature representation, enhancing the model's coherence and discriminative power in representing multi-scale targets. This cross-level fusion achieved through convolution and upsampling operations ultimately outputs a set of enhanced multi-scale feature maps. These multi-scale feature maps simultaneously contain rich spatial details and high-level semantic information, laying the foundation for subsequent accurate localization and recognition.

[0036] Next, the region proposal network slides a set of pre-defined anchor boxes across the enhanced feature maps to generate candidate regions. The size and aspect ratio of these anchor boxes are preset based on typical size ranges derived from historical wear particle data (e.g., covering 3μm to 500μm), serving as initial hypotheses about wear particle locations and efficiently nominating all regions in the image containing wear particles. Then, the RoIAlign layer is used to extract fixed-dimensional features from each irregularly shaped candidate region. This step addresses the spatial misalignment problem caused by direct pooling, ensuring the accuracy of subsequent classification. The classification and regression heads, based on these aligned fixed-size features, discriminate between candidate regions, selecting those confirmed as true wear particles and outputting their precise bounding box coordinates, which accurately pinpoint the location and extent of the wear particles in the image. Finally, the masking head generates a pixel-level binary mask for each confirmed wear particle candidate region. The wear particle mask accurately delineates the outline shape of each wear particle, achieving instance-level pixel-level segmentation. This completely separates adhered or overlapping wear particles in space, providing the most direct and accurate object region for subsequent precise extraction and classification of geometric and texture features based on pattern recognition.

[0037] S4-1: Perform the Canny edge detection algorithm on each wear particle mask to extract sub-pixel precision contours, determine feature vectors based on the sub-pixel precision contours, and output the final category, including: Multiply the total number of pixels within the subpixel precision contour by the physical area of ​​a single pixel to obtain the area feature of each wear particle; fit the subpixel precision contour to an ellipse, and use the length of the major axis of the ellipse as the major axis dimension feature of each wear particle; multiply the area feature by 4π and then divide by the square of the perimeter of the subpixel precision contour to obtain the roundness feature of each wear particle. Based on the original image within the sub-pixel precision contour region, a gray-level co-occurrence matrix is ​​calculated; three statistics—contrast, entropy, and energy—are extracted from the gray-level co-occurrence matrix; the three statistics—contrast, entropy, and energy—are arranged in order into a multi-dimensional vector to obtain the texture features of each wear particle. The area features, major axis size features, roundness features, and texture features of each wear particle are combined into a feature vector; the feature vector is input into a post-processing classifier; the post-processing classifier outputs the category of each wear particle, including normal wear particles, cutting wear particles, severe sliding wear particles, spherical wear particles, and other wear particles; If the major axis dimension of the wear particle is smaller than a preset size threshold, the final category of the wear particle is determined as a normal wear particle; otherwise, the initial category output by the post-processing classifier is used as the final category of the wear particle.

[0038] It should be noted that, to overcome the limitations of traditional wear particle identification which relies on manual measurement and subjective experience for classification, this step achieves objective quantitative analysis and intelligent classification of wear particles through systematic digital image processing and feature engineering. The Canny edge detection algorithm is executed on each wear particle mask generated by Mask-R-CNN. This algorithm first performs Gaussian filtering for noise reduction, then calculates the image gradient magnitude and direction, and uses non-maximum suppression and double threshold hysteresis to locate precise edge points, ultimately obtaining continuous, single-pixel-wide edge contours. Furthermore, sub-pixel interpolation is used to obtain contour coordinates with higher accuracy than a single pixel, i.e., sub-pixel precision contours, which achieve high-precision quantization of the geometric dimensions of wear particles.

[0039] Based on sub-pixel precision contours, the total number of pixels within the contour is multiplied by the actual physical area corresponding to a single pixel to obtain an area feature characterizing the size of the wear particles. The contour point set is fitted into an ellipse, and the length of the ellipse's major axis is used as the major axis dimension feature of the wear particles to describe their maximum extension scale. The roundness feature is calculated using the formula (area * 4π) / (perimeter squared). The closer this value is to 1, the closer the shape is to a circle, used to distinguish spherical wear particles. These three geometric features characterize the morphology of the wear particles from different dimensions.

[0040] To supplement surface information, a gray-level co-occurrence matrix (GLCM) is calculated within the original image region defined by the sub-pixel contours. The GLCM statistically analyzes the joint distribution of gray values ​​among pixel pairs with certain positional relationships in the image, forming the basis for texture quantization. Three statistical measures—contrast, entropy, and energy—are extracted from the GLCM: contrast reflects the drasticness of local image changes, entropy measures the randomness and complexity of the texture, and energy characterizes the uniformity of the image's gray-level distribution. These three scalars are arranged in a fixed order to form a multi-dimensional vector, the texture feature vector, which comprehensively describes the microstructure of the wear particle surface.

[0041] Geometric and texture feature vectors are combined to form a comprehensive feature vector that fully describes a single wear particle. This comprehensive feature vector is then input into a pre-trained post-processing classifier (such as a support vector machine or random forest). Based on the learned decision boundary, the post-processing classifier outputs the initial category of the wear particle, such as cutting wear particles (slender strips), severe sliding wear particles (scratched surfaces), and spherical wear particles (nearly round). Each category corresponds to a typical wear mechanism.

[0042] The preset size threshold (e.g., 20 micrometers) is an empirical value determined based on extensive engineering practice and is used for rapid screening. If the long axis dimension of the wear particle is smaller than the preset size threshold, it is directly identified as a normal wear particle, because small particles are usually the product of normal wear. Otherwise, the initial category determined by the post-processing classifier is adopted as the final category. This final category is the key qualitative basis for judging the type and severity of equipment wear, marking the completion of the pattern recognition process for a single wear particle (from mask generation, feature extraction to classification decision), and together with subsequent quantitative statistics, it supports the comprehensive diagnosis of wear status.

[0043] S4-2: Based on the final category, obtain the number of various characteristic wear particles, combine the total wear particle content, the number of large wear particles, and the size of the largest wear particle to evaluate the wear state and automatically generate a detection report, including: Based on the sum of the area characteristics of all wear particles and the effective area of ​​the deposition region, the total content of wear particles is calculated using the formula: Total content of wear particles = (Sum of the area characteristics of all wear particles / Effective area of ​​the deposition region) * Calibration coefficient; the number of wear particles with a major axis dimension greater than or equal to the preset large wear particle size threshold is counted to obtain the number of large wear particles; the maximum value of the major axis dimension among all wear particles is selected to obtain the maximum wear particle size; the number of wear particles in the categories of cutting wear particles, severe sliding wear particles, and spherical wear particles is counted separately to obtain the number of wear particles of each type. The system evaluation database is constructed in the following way: historical detection data is continuously collected and stored, including component identification, wear particle data and equipment running time; when the historical detection data for a specific component accumulates to a preset amount, warning line thresholds and danger line thresholds are calculated based on the three-line value method for the total content of wear particles, the number of large wear particles, the maximum wear particle size and the number of characteristic wear particles, and the warning line thresholds and danger line thresholds are stored in the system database; Read the warning line threshold and danger line threshold corresponding to the currently tested component from the system database; compare the total wear particle content, number of large wear particles, maximum wear particle size, and number of characteristic wear particles calculated in this test with the read warning line threshold and danger line threshold respectively; If any indicator is greater than or equal to the danger threshold, the wear condition of the component is determined to be abnormal; if no indicator reaches the danger threshold, but any indicator is greater than or equal to the warning threshold, the wear condition of the component is determined to be cautious; if all indicators are less than the warning threshold, the wear condition of the component is determined to be normal. Automatically generate a PDF oil testing report; the oil testing report includes the component wear status conclusion, total wear particle content, number of large wear particles, maximum wear particle size, number of characteristic wear particles, wear particle distribution heat map, typical wear particle images, and historical test trend curves; output the oil testing report to the specified storage path and display it on the interface.

[0044] It should be noted that, to overcome the limitations of traditional wear condition assessments that rely on expert subjective experience, lack quantitative standards, and are difficult to dynamically track the aging process of individual equipment, this invention introduces an intelligent diagnosis and report generation process based on multi-dimensional quantitative indicators and data-driven dynamic thresholds. This process first aggregates the individual wear particle information identified in the previous steps into four core macroscopic statistical indicators: total wear particle content, number of large wear particles, maximum wear particle size, and the number of various characteristic wear particles (cutting, severe sliding, spherical). The total content of wear particles is calculated by summing the area characteristics of all wear particles to obtain the total area, then dividing it by the effective area of ​​the deposition region (i.e., the area of ​​the rectangular region calculated based on the length and width values ​​of the deposition region read in steps S1-2), and multiplying it by a calibration coefficient determined through standard sample experiments (the calibration coefficient is determined experimentally by measuring the correspondence between the deposition projection area of ​​standard oil sample particles of known concentration and their standard concentration, and is updated based on the calibration data accumulated during system operation to maintain the accuracy of concentration calculation; for example, "using a set of standard wear particle oil samples with a concentration gradient, imaging and area statistics are performed respectively, and a curve relating the projection area and concentration is established through linear fitting, the slope or coefficient of the fitting function is the calibration coefficient"). This index reflects the overall concentration of wear products in the oil. The number of large wear particles is obtained by counting the total number of particles whose long axis dimension exceeds a preset threshold (e.g., 20 micrometers set based on tribological theory and historical data), the appearance of which often indicates abnormal wear. The maximum wear particle size is directly extracted by finding the maximum value of the long axis dimension of all particles, used to evaluate extreme wear events. The statistical analysis of the number of wear particles in the three characteristic categories of cutting, severe sliding, and spherical wear can be directly related to specific failure mechanisms of cutting wear, adhesive wear, and fatigue spalling, and is crucial for the qualitative diagnosis of fault types.

[0045] To achieve dynamic adaptation of the evaluation threshold, a continuously growing historical database is constructed to archive the component identification, detailed wear particle data, and equipment uptime for each inspection. Component identification ensures accurate data classification. Wear particle data includes the area, major axis dimension, roundness, texture features, and category of each wear particle, as well as the total wear particle content, the number of large wear particles, the largest wear particle size, and the number of wear particles of various characteristics. Equipment uptime provides a timeline for analyzing wear trends. When the accumulated valid historical inspection data for a specific component reaches a preset quantity (the preset quantity is determined based on statistical process control theory and must ensure sufficient significance for subsequent statistical analysis; for example, it can be comprehensively set by combining the central limit theorem with the wear characteristics and maintenance cycle of specific equipment; for example, based on the central limit theorem, the sample size n ≥ 30 is generally considered; or based on the specific equipment maintenance cycle, it is set to no less than 5 valid inspection data), statistical analysis is then performed on each key indicator based on the trilinear value method. This method calculates the warning threshold (usually the historical mean plus two standard deviations) and the danger threshold (historical mean plus three standard deviations) for each indicator. Based on the normal distribution assumption, the warning line (mean + two standard deviations) and the danger line (mean + three standard deviations) correspond to confidence intervals of approximately 95.4% and 99.7%, respectively, thereby quantifying the individualized performance of the equipment during normal operation into "attention intervals" and "abnormal boundaries".

[0046] During actual diagnosis, the system automatically retrieves the dynamic threshold corresponding to the currently tested component and compares it in real time with the measured values ​​of the four key indicators obtained in this test. Based on the comparison results, a tiered judgment is performed: if any indicator exceeds the danger threshold, the wear state is judged as "abnormal," indicating that the wear has significantly deviated from its historical normal pattern; if no indicator touches the danger line, but any indicator exceeds the warning line, the state is judged as "attention," indicating that the wear is showing an aggravating trend and requires attention; if all indicators remain below the warning line, the state is judged as "normal." This multi-indicator fusion and tiered judgment mechanism effectively avoids the risk of misjudgment based on a single indicator, achieving progressive intelligent diagnosis from trend warning to explicit alarm.

[0047] Finally, the system automatically integrates diagnostic conclusions, all quantitative indicators, wear particle distribution heatmaps, typical particle images, and their historical trend curves to generate a structured PDF oil testing report, which is then output to a designated storage path and displayed on a human-machine interface. Thus, this invention transforms the traditional wear analysis process, which relies on manual interpretation and is highly subjective, into a full-process pattern recognition intelligent diagnostic system encompassing image acquisition, intelligent segmentation, pattern recognition classification, and dynamic evaluation. This achieves an objective, traceable, and standardized data-driven intelligent diagnostic closed loop, significantly improving the reliability, consistency, and efficiency of condition monitoring.

[0048] Example 2: Based on Example 1, an artificial intelligence-based condition monitoring and fault diagnosis system, such as... Figure 2 As shown, it includes: The sample preparation and imaging control module is used to perform oil sample type determination and background purification, and to control the biaxial motion platform to complete the scanning path planning of the sedimentation area and the acquisition of multi-focal images. The panoramic image generation module is used to perform depth fusion and panoramic stitching processing on multi-focal plane image sequences to generate panoramic deep panoramic images. The wear particle intelligent recognition module is used to extract multi-scale primary feature maps through an improved Mask-R-CNN pattern recognition model, and output a wear particle mask after feature fusion and mask head processing. The status assessment and report generation module is used to calculate and classify the geometric and texture features of wear particles, and combine them with historical database thresholds to complete the wear status assessment and output of the detection report.

[0049] The above description is only a preferred embodiment of the present invention and is not intended to limit the present invention. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the protection scope of the present invention.

Claims

1. A condition monitoring and fault diagnosis method based on artificial intelligence, characterized in that, Includes the following steps: S1: After acquiring oil sample type data and performing background purification processing, execute the high-precision panoramic imaging process; The number of images captured in the X and Y directions is obtained based on a two-axis motion platform; S2: Generate a grid coordinate sequence covering the deposition area based on the number of images taken in the X direction and the number of images taken in the Y direction, forming a line-by-line scanning path; synthesize a full-focal-area image based on the scanning path, and output a fused panoramic depth-of-view image based on the synthesized full-focal-area image; S3: Construct an improved Mask-R-CNN pattern recognition model and use the improved Mask-R-CNN pattern recognition model to obtain multi-scale primary feature maps; use the multi-scale feature fusion module of the improved Mask-R-CNN pattern recognition model to process the multi-scale primary feature maps, and determine the wear particle mask based on the multi-scale primary feature maps. S4: Execute the Canny edge detection algorithm on each wear particle mask to extract sub-pixel precision contours, determine feature vectors based on the sub-pixel precision contours and output the final category; obtain the number of wear particles of each type according to the final category, and judge the wear status by combining the total content of wear particles, the number of large wear particles and the size of the largest wear particle, and automatically generate a detection report.

2. The artificial intelligence-based condition monitoring and fault diagnosis method according to claim 1, characterized in that, The process of acquiring oil sample type data, performing background purification processing, and then executing a high-precision panoramic imaging workflow includes: Connect the oil sample bottle to the inlet and the transparent auxiliary oil bottle to the auxiliary oil inlet. Obtain the oil sample type data. If the oil sample type data is not a dark oil sample, execute the high-precision panoramic imaging process. If the oil sample type data is a dark oil sample, keep the excitation coil energized, close the oil sample passage of the inlet through the three-way valve, open the auxiliary oil passage of the auxiliary oil inlet through the three-way valve, and start the peristaltic pump to pump the transparent auxiliary oil into the detection channel at a preset flow rate. The transparent auxiliary oil flows through the surface of the deposition glass slide in the detection channel. When the volume of the pumped transparent auxiliary oil reaches a preset threshold, stop the peristaltic pump, complete the background purification, and execute the high-precision panoramic imaging process.

3. The artificial intelligence-based condition monitoring and fault diagnosis method according to claim 1, characterized in that, The acquisition of the number of images in the X and Y directions based on the two-axis motion platform includes: The two-axis motion platform is controlled to move to the mechanical origin position to establish a motion reference; the ring-shaped LED surface light source is turned on to provide illumination for the deposition slide; the length and width values ​​of the deposition area on the deposition slide are read; the length and width values ​​of the single-shot field of view of the camera are read; the length value of the deposition area is divided by the length value of the single-shot field of view of the camera, and the result is rounded up to obtain the number of shots in the X direction; the width value of the deposition area is divided by the width value of the single-shot field of view of the camera, and the result is rounded up to obtain the number of shots in the Y direction.

4. The artificial intelligence-based condition monitoring and fault diagnosis method according to claim 1, characterized in that, The step of generating a grid coordinate sequence covering the deposition area based on the number of images taken in the X direction and the number of images taken in the Y direction, forming a line-by-line scanning path, includes: Using the mechanical origin as the starting point for scanning, the step spacing in the X direction is the length of the field of view captured by the camera in a single shot, and the step spacing in the Y direction is the width of the field of view captured by the camera in a single shot. Based on the number of shots captured in the X and Y directions, a grid coordinate sequence covering the deposition area is generated using the grid point coordinate calculation formulas in the X and Y directions, forming a line-by-line scanning path. The formula for calculating the coordinates of grid points in the X direction is: ;in, For the first The X coordinates of the grid points The X-coordinate of the starting point of the scan, i.e., the X-coordinate of the machine origin position. This represents the length of the field of view captured by the camera in a single shot. Index the grid points, starting from 1 and counting the number of images taken in the X direction; The formula for calculating the coordinates of grid points in the Y direction is: ;in, For the first The Y coordinate of the grid point The Y-coordinate of the starting point of the scan, i.e., the Y-coordinate of the machine origin position. This represents the width of the field of view in a single shot taken by the camera. Index the grid points, starting from 1 and counting the number of images taken in the Y direction; Traverse all ( , Combined, a grid coordinate sequence is generated.

5. A condition monitoring and fault diagnosis method based on artificial intelligence according to claim 1, characterized in that, The process of synthesizing a full-focal-plane image based on the scanning path and outputting a fused panoramic depth-of-view image based on the synthesized full-focal-plane image includes: For each grid coordinate point in the scanning path, the following steps are performed: control the two-axis motion platform to move to the corresponding position of the grid coordinate point, control the two-axis motion platform to move step by step in the Z-axis direction according to the preset step size, trigger the camera to acquire images at each Z-axis position, construct a multi-focal image sequence according to the Z-axis position order, and select the pixel value with the highest contrast from all multi-focal images for each pixel position in the multi-focal image sequence to synthesize a full-focal image. The SIFT feature point matching algorithm is used to calculate the overlapping region and transformation matrix between adjacent full-focal plane images; all full-focal plane images are projected and aligned according to the transformation matrix; a weighted average fusion algorithm is performed on the pixels of the aligned overlapping region; and the fused panoramic depth image is output.

6. The artificial intelligence-based condition monitoring and fault diagnosis method according to claim 1, characterized in that, The construction of the improved Mask-R-CNN pattern recognition model, and the use of the improved Mask-R-CNN pattern recognition model to obtain multi-scale primary feature maps, includes: The improved Mask-R-CNN pattern recognition model includes a backbone network, a multi-scale feature fusion module, a region proposal network, a RoIAlign layer, a classification and regression head, and a mask head; The improved Mask-R-CNN pattern recognition model was trained using a large number of sample images containing labeled wear particle outlines and categories. A contrast stretching algorithm is applied to the panoramic depth image to obtain preprocessed image one; a noise reduction filtering algorithm is applied to preprocessed image one to obtain preprocessed image two; preprocessed image two is input into the trained improved Mask-R-CNN pattern recognition model; and multi-scale primary feature maps of preprocessed image two are extracted using the backbone network of the improved Mask-R-CNN pattern recognition model.

7. The artificial intelligence-based condition monitoring and fault diagnosis method according to claim 1, characterized in that, The multi-scale feature fusion module utilizing the improved Mask-R-CNN pattern recognition model processes multi-scale primary feature maps and determines wear particle masks based on the multi-scale primary feature maps, including: The multi-scale feature fusion module first runs the top-down path of the feature pyramid network, and then runs the bottom-up enhancement path. Through convolution and upsampling operations, it fuses the low-level high-resolution features with the high-level high-semantic features step by step, and outputs a set of enhanced multi-scale feature maps. A region proposal network using an improved Mask-R-CNN pattern recognition model generates candidate regions on an enhanced multi-scale feature map; the region proposal network uses a set of anchor box sizes and aspect ratios specifically pre-defined to cover the range of wear particle sizes; The improved Mask-R-CNN pattern recognition model extracts fixed-size features from each candidate region using the RoIAlign layer. The improved Mask-R-CNN pattern recognition model's classification and regression heads filter wear particle candidate regions identified as wear particles based on the fixed-size features, and output precise bounding boxes. The improved Mask-R-CNN pattern recognition model's mask head outputs a pixel-level binary mask for each selected wear particle candidate region, denoted as the wear particle mask.

8. A condition monitoring and fault diagnosis method based on artificial intelligence according to claim 1, characterized in that, The process of performing the Canny edge detection algorithm on each wear particle mask to extract sub-pixel precision contours, determining feature vectors based on the sub-pixel precision contours, and outputting the final category includes: Multiply the total number of pixels within the subpixel precision contour by the physical area of ​​a single pixel to obtain the area feature of each wear particle; fit the subpixel precision contour to an ellipse, and use the length of the major axis of the ellipse as the major axis dimension feature of each wear particle; multiply the area feature by 4π and then divide by the square of the perimeter of the subpixel precision contour to obtain the roundness feature of each wear particle. Based on the original image within the sub-pixel precision contour region, a gray-level co-occurrence matrix is ​​calculated; three statistics—contrast, entropy, and energy—are extracted from the gray-level co-occurrence matrix; the three statistics—contrast, entropy, and energy—are arranged in order into a multi-dimensional vector to obtain the texture features of each wear particle. The area features, major axis size features, roundness features, and texture features of each wear particle are combined into a feature vector; the feature vector is input into a post-processing classifier; the post-processing classifier outputs the category of each wear particle, including normal wear particles, cutting wear particles, severe sliding wear particles, spherical wear particles, and other wear particles; If the major axis dimension of the wear particle is smaller than a preset size threshold, the final category of the wear particle is determined as a normal wear particle; otherwise, the initial category output by the post-processing classifier is used as the final category of the wear particle.

9. A condition monitoring and fault diagnosis method based on artificial intelligence according to claim 1, characterized in that, The process involves obtaining the number of various characteristic wear particles based on the final category, combining the total wear particle content, the number of large wear particles, and the size of the largest wear particle to assess the wear state and automatically generate a detection report, including: Based on the sum of the area characteristics of all wear particles and the effective area of ​​the deposition region, the total content of wear particles is calculated using the formula: Total content of wear particles = (Sum of the area characteristics of all wear particles / Effective area of ​​the deposition region) * Calibration coefficient; the number of wear particles with a major axis dimension greater than or equal to the preset large wear particle size threshold is counted to obtain the number of large wear particles; the maximum value of the major axis dimension among all wear particles is selected to obtain the maximum wear particle size; the number of wear particles in the categories of cutting wear particles, severe sliding wear particles, and spherical wear particles is counted separately to obtain the number of wear particles of each type. The system evaluation database is constructed in the following way: historical detection data is continuously collected and stored, including component identification, wear particle data and equipment running time; when the historical detection data for a specific component accumulates to a preset amount, warning line thresholds and danger line thresholds are calculated based on the three-line value method for the total content of wear particles, the number of large wear particles, the maximum wear particle size and the number of characteristic wear particles, and the warning line thresholds and danger line thresholds are stored in the system database; Read the warning line threshold and danger line threshold corresponding to the currently tested component from the system database; compare the total wear particle content, number of large wear particles, maximum wear particle size, and number of characteristic wear particles calculated in this test with the read warning line threshold and danger line threshold respectively; If any indicator is greater than or equal to the danger threshold, the wear condition of the component is determined to be abnormal; if no indicator reaches the danger threshold, but any indicator is greater than or equal to the warning threshold, the wear condition of the component is determined to be cautious; if all indicators are less than the warning threshold, the wear condition of the component is determined to be normal. Automatically generate a PDF oil testing report; the oil testing report includes the component wear status conclusion, total wear particle content, number of large wear particles, maximum wear particle size, number of characteristic wear particles, wear particle distribution heat map, typical wear particle images, and historical test trend curves; output the oil testing report to the specified storage path and display it on the interface.

10. An artificial intelligence-based condition monitoring and fault diagnosis system, used to implement the artificial intelligence-based condition monitoring and fault diagnosis method according to any one of claims 1-9, characterized in that, include: The sample preparation and imaging control module is used to perform oil sample type determination and background purification, and to control the biaxial motion platform to complete the scanning path planning of the sedimentation area and the acquisition of multi-focal images. The panoramic image generation module is used to perform depth fusion and panoramic stitching processing on multi-focal plane image sequences to generate panoramic deep panoramic images. The wear particle intelligent recognition module is used to extract multi-scale primary feature maps through an improved Mask-R-CNN pattern recognition model, and output a wear particle mask after feature fusion and mask head processing. The status assessment and report generation module is used to calculate and classify the geometric and texture features of wear particles, and combine them with historical database thresholds to complete the wear status assessment and output of the detection report.

Images