A deep learning-based oil particle dynamic monitoring method and system

Abstract

The application discloses a kind of based on deep learning's oil liquid abrasive particle dynamic monitoring method and system, belong to oil liquid abrasive particle dynamic monitoring technical field.S1.Through multiple sensors in the equipment lubrication system Real-time acquisition oil sample set;S2.Extract abrasive particle image in oil sample set;S3.Construct double-layer generative adversarial network model;S4.The trained double-layer generative adversarial network model is deployed in monitoring system, real-time acquisition and input abrasive particle image, and generate wear characteristic data;S5.Based on the wear characteristic data of identification, calculate the wear state index of equipment, judge whether equipment is in abnormal state;S6.If wear state index is detected to exceed set threshold, system automatically triggers early warning mechanism, and generates early warning information and informs relevant personnel to take maintenance measures;S7.Monitoring result is stored in database.The present application realizes accurate analysis and classification to oil liquid abrasive particle data, greatly improves the real-time monitoring and early warning capability of system.

Description

Technical Field

[0001] This invention belongs to the field of dynamic monitoring technology of oil abrasive particles, and particularly relates to a method and system for dynamic monitoring of oil abrasive particles based on deep learning. Background Technology

[0002] In the existing technology, oil wear monitoring mainly relies on wear detection and analysis based on traditional equipment such as optical microscopes and laser particle size analyzers. Traditional methods require manual intervention, and offline analysis is performed after collecting oil samples to determine the lubrication status and wear of the equipment. While this method can monitor the wear of the equipment to a certain extent, it has several limitations.

[0003] Traditional oil wear monitoring methods mostly require manual operation, have low data processing efficiency, long monitoring cycles, and are difficult to achieve real-time equipment status monitoring. Wear problems cannot be detected in the early stages, which may lead to maintenance measures being taken only when serious failures occur, resulting in unnecessary downtime and economic losses.

[0004] In existing technologies, the detection of oil abrasive particles is mostly based on simple particle counting or size measurement methods. These technical means cannot fully capture the key information of the shape, composition, and type of abrasive particles. This key information is the key indicator of equipment wear characteristics. Traditional methods are difficult to comprehensively and accurately characterize the wear state of equipment. In addition, due to the complexity of the morphology and chemical composition of abrasive particles, traditional detection methods lack sufficient intelligent means to automatically analyze and classify complex data, which limits their application scope and detection accuracy.

[0005] Traditional oil wear monitoring systems cannot achieve automation and intelligence. Existing monitoring systems often rely on a single data acquisition and analysis method, lacking the processing capabilities based on big data and artificial intelligence. They are unable to cope with complex and ever-changing industrial environments. In practical applications, as equipment operating time increases and operating conditions change, the wear patterns of the equipment often change dynamically. Existing monitoring methods lack adaptability and responsiveness to dynamic changes.

[0006] In summary, the existing technologies have the following main drawbacks: First, traditional oil abrasive monitoring methods cannot achieve real-time monitoring and are difficult to provide early warning of equipment wear problems; second, existing technologies cannot fully capture the multi-dimensional characteristics of abrasive particles, resulting in limited detection accuracy; third, existing systems lack intelligent and automated processing capabilities, making them difficult to adapt to complex industrial application environments. Summary of the Invention

[0007] One objective of this invention is to propose a method and system for dynamic monitoring of oil wear particles based on deep learning. This invention enables accurate analysis and classification of oil wear particle data, significantly improving the system's real-time monitoring and early warning capabilities.

[0008] A deep learning-based method for dynamic monitoring of oil wear particles according to an embodiment of the present invention includes the following steps:

[0009] S1. Collect oil sample sets in real time through multiple sensors deployed in the equipment lubrication system, and transmit the oil sample sets to the data processing module;

[0010] S2. Preprocess the oil sample set transmitted to the data processing module and extract the abrasive images from the oil sample set;

[0011] S3. Construct a two-layer generative adversarial network model, including a first-layer generator, a second-layer generator, and a discriminator;

[0012] S4. Deploy the trained two-layer generative adversarial network model in the monitoring system, collect and input abrasive images in real time, and use the two-layer generative adversarial network model to automatically identify the type, size and number of abrasive particles and generate wear feature data.

[0013] S5. Based on the identified wear characteristic data, calculate the wear status index of the equipment and compare it with the set threshold to determine whether the equipment is in an abnormal state. If the wear status index is detected to exceed the set threshold, proceed to S6. If the wear status index is detected not to exceed the set threshold, proceed to S7.

[0014] S6. Automatically trigger the early warning mechanism and generate early warning information to notify relevant personnel to take maintenance measures;

[0015] S7. Store the monitoring results in the database and display the abrasive characteristic change trend and equipment operating status in chart form through the data visualization module.

[0016] Optionally, S1 specifically includes:

[0017] S11. Install sensor arrays at key nodes in the equipment lubrication system. The sensor array includes n sensors T1, T2, ..., Tn. n Each sensor collects real-time data on oil abrasive particles in the oil sample, including the oil volumetric flow rate Q. i (t), Number of abrasive particles Q i (t), abrasive grain size D i (t) and the chemical composition C of the abrasive particles i (t), where i is the sensor number and t is time;

[0018] Sl2. The oil abrasive data collected by the n sensors are transmitted to the data processing module through the data transmission module;

[0019] S13. Perform preliminary formatting processing on the collected oil abrasive data to form an oil sample set:

[0020] D={(Q i (t), N i (t), D i (t), C i (t))|i=1,2,...,n;t=t1,t2,...,t m};

[0021] Where D represents all oil abrasive data collected at different time points t and multiple sensor positions i, and m is the number of time points.

[0022] Optionally, S2 specifically includes:

[0023] S21. Denoise the wear particle data in the oil sample set D transmitted to the data processing module by using a bandpass filter to eliminate random noise and system noise during the sampling process, thus obtaining the denoised oil sample set D. denoise ;

[0024] S22. For the denoised oil sample set D denoise Standardization is performed by processing each time step t. j and each sensor T i abrasive grain size D i (t j ), number of abrasive particles N i (t j ) and abrasive chemical composition C i (t j The oil sample set D was obtained by normalization. norm ;

[0025] S23. Based on standardized oil sample set D norm Extract abrasive grain images, for each abrasive grain size acquired by sensor Ti. and number of abrasive grains Through the two-dimensional Laplacian operator Perform image gradient enhancement:

[0026]

[0027] in, The results are from Laplacian edge detection. Indicates time t j Sensor T i The collected standardized abrasive grain size, with x and y representing the spatial coordinates of the image;

[0028] S24. For each time point t jBased on the detected edge information, the abrasive image is analyzed. Perform energy function Optimization, followed by abrasive grain region segmentation:

[0029]

[0030] Among them, Ω i The image shows the region containing abrasive particles, where α and β are weighting parameters. The image gradient represents the intensity of the edges. This indicates chemical composition data. A weighting function related to chemical composition;

[0031] S25. Based on energy function The minimized result is used to label and classify each segmented region using a fast matching algorithm, generating the final abrasive grain image matrix:

[0032]

[0033] Among them, M i (t j ) represents the matrix representation of the abrasive image, and Label is the region labeling function, which represents the classification of the optimal segmentation region, and finally outputs the labeled and classified abrasive image.

[0034] Optionally, S3 specifically includes:

[0035] S31. Construct a visual transformer model Vi(t) based on the preprocessed abrasive particle image matrix. j The abrasive grain image matrix is ​​divided into several image blocks Pk, where k is the index of the image block, and the size of each image block is p×p. An embedding operation is performed on each image block to obtain the embedded feature vector. The global dependencies between image patches are extracted using a self-attention mechanism, and the output is a transformer model representation containing global features of the abrasive image.

[0036]

[0037] S32. Construct a two-layer generative adversarial network model, including a first-layer generator G1, a second-layer generator G2, and a discriminator D:

[0038] The first-layer generator G1 generates abrasive grain images M of different types and sizes by inputting noise vector z1~N(0,1). gen1 (t j ) = G1(z1), which performs preliminary feature generation on the generated abrasive grain images of different types and sizes through a convolutional neural network;

[0039] The second-layer generator G2 generates the abrasive image M output by the first-layer generator G1. gen1 (t j Based on this, abrasive grain image moments M with more detailed features are generated. gen2 (t j ) = G2(M gen1 (t j This makes the generated abrasive grain image matrix approximate the real abrasive grain image matrix;

[0040] The discriminator D inputs the generated abrasive grain image matrix M gen2 (t j ) and the actual acquired abrasive particle image matrix M i (t j The authenticity of the two is determined through adversarial training:

[0041]

[0042] in, Let the loss function of the discriminator be , For the desired operation, D(·) represents the classification probability output of the discriminator for the input image;

[0043] S33. Using a pre-labeled abrasive grain image dataset D label For the visual transformer model Vi(t) j The two-layer generative adversarial network model (GAN) and discriminator (D) are jointly trained to optimize the feature extraction capability and abrasive particle feature classification accuracy.

[0044]

[0045] in, Let λ be the classification loss function, and λ be the loss weight factor.

[0046] Optionally, S4 specifically includes:

[0047] S41. The trained two-layer generative adversarial network model is deployed in the central data processing module of the monitoring system. The monitoring system acquires the wear particle image matrix in the oil sample in real time through the sensor array and inputs the wear particle image matrix into the two-layer generative adversarial network model.

[0048] S42. The first layer generator generates the abrasive grain image matrix M based on the acquired abrasive grain image matrix. i (t j Generate preliminary abrasive grain image matrix features F gen1 (t j ) = G1(M i (t j The generated abrasive grain image matrix M i (t jThis includes the initial abrasive grain size, abrasive grain shape, and number of abrasive grains;

[0049] S43. The second-layer generator generates detailed abrasive image matrix features F based on the preliminary abrasive image matrix features output by the first-layer generator. gen2 (t j )=G2(F gen1 (t j This makes the enhanced abrasive image matrix closer to the features of the real abrasive image matrix;

[0050] S44. Input the generated abrasive grain image matrix features and the actually acquired abrasive grain image matrix into the discriminator. The discriminator judges the authenticity and classification of the abrasive grain image matrix and generates a classification result C. i (t j );

[0051] S45. Based on classification result C i (t j ), generate wear characteristic data W i (t j This includes the type of abrasive grain, the size of the abrasive grain, the number of abrasive grains, and the chemical composition of the abrasive grains.

[0052] Optionally, S5 specifically includes:

[0053] S51. Based on the generated wear feature data W i (t j Calculate the wear condition index of the equipment. The wear condition index is defined by combining abrasive type, abrasive size, abrasive number, and abrasive chemical composition:

[0054]

[0055] Where α1, α2, and α3 are weight parameters. This refers to the standardized number of abrasive grains. The standardized abrasive grain size;

[0056] S52. Calculate the wear condition index. With the set threshold T w The comparison is performed, and the threshold is set based on the equipment's historical wear data and operating conditions to determine whether the equipment is in an abnormal state. The judgment criteria are as follows:

[0057]

[0058] Among them, Status(t) j ) indicates that the device is at time t j The state of T at that time w The threshold value is set.

[0059] A deep learning-based dynamic monitoring system for oil wear particles includes the following modules:

[0060] The sensor module is used to collect real-time data on abrasive particles in the lubrication system of the equipment, including the number of abrasive particles, their size, shape, and chemical composition.

[0061] The data processing module is configured to receive oil abrasive data collected by the sensor module and to preprocess, standardize, and denoise the oil abrasive data.

[0062] A two-layer generative adversarial network model, consisting of a first-layer generator, a second-layer generator, and a discriminator, generates and discriminates abrasive image features and classifies abrasive feature data.

[0063] The wear condition assessment module calculates the wear condition index of the equipment based on abrasive particle characteristic data and determines whether the equipment is in an abnormal state.

[0064] The early warning module automatically generates an abnormal warning message and notifies maintenance personnel when it detects abnormal wear and tear on the equipment.

[0065] The data storage and visualization module stores monitoring data and uses visualization tools to display the equipment's operating status and the changing trends of abrasive characteristics.

[0066] The beneficial effects of this invention are:

[0067] (1) The present invention introduces a visual transformer model in abrasive image processing. It uses a self-attention mechanism to realize global feature extraction of abrasive images. Unlike traditional local image analysis methods, the visual transformer divides the abrasive image into several image blocks and embeds each image block. Then, it extracts the global dependency of the abrasive image through a self-attention mechanism, which can capture the multidimensional features of abrasives more comprehensively and accurately, including the shape, size and surface details of abrasives.

[0068] (2) The present invention adopts a two-layer generative adversarial network model, including two generators and one discriminator. The first-layer generator generates a preliminary abrasive image, and the second-layer generator enhances the details of the abrasive image based on the first-layer generator, making it closer to the real abrasive image. Through adversarial training, the discriminator can effectively distinguish between real and generated abrasive images, improving the system's ability to identify abrasives in complex industrial environments. Compared with traditional rule-based recognition methods, the introduction of the two-layer generative adversarial network not only improves the detail fidelity of the abrasive image, but also optimizes the accuracy and reliability of abrasive feature classification.

[0069] (3) By deploying a trained two-layer generative adversarial network model in the monitoring system, the present invention realizes real-time monitoring of the lubrication status and wear of the equipment. The system can dynamically calculate the wear status index of the equipment based on the type, size, quantity and chemical composition characteristics of the abrasive particles, and compare it with the set threshold in real time. Once the threshold is exceeded, the system will automatically trigger an early warning mechanism to notify relevant maintenance personnel to take timely maintenance measures, thereby improving the real-time performance and automation level of the monitoring system and avoiding economic losses caused by equipment downtime due to malfunctions. Attached Figure Description

[0070] The accompanying drawings are provided to further illustrate the invention and form part of the specification. They are used in conjunction with embodiments of the invention to explain the invention and do not constitute a limitation thereof. In the drawings:

[0071] Figure 1 This is a flowchart of a deep learning-based method and system for dynamic monitoring of oil wear particles proposed in this invention. Detailed Implementation

[0072] The present invention will now be described in further detail with reference to the accompanying drawings. These drawings are simplified schematic diagrams, illustrating only the basic structure of the invention, and therefore only show the components relevant to the invention.

[0073] Example 1: Reference Figure 1 A deep learning-based method for dynamic monitoring of oil wear particles includes the following steps:

[0074] S1. Collect oil sample sets in real time through multiple sensors deployed in the equipment lubrication system, and transmit the oil sample sets to the data processing module;

[0075] S2. Preprocess the oil sample set transmitted to the data processing module and extract the abrasive images from the oil sample set;

[0076] S3. Construct a two-layer generative adversarial network model, including a first-layer generator, a second-layer generator, and a discriminator;

[0077] S4. Deploy the trained two-layer generative adversarial network model in the monitoring system, collect and input abrasive images in real time, and use the two-layer generative adversarial network model to automatically identify the type, size and number of abrasive particles and generate wear feature data.

[0078] S5. Based on the identified wear characteristic data, calculate the wear status index of the equipment and compare it with the set threshold to determine whether the equipment is in an abnormal state. If the wear status index is detected to exceed the set threshold, proceed to S6; if the wear status index is detected not to exceed the set threshold, proceed to S7.

[0079] S6. The system automatically triggers the early warning mechanism and generates early warning information to notify relevant personnel to take maintenance measures;

[0080] S7. Store the monitoring results in the database and display the abrasive characteristic change trend and equipment operating status in chart form through the data visualization module.

[0081] In this embodiment, S1 specifically includes:

[0082] S11. Install sensor arrays at key nodes in the equipment lubrication system. The sensor array includes n sensors T1, T2, ..., Tn. n Each sensor collects real-time data on oil abrasive particles in the oil sample, including the oil volumetric flow rate Q. i (t), Number of abrasive particles Q i (t), abrasive grain size D i (t) and the chemical composition C of the abrasive particles i (t), where i is the sensor number and t is time;

[0083] S12. Transmit the oil abrasive data collected by n sensors to the data processing module through the data transmission module;

[0084] S13. Perform preliminary formatting processing on the collected oil abrasive data to form an oil sample set:

[0085] D={(Q i (t), N i (t), D i (t), C i (t))|i=1,2,...,n;t=t1,t2,...,t m};

[0086] Where D represents all oil abrasive data collected at different time points t and multiple sensor positions i, and m is the number of time points.

[0087] In this embodiment, S2 specifically includes:

[0088] S21. Denoise the wear particle data in the oil sample set D transmitted to the data processing module by using a bandpass filter to eliminate random noise and system noise during the sampling process, thus obtaining the denoised oil sample set D. denoise ;

[0089] S22. For the denoised oil sample set D denoise Standardization is performed by processing each time step t. j and each sensor T i abrasive grain size D i (tj ), number of abrasive particles N i (t j ) and abrasive chemical composition C i (t j The oil sample set D was obtained by normalization. norm ;

[0090] S23. Based on standardized oil sample set D norm Extract abrasive grain images, for each abrasive grain size acquired by sensor Ti. and number of abrasive grains Through the two-dimensional Laplacian operator Perform image gradient enhancement:

[0091]

[0092] in, The results are from Laplacian edge detection. Indicates time t j Sensor T i The collected standardized abrasive grain size, with x and y representing the spatial coordinates of the image;

[0093] S24. For each time point t j Based on the detected edge information, the abrasive image is analyzed. Perform energy function Optimization, followed by abrasive grain region segmentation:

[0094]

[0095] Among them, Ω i The image shows the region containing abrasive particles, where α and β are weighting parameters. The image gradient represents the intensity of the edges. This indicates chemical composition data. A weighting function related to chemical composition;

[0096] S25. Based on energy function The minimized result is used to label and classify each segmented region using a fast matching algorithm, generating the final abrasive grain image matrix:

[0097]

[0098] Among them, M i (t j ) represents the matrix representation of the abrasive image, and Label is the region labeling function, which represents the classification of the optimal segmentation region, and finally outputs the labeled and classified abrasive image.

[0099] In this embodiment, S3 specifically includes:

[0100] S31. Construct a visual transformer model Vi(t) based on the preprocessed abrasive particle image matrix. j The abrasive grain image matrix is ​​divided into several image blocks Pk, where k is the index of the image block. Each image block has a size of p×p. An embedding operation is performed on each image block to obtain the embedded feature vector εk. The global dependencies between image blocks are extracted through a self-attention mechanism, and the transformer model representation containing the global features of the abrasive grain image is output.

[0101] Vi(t j ) = Attention(εk);

[0102] S32. Construct a two-layer generative adversarial network model, including a first-layer generator G1, a second-layer generator G2, and a discriminator D:

[0103] The first-layer generator G1 generates abrasive grain images M of different types and sizes by inputting noise vector z1~N(0,1). gen1 (t j ) = G1(z1), which performs preliminary feature generation on the generated abrasive grain images of different types and sizes through a convolutional neural network;

[0104] The second-layer generator G2 generates the abrasive image M output by the first-layer generator G1. gen1 (t j Based on this, abrasive grain image moments M with more detailed features are generated. gen2 (t j ) = G2(M gen1 (t j This makes the generated abrasive grain image matrix approximate the real abrasive grain image matrix;

[0105] The discriminator D inputs the generated abrasive grain image matrix M gen2 (t j ) and the actual acquired abrasive particle image matrix M i (t j The authenticity of the two is determined through adversarial training:

[0106]

[0107] in, Let the loss function of the discriminator be , For the desired operation, D(·) represents the classification probability output of the discriminator for the input image;

[0108] S33. Using a pre-labeled abrasive grain image dataset D label For the visual transformer model Vi(t) jThe two-layer generative adversarial network model (GAN) and discriminator (D) are jointly trained to optimize the feature extraction capability and abrasive particle feature classification accuracy.

[0109]

[0110] in, Let λ be the classification loss function, and λ be the loss weight factor.

[0111] In this embodiment, S4 specifically includes:

[0112] S41. The trained two-layer generative adversarial network model is deployed in the central data processing module of the monitoring system. The monitoring system acquires the wear particle image matrix in the oil sample in real time through the sensor array and inputs the wear particle image matrix into the two-layer generative adversarial network model.

[0113] S42. The first layer generator generates the abrasive grain image matrix M based on the acquired abrasive grain image matrix. i (t j Generate preliminary abrasive grain image matrix features F gen1 (t j ) = G1(M i (t j The generated abrasive grain image matrix M i (t j This includes the initial abrasive grain size, abrasive grain shape, and number of abrasive grains;

[0114] S43. The second-layer generator generates detailed abrasive image matrix features F based on the preliminary abrasive image matrix features output by the first-layer generator. gen2 (t j )=G2(F gen1 (t j This makes the enhanced abrasive image matrix closer to the features of the real abrasive image matrix;

[0115] S44. Input the generated abrasive grain image matrix features and the actually acquired abrasive grain image matrix into the discriminator. The discriminator judges the authenticity and classification of the abrasive grain image matrix and generates a classification result C. i (t j );

[0116] S45. Based on classification result C i (t j ), generate wear characteristic data W i (t j This includes the type of abrasive grain, the size of the abrasive grain, the number of abrasive grains, and the chemical composition of the abrasive grains.

[0117] In this embodiment, S5 specifically includes:

[0118] S51. Based on the generated wear feature data W i (t j Calculate the wear condition index of the equipment. The wear condition index is defined by combining abrasive type, abrasive size, abrasive number, and abrasive chemical composition:

[0119]

[0120] Where α1, α2, and α3 are weight parameters. This refers to the standardized number of abrasive grains. The standardized abrasive grain size;

[0121] S52. Calculate the wear condition index. With the set threshold T w The comparison is performed, and the threshold is set based on the equipment's historical wear data and operating conditions to determine whether the equipment is in an abnormal state. The judgment criteria are as follows:

[0122]

[0123] Among them, Status(t) j ) indicates that the device is at time t j The state of T at that time w The threshold value is set.

[0124] A deep learning-based dynamic monitoring system for oil wear particles includes the following modules:

[0125] The sensor module is used to collect real-time data on abrasive particles in the lubrication system of the equipment, including the number of abrasive particles, their size, shape, and chemical composition.

[0126] The data processing module is configured to receive oil abrasive data collected by the sensor module and to preprocess, standardize, and denoise the oil abrasive data.

[0127] A two-layer generative adversarial network model, consisting of a first-layer generator, a second-layer generator, and a discriminator, generates and discriminates abrasive image features and classifies abrasive feature data.

[0128] The wear condition assessment module calculates the wear condition index of the equipment based on abrasive particle characteristic data and determines whether the equipment is in an abnormal state.

[0129] The early warning module automatically generates an abnormal warning message and notifies maintenance personnel when it detects abnormal wear and tear on the equipment.

[0130] The data storage and visualization module stores monitoring data and uses visualization tools to display the equipment's operating status and the changing trends of abrasive characteristics.

[0131] Example 2: Example 2 illustrates the application of the present invention in wind power generation equipment. In a coastal wind farm, hundreds of wind turbines operate daily under harsh climatic conditions. Moisture, salt spray, and high wind speeds in the environment continuously affect the lubrication system of the equipment. Monitoring abrasive particles in the lubricating oil is crucial to ensuring the normal operation of the wind turbines. However, traditional abrasive particle monitoring methods often rely on manual sampling, which has a long detection cycle and makes it difficult to detect early wear problems in the equipment in a timely manner.

[0132] To address this issue, the wind farm decided to introduce this invention to achieve real-time monitoring of the wear status of the lubrication system. The system is deployed on each wind turbine in the wind farm and collects lubricating oil samples in real time through sensors to monitor the characteristics of abrasive particles in the oil. During operation, the system continuously analyzes multi-dimensional data on the number, size, and chemical composition of abrasive particles, thereby dynamically assessing the wear status of the equipment.

[0133] During a monitoring process in May 2023, the monitoring system detected an anomaly in the wind turbine generator set numbered "W001". The system collected data in real time and found that the number of abrasive particles in the lubricating oil began to increase rapidly, from 50 to 75 per milliliter. At the same time, the average size of the abrasive particles also increased from 5.2 micrometers to 6.2 micrometers. The system automatically determined that the abrasive particle data exceeded the preset threshold and generated the first abnormal wear warning report at 14:32 that day.

[0134] Monitoring data showed a significant increase in copper content in the abrasive particles' chemical composition. Analysis indicated that the bearings inside the equipment might be experiencing abnormal wear. Equipment monitoring personnel confirmed this anomaly through a system-generated report and received a warning signal at 14:35. Based on the real-time data provided by the system, the maintenance team decided to shut down the equipment for inspection at 17:00 that afternoon.

[0135] Inspection revealed that the bearings of equipment W001 had suffered severe wear, and the lubricating oil contained excessive levels of metal particles. Further analysis of the abrasive particle shape, size, and composition determined that the primary cause of the wear was long-term operation leading to bearing lubrication failure and increased metal-to-metal friction. The maintenance team immediately repaired the lubrication system and replaced the damaged bearing assembly, successfully preventing a larger-scale equipment failure.

[0136] During this period, another device, numbered "W002", also showed a similar wear warning. After the system detected that the number of abrasive particles in device W002 increased from 65 per milliliter to 85, and the abrasive particle size increased from 5.8 micrometers to 7.0 micrometers, it generated an anomaly report again at 10:12 on May 18. Through the report automatically generated by the system, the maintenance personnel quickly located the wear problem and decided to conduct further testing on the equipment.

[0137] The test results of equipment W002 showed that the iron content in the abrasive particles was significantly high. System analysis showed that the external material of the bearing may have generated a large number of wear particles due to friction. The monitoring system recorded the growth trend of iron particles in the lubricating oil of the equipment in real time and sent an emergency warning at 10:15. Based on the data, the maintenance team took timely shutdown maintenance and successfully repaired the damaged parts of the equipment, avoiding more serious failures.

[0138] To better verify the effectiveness of the system of the present invention, we selected 10 devices in a wind farm and conducted dynamic monitoring of oil abrasive particles for three months, and compared the results with traditional methods. Table 1 below shows the monitoring data of some devices:

[0139] Table 1 Comparison of data from the oil abrasive dynamic monitoring system

[0140]

[0141] As can be seen from the data in Table 1 above, the system of the present invention can respond quickly when abnormal wear is detected. In the monitoring of equipment W001, when the system detected a rapid increase in the number and size of abrasive particles, it generated an abnormal wear warning report at 14:32. The equipment inspection conducted at 17:00 confirmed the wear of the bearing. In contrast, the traditional manual inspection method requires sampling and analysis once a month, with a cycle of up to 30 days. This means that in the traditional method, the wear problem of equipment W001 may not be discovered until 30 days later, and the equipment will be subject to a greater risk of damage.

[0142] During the monitoring of equipment W002, the system detected that the number of abrasive particles increased from 65 per milliliter to 85, and generated an abnormality warning report at 10:12. Based on this data, the maintenance team quickly shut down the machine for maintenance and repaired the wear problem of the equipment at 13:00 that afternoon. Traditional methods, due to their long detection cycle, cannot detect early signs of equipment wear in a timely manner.

[0143] The system of this invention can not only monitor the changes in the number and size of abrasive particles in real time, but also further identify the root cause of equipment wear by analyzing the chemical composition of the abrasive particles. In the monitoring of W004 equipment, the system found that the metal particles in the lubricating oil exceeded the standard by analyzing the copper content in the abrasive particles, and generated an abnormality warning report at 11:00 to remind maintenance personnel to perform emergency maintenance operations. Traditional methods cannot provide real-time data due to long sampling cycles, which may delay the resolution of equipment wear problems.

[0144] Through practical application in wind power generation equipment, the oil wear particle dynamic monitoring system based on a two-layer generative adversarial network model of this invention has demonstrated its superior real-time monitoring capabilities. The system can generate timely anomaly warning reports by analyzing multi-dimensional data on the number, size, shape, and chemical composition of wear particles at an early stage of lubricating oil wear problems, guiding maintenance personnel to perform precise equipment maintenance and preventing major equipment failures. Compared with traditional detection methods, this system has higher response speed and detection accuracy, effectively reducing equipment maintenance costs and improving equipment operational reliability.

[0145] This invention introduces a visual transformer model into abrasive image processing. It utilizes a self-attention mechanism to achieve global feature extraction of abrasive images. Unlike traditional local image analysis methods, the visual transformer divides the abrasive image into several image blocks and embeds each image block. Then, it extracts the global dependencies of the abrasive image through a self-attention mechanism, which can capture the multidimensional features of abrasives more comprehensively and accurately, including the shape, size and surface details of the abrasives.

[0146] This invention employs a two-layer generative adversarial network model, comprising two generators and one discriminator. The first-layer generator generates an initial abrasive grain image, while the second-layer generator enhances the details of the abrasive grain image based on the first-layer generator, making it closer to a real abrasive grain image. Through adversarial training, the discriminator can effectively distinguish between real and generated abrasive grain images, improving the system's ability to identify abrasive grains in complex industrial environments. Compared to traditional rule-based recognition methods, the introduction of the two-layer generative adversarial network not only improves the detail fidelity of the abrasive grain image but also optimizes the accuracy and reliability of abrasive grain feature classification.

[0147] This invention deploys a trained two-layer generative adversarial network model into a monitoring system, enabling real-time monitoring of equipment lubrication status and wear. The system can dynamically calculate the wear status index of the equipment based on the type, size, quantity, and chemical composition characteristics of abrasive particles, and compare it with a set threshold in real time. Once the threshold is exceeded, the system will automatically trigger an early warning mechanism to notify relevant maintenance personnel to take timely maintenance measures, thereby improving the real-time performance and automation level of the monitoring system and avoiding economic losses caused by equipment downtime due to malfunctions.

[0148] The above description is only a preferred embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any equivalent substitutions or modifications made by those skilled in the art within the scope of the technology disclosed in the present invention, based on the technical solution and inventive concept of the present invention, should be covered within the scope of protection of the present invention.

Claims

1. A method for dynamic monitoring of oil wear particles based on deep learning, characterized in that, Includes the following steps: S1. Real-time oil sample data is collected by multiple sensors installed in the equipment lubrication system, and the oil sample data is transmitted to the data processing module to form an oil sample set; S2. Preprocess the oil sample set transmitted to the data processing module and extract the abrasive images from the oil sample set; S3. Construct a two-layer generative adversarial network model, including a first-layer generator, a second-layer generator, and a discriminator; S4. Deploy the trained two-layer generative adversarial network model in the monitoring system, collect and input abrasive images in real time, and use the two-layer generative adversarial network model to automatically identify the type, size and number of abrasive particles in the abrasive images and generate wear feature data. S5. Based on the identified wear characteristic data, calculate the wear status index of the equipment and compare it with the set threshold to determine whether the equipment is in an abnormal state. If the wear status index is detected to exceed the set threshold, proceed to S6. If the wear status index is detected not to exceed the set threshold, proceed to S7. S6. Automatically trigger the early warning mechanism and generate early warning information to notify relevant personnel to take maintenance measures; S7. Store the monitoring results in the database and display the abrasive particle characteristic change trend and equipment operating status in chart form through the data visualization module; S2 specifically includes: S21. The oil sample set transmitted to the data processing module. The wear particle data in the sample is denoised by using a bandpass filter to eliminate random noise and system noise during the sampling process, resulting in a denoised oil sample set. ; S22. For the denoised oil sample set Standardization is performed by processing each time step. and each sensor abrasive grain size Number of abrasive grains and abrasive chemical composition Normalization was performed to obtain a standardized oil sample set. ; S23. Based on standardized oil sample sets Extract abrasive images for each sensor. In time Standardized abrasive grain size collected and number of abrasive grains Through the two-dimensional Laplacian operator Perform image gradient enhancement: ； in, The results are Laplacian edge detection results, where x and y are the spatial coordinates of the image, respectively. S24. For each time point Based on the detected edge information, the abrasive image is analyzed. Perform energy function Optimization, followed by abrasive grain region segmentation: ； in, This represents the area of ​​abrasive particles in the image. and For weight parameters, The image gradient represents the intensity of the edges. This indicates chemical composition data. A weighting function related to chemical composition; S25. Based on energy function The minimized result is used to label and classify each segmented region using a fast matching algorithm, generating the final abrasive grain image matrix: ； in, This is a matrix representation of the abrasive grain image. This is the region labeling function, which classifies the optimal segmented region and ultimately outputs the labeled and classified abrasive grain images. S3 specifically includes: S31. Constructing a visual transformer model based on the preprocessed abrasive particle image matrix. The abrasive particle image matrix is ​​divided into several image blocks. Where k is the index of the image patch, and the size of each image patch is . Embedding is performed on each image patch to obtain the embedded feature vector. The global dependencies between image patches are extracted using a self-attention mechanism, and the output is a transformer model representation containing global features of the abrasive image. ； S32. Construct a two-layer generative adversarial network model, including a first-layer generator. Second layer generator And a discriminator D: First-level generator By input noise vector Generate images of abrasive particles of different types and sizes. Preliminary feature generation is performed on the generated abrasive grain images of different types and sizes using a convolutional neural network; Second-level generator First-level generator Output abrasive grain image Based on this, generate abrasive grain image moments with more detailed features. This makes the generated abrasive grain image matrix approximate the real abrasive grain image matrix; The discriminator D input generates abrasive grain image matrix. and the matrix of real-world abrasive particle images The authenticity of the two can be determined through adversarial training: ； in, Let the loss function be that of the discriminator. For the desired operation, This represents the classification probability output of the discriminator for the input image; S33. Using a pre-labeled abrasive grain image dataset For the visual transformer model Jointly train the two-layer generative adversarial network model with the discriminator D to optimize its feature extraction capability and abrasive particle feature classification accuracy. ； in, For classification loss function, This is the loss weighting factor.

2. The method for dynamic monitoring of oil wear particles based on deep learning according to claim 1, characterized in that, S1 specifically includes: S11. Install sensor arrays at key nodes of the equipment lubrication system. The sensor array includes n sensors. Each sensor collects real-time data on oil abrasive particles in the oil sample, including the oil volumetric flow rate. Number of abrasive grains Abrasive grain size and the chemical composition of abrasive particles Where i is the sensor number and t is the time; S12. The oil abrasive data collected by the n sensors are transmitted to the data processing module through the data transmission module; S13. Perform preliminary formatting processing on the collected oil abrasive data to form an oil sample set: ； in, This represents all oil abrasive data collected at different time points t and multiple sensor locations i, where m is the number of time points.

3. The method for dynamic monitoring of oil wear particles based on deep learning according to claim 1, characterized in that, S4 specifically includes: S41. The trained two-layer generative adversarial network model is deployed in the central data processing module of the monitoring system. The monitoring system acquires the wear particle image matrix in the oil sample in real time through the sensor array and inputs the wear particle image matrix into the two-layer generative adversarial network model. S42. The first layer generator generates the abrasive grain image matrix based on the acquired abrasive grain image matrix. Generate preliminary abrasive particle image matrix features The generated abrasive grain image matrix This includes the initial abrasive grain size, abrasive grain shape, and number of abrasive grains; S43. The second-layer generator generates detailed abrasive image matrix features based on the preliminary abrasive image matrix features output by the first-layer generator. This makes the enhanced abrasive image matrix closer to the features of the real abrasive image matrix; S44. Input the generated abrasive grain image matrix features and the actually acquired abrasive grain image matrix into the discriminator. The discriminator judges the authenticity and classification of the abrasive grain image matrix and generates a classification result. ; S45. Based on classification results Generate wear characteristic data This includes the type of abrasive grain, the size of the abrasive grain, the number of abrasive grains, and the chemical composition of the abrasive grains.

4. The method for dynamic monitoring of oil wear particles based on deep learning according to claim 1, characterized in that, S5 specifically includes: S51. Based on the generated wear feature data Calculate the wear condition index of the equipment The wear condition index is defined by combining the abrasive type, abrasive size, abrasive number, and abrasive chemical composition: ； in, , , For weight parameters, This refers to the standardized number of abrasive grains. The standardized abrasive grain size; S52. Calculate the wear condition index. With the set threshold The comparison is performed, and the threshold is set based on the equipment's historical wear data and operating conditions to determine whether the equipment is in an abnormal state. The judgment criteria are as follows: ； in, Indicates the device at time The state at that time, The threshold value is set.

5. A deep learning-based dynamic monitoring system for oil wear particles, characterized in that, Includes the following modules: The sensor module is used to collect real-time data on abrasive particles in the lubrication system of the equipment, including the number of abrasive particles, their size, shape, and chemical composition. The data processing module is configured to receive oil abrasive data collected by the sensor module and to preprocess, standardize, and denoise the oil abrasive data. A two-layer generative adversarial network model, consisting of a first-layer generator, a second-layer generator, and a discriminator, generates and discriminates abrasive image features and classifies abrasive feature data. The wear condition assessment module calculates the wear condition index of the equipment based on abrasive particle characteristic data and determines whether the equipment is in an abnormal state. The early warning module automatically generates an abnormal warning message and notifies maintenance personnel when it detects abnormal wear and tear on the equipment. The data storage and visualization module stores monitoring data and uses visualization tools to display the equipment's operating status and the changing trends of abrasive characteristics.

Images