Oil impurity detection system based on image recognition

Abstract

The invention discloses an oil impurity detection system based on image recognition, and relates to the technical field of industrial oil detection and intelligent operation and maintenance. A multi-dimensional image preprocessing module; a weak supervision dual-path detection module; a multi-source data traceability module; an intelligent decision output module; according to the invention, through fusion of visible light and dynamic visual data, single-mode detection limitation is broken through, known and unknown impurities are accurately identified through a multi-mode cooperation mechanism, detection robustness in a complex industrial scene is improved, and comprehensive data support is provided for equipment fault early warning; meanwhile, an impurity traceability system is constructed, dynamic topology matching and a time decay model are combined, an impurity source is rapidly positioned, equipment abrasion is quantified, an intelligent decision module generates a maintenance report and triggers load reduction protection, whole-process intelligent management is achieved, the diagnosis period is shortened, and the shutdown risk and the maintenance cost are reduced.

Description

Technical Field

[0001] This invention relates to the field of industrial oil detection and intelligent operation and maintenance technology, specifically an oil impurity detection system based on image recognition. Background Technology

[0002] In modern industrial systems, the efficient and stable operation of machinery is crucial for ensuring production safety and improving economic efficiency. As the core medium for mechanical transmission and lubrication, the cleanliness of hydraulic fluid directly affects the performance and lifespan of equipment. With the continuous improvement of industrial automation, the trend towards precision and complexity in machinery is becoming increasingly significant, placing increasingly stringent requirements on the purity of hydraulic fluids. Impurities mixed in the fluid, such as metal particles, fibers, and dust, not only accelerate the wear of mechanical parts but may also cause system failures, leading to production interruptions or even safety accidents. Therefore, real-time and accurate detection of the content and types of impurities in the fluid has become an important means of ensuring the healthy operation of equipment and preventing potential failures.

[0003] Traditional methods for detecting oil impurities mainly rely on chemical analysis, microscopic observation, or single-sensor-based detection technologies. While each method has its own characteristics, they all have significant limitations. Chemical analysis methods can accurately determine the composition of impurities, but they are complex to operate and time-consuming, making it difficult to meet the needs of real-time monitoring. Microscopic observation methods are limited by the observation range and resolution, and have limited ability to identify tiny or transparent impurities. While single-sensor-based detection technologies, such as turbidity sensors, can quickly reflect the degree of oil contamination, they cannot distinguish the type and source of impurities, making it difficult to provide accurate early warnings of equipment failures. In addition, traditional methods often lack the ability to identify unknown impurities, and cannot comprehensively assess the health status of equipment, limiting their application effectiveness in complex industrial environments. Summary of the Invention

[0004] The purpose of this invention is to overcome the shortcomings of existing technologies and provide an image recognition-based oil impurity detection system. This system breaks through the limitations of traditional single-modal detection by fusing visible light images and dynamic visual sensor data. It uses a multi-modal collaborative mechanism to accurately identify known and unknown impurities in oil. The system constructs an impurity tracing system and combines dynamic topology matching and time decay models to quickly locate the source of impurities and quantify equipment wear. The intelligent decision-making module generates maintenance reports, realizing intelligent management of the entire process, effectively improving the equipment fault early warning capability, and reducing the risk of unplanned downtime and maintenance costs.

[0005] To solve the above-mentioned technical problems, the present invention provides the following technical solution: an oil impurity detection system based on image recognition, the system comprising: Adaptive dual-modal acquisition module: The miniature controllable oil pump adjusts the sampling flow rate based on the oil contamination data fed back by the built-in turbidity sensor, and the dual-modal image acquisition unit simultaneously acquires visible light images of the oil and event stream data from the dynamic visual sensor; Multi-dimensional image preprocessing module: The FPGA unit performs noise reduction processing on the acquired visible light image and the dynamic visual sensor event stream data respectively. The GPU unit performs image enhancement and dual-modal image alignment on the noise-reduced data in sequence, and then calculates the fused feature vector through the multi-modal feature hypergraph fusion formula. Weakly supervised dual-path detection module: Based on fused feature vectors, the Mask-RCNN model is improved to perform instance segmentation on known impurities; for unidentified unknown impurity regions, the topological distortion of the feature manifold in the region is first calculated using the topological distortion calculation formula of the feature manifold, and then the topological distortion is processed by the variational autoencoder to generate the topological feature vector of the unknown impurity. Multi-source data tracing module: retrieves the material topology feature library from the equipment operation and maintenance database, collects the cumulative running time of the equipment being tested and calculates the time decay factor, and then calculates the similarity between the topology feature vector of the unknown impurity and the material features in the material topology feature library through the dynamic topology similarity calculation formula, and outputs the top three materials with the highest similarity and their corresponding equipment components. Intelligent decision output module: Integrates known impurity segmentation results with unknown impurity source tracing results, calculates equipment health index, generates inspection report containing known impurity statistics, unknown impurity feature maps and suspected wear parts markings, and sends the inspection report and control signals to external equipment control system through industrial communication protocol.

[0006] Furthermore, in the adaptive dual-modal acquisition module, the sampling flow rate adjustment range of the micro controllable oil pump is 0.5 to 5 ml per minute, the measurement accuracy of the turbidity sensor is ±2 NTU, the data refresh cycle is 50 milliseconds, and the dual-modal image acquisition unit includes an industrial CMOS camera and a dynamic vision sensor. The dual-modal image acquisition unit adopts an oil-electric separation structure.

[0007] Furthermore, the multimodal feature hypergraph fusion calculation formula in the multidimensional image preprocessing module is as follows: ,in, The fused feature vector; For adaptive modal weights; V represents visible light image features; The result of graph convolution processing for V; For dynamic visual sensor event flow characteristics, For E-commerce The global feature vector after encoder processing.

[0008] Furthermore, in the multi-dimensional image preprocessing module, the FPGA unit uses adaptive median filtering to perform noise reduction processing, and the filtering window size can be switched between 3×3 and 7×7; the GPU unit uses multi-scale Retinex decomposition to perform image enhancement, with a decomposition scale of 3 layers; and the dual-modal image alignment uses SIFT feature point matching.

[0009] Furthermore, the formula for calculating the topological distortion of the feature manifold in the weakly supervised dual-path detection module is as follows: ,in, The topological distortion of the characteristic manifold; The number of local feature points sampled from the fused feature vector; i and j are the indices of the local feature points, representing the i-th and j-th local feature points, respectively; Let be the feature vector of the i-th local feature point; Let j be the feature vector of the j-th local feature point; for The Riemann curvature tensor, derived from the second derivative matrix of its neighborhood features, is used to describe the curvature of the feature manifold at that point. for Riemann curvature tensor It is the square root of the sum of the squares of the matrix elements; It is the square root of the sum of squares of the vector elements.

[0010] Furthermore, in the weakly supervised dual-path detection module, the improved Mask-RCNN model includes a channel attention module and a spatial attention module; the variational autoencoder includes a 3-layer fully connected encoder and a 3-layer fully connected decoder, the encoder input dimension is 100 and the output dimension is 2048, the decoder and encoder have symmetrical structures, and the activation function used for both is ReLU.

[0011] Furthermore, the formula for calculating the time decay factor in the multi-source data tracing module is as follows: ,in, This is the time decay factor; The cumulative operating time of the equipment being tested; This is the attenuation coefficient, with a value ranging from 0.0008 to 0.0012, used to control the attenuation rate.

[0012] Furthermore, the dynamic topological similarity calculation formula in the multi-source data tracing module is as follows: ,in, The similarity between the unknown impurity and the k-th type of material; The topological feature vector of the unknown impurity; This is the feature vector of the k-th material in the material topology feature library; Cosine similarity; This is the time decay factor; This is the distribution weight coefficient, with a value of 0.3; for Divergence; used to measure the difference between two probability distributions; for The probability distribution; for The probability distribution.

[0013] Furthermore, in the multi-source data traceability module, the equipment operation and maintenance database adopts a 3-node distributed architecture, with each node having a storage capacity of no less than 1TB; the material topology feature library has no less than 1,000 samples for each type of material, covering three stages: initial wear, intermediate wear, and severe wear, and the various materials include gear steel, bearing steel, and other commonly used industrial materials.

[0014] Furthermore, the equipment health index calculation formula in the intelligent decision output module is: Health = (1 - Known impurity concentration exceedance rate × 0.6) × 100 - Unknown impurity traceability risk level × 5, where the known impurity concentration exceedance rate is the ratio of the number of exceedance impurities to the total number of impurities; the unknown impurity traceability risk level is divided according to the value of the first-ranked similarity, with a value not lower than 0.9 being level 5, 0.8 to 0.9 being level 3, and below 0.8 being level 1; the industrial communication protocol includes Modbus TCP protocol and OPCUA protocol, and the control signals include load reduction operation signals.

[0015] Compared with existing technologies, this image recognition-based oil impurity detection system has the following advantages: I. This invention overcomes the limitations of traditional single-modal detection in terms of sensitivity to ambient light and impurity morphology by fusing visible light images and dynamic visual sensor data. The adaptive dual-modal acquisition module can simultaneously capture the static morphology and dynamic motion characteristics of impurities in oil. Combined with noise reduction, enhancement, and alignment algorithms in multi-dimensional image preprocessing technology, it effectively eliminates data noise and improves feature consistency. The weakly supervised dual-path detection model, through improved instance segmentation algorithms and topological feature analysis methods, can accurately identify known types of impurities and discover unknown new impurities and extract their topological features. This multi-modal collaborative detection mechanism significantly improves the robustness of impurity detection in complex industrial scenarios and provides more comprehensive data support for early warning of equipment failures.

[0016] Second, the impurity tracing system constructed in this invention achieves rapid location of impurity sources and quantitative assessment of equipment wear status through dynamic topological similarity matching and time decay models. The system can automatically associate unknown impurity characteristics with equipment material libraries and, combined with operating time decay factors, accurately deduce the component location and wear stage where impurities are generated. The intelligent decision-making module integrates the detection results of known and unknown impurities to generate a decision report containing risk level, component labeling, and maintenance suggestions, and supports direct triggering of load reduction protection through industrial protocols. This intelligent management of the entire process from detection to tracing to maintenance decision-making significantly shortens the fault diagnosis cycle, shifts equipment maintenance from passive response to proactive prevention, and effectively reduces the risk of unplanned downtime and maintenance costs.

[0017] Other advantages, objectives and features of the invention will be set forth in part in the description which follows, and in part will be apparent to those skilled in the art from the following examination or study, or may be learned from the practice of the invention. Attached Figure Description

[0018] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the accompanying drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are merely some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without any creative effort.

[0019] Figure 1 This is a flowchart of an image recognition-based oil impurity detection system; Figure 2 This is a framework diagram of an image recognition-based oil impurity detection system. Detailed Implementation

[0020] To further illustrate the technical means and effects of the present invention in achieving its intended purpose, the following detailed description of the specific implementation methods, structures, features, and effects of the present invention, in conjunction with the accompanying drawings and preferred embodiments, is provided below.

[0021] Example 1: Industrial Gearbox Oil Impurity Detection Scenario During gearbox maintenance at a heavy machinery manufacturing plant, staff activated an image recognition-based oil impurity detection system to inspect the ZLY180 industrial gearbox for oil impurities. This was done to assess the gearbox's operational health and identify potential fault risks. The specific implementation process is as follows: Figure 1 As shown: The adaptive dual-modal acquisition module operates as follows: The system first activates the adaptive dual-modal acquisition module. A miniature controllable oil pump is connected to the oil circulation pipeline of the gearbox. A built-in turbidity sensor monitors the oil contamination level in real time and feeds the data back to the miniature controllable oil pump. Based on the feedback contamination data, the miniature controllable oil pump automatically adjusts the sampling flow rate of oil drawn from the gearbox, ensuring that an appropriate amount of oil sample is obtained for subsequent testing regardless of the degree of contamination. Simultaneously, the dual-modal image acquisition unit begins operation. This unit includes an industrial CMOS camera and a dynamic vision sensor. Its oil-electricity separation structure avoids the impact of electrical interference on image acquisition. Both simultaneously acquire visible light images of the oil sample and event stream data from the dynamic vision sensor, providing the foundational data for subsequent processing.

[0022] The multi-dimensional image preprocessing module works as follows: Acquired visible light images and dynamic visual sensor event stream data are transmitted to the multi-dimensional image preprocessing module. First, the FPGA unit performs noise reduction on both types of data separately, using adaptive median filtering to automatically adjust the filtering window based on the noise level, effectively removing noise interference. After noise reduction, the GPU unit takes over, performing image enhancement on the denoised data. Multi-scale Retinex decomposition is used to improve image contrast and detail clarity, making impurity features in the oil easier to identify. Next, SIFT feature point matching is used to align the two modal images, ensuring spatial consistency between the two types of images. Finally, the multi-modal feature hypergraph fusion formula is applied to fuse the processed two-modal data, obtaining a fused feature vector that comprehensively reflects the characteristics of oil impurities. The multi-modal feature hypergraph fusion formula is as follows: ,in, The fused feature vector; For adaptive modal weights; V represents visible light image features; The result of graph convolution processing for V; For dynamic visual sensor event flow characteristics, For E-commerce The global feature vector after encoder processing.

[0023] The weakly supervised dual-path detection module works as follows: The fused feature vector is fed into the weakly supervised dual-path detection module. On one hand, the improved Mask-RCNN model processes the fused feature vector. This model includes channel attention and spatial attention modules, enabling more precise focusing on impurity features in the oil. It performs instance segmentation on known types of impurities, clearly identifying the location, outline, and quantity of known impurities. On the other hand, for unknown impurity regions not identified by the improved Mask-RCNN model, the system first calculates the topological distortion of the feature manifold in that region using the feature manifold topological distortion calculation formula. This describes the spatial distribution characteristics of the features in the unknown impurity region. The feature manifold topological distortion calculation formula is: ,in, The topological distortion of the characteristic manifold; The number of local feature points sampled from the fused feature vector; i and j are the indices of the local feature points, representing the i-th and j-th local feature points, respectively; Let be the feature vector of the i-th local feature point; Let j be the feature vector of the j-th local feature point; for The Riemann curvature tensor; for Riemann curvature tensor It is the square root of the sum of the squares of the matrix elements; The calculated topological distortion is then input into a variational autoencoder for processing, generating an unknown impurity topological feature vector that can characterize the features of the unknown impurities. This variational autoencoder consists of a 3-layer fully connected encoder and a 3-layer fully connected decoder, with ReLU activation function used for all layers.

[0024] The multi-source data tracing module works as follows: The generated unknown impurity topological feature vector is transmitted to the multi-source data tracing module. The system retrieves the material topological feature library from the equipment operation and maintenance database, which adopts a 3-node distributed architecture. Each node has a storage capacity of no less than 1TB, and the sample size for each material type in this material topological feature library is no less than 1000, covering three stages: initial wear, intermediate wear, and severe wear. The samples include feature data of commonly used industrial materials such as gear steel and bearing steel. Simultaneously, the system collects the cumulative running time of the currently inspected ZLY180 gearbox and calculates the corresponding time decay factor according to the time decay factor calculation formula. The time decay factor calculation formula is: ,in, This is the time decay factor; The cumulative operating time of the equipment being tested; The decay coefficient is then used. Next, the dynamic topological similarity calculation formula is applied to calculate the similarity between the topological feature vector of the unknown impurity and the feature vectors of each material in the material topological feature library. Taking into account factors such as cosine similarity, time decay factor, and KL divergence, the top three similarity materials and their corresponding device components are finally output. The dynamic topological similarity calculation formula is: ,in, The similarity between the unknown impurity and the k-th type of material; The topological feature vector of the unknown impurity; This is the feature vector of the k-th material in the material topology feature library; Cosine similarity; This is the time decay factor; This is the distribution weight coefficient, with a value of 0.3; for Divergence; for The probability distribution; for The probability distribution.

[0025] The intelligent decision output module integrates the known impurity segmentation results from the weakly supervised dual-path detection module with the unknown impurity tracing results from the multi-source data tracing module. It then calculates the gearbox's health index using the equipment health index calculation formula. The known impurity concentration exceedance rate is the ratio of the number of exceedance impurities to the total number of impurities. The unknown impurity tracing risk level is determined by the highest similarity ranking: level 5 (not lower than 0.9), level 3 (0.8 to 0.9), and level 1 (below 0.8). The system then generates a detailed inspection report containing known impurity statistics, unknown impurity feature maps, and annotations of suspected worn parts. Finally, the inspection report and control signals are sent to the external equipment control system via industrial communication protocols, including Modbus TCP and OPCUA. If the gearbox health index is low, a load reduction signal can be sent to prevent further escalation of the fault, completing the entire gearbox oil impurity detection process.

[0026] In summary, in the scenario of detecting oil impurities in industrial gearboxes, the image recognition-based oil impurity detection system completes the detection through the orderly collaboration of five modules. The adaptive dual-modal acquisition module adjusts the sampling flow rate based on turbidity sensor data, simultaneously acquiring dual-modal data; the multi-dimensional image preprocessing module obtains a fused feature vector through noise reduction, enhancement, alignment, and fusion; the weakly supervised dual-path detection module achieves known impurity segmentation and unknown impurity topological feature vector generation; the multi-source data tracing module, combined with the equipment maintenance database, completes the tracing of unknown impurities; and the intelligent decision output module calculates the health index, generates reports, and sends signals, providing accurate data support for gearbox maintenance and ensuring its stable operation.

[0027] Example 2: Detection of Impurities in Wind Turbine Main Shaft Bearing Oil During the operation and maintenance of wind power equipment at a wind farm, maintenance personnel used an image recognition-based oil impurity detection system to detect impurities in the main shaft bearing oil of a 1.5MW wind turbine, ensuring the safe and stable operation of the wind turbine. The specific implementation steps are as follows: Figure 2 As shown: The adaptive dual-modal acquisition module operates as follows: Maintenance personnel connect the system's adaptive dual-modal acquisition module to the oil sampling port of the wind turbine's main shaft bearing. After starting the module, the miniature controllable oil pump begins operation. The built-in turbidity sensor continuously monitors the contamination level of the main shaft bearing oil and feeds real-time data back to the miniature controllable oil pump. Based on this data, the miniature controllable oil pump flexibly adjusts the sampling flow rate to ensure that the collected oil sample accurately reflects the actual contamination status of the oil. Simultaneously, the dual-modal image acquisition unit starts up. This unit consists of an industrial CMOS camera and a dynamic vision sensor, employing an oil-electric separation structure. The two respectively acquire visible light images of the oil and event stream data from the dynamic vision sensor, providing raw data support for subsequent impurity detection.

[0028] The multi-dimensional image preprocessing module works as follows: The acquired raw data is transmitted to the multi-dimensional image preprocessing module. First, the FPGA unit performs noise reduction processing on the visible light image and the dynamic visual sensor event stream data separately. Adaptive median filtering technology is used to automatically switch the filter window size according to the different noise intensities in the data, efficiently filtering out noise and improving data quality. The two types of data after noise reduction are then passed to the GPU unit. The GPU unit first uses the multi-scale Retinex decomposition method to enhance the image of the data, enhancing the visual features of impurities in the oil and making the impurities clearer. Then, SIFT feature point matching is used to align the two modal images to ensure accurate matching of the two images in spatial dimensions. Finally, the multi-modal feature hypergraph fusion formula is used to fuse the aligned two modal data to generate a fused feature vector. This vector integrates the advantages of the two image data and can more comprehensively reflect the impurity information in the oil.

[0029] The weakly supervised dual-path detection module operates as follows: After the fused feature vectors enter the module, the dual-path detection process begins. In one path, an improved Mask-RCNN model analyzes the fused feature vectors. This model includes channel attention and spatial attention modules, accurately identifying and segmenting known impurities in the oil, such as metal particles from bearing wear and rubber debris from aging seals, recording their location, size, quantity, and other relevant information. In the other path, for unknown impurity regions not identified by the aforementioned model, the topological distortion of the feature manifold in the region is calculated using the feature manifold topological distortion calculation formula. This distortion is then input into a variational autoencoder (VAE), a 3-layer fully connected encoder and decoder structure employing the ReLU activation function. After processing, an unknown impurity topological feature vector is generated for subsequent impurity source tracing.

[0030] The multi-source data tracing module works as follows: After receiving the topological feature vector of an unknown impurity, the module retrieves the material topological feature library from the equipment operation and maintenance database. This database is a 3-node distributed architecture with a single node storage capacity of no less than 1TB. The material topological feature library covers various materials commonly used in wind turbine equipment, such as bearing steel commonly used in main shaft bearings and gear steel related to gearboxes. Each material category has no fewer than 1000 samples, covering different stages of material wear from initial wear to severe wear, comprehensively reflecting the characteristics of various materials under different wear states. Simultaneously, the system obtains the cumulative operating time of the current wind turbine main shaft bearing and calculates the time decay factor using the time decay factor calculation formula. Then, the dynamic topological similarity calculation formula is used to compare the topological feature vector of the unknown impurity with the feature vectors of various materials in the material topological feature library. After comprehensively considering various influencing factors, the top three materials with the highest similarity to the unknown impurity and their corresponding wind turbine components, such as the inner and outer rings of the main shaft bearing and the cage, are determined and output.

[0031] The intelligent decision-making output module integrates the segmentation results of known impurities and the source tracing results of unknown impurities. The segmentation results of known impurities include the types, quantities, and distribution of known impurities, while the source tracing results of unknown impurities include the top three materials with the highest similarity and their corresponding components. Based on the equipment health index calculation formula, and combining the known impurity concentration exceedance rate and the source tracing risk level of unknown impurities, the health index of the wind turbine main shaft bearing is calculated. The known impurity concentration exceedance rate is the ratio of the number of exceedance impurities to the total number of impurities, and the source tracing risk level of unknown impurities is determined based on the value ranked first in similarity. The system then generates a detailed inspection report, which includes statistical data on known impurities, characteristic maps of unknown impurities, and annotations of components suspected of wear problems. Finally, using the Modbus TCP and OPCUA industrial communication protocols, the inspection report and corresponding control signals are sent to the wind farm's external equipment control system. If the health index is too low, a load reduction signal can be sent to prevent serious failures caused by excessive wear of the main shaft bearing, providing a reliable basis for the wind turbine's operation and maintenance decisions and completing the wind turbine main shaft bearing oil impurity detection work.

[0032] In summary, in the scenario of detecting impurities in the oil of wind turbine main shaft bearings, this detection system also relies on five core modules to operate. The adaptive dual-modal acquisition module ensures that oil sample collection is adapted to the contamination status; the multi-dimensional image preprocessing module optimizes data and fuses it to generate feature vectors; the weakly supervised dual-path detection module accurately identifies known impurities and acquires features of unknown impurities; the multi-source data tracing module combines equipment operation data and a material library to complete tracing; and the intelligent decision output module integrates results to generate reports and control signals, providing a reliable basis for the operation and maintenance of wind turbine main shaft bearings, effectively preventing serious failures caused by wear, and ensuring the safe and stable operation of wind farm equipment.

[0033] The above description is merely a preferred embodiment of the present invention and is not intended to limit the present invention in any way. Although the present invention has been disclosed above with reference to preferred embodiments, it is not intended to limit the present invention. Any person skilled in the art can make some modifications or alterations to the above-disclosed technical content to create equivalent embodiments without departing from the scope of the present invention. Any simple modifications, equivalent changes and alterations made to the above embodiments based on the technical essence of the present invention without departing from the scope of the present invention shall still fall within the scope of the present invention.

Claims

1. An image recognition-based oil impurity detection system, characterized by, The system includes: Adaptive dual-modal acquisition module: The miniature controllable oil pump adjusts the sampling flow rate based on the oil contamination data fed back by the built-in turbidity sensor, and the dual-modal image acquisition unit simultaneously acquires visible light images of the oil and event stream data from the dynamic visual sensor; Multi-dimensional image preprocessing module: The FPGA unit performs noise reduction processing on the acquired visible light image and the dynamic visual sensor event stream data respectively. The GPU unit performs image enhancement and dual-modal image alignment on the noise-reduced data in sequence, and then calculates the fused feature vector through the multi-modal feature hypergraph fusion formula. Weakly supervised dual-path detection module: Based on fused feature vectors, the Mask-RCNN model is improved to perform instance segmentation on known impurities; for unidentified unknown impurity regions, the topological distortion of the feature manifold in the region is first calculated using the topological distortion calculation formula of the feature manifold, and then the topological distortion is processed by the variational autoencoder to generate the topological feature vector of the unknown impurity. Multi-source data tracing module: retrieves the material topology feature library from the equipment operation and maintenance database, collects the cumulative running time of the equipment being tested and calculates the time decay factor, and then calculates the similarity between the topology feature vector of the unknown impurity and the material features in the material topology feature library through the dynamic topology similarity calculation formula, and outputs the top three materials with the highest similarity and their corresponding equipment components. Intelligent decision output module: Integrates known impurity segmentation results with unknown impurity source tracing results, calculates equipment health index, generates inspection report containing known impurity statistics, unknown impurity feature maps and suspected wear parts markings, and sends the inspection report and control signals to external equipment control system through industrial communication protocol.

2. The oil impurity detection system based on image recognition according to claim 1, characterized in that, In the adaptive dual-modal acquisition module, the sampling flow rate adjustment range of the micro controllable oil pump is 0.5 to 5 ml per minute, the measurement accuracy of the turbidity sensor is ±2 NTU, the data refresh cycle is 50 milliseconds, and the dual-modal image acquisition unit includes an industrial CMOS camera and a dynamic vision sensor. The dual-modal image acquisition unit adopts an oil-electric separation structure.

3. The oil impurity detection system based on image recognition according to claim 1, characterized in that, The multimodal feature hypergraph fusion calculation formula in the multidimensional image preprocessing module is as follows: ,in, The fused feature vector; For adaptive modal weights; V represents visible light image features; The result of graph convolution processing for V; For dynamic visual sensor event flow characteristics, For E-commerce The global feature vector after encoder processing.

4. The oil impurity detection system based on image recognition according to claim 1, characterized in that, In the multi-dimensional image preprocessing module, the FPGA unit uses adaptive median filtering to perform noise reduction, and the filter window size can be switched between 3×3 and 7×7; the GPU unit uses multi-scale Retinex decomposition to perform image enhancement, with a decomposition scale of 3 layers; and the dual-modal image alignment uses SIFT feature point matching.

5. The oil impurity detection system based on image recognition according to claim 1, characterized in that, The formula for calculating the topological distortion of the feature manifold in the weakly supervised dual-path detection module is as follows: ,in, The topological distortion of the characteristic manifold; The number of local feature points sampled from the fused feature vector; i and j are the indices of the local feature points, representing the i-th and j-th local feature points, respectively; Let be the feature vector of the i-th local feature point; Let j be the feature vector of the j-th local feature point; for The Riemann curvature tensor; for Riemann curvature tensor It is the square root of the sum of the squares of the matrix elements; It is the square root of the sum of squares of the vector elements.

6. The oil impurity detection system based on image recognition according to claim 1, characterized in that, In the weakly supervised dual-path detection module, the improved Mask-RCNN model includes a channel attention module and a spatial attention module; the variational autoencoder includes a 3-layer fully connected encoder and a 3-layer fully connected decoder. The encoder has an input dimension of 100 and an output dimension of 2048. The decoder and encoder have symmetrical structures, and both use ReLU as the activation function.

7. The oil impurity detection system based on image recognition according to claim 1, characterized in that, The formula for calculating the time decay factor in the multi-source data tracing module is as follows: ,in, This is the time decay factor; The cumulative operating time of the equipment being tested; This is the attenuation coefficient.

8. The oil impurity detection system based on image recognition according to claim 1, characterized in that, The dynamic topological similarity calculation formula in the multi-source data tracing module is as follows: ,in, The similarity between the unknown impurity and the k-th type of material; The topological feature vector of the unknown impurity; This is the feature vector of the k-th material in the material topology feature library; Cosine similarity; This is the time decay factor; This is the distribution weight coefficient, with a value of 0.3; for Divergence; for The probability distribution; for The probability distribution.

9. The oil impurity detection system based on image recognition according to claim 1, characterized in that, In the multi-source data traceability module, the equipment operation and maintenance database adopts a 3-node distributed architecture, with each node having a storage capacity of no less than 1TB; the material topology feature library has no less than 1,000 samples for each type of material, covering three stages: initial wear, intermediate wear, and severe wear. The various materials include gear steel and bearing steel, which are commonly used industrial materials.

10. The oil impurity detection system based on image recognition according to claim 1, characterized in that, The equipment health index calculation formula in the intelligent decision output module is: Health = (1 - known impurity concentration exceedance rate × 0.6) × 100 - unknown impurity source traceability risk level × 5, where the known impurity concentration exceedance rate is the ratio of the number of exceeding impurities to the total number of impurities; the unknown impurity source traceability risk level is divided according to the value of the first-ranked similarity, with level 5 being no less than 0.9, level 3 being between 0.8 and 0.9, and level 1 being less than 0.8; the industrial communication protocol includes Modbus TCP protocol and OPCUA protocol, and the control signals include load reduction operation signals.

Images