Search

CN-121999327-A - Oil impurity detecting system based on image recognition

CN121999327ACN 121999327 ACN121999327 ACN 121999327ACN-121999327-A

Abstract

The invention discloses an oil impurity detection system based on image recognition, which relates to the technical field of industrial oil detection and intelligent operation and maintenance, and comprises a self-adaptive bimodal acquisition module; a multidimensional image preprocessing module; the intelligent decision-making system comprises a weak supervision dual-path detection module, a multi-source data tracing module, an intelligent decision-making output module, a multi-mode cooperative mechanism, a dynamic topology matching and time attenuation model, an impurity tracing system, an intelligent decision-making module, a maintenance report generation and load reduction protection triggering module, a full-flow intelligent management module, a diagnosis period shortening module and a shutdown risk and maintenance cost reduction module.

Inventors

  • WANG YUESHENG
  • ZOU WENLIANG

Assignees

  • 北京菲舍波特科技发展有限公司

Dates

Publication Date
20260508
Application Date
20251204

Claims (10)

  1. 1. Oil impurity detecting system based on image recognition, its characterized in that, this system includes: The self-adaptive bimodal acquisition module is used for adjusting the sampling flow of the miniature controllable oil pump according to the oil pollution data fed back by the built-in turbidity sensor, and synchronously acquiring the visible light image of the oil and the event flow data of the dynamic vision sensor by the bimodal image acquisition unit; The multi-dimensional image preprocessing module comprises an FPGA unit, a GPU unit, a multi-mode feature hypergraph fusion formula and a multi-mode feature hypergraph fusion module, wherein the FPGA unit respectively executes noise reduction processing on the acquired visible light image and the event stream data of the dynamic vision sensor, the GPU unit sequentially executes image enhancement and dual-mode image alignment on the noise reduced data, and then fusion feature vectors are obtained through calculation of the multi-mode feature hypergraph fusion formula; For unrecognized unknown impurity regions, calculating the topological torsion of the characteristic manifold of the region through a characteristic manifold topological torsion calculation formula, and then processing the topological torsion through a variation self-encoder to generate unknown impurity topological feature vectors; The multi-source data tracing module is used for calling a material topology feature library from an equipment operation and maintenance database, collecting the accumulated running time of the equipment to be detected, calculating a time attenuation factor, calculating the similarity between an unknown impurity topology feature vector and each material feature in the material topology feature library through a dynamic topology similarity calculation formula, and outputting the material with the top three ranks of the similarity and the corresponding equipment components; And the intelligent decision output module integrates the known impurity segmentation result and the unknown impurity tracing result, calculates the equipment health index, generates a detection report containing the known impurity statistical data, the unknown impurity characteristic map and the suspected wear part label, and sends the detection report and a control signal to an external equipment control system through an industrial communication protocol.
  2. 2. The oil impurity detection system based on image recognition according to claim 1, wherein in the self-adaptive bimodal collection module, the sampling flow adjustment range of the miniature controllable oil pump is 0.5-5 milliliters per minute, the measurement precision of the turbidity sensor is +/-2 NTU, the data refreshing period is 50 milliseconds, the bimodal image collection unit comprises an industrial CMOS camera and a dynamic vision sensor, and the bimodal image collection unit adopts an oil-electricity separation structure.
  3. 3. The oil impurity detection system based on image recognition according to claim 1, wherein the multi-modal feature hypergraph fusion calculation formula in the multi-dimensional image preprocessing module is as follows: , wherein, The feature vector is the feature vector after fusion; is an adaptive modal weight; V is the visible light image characteristic; the result of the graph convolution processing is V; for dynamic visual sensor event stream features, Is E channel The global feature vector processed by the encoder.
  4. 4. The oil impurity detection system based on image recognition according to claim 1, wherein in the multi-dimensional image preprocessing module, the FPGA unit performs noise reduction processing by adopting adaptive median filtering, the size of a filtering window can be switched in a range of 3×3 to 7×7, the GPU unit performs image enhancement by adopting multi-scale Retinex decomposition, the decomposition scale is 3 layers, and the bimodal image alignment adopts a SIFT feature point matching mode.
  5. 5. The image recognition-based oil impurity detection system according to claim 1, wherein the characteristic manifold topology torsion degree calculation formula in the weak supervision dual-path detection module is: , wherein, Topological torsion degree of characteristic manifold; I and j are indexes of local feature points and respectively represent the ith and jth local feature points; the feature vector is the i-th local feature point; The feature vector is the feature vector of the jth local feature point; Is that A Riemann curvature tensor; Is that Riemann curvature tensor Square root of sum of squares of matrix elements; Is the square root of the sum of the squares of the vector elements.
  6. 6. The oil impurity detection system based on image recognition according to claim 1, wherein in the weak supervision dual-path detection module, the improved Mask-RCNN model comprises a channel attention module and a space attention module, the variable self-encoder comprises a 3-layer full-connection encoder and a 3-layer full-connection decoder, the input dimension of the encoder is 100, the output dimension of the encoder is 2048, the decoder is symmetrical to the encoder structure, and the activation functions are all ReLU.
  7. 7. The oil impurity detection system based on image recognition according to claim 1, wherein the time attenuation factor calculation formula in the multi-source data tracing module is as follows: , wherein, Is a time decay factor; accumulated running time for the detection target device; Is the attenuation coefficient.
  8. 8. The oil impurity detection system based on image recognition according to claim 1, wherein a dynamic topological similarity calculation formula in the multi-source data tracing module is as follows: , wherein, The similarity between unknown impurities and the k-th material; is the topological feature vector of the unknown impurity; The feature vector is the feature vector of the kth class of materials in the material topology feature library; cosine similarity; Is a time decay factor; The value of the distributed weight coefficient is 0.3; Is that Divergence; Is that Probability distribution of (2); Is that Is a probability distribution of (c).
  9. 9. The oil impurity detection system based on image recognition according to claim 1, wherein in the multi-source data tracing module, a 3-node distributed architecture is adopted in an equipment operation and maintenance database, the storage capacity of each node is not lower than 1TB, the sample size of each type of material in a material topology feature library is not lower than 1000, and the sample covers three stages of initial abrasion, medium abrasion and serious abrasion, and various materials comprise gear steel and bearing steel industry common materials.
  10. 10. The oil impurity detection system based on image recognition according to claim 1, wherein the equipment health index calculation formula in the intelligent decision output module is health= (1-known impurity concentration exceeding rate×0.6) ×100-unknown impurity tracing risk level×5, wherein the known impurity concentration exceeding rate is the ratio of the number of exceeding impurities to the total number of impurities, the unknown impurity tracing risk level is divided according to the value of the first bit of similarity ranking, is 5 level when not lower than 0.9, is 3 level when 0.8 to 0.9, is 1 level when lower than 0.8, and the industrial communication protocol comprises modbusTCP protocol and OPCUA protocol, and the control signal comprises a descending running signal.

Description

Oil impurity detecting system based on image recognition Technical Field The invention relates to the technical field of industrial oil detection and intelligent operation and maintenance, in particular to an oil impurity detection system based on image recognition. Background In modern industrial systems, efficient and stable operation of mechanical equipment is a key for guaranteeing production safety and improving economic benefits. The cleanliness of the oil, which serves as a core medium for mechanical transmission and lubrication, is directly related to the performance and service life of the equipment. Along with the continuous improvement of the industrial automation degree, the trend of the precision and the complexity of mechanical equipment is increasingly remarkable, and the requirement on the purity of oil is also more severe. Impurities mixed in the oil, such as metal particles, fibers, dust and the like, not only can aggravate the abrasion of mechanical parts, but also can cause system faults, thereby causing production interruption and even safety accidents. Therefore, the method for detecting the impurity content and the impurity type in the oil liquid accurately in real time becomes an important means for guaranteeing the healthy operation of equipment and preventing potential faults. The traditional oil impurity detection method mainly depends on chemical analysis, microscopic observation or detection technology based on a single sensor, although the methods have the characteristics, the method has obvious limitations, the chemical analysis method can accurately measure impurity components, but is complex in operation and long in time consumption, the real-time monitoring requirement is difficult to meet, the microscopic observation rule is limited by the observation range and resolution, the identification capability of tiny or transparent impurities is limited, the detection technology based on the single sensor, such as a turbidity sensor, can rapidly reflect the pollution degree of oil, the types and sources of the impurities cannot be distinguished, the accurate early warning of equipment faults is difficult to perform, in addition, the traditional method often lacks the identification capability of unknown impurities, the health state of equipment cannot be comprehensively estimated, and the application effect of the method in complex industrial environments is limited. Disclosure of Invention The invention aims to make up the defects of the prior art, provides an oil impurity detection system based on image recognition, which breaks through the limitation of the traditional single-mode detection by fusing visible light images and dynamic vision sensor data, accurately recognizes known and unknown impurities in oil by utilizing a multi-mode cooperative mechanism, the system builds an impurity traceability system, combines a dynamic topology matching and time attenuation model, rapidly locates impurity sources and quantifies equipment abrasion, and the intelligent decision module generates a maintenance report, so that full-flow intelligent management is realized, the fault early warning capability of the equipment is effectively improved, and the unplanned shutdown risk and maintenance cost are reduced. The invention provides a technical scheme for solving the technical problems, which comprises the following steps of: The self-adaptive bimodal acquisition module is used for adjusting the sampling flow of the miniature controllable oil pump according to the oil pollution data fed back by the built-in turbidity sensor, and synchronously acquiring the visible light image of the oil and the event flow data of the dynamic vision sensor by the bimodal image acquisition unit; The multi-dimensional image preprocessing module comprises an FPGA unit, a GPU unit, a multi-mode feature hypergraph fusion formula and a multi-mode feature hypergraph fusion module, wherein the FPGA unit respectively executes noise reduction processing on the acquired visible light image and the event stream data of the dynamic vision sensor, the GPU unit sequentially executes image enhancement and dual-mode image alignment on the noise reduced data, and then fusion feature vectors are obtained through calculation of the multi-mode feature hypergraph fusion formula; For unrecognized unknown impurity regions, calculating the topological torsion of the characteristic manifold of the region through a characteristic manifold topological torsion calculation formula, and then processing the topological torsion through a variation self-encoder to generate unknown impurity topological feature vectors; The multi-source data tracing module is used for calling a material topology feature library from an equipment operation and maintenance database, collecting the accumulated running time of the equipment to be detected, calculating a time attenuation factor, calculating the similarity between an unknown impurity topology featu