A method and system for oil particle image recognition analysis based on reinforcement learning

Abstract

The present application relates to the technical field of oil particle identification, in particular to an oil particle image recognition and analysis method and system based on reinforcement learning; when the data of the reinforcement learning model is used for real-time monitoring of the oil to be monitored, the trained reinforcement learning model receives the infrared image to be detected, automatically separates the single particles and large particle substances in the recognition image, combines the large particle substances with their projection area and the average particle size threshold value S3, calculates the number of single particles by dividing the large particle area by the small particle area, and obtains how many small particles constitute a large particle, thereby avoiding the counting error of small particle size; after calculation, the results are output, and the output content is the actual count of particles in the output oil (the sum of the number of single particles and the number of single particles after the agglomerate is split), which provides data support for equipment damage diagnosis; and the problem of low count caused by misjudgment of agglomerates as single particles due to particle agglomeration in high-viscosity oil is avoided.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of oil particle identification, in particular to an oil particle image identification analysis method and system based on reinforcement learning. BACKGROUND

[0002] In the field of industrial equipment operation and maintenance, oil is the core medium for equipment lubrication and cooling. The state of particles in the oil directly reflects the wear degree and operation health of the equipment. For example, the increase in the number and size of metal wear particles in the hydraulic system often indicates early failure of key components such as bearings and gears. Contaminants in hydraulic oil and gear oil, such as dust and fibers, can exacerbate seal wear and cause system leaks. Therefore, accurate identification (type differentiation), counting (number statistics), and size measurement of particles in oil are the core technical support for equipment fault diagnosis and predictive maintenance, and are also a key link to ensure long-term stable operation of industrial equipment.

[0003] With the development of image processing and artificial intelligence technologies, oil particle detection has gradually evolved from traditional offline weighing methods and manual counting methods under a microscope to automated image recognition and analysis technologies based on machine vision.

[0004] Existing automated detection solutions typically use high-resolution optical imaging systems (such as microscopes and laser particle analyzers) to capture oil sample images, and then use image processing algorithms (such as threshold segmentation and edge detection) or machine learning models (such as convolutional neural networks and support vector machines) to extract features, classify, and calculate parameters of particles in the images. The type, number, and size distribution of particles are finally output, significantly improving detection efficiency and data objectivity.

[0005] However, in actual industrial scenarios, the viscosity of commonly used lubricating oil and hydraulic oil for industrial equipment is relatively high at room temperature or low temperature, and the adsorption force between oil molecules is strong, which causes particles in the oil to easily form agglomerates through the bridging action of oil molecules. This agglomeration phenomenon directly leads to errors in particle counting and size measurement using existing image recognition and analysis technologies.

[0006] To address the technical problems presented in this paper, the present application proposes an oil particle image identification analysis method and system based on reinforcement learning. SUMMARY

[0007] The present application proposes an oil particle image identification analysis method based on reinforcement learning, comprising the following steps:

[0008] S1: Obtain the oil particle image by infrared acquisition mode, select a transistor as the oil flow channel, and the oil flows through the transistor; select an infrared light source as the detection light, and the infrared light irradiates the transistor; select an infrared imaging assembly as the imaging system, and the transistor is located between the infrared light source and the infrared imaging assembly;

[0009] S2: Preprocess the infrared image to obtain a standardized image, and the preprocessing at least includes graying, adaptive contrast enhancement and edge sharpening, so as to highlight the gray difference between the particles and the background and the outline boundary of the single particles in the agglomerates;

[0010] S3: In the process of obtaining the oil particles, the average particle size threshold in the oil is calculated in real time according to the infrared data collected in step S1, and the average particle size threshold is the statistical average of the equivalent diameters of the single particles;

[0011] S4: Construct a reinforcement learning model, and the reinforcement learning model takes the connected regions in the standardized image as the observation object, is used for distinguishing the single particles and the large particle substances, and averages the large particle substances into single particles according to the average value in S3;

[0012] S5: Train the reinforcement learning model by using a labeled data set, and the labeled data set includes oil infrared images of different agglomeration degrees and corresponding labeled information, and the labeled information at least includes the position of the large particle substance, the number of single particles in the large particle substance, the actual size of each single particle and the average particle size threshold;

[0013] S6: Input the infrared image of the oil to be detected into the trained reinforcement learning model, when the large particle substance adhered together irradiated by the infrared light is detected, the model calculates how many small particles are combined to form the large particle substance according to the size of the large particle substance and the average particle size threshold calculated in step S3, and outputs the actual counting result of the particles in the oil to be detected and the actual size distribution of each single particle.

[0014] As a preferred scheme of the present application, in S1, three infrared light sources distributed upward and downward are selected as the detection light, wherein the middle light source horizontally irradiates the transistor, the upper light source obliquely downward irradiates the transistor, and the lower light source obliquely upward irradiates the transistor, the middle light source and the axis of the transistor have an angle of 0°, the upper light source and the axis have an angle of 30° to 45°, and the lower light source and the axis have an angle of -30° to -45°, and the three light sources are synchronously triggered.

[0015] As a preferred scheme of the present application, the cross section of the transistor is a flat rectangle, and the material of the transistor is fused quartz glass.

[0016] As a preferred scheme of the present application, the real-time calculation logic of the average particle size threshold in step S3 is: the equivalent diameter of all regions determined as single particles in the collected infrared image is measured, the first 5% maximum value and the last 5% minimum value in the measured value are removed, and the arithmetic mean value of the remaining measured values is taken as the average particle size threshold.

[0017] As a preferred scheme of the present application, the single particle number calculation method of the large particle substance in step S6 is: the three-dimensional contour volume V of the large particle substance is extracted through the reinforcement learning model, and the single particle volume V0 corresponding to the average particle size threshold is combined, V0=πd 3 / 6, d is the average particle size threshold, and the single particle number n=round(V / V0K) is calculated, wherein k is a shape correction coefficient, the value range is 0.8-1.2, and the irregularity of the large particle substance is dynamically adjusted.

[0018] An oil particle image recognition and analysis system based on reinforcement learning, which is used to realize the oil particle image recognition and analysis method based on reinforcement learning; the system comprises:

[0019] An infrared acquisition module, comprising a flat rectangular straight channel transistor made of fused quartz glass, three infrared light sources and an infrared imaging assembly; the transistor is an oil flow channel, and the cross section is a flat rectangle; the three infrared light sources are distributed in the upper, middle and lower positions, the middle light source horizontally irradiates the transistor, and the upper and lower light sources obliquely irradiate the transistor downward and upward respectively; the infrared imaging assembly is arranged opposite to the transistor, and is used to receive infrared light transmitted through the transistor and generate an oil particle infrared image;

[0020] A preprocessing module connected with the infrared acquisition module, used to perform grayscale, contrast enhancement and edge sharpening processing on the infrared image, and highlight the single particle contour inside the large particle substance;

[0021] A particle size threshold calculation module connected with the infrared acquisition module, used to calculate and update the average particle size threshold of the oil particle according to the real-time collected infrared data;

[0022] A reinforcement learning model module connected with the preprocessing module and the particle size threshold calculation module respectively, used to receive the standardized image and the average particle size threshold, and output the single particle splitting result, the actual count and the size distribution of the large particle substance;

[0023] A result output module connected with the reinforcement learning model module, used to display the particle count, the size distribution and the agglomeration ratio of the large particle substance.

[0024] As a preferred scheme of the present application, the infrared imaging assembly adopts a refrigeration type infrared focal plane array detector, the frame rate is matched with the light source acquisition frequency, and the imaging assembly further has a three-dimensional reconstruction unit built-in, which is used to generate a three-dimensional contour model of the particle in combination with the image data of the three light sources.

[0025] As a preferred scheme of the application, the reinforcement learning model module comprises a multi-light-source feature fusion unit for weighting and fusing the particle gray scale features and edge features under the irradiation of three light sources, and the weight coefficients are dynamically adjusted according to the light source angle and the particle position, wherein the feature weight of the corresponding light source is increased by 1.2-1.5 times for the overlapping area of the large particle matter.

[0026] The beneficial effects of the application are as follows:

[0027] When the data of the reinforcement learning model is used for real-time monitoring of the oil liquid to be monitored, after the trained reinforcement learning model receives the infrared image to be detected, the single particles and the large particle matter in the identified image are automatically recognized, and the large particle matter is combined with the average particle size threshold S3 according to the projection area to calculate the number of single particles. The calculation method is that the large particle area is divided by the small particle area to obtain how many small particles constitute the large particle, thereby avoiding the counting error of the small particle size; after the calculation, the result is output, and the output content is the actual count of the particles in the output oil liquid (the sum of the number of single particles and the number of single particles after the agglomerate is split), which provides data support for equipment damage diagnosis; thereby avoiding the problem that the particle agglomeration in the high-viscosity oil liquid causes the agglomerate to be misjudged as a single particle, resulting in low counting. BRIEF DESCRIPTION OF DRAWINGS

[0028] Figure 1 is the method flowchart in the application; DETAILED DESCRIPTION

[0029] In order to make the technical means, creative features, purposes and effects realized by the application easy to understand, the application is further described below in combination with specific embodiments.

[0030] Embodiment one:

[0031] An oil liquid particle image recognition and analysis method based on reinforcement learning, comprising the following steps:

[0032] S1: obtaining an oil liquid particle image by an infrared acquisition method, selecting a transistor as an oil liquid flow channel, and allowing the oil liquid to flow through the transistor; selecting an infrared light source as a detection light, and irradiating the transistor with the infrared light; selecting an infrared imaging assembly as an imaging system, and locating the transistor between the infrared light source and the infrared imaging assembly;

[0033] S2: pre-processing the infrared image to obtain a standardized image, and the pre-processing at least includes gray scale, adaptive contrast enhancement and edge sharpening to highlight the gray scale difference between the particles and the background and the outline boundary of the single particles in the agglomerate;

[0034] S3: In the process of acquiring oil particles, the average particle size threshold in the oil is calculated in real time according to the infrared data collected in step S1, and the average particle size threshold is the statistical average of the equivalent diameters of single particles;

[0035] S4: A reinforcement learning model is constructed, the reinforcement learning model takes the connected region in the standardized image as the observation object, is used for distinguishing single particles and large particle substances, and according to the average value in S3, the large particle substances are averaged as single particles;

[0036] S5: The reinforcement learning model is trained by using a labeled data set, the labeled data set contains oil infrared images of different agglomeration degrees and corresponding labeled information, and the labeled information at least includes the position of the large particle substance, the number of single particles in the large particle substance, the actual size of each single particle and the average particle size threshold;

[0037] S6: The infrared image of the oil to be detected is input into the trained reinforcement learning model, when the large particle substance adhered together irradiated by infrared light is detected, the model calculates how many small particles are bonded together to form the large particle substance according to the size of the large particle substance obtained and the average particle size threshold calculated in step S3, and outputs the actual counting result of the particles in the oil to be detected and the actual size distribution of each single particle;

[0038] The principle is as follows;

[0039] The transistor is selected as an oil flow channel to ensure stable flow of oil through the detection area; the infrared light source and the infrared imaging assembly are respectively arranged on both sides of the oil pipe to form a straight light path of "light source-transistor-detecting device", ensuring the signal integrity of the infrared light when penetrating the oil and irradiating the particles in the oil, and the scattering signal of the particles to the infrared light is captured by the infrared imaging assembly to generate an infrared image containing single particles and agglomerates (large particle substances formed by the adhesion of multiple particles), providing original data for subsequent analysis; then the infrared image in S2 is preprocessed to obtain a standardized image, and the preprocessing at least includes grayscale, adaptive contrast enhancement and edge sharpening to highlight the grayscale difference between the particles and the background and the outline boundary of the single particles inside the agglomerates; in order to solve the problem that the infrared image is easily affected by noise and the grayscale difference between the particles and the background is small, the multi-channel infrared image is converted into a single-channel grayscale image to unify the brightness dimension and reduce the residual information; the CLAHE algorithm is used to dynamically adjust the contrast of the local area of the image, amplify the grayscale difference between the particles (especially the small particles and the edge of the agglomerates) and the oil background, and avoid the "submersion" of the particles caused by low contrast; the Laplace operator is used to calculate the pixel grayscale gradient to strengthen the outline boundary of the single particles inside the agglomerates, providing a clear feature basis for subsequent "distinguishing single particles and agglomerates"; at the same time, the arithmetic mean value of the equivalent diameters of the single particles is dynamically calculated based on the infrared data collected in S1 as the single particle size reference; then the connected regions (regions with continuous grayscale in the image, single particles correspond to independent connected regions, and agglomerates correspond to large-size connected regions) in the preprocessed standardized image are taken as the observation object, so that the reinforcement learning model is constructed to intelligently distinguish the single particle connected region and the large particle substance connected region; and the average particle size threshold value calculated in S3 is taken as the reference to split the size of the large particle substance according to the single particle size, convert the number of single particles inside the agglomerates, and then obtain the number of small particles in the large agglomerated particles; the particle counting error is avoided; and before the formal detection of the oil particles, an oil infrared image labeling data set containing different agglomeration degrees is constructed; such as the distribution of particles in the process of oil viscosity from bottom to high, the situation of large particle agglomeration from less to more, and the distribution of particle types, to ensure that the model adapts to diversified scenarios; during the process, the pixel coordinates of the large particle substances, the number of internal single particles (verified by artificial microscope counting and ultrasonic dispersion), the actual size of each single particle (calibrated by a laser particle size instrument), and the average particle size value under the corresponding working condition are accurately labeled to provide data support for model training, thereby supplementing the database of the reinforcement learning model, and facilitating subsequent formal identification and monitoring of the oil;

[0040] In S6, the data of the reinforcement learning model is used for real-time monitoring of the oil liquid to be monitored. After the trained reinforcement learning model receives the infrared image to be detected, it automatically identifies the single particles and large particle substances in the image, combines the large particle substances with their projection areas, and calculates the average particle size threshold in S3. The calculation method for the number of single particles is the area of large particles divided by the area of small particles, which determines how many small particles make up a large particle, thereby avoiding the counting error of small particle size. After calculation, the results are output, and the output content is the actual count of particles in the output oil liquid (the sum of the number of single particles and the number of single particles after the agglomerate is split), which provides data support for equipment fault diagnosis. In this way, the problem of low count caused by particle agglomeration in high-viscosity oil is avoided.

[0041] Example Two:

[0042] In S1, three infrared light sources distributed above and below are selected as detection light, among which the middle light source horizontally irradiates the transistor, the upper light source obliquely downward irradiates the transistor, and the lower light source obliquely upward irradiates the transistor. The middle light source has an angle of 0° with the axis of the transistor, the upper light source has an angle of 30°-45° with the axis, and the lower light source has an angle of -30° to -45° with the axis. The three light sources are synchronously triggered.

[0043] The cross section of the transistor is a flat rectangle, and the transistor is made of fused quartz glass material.

[0044] The principle is as follows.

[0045] On the basis of the above-mentioned infrared light source irradiation, three infrared light sources distributed above and below are selected, among which the middle light source has an angle of 0° with the axis of the transistor, which is horizontal irradiation, covering the middle area of the transistor, and capturing the particle signal in the middle area of the oil liquid. The upper light source has an angle of 30° to 45° with the axis, which is obliquely downward irradiation, and the light obliquely penetrates the upper half area of the transistor, covering the top of the particle which is difficult to irradiate by the middle light source. It is used to obtain the upper blind area of the particle. The lower light source has an angle of -30° to -45° with the axis, which is obliquely upward irradiation, and the light obliquely penetrates the lower half area of the transistor, covering the bottom of the particle which is difficult to irradiate by the middle light source, and is used to obtain the lower blind area of the particle. The three light sources are synchronously triggered by the controller to ensure that the three-view infrared images of the particles in the horizontal direction and the up-down direction are obtained at the same time. For example, when three 5um particles are arranged in a straight line along the horizontal direction, the middle light source can only capture the signal of one large particle, while the upper and lower inclined light sources can capture the side scattering signals of each particle, avoiding the problem of only one particle signal being obtained when multiple particles are arranged in a straight line.

[0046] In the above process, by making the transistor of fused quartz glass material, the absorption and reflection loss of infrared light on the tube wall can be reduced, ensuring that the particle scattering signal is completely transmitted to the imaging assembly; the surface of the fused quartz glass material is not easy to adsorb oil film, avoiding the image background noise caused by oil stains on the tube wall; at the same time, the transistor is a rectangular tube with a flat rectangular cross section; compared with the circular cross section, the flat channel design can reduce the thickness of the oil in the vertical direction and reduce the probability of the particles presenting a straight line in the horizontal direction; and the parallel side walls of the rectangular tube ensure that the central horizontal light source is vertically incident, and the reflection angle of the light of the upper and lower inclined light sources on the side wall is stable, avoiding the problem of light path disorder caused by irregular tube wall shape.

[0047] Embodiment three:

[0048] The real-time calculation logic of the average particle size threshold in step S3 is that the equivalent diameter of the region determined as a single particle in the collected infrared image is measured, the first 5% maximum value and the last 5% minimum value in the measured value are removed, and the arithmetic mean value of the remaining measured value is taken as the average particle size threshold.

[0049] The single particle number calculation method of the large particle substance in step S6 is that the three-dimensional contour volume V of the large particle substance is extracted through the reinforcement learning model, and the single particle volume v0 corresponding to the average particle size threshold is combined, v0=πd 3 / 6, d is the average particle size threshold, and the single particle number n=round(V / v0k) is calculated, where k is a shape correction coefficient, and the value range is 0.8-1.2, which is dynamically adjusted according to the irregularity of the large particle substance;

[0050] The principle is as follows:

[0051] In S3, the equivalent diameter of all regions determined as single particles in the collected infrared image is measured, and after measuring the equivalent diameter, the first 5% maximum value (mostly impurity particles, oil additive precipitates) and the last 5% minimum value (mostly "pseudo-particles" caused by image noise) are removed to avoid the interference of extreme values on the threshold; if 1 impurity particle of 20um is mixed in the oil, the average particle size threshold will rise from 5um to 6.5um if not removed, resulting in less single particle number when the agglomerate is split; and the threshold is updated once every 100 frames of images to avoid the hysteresis of the fixed threshold;

[0052] Through the above embodiment two, the three infrared light sources of upper, middle and lower distribution, according to the three direction synchronous irradiation, the infrared imaging assembly receives the irradiation result, and the three-dimensional model reconstruction is carried out on the irradiated particles, relying on the three-dimensional reconstruction of the infrared imaging assembly, combining the multi-view image of the three light sources, the three-dimensional contour model of the large particle material is generated through the SIFT feature matching, and the error of the actual volume compared with the two-dimensional projection area conversion volume is calculated, wherein the single particle volume reference calculation is based on the average particle size threshold d of S3, assuming that the single particle is a regular sphere, the sphere volume formula v0=πd 3 / 6 is used to calculate the theoretical volume of single particle (such as d=5um, v0≈65.45um 3 ); At the same time, considering that the agglomerate is irregular in shape, the shape correction coefficient k (k takes the value of 0.8-1.2) is introduced; At the same time, if there are gaps in the loose agglomerate, the actual number of single particles is less than the theoretical value of V0, k takes 0.8-0.9; If the actual agglomerate is relatively tight and there is no obvious gap inside, k takes 1.0-1.2; Wherein the number of single particles n=round(V / V0×k), the round function ensures that the result is an integer; For example, if 8 particles form a loose agglomerate, the three-dimensional contour volume is V=523.6um, the average particle size threshold is d=5um, the single particle volume is v0=65.45um, the value of k in this embodiment is 0.8, n=round(523.6 / 65.45×0.8)=round (6.4)=6, if n is 6, the error of 2 small particles is formed with the actual 8 loose agglomerates, and the error rate is 25%;

[0053] Further, on the basis of this embodiment, for the agglomerated particles with voids, the void ratio in the loose agglomerate is added on the basis of the agglomerated particles, V=8×v0×(1+void ratio), if the void ratio is 30%, then V=8×65.45×1.3=680.68um 3 , similarly, k=0.8, n=round(680.68 / 65.45×0.8)=round(8.32)=8; If n is 8, the calculated result is relatively small compared with the actual result, and 2 small particles are compared; In the actual monitoring, the monitoring result of small particles is more accurate.

[0054] Embodiment four:

[0055] An oil particle image recognition and analysis system based on reinforcement learning is used to realize the oil particle image recognition and analysis method based on reinforcement learning, and the system comprises:

[0056] The infrared acquisition module comprises a flat rectangular straight channel transistor of fused quartz glass material, three infrared light sources and an infrared imaging assembly; the transistor is an oil flow channel, and the cross section is a flat rectangle; the three infrared light sources are distributed in the upper, middle and lower positions, the middle light source horizontally irradiates the transistor, and the upper and lower light sources obliquely irradiate the transistor downward and upward respectively; the infrared imaging assembly is arranged opposite to the transistor, and is used for receiving infrared light transmitted through the transistor and generating an infrared image of oil particle;

[0057] The preprocessing module is connected with the infrared acquisition module, and is used for performing grayscale, contrast enhancement and edge sharpening processing on the infrared image to highlight the single particle outline inside the large particle material;

[0058] The particle size threshold calculation module is connected with the infrared acquisition module, and is used for calculating and updating the average particle size threshold of the oil particle according to the real-time collected infrared data;

[0059] The reinforcement learning model module is connected with the preprocessing module and the particle size threshold calculation module respectively, and is used for receiving the standardized image and the average particle size threshold, and outputting the single particle splitting result of the large particle material, the actual count and the size distribution;

[0060] The result output module is connected with the reinforcement learning model module, and is used for displaying the particle count, the size distribution and the agglomeration ratio of the large particle material;

[0061] The infrared imaging assembly adopts a refrigeration type infrared focal plane array detector, the frame rate is matched with the light source acquisition frequency, and the imaging assembly further internally arranges a three-dimensional reconstruction unit, which is used for generating a three-dimensional outline model of the particle in combination with the image data of the three light sources;

[0062] The reinforcement learning model module comprises a multi-light-source feature fusion unit, which is used for weightedly fusing the particle grayscale features and edge features under the irradiation of the three light sources, and the weight coefficients are dynamically adjusted according to the light source angle and the particle position, wherein the feature weight of the corresponding light source is increased by 1.2-1.5 times for the overlapping area of the large particle material;

[0063] The principle is as follows;

[0064] In the principles of the above embodiment one to embodiment three, by setting the infrared acquisition module and the flat rectangular transistor; and the three infrared light sources distributed in the upper, middle and lower, the infrared imaging assembly, the particle infrared image without shielding and high signal-to-noise ratio provides high-quality raw data for subsequent analysis; and through the pretreatment module, such as eliminating image sound in S2, strengthening the characteristics of the agglomerate, converting the fuzzy image into a standardized feature image to reduce the difficulty of model analysis; through the particle size threshold calculation module, in S3, the single particle size reference is dynamically generated, avoiding the splitting error caused by the fixed threshold due to the change of working condition, by setting the reinforcement learning model module, in S4, multi-dimensional data is fused to realize the accurate identification and splitting of the agglomerate; by setting the result output module, the abstract data is converted into intuitive information to meet the actual needs of the operation and maintenance personnel to quickly judge the oil pollution state.

[0065] The infrared imaging assembly adopts a refrigeration type outer focal plane array detector, the noise equivalent temperature difference of the traditional uncooled detector can easily mask the weak scattering signal of the small particle; the refrigeration type detector reduces the temperature to-75 DEG C through semiconductor refrigeration, K, the noise is reduced, even the scattering signal of the small particle can be clearly imaged, and then the particle trailing error caused by imaging blur is avoided to be misjudged as the agglomerate.

[0066] And by embedding a three-dimensional reconstruction unit in the infrared imaging assembly, compared with the traditional two-dimensional imaging, only the projection area of the particle can be obtained, the flat agglomerate and the thick agglomerate cannot be distinguished, and the volume calculation error is large; the three-dimensional reconstruction unit extracts the particle feature points such as the convex and concave of the agglomerate edge of the three light source image through the SIFT algorithm, calculates the three-dimensional coordinates to generate the contour model, so that the volume measurement error is reduced, and accurate volume data is provided for subsequent agglomerate splitting.

[0067] The above shows and describes the basic principles, main features and advantages of the present application; those skilled in the art should understand that the present application is not limited to the above embodiments, the above embodiments and the description in the specification are only to illustrate the principles of the present application, without departing from the spirit and scope of the present application, the present application can have various changes and improvements, and these changes and improvements all fall within the scope of the claimed present application. The scope of protection of the present application is defined by the appended claims and their equivalents.

Claims

1. A method for oil particle image recognition analysis based on reinforcement learning, characterized in that, The method comprises the following steps: S1: obtaining an oil particle image by infrared acquisition, selecting a transistor as an oil flow channel, and allowing oil to flow through the transistor; using an infrared light source as a detection light, and irradiating the transistor with the infrared light; Selecting an infrared imaging assembly as an imaging system, and the transistor is located between the infrared light source and the infrared imaging assembly; S2: preprocessing the infrared image to obtain a standardized image, the preprocessing at least including grayscale, adaptive contrast enhancement and edge sharpening, so as to highlight the gray difference between the particles and the background and the outline boundary of the single particles in the agglomerates; S3: in the process of obtaining the oil particle, calculating the average particle size threshold of the particles in the oil in real time according to the infrared data collected in step S1, and the average particle size threshold is the statistical average of the equivalent diameters of the single particles; S4: constructing a reinforcement learning model, the reinforcement learning model taking the connected regions in the standardized image as the observation object, and being used for distinguishing the single particles and the large particle substances, and according to the average value in S3, the large particle substances are averaged as single particles; S5: training the reinforcement learning model by using a labeled data set, the labeled data set containing oil infrared images of different agglomeration degrees and corresponding labeled information, and the labeled information at least including the position of the large particle substances, the number of single particles in the large particle substances, the actual size of each single particle and the average particle size threshold; S6: inputting the infrared image of the oil to be detected into the trained reinforcement learning model, when the large particle substances adhered together irradiated by the infrared light are detected, the model calculates how many small particles are combined to form the large particle substances according to the size of the large particle substances and the average particle size threshold calculated in step S3, and outputs the actual counting result of the particles in the oil to be detected and the actual size distribution of each single particle.

2. The oil particle image recognition analysis method based on reinforcement learning according to claim 1, characterized in that: In S1, three infrared light sources distributed upward and downward are selected as the detection light, wherein the middle light source horizontally irradiates the transistor, the upper light source obliquely downward irradiates the transistor, and the lower light source obliquely upward irradiates the transistor, wherein the middle light source and the axis of the transistor have an angle of 0°, the upper light source and the axis have an angle of 30° to 45°, and the lower light source and the axis have an angle of -30° to -45°, and the three light sources are synchronously triggered.

3. The oil particle image recognition analysis method based on reinforcement learning according to claim 2, characterized in that: The cross section of the transistor is a flat rectangle, and the material of the transistor is fused quartz glass.

4. The oil particle image recognition analysis method based on reinforcement learning according to claim 3, characterized in that: In step S3, the real-time calculation logic of the average particle size threshold is that the equivalent diameters of all regions determined as single particles in the collected infrared image are measured, the first 5% maximum value and the last 5% minimum value in the measured values are removed, and the arithmetic average value of the remaining measured values is taken as the average particle size threshold.

5. The oil particle image recognition analysis method based on reinforcement learning according to claim 4, characterized in that: The single particle number calculation method of the large particle material in step S6 is as follows: the three-dimensional contour volume V of the large particle material is extracted through the reinforcement learning model, and the single particle volume V0 corresponding to the average particle size threshold is combined, V0=πd 3 / 6, d is the average particle size threshold, the single particle number n=round(V / V0K) is calculated, wherein k is a shape correction coefficient, the value range is 0.8-1.2, and the irregularity of the large particle material is dynamically adjusted.

6. A system for oil particle image recognition and analysis based on reinforcement learning, which is used to implement the method for oil particle image recognition and analysis based on reinforcement learning in any one of claims 1-5; characterized in that, The system comprises: An infrared acquisition module comprising a flat rectangular straight channel transistor of fused quartz glass material, three infrared light sources and an infrared imaging assembly; the transistor is an oil flow channel, and the cross section is a flat rectangle; the three infrared light sources are distributed upward and downward, the middle light source horizontally irradiates the transistor, and the upper and lower light sources obliquely downward and upward irradiate the transistor respectively; the infrared imaging assembly is arranged opposite to the transistor, and is used for receiving the infrared light transmitted through the transistor and generating an oil particle infrared image; The preprocessing module is connected with the infrared acquisition module, and is used for performing grayscale, contrast enhancement and edge sharpening processing on the infrared image, so as to highlight the single particle outline inside the large particle matter; The particle size threshold calculation module is connected with the infrared acquisition module, and is used for calculating and updating the average particle size threshold of the oil liquid particle according to the real-time collected infrared data; The reinforcement learning model module is connected with the preprocessing module and the particle size threshold calculation module respectively, and is used for receiving the standardized image and the average particle size threshold, and outputting the single particle splitting result, the actual counting and the size distribution of the large particle matter; The result output module is connected with the reinforcement learning model module, and is used for displaying the particle counting, the size distribution and the agglomeration proportion of the large particle matter.

7. The oil particle image recognition and analysis system based on reinforcement learning according to claim 6, wherein: The infrared imaging assembly adopts a refrigeration type infrared focal plane array detector, the frame rate is matched with the light source acquisition frequency, and the imaging assembly further internally arranges a three-dimensional reconstruction unit, which is used for generating a three-dimensional outline model of the particle in combination with the image data of the three light sources.

8. The oil particle image recognition and analysis system based on reinforcement learning according to claim 6, wherein: The reinforcement learning model module comprises a multi-light-source feature fusion unit, which is used for performing weighted fusion on the particle grayscale features and the edge features under the irradiation of the three light sources, and the weight coefficients are dynamically adjusted according to the light source angle and the particle position, wherein the feature weight of the corresponding light source is increased by 1.2-1.5 times for the overlapping area of the large particle matter.

Images