Image sensor for online detection of oil quality

Abstract

The embodiment of the application provides an image sensor for online detection of oil quality, comprising a sensor body, a multi-view imaging unit, an acoustic wave separation unit, a dynamic focusing unit and an analysis unit. The sensor body comprises a multi-prism frustum-shaped multi-view observation chamber. The multi-view imaging unit acquires a top view image and a side view angle image of abrasive particles through a main imaging assembly and an auxiliary imaging assembly respectively. The acoustic wave separation unit comprises an ultrasonic wave transmitter array, a phase control circuit and an acoustic wave focusing lens, and implements physical separation on the gathered abrasive particles by generating focused acoustic waves. The dynamic focusing unit comprises a liquid lens assembly and a focal length control circuit, and realizes clear imaging of different depth levels by adjusting the optical focal length of the liquid lens. The analysis unit comprises a gathering detection network, a separation planning network and a feature extraction network, and is used for identifying abrasive particle gathering, generating a separation scheme and extracting abrasive particle features. The application solves the problem of inaccurate abrasive particle counting.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of oil quality detection, and particularly relates to an image sensor for online detection of oil quality. BACKGROUND

[0002] In the field of condition monitoring and fault diagnosis of modern industrial equipment, online detection of oil quality plays an important role. By analyzing the characteristic parameters such as the number, size, shape and material of wear particles in oil, the wear state of equipment can be effectively evaluated, and key information can be provided for equipment maintenance and fault prediction. At present, the existing online detection image sensor of oil quality adopts a technical route combining magnetic adsorption and image acquisition, as shown in FIG. 1. The ferromagnetic wear particles in the oil are adsorbed onto the surface of a transparent protective plate by a magnetic adsorption assembly, and then a two-dimensional image of the wear particles is acquired by an image acquisition assembly. Finally, the wear particles are identified and analyzed by an image processing unit. Figure 1

[0003] However, in actual industrial application environment, the wear particles in the oil are not always dispersed alone, but often appear in the phenomenon of aggregation, overlap and mutual occlusion. When multiple wear particles are simultaneously adsorbed onto the surface of the transparent protective plate by the magnetic field, due to the unevenness of the magnetic field distribution and the interaction force between the wear particles, the wear particles are easy to form a cluster-like aggregation form. This aggregation phenomenon is more obvious in small-sized wear particles, because small wear particles have a larger surface area to volume ratio and higher surface energy, and are more likely to adhere to each other.

[0004] When the wear particles aggregate, the existing image segmentation and identification method faces great difficulties. The traditional edge detection and contour extraction algorithm is difficult to accurately identify the independent boundaries of each wear particle in the aggregation cluster, resulting in inaccurate wear particle counting results. At the same time, due to the partial occlusion or deformation of the real topographic features of the wear particles in the aggregation state, the key feature parameters such as the shape, size and surface texture of the single wear particle cannot be accurately extracted, which affects the accurate judgment of the wear state of the equipment.

[0005] In addition, when the wear particles overlap in the thickness direction, the wear particles in the lower layer are completely or partially occluded by the wear particles in the upper layer. The traditional two-dimensional image acquisition method can only obtain the information of the surface visible wear particles, and cannot obtain the complete topographic features of the occluded wear particles, further exacerbating the inaccuracy of the detection results. This information loss problem is particularly prominent in high-concentration oil samples, seriously affecting the reliability and practicability of the detection system.

[0006] Therefore, it is necessary to provide a new technical scheme to solve the above technical problems. SUMMARY

[0007] ​The embodiment of the present application aims to provide an image sensor for online detection of oil quality, aiming to solve the technical problem of inaccurate particle counting existing in the prior art.

[0008] The embodiment of the present application provides an image sensor for online detection of oil quality, comprising a sensor body, wherein the sensor body comprises a multi-view observation chamber, the multi-view observation chamber is a multi-truncated pyramid, the multi-view observation chamber is provided with a multi-layer optical window assembly, the multi-layer optical window assembly comprises a main observation window located at the center of the top of the chamber and a plurality of auxiliary observation windows radially distributed around the main observation window, and the plurality of auxiliary observation windows are all inclined to the inside of the chamber; a multi-view imaging unit comprising a main imaging assembly and an auxiliary imaging assembly, the main imaging assembly is used for acquiring a top view image of the particles through the main observation window, and the auxiliary imaging assembly is used for acquiring a side view angle image of the particles through the auxiliary observation window; an acoustic wave separation unit comprising an ultrasonic wave transmitter array, a phase control circuit and an acoustic wave focusing lens, the ultrasonic wave transmitter array is located at the bottom of the detection chamber, the phase control circuit is used for calculating the phase delay value of each transducer according to the target focusing position and generating a multi-channel driving signal with a specific phase relationship, and the acoustic wave focusing lens is used for performing a phase modulation operation on the incident acoustic wave; a dynamic focusing unit comprising a liquid lens assembly and a focal length control circuit, the liquid lens assembly comprises a double-layer liquid medium of conductive liquid and insulating oil, and a transparent electrode surrounds the lens cavity to form an electric field control area, and the focal length control circuit is used for applying a variable voltage to the transparent electrode to generate an electric field intensity change to adjust the optical focal length of the liquid lens; and an analysis unit comprising an aggregation detection network, a separation planning network and a feature extraction network, the aggregation detection network is used for performing feature mapping extraction operation on the input image to identify the aggregation phenomenon of the particles, the separation planning network is used for generating an execution scheme of acoustic wave separation according to the aggregation detection result, and the feature extraction network is used for performing deep feature analysis on the separated independent particles.

[0009] In the above image sensor, the circumferential angle interval between every two adjacent auxiliary windows is 45 degrees, and the included angle between the window plane of each auxiliary observation window and the top surface of the chamber is 30 degrees.

[0010] In the above image sensor, the multi-view imaging unit further comprises a field of view calculation unit and a stereo vision processing module, the field of view calculation unit is used for calculating the field of view boundary data of each window, and the stereo vision processing module is used for distributing the image data to the corresponding processing channel according to the view angle identifier.

[0011] In the image sensor, the acoustic wave separation unit further comprises an amplitude modulator and an acoustic field distribution controller, the amplitude modulator is used to calculate the required acoustic radiation force size according to the abrasive particle aggregation information and convert it into the transducer driving power requirement, and the acoustic field distribution controller is used to control the acoustic wave focus to move in the detection area according to the predetermined path.

[0012] In the image sensor, the dynamic focusing unit further comprises a focus position detector and a multi-layer image acquisition module, the focus position detector is used to emit a reference laser beam to the detection area and calculate the focus distance according to the reflected laser, and the multi-layer image acquisition module is used to continuously acquire image data at different focus positions and construct a three-dimensional image data stack.

[0013] In the image sensor, a spectrum analysis unit is further included, the spectrum analysis unit comprises a multi-wavelength LED array and a spectrum filter group, the multi-wavelength LED array is used to activate light sources of different wavelengths in turn according to a time sequence configuration table, and the spectrum filter group is used to receive a wavelength synchronization signal and rotate to a predetermined angle to realize wavelength switching.

[0014] In the image sensor, the aggregation detection network comprises a spatial density analysis module, a shape abnormality evaluation module and a boundary continuity detection module, the spatial density analysis module is used to calculate the density value by counting the number of abrasive particle feature points in the local area, the shape abnormality evaluation module is used to calculate the geometric feature parameters of the abrasive particle contour and compare them with the standard template, and the boundary continuity detection module is used to analyze the fracture points and curvature changes of the contour to judge the tightness of the aggregation.

[0015] In the image sensor, the separation planning network comprises a frequency selection module, a power distribution calculation module and an action time planning module, the frequency selection module is used to calculate the optimal acoustic wave frequency according to the abrasive particle size distribution, the power distribution calculation module is used to determine the required acoustic wave power level according to the aggregation degree, and the action time planning module is used to calculate the duration of the acoustic wave action and divide it into multiple short pulse periods.

[0016] In the image sensor, the feature extraction network comprises a boundary detection convolution layer, a contour refinement module and a multi-scale feature fusion module, the boundary detection convolution layer is used to detect the irregular shape contour of the abrasive particle using a deformable convolution kernel, the contour refinement module is used to perform sub-pixel level processing on the boundary contour, and the multi-scale feature fusion module is used to process image data of different resolutions and extract detail features and shape features.

[0017] In the image sensor, an adaptive learning unit is further included, the adaptive learning unit includes an aggregation pattern recognition module, a parameter optimization module and a knowledge base updating module, the aggregation pattern recognition module is used for establishing a feature data index table of the abrasive particle aggregation pattern, the parameter optimization module is used for dynamically adjusting the sound wave parameters according to the separation effect, and the knowledge base updating module is used for identifying high-quality processing cases and updating the database.

[0018] The image sensor for online detection of oil quality provided in the application effectively solves the technical problem of inaccurate abrasive particle counting in the prior art through the cooperative matching of a multi-view observation chamber, a multi-view imaging unit, a sound wave separation unit, a dynamic focusing unit and an analysis unit.

[0019] The image sensor adopts a multi-prism platform-shaped multi-view observation chamber matched with a multi-layer optical window assembly, a main observation window obtains a complete overhead image of the abrasive particle, and multiple auxiliary observation windows observe the abrasive particle from different side angles, so that the three-dimensional appearance information of the abrasive particle can be obtained from multiple dimensions. This multi-view imaging mode overcomes the limitation of traditional single-view observation. When the abrasive particles are mutually occluded, the occluded area in some view is still visible in other views. Through information fusion of the multi-view images, the complete contour of the occluded abrasive particle can be reconstructed, so that the overall appearance characteristics of the abrasive particle can be obtained. The stereo vision processing module of the multi-view imaging unit can determine the position coordinates of the abrasive particle in the three-dimensional space by calculating the parallax values of the feature points in different view images, so that the spatial positioning and hierarchical recognition of the overlapping abrasive particles can be realized, and the problem that the traditional two-dimensional image cannot obtain depth information can be solved.

[0020] The sound wave separation unit generates a controllable sound wave phase distribution through an ultrasonic wave emitter array. A phase control circuit calculates the phase delay value of each transducer according to the abrasive particle aggregation position, so that the sound wave energy forms constructive interference at the target position, realizing the focusing effect of the sound wave. The sound wave focusing lens performs a phase modulation operation on the incident sound wave, converts the plane sound wave into a focused sound wave with a specific wave front shape, and generates concentrated acoustic radiation force in the abrasive particle aggregation area. When the ultrasonic wave encounters the abrasive particle, the acoustic scattering phenomenon occurs, and the scattered sound wave and the incident sound wave form an uneven sound pressure distribution around the abrasive particle, generating a net radiation force on the abrasive particle. This radiation force can overcome the adhesion between the abrasive particles, realizing the physical separation of the aggregated abrasive particles. The amplitude modulator adjusts the driving power of the transducer according to the bonding strength between the abrasive particles, ensuring that sufficient separation force is generated. The sound field distribution controller controls the movement trajectory of the sound wave focus in the detection area, realizing selective processing of different aggregation areas, thereby effectively solving the problem of inaccurate counting caused by abrasive particle aggregation.

[0021] The dynamic focusing unit realizes clear imaging of abrasive particles at different depth levels through a liquid lens assembly. The liquid lens includes a double-layer medium of conductive liquid and insulating oil. A focal length control circuit applies a variable voltage to the transparent electrode to change the electric field strength. The change in electric field strength adjusts the wetting properties of the conductive liquid surface, thereby changing the geometry of the liquid interface and the optical focal length of the lens. By continuously adjusting the focal length, the multi-layer image acquisition module can collect clear abrasive particle images at different depth levels, constructing a three-dimensional image data stack containing multiple depth information. The focal point position detector realizes dynamic correction of the focal point position through the reflection of the laser beam, ensuring the best focusing effect at each depth level. This dynamic focusing capability enables each individual in the overlapping abrasive particles to be clearly imaged at the corresponding depth level. The overlapping abrasive particle recognition module analyzes the pixel intensity distribution at each depth level to identify abrasive particle individuals at different depths and reconstruct the complete profile of the obscured abrasive particles, effectively solving the information loss problem in overlapping abrasive particle detection.

[0022] The aggregation detection network of the analysis unit accurately identifies abrasive particle aggregation phenomena through deep learning processing. The spatial density analysis module calculates the density distribution of feature points within a local region. The shape abnormality evaluation module calculates the degree to which the abrasive particle profile deviates from the standard shape. The boundary continuity detection module analyzes the breaking and curvature changes of the profile. The coordinated action of the three modules can comprehensively evaluate the aggregation type and tightness of the abrasive particles. The separation planning network generates a targeted separation scheme based on the aggregation detection results. The frequency selection module determines the optimal acoustic frequency based on the size distribution of the abrasive particles. The power allocation calculation module calculates the required acoustic power based on the aggregation degree. The action time planning module divides the continuous action time into multiple pulse periods to avoid over-action, ensuring that the separation parameters are highly matched to the actual situation. The feature extraction network performs deep feature analysis on the separated independent abrasive particles. The boundary detection convolution layer uses a deformable convolution kernel to adapt to the irregular shape of the abrasive particles. The profile refinement module improves the accuracy of boundary recognition through sub-pixel level processing. The multi-scale feature fusion module simultaneously extracts detailed features and overall features. BRIEF DESCRIPTION OF DRAWINGS

[0023] Figure 1 is a schematic diagram of a conventional image sensor for online detection of oil quality.

[0024] Figure 2 is a schematic diagram of an image sensor for online detection of oil quality provided by an embodiment of the present application.

[0025] Figure 3 is a block diagram of an image sensor for online detection of oil quality provided by an embodiment of the present application.

[0026] Figure 4 is Figure 3A block diagram of a multi-view imaging unit of an image sensor for online detection of oil quality is shown.

[0027] Figure 5 is Figure 3 A block diagram of a spectral analysis unit of an image sensor for online detection of oil quality is shown.

[0028] Figure 6 is Figure 3 A block diagram of a dynamic focusing unit of an image sensor for online detection of oil quality is shown.

[0029] Figure 7 is Figure 3 A block diagram of a sound wave separation unit of an image sensor for online detection of oil quality is shown.

[0030] Figure 8 is Figure 3 A block diagram of an analysis unit of an image sensor for online detection of oil quality is shown.

[0031] Figure 9 is Figure 3 A block diagram of a real-time monitoring feedback unit of an image sensor for online detection of oil quality is shown.

[0032] Figure 10 is Figure 3 A block diagram of an adaptive learning unit of an image sensor for online detection of oil quality is shown. DETAILED DESCRIPTION

[0033] The specific embodiments of the present application will be described in detail below with reference to the accompanying drawings.

[0034] The terms "first", "second", and similar terms do not denote any order, quantity, or importance, but are used to distinguish different technical features. The term "multiple" and similar terms mean two or more, unless otherwise explicitly limited.

[0035] The embodiments of the present application can be combined with each other.

[0036] As Figure 1As shown, the existing image sensor for online detection of oil quality adsorbs ferromagnetic wear particles to the lower surface of the transparent protective plate through the magnetic adsorption assembly, then obtains the particle image through the image acquisition assembly, and then identifies and analyzes the particles through the image processing unit and feature analysis unit in the processing circuit. However, in actual industrial applications, wear particles in oil often do not exist in isolation, but often appear to be aggregated, overlapped and mutually occluded. When multiple wear particles are simultaneously adsorbed to the surface of the transparent protective plate by the magnetic field, due to the unevenness of the magnetic field distribution and the interaction force between the wear particles, the wear particles are prone to form cluster aggregation. This aggregation phenomenon is more pronounced in small-sized wear particles, because the surface area to volume ratio of small wear particles is larger, and the surface energy is higher, so they are more likely to adhere to each other. When wear particles aggregate, the traditional image segmentation method cannot accurately identify the individual wear particle boundaries in the aggregated cluster, resulting in inaccurate wear particle counting. At the same time, due to the occlusion or deformation of the morphology features of the aggregated wear particles, the true shape, size and surface texture features of individual wear particles cannot be accurately extracted, affecting the judgment of the equipment wear state. In addition, when wear particles overlap in the thickness direction, the lower wear particles are completely or partially occluded by the upper wear particles, and the traditional two-dimensional image acquisition method cannot obtain the complete information of the occluded wear particles, further exacerbating the inaccuracy of the detection results.

[0037] To solve the above technical problems, the image sensor for online detection of oil quality is constructed on the basis of the original image sensor, as shown in Figure 3 The image sensor realizes effective separation and accurate detection of aggregated wear particles by integrating a multi-view imaging unit, a spectral analysis unit, a dynamic focusing unit, a sound wave separation unit and an analysis unit.

[0038] As shown in Figure 2 The sensor body of the image sensor is improved to a multi-view observation chamber on the basis of the original detection chamber, and the chamber adopts a stereomultifaceted configuration, and the overall shape of the chamber is expanded from a simple cylindrical shape to a multi-prism frustum shape. The chamber is provided with a plurality of optical window assemblies, which include a main observation window located at the top center of the chamber and a plurality of (for example, eight) auxiliary observation windows distributed radially around the main observation window. The main observation window maintains a vertical downward observation angle, and the window plane is parallel to the bottom surface of the detection chamber, which is used to obtain a complete overhead image of the wear particles, and the effective observation area of the window covers the entire magnetic adsorption area.

[0039] Eight auxiliary observation windows are arranged at eight equiangular positions around the main observation window, and the circumferential angle interval between every two adjacent auxiliary windows is 45 degrees. Each auxiliary observation window is inclined towards the inside of the chamber, and the included angle between the window plane and the top surface of the chamber is set to 30 degrees, and the normal vector of the window points to the center position of the detection area. Through this inclined configuration, each auxiliary observation window can observe the abrasive particles adsorbed on the surface of the transparent protective plate from a specific side view angle and obtain the three-dimensional topographic information of the abrasive particles.

[0040] The sensor body comprises a multi-view observation chamber, which is a multi-pyramid platform. The multi-view observation chamber is provided with a multi-layer optical window assembly, which comprises a main observation window located at the center of the top of the chamber and a plurality of auxiliary observation windows distributed radially around the main observation window. The plurality of auxiliary observation windows are arranged inclined towards the inside of the chamber.

[0041] As shown in Figure 4 , in order to ensure that there is no visual interference between the observation windows, the multi-view imaging unit further comprises a field of view calculation unit and a stereo vision processing module. The field of view calculation unit calculates the field of view boundary data of each window, and the stereo vision processing module distributes the image data to the corresponding processing channel according to the view angle identifier. The field of view calculation unit of the multi-view imaging unit calculates the field of view boundary data of each window, and performs spatial intersection operation on the field of view conical region coordinates of each window at the center position of the detection area. The field of view conical region of each window converges at the center position of the detection area, but remains independent near the window to avoid physical interference between the imaging assemblies. The window spacing baffle of the multi-view observation chamber is used to block the cross irradiation of different window illumination light sources, and to absorb excess light by light-absorbing material to convert light energy into heat energy and dissipate, so as to prevent the light intensity superposition interference of each illumination light source in the adjacent window area, thereby improving the clarity and contrast of image acquisition.

[0042] As shown in Figure 2 and Figure 4 , the multi-view imaging unit comprises a main imaging assembly and an auxiliary imaging assembly. The main imaging assembly obtains the overhead image of the abrasive particles through the main observation window, and the auxiliary imaging assembly obtains the side view angle image of the abrasive particles through the auxiliary observation window.

[0043] The trapezoidal side wall of the multi-view observation chamber is used for geometric optimization of light path propagation, and the chamber gradually expands from the bottom to the top to ensure that the optical axes of the multiple observation windows and imaging assemblies do not block each other. The inner wall of the multi-view observation chamber is coated with a low reflectivity black coating to reduce the secondary reflection of stray light and absorb incident light, reducing the reflected light intensity to a negligible level, thereby ensuring that the imaging quality is not disturbed by ambient light, improving the image clarity and contrast, and achieving high-quality acquisition of abrasive particle images.

[0044] The main imaging assembly of the multi-view imaging unit performs distributed image acquisition and data transmission operations, converts the optical signal into an electric charge signal, and converts the electric charge signal into a digital pixel value through an analog-to-digital conversion circuit, encapsulates the digital image data into a data packet, and transmits the data packet to an internal image processing module. After receiving the data packet, the multi-view imaging unit analyzes the pixel matrix of the image, performs convolution operation on the pixel matrix through edge detection processing to identify the abrasive grain boundary, calculates the gray difference value between adjacent pixels using a gradient operator, marks the pixels with a gray difference value exceeding a threshold as edge points, and connects adjacent edge points to form a continuous contour line. The multi-view imaging unit further performs chain code encoding on the contour line data, converts the coordinates of each point on the contour into a direction encoding sequence, calculates the geometric center point coordinates of the contour as the abrasive grain position, and outputs the abrasive grain position coordinates and the encoding data of the boundary contour, realizing the position positioning and shape description of a single abrasive grain, and effectively solving the deficiency of the traditional method in abrasive grain boundary identification.

[0045] The auxiliary imaging assembly of the multi-view imaging unit synchronously receives the trigger signal for image acquisition, sends a uniform clock pulse to the eight auxiliary imaging assemblies through a synchronous trigger circuit, so that the auxiliary imaging assemblies start exposure and data reading at the same time as the rising edge of the clock pulse. The multi-view imaging unit transmits the oblique view image data to the stereo vision processing module after adding a view identifier, and distributes the image data to the corresponding processing channels according to the view identifier. The multi-view imaging unit performs corner point detection operation on each channel image, calculates the eigenvalue of the gradient matrix of each pixel, and regards the pixels with eigenvalues satisfying the corner point condition as candidate feature points, and performs local extreme value screening on the candidate points to obtain stable feature points. The multi-view imaging unit calculates a description vector for each feature point, extracts the gradient direction histogram of the neighborhood around the feature point, and normalizes the histogram data into a fixed length description vector. The multi-view imaging unit calculates the similarity of the description vectors of the feature points in different view images, uses the Euclidean distance to measure the difference between the description vectors, and regards the feature point pair with the smallest distance as the matching result. The multi-view imaging unit calculates the pixel coordinate difference value according to the matching point pair, subtracts the pixel coordinates of the same feature point in different views to obtain the parallax value, and converts the parallax value into a depth distance combined with the camera intrinsic matrix, combines the depth distance with the plane coordinates of the main imaging assembly, constructs a grain position vector containing three-dimensional coordinates, and outputs a list of spatial position data of the abrasive grain, realizing the three-dimensional spatial positioning of the abrasive grain, and effectively solving the problem that the traditional two-dimensional image detection technical solution cannot obtain depth information.

[0046] The three-dimensional reconstruction module of the multi-view imaging unit receives the spatial position data of the abrasive particles, calculates the spatial distances between the abrasive particles and constructs a distance matrix, uses a clustering discrimination function to classify the abrasive particles with similar distances into clusters, and assigns a unique identification code to each cluster. The three-dimensional reconstruction module performs depth sorting on the abrasive particles in the cluster, establishes the front and rear order relationship of the abrasive particles according to the Z coordinate value, identifies the pixel area of the rear abrasive particles blocked by the front abrasive particles, and calculates the proportion of the blocked area to the total area of the abrasive particles. The three-dimensional reconstruction module calculates the sound wave action parameters according to the cluster analysis results, determines the sound wave frequency and power configuration for each cluster, encodes the parameter data into a control instruction format, and transmits the instruction data to the sound wave separation control module to realize targeted separation operation on the aggregated abrasive particles, solving the problems of inaccurate counting and difficult extraction of morphology characteristics caused by abrasive particle aggregation.

[0047] The spectral analysis unit stores a time sequence configuration table of multi-wavelength scanning, generates a time reference signal according to the configuration table, and distributes the reference signal to each LED driving circuit as an activation trigger. As shown in Figure 5 The spectral analysis unit includes a multi-wavelength LED array and a spectral filter group. The multi-wavelength LED array activates light sources of different wavelengths in sequence according to the time sequence configuration table, and the spectral filter group receives a wavelength synchronization signal and rotates to a predetermined angle to realize wavelength switching. The multi-wavelength LED array of the spectral analysis unit receives a trigger signal and performs current driving on the corresponding LED, adjusts the current size to control the light intensity output of the LED, and performs gradual control when switching wavelengths to avoid sudden changes in light intensity. The spectral filter group of the spectral analysis unit receives a wavelength synchronization signal, calculates the target angle position of the filter, drives the stepper motor to rotate the filter to a predetermined angle, and feeds back the ready state after confirming that the filter is in place through the position encoder, realizing continuous switching of wavelengths and spectral scanning, and providing a spectral basis for accurate identification of abrasive particles of different materials.

[0048] The spectral data acquisition module of the spectral analysis unit reads pixel data from the image sensor during each wavelength illumination, classifies and stores the pixel data according to the wavelength identifier, and constructs a multi-dimensional spectral data matrix containing image information of each wavelength. The spectral data acquisition module extracts the pixel values of the abrasive particle region from the spectral data matrix, performs ratio operation on the abrasive particle pixel values and the reference region pixel values, and generates a spectral feature vector containing multi-wavelength reflectivity for each abrasive particle. The spectral data acquisition module calculates the correlation coefficient of the spectral feature vector and each template by correlating the spectral feature vector with the pre-stored material spectral template, selects the template with the highest correlation coefficient as the material identification result, and outputs the material type identification of the abrasive particle, realizing accurate classification of abrasive particles of different materials and solving the problem that traditional methods cannot distinguish abrasive particles of different materials.

[0049] The spectral depth analysis module of the spectral analysis unit establishes a depth relationship model according to the attenuation characteristics of light rays of various wavelengths in the oil, calculates the effective penetration distance of light rays of different wavelengths in the oil, and establishes a corresponding relationship between depth and light intensity attenuation for each wavelength. The spectral depth analysis module analyzes the change rule of the reflected light intensity of the abrasive particles under different wavelengths, compares the light intensity attenuation degrees of various wavelengths, calculates the depth level position of the abrasive particles according to the attenuation difference, assigns the front and rear depth labels of the overlapping abrasive particles, and outputs the sorting information of the depth level of the abrasive particles, thereby realizing the depth level identification and three-dimensional spatial distribution analysis of the overlapping abrasive particles, and effectively solving the problem of information loss caused by the shielding of lower abrasive particles in the detection of overlapping abrasive particles.

[0050] As shown in Figure 6 The dynamic focusing unit includes a liquid lens assembly, a focal length control circuit, a focal point position detector, and a multi-layer image acquisition module. The liquid lens assembly of the dynamic focusing unit contains a double-layer liquid medium of conductive liquid and insulating oil, a transparent electrode surrounds the lens cavity to form an electric field control area, and a variable voltage is applied to the transparent electrode through the focal length control circuit to generate an electric field intensity change to adjust the optical focal length of the liquid lens. The change of the electric field intensity changes the wetting characteristics of the surface of the conductive liquid, and the change of the wetting characteristics directly affects the geometric shape of the interface between the conductive liquid and the insulating oil, and the change of the curvature of the interface shape adjusts the optical focal length of the liquid lens. The dynamic focusing unit stores a plurality of focal length preset value data tables, the data table contains the voltage configuration corresponding to different depth levels in the detection area, and outputs the corresponding control voltage signals in a preset scanning order, realizes accurate focusing of abrasive particles of different depth levels, and solves the problem that the traditional fixed focal length method cannot simultaneously observe multiple overlapping abrasive particles.

[0051] The focal point position detector of the dynamic focusing unit emits a reference laser beam to the detection area and calculates the focal point distance according to the reflected laser, and the multi-layer image acquisition module continuously acquires image data at different focal length positions and constructs a three-dimensional image data stack. The focal point position detector emits a reference laser beam to the detection area, the laser beam is reflected after passing through the surface of the abrasive particle and is received by a position sensitive detector, and the position offset of the reflected laser is converted into an electrical signal. By calculating the geometric relationship between the laser incidence angle and the reflection angle, the actual distance from the focal point distance to the surface of the transparent protective plate is calculated according to the triangulation formula. The focal point position detector feeds back the distance measurement data to the focal length control circuit after digital conversion, compares the deviation between the measured distance and the target distance, adjusts the voltage output according to the deviation size to realize dynamic correction of the focal point position, and ensures the best focusing effect of each depth level.

[0052] The multi-layer image acquisition module of the dynamic focusing unit sends an acquisition instruction to the image sensor after each focal length is stabilized, continuously acquires clear focused image data at different focal length positions, classifies and stores the image data of each focal length level according to depth identification, and constructs a three-dimensional image data stack containing multiple depth level information. The multi-layer image acquisition module performs gradient amplitude calculation on each layer of image in the three-dimensional image stack, counts the gradient intensity of each pixel position at different depth levels, and selects the pixel with the maximum gradient intensity as the best focus pixel at the position. The multi-layer image acquisition module extracts the best focus pixels from each depth level to compose a panoramic depth image, records the depth source information of each pixel, and outputs the fused clear image and pixel depth mapping table, realizing clear imaging and depth information extraction of overlapping abrasive grains.

[0053] The overlapping abrasive grain recognition module of the dynamic focusing unit reads the pixel intensity distribution data of the three-dimensional image stack, calculates the intensity change curve of each pixel at different depth levels, and identifies the focal depth of the abrasive grain corresponding to the peak position in the intensity curve. The overlapping abrasive grain recognition module classifies pixels with similar focal depths into the same abrasive grain individual, analyzes the abrasive grain distribution pattern at different depth levels, and identifies the spatial relationship between foreground abrasive grains and background abrasive grains in the overlapping area. The overlapping abrasive grain recognition module reconstructs the complete contour of the blocked abrasive grain according to the clear pixels at each depth level, completes the boundary information of the background abrasive grain blocked by the foreground abrasive grain, and outputs the depth coordinates and complete contour data of each abrasive grain individual, effectively solving the incomplete topography feature extraction and counting error problem caused by overlapping abrasive grains.

[0054] As Figure 7As shown, the acoustic separation unit includes an ultrasonic transmitter array, a phase control circuit, an acoustic focusing lens, an amplitude modulator, and an acoustic field distribution controller. The ultrasonic transmitter array is located at the bottom of the detection chamber, the phase control circuit calculates the phase delay value of each transducer according to the target focusing position and generates a multi-channel driving signal with a specific phase relationship, and the acoustic focusing lens performs a phase modulation operation on the incident acoustic wave. The ultrasonic transmitter array of the acoustic separation unit converts electrical signals into mechanical vibrations, generates a controllable acoustic wave phase distribution through the matrix arrangement of the transducers, and forms an acoustic wave emission surface at the bottom of the detection chamber. The amplitude modulator calculates the required acoustic radiation force according to the abrasive particle aggregation information and converts it into the transducer driving power requirement, and the acoustic field distribution controller controls the acoustic wave focus to move in the detection area according to the predetermined path. The acoustic separation unit calculates the required phase delay value of each transducer according to the target focusing position, determines the phase configuration of the acoustic wave forming constructive interference at the target point using the wavefront reconstruction method, and generates independent phase control parameters for each transducer. After receiving the phase control parameters, the acoustic separation unit generates the corresponding driving signal, shifts the reference signal according to the calculated phase delay, and outputs a multi-channel driving signal with a specific phase relationship for the transducer array, realizing the accurate focusing of acoustic wave energy at the abrasive particle aggregation position.

[0055] The acoustic focusing lens of the acoustic separation unit performs a phase modulation operation on the incident acoustic wave, determines the phase delay amount of the acoustic wave through the geometric size and arrangement of the multi-layer metamaterial units, and converts the plane acoustic wave into a focused acoustic wave with a specific wavefront shape. The acoustic focusing lens calculates the propagation path of the acoustic wave according to the metamaterial parameters, predicts the focusing effect of the acoustic wave in the oil medium, and determines the spatial position of the acoustic wave focus and the acoustic pressure distribution. The acoustic focusing lens adjusts the working state of the metamaterial units, changes the position of the acoustic wave focus in the detection area, and realizes the concentration of acoustic wave energy at the target abrasive particle position, improving the targeting and efficiency of acoustic separation.

[0056] The amplitude modulator of the acoustic separation unit receives the abrasive particle aggregation information output by the image analysis module, calculates the required acoustic radiation force according to the strength of the abrasive particle bonding force and the tightness of the aggregation, and converts the acoustic radiation force requirement into the transducer driving power requirement. The amplitude modulator allocates driving power according to the position and effect of each transducer in the array, adjusts the voltage amplitude of each transducer to achieve the required power output, and outputs a power control scheme matching the target acoustic field. The amplitude modulator amplifies the power control signal to the voltage level required by the transducer, keeps the phase relationship of each signal unchanged, and provides stable driving power to the transducer array, ensuring the controllability and effectiveness of the acoustic separation process.

[0057] The sound field distribution controller of the acoustic separation unit synchronously manages the working states of the phase control circuit and the amplitude modulator, generates a path planning for acoustic wave scanning according to the spatial distribution of the abrasive particle separation task, and controls the acoustic wave focal point to move in the detection area according to the predetermined path. The sound field distribution controller calculates the moving track and the residence time of the acoustic wave focal point, optimizes the scanning sequence to improve the separation efficiency, and ensures that the abrasive particles gathered at different positions receive appropriate acoustic wave action. The sound field distribution controller manages the time allocation of the acoustic wave action, coordinates the processing sequence of each gathering area, and realizes selective and time-sequential acoustic wave processing of the abrasive particle gathering area, effectively avoiding unnecessary waste of acoustic wave energy.

[0058] The working process of the acoustic separation unit is based on the physical action of acoustic radiation force on micro particles. When ultrasonic waves propagate in the oil medium and encounter abrasive particles, acoustic scattering occurs, and the scattered acoustic waves interfere with the incident acoustic waves in the space around the abrasive particles. The inhomogeneity of the acoustic pressure distribution formed by the wave interference produces a net radiation force on the abrasive particles. The acoustic separation unit establishes a scattering model based on the geometric size and material properties of the abrasive particles, calculates the interaction strength of the scattered waves and the incident waves, and determines the vector direction and magnitude of the acoustic radiation force. The acoustic separation unit decomposes the acoustic radiation force into components along different spatial directions, analyzes the contribution of each directional component to the motion of the abrasive particles, and predicts the motion trajectory of the abrasive particles under the action of the acoustic field, thereby realizing accurate control of the abrasive particle separation process.

[0059] The multi-frequency acoustic wave controller of the acoustic separation unit simultaneously generates multiple ultrasonic signals of different frequencies, modulates the amplitude and phase of each frequency signal, and generates a sound field distribution of composite frequencies in the abrasive particle gathering area. The multi-frequency acoustic wave controller selects appropriate combinations of acoustic wave frequencies according to the size distribution of the abrasive particles, utilizes the selective action of acoustic waves of different frequencies on abrasive particles of different sizes, and realizes the preferential separation of abrasive particles of specific sizes. The multi-frequency acoustic wave controller analyzes the spatial superposition effect of multi-frequency acoustic waves, calculates the amplitude and phase distribution of the acoustic pressure at each position in the composite sound field, and determines the optimal frequency configuration and power allocation scheme to improve the separation adaptability of abrasive particles of different sizes.

[0060] The acoustic separation unit calculates the minimum separation force threshold required to overcome the adhesion between abrasive particles, analyzes the influence of van der Waals force and surface tension between abrasive particles, and determines the acoustic wave parameter configuration to generate sufficient separation force. The acoustic separation unit adjusts the combination of acoustic wave frequency, phase, and power to generate the optimal acoustic radiation force distribution at the target position, and controls the time length and intermittent mode of acoustic wave action. The acoustic separation unit monitors the real-time position changes of the abrasive particles, adjusts the acoustic wave action parameters according to the separation progress, and realizes the gradual and controllable physical separation of the gathered abrasive particles, effectively solving the problem of reduced detection accuracy caused by abrasive particle gathering.

[0061] The separation effect monitoring module of the acoustic wave separation unit collects the images of abrasive particles in real time during the action of acoustic waves, compares the spatial distribution changes of abrasive particles before and after separation, and counts the number of successfully separated abrasive particles and the remaining aggregation degree. The separation effect monitoring module analyzes the position change trajectory of each abrasive particle, calculates the abrasive particle moving distance and separation speed, and evaluates the efficiency and integrity of the separation process. The separation effect monitoring module compares and analyzes the separation results with the expected target, generates quantitative evaluation indicators of the separation effect, and feeds back optimization suggestions to the acoustic wave controller. The acoustic wave controller dynamically adjusts the acoustic wave parameter configuration according to the feedback information to improve the separation effect, forming a closed-loop optimized separation process control.

[0062] As shown in Figure 8 The analysis unit includes an aggregation detection network, a separation planning network, and a feature extraction network, which are specially designed for deep learning processing of abrasive particle aggregation problems. The aggregation detection network performs feature mapping extraction operations on input images to identify abrasive particle aggregation phenomena, the separation planning network generates an execution plan for acoustic wave separation based on the aggregation detection results, and the feature extraction network performs deep feature analysis on the separated independent abrasive particles. The aggregation detection network of the analysis unit receives image data from the multi-view imaging unit, performs feature mapping extraction operations on the input images, performs gradient filtering operations on the image pixel matrix through the first layer of convolution kernels, calculates the gray difference between adjacent pixels to generate an edge response map, and highlights the boundary information between abrasive particles and the background. The aggregation detection network performs texture pattern recognition operations on the feature maps output by the first layer through the second layer of convolution kernels, uses Gabor filter groups to extract texture features of different directions and frequencies, and generates texture feature maps to distinguish the roughness and smoothness characteristics of abrasive particle surfaces. The aggregation detection network performs spatial correlation analysis on the feature maps of the previous layer through the third layer of convolution kernels, calculates the spatial distance and direction relationship between feature points, and generates spatial correlation feature maps to display the aggregation mode and distribution law between abrasive particles, thereby accurately identifying the initial position and range of abrasive particle aggregation and effectively solving the problem that traditional methods cannot identify complex aggregation patterns.

[0063] The aggregation detection network includes a spatial density analysis module, a shape abnormality evaluation module, and a boundary continuity detection module. The spatial density analysis module calculates the density value by counting the number of abrasive particle feature points in the local area, the shape abnormality evaluation module calculates the geometric feature parameters of the abrasive particle contour and compares them with the standard template, and the boundary continuity detection module analyzes the breaking points and curvature changes of the contour to judge the aggregation tightness.

[0064] The spatial density analysis module of the aggregation detection network reads the feature map data output by the convolutional layer, sets a circular window with a fixed radius at each pixel position, and calculates the local density value by counting the number of detected abrasive feature points within the window. The spatial density analysis module compares the local density value with a preset threshold value, marks the pixel positions with density values exceeding the threshold value as high-density candidate points, and connects adjacent high-density candidate points to form the boundary contour of the aggregation region. The spatial density analysis module calculates the geometric center point coordinates from the aggregation region contour and outputs the position information of the high-density region for subsequent processing, thereby quickly locating the key region of abrasive dense distribution and improving the efficiency and accuracy of aggregation detection.

[0065] The shape abnormality evaluation module of the aggregation detection network reads the pixel coordinate sequence of the abrasive profile, calculates the geometric feature parameters of the profile, and obtains the aspect ratio value by dividing the length of the long side of the minimum circumscribed rectangle of the profile by the length of the short side. The shape abnormality evaluation module measures the perimeter and area of the profile, calculates the similarity of the profile to a standard circle using the circularity formula, and constructs a convex hull polygon of the profile. The shape abnormality evaluation module calculates the difference between the calculated geometric parameters and the parameters of the standard abrasive template, identifies abrasive profiles with significantly deviated geometric parameters, and outputs the identification of suspected aggregation regions with shape abnormalities, thereby detecting abrasive aggregation from the perspective of shape change and improving the accuracy of aggregation identification.

[0066] The boundary continuity detection module of the aggregation detection network performs boundary tracking processing on the abrasive profile, moves pixel by pixel along the boundary direction from the starting point of the profile, records the coordinates and moving direction of each boundary point, and detects positions with direction changes exceeding a threshold value and marks them as turning points. The boundary continuity detection module analyzes the distance between adjacent boundary points, identifies positions with distances exceeding a continuity threshold value and marks them as broken points, and counts the number and distribution of broken points in the profile. The boundary continuity detection module calculates the local curvature value for each point on the profile, calculates the curvature size using a three-point circular arc fitting method, and calculates the standard deviation of curvature variation as a smoothness indicator. The boundary continuity detection module judges the tightness of abrasive aggregation based on the number of broken points and the degree of curvature variation, and outputs a quantitative score of aggregation tightness, thereby evaluating the binding strength of abrasive aggregation and providing an important reference for parameter selection.

[0067] The aggregation type classifier of the aggregation detection network receives the output data of the above three modules, combines the density value, shape parameter and boundary feature into a multi-dimensional feature vector, and performs branch judgment according to the numerical range of the feature vector. The aggregation type classifier first judges whether there is an aggregation phenomenon according to the density value, then judges the degree of aggregation according to the shape abnormality degree, and finally judges the type of aggregation according to the boundary continuity. The aggregation type classifier divides the aggregation state into four levels of slight contact, partial overlap, complete aggregation and multi-layer stacking, assigns a digital code identifier to each level, and outputs the aggregation level identifier and the boundary box coordinate data of the corresponding region, providing accurate target positioning information for subsequent separation operation, realizing accurate classification and positioning of aggregated abrasive particles.

[0068] The separation planning network of the analysis unit receives the aggregation region information output by the aggregation detection network and the material data provided by the spectral analysis unit, generates an execution scheme of acoustic wave separation according to the input data, and associates and matches the aggregation information and the material information. The separation planning network establishes a correspondence table of aggregation regions and material types, assigns a corresponding material attribute identifier to each aggregation region, and queries a preset separation parameter template according to the association table to select the parameter configuration that best matches the current aggregation type and material combination, and outputs a complete separation scheme including frequency, power and timing, ensuring that the separation parameters are highly matched with the actual situation and improving the targeting and success rate of the separation operation.

[0069] The separation planning network includes a frequency selection module, a power distribution calculation module and an action time planning module. The frequency selection module calculates the optimal acoustic wave frequency according to the abrasive particle size distribution, the power distribution calculation module determines the required acoustic wave power level according to the aggregation degree, and the action time planning module calculates the duration of acoustic wave action and divides it into multiple short pulse periods.

[0070] The frequency selection module of the separation planning network calculates the optimal acoustic wave frequency according to the size distribution and material density of the abrasive particles, extracts the equivalent diameter of each abrasive particle from the image data, counts the size distribution histogram of the abrasive particles, and calculates the mean and standard deviation of the size distribution. The frequency selection module calculates the resonance frequency of each abrasive particle using the acoustic resonance formula according to the size and material density of the abrasive particle, establishes a functional relationship between the size and the resonance frequency, and generates a size-frequency correspondence data table. The frequency selection module queries the data table according to the size range of the abrasive particles in the aggregation region, determines the optimal frequency interval covering the size range, and outputs a frequency configuration scheme of multiple frequency point combinations, realizing selective separation of abrasive particles of different sizes and improving the targeting and efficiency of the separation process.

[0071] The power distribution calculation module of the separation planning network determines the required acoustic power level according to the aggregation degree and inter-particle bonding strength, calculates the adhesion force between abrasive particles according to the material and surface characteristics of the particles, calculates the bonding force value using the van der Waals force model and surface energy theory, and establishes a mapping relationship between aggregation type and bonding strength. The power distribution calculation module calculates the acoustic radiation force required to overcome adhesion according to the bonding force, converts the force requirement into acoustic power requirement using the acoustic radiation force formula, and allocates the corresponding power coefficient to each transducer. The power distribution calculation module checks whether the power distribution of each transducer exceeds the safety limit, adjusts the power distribution ratio under the premise of ensuring separation effect, and outputs a power distribution scheme that meets the safety requirements, avoiding equipment overload while ensuring separation effect and ensuring safe and stable operation of the equipment.

[0072] The action time planning module of the separation planning network calculates the duration of acoustic action according to the stability of particle aggregation and the strength of adhesion force, evaluates the response characteristics of aggregated particles under acoustic action, calculates the energy and time required for particles to separate from the aggregated state, and establishes a relationship model between aggregation stability and separation time. The action time planning module divides the continuous action time into multiple short pulse periods, inserts an interval period between the pulses for effect observation and parameter adjustment, and calculates the optimal ratio of pulse width and interval time. The action time planning module encodes the pulse timing information into digital control instructions, generates timing data packets containing start time, duration and interval time, and outputs timing instructions that can be directly used for hardware control, realizing time control of the separation process and avoiding particle damage caused by excessive action.

[0073] The action sequence optimization module of the separation planning network analyzes the spatial distribution and mutual influence relationship of the aggregated areas, calculates the relative position and distance of each aggregated area, identifies the possibility of acoustic interference between adjacent areas, and establishes a relationship diagram of mutual influence between areas. The action sequence optimization module determines the processing sequence according to the independence and separation difficulty of the aggregated areas, preferentially selects areas with strong independence and low separation difficulty, and avoids simultaneous processing of adjacent areas to prevent acoustic interference. The action sequence optimization module generates a processing queue of aggregated areas according to the priority order, allocates a processing time window to each area, and outputs an execution schedule containing the processing sequence and time arrangement, maximizing separation efficiency and avoiding mutual interference, improving the effect and efficiency of the overall separation operation.

[0074] The safety evaluation module of the separation planning network checks the safety impact of the generated separation scheme on the abrasive particles, calculates the acoustic pressure stress on the surface of the abrasive particles according to the acoustic wave parameters, predicts the distribution and size of the stress using acoustic theory models, and compares the predicted stress value with the mechanical strength of the abrasive particle material. The safety evaluation module determines whether the acoustic stress exceeds the yield strength or the fracture strength of the material, identifies risk parameters that may cause the abrasive particles to break or deform, and triggers parameter adjustment when safety risks are detected. The safety evaluation module automatically reduces the power parameters that exceed the limit or modifies the action mode, recalculates the stress level after adjustment to ensure safety, and outputs the corrected safety separation scheme, which achieves the separation target while protecting the integrity of the abrasive particles, ensuring the accuracy and reliability of the detection results.

[0075] The feature extraction network of the analysis unit performs deep feature analysis on the separated independent abrasive particles, realizes pixel-level abrasive particle boundary recognition, performs foreground and background classification on each pixel in the image, extracts the context features of the pixels using a deep convolutional network, and assigns the pixels to corresponding abrasive particle instances according to feature similarity. The feature extraction network assigns a unique instance identifier to each abrasive particle, ensuring that pixels of the same abrasive particle have the same identifier, and generates an instance segmentation mask map for subsequent boundary extraction, achieving accurate separation and identification of individual abrasive particles and solving the problem of individual identification after separation of aggregated abrasive particles.

[0076] The feature extraction network includes a boundary detection convolutional layer, a contour refinement module, and a multi-scale feature fusion module. The boundary detection convolutional layer uses a deformable convolution kernel to detect the irregular shape contour of the abrasive particle. The contour refinement module performs sub-pixel level processing on the boundary contour. The multi-scale feature fusion module processes image data of different resolutions and extracts detailed features and shape features.

[0077] The boundary detection convolutional layer of the feature extraction network uses a deformable convolution kernel to detect the irregular shape contour of the abrasive particle. It dynamically adjusts the sampling point position according to local gradient information, calculates the gradient direction and intensity to determine the optimal sampling position, and concentrates the sampling points of the convolution kernel in the boundary vicinity area. The boundary detection convolutional layer uses the adjusted sampling points to extract boundary features, captures detailed feature information of sharp edges and small protrusions, and generates a high-resolution boundary feature map, accurately detecting the boundary details of complex-shaped abrasive particles and improving the accuracy and integrity of boundary recognition.

[0078] The edge enhancement processing module of the feature extraction network performs morphological operations on the boundary detection result, performs morphological dilation on the boundary image, expands the width of the boundary line using a circular element, and connects the broken boundary line segments close in distance. The edge enhancement processing module performs morphological erosion on the dilation result, shrinks the boundary line width using an element of the same size, restores the original thickness of the boundary line and maintains the connectivity. The edge enhancement processing module identifies and removes pseudo-boundaries caused by noise, filters valid boundaries according to the length and continuity of the boundary line segments, and outputs clean abrasive grain boundary contours to obtain complete and accurate abrasive grain boundary information and provide reliable basic data for subsequent feature extraction.

[0079] The contour refinement module of the feature extraction network performs sub-pixel level processing on the rough boundary contour, extracts the pixel coordinate points of the contour from the boundary image, sorts the coordinate points according to the contour direction, and generates an ordered contour point sequence. The contour refinement module performs spline curve fitting on the contour point sequence, uses B-spline function to reconstruct a smooth and continuous contour curve, and interpolates to generate sub-pixel precision contour point coordinates. The contour refinement module performs local optimization on the fitting result to eliminate oscillation and overfitting phenomena in the fitting process, and outputs a high-precision abrasive grain contour coordinate sequence to provide sub-pixel level contour measurement accuracy and significantly improve the accuracy of abrasive grain size and shape measurement.

[0080] The multi-scale feature fusion module of the feature extraction network processes image data of different resolutions, decomposes the input image into multiple image layers of different resolutions, generates a multi-scale representation of the image using the Gaussian pyramid method, and maintains the spatial correspondence of the images at each scale. The multi-scale feature fusion module extracts texture features of the abrasive grain surface from the detail image, identifies surface scratches, particles and reflection characteristics, generates a detail feature vector, and simultaneously extracts shape features of the abrasive grain from the overall image, calculates geometric parameters and shape descriptors of the contour, and generates a shape feature vector. The multi-scale feature fusion module combines the detail features and shape features by weighting, assigns weight coefficients according to the importance of the features, and generates a complete feature vector containing multi-level information, while capturing the local details and overall characteristics of the abrasive grain, providing a comprehensive and accurate description of the abrasive grain features.

[0081] The local detail analysis module of the feature extraction network is specifically designed to handle the microscopic features of the abrasive particle surface, identify scratches, pits and oxidation spots on the abrasive particle surface, locate surface defects using edge detection and region segmentation methods, and calculate the geometric size and distribution density of the defects. The local detail analysis module measures the height variation of the abrasive particle surface, estimates the surface roughness using the gray level gradient variation, calculates the root mean square value of the surface roughness, and calculates the main direction of the surface texture, determines the dominant direction of the texture using the gradient statistical method, and outputs the directionality index of the texture. The local detail analysis module converts the surface features into numerical evaluation results, generates a comprehensive score of the surface quality, and outputs a standardized description of the abrasive particle surface state, providing an important basis for quantitative evaluation of the surface quality and wear degree of the abrasive particle and equipment state diagnosis.

[0082] The overall shape analysis module of the feature extraction network calculates the geometric shape parameters of the abrasive particle, calculates the equivalent circle diameter based on the abrasive particle area, converts the abrasive particle area to the diameter of a circle with the same area, and provides a standardized representation of the abrasive particle size. The overall shape analysis module calculates the maximum length and maximum width of the abrasive particle, fits the smallest circumscribed ellipse of the abrasive particle, extracts the lengths of the major and minor axes of the ellipse, and evaluates the degree of deviation of the abrasive particle shape from the regular geometric shape, compares the difference between the abrasive particle contour and the standard geometric shape, and outputs the shape irregularity coefficient. The overall shape analysis module identifies the basic shape type of the abrasive particle based on the geometric parameters, classifies the abrasive particle into spherical, flaky, needle-like or irregular shape, and calculates the similarity index of each type. The overall shape analysis module converts all geometric parameters into a unified format of description data, generates a digital archive of the abrasive particle shape, and outputs standard data that can be used for database storage and comparative analysis, establishing a complete database of abrasive particle shape features for equipment state monitoring and trend analysis, and providing a scientific basis for equipment maintenance decision-making.

[0083] The workflow of the image sensor for online detection of oil quality is processed in an iterative optimization manner. The image sensor first sends a start instruction to the multi-view imaging unit, and the multi-view imaging unit starts image data acquisition operation after receiving the instruction and establishes a data communication link with the aggregation detection network. The image sensor encapsulates the initial image data into a data packet format, which contains image pixel matrix, time stamp and view identification information, and sends the data packet to the input buffer of the aggregation detection network. The aggregation detection network of the analysis unit reads the image data packet from the buffer, extracts the pixel matrix and reconstructs the complete image data, starts the aggregation recognition processing flow, thereby ensuring the integrity and time sequence correctness of the image data.

[0084] The analysis unit's aggregation detection network processes results containing pixel coordinate lists of aggregation areas and numerical encodings of aggregation degrees, converts the processing results into a standardized data format, and generates data packets containing spatial coordinates and level identifiers. The image sensor routes the aggregation detection results to the spectral analysis control module according to the type identifier of the data packets, establishes a priority queue for data transmission, and ensures timely transmission of critical data, thereby reducing data transmission delays and improving processing efficiency, achieving efficient coordination between functional modules.

[0085] The spectral analysis control module of the image sensor analyzes the aggregation area information, extracts the spatial coordinates and range data of the aggregation area, and generates a target area list for spectral scanning. The spectral analysis control module generates an execution sequence for spectral analysis based on the target area list, assigns a scanning time window to each area, and sends execution instructions to the spectral analysis unit. The spectral analysis unit activates different wavelengths of light sources in sequence according to the instruction sequence, records the reflected light intensity data of the target area at each wavelength, and constructs a multi-wavelength spectral data matrix. The spectral analysis unit performs material recognition operations on the spectral data matrix, calculates the spectral feature vector of each pixel position, and matches the feature vector with the material database to obtain the material recognition result, thereby obtaining the material composition information of the abrasive particles in the aggregation area.

[0086] The image sensor converts the material recognition result and spectral data into a data format recognizable by the separation planning network, compresses the spectral data and extracts key feature parameters, and generates data packets containing material identifiers and feature parameters. The image sensor receives aggregation information and material data, establishes an association mapping table between aggregation areas and material types, and provides a complete input data set for the separation planning network. The separation planning network of the analysis unit performs optimization calculations of acoustic parameters based on the input data set, uses a multi-objective optimization method to solve the optimal parameter combination, generates parameter configuration instructions containing frequency, power, and timing, and ensures the best match between acoustic parameters and actual aggregation conditions, significantly improving the success rate of separation operations.

[0087] The image sensor encodes the parameter configuration instruction into a hardware control instruction format, converts the floating-point parameters into hardware recognizable digital control codes, and adds a check code to ensure the accuracy of instruction transmission. The image sensor sends the control instruction to the dynamic focusing control module and the acoustic separation control module, establishes a bidirectional communication channel to ensure reliable transmission of the instruction, and monitors the instruction execution state to provide feedback information. The control module of the dynamic focusing unit parses the received control instruction, extracts the focusing depth and scanning range parameters, and sends the focusing control signal to the dynamic focusing unit. The dynamic focusing unit collects image data at different depth levels according to the control signal, records the focusing quality evaluation index of each depth level, and constructs an image data stack in three-dimensional space. The dynamic focusing unit performs spatial coordinate calculation on the image stack to determine the position coordinates of each abrasive particle in three-dimensional space, and generates a digital model of the spatial distribution of abrasive particles.

[0088] The image sensor converts the three-dimensional position data into an acoustic control coordinate system, performs coordinate system rotation and translation operations, and converts the image coordinates into physical coordinates of the acoustic transmitter array. The image sensor merges the three-dimensional position data with the separation parameter data to generate comprehensive control instructions containing target position and action parameters, and transmits complete separation task data to the acoustic separation control module. The control module of the acoustic separation unit analyzes the comprehensive control instruction, extracts the target coordinate and acoustic parameter information, and generates the drive signal configuration of the transducer array. The acoustic separation unit generates multiple electric signals with controllable phase and amplitude according to the drive signal configuration, outputs the electric signals to the transducer array, and the transducer array converts the electric signals into acoustic energy and generates acoustic radiation force in the target area, realizing directional separation of the gathered abrasive particles and effectively solving the detection accuracy problem caused by abrasive particle gathering.

[0089] After the separation operation is completed, the image sensor detects the work completion signal of the acoustic separation unit, triggers a new round of image acquisition process, and restarts the multi-view imaging unit to obtain the image data after separation. The image sensor receives the image data before and after separation, performs pixel-level difference operation on the two groups of image data, identifies the changed areas and pixel positions in the image. The image sensor performs region segmentation and labeling on the difference result, calculates the area and position information of each changed area, and judges whether the changed area corresponds to the successfully separated abrasive particles, thereby quantitatively evaluating the effect of the separation operation and providing accurate feedback data for iterative optimization.

[0090] The effect evaluation module of the image sensor performs numerical statistics on the separation results, counts the number of successfully separated abrasive particles and the number of remaining aggregated areas, and calculates the separation success rate and the percentage of aggregation degree improvement. The effect evaluation module scans the separated image data, identifies the coordinates of the areas where aggregation still exists, and analyzes the characteristics and distribution patterns of the residual aggregated areas. The effect evaluation module encodes the statistical data and residual aggregation information into an evaluation report, generates a data package containing the success rate index and the coordinates of the failure areas, and transmits the evaluation report to the parameter adjustment module to provide quantitative basis for subsequent parameter optimization, realizing closed-loop control and continuous improvement of the separation process.

[0091] The parameter adjustment module of the image sensor receives the evaluation report data, compares the difference between the expected separation effect and the actual effect, and identifies the key factors that cause separation failure. The parameter adjustment module recalculates the acoustic wave parameters based on the failure cause analysis results, increases the power value or modifies the frequency combination to improve the separation effect, and optimizes the parameter configuration using feedback control methods. The parameter adjustment module converts the adjusted parameters into new control instructions, generates a special separation scheme for the remaining aggregated areas, and transmits the secondary separation instructions to the acoustic wave separation control module to specifically address the areas that have not been completed in the first separation, thereby improving the overall separation completion rate.

[0092] The control module of the acoustic wave separation unit receives the secondary separation instructions, replaces the original control configuration with new parameters, and reconfigures the driving parameters of the transducer array. The acoustic wave separation unit restarts the separation operation under the adjusted parameter configuration, specifically targets the remaining aggregated areas for acoustic wave action, and monitors the real-time progress of the secondary separation. The iterative control module of the image sensor records the number of separation operation executions and monitors whether the cycle count reaches the preset upper limit, as well as whether the separation success rate reaches the target threshold. The iterative control module decides whether to terminate the iteration process based on the cycle count and success rate data, sends a stop signal to the flow control module when the termination condition is met, and triggers the final feature extraction processing flow to ensure that the separation task is completed within a reasonable time and to avoid excessive processing.

[0093] After the iteration process is terminated, the image sensor starts the feature extraction network, transmits the separated abrasive particle image data to the input port of the feature extraction network, and establishes a processing queue for feature extraction. The feature extraction network of the analysis unit performs feature analysis operations on each independent abrasive particle, extracts the shape, texture, and surface features of the abrasive particle, and generates a feature vector containing multi-dimensional feature parameters. The image sensor stores the feature vectors of each abrasive particle in the feature database, establishes an index relationship between the abrasive particle identifier and the feature vector, and outputs a complete abrasive particle feature data set for comprehensive evaluation of oil quality, and establishes a complete abrasive particle feature archive for equipment state diagnosis, providing a scientific basis for equipment maintenance decisions.

[0094] The image sensor for online detection of oil quality integrates a real-time monitoring feedback unit, ensuring the safety and effectiveness of the separation process. As shown in Figure 9 the real-time monitoring feedback unit includes a high-speed image acquisition module, a motion state analysis module, a force sensing monitoring module, and a safety control module. The high-speed image acquisition module of the real-time monitoring feedback unit receives continuous image frames collected by a high-frame-rate camera in real time, arranges and stores the image frames in time sequence, establishes a ring buffer queue to ensure data continuity. The high-speed image acquisition module transmits continuous image frame data to the motion state analysis module, maintains the time sequence relationship of the image frames, adds a timestamp to ensure the time accuracy of motion analysis, and realizes real-time tracking of abrasive particle motion state.

[0095] The motion state analysis module of the real-time monitoring feedback unit performs pixel-level difference operation on continuous image frames, calculates the pixel gray difference between adjacent frames, and generates a binary mask image of the motion area. The motion state analysis module identifies and segments the moving abrasive particle target from the mask image, calculates the centroid coordinates and boundary contour of the moving target, and establishes the identification and tracking record of the abrasive particle target. The motion state analysis module calculates the motion trajectory according to the position change of the abrasive particle in the continuous frames, calculates the velocity vector using the position difference method, and calculates the acceleration vector using the velocity difference method. The motion state analysis module monitors whether the motion parameters of the abrasive particle exceed the normal range, identifies abnormal high-speed motion or abnormal rotation behavior, judges whether the abrasive particle deviates from the expected separation trajectory, and discovers abnormal conditions in the separation process in a timely manner to ensure the safety and controllability of the separation operation.

[0096] The motion state analysis module generates a warning signal when detecting abnormal motion, encodes the warning signal into a digital message format, and sends the message to the signal receiving port of the safety control module. The aggregation warning module of the motion state analysis module calculates the mutual distance between the abrasive particles after separation, monitors the distance change trend between the abrasive particles, and identifies the abrasive particle pair with rapidly reduced distance. The aggregation warning module activates the warning when detecting the risk of secondary aggregation, generates a secondary aggregation risk signal, transmits the risk signal to the safety control module, prevents the re-aggregation of abrasive particles after separation, and ensures the persistence of the separation effect.

[0097] The force sensing monitoring module of the real-time monitoring feedback unit receives the analog signal output by the pressure sensor, amplifies and filters the signal, and converts the analog signal to digital data. The force sensing monitoring module performs frequency spectrum analysis on the digital data, extracts the frequency component and amplitude information of the vibration signal, and monitors the mechanical stress change of the detection chamber. The force sensing monitoring module compares the measured data with the preset safety threshold value, identifies the stress level exceeding the safety limit, generates an overload warning signal when detecting overload stress, protects the detection equipment from excessive stress damage, and ensures the long-term stable operation of the equipment.

[0098] The safety control module of the real-time monitoring feedback unit receives state information from each monitoring module, prioritizes and sorts the signals, and establishes a priority queue for signal processing. The safety control module selects the corresponding safety response measures according to the type of warning signal, generates a power reduction instruction when detecting abnormal movement, generates a pause operation instruction when detecting secondary aggregation risk, and generates an emergency stop instruction when detecting overload stress. The safety control module sends the response instruction to the corresponding control module to ensure the priority execution of the safety instruction and monitor the execution effect of the safety measures. In the event of an abnormal situation, the safety control module quickly implements protection measures to ensure the safety of the equipment and the sample, and realizes comprehensive safety protection for the separation process.

[0099] The image sensor for online detection of oil quality is used to automatically optimize the processing parameters according to the separation effect of different types of abrasive particles. As shown in Figure 10 The adaptive learning unit of the image sensor includes an aggregation pattern recognition module, a parameter optimization module, an effect evaluation module, and a knowledge base updating module. The aggregation pattern recognition module establishes a feature data index table of abrasive particle aggregation patterns, the parameter optimization module dynamically adjusts the acoustic wave parameters according to the separation effect, and the knowledge base updating module identifies high-quality processing cases and updates the database. The aggregation pattern recognition module of the adaptive learning unit establishes a feature data index table of abrasive particle aggregation patterns, stores the image feature vectors of different aggregation types and the corresponding optimal separation parameter combinations in association, and uses a hash table to establish a fast retrieval relationship between the feature vectors and the parameter configurations. The aggregation pattern recognition module performs multi-layer feature extraction operations on newly encountered aggregation conditions, calculates the spatial distribution features, shape features, and density features of the aggregation area, and organizes the extracted feature data into a standardized feature vector format. The aggregation pattern recognition module compares the new feature vector with the known patterns in the database one by one, performs vector inner product and module length calculations to obtain similarity values, and quickly identifies the similarity degree of the current aggregation pattern and the historical pattern.

[0100] The similarity calculation module of the adaptive learning unit calculates the Euclidean distance between the new feature vector and the database vectors, takes the square root of the sum of the feature difference values in each dimension to obtain the distance value, and calculates the cosine similarity between the vectors as an angle similarity index. The similarity calculation module calculates the final matching degree score by combining the distance and angle indexes using a weighted summation method, outputs the highest matching degree score and the corresponding historical aggregation pattern identifier. The similarity calculation module compares the matching degree score with the preset threshold value, activates the new pattern learning process when the matching degree is lower than the threshold value, triggers the expansion and update operation of the feature database, automatically identifies unknown aggregation pattern types, and ensures effective processing capability for new aggregation patterns.

[0101] The parameter optimization module of the adaptive learning unit extracts the corresponding separation parameter configuration from the database according to the matching result, takes the historical optimal parameters as the starting parameters of the current separation operation, and establishes the mapping relationship between the parameter configuration and the separation task. The parameter optimization module monitors the real-time changes of the separation effect during the separation process, dynamically adjusts the parameters of the sound wave frequency, power and action time according to the effect feedback data, and records the numerical changes and effect responses of each parameter adjustment. The parameter optimization module analyzes the mathematical relationship between parameter changes and effect improvement, calculates the partial derivatives of the effect function with respect to each parameter, determines the search direction and step size of parameter optimization, and saves the complete historical trajectory of parameter adjustment. The parameter optimization module establishes the time correspondence relationship between the parameter sequence and the effect sequence, generates learning sample data for parameter optimization, continuously improves the accuracy of parameter selection and separation effect, and realizes the continuous optimization of the separation process.

[0102] The effect evaluation module of the adaptive learning unit performs numerical calculation on multiple performance indicators of the separation operation, counts the number of successfully separated abrasive particles and the total processing time, and calculates the percentage values of the separation success rate and processing efficiency. The effect evaluation module records the power consumption of each device module during the separation process, accumulates the total energy consumption value, and analyzes the relationship mode between energy consumption and separation effect. The effect evaluation module checks the shape integrity and surface integrity of the abrasive particles after separation, compares the geometric parameter changes of the abrasive particles before and after separation, and calculates the quantitative indicators of abrasive particle damage. The effect evaluation module statistically compares the current evaluation results with historical data, identifies the performance indicator change trend and improvement amplitude, generates a performance evaluation report for optimization reference, and comprehensively evaluates the comprehensive performance level of the separation operation.

[0103] The knowledge base updating module of the adaptive learning unit identifies processing cases with high separation success rate and excellent efficiency, filters high-quality cases according to the preset performance threshold, and extracts feature vectors and parameter configuration data of high-quality cases. The knowledge base updating module integrates new high-quality cases with similar cases in the database, calculates the mean of the feature vectors of similar cases, and updates the feature vectors and parameter configurations of the corresponding entries in the database. The knowledge base updating module identifies historical records with poor effect or low usage frequency in the database, deletes redundant data according to performance scores and access frequencies, and maintains the storage efficiency and query speed of the database. The knowledge base updating module regularly checks the practicality and accuracy of each entry in the database, verifies the applicability of historical data in the current environment, updates the index and retrieval efficiency of the database, and ensures the continuous optimization and practical value of the knowledge base.

[0104] The new mode learning operation mechanism of the image sensor activates the learning process when encountering unknown aggregation types, analyzes the feature vector of the current aggregation mode, and searches the closest historical mode in the database as the basis for learning. The new mode learning operation mechanism establishes an initial processing scheme based on the most similar historical case, uses the parameter configuration of the similar case as the starting point, assigns a temporary identification code and parameter configuration to the new mode. The new mode learning operation mechanism performs small-scale parameter changes during the separation process, makes exploratory adjustments to the frequency, power, and timing parameters, and monitors the impact of each adjustment on the separation effect. The new mode learning operation mechanism establishes a new feature mode record based on the effect of the exploratory adjustment, associates and stores the optimal parameter combination with the feature vector, adds the new mode to the knowledge base for subsequent use, continuously expands the processing capability and adapts to new aggregation mode types, and improves the adaptability of the image sensor to complex aggregation conditions.

[0105] The multi-dimensional separation detection technology of the image sensor for online detection of oil quality effectively solves the problems of abrasive particle aggregation, overlap, and occlusion. The aggregation problem solution of the image sensor significantly improves the accuracy of abrasive particle counting through the synergistic effect of spatial analysis and acoustic wave separation. The aggregation detection network identifies the location of the aggregation area, and the acoustic wave separation unit performs directional separation operation on the aggregation area. The separated independent abrasive particles can be accurately counted and analyzed. The image sensor eliminates the counting errors in traditional methods where multiple abrasive particles are mistaken for a single large particle. In processing dense aggregation areas, the image sensor maintains stable recognition performance through multiple rounds of iterative separation, significantly improves the quantity accuracy and reliability of abrasive particle detection, and solves the core problem of insufficient counting accuracy in traditional methods.

[0106] The overlap problem solving capability of the image sensor is achieved through dynamic focusing and three-dimensional reconstruction technology to effectively handle deep separation. The dynamic focusing unit collects clear images at different depth levels, and the three-dimensional reconstruction module of the multi-layer image acquisition module identifies the spatial position relationship of the front and back of the multi-layer overlapping abrasive particles. The image sensor completes the outline reconstruction of the occluded abrasive particles based on the information fusion of multi-depth images, extracts the visible part of the abrasive particles from each depth level, splices the visible part into the complete abrasive particle boundary, extracts the shape features of the abrasive particles from the reconstructed outline, and obtains the complete topographic information of the occluded abrasive particles, effectively solving the information missing problem in overlapping abrasive particle detection.

[0107] The occlusion problem processing of the image sensor is achieved through multi-view imaging and stereo vision reconstruction for all-around observation. The multi-view imaging unit collects image data of the same abrasive particle from different angles, and the stereo vision processing module calculates the spatial correspondence of the multi-view images. The image sensor constructs a complete spatial model of the abrasive particle based on the multi-view image data, uses three-dimensional point cloud data to represent the surface geometric information of the abrasive particle, generates a three-dimensional mesh model of the abrasive particle for feature analysis, overcomes the limitations of single-view observation, and obtains comprehensive topographic information of the abrasive particle.

[0108] The reliability of topographic feature extraction of the image sensor is improved based on multi-dimensional information fusion and high-precision image processing technology. The feature extraction network extracts the detailed texture features of the abrasive particle surface from the high-resolution image, and uses a multi-scale filter to maintain the original detail information of the texture features. The high precision of the geometric parameter measurement of the image sensor is derived from the sub-pixel level boundary detection and contour fitting technology. The sub-pixel interpolation method is used to improve the boundary positioning accuracy, and the curve fitting is used to reduce the measurement error, so as to ensure the high fidelity and measurement accuracy of the abrasive particle feature information.

[0109] The comparison of the image sensor with the traditional two-dimensional image analysis method shows that more comprehensive abrasive particle information is obtained. The projection contour information of the traditional method only reflects the projection features of the abrasive particle in a single plane, and the three-dimensional reconstruction technology obtains the complete spatial shape information of the abrasive particle. The image sensor realizes material component recognition through multi-wavelength spectral analysis. The spectral analysis unit extracts the reflection characteristics of the abrasive particle under different wavelengths, and matches and identifies the spectral features with the material database. The comprehensive information acquisition of the image sensor includes multi-dimensional data of shape, surface texture, material component and spatial position, which provides more complete and accurate abrasive particle feature information for oil quality analysis, and significantly improves the comprehensiveness and accuracy of detection and analysis.

[0110] To verify the technical effects of the image sensor of the present application, a standardized test scheme is designed to compare and evaluate the technical scheme of the conventional two-dimensional image detection and the technical scheme of the multi-dimensional detection of the present application. The test environment uses a standard oil quality detection experimental platform, which is equipped with a constant temperature control system to stabilize the oil temperature at 25±1℃. A standard abrasive particle sample library is used, which includes ferromagnetic abrasive particles, copper abrasive particles and aluminum abrasive particles of different materials, and the abrasive particle size range covers 5μm to 50μm. During the preparation of the test sample, a known number and characteristics of standard abrasive particles are dispersed in transparent carrier oil, and different degrees of aggregation states are formed by magnetic field control, including four typical aggregation modes of slight contact, partial overlap, complete aggregation and multi-layer stacking. The comparative test uses a conventional single-view CCD image sensor combined with conventional image processing software, the test of Example 1 uses the complete technical scheme of the present application but uses fixed parameter configuration, and the test of Example 2 uses the technical scheme of the present application and activates the parameter optimization function of the adaptive learning unit. During the test, each method processes the same 100 groups of standard samples, each group of samples contains 20 to 80 abrasive particles of different sizes and materials, and the standard answer database is established by artificial microscope inspection for accuracy evaluation. The performance index calculation method is: the abrasive particle recognition accuracy is equal to the number of correctly recognized abrasive particles divided by the total number of standard abrasive particles, the success rate of separating aggregated abrasive particles is equal to the number of successfully separated aggregated abrasive particles divided by the total number of initial aggregated abrasive particles, the detection integrity of overlapping abrasive particles is equal to the number of detected overlapping abrasive particle complete contours divided by the total number of actual overlapping abrasive particles, and the morphology feature extraction accuracy is calculated by the relative error between the measured result and the standard value. The test results are as follows:

[0111]

[0112] Table 1 Performance comparison test results of the technical scheme of the present application and the conventional technical scheme

[0113] As shown in Table 1, the test data shows that the multi-dimension detection technical solution of the present application has achieved significant improvement in various key performance indicators compared with the traditional two-dimensional image detection technical solution. Embodiment 1 of the present application adopts a standard configuration of a multi-view imaging unit, a spectral analysis unit, a dynamic focusing unit and a sound wave separation unit, and the abrasive grain recognition accuracy is improved from 72.3% of the traditional method to 94.6%, and the success rate of separating aggregated abrasive grains is greatly improved from 31.5% to 87.2%. Embodiment 2 further improves the abrasive grain recognition accuracy to 96.8% and the success rate of separating aggregated abrasive grains to 91.4% based on the optimization configuration of the adaptive learning unit of Embodiment 1. The present application effectively solves the information missing problem in overlapping abrasive grain detection through multi-view stereoscopic imaging, and the completeness of overlapping abrasive grain detection is improved from 45.8% of the traditional method to 92.7% of Embodiment 2. The spectral analysis unit enables the present application to have material recognition capability, achieving a material recognition accuracy of 88.7% to 91.3%, while the traditional method does not have this function. The dynamic focusing unit enables the present application to support up to 12 layers of depth level recognition, and the three-dimensional spatial positioning accuracy reaches ±1.8μm, significantly improving the stereoscopic level and accuracy of detection.

[0114] The technical support of the image sensor for oil quality evaluation and equipment state diagnosis is significantly enhanced, and the output multi-dimensional abrasive grain data is used to establish a quantitative analysis model of the wear process, and the wear state of the equipment parts is inferred according to the size distribution, shape characteristics and material composition of the abrasive grains. The image sensor judges the type and severity of wear according to the characteristics of the abrasive grains, distinguishes different states of normal wear, abnormal wear and severe wear, trains a prediction model of the equipment health state using historical abrasive grain data, establishes a relationship model between abrasive grain characteristics and equipment remaining life, and realizes equipment state prediction and maintenance decision support based on abrasive grain analysis.

[0115] The above has introduced the embodiments of the present application in detail, and the content of the specification should not be understood as limiting the protection scope of the present application.

Claims

1. An image sensor for online detection of oil quality, characterized in that, The sensor body comprises a multi-view observation chamber, which is a multi-pyramid frustum, and is provided with a multi-layer optical window assembly comprising a main observation window located at the center of the top of the chamber and a plurality of auxiliary observation windows radially distributed around the main observation window, each of the auxiliary observation windows being inclined to the inside of the chamber. The multi-view imaging unit comprises a main imaging assembly for acquiring a top view image of the abrasive particles through the main observation window and an auxiliary imaging assembly for acquiring a side view angle image of the abrasive particles through the auxiliary observation window. The acoustic wave separation unit comprises an ultrasonic transmitter array located at the bottom of the multi-view observation chamber, a phase control circuit for calculating the phase delay value of each transducer according to the target focusing position and generating a multi-channel driving signal with a specific phase relationship, an acoustic wave focusing lens for performing a phase modulation operation on the incident acoustic wave to convert a plane acoustic wave into a focused acoustic wave with a specific wave front shape, an amplitude modulator for calculating the required acoustic radiation force size according to the abrasive particle aggregation information and converting it into the transducer driving power requirement, and an acoustic field distribution controller for controlling the acoustic wave focus to move in the detection area according to the predetermined path. The dynamic focusing unit comprises a liquid lens assembly containing a double-layer liquid medium of conductive liquid and insulating oil, a transparent electrode surrounding the lens cavity to form an electric field control area, a focal length control circuit for applying a variable voltage to the transparent electrode to generate an electric field intensity change to adjust the optical focal length of the liquid lens, a focal point position detector for emitting a reference laser beam to the detection area and calculating the focal point distance according to the reflected laser, and a multi-layer image acquisition module for continuously acquiring image data at different focal length positions and constructing a three-dimensional image data stack. The analysis unit comprises an aggregation detection network for performing feature mapping extraction operation on the input image to identify abrasive particle aggregation phenomenon, a separation planning network for generating an execution scheme of acoustic wave separation according to the aggregation detection result, and a feature extraction network for performing deep feature analysis on the separated independent abrasive particles. The circumferential angle interval between every two adjacent auxiliary observation windows is 45 degrees, and the angle between the window plane of each auxiliary observation window and the top surface of the chamber is 30 degrees. The multi-view imaging unit further comprises a field of view calculation unit for calculating the field of view boundary data of each window and a stereo vision processing module for distributing the image data to the corresponding processing channel according to the view angle identifier.

2. The image sensor of claim 1, wherein, ​ 3. The image sensor of claim 1, wherein, ​ 4. The image sensor of claim 1, wherein, Also included is a spectrum analysis unit, which includes a multi-wavelength LED array for sequentially activating light sources of different wavelengths according to a timing configuration table, and a spectrum filter set for receiving a wavelength synchronization signal and rotating to a predetermined angle to achieve wavelength switching.

5. The image sensor of claim 1, wherein, The aggregation detection network includes a spatial density analysis module for calculating density values by counting the number of feature points of abrasive particles in a local area, a shape abnormality evaluation module for calculating geometric feature parameters of the abrasive particle contour and comparing them with standard templates, and a boundary continuity detection module for analyzing the breaking points and curvature changes of the contour to determine the degree of close aggregation.

6. The image sensor of claim 1, wherein, The separation planning network includes a frequency selection module for calculating the optimal acoustic wave frequency according to the size distribution of abrasive particles, a power allocation calculation module for determining the required acoustic wave power level according to the degree of aggregation, and an action time planning module for calculating the duration of acoustic wave action and dividing it into multiple short pulse periods.

7. The image sensor of claim 1, wherein, The feature extraction network includes a boundary detection convolution layer for detecting the irregular shape contour of abrasive particles using a deformable convolution kernel, a contour refinement module for sub-pixel level processing of the boundary contour, and a multi-scale feature fusion module for processing image data of different resolutions and extracting detailed features and shape features.

8. The image sensor of claim 1, wherein, Also included is an adaptive learning unit, which includes an aggregation pattern recognition module for establishing a feature data index table of abrasive particle aggregation patterns, a parameter optimization module for dynamically adjusting acoustic wave parameters according to separation effects, and a knowledge base update module for identifying high-quality processing cases and updating the database.

Images