A laser illumination-based microscopic imaging method and probe device

Abstract

The application provides a kind of microscopic imaging method and probe device based on laser illumination, belong to microscopic imaging technical field, the present application is generated laser by semiconductor laser and is conducted to lens system by optical fiber and forms transmission type illumination, utilizes rotation diffusion sheet and carries out speckle time domain average processing to reduce speckle contrast to threshold value below, adopts pulsed laser stroboscopic illumination mode to eliminate motion blur, by telecentric lens and industrial camera collection through the groove in the particle image, with high density particle adhesion segmentation processing based on graph cut accurately separates adhesion particle from high solid content image and extracts particle shape parameter and particle size distribution data, finally by laser power self-adaptive regulation and control artificial intelligence model, according to image quality evaluation function, realize illumination power closed loop control, solve the technical problem that particle adhesion segmentation precision is seriously reduced under high solid content scene caused by laser speckle noise.

Description

Technical Field

[0001] This invention belongs to the field of microscopic imaging technology, and more specifically, relates to a microscopic imaging method and probe device based on laser illumination. Background Technology

[0002] Online microscopy is an important tool for monitoring industrial crystallization processes. By directly inserting a probe into the crystallizer solution, it acquires real-time images of suspended particles, thereby obtaining particle shape parameters and particle size distribution data. Due to the advantages of semiconductor lasers, such as high brightness, strong directivity, and ease of fiber optic transmission, laser illumination schemes are widely used in the illumination systems of online microscopy probes. However, the high coherence of lasers causes random coherent superposition of the illumination beams on the imaging surface, generating speckle noise. This noise manifests as randomly distributed bright and dark spots in the image grayscale field, with a speckle contrast as high as 0.9, severely interfering with the true grayscale gradient information at particle edges.

[0003] In high-solids-content industrial crystallization scenarios, particles are in close contact, forming numerous aggregated particle clusters. The particle spacing in the image is extremely small, and the gray-level gradient is not obvious, making particle aggregation segmentation inherently challenging. Existing technologies typically employ thresholding and watershed algorithms to handle aggregated particles. These methods rely on clear gray-level gradient boundaries to accurately locate particle boundaries, placing high demands on image contrast. When speckle noise is superimposed on the particle edge region, the pseudo-gray-level gradient and the real particle edge gradient become confused, leading to segmentation boundary shifts, misjudgment of aggregated particles, and severe distortion of particle shape parameters and particle size distribution data. In other words, existing technologies suffer from a technical problem where laser speckle noise severely degrades the accuracy of aggregated particle segmentation in high-solids-content scenarios. Summary of the Invention

[0004] In view of this, the present invention provides a laser-illuminated microscopic imaging method and probe device, which can solve the technical problem that speckle noise caused by laser coherence leads to unstable image contrast in online microscopic imaging scenarios of industrial crystals with high solid content, resulting in a serious decrease in the accuracy of particle adhesion and segmentation.

[0005] This invention is implemented as follows: This invention provides a laser-illuminated microscopic imaging method, comprising the following steps:

[0006] A semiconductor laser generates a laser beam, which is coupled into a laser fiber via an optical fiber coupler. The laser beam is then transmitted through the laser fiber to a lens system. The laser beam is shaped into a parallel beam by a plano-convex lens, its propagation direction is changed by two right-angle reflections by a reflecting prism, and it is converged into a transparent slot by a convex lens to form transmissive illumination.

[0007] The rotating diffuser performs speckle time-domain averaging on the laser beam in the optical path of the lens system, reducing the speckle contrast to below the speckle contrast threshold. The semiconductor laser operates in pulsed laser stroboscopic illumination mode, with the illumination pulse width set to 0.02ms. The exposure time of the industrial camera is synchronously matched with the illumination pulse width.

[0008] The probe is placed in the solution to be tested so that the solution covers the permeable channel. The telecentric lens acquires images of tiny particles in the permeable channel. The industrial camera transmits the image information to the network switch via the camera data cable. The network switch then uploads the image information to the computer via the network cable.

[0009] The computer performs high-density particle adhesion segmentation processing on image information based on graph cut. The pixels in the image information are modeled as graph nodes. The edge weights are formed by the weighted sum of the pixel intensity gradient field, the distance transformation field and the orientation constraint. The particle convexity prior is introduced as a higher-order penalty term and added to the energy function. The maximum flow algorithm is used to solve the global optimal segmentation and obtain the particle shape parameters and particle size distribution data.

[0010] The computer uses Wasserstein distance to measure the difference between granular distribution data in two consecutive time windows, employs the Sinkhorn iterative algorithm to solve the discrete optimal transmission problem, constructs a granular distribution time series Wasserstein distance matrix, and uses a sliding window statistical test method to identify granular distribution drift events. When an abnormal drift is detected, the transmission plan matrix is ​​solved in reverse to locate the granularity range that caused the change in granular distribution.

[0011] The laser power adaptive control model collects the histogram statistics of the current frame image, the laser power current and solution concentration sensor readings, calculates the illumination quality assessment function value, and adjusts the output power of the semiconductor laser according to the interval to which the illumination quality assessment function value belongs, so that the image contrast is maintained within the optimal range.

[0012] The lens system consists of a plano-convex lens, a reflecting prism, and a convex lens. The plano-convex lens shapes the diverging laser beam emitted from the laser fiber into a parallel beam. The reflecting prism performs two right-angle reflections on the parallel beam to change its propagation direction. The convex lens focuses the parallel beam onto the test solution area within the transparent groove.

[0013] The speckle contrast is the ratio of the standard deviation to the mean of the random light intensity fluctuations generated by the coherent superposition of laser light on the imaging surface. The speckle contrast threshold is 0.1. The rotating diffuser is continuously rotated within the single-frame exposure time of the industrial camera, so that the speckle pattern at different times is averaged in the speckle time domain within the single-frame exposure integral.

[0014] The pulsed laser strobe illumination mode is an illumination mode in which the semiconductor laser emits light periodically with an illumination pulse width of 0.02ms. The exposure trigger signal of the industrial camera is synchronized with the pulse output signal of the semiconductor laser to ensure that image acquisition is completed only during the pulse emission of the semiconductor laser.

[0015] The high-density particle adhesion segmentation process based on graph cut models the particle image segmentation problem as a minimum graph cut energy optimization problem. In the graph, nodes correspond to image pixels, and the edge weights between adjacent nodes are composed of the weighted sum of three terms: pixel intensity gradient field, distance transformation field, and orientation constraint. The particle convexity prior is introduced as a higher-order penalty term and added to the energy function to penalize non-convex segmentation boundaries.

[0016] In particular, for images with high solids content and a large number of particles, a block-based parallel graph cutting strategy is adopted. The results of each block are merged with the consistency constraint of overlapping regions, so that the computational complexity of segmentation increases linearly with the number of particles.

[0017] The distance transformation field is a scalar field composed of the Euclidean distance from each pixel in the image information to the nearest background pixel, which is used to guide the segmentation boundary to shrink towards the grain edge; the grain convexity prior is to apply an energy penalty to the region where the segmentation result has a concave boundary, so that the segmentation boundary tends to be convex; the grain shape parameters include aspect ratio and roundness.

[0018] The Wasserstein distance defines the difference between two granularity distributions as the minimum total transmission cost required to move the mass of one distribution to the other. The Sinkhorn iterative algorithm converges to an approximately optimal transmission plan by alternately performing row and column normalization operations on the transmission plan matrix, with a computational complexity of O(n log n). ,in Number of particle size ranges is the regularization coefficient.

[0019] The particle size distribution time series Wasserstein distance matrix is ​​a matrix formed by calculating the Wasserstein distance between each pair of particle size distribution data in a continuous time window with time as the axis. The reverse solution of the transport plan matrix is ​​to locate the particle size range that causes the change in particle size distribution by analyzing the row and column positions of the mass transport volume concentration in the transport plan matrix.

[0020] It consists of three parts: a probe, a host, and a computer. The probe is made of Hastelloy as the shell material and integrates a telecentric lens, an industrial camera, a laser fiber, and a lens system. A transparent slot is opened on the side of the probe, and a sapphire glass window is installed on each side of the transparent slot. The upper window faces the lens system side, and the lower window faces the telecentric lens side.

[0021] The state space of the laser power adaptive control model is composed of the observation vector formed by splicing the histogram statistics of the current frame image, the laser power current and the solution concentration sensor readings. The action space is a discretized set of laser power adjustment values. The policy network uses a long short-term memory network module to encode the historical state sequence. The output layer connects the policy head and the value head of the dual-delay deep deterministic policy gradient algorithm.

[0022] The calculation formula for the lighting quality evaluation function is as follows: ,in This is the value of the lighting quality assessment function. The contrast of the current frame image. This is a standard reference value for image contrast. For the ratio of overexposed to underexposed pixels, This is the standard reference value for the overexposure / underexposure pixel ratio. and These are weighting coefficients. .

[0023] Among them, when When, maintain the current laser power and current unchanged; when When, increase the laser power current in small steps; when At that time, the laser power current is increased in medium step sizes; when At this time, the laser power adaptive control model takes over the control and increases the laser power current in large steps, while simultaneously triggering the adaptive adjustment process of the rotating diffuser speed.

[0024] The main unit is equipped with a semiconductor laser. The laser beam is coupled into the laser fiber through the fiber coupler and then transmitted to the lens system. The plano-convex lens in the lens system shapes the diverging beam into a parallel beam. After being reflected twice at right angles by the reflecting prism, the parallel beam is focused into the solution observation area in the transparent tank by the convex lens. The laser illumination main unit supports continuous adjustment of laser power in the range of 1mW to 1W.

[0025] The industrial camera uses a 1 / 1.8-inch complementary metal-oxide-semiconductor image sensor with a minimum exposure time of 0.03ms and a minimum pixel size of 2.20μm. The telecentric lens is connected to the industrial camera via a precision thread, and focusing is achieved by adjusting the number of turns of the precision thread to change the focal plane position of the telecentric lens.

[0026] The industrial camera transmits image data to the network switch in the host computer via a camera data cable. The network switch then uploads the data to the computer via a network cable. The computer performs closed-loop control based on graph cut high-density particle adhesion segmentation, particle size distribution temporal Wasserstein distance matrix construction, and laser power adaptive control model.

[0027] This invention addresses the technical problem of severely reduced particle adhesion segmentation accuracy in high-solid-content scenarios caused by laser speckle noise by introducing a rotating diffuser to perform speckle time-domain averaging on the laser beam and combining it with a graph-cut-based high-density particle adhesion segmentation algorithm. The rotating diffuser continuously rotates within a single frame exposure time, causing the laser wavefront phase distribution to change continuously over time. The speckle patterns at different times superimpose and cancel each other out within the single-frame integration, reducing the speckle contrast from the order of 0.9 for highly coherent lasers to below 0.1. This allows for a clear representation of the true grayscale gradient at particle edges, providing a high-quality input image for subsequent segmentation algorithms. In summary, this invention solves the technical problem mentioned in the background art where laser speckle noise severely reduces particle adhesion segmentation accuracy in high-solid-content scenarios. Attached Figure Description

[0028] Figure 1 This is a flowchart of the method of the present invention.

[0029] Figure 2 This is a graph showing the changes in particle size distribution data at different time windows during the crystallization process.

[0030] Figure 3 The graph shows the time-series Wasserstein distance matrix for particle size distribution and the results of abnormal drift detection.

[0031] Figure 4 This is a schematic diagram of the probe device in this embodiment.

[0032] The components represented by the reference numerals in the attached figures are as follows: 1. Probe; 2. Main unit; 3. Image acquisition system; 4. Lens system; 5. Transmission slot; 6. Laser fiber; 7. Plano-convex lens; 8. Reflecting prism; 9. Convex lens; 10. Telecentric lens; 11. Industrial camera; 12. Camera data cable; 13. Upper window lens; 14. Lower window lens; 15. Data cable interface; 16. Network cable; 17. Fiber optic interface; 18. Network switch; 19. Laser. Detailed Implementation

[0033] To make the objectives, technical solutions, and advantages of the embodiments of the present invention clearer, the technical solutions in the embodiments of the present invention will be clearly and completely described below.

[0034] like Figure 1 The diagram shown is a flowchart of a laser-illuminated microscopic imaging method provided by the first aspect of the present invention. This method includes the following steps:

[0035] S01. A semiconductor laser generates a laser beam, which is coupled into a laser fiber via a fiber coupler. The laser beam is then transmitted to a lens system via the laser fiber. The laser beam is shaped into a parallel beam by a plano-convex lens, its propagation direction is changed by two right-angle reflections by a reflecting prism, and it is converged into a transparent slot by a convex lens to form a transmissive illumination.

[0036] S02. The rotating diffuser performs speckle time-domain averaging on the laser beam in the optical path of the lens system, reducing the speckle contrast to below the speckle contrast threshold. The semiconductor laser operates in pulsed laser stroboscopic illumination mode, with the illumination pulse width set to 0.02ms. The exposure time of the industrial camera is synchronously matched with the illumination pulse width.

[0037] S03. The probe is placed in the solution to be tested so that the solution covers the transparent groove. The telecentric lens acquires images of tiny particles in the transparent groove. The industrial camera transmits the image information to the network switch via the camera data cable. The network switch then uploads the image information to the computer via the network cable.

[0038] S04. The computer performs high-density particle adhesion segmentation processing on the image information based on graph cut. The pixels in the image information are modeled as graph nodes. The edge weights are formed by the weighted sum of the pixel intensity gradient field, the distance transformation field and the orientation constraint. The particle convexity prior is introduced as a higher-order penalty term and added to the energy function. The maximum flow algorithm is used to solve the global optimal segmentation and obtain the particle shape parameters and particle size distribution data.

[0039] S05. The computer uses Wasserstein distance to measure the difference between granular distribution data in two consecutive time windows, uses the Sinkhorn iterative algorithm to solve the discrete optimal transmission problem, constructs the granular distribution time series Wasserstein distance matrix, uses the sliding window statistical test method to identify granular distribution drift events, and when abnormal drift is detected, reversely solves the transmission plan matrix to locate the granularity range that causes the change in granular distribution.

[0040] S06. The laser power adaptive control model collects the histogram statistics of the current frame image, the laser power current and solution concentration sensor readings, calculates the illumination quality evaluation function value, and adjusts the output power of the semiconductor laser according to the interval to which the illumination quality evaluation function value belongs, so that the image contrast is maintained within the optimal range.

[0041] The lens system consists of a plano-convex lens, a reflecting prism, and a convex lens. The plano-convex lens shapes the diverging laser beam emitted from the laser fiber into a parallel beam. The reflecting prism performs two right-angle reflections on the parallel beam to change its propagation direction. The convex lens focuses the parallel beam onto the test solution area within the transparent groove.

[0042] The speckle contrast refers to the ratio of the standard deviation to the mean of the random light intensity fluctuations generated by the coherent superposition of laser light on the imaging surface; the speckle contrast threshold is 0.1; the rotating diffuser continuously rotates within the single-frame exposure time of the industrial camera, causing the phase distribution of the laser wavefront to change continuously in the time dimension. The speckle patterns at different times are averaged in the speckle time domain within the single-frame exposure integration, reducing the speckle contrast to below the speckle contrast threshold without reducing the particle imaging resolution.

[0043] The pulsed laser strobe illumination mode refers to the illumination mode in which the semiconductor laser emits light periodically with an extremely short illumination pulse width; the synchronous matching of the industrial camera exposure time and the illumination pulse width means that the exposure trigger signal of the industrial camera is synchronized with the pulse output signal of the semiconductor laser, ensuring that image acquisition is completed only during the period when the semiconductor laser emits pulses.

[0044] The high-density particle adhesion segmentation process based on graph cut models the particle image segmentation problem as a minimum graph cut energy optimization problem. Nodes in the graph correspond to image pixels, and the edge weights between adjacent nodes are composed of a weighted sum of the pixel intensity gradient field, the distance transformation field, and the orientation constraint. A priori particle convexity is introduced as a higher-order penalty term added to the energy function to penalize non-convex segmentation boundaries. The maximum flow algorithm is used to solve for the globally optimal segmentation. For images with high solids content and a large number of particles, a block-parallel graph cut strategy is adopted, merging the results of each block with consistency constraints in overlapping regions, making the segmentation computational complexity linearly expandable with the number of particles. The distance transformation field refers to the distance in the image information... The scalar field formed by the Euclidean distance from each pixel to the nearest background pixel is used to guide the segmentation boundary to shrink towards the particle edge; the particle convexity prior refers to applying an energy penalty to the region where a concave boundary appears in the segmentation result, so that the segmentation boundary tends to be convex, in order to conform to the geometric shape characteristics of crystal particles; the particle shape parameters include aspect ratio and roundness; the high-density particle adhesion segmentation processing based on graph cut distinguishes adhered particles through global energy optimization, and can accurately recover the independent contour of each particle in high solid content scenarios with extremely small particle spacing and insignificant brightness gradient. The segmentation result directly reflects the true morphology of the particles, providing accurate particle boundary information for subsequent particle shape parameter calculation.

[0045] The Wasserstein distance defines the difference between two granularity distributions as the minimum total transmission cost required to move the mass of one distribution to the other. The Sinkhorn iterative algorithm converges to an approximately optimal transmission plan by alternately performing row and column normalization operations on the transmission plan matrix, with a computational complexity of O(n log n). ,in Number of particle size ranges The regularization coefficient is used. The particle size distribution time series Wasserstein distance matrix refers to the matrix formed by calculating the Wasserstein distance between each pair of particle size distribution data in a continuous time window with time as the axis. The transport plan matrix refers to the matrix obtained after the convergence of the Sinkhorn iterative algorithm, which describes the mass transport scheme between two particle size distribution data. The reverse solution of the transport plan matrix is ​​to locate the particle size range that causes the particle size distribution change by analyzing the row and column positions of the mass transport volume concentration in the transport plan matrix. The particle size distribution change detection based on the optimal transport theory captures the overall morphological change of the particle size distribution data through the geometrically clear optimal transport cost. It can stably detect particle size distribution drift in industrial crystallization scenarios with wide particle size distribution and sparse particle size range, and promptly detect crystallization process anomalies, providing real-time quantitative basis for process adjustment.

[0046] The specific structure of the laser power adaptive control model is as follows: the state space is composed of the current frame image histogram statistics, laser power current, and solution concentration sensor readings to form an observation vector; the action space is a discretized set of laser power adjustment values; the policy network uses a long short-term memory network module to encode historical state sequences to capture the trend of crystal precipitation rate changes; and the output layer connects the policy head and value head of the dual-delay deep deterministic policy gradient algorithm. The steps for establishing the training dataset for the laser power adaptive control model specifically include: acquiring probe image sequences and corresponding solution concentration sensor readings under different solution concentrations, different solid contents, and different combinations of laser power current; labeling each frame image with image contrast values ​​and overexposure / underexposure pixel ratios; constructing state transition samples using frame sequences as time steps; and establishing a training dataset covering a wide range of solution concentrations and laser power currents. The steps for training the laser power adaptive control model specifically include: defining a reward function using the weighted negative values ​​of image contrast and overexposure / underexposure pixel ratios; and using... The dual-delay deep deterministic policy gradient algorithm alternately updates the policy network and the value network. It suppresses overestimation of value during training through delayed policy updates and a soft update mechanism for the target network, training until the average reward of the policy network converges on the validation set. The adaptive laser power control model encodes the trends of laser power current and image contrast changes in historical frame sequences through a long short-term memory network module. This allows the policy network to increase the laser power current output in advance during the rapid increase in solid content and decrease the laser power current in advance during the decrease in solid content to avoid overexposure. Feedforward predictive dimming reduces fluctuations in image contrast during process state switching, ensuring the stability of image contrast required for particle shape parameters and particle size distribution data analysis. Simultaneously, precise control of laser power current reduces unnecessary laser energy consumption and extends the lifespan of the semiconductor laser. The overexposure / underexposure pixel ratio refers to the ratio of the sum of the number of pixels with gray values ​​exceeding the overexposure threshold and the number of pixels with gray values ​​below the underexposure threshold to the total number of pixels in the image information.

[0047] The calculation formula for the lighting quality evaluation function is as follows:

[0048] ;

[0049] in This is the value of the lighting quality assessment function. The contrast of the current frame image. This is a standard reference value for image contrast. For the ratio of overexposed to underexposed pixels, This is the standard reference value for the overexposure / underexposure pixel ratio. and These are weighting coefficients. ;when When, maintain the current laser power and current unchanged; when When, increase the laser power current in small steps; when At that time, the laser power current is increased in medium step sizes; when At this time, the laser power adaptive control model takes over control and increases the laser power current in large steps, while simultaneously triggering the adaptive adjustment process of the rotating diffuser speed. The adaptive adjustment process of the rotating diffuser speed refers to calculating the current speckle contrast based on the histogram statistics of the current frame image. When the speckle contrast is higher than the speckle contrast threshold, the rotating diffuser speed is increased until the speckle contrast decreases below the speckle contrast threshold. The image contrast standard reference value... Standard reference value for the ratio of overexposed to underexposed pixels These are reference values ​​obtained statistically from image information acquired under standard solution concentration and standard laser power and current conditions.

[0050] The probe is made of Hastelloy, which is corrosion-resistant, pressure-resistant, temperature-resistant, and explosion-proof, allowing it to be directly inserted into the solution to be tested for online observation. The windows on both sides of the transparent slot are made of sapphire glass. The industrial camera is a 1 / 1.8-inch complementary metal-oxide-semiconductor image sensor with a minimum exposure time of 0.03ms and a minimum pixel size of 2.20μm. The telecentric lens is connected to the industrial camera via a precision thread, and focusing is achieved by adjusting the number of turns of the precision thread to change the focal plane position of the telecentric lens. The laser power current is the driving current of the semiconductor laser, and the output power of the semiconductor laser is changed by adjusting the laser power current, with an adjustment range of 1mW to 1W.

[0051] A second aspect of this invention provides a laser-illuminated probe device, comprising a probe, a main unit, and a computer. The probe uses Hastelloy as its housing material and integrates a telecentric lens, an industrial camera, a laser fiber, and a lens system. A transparent groove is formed on the side of the probe, with a sapphire glass window installed on each side of the groove. The upper window faces the lens system side, and the lower window faces the telecentric lens side. The transparent groove allows the solution to be tested to flow naturally into the observation area after the probe is inserted.

[0052] The illumination optical path is generated by a semiconductor laser within the main unit. The laser beam is coupled into a laser fiber via a fiber coupler and then transmitted to a lens system inside the probe. In the lens system, a plano-convex lens first shapes the diverging beam from the laser fiber into a parallel beam. This parallel beam then enters a reflecting prism, undergoes two right-angle reflections within the prism, changes its propagation direction to be perpendicular to the plane of the transparent groove, and finally, a convex lens converges the parallel beam onto the solution observation area within the transparent groove, forming transmissive illumination. The laser illumination main unit supports continuous adjustment of the laser power within the range of 1mW to 1W, and also supports a pulsed laser strobe illumination mode with a minimum pulse width of 0.02ms.

[0053] To eliminate speckle noise caused by laser coherence, a rotating diffuser is added to the optical path of the lens system. The rotating diffuser rotates continuously within the single-frame exposure time of the industrial camera, causing the laser wavefront phase distribution to change continuously over time. The speckle pattern at different times is time-domain averaged within the single-frame exposure integration, reducing the speckle contrast from 0.9 for highly coherent lasers to below 0.1, without affecting particle imaging resolution. To address motion blur caused by rapid particle movement within the industrial crystallizer, the system employs a pulsed laser stroboscopic illumination mode, compressing the illumination pulse width to 0.02 ms. The industrial camera's exposure trigger signal is synchronized with the laser pulse, ensuring image acquisition is completed only during the laser pulse emission period. This reduces the displacement of particles with a flow velocity of 0.5 m / s to the order of 10 nm within a single exposure, completely eliminating motion blur.

[0054] Image acquisition is accomplished by a telecentric lens and an industrial camera. The telecentric lens's focal plane is aligned with the solution observation area within the transparent tank, acquiring clear images of minute particles and projecting them onto the industrial camera's complementary metal-oxide-semiconductor (CMOS) image sensor. The industrial camera transmits the image data to a network switch within the host computer via a camera data cable, and the network switch then uploads the data to a computer via a network cable.

[0055] The computer first performs a high-density particle adhesion segmentation algorithm based on graph cut on the acquired images. Image pixels are modeled as graph nodes, with edge weights composed of a weighted average of pixel intensity gradient field, distance transformation field, and orientation constraint. Particle convexity prior is introduced as a higher-order penalty term. The maximum flow algorithm is used to solve for the globally optimal segmentation, accurately separating adherent particles from high-solids-content images. The independent contour of each particle is extracted, and particle shape parameters such as aspect ratio and roundness, as well as particle size distribution data, are calculated. To address the large number of particles in high-solids-content scenarios, a block-parallel graph cut strategy is adopted, merging block results with overlapping region consistency constraints, making the computational complexity linearly expandable with the number of particles.

[0056] After acquiring the granularity distribution data, the system uses the Wasserstein distance to measure the difference in granularity distribution between two consecutive time windows, and employs the Sinkhorn iterative algorithm to efficiently solve the discrete optimal transmission problem, with a computational complexity of O(n log n). The system constructs a time-series Wasserstein distance matrix for particle size distribution and continuously monitors particle size distribution drift events using a sliding window statistical test method, distinguishing between expected changes caused by process parameter adjustments and abnormal changes caused by equipment failures. When abnormal drift is detected, the system reverse-engineers the transport plan matrix to locate the particle size range causing the distribution change, providing operators with quantitative evidence for fault diagnosis.

[0057] The laser power closed-loop control is implemented by a laser power adaptive adjustment model. This model collects the histogram statistics of the current frame image, laser power current, and solution concentration sensor readings. It encodes historical state sequences through a long short-term memory network to capture the trend of crystal precipitation rate changes, and outputs the laser power adjustment amount using a dual-delay depth deterministic gradient algorithm. The illumination quality evaluation function calculates a comprehensive evaluation value based on image contrast, overexposed pixel ratio, and underexposed pixel ratio. When the evaluation value falls into different ranges, power adjustment strategies of maintaining, small step increase, medium step increase, and large step increase are adopted respectively. When the evaluation value falls below a minimum threshold, the rotation speed of the diffusing plate is adaptively adjusted to ensure that the image quality always meets the requirements of grain size analysis. The entire system realizes a complete online microscopic imaging process from illumination source generation, speckle suppression, motion blur elimination, image acquisition and transmission, particle segmentation and recognition, particle size distribution monitoring to adaptive closed-loop control of illumination power.

[0058] The specific implementation of step S01 is as follows: A semiconductor laser serves as the illumination source. Its output laser light is coupled into a laser fiber after spatial mode matching via a fiber coupler. The laser fiber transmits the laser energy from the host to the lens system entrance inside the probe. In the lens system, the plano-convex lens has its flat surface facing the laser fiber's output end and its curved surface facing the beam propagation direction. The curved surface refraction collimates the divergent conical beam emitted from the laser fiber into a parallel beam with a uniform cross-section. The parallel beam then enters a reflecting prism, undergoing two total internal reflections inside the prism. Each reflection deflects the beam propagation direction by 90°, and the combination of the two right-angle reflections deflects the beam by a total of 180°, converting the horizontally propagating beam into a direction perpendicular to the plane of the transparent slot, thus adapting to the probe's lateral slot geometry. The direction-converted parallel beam is converged by the convex lens, focusing on the solution observation area within the transparent slot. This forms a concentrated, transmissive illumination spot within the area, providing uniform backlight transmitted from below to the telecentric lens, allowing the particles to be clearly presented in a dark-field outline form.

[0059] The specific implementation of step S02 is as follows: The rotating diffuser is driven by a stepper motor and installed in the parallel beam propagation section of the lens system's optical path. The surface of the diffuser has randomly distributed micro-rough structures. When the laser beam penetrates the diffuser, the micro-rough structures introduce random phase modulation onto the laser wavefront, generating a speckle pattern related to the current rotation angle of the diffuser. Since the diffuser rotates continuously within the single-frame exposure time of the industrial camera, the rotation angle changes continuously, and the phase modulation state is different at each moment. The corresponding speckle patterns are statistically independent of each other in space. During the single-frame exposure integration process, the sensor accumulates the superposition of speckle patterns at all moments throughout the exposure time, which is equivalent to achieving temporal averaging of a large number of independent speckles. The ratio of the standard deviation to the mean of the random light intensity fluctuations, i.e., the speckle contrast, decreases from the order of 0.9 for high-coherence lasers to below the speckle contrast threshold of 0.1. Meanwhile, the system adopts a pulsed laser strobe illumination mode, which compresses the illumination pulse width of the semiconductor laser to 0.02ms. The exposure trigger signal of the industrial camera is hardware synchronized with the laser pulse output signal, ensuring that the sensor completes image integration only during the laser pulse emission period. The extremely short equivalent exposure time reduces the displacement of high-speed moving particles in a single exposure to the nanometer level, thereby eliminating motion blur.

[0060] The specific implementation of step S03 is as follows: After the probe is inserted into the solution to be tested, the open area between the sapphire glass windows on both sides of the transparent groove is naturally filled by the solution to be tested, and the suspended particles flow into the observation area of ​​the transparent groove with the solution. The telecentric lens adopts an object-side telecentric optical path design, with its entrance pupil located at infinity. The imaging magnification does not change with the object distance, ensuring that particles at different depths within the transparent groove are mapped to the sensor plane with the same geometric size, eliminating the influence of perspective distortion on particle shape parameters. The telecentric lens projects the particle image onto the complementary metal-oxide-semiconductor image sensor of the industrial camera. After the industrial camera completes image acquisition, the image data is transmitted to the network switch inside the host computer via the camera data cable. The network switch then uploads the image information to the computer via a network cable for use by subsequent image processing algorithms.

[0061] The specific implementation of step S04 is as follows: After receiving the image information, the computer models the particle image segmentation problem as a minimum graph cut energy optimization problem. Each pixel in the image corresponds to a node in the graph, and the edge weight between adjacent pixel nodes is composed of a weighted sum of three terms: the pixel intensity gradient field reflects the degree of gray-level abrupt change at the particle edge, with smaller edge weights at locations with larger gradients, guiding the cut surface to preferentially pass through gray-level abrupt change regions; the distance transformation field provides Euclidean distance information from each pixel to the nearest background pixel, giving the cut surface high edge weights in regions inside the particle that are far from the edge, constraining the cut surface to shrink towards the true edge of the particle; the orientation constraint term assigns low weights to edges along the tangent direction of the particle contour, causing the cut surface to extend along the contour tangent and inhibiting the cut surface from traversing the interior of the particle. The energy function introduces a priori particle convexity as a higher-order penalty term, applying additional energy costs to regions with concave boundaries in the segmentation result, causing the segmentation boundary to tend towards convexity, which conforms to the geometric shape characteristics of crystal particles. Finally, the maximum flow algorithm is used to solve the globally optimal graph cut problem. This algorithm is dual to the minimum cut problem; the value of the maximum flow equals the cost of the minimum cut. The globally optimal segmentation result is obtained by iterating through augmented paths until the network saturates. For high-solids-content images with a large number of particles, a block-parallel graph cut strategy is adopted. The image is divided into several sub-blocks with overlapping boundaries, and graph cuts are performed independently on each sub-block. The segmentation results of overlapping regions are merged with consistency constraints, making the overall computational complexity linearly expandable with the number of particles. After segmentation, particle shape parameters such as aspect ratio and roundness are extracted from the independent contour of each particle, and the equivalent particle size distribution of all particles in the image is statistically analyzed, outputting the particle size distribution data.

[0062] The specific implementation of step S05 is as follows: The computer segments the continuously collected granular distribution data into segments using fixed time windows. The Wasserstein distance is calculated for the granular distribution data in two adjacent time windows. The physical meaning of the Wasserstein distance is the minimum total transmission cost required to move the mass of one distribution to another. Its measurement method considers the geometric distance relationship between particle size intervals and has high sensitivity to overall changes in distribution morphology. The Sinkhorn iterative algorithm converges to an approximately optimal transmission plan under entropy regularization constraints by alternately performing row and column normalization operations on the transmission plan matrix. The computational complexity is O(n log n). ,in Number of particle size ranges The regularization coefficient is between 0.01 and 0.1. A smaller regularization coefficient results in higher approximate accuracy but requires more iterations. The Wasserstein distances between all pairs of time windows are arranged to form a granularity distribution time-series Wasserstein distance matrix. A sliding window statistical test is used to scan the matrix to check if the Wasserstein distance sequence exceeds the statistical test threshold. The threshold reference value is determined based on the 95th quantile of the Wasserstein distance distribution under historical normal operating conditions. When an abnormal drift event is detected, the Sinkhorn convergence transport plan matrix for the corresponding time window is retrieved. The row and column positions of the mass transfer concentration in the matrix are analyzed. The particle size intervals corresponding to these row and column positions are the key particle size intervals causing changes in particle size distribution, providing operators with quantitative fault diagnosis information.

[0063] The specific implementation of step S06 is as follows: The laser power adaptive control model runs in real time after each frame of image acquisition. The input observation vector is composed of three parts: the histogram statistics of the current frame image, the current laser power current, and the solution concentration sensor reading. The long short-term memory network module of the policy network receives the observation vector of the historical frame sequence, extracts the changing trends of laser power current and image contrast over time through a gating mechanism, and outputs the hidden state vector, which is then fed into the policy head of the dual-delay deep deterministic policy gradient algorithm. The policy head outputs the laser power adjustment amount. Before the policy network outputs the laser power adjustment amount, the illumination quality evaluation function value is calculated first. , Image contrast standard reference value Standard reference value for the ratio of overexposed to underexposed pixels The weighting coefficients were obtained statistically after images were acquired under standard solution concentration and standard laser power / current conditions. and satisfy The reference value is , .when When the current laser power and current remain constant; when Increase the laser power current in small steps; when Increase the laser power current with medium step size; when The laser power adaptive control model takes over the control and increases the laser power current in large steps. At the same time, it triggers the adaptive adjustment process of the rotating diffuser speed. The speckle contrast is calculated in real time based on the histogram statistics of the current frame image. When the speckle contrast is higher than the threshold of 0.1, the rotating diffuser speed is continuously increased until the speckle contrast is reduced to below the threshold, thereby maintaining stable image contrast during the rapid change of solution concentration and solid content.

[0064] It should be explained that the first key technical concept of this invention is the temporal averaging mechanism of speckle patterns using a rotating diffuser. Traditional laser illumination schemes directly irradiate the imaging area with highly coherent laser light, resulting in speckle contrast close to the theoretical upper limit, and spurious gradient information severely interfering with particle edge recognition. The rotating diffuser, by continuously rotating within a single frame exposure time, superimposes and averages the temporally independent speckle patterns in the integral domain. Utilizing statistical independence, random light intensity fluctuations cancel each other out, achieving significant compression of speckle contrast without reducing the imaging spatial resolution, thus accurately representing the true grayscale gradient of particle edges.

[0065] The second key technical approach is a global energy optimization segmentation method based on graph cut. Traditional watershed algorithms rely on local gray-level extrema, which can easily lead to over-segmentation or under-segmentation in high-solid-content scenarios due to small particle spacing. The graph cut method models the segmentation problem as a global energy minimization problem, and solves for the globally optimal cut surface in one go using the maximum flow algorithm. It introduces a priori particle convexity to make the cut surface tend to a convex boundary that conforms to the crystal geometry, thus accurately recovering the independent contour of each particle even in scenarios where the gray-level gradient between particles is extremely insignificant.

[0066] The synergistic effect of the two technical approaches is reflected in the following aspects: the speckle temporal averaging provides a high-quality gradient field input for the graph cut algorithm, enabling the pixel intensity gradient field in the edge weight construction to accurately reflect the true boundary of the particles and eliminate the interference of speckle pseudo-gradients on the graph cut energy function. The global optimization capability of the graph cut further utilizes the distance transformation field and convexity prior to robustly suppress the influence of speckle residual noise. The two complement each other in terms of signal quality improvement and algorithm robustness, jointly ensuring the accuracy of particle adhesion segmentation in high solid content scenarios.

[0067] It should be noted that this invention also solves the following technical problem: In online monitoring scenarios of industrial crystallization with high solids content, existing particle size distribution monitoring methods typically use statistics based on histogram differences or KL divergence to detect changes in particle size distribution. These methods treat particle size intervals as mutually independent discrete probability points, ignoring the geometric distance information between particle size intervals. This results in insufficient sensitivity of the statistics when particle size distribution undergoes changes with clear physical meaning, such as morphological translation or widening, making it difficult to stably detect particle size distribution drift events in sparse particle size interval scenarios. This invention defines particle size distribution difference as the optimal transmission cost through Wasserstein distance. Its metric naturally incorporates the geometric distance between particle size intervals, enabling global changes such as morphological translation to produce a distance response that matches their physical meaning. Combined with the Sinkhorn iterative algorithm, it efficiently solves the discrete optimal transmission problem, thus solving the technical problem of insufficient sensitivity in particle size distribution drift detection in sparse particle size interval scenarios.

[0068] This invention also solves the technical problem of the difficulty in real-time adaptive adjustment of laser illumination power in industrial crystallization processes due to dynamic changes in solution concentration and solid content. Existing technologies typically employ fixed laser power or simple feedback control based on the average grayscale value of a single frame image, which cannot predict rapid changes in solid content, leading to frequent overexposure or underexposure of images during process state transitions. This invention encodes the laser power current and image contrast change trends in historical frame sequences using a long short-term memory network. This allows the strategy network to increase the laser power current in advance when the solid content begins to rise and decrease the power in advance when the solid content decreases to avoid overexposure. This achieves closed-loop laser power control with feedforward prediction capabilities, solving the technical problem of the difficulty in real-time adaptive adjustment of laser illumination power to follow dynamic changes in solution state.

[0069] Specifically, the principle of this invention is as follows: The fundamental reason why this invention can solve the above-mentioned technical problems lies in the inherent logical consistency between the speckle temporal averaging mechanism and the global energy optimization segmentation method. Speckle noise is essentially an interference pattern formed by the spatial random distribution of the laser wavefront phase on the imaging surface. When the rotating diffuser rotates continuously within a single frame exposure time, the random phase structures of different regions on the diffuser surface participate in wavefront modulation sequentially, making the speckle pattern at each moment spatially independent. Single-frame integration is equivalent to temporal averaging a large number of independent speckle patterns. The randomly fluctuating light intensity tends to be uniform in a statistical sense, and the speckle contrast is significantly reduced. After speckle suppression, the gray-level gradient field of the particle edges is restored to its true form. The graph-cut-based segmentation algorithm uses a weighted sum of the pixel intensity gradient field, distance transformation field, and orientation constraint to form the edge weights, and introduces a priori particle convexity as a higher-order penalty term. The global optimal segmentation is solved using the maximum flow algorithm, which can accurately recover the independent contour of each particle in high-solids content scenarios with extremely small particle spacing, providing accurate particle boundary information for particle shape parameters and particle size distribution data.

[0070] The following provides a specific embodiment 1 of the present invention, and the specific implementation of each step in this embodiment 1 is described in detail below.

[0071] The specific implementation of step S01 is as follows: A semiconductor laser generates laser light, which is coupled into a laser fiber via an optical fiber coupler. The laser light is then propagated along the laser fiber to a lens system inside the probe. In the lens system, a plano-convex lens first shapes the diverging beam emitted from the laser fiber into a parallel beam. The parallel beam then enters a reflecting prism, undergoes two right-angle reflections inside the prism to change its propagation direction, and finally is converged by a convex lens onto the area of ​​the solution to be tested within the transparent groove, forming transmissive illumination.

[0072] The specific implementation of step S02 is as follows: A rotating diffuser is placed in the optical path of the lens system and continuously rotated within the single-frame exposure time of the industrial camera, causing the laser wavefront phase distribution to change continuously in the time dimension. The speckle pattern at different times is averaged in the time domain within the single-frame exposure integration, reducing the speckle contrast to below the threshold. (Speckle contrast) The calculation formula is expressed as follows:

[0073] ;

[0074] In the formula, This represents the standard deviation of the light intensity distribution on the imaging surface, expressed in grayscale values. This represents the mean value of the light intensity distribution on the imaging surface, expressed in grayscale values. The ratio of the two is a dimensionless quantity, and the speckle contrast threshold is set to 0.1. The semiconductor laser operates in pulsed laser stroboscopic illumination mode, with the illumination pulse width set to 0.02 ms. The industrial camera's exposure trigger signal is synchronized with the laser pulse output signal to ensure that image acquisition is completed only during the laser pulse emission period, thereby eliminating motion blur.

[0075] The specific implementation of step S03 is as follows: After the probe is inserted into the solution to be tested, the solution to be tested flows naturally into the transparent tank. The focal plane of the telecentric lens is aligned with the observation area of ​​the solution in the transparent tank to collect images of tiny particles. The industrial camera transmits the image data to the network switch in the host computer via the camera data cable. The network switch then uploads the data to the computer via the network cable.

[0076] The specific implementation of step S04 is as follows: The computer performs high-density particle adhesion segmentation processing based on graph cut on the acquired image. Each pixel in the image is modeled as a graph node, and the edge weights between adjacent pixel nodes are... It consists of a weighted sum of the pixel intensity gradient field, the distance transformation field, and the orientation constraint, as expressed in the following formula:

[0077] ;

[0078] In the formula, and pixels With pixels The grayscale gradient magnitude, expressed in grayscale value per pixel; It is the absolute value of the difference between the two, in grayscale value per pixel; The gradient magnitude is a standard reference value, obtained statistically from a standard image, with units of grayscale values ​​per pixel. It is a dimensionless quantity; For pixels The Euclidean distance to the nearest background pixel, i.e., the distance transformation field at the pixel. The value at that location is in pixels; The distance transformation field standard reference value is in pixels, making It is a dimensionless quantity; the distance transformation field term is taken as This makes the distance transformation value at the particle edge smaller, thus resulting in a larger weight, and guides the segmentation boundary to shrink towards the particle edge; Adjacent pixels and The angle between the direction of the connecting line and the normal vector of the particle boundary, in radians; Map the orientation constraints to The interval is a dimensionless quantity. , , These are the weighting coefficients for the gradient field, the distance transformation field, and the orientation constraint, respectively. All three are dimensionless, with empirical values ​​of 0.4, 0.4, and 0.2, respectively. ; It is a weighted sum of three dimensionless quantities, and the whole is a dimensionless quantity; and This represents the index of a neighboring pixel in the image. Energy function. The particle convexity prior is introduced as a higher-order penalty term, and the formula is expressed as follows:

[0079] ;

[0080] In the formula, The edge set of the graph is the set of all adjacent pixel pairs. This is an indicator function; it takes the value 1 when the condition inside the parentheses is true, and 0 otherwise. For pixels The segmentation labels are 0 for background and 1 for particles; is the convexity prior penalty coefficient, which is a dimensionless quantity with an empirical value of 0.15; For the first The perimeter of the concave boundary segment in each particle segmentation contour, in pixels; For the first The total perimeter of the segmented contour of each particle, in pixels; It is a dimensionless ratio; For granular indexing; The entire quantity is dimensionless. The maximum flow algorithm is used to solve for the globally optimal segmentation of the energy function, obtaining the independent profile of each particle. Then, the particle shape parameters and aspect ratio are calculated. with roundness The calculation formulas are expressed as follows:

[0081] ;

[0082] ;

[0083] In the formula, The length of the minor axis of the smallest circumscribed ellipse of the particle, in units of ; The length of the minor axis of the smallest circumscribed ellipse of the particle, in units of ; It is the ratio of the two, and is a dimensionless quantity; The area enclosed by the particle outline is expressed in units of... ; Perimeter of particle outline, in units of ; Middle molecule Units are denominator Units are , It is a dimensionless quantity; here, the perimeter of the particle outline is denoted as... With respect to the particle convexity penalty term , To differentiate between them, a block-based parallel graph cut strategy is adopted for high solids content scenarios. The results of each block are merged with the consistency constraint of overlapping regions, so that the computational complexity of segmentation increases linearly with the number of particles.

[0084] The specific implementation of step S05 is as follows: The computer uses the Wasserstein distance to measure the difference between the granular distribution data of two consecutive time windows. Let the granular distribution data of the two time windows be discrete probability vectors. and ,in and They are the particle size ranges Particle size range The corresponding particle number percentages are all dimensionless quantities and satisfy the following conditions: , ; This represents the total number of particle size ranges. Particle size cost matrix. The definition is as follows:

[0085] ;

[0086] In the formula, For the first The representative particle size values ​​for each particle size range, in units of ; Particle size range Particle size range The absolute value of the particle size difference between them, in units of ; for The symmetric matrix is ​​used. The Sinkhorn iterative algorithm (i.e., the optimal transmission iterative solution algorithm with entropy regularization) is employed to solve the discrete optimal transmission problem. The regularized optimal transmission objective function is expressed as follows:

[0087] ;

[0088] In the formula, For the transmission plan matrix, To be from the particle size range Towards particle size range The mass percentage of the transported goods is a dimensionless quantity. To meet marginal constraints and The feasible domain of the transmission plan, in which for A column vector of all 1s; This is a standard reference value for particle size, in units of... ,make It is a dimensionless quantity; is the regularization coefficient, a dimensionless quantity, typically ranging from 0.01 to 0.1; middle It is a dimensionless quantity. It is also a dimensionless quantity, and the entropy regularization term is dimensionless overall; The regularized Wasserstein distance approximation is a dimensionless quantity. The Sinkhorn iterative algorithm iteratively applies the transfer plan matrix... Performing row and column normalization operations converges to an approximately optimal transmission plan, with a computational complexity of O(n log n). After convergence, the optimal transmission plan matrix is ​​obtained. and the approximate value of Wasserstein distance Construct the temporal Wasserstein distance matrix for granular distribution. , for A symmetric matrix, whose th... The element is defined as follows:

[0089] ;

[0090] In the formula, and The first With the Discrete probability vectors of granular distribution within a time window; Total number of time windows; and For time window indexing, ; It is a dimensionless quantity. The sliding window statistical test method is used to... Continuous monitoring is conducted to identify granularity distribution drift events. When abnormal drift is detected, analysis is performed... The particle size range in which the position of the row and column where the mass transport volume is concentrated causes changes in particle size distribution.

[0091] The specific implementation of step S06 is as follows: The laser power adaptive control model collects the histogram statistics of the current frame image, the laser power current and solution concentration sensor readings, and calculates the illumination quality assessment function value. The formula is expressed as follows:

[0092] ;

[0093] In the formula, The value of the lighting quality assessment function is a dimensionless quantity. The contrast of the current frame image, in grayscale values, is calculated by multiplying the ratio of the standard deviation to the mean of the grayscale values ​​of the current frame image by the mean. The image contrast standard reference value is expressed in grayscale values ​​and is obtained by statistical analysis of images acquired under standard solution concentration and standard laser power and current conditions. It is a dimensionless ratio; The ratio of overexposed to underexposed pixels is a dimensionless quantity, and its calculation formula is expressed as follows:

[0094] ;

[0095] In the formula, For images where grayscale values ​​exceed the overexposure threshold The number of pixels, The empirical value is 240 (grayscale range 0-255); For images with grayscale values ​​below the underexposure threshold The number of pixels, The experience value is 15; This represents the total number of pixels in the image. , , All units are pixels. It is a ratio of the three, and is a dimensionless quantity; The overexposure / underexposure pixel ratio is a standard reference value, obtained by statistical analysis of images acquired under standard solution concentration and standard laser power current conditions, and is a dimensionless quantity. It is a dimensionless ratio; and These are weighting coefficients, all dimensionless, with empirical values ​​of 0.6 and 0.4 respectively. ; It is obtained by weighted combination of two dimensionless ratios, and the whole is a dimensionless quantity. When When, maintain the current laser power and current unchanged; when When, increase the laser power current in small steps; when At that time, the laser power current is increased in medium step sizes; when At this time, the laser power adaptive control model increases the laser power current in large steps, while simultaneously triggering the adaptive adjustment process of the rotating diffuser speed, based on the current speckle contrast. If the value is higher than the threshold of 0.1, gradually increase the rotation speed of the diffuser until... Reduce the image quality to below the threshold to ensure that the image quality always meets the requirements for particle shape parameters and particle size distribution data analysis.

[0096] Regarding the probe device provided in the second aspect of the present invention, such as Figure 4As shown, the device includes a probe 1 and a host 2. A through groove 5 is provided on one side of the surface of the probe 1. A lens system 4 is provided on one side of the through groove 5. A telecentric lens 10 is provided on the other side of the through groove 5. One end of the telecentric lens 10 is connected to one end of an industrial camera 11 by a thread. The industrial camera 11 is fixed inside the probe 1. The camera data cable 13 of the industrial camera 11 is connected to the data cable interface 15 of the host 3. A laser fiber 6 is fixed inside the probe 1. One end of the laser fiber 6 is connected to the lens system 4. The other end of the laser fiber 6 is connected to the fiber optic interface 17 of the host 3. The fiber optic interface 17 is connected to the laser 19 inside the host 3.

[0097] Among them, host 3 is the laser lighting host. Laser lighting host: LA-EP-1; Laser model: LC-4; Network switch model: NS-I-2.

[0098] A laser illumination host is a device capable of providing controllable laser illumination while simultaneously transmitting image data. In this invention, the laser illumination host is primarily used to receive and process image information transmitted from the probe, while also emitting laser light for probe illumination.

[0099] Telecentric lenses can use any of the following magnifications: 0.2, 0.5, 0.75, 1, 2, 3, 4, 6, 8, 10.

[0100] In the above technical solution, the camera data cable 13 of the industrial camera 11 and the laser fiber 6 are wrapped by a protective tube after they come out from the rear end of the probe 1, protecting the lines from wear.

[0101] Furthermore, in the above technical solution, the host 2 has a built-in network switch 18 and a laser 19. The network switch 18 and the laser 19 are connected in communication. The network switch 18 is connected to the computer through the network cable interface 16.

[0102] Furthermore, in the above technical solution, the laser fiber 6 is communicatively connected to the laser 19 inside the host 3 via the fiber optic interface 17.

[0103] Furthermore, in the above technical solution, the industrial camera 11 is connected to the network switch 18 via the camera data cable 12.

[0104] Furthermore, in the above technical solution, the laser fiber 6 is located inside the probe 1, with one end connected to the lens system 4 and the other end connected to the fiber optic interface 17 of the host 2.

[0105] Furthermore, in the above technical solution, the lens system 4 is provided with a plano-convex lens 7, a reflecting prism 8, and a convex lens 9.

[0106] Furthermore, in the above technical solution, an upper window glass 13 is provided on one side of the transparent groove 5, and a lower window glass 14 is provided on the other side of the transparent groove 5.

[0107] Furthermore, in the above technical solution, the probe 1 is made of any one of stainless steel, Hastelloy, niobium-tantalum alloy, titanium alloy and zirconium alloy, and has an explosion-proof design.

[0108] To better understand and implement this invention, a specific application scenario of the invention is provided below as an embodiment 2: In this embodiment, the probe device 1 is specifically composed as follows. The probe 1 uses a Hastelloy housing, which is resistant to solution corrosion and crystallizer internal pressure. A through-slot 5 is formed on the side of the probe 1, and an upper window lens 13 and a lower window lens 14 are respectively installed on both sides of the through-slot 5. Both lenses are made of sapphire glass, which has a Mohs hardness of 9, and can withstand particle impact without causing scratches that affect image quality. An image acquisition system 3 is integrated inside the probe 1. The image acquisition system 3 consists of a telecentric lens 10 and an industrial camera 11. The industrial camera 11 uses a 1 / 1.8-inch complementary metal-oxide-semiconductor image sensor camera with a minimum exposure time of 0.03ms and a minimum pixel size of 2.20μm. The telecentric lens 10 is connected to the industrial camera 11 via a precision thread. Adjusting the number of turns of the precision thread aligns the focal plane with the observation area within the through-slot 5. A lens system 4 is also integrated inside the probe 1. The lens system 4 consists of a plano-convex lens 7, a reflecting prism 8, and a convex lens 9. Laser fiber 6 enters the probe 1 through fiber optic interface 17, transmitting the laser from host 2 to the entrance of plano-convex lens 7. Plano-convex lens 7 collimates the diverging laser into a parallel beam. After two right-angle reflections by reflecting prism 8, the beam propagation direction becomes perpendicular to the plane of the transparent groove 5, and is converged by convex lens 9 into the solution observation area within the transparent groove 5, forming transmissive illumination. Laser 19 is installed inside host 2. After generating laser light, laser 19 couples it into laser fiber 6 via fiber optic coupler. Laser 19 supports continuous adjustment of laser power within the range of 1mW to 1W, with a minimum pulse width of 0.02ms. A rotating diffuser is installed in the optical path of lens system 4, driven by a stepper motor. Industrial camera 11 is connected to data interface 15 via camera data cable 12. Data is transmitted via network cable 16 to network switch 18 inside host 2, and network switch 18 uploads image information to computer.

[0109] The test used an aqueous solution with an initial potassium sulfate mass fraction of 18% as the test object. The initial temperature of the solution was 60℃, and it was cooled to 20℃ at a cooling rate of 1℃ / min. The crystallization process lasted for about 40 minutes. The illumination pulse width was set to 0.02ms, and the exposure trigger signal of the industrial camera 11 was hardware synchronized with the pulse output signal of the laser 19. The speckle contrast was maintained below 0.08, and the image acquisition frame rate was 10 frames / s.

[0110] In the image processing stage, the computer performs high-density particle adhesion segmentation based on graph cut on the acquired images. Pixels in the image are modeled as graph nodes, and the edge weights are composed of a weighted sum of three terms: pixel intensity gradient field, distance transformation field, and orientation constraint, weighted in a 3:5:2 ratio. The prior penalty coefficient for particle convexity is set to 0.5, and the maximum flow algorithm is used to solve for the globally optimal segmentation. For cases with high solid content in the later stages of crystallization and more than 800 particles in a single frame, a block-parallel graph cut strategy is adopted. The image is divided into 16 4×4 sub-blocks with 10% overlap. Graph cuts are performed in parallel on each sub-block, and the results are merged based on the consistency constraint of the overlapping area. After segmentation of each frame, particle shape parameters such as aspect ratio and roundness are extracted, and the particle size distribution data of equivalent sphere diameter are statistically analyzed, such as... Figure 2 As shown.

[0111] During the particle size distribution monitoring phase, the particle size distribution data was segmented with a time window of 60 seconds. The Wasserstein distance and regularization coefficient between adjacent windows were calculated using the Sinkhorn iterative algorithm. Take 0.05, number of particle size intervals The value is 50, the convergence iterations are approximately 200, and the computational complexity is O(n log n). After constructing the granular distribution time-series Wasserstein distance matrix, a sliding window statistical test method is used for continuous scanning. The test threshold is determined based on the 95th quantile of the historical normal operating condition Wasserstein distance sequence. Figure 3 As shown, when the temperature dropped to around 35℃, the impeller speed unexpectedly decreased from 150 r / min to 80 r / min. The system detected an over-threshold drift in the Wasserstein distance sequence. After reversing the transfer plan matrix, it was found that the mass transport volume was significantly concentrated in the 100–200 μm particle size range, indicating an abnormal increase in the number of particles in this range. This is consistent with the physical mechanism of particle agglomeration caused by weakened stirring. The anomaly detection time lag was approximately 120 s, providing operators with quantitative fault diagnosis information. The key parameters of particle size distribution at each stage of the test are shown in Table 1.

[0112] Table 1 Key parameters of particle size distribution at each cooling stage

[0113]

[0114] During the laser power closed-loop control phase, the laser power adaptive control model runs in real time at a frequency of 10 frames / s, calculating the illumination quality evaluation function value for each frame. The weighting coefficients are taken as follows: , During the rapid increase in solid content (cooling down to around 35℃), the strategy network, through its long short-term memory network module, recognizes the continuous decreasing trend in image contrast and proactively increases the laser power current in medium steps, thus... The value recovered from 0.62 to 0.88, and the image contrast remained stable during the process state switching, without overexposure or underexposure. During the abnormal event, The value briefly dropped below 0.28, triggering an adaptive adjustment process for the rotating diffuser speed. The speed increased from 1200 r / min to 1800 r / min, and the speckle contrast decreased back below 0.1 within 3 seconds. Figure 4 The diagram shown is a structural schematic of the probe device 1 in this embodiment. The probe 1, host 2, image acquisition system 3, lens system 4, through slot 5, laser fiber 6, plano-convex lens 7, reflecting prism 8, convex lens 9, telecentric lens 10, industrial camera 11, camera data cable 12, upper window lens 13, lower window lens 14, data cable interface 15, network cable 16, fiber optic interface 17, network switch 18, and laser 19 are arranged and connected in the diagram. Figure 4 The annotation is in Chinese.

[0115] Compared with traditional methods, this invention achieves the following technical advancements: Traditional laser illumination schemes cannot suppress speckle noise while maintaining particle imaging resolution. This invention, through the time-domain averaging mechanism of a rotating diffuser, utilizes statistical independence to naturally eliminate random speckle fluctuations in the integral domain, fundamentally removing the constraint between laser coherence and imaging quality. Traditional watershed segmentation algorithms rely on manual intervention to adjust parameters when the gray-level gradient between particles is not significant. This invention, through a combination of global graph cut energy optimization and particle convexity priors, embeds particle geometric characteristics into the segmentation framework in the form of energy penalties, enabling the algorithm to adaptively handle high-solids-content scenarios without manual intervention. Traditional statistical measures such as KL divergence treat particle size intervals as independent discrete points, failing to perceive geometric changes in distribution patterns. This invention introduces optimal transport theory, allowing the measurement of particle size distribution differences to naturally integrate with the geometric structure of the particle size space, improving the physical interpretability of anomaly detection in industrial crystallization processes.

[0116] It should be noted that the variables involved in this invention are explained in detail in Tables 2 and 3.

[0117] Table 2. Variable Explanation Table (Part 1)

[0118]

[0119] Table 3. Variable Explanation Table (Part Two)

[0120]

[0121] The above description is merely a specific embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any changes or substitutions that can be easily conceived by those skilled in the art within the scope of the technology disclosed in the present invention should be included within the scope of protection of the present invention.

Claims

1. A laser-illuminated microscopic imaging method, characterized in that, Includes the following steps: A semiconductor laser generates a laser beam, which is coupled into a laser fiber via an optical fiber coupler. The laser beam is then transmitted through the laser fiber to a lens system. The laser beam is shaped into a parallel beam by a plano-convex lens, its propagation direction is changed by two right-angle reflections by a reflecting prism, and it is converged into a transparent slot by a convex lens to form transmissive illumination. The rotating diffuser performs speckle time-domain averaging on the laser beam in the optical path of the lens system, reducing the speckle contrast to below the speckle contrast threshold. The semiconductor laser operates in pulsed laser stroboscopic illumination mode, with the illumination pulse width set to 0.02ms. The exposure time of the industrial camera is synchronously matched with the illumination pulse width. The probe is placed in the solution to be tested so that the solution covers the permeable channel. The telecentric lens acquires images of tiny particles in the permeable channel. The industrial camera transmits the image information to the network switch via the camera data cable. The network switch then uploads the image information to the computer via the network cable. The computer performs high-density particle adhesion segmentation processing on image information based on graph cut. The pixels in the image information are modeled as graph nodes. The edge weights are formed by the weighted sum of the pixel intensity gradient field, the distance transformation field and the orientation constraint. The particle convexity prior is introduced as a higher-order penalty term and added to the energy function. The maximum flow algorithm is used to solve the global optimal segmentation and obtain the particle shape parameters and particle size distribution data. The computer uses Wasserstein distance to measure the difference between granular distribution data in two consecutive time windows, employs the Sinkhorn iterative algorithm to solve the discrete optimal transmission problem, constructs a granular distribution time series Wasserstein distance matrix, and uses a sliding window statistical test method to identify granular distribution drift events. When an abnormal drift is detected, the transmission plan matrix is ​​solved in reverse to locate the granularity range that caused the change in granular distribution. The laser power adaptive control model collects the histogram statistics of the current frame image, the laser power current and solution concentration sensor readings, calculates the illumination quality assessment function value, and adjusts the output power of the semiconductor laser according to the interval to which the illumination quality assessment function value belongs, so that the image contrast is maintained within the optimal range.

2. The laser-illuminated microscopic imaging method according to claim 1, characterized in that, The lens system consists of a plano-convex lens, a reflecting prism, and a convex lens. The plano-convex lens shapes the diverging laser beam emitted from the laser fiber into a parallel beam. The reflecting prism performs two right-angle reflections on the parallel beam to change its propagation direction. The convex lens focuses the parallel beam onto the test solution area within the transparent groove.

3. The laser-illuminated microscopic imaging method according to claim 2, characterized in that, The speckle contrast is the ratio of the standard deviation to the mean of the random light intensity fluctuations generated by the coherent superposition of laser light on the imaging surface. The speckle contrast threshold is 0.

1. The rotating diffuser is continuously rotated within the single-frame exposure time of the industrial camera, so that the speckle pattern at different times is averaged in the speckle time domain within the single-frame exposure integral.

4. The laser-illuminated microscopic imaging method according to claim 3, characterized in that, The pulsed laser strobe illumination mode is an illumination mode in which the semiconductor laser emits light periodically with an illumination pulse width of 0.02ms. The exposure trigger signal of the industrial camera is synchronized with the pulse output signal of the semiconductor laser to ensure that image acquisition is completed only during the period when the semiconductor laser emits pulses.

5. The laser-illuminated microscopic imaging method according to claim 4, characterized in that, The high-density particle adhesion segmentation processing based on graph cut models the particle image segmentation problem as a minimum graph cut energy optimization problem. In the graph, nodes correspond to image pixels, and the edge weights between adjacent nodes are composed of the weighted sum of three terms: pixel intensity gradient field, distance transformation field, and orientation constraint. The particle convexity prior is introduced as a higher-order penalty term added to the energy function to penalize non-convex segmentation boundaries.

6. The laser-illuminated microscopic imaging method according to claim 5, characterized in that, For images with high solids content and a large number of particles, a block-based parallel graph cutting strategy is adopted. The results of each block are merged with the consistency constraint of overlapping regions, so that the computational complexity of segmentation increases linearly with the number of particles.

7. The laser-illuminated microscopic imaging method according to claim 6, characterized in that, The distance transformation field is a scalar field composed of the Euclidean distance from each pixel in the image information to the nearest background pixel, which is used to guide the segmentation boundary to shrink towards the grain edge; the grain convexity prior is to apply an energy penalty to the region where the segmentation result has a concave boundary, so that the segmentation boundary tends to be convex; the grain shape parameters include aspect ratio and roundness.

8. The laser-illuminated microscopic imaging method according to claim 7, characterized in that, The Wasserstein distance defines the difference between two granularity distributions as the minimum total transmission cost required to move the mass of one distribution to the other. The Sinkhorn iterative algorithm converges to an approximately optimal transmission plan by alternately performing row and column normalization operations on the transmission plan matrix, with a computational complexity of O(n log n). ,in Number of particle size ranges This is the regularization coefficient.

9. The laser-illuminated microscopic imaging method according to claim 8, characterized in that, The particle size distribution time series Wasserstein distance matrix is ​​a matrix formed by calculating the Wasserstein distance between each pair of particle size distribution data in a continuous time window with time as the axis. The reverse solution of the transport plan matrix is ​​to locate the particle size range that causes the change in particle size distribution by analyzing the row and column positions of the mass transport volume concentration in the transport plan matrix.

10. A probe device based on laser illumination, characterized in that, It consists of three parts: a probe, a main unit, and a computer. The probe is made of Hastelloy as the shell material and integrates a telecentric lens, an industrial camera, a laser fiber, and a lens system. A transparent slot is opened on the side of the probe, and a sapphire glass window is installed on each side of the transparent slot. The upper window faces the lens system side, and the lower window faces the telecentric lens side.

Images