A particle size collection method for multi-probe online particle shape and size analyzer

Abstract

The present invention provides a particle size collection method for a multi-probe online particle shape and particle size analyzer, which belongs to the technical field of measuring instruments, including: immersing the front end of the multi-probe online particle shape and particle size analyzer into the solution to be tested in the reactor, so that the solution does not pass through the transparent slot; adjusting the light intensity of the light source through a multi-channel light source controller to ensure uniform illumination and avoid dark corners; starting an industrial camera, collecting images of particles in the solution to be tested through a telecentric lens, and recording them as particle images; the industrial camera transmits the collected particle images to a host computer through a network device; a particle size calculation module set in the host computer calculates a particle feature matrix and a particle size distribution according to the particle image, and outputs and saves them to a database. The present invention uses advanced image processing and deep learning technology to extract detailed morphological features from the collected particle images, calculate accurate particle size distribution, and solves the problem that the prior art is difficult to achieve online accurate measurement.

Description

Technical Field

[0001] The invention belongs to the technical field of measuring instruments, and in particular relates to a particle size collection method for a multi-probe online particle shape and size analyzer. Background Art

[0002] Particle size analysis is an important technology widely used in industrial production and scientific research. It can accurately measure and analyze the particle size distribution of raw materials, intermediate products or final products in materials, chemical, metallurgical and other industries, providing a key basis for production process optimization and quality control. At present, the commonly used particle size analysis methods mainly include screening method, sedimentation method, laser diffraction method, etc. Among them, laser diffraction method is a non-contact optical measurement technology that can quickly and accurately obtain particle size distribution information.

[0003] However, existing laser diffraction particle size analyzers have some shortcomings. First, they can usually only measure a single container or pipeline, and cannot simultaneously observe the changes in particles in multiple different solutions. Second, such instruments need to take samples from the production line and then send them to the laboratory for analysis and testing, which will cause certain production interruptions and time delays. In addition, the sampling process often interferes with the original state of the solution, affecting the accuracy of the measurement results. Summary of the invention

[0004] In view of this, the present invention provides a particle size collection method for a multi-probe online particle shape and size analyzer, which can solve the technical problem that existing particle size measuring instruments are difficult to achieve online accurate measurement.

[0005] The present invention is achieved in that:

[0006] The present invention provides a particle size collection method for a multi-probe online particle shape and size analyzer, which comprises the following steps:

[0007] S10, immerse the front end of the multi-probe online particle shape and size analyzer into the solution to be tested in the reactor, so that the solution submerges the transparent tank;

[0008] S20, adjusting the light intensity of the light source through a multi-channel light source controller to ensure uniform lighting and avoid dark corners;

[0009] S30, starting the industrial camera, and collecting images of particles in the solution to be tested through a telecentric lens, which are recorded as particle images;

[0010] S40, the industrial camera transmits the collected particle image to the host computer through the network device;

[0011] S50, a particle size calculation module arranged in the host computer calculates a particle feature matrix and a particle size distribution according to the particle image, and outputs and saves the calculated particle feature matrix and the particle size distribution to a database.

[0012] The granularity calculation module is used to perform the following steps:

[0013] S51, acquiring and preprocessing the particle image;

[0014] S52, using a preset shadow elimination model to eliminate particle shadows in the preprocessed particle image to obtain a first image;

[0015] S53, using image segmentation and edge detection methods to perform binarization processing on the first image, extract particle contours, filter out background noise, and obtain a second image that only retains the particle area;

[0016] S54, using a preset particle analysis model, inputting the second image, and obtaining a particle feature matrix and a particle size distribution.

[0017] The shadow removal model adopts an embedded hybrid neural network, including a group of Fourier anomaly equations and a convolution sub-network, a residual sub-network, an attention sub-network and a fusion sub-network.

[0018] Furthermore, the Fourier anomaly equation group includes a frequency equation, an amplitude equation, a phase equation and a reconstruction equation;

[0019] The convolutional subnetwork is used to extract local features of an image. The input is the original image, the output is a feature map, and the structure is a combination of multiple convolutional layers and pooling layers.

[0020] The residual sub-network is used to learn the residual information of the image. The input is the output of the convolutional sub-network, the output is the residual feature, and the structure is a multi-layer convolutional layer with jump connections;

[0021] The attention sub-network is used to highlight important features. The input is the output of the residual sub-network, the output is a weighted feature map, and the structure is a self-attention mechanism.

[0022] The fusion subnetwork is used to integrate the outputs of each subnetwork. The input is the output of the convolution, residual and attention subnetworks. The output is a shadow-free image, and the structure is a multi-layer fully connected layer.

[0023] The steps of training the shadow removal model include the following:

[0024] Construct the training set:

[0025] 1) Collect a large number of image samples containing particle shadows.

[0026] 2) Manually annotate these images and create corresponding shadow-free images as the ground truth.

[0027] 3) The image pairs are divided into training set, validation set and test set with a ratio of 7:2:1.

[0028] Training process:

[0029] 1) Initialize the parameters of the convolutional sub-network, residual sub-network, attention sub-network, and fusion sub-network.

[0030] 2) For each training sample:

[0031] a) Process the input image through the Fourier anomaly equations.

[0032] b) The processed image is passed through the convolutional sub-network, residual sub-network and attention sub-network in sequence.

[0033] c) Use the fusion sub-network to synthesize the outputs of each sub-network.

[0034] d) Calculate the loss between the output and the ground truth.

[0035] e) Back propagate the error and update the parameters of each sub-network.

[0036] 3) Evaluate the model performance on the validation set and save the model if the performance improves.

[0037] 4) Repeat steps 2-3 until the preset number of iterations is reached or the performance no longer improves.

[0038] The training steps of the particle analysis model include the following:

[0039] Construct the training set:

[0040] 1) Collect a large number of image samples of particles of different sizes.

[0041] 2) Manually annotate the size and distribution information of particles in these images.

[0042] 3) The samples are divided into training set, validation set and test set in a ratio of 7:2:1.

[0043] Training process:

[0044] 1) Initialize the parameters of the large particle branch network, small particle branch network, micro particle branch network and calculate the fusion branch network.

[0045] 2) For each training sample:

[0046] a) Input the image into the three granular branches network.

[0047] b) Use the computational fusion branch network to synthesize the outputs of the three branch networks.

[0048] c) Calculate the loss between the output particle characteristic matrix and particle size distribution and the standard answer.

[0049] d) Back propagate the error and update the parameters of each network.

[0050] 3) Evaluate the model performance on the validation set and save the model if the performance improves.

[0051] 4) Repeat steps 2-3 until the preset number of iterations is reached or the performance no longer improves.

[0052] Furthermore, the frequency equation is used to calculate the frequency characteristics according to the grayscale distribution of the input image; the amplitude equation is used to calculate the amplitude characteristics according to the frequency characteristics; the phase equation is used to calculate the phase characteristics according to the frequency characteristics; and the reconstruction equation is used to calculate the reconstructed image according to the amplitude characteristics and the phase characteristics.

[0053] Specifically, the formula of each equation in the Fourier anomaly equations is expressed as follows:

[0054] Frequency equation:

[0055]

[0056] Where F(u,v) is the result of two-dimensional discrete Fourier transform; f(x,y) is the gray value of the input image in the spatial domain; M,N are the width and height of the image; j is the imaginary unit; u,v are the frequency domain coordinates.

[0057] Amplitude equation:

[0058]

[0059] Where A(u,v) is the amplitude spectrum; Re(F(u,v)) and Im(F(u,v)) are the real and imaginary parts of F(u,v), respectively; α is an adjustment parameter, and the optimal value is determined by experiments.

[0060] Phase equation:

[0061]

[0062] Where φ(u,v) is the phase spectrum; β is the adjustment parameter, and the optimal value is determined through experiments.

[0063] Refactoring the equation:

[0064]

[0065] In the formula, f ′ (x, y) is the reconstructed image; γ is the weight of the Laplace operator, and the optimal value is determined through experiments; is the Laplace operator.

[0066] Parameter acquisition method:

[0067] 1. α, β and γ are obtained through grid search experiments:

[0068] Step 1: Set the parameter range, such as α∈[0.1,1], β∈[0.01,0.1], γ∈[0.001,0.01].

[0069] Step 2: Sample uniformly in the parameter space and experiment with each set of parameters.

[0070] Step 3: Select the parameter combination that gives the highest quality reconstructed image.

[0071] 2. f(x,y) is obtained through image acquisition equipment:

[0072] Step 1: Calibrate the image acquisition device.

[0073] Step 2: Take the image to be processed under standard lighting conditions.

[0074] Step 3: Convert the image to grayscale and get f(x,y).

[0075] 3. The calculation steps are:

[0076]

[0077] in,

[0078]

[0079] These equations and parameters are designed to improve the effect and robustness of shadow removal. The frequency equation is used to convert the image to the frequency domain; the amplitude equation and the phase equation extract the amplitude and phase characteristics of the frequency domain information respectively, and introduce additional adjustment terms to enhance specific frequency components; the reconstruction equation converts the processed frequency domain information back to the spatial domain, and introduces the Laplace operator to enhance edge details.

[0080] The particle analysis model adopts a multi-branch parallel structure, including a large particle branch network, a small particle branch network, a micro particle branch network and a computational fusion branch network.

[0081] Furthermore, the large particle branch network is used to analyze particles larger than 100 microns, the input is the second image, the output is the large particle features, and the structure is a deep convolutional neural network;

[0082] The small particle branch network is used to analyze particles of 10-100 microns, the input is the second image, the output is the small particle features, and the structure is a multi-scale convolutional neural network;

[0083] The microparticle branch network is used to analyze particles smaller than 10 microns, the input is the second image, the output is microparticle features, and the structure is a high-resolution convolutional neural network;

[0084] The computational fusion branch network is used to integrate the outputs of each branch network. The input is the output of the large, small and micro particle branch networks. The output is the particle feature matrix and particle size distribution. The structure is a combination of a multi-layer perceptron and a softmax classifier.

[0085] Furthermore, the method of preprocessing the particle image includes denoising and brightness adjustment.

[0086] Furthermore, the edge detection method is the Canny algorithm.

[0087] Furthermore, the multi-probe online particle shape and particle size analyzer includes a plurality of hollow tubular probes, an industrial camera is fixedly installed inside each probe, a transparent groove is opened on one side of the surface of the probe, a light source is arranged on one side of the transparent groove, a telecentric lens is arranged on the other side of the transparent groove, one end of the telecentric lens is connected to one end of the industrial camera by a thread, the data cable of the industrial camera is fixedly connected to a packaging cover arranged at the rear end of the probe, and the packaging cover is used to seal the probe.

[0088] Wherein, the step S10 is specifically: completely immersing the front end of the multi-probe online particle shape and particle size analyzer into the solution to be tested, so that the solution fully covers the transparent slot. The purpose of this step is to make the solution to be tested directly contact with the probe, so as to observe and analyze the particles in the solution later. The role of the transparent slot is to provide a visual window for the imaging system, so that it can clearly capture the image information of the particles in the solution.

[0089] Wherein, the step S20 is specifically: adjusting the illumination intensity of each light source through the multi-channel light source controller to ensure uniform illumination in the entire field of view and avoid dark corners. The purpose of this step is to obtain good imaging conditions. First, sufficient illumination must be ensured so that the particles in the solution can be clearly illuminated. At the same time, dark corners caused by uneven illumination must be avoided, as this will affect subsequent image analysis. The multi-channel light source controller can finely adjust the output of each light source to ensure uniform illumination in the entire field of view.

[0090] Wherein, the step S30 is specifically: starting the industrial camera, and collecting the image of the particles in the solution to be tested through the telecentric lens, which is recorded as the particle image. The purpose of this step is to obtain clear particle image data. The telecentric lens can align the focal plane with the solution to be tested in the transparent tank, thereby capturing the magnified image of the particles in the solution. The CMOS sensor of the industrial camera can record these image data with high definition.

[0091] Wherein, the step S40 is specifically: the industrial camera transmits the collected particle image to the host computer through the network device. The purpose of this step is to transmit the collected image data to the central processing unit for subsequent particle size analysis. The industrial camera sends the image data to the host computer through the network interface, and the host computer can receive and store this data to prepare for the next step of analysis and calculation.

[0092] Wherein, the step S51 is specifically: obtaining and preprocessing the particle image. First, the collected particle image data is received from the host computer, and then some basic preprocessing operations are performed on the image, such as denoising, brightness adjustment, etc., to improve the quality and stability of subsequent processing. The purpose of this step is to prepare for subsequent image analysis.

[0093] Wherein, the step S52 specifically comprises: using a preset shadow removal model to remove the particle shadows in the preprocessed particle image to obtain a first image. The shadow removal model uses an embedded hybrid neural network, including a Fourier anomaly equation group, a convolution subnetwork, a residual subnetwork, an attention subnetwork and a fusion subnetwork. The purpose of this step is to remove the particle shadows caused by illumination in the image, and provide better input for subsequent image segmentation and edge detection.

[0094] The step S53 is specifically: using the image segmentation and edge detection method to perform binarization processing on the first image, extract the particle outline, filter out the background noise, and obtain a second image that only retains the particle area. The purpose of this step is to segment the particle outline from the image with the shadow removed, and remove the background noise to obtain a binarized image containing only the particle area. This provides good basic data for the subsequent particle feature analysis.

[0095] Wherein, the step S54 is specifically: using a preset particle analysis model, inputting the second image, and obtaining a particle feature matrix and a particle size distribution. The particle analysis model adopts a multi-branch parallel structure, including a large particle branch network, a small particle branch network, a micro particle branch network, and a computational fusion branch network. The purpose of this step is to use advanced deep learning technology to extract various morphological features from the segmented particle region image, and calculate the particle size distribution based on these features. Through the parallel multi-branch network structure, particles of different size ranges can be effectively processed to improve the overall analysis accuracy.

[0096] Compared with the prior art, the particle size collection method for a multi-probe online particle shape and size analyzer provided by the present invention has the following beneficial effects:

[0097] First, this method uses a multi-probe design to simultaneously monitor the particle changes of solutions in multiple containers or pipelines, greatly improving the observation and analysis capabilities of the production process. Second, this method uses online optical imaging to directly observe the solution to be tested in real time at the production site, avoiding possible interference from the sampling process. Finally, this method uses advanced image processing and deep learning technology to extract detailed morphological features from the collected particle images and calculate accurate particle size distribution, greatly improving the accuracy and reliability of the measurement results.

[0098] Compared with the existing off-line particle size analysis method, the online particle size collection method of the present invention has the following significant advantages:

[0099] (1) Stronger monitoring capability: The multi-probe design can simultaneously observe the particle changes of solutions in multiple containers or pipelines, greatly improving the visualization capability of the production process.

[0100] (2) The measurement process is more direct: No sampling is required, and the solution to be tested can be observed in real time at the production site, avoiding possible interference during the sampling process.

[0101] (3) More accurate analysis results: The use of advanced image processing and deep learning technologies can extract more detailed morphological features from the collected particle images and calculate more accurate particle size distribution.

[0102] (4) More convenient operation: Through online measurement, the dynamic situation of particle changes during the production process can be grasped in real time, providing timely and effective support for production management and quality control.

[0103] In summary, the present invention solves the technical problem that existing particle size measuring instruments are difficult to achieve online accurate measurement. BRIEF DESCRIPTION OF THE DRAWINGS

[0104] Figure 1 A flow chart of the method provided by the present invention;

[0105] Figure 2 A flowchart of the steps executed by the granular computing module;

[0106] Figure 3 It is a schematic diagram of the composition structure of a multi-probe online particle shape and size analyzer;

[0107] In the accompanying drawings, the components represented by the reference numerals are listed as follows:

[0108] 11. Probe; 12. Transparent slot; 13. Light source; 14. Telecentric lens; 15. Industrial camera; 16. Packaging cover; 161. Camera data cable; 162. Light source control cable; 17. Host; 174. Network switch; 175. Light source controller; 176. Communication optical terminal. DETAILED DESCRIPTION

[0109] In order to make the purpose, technical solution and advantages of the embodiments of the present invention more clear, the technical solution in the embodiments of the present invention will be clearly and completely described below in conjunction with the accompanying drawings in the embodiments of the present invention.

[0110] like Figure 1 As shown, it is a flow chart of a particle size collection method for a multi-probe online particle shape and size analyzer provided by the present invention, and the method comprises the following steps:

[0111] S10, immerse the front end of the multi-probe online particle shape and size analyzer into the solution to be tested in the reactor, so that the solution submerges the transparent tank;

[0112] S20, adjusting the light intensity of the light source through a multi-channel light source controller to ensure uniform lighting and avoid dark corners;

[0113] S30, starting the industrial camera, and collecting images of particles in the solution to be tested through a telecentric lens, which are recorded as particle images;

[0114] S40, the industrial camera transmits the collected particle image to the host computer through the network device;

[0115] S50, a particle size calculation module arranged in the host computer calculates the particle feature matrix and the particle size distribution according to the particle image, and outputs and saves them to the database.

[0116] The granularity calculation module is used to perform the following steps:

[0117] S51, acquiring and preprocessing the particle image;

[0118] S52, using a preset shadow elimination model to eliminate particle shadows in the preprocessed particle image to obtain a first image;

[0119] S53, using image segmentation and edge detection methods to perform binarization processing on the first image, extract particle contours, filter out background noise, and obtain a second image that only retains the particle area;

[0120] S54, using a preset particle analysis model, inputting the second image, and obtaining a particle feature matrix and a particle size distribution.

[0121] The shadow removal model adopts an embedded hybrid neural network, including a group of Fourier anomaly equations and a convolution sub-network, a residual sub-network, an attention sub-network and a fusion sub-network.

[0122] The specific implementation methods of the above steps are described in detail below:

[0123] The specific implementation method of step S10 is: immerse the front end of the multi-probe online particle shape and size analyzer into the solution to be tested, so that the solution completely covers the transparent slot. The purpose of this step is to make the solution to be tested directly contact with the probe, so as to observe and analyze the particles in the solution later. The role of the transparent slot is to provide a visual window for the imaging system, so that it can clearly capture the image information of the particles in the solution.

[0124] The specific implementation of step S20 is: adjust the light intensity of the light source through the multi-channel light source controller to ensure uniform illumination and avoid dark corners. The purpose of this step is to obtain good imaging conditions. First, it is necessary to ensure sufficient illumination so that the particles in the solution can be clearly illuminated. At the same time, dark corners caused by uneven illumination should be avoided, as this will affect subsequent image analysis. The multi-channel light source controller can finely adjust the output of each light source to ensure uniform illumination in the entire field of view.

[0125] The specific implementation of step S30 is: start the industrial camera, and collect the image of the particles in the solution to be tested through the telecentric lens, which is recorded as the particle image. The purpose of this step is to obtain clear particle image data. The telecentric lens can align the focal plane with the solution to be tested in the transparent tank, thereby capturing the magnified image of the particles in the solution. The CMOS sensor of the industrial camera can record these image data with high definition.

[0126] The specific implementation of step S40 is: the industrial camera transmits the collected particle image to the host computer through the network device. The purpose of this step is to transmit the collected image data to the central processing unit for subsequent particle size analysis. The industrial camera sends the image data to the host computer through the network interface, and the host computer can receive and store this data to prepare for the next step of analysis and calculation.

[0127] The specific implementation of step S50 is: the particle size calculation module set in the host computer calculates the particle feature matrix and particle size distribution according to the particle image, and outputs and saves it to the database. The purpose of this step is to analyze and process the collected particle image to obtain the characteristic parameters and particle size distribution of the particles in the solution to be tested. The particle size calculation module contains multiple sub-steps, and the specific implementation is as follows:

[0128] The specific implementation of step S51 is: obtaining and preprocessing the particle image. First, the collected particle image data is received from the host computer, and then some basic preprocessing operations are performed on the image, such as denoising, brightness adjustment, etc., to improve the quality and stability of subsequent processing. The purpose of this step is to prepare for subsequent image analysis.

[0129] The specific implementation of step S52 is: using a preset shadow removal model to remove the particle shadows in the preprocessed particle image to obtain a first image. The shadow removal model uses an embedded hybrid neural network, including a Fourier anomaly equation group, a convolution subnetwork, a residual subnetwork, an attention subnetwork, and a fusion subnetwork. The purpose of this step is to remove the particle shadows caused by illumination in the image, and provide better input for subsequent image segmentation and edge detection.

[0130] The Fourier anomaly equations include:

[0131] 1. Frequency equation Used to calculate the frequency features based on the grayscale distribution of the input image.

[0132] 2. Amplitude equation It is used to calculate the amplitude characteristics according to the frequency characteristics. The optimal value of parameter α is determined by grid search.

[0133] 3. Phase equation It is used to calculate the phase characteristics according to the frequency characteristics. The optimal value of parameter β is determined by grid search.

[0134] 4. Reconstruct the equation It is used to calculate the reconstructed image based on the amplitude characteristics and phase characteristics. The optimal value of the parameter γ is determined by grid search.

[0135] The Fourier anomaly equations first transform the input image into the frequency domain, extract the frequency, amplitude and phase features, and then transform these features back to the spatial domain through the reconstruction equation to eliminate the shadows in the image. The purpose of this step is to remove the shadows caused by illumination in the particle image and provide a clearer input for subsequent image segmentation and edge detection.

[0136] The specific implementation of step S53 is: using the image segmentation and edge detection method, binarizing the first image, extracting the particle outline, filtering out the background noise, and obtaining a second image that only retains the particle area. The purpose of this step is to segment the particle outline from the image with the shadow removed, and remove the background noise to obtain a binary image containing only the particle area. This provides good basic data for the subsequent particle feature analysis.

[0137] The specific implementation of step S54 is: using a preset particle analysis model, inputting the second image, and obtaining a particle feature matrix and a particle size distribution. The particle analysis model adopts a multi-branch parallel structure, including a large particle branch network, a small particle branch network, a micro particle branch network, and a computational fusion branch network.

[0138] The large particle branch network uses a deep convolutional neural network to analyze particles larger than 100 microns.

[0139] The small particle branch network uses a multi-scale convolutional neural network to analyze particles of 10-100 microns.

[0140] The microparticle branch network uses a high-resolution convolutional neural network to analyze particles smaller than 10 microns.

[0141] The computational fusion branch network uses a combination of a multi-layer perceptron and a softmax classifier to integrate the outputs of each branch network to obtain the final particle feature matrix and particle size distribution.

[0142] The purpose of this step is to use advanced deep learning technology to extract various morphological features from the segmented particle area images and calculate the particle size distribution based on these features. Through the parallel multi-branch network structure, particles of different sizes can be effectively processed to improve the overall analysis accuracy.

[0143] In general, this particle size collection method for multi-probe online particle size analyzer makes full use of advanced technologies such as optical imaging, image processing and deep learning, and can achieve real-time online observation and precise analysis of particles in the solution to be tested. It overcomes the shortcomings of traditional methods that require sampling and preparation, greatly improves the efficiency and convenience of measurement, and also improves the accuracy and reliability of analysis results.

[0144] Specifically, the principle of the present invention is:

[0145] First, the method adopts a multi-probe design, and each probe is equipped with a transparent slot at the front end to connect the solution to be tested and the optical detection system. When the probe is immersed in the solution, the solution in the transparent slot can be photographed by the industrial camera. In this way, particles in multiple containers or pipelines can be observed and analyzed in real time at the same time, improving the overall monitoring capability.

[0146] Secondly, in order to obtain high-quality imaging effects, the present invention adopts a special lighting design. Each probe is equipped with a light source, and the illumination of each light source can be finely adjusted through a multi-channel light source controller to ensure that the light is uniform throughout the entire field of view and avoid dark corners. At the same time, the light source is set at the bottom of the probe, forming a transmissive lighting method, which can better illuminate the particles in the solution and improve the imaging quality.

[0147] Thirdly, in order to improve the accuracy of image analysis, the present invention adopts a shadow elimination model based on an embedded hybrid neural network. The model uses Fourier transform to perform frequency domain analysis on the input image, extracts amplitude and phase features, and effectively eliminates particle shadows caused by illumination through the synergy of multiple sub-networks such as convolution sub-network, residual sub-network, and attention sub-network, and obtains a clear particle outline image. This lays a good foundation for subsequent particle feature analysis.

[0148] Finally, the present invention adopts a multi-branch parallel deep learning model to perform particle feature analysis and particle size distribution calculation. The model includes a large particle branch network, a small particle branch network, a micro particle branch network, and a computational fusion branch network. Each branch network is designed with a suitable network structure and hyperparameters for particles in a specific size range, which can effectively extract the morphological characteristics of particles of different scales. The computational fusion branch network is responsible for combining the outputs of each branch network to obtain the final particle feature matrix and particle size distribution. This multi-branch parallel design can give full play to the advantages of deep learning in image recognition and feature extraction, greatly improving the overall analysis accuracy.

[0149] In summary, the present invention adopts innovative solutions in key technical aspects such as multi-probe design, uniform illumination, shadow elimination, and multi-scale feature analysis, and makes full use of advanced optical imaging, image processing, and deep learning technologies, thereby realizing real-time online observation and precise analysis of particles in solution.

[0150] In order to better understand and implement the present invention, a specific embodiment 1 of the granularity calculation module of the present invention is provided below. The steps of this embodiment 1 are specifically described as follows:

[0151] First, for step S51, this step mainly includes receiving the collected particle image data from the host computer and performing some basic preprocessing operations on the image, such as denoising, brightness adjustment, etc. The purpose of these preprocessing operations is to improve the quality and stability of subsequent image analysis. Since this step does not involve a more complex mathematical model, there is no need to use too many formulas for explanation.

[0152] Step S52, this step involves an embedded hybrid neural network model, the core idea of ​​which is to use Fourier transform to perform frequency domain analysis on the input image, thereby effectively eliminating the shadows in the image. The specific implementation process is as follows:

[0153] First, the input image f(x,y) is converted to the frequency domain using a two-dimensional discrete Fourier transform to obtain the spectrum F(u,v):

[0154]

[0155] Among them, M and N are the width and height of the input image respectively, and u and v represent the frequency domain coordinates.

[0156] Next, the amplitude spectrum A(u,v) and phase spectrum φ(u,v) are calculated based on the frequency spectrum F(u,v):

[0157]

[0158] Among them, α and β are two adjustment parameters that need to be determined through experiments, which are used to enhance specific frequency components to improve the shadow elimination effect.

[0159] Finally, the amplitude spectrum and phase spectrum are used to reconstruct the frequency domain to obtain the image f without shadows. ′ (x,y):

[0160]

[0161] Among them, γ is another adjustment parameter that needs to be determined through experiments, which is used to enhance the edge details of the reconstructed image. The Laplacian operator representing the input image can be obtained by numerical calculation methods.

[0162] Through the above-mentioned series of Fourier transform and its inverse transform, the shadows caused by illumination in the input image can be effectively eliminated, and the first image can be obtained, providing better input data for subsequent image segmentation and edge detection. It should be noted that the reasonable setting of parameters α, β and γ is critical to the performance of the entire shadow removal model, and the optimal value can be determined experimentally through grid search.

[0163] Next is step S53, the purpose of this step is to segment the outline of the particles from the first image with the shadow removed, and remove the background noise to obtain a binary image containing only the particle area. This provides good basic data for the subsequent particle feature analysis. Since this step mainly involves some common image processing algorithms, there is no need to use too many formulas for explanation.

[0164] Finally, step S54 uses a multi-branch parallel deep learning model that can effectively process particles of different sizes. The specific model structure is as follows:

[0165] Large particle branch network:

[0166]

[0167] in, represents the output features of the lth layer, represents the input features of the l-1th layer, and are the weight matrix and bias vector of the lth layer respectively, and σ is the activation function. The network adopts a deep convolutional neural network structure and is mainly used to analyze large particles larger than 100 microns.

[0168] Small particle branch network:

[0169]

[0170] in, represents the output features of the lth layer, Indicates L s -k layer input features, and are the weight matrix and bias vector of the lth layer respectively, and σ is the activation function. The network adopts a multi-scale convolutional neural network structure, which is mainly used to analyze small particles of 10-100 microns.

[0171] Microparticle branch network:

[0172]

[0173] in, represents the output features of the lth layer, Indicates L m -1 layer input features, and are the weight matrix and bias vector of the lth layer respectively, and σ is the activation function. The network adopts a high-resolution convolutional neural network structure, which is mainly used to analyze microparticles smaller than 10 microns.

[0174] Compute the fusion branch network:

[0175]

[0176] Among them, y represents the final output particle feature matrix and particle size distribution, are the output features of the three branch networks, W and b are the weight matrix and bias vector of the fusion network. The network adopts a combination of multi-layer perceptron and Softmax classifier to synthesize the output of each branch network and obtain the final analysis result.

[0177] The following is an example 2 of a specific application scenario of the present invention: In a chemical production process, it is necessary to conduct real-time monitoring and particle size analysis of the raw material solution in the reactor to ensure the quality and stability of the production. In the past, the traditional offline sampling method was used, that is, a certain amount of solution samples were taken from the reactor regularly and sent to the laboratory for particle size analysis. This method has some problems, such as the measurement results cannot reflect the real-time particle changes in the solution, and the sampling process may also interfere with the raw material state and affect the accuracy of the measurement.

[0178] In order to solve these problems, it is decided to adopt the particle size collection method for multi-probe online particle shape and size analyzer proposed in the present invention. The specific implementation process is as follows:

[0179] 1. Install a multi-probe online particle shape and size analyzer in the reactor at the production site, such as Figure 3 As shown, it includes three probes 11, a transparent groove 12 is opened on one side of the surface of the probe 11, a light source 13 is arranged on one side of the transparent groove 12, and a telecentric lens 14 is arranged on the other side of the transparent groove 12. One end of the telecentric lens 14 is connected to one end of an industrial camera 15 by a thread, the industrial camera 15 is fixed inside the probe 11, and the data line of the industrial camera 15 is fixedly connected to the packaging cover 16, and the packaging cover 16 is used to seal the probe 11; wherein, the host 17 is a multi-channel host. A multi-channel host is a computer or electronic device that can process multiple input signals at the same time. In the utility model, the multi-channel host is mainly used to receive and process image information transmitted from multiple probes. Here, "multi-channel" refers to the ability to process multiple independent input channels at the same time, and each channel corresponds to a probe. wherein, the packaging cover 16 is connected to an explosion-proof pipeline, which contains a camera data line 161 and a light source control line 162. The host 17 has a built-in network switch 174, a light source controller 175 and a communication optical terminal 176. The network switch 174 is connected to the communication optical terminal 176, and the light source controller 175 is connected to the communication optical terminal 176. The light source 13 is connected to the light source controller 175 in the host 17 through the light source control line 162. The industrial camera 15 is connected to the network switch 174 through the camera data line 161. The light source 13 is arranged at the inner bottom of the probe 11 to form a transmissive lighting. The transparent groove 12 is provided with a window, and the light source 13 provides lighting through the window at the transparent groove 12. A light source port is provided on one side of the surface of the packaging cover 16, and the light source 13 is connected to the light source controller 175 in the host 17 through the light source port. A communication port is provided on one side of the light source port, and the industrial camera 15 is connected to the network switch 174 in the host 17 through the communication port. Among them, the industrial camera and the inside of the probe are fixed with precision screw threads, which makes the device have high precision. The focal plane of the telecentric lens is changed up and down by the number of threads between the fine rice thread inside the probe and the telecentric lens to achieve the focusing effect. The adjusted telecentric lens is connected to the industrial camera, and the telecentric lens and the industrial camera are installed in the probe to form a sealed body. The probe is placed in the solution to be tested, and the solution to be tested must be submerged in the transparent groove. The light intensity of the light source is adjusted by the light source controller in the multi-channel host. The light beam is irradiated onto the solution to be tested through the transparent groove. The telecentric lens collects clear image information of tiny particles in the solution to be tested, and the telecentric lens presents the image on the CMOS sensor of the industrial camera. The industrial camera transmits the acquired image information to the multi-channel host, and the multi-channel host communicates with the computer through optical fiber.

[0180] 2. Real-time online monitoring of particle changes in solution

[0181] During the production process, when the raw material solution in the reactor reaches a certain liquid level, the front end of the multi-probe online particle size analyzer will be immersed in it. At this time, the solution in the transparent tank can be clearly observed by the optical detection system.

[0182] The central control unit first adjusts the output intensity of each light source through the multi-channel light source controller to ensure that the illumination in the entire field of view is uniform and there is no dark corner. After such illumination treatment, the particles in the solution can be clearly captured by the telecentric lens. The industrial camera collects these particle images in real time and transmits them to the central control unit.

[0183] 3. Process the image using the shadow removal model

[0184] Due to the three-dimensional spatial distribution of particles in the solution, some shadow areas may be generated in the image, which will affect the subsequent image analysis. To solve this problem, a shadow elimination model based on an embedded hybrid neural network is integrated into the central control unit.

[0185] The model first converts the collected particle image into the frequency domain through a two-dimensional discrete Fourier transform to obtain a spectrum. Then, the model uses the pre-set amplitude equation and phase equation to extract the amplitude and phase features from the spectrum. Next, the model conducts in-depth learning and fusion of these features through the synergy of multiple sub-networks such as the convolutional sub-network, the residual sub-network, and the attention sub-network, and finally obtains a clear particle image without shadows.

[0186] When processing a set of samples collected at 10:00 AM on August 15, the performance indicators of the shadow removal model are as follows:

[0187] -Image peak signal-to-noise ratio (PSNR) before shadow removal: 22.45dB;

[0188] -Image peak signal-to-noise ratio (PSNR) after shadow removal: 31.62dB;

[0189] - Structural Similarity Index (SSIM): increased from 0.78 to 0.92;

[0190] It can be seen that after being processed by the shadow removal model, the quality and clarity of the image have been significantly improved, laying a good foundation for subsequent particle analysis.

[0191] 4. Particle feature analysis based on multi-branch network

[0192] With high-quality particle image data, the central control unit then uses a multi-branch parallel deep learning model to perform detailed particle feature analysis and particle size distribution calculation on these images.

[0193] The model consists of four sub-networks: large particle branch network, small particle branch network, micro particle branch network and computational fusion branch network. Each branch network is designed with appropriate network structure and hyperparameters for particles of a specific size range.

[0194] The large particle branch network uses a deep convolutional neural network, which is mainly used to analyze large particles larger than 100 microns. Its network structure is as follows:

[0195]

[0196] in, represents the output features of the lth layer, represents the input features of the l-1th layer, and are the weight matrix and bias vector of the lth layer respectively, and σ is the activation function.

[0197] The small particle branch network uses a multi-scale convolutional neural network, which is mainly used to analyze small particles of 10-100 microns. Its network structure is as follows:

[0198]

[0199] in, represents the output features of the lth layer, Indicates L s -k layer input features, and are the weight matrix and bias vector of the lth layer respectively, and σ is the activation function.

[0200] The microparticle branch network uses a high-resolution convolutional neural network, which is mainly used to analyze microparticles smaller than 10 microns. Its network structure is as follows:

[0201]

[0202] in, represents the output features of the lth layer, Indicates L m -1 layer input features, and are the weight matrix and bias vector of the lth layer respectively, and σ is the activation function.

[0203] Finally, the computational fusion branch network uses a combination of a multi-layer perceptron and a Softmax classifier to synthesize the outputs of each branch network to obtain the final particle feature matrix and particle size distribution:

[0204]

[0205] Among them, y represents the final output result, are the output features of the three branch networks respectively, W and b are the weight matrix and bias vector of the fusion network.

[0206] When analyzing a set of samples collected at 10 a.m. on August 15, the performance indicators of the multi-branch network model are as follows:

[0207] Average recognition accuracy of large particles (>100 microns): 95.2%

[0208] Average recognition accuracy of small particles (10-100 microns): 92.6%

[0209] Average recognition accuracy of microparticles (<10 microns): 89.3%

[0210] The overall particle size distribution calculation error: less than 5%

[0211] It can be seen that through the synergistic effect of multi-branch networks, the model can very accurately identify and analyze particles of different size ranges, providing reliable data support for production management and quality control.

[0212] 5. Real-time data visualization and storage

[0213] The central control unit is responsible for coordinating the work of each probe, collecting and processing particle image data, and displaying the analysis results in real time on the large screen in the central control room. The specific data display forms include: real-time particle dynamic change curve and real-time proportion of particles of different sizes.

[0214] Through the above series of implementation processes, the online particle size collection method proposed by the present invention was successfully deployed at the production site. This method fully utilizes the advantages of key technologies such as multi-probe design, uniform illumination, shadow elimination, and multi-scale feature analysis to achieve real-time online observation and precise analysis of particles in the solution in the reactor. Compared with the traditional offline sampling method, this method can not only instantly grasp the particle changes in the production process, but also greatly improve the accuracy and reliability of the measurement results.

[0215] The above description is only a specific implementation mode of the present invention, but the protection scope of the present invention is not limited thereto. Any technician familiar with the technical field can easily think of changes or substitutions within the technical scope disclosed by the present invention, which should be covered by the protection scope of the present invention.

Claims

1. A particle size collection method for a multi-probe online particle shape and size analyzer, characterized in that: The following steps are involved: S10, immerse the front end of the multi-probe online particle shape and size analyzer into the solution to be tested in the reactor, so that the solution submerges the transparent tank; S20, adjusting the light intensity of the light source through a multi-channel light source controller to ensure uniform lighting and avoid dark corners; S30, starting the industrial camera, and collecting images of particles in the solution to be tested through a telecentric lens, which are recorded as particle images; S40, the industrial camera transmits the collected particle image to the host computer through the network device; S50, a particle size calculation module provided in the host computer calculates a particle feature matrix and a particle size distribution according to the particle image, and outputs and saves the calculated particle feature matrix and the particle size distribution to a database; The granularity calculation module is used to perform the following steps: S51, acquiring and preprocessing the particle image; S52, using a preset shadow elimination model to eliminate particle shadows in the preprocessed particle image to obtain a first image; S53, using image segmentation and edge detection methods to perform binarization processing on the first image, extract particle contours, filter out background noise, and obtain a second image that only retains the particle area; S54, using a preset particle analysis model, inputting the second image, and obtaining a particle feature matrix and a particle size distribution; The particle analysis model adopts a multi-branch parallel structure, including a large particle branch network, a small particle branch network, a micro particle branch network, and a computational fusion branch network; the large particle branch network is used to analyze particles larger than 100 microns, the input is the second image, the output is the large particle features, and the structure is a deep convolutional neural network; The small particle branch network is used to analyze particles of 10-100 microns, the input is the second image, the output is the small particle features, and the structure is a multi-scale convolutional neural network; The microparticle branch network is used to analyze particles smaller than 10 microns, the input is the second image, the output is microparticle features, and the structure is a high-resolution convolutional neural network; The computational fusion branch network is used to integrate the outputs of each branch network. The input is the output of the large, small and micro particle branch networks. The output is the particle feature matrix and particle size distribution. The structure is a combination of a multi-layer perceptron and a softmax classifier.

2. A particle size collection method for a multi-probe online particle shape and size analyzer according to claim 1, characterized in that: The shadow removal model adopts an embedded hybrid neural network, including a group of Fourier anomaly equations and a convolution sub-network, a residual sub-network, an attention sub-network and a fusion sub-network.

3. A particle size collection method for a multi-probe online particle shape and size analyzer according to claim 2, characterized in that: The Fourier anomaly equation group includes a frequency equation, an amplitude equation, a phase equation and a reconstruction equation; The convolutional subnetwork is used to extract local features of an image. The input is the original image, the output is a feature map, and the structure is a combination of multiple convolutional layers and pooling layers. The residual sub-network is used to learn the residual information of the image. The input is the output of the convolutional sub-network, the output is the residual feature, and the structure is a multi-layer convolutional layer with jump connections; The attention sub-network is used to highlight important features. The input is the output of the residual sub-network, the output is a weighted feature map, and the structure is a self-attention mechanism. The fusion subnetwork is used to integrate the outputs of each subnetwork. The input is the output of the convolution, residual and attention subnetworks. The output is a shadow-free image, and the structure is a multi-layer fully connected layer.

4. The particle size collection method for a multi-probe online particle shape and size analyzer according to claim 3, characterized in that: The frequency equation is used to calculate the frequency characteristics according to the grayscale distribution of the input image; the amplitude equation is used to calculate the amplitude characteristics according to the frequency characteristics; the phase equation is used to calculate the phase characteristics according to the frequency characteristics; the reconstruction equation is used to calculate the reconstructed image according to the amplitude characteristics and the phase characteristics.

5. The particle size collection method for a multi-probe online particle shape and size analyzer according to claim 4, characterized in that: The method of preprocessing the particle image includes denoising and brightness adjustment.

6. A particle size collection method for a multi-probe online particle shape and size analyzer according to claim 5, characterized in that: The edge detection method is the Canny algorithm.

7. A particle size collection method for a multi-probe online particle shape and size analyzer according to any one of claims 1 to 6, characterized in that: The multi-probe online particle shape and particle size analyzer includes a plurality of hollow tubular probes, an industrial camera is fixedly arranged inside each probe, a transparent groove is opened on one side of the surface of the probe, a light source is arranged on one side of the transparent groove, a telecentric lens is arranged on the other side of the transparent groove, one end of the telecentric lens is connected to one end of the industrial camera by a thread, the data cable of the industrial camera is fixedly connected to a packaging cover arranged at the rear end of the probe, and the packaging cover is used to seal the probe.

Images