An aggregate particle size laser online monitoring system based on AI model

Abstract

The present invention relates to the technical field of aggregate detection, and discloses an aggregate particle size laser online monitoring system based on an AI model, comprising: an image processing module, used for acquiring digital images of aggregates; a deep learning module, used for predicting particle sizes of digital images of aggregates through a low-rank CNN model; the low-rank CNN model comprises: a model compression unit and an incremental learning unit; an adaptive acquisition module, used for acquiring digital images of bone particles at multiple angles and multiple times; a real-time feedback adjustment module, used for reporting and marking according to the output of the deep learning module, automatically adjusting the sampling frequency and scanning angle of laser scanning according to the current state and demand, and updating the parameters of the model, and after receiving new data, continuously adapting to new data and environmental changes, so that the system can better adapt to different working environments and demands, and improve monitoring efficiency and accuracy.

Description

Technical Field

[0001] The present invention relates to the field of aggregate detection technology, and in particular to an aggregate particle size laser online monitoring system based on an AI model. Background Art

[0002] With the development of laser technology, the performance of laser particle size analyzers has been improved, the accuracy and resolution have been improved, and the measurement speed has also been significantly improved. People have begun to try to apply it to the field of particle size analysis, and combine it with artificial intelligence technology to automatically optimize and expand the performance and functions of aggregate particle size laser monitoring systems.

[0003] Traditional systems often find it difficult to accurately describe the dynamic characteristics of complex nonlinear systems or highly changing environments, and usually require models and parameters to be determined during the design phase. It is difficult to dynamically adjust and optimize over time, and it is difficult to dynamically adjust monitoring parameters and algorithms based on real-time data to cope with complex and changing production environments. They lack adaptability and intelligence.

[0004] After searching, the Chinese patent number CN117911781A discloses a video-based aggregate particle size statistics method, which uses advanced image segmentation algorithms and deep learning models to accurately segment aggregates, combined with geometric calculation methods such as the minimum circumscribed rectangle, and can accurately measure the maximum particle size of aggregates. The real-time image acquisition system can continuously and stably obtain aggregate information, which is conducive to timely discovering abnormal conditions in the production process and making corresponding adjustments. It uses advanced image segmentation algorithms and deep learning models to accurately segment aggregates, combined with geometric calculation methods such as the minimum circumscribed rectangle, and can accurately measure the maximum particle size of aggregates.

[0005] However, the above-mentioned patent occupies a large storage space for the model, has a high computational complexity, and requires a large amount of computing resources for reasoning, which increases the time and energy consumption of the reasoning process, making it difficult to respond in a short period of time, resulting in reduced detection efficiency. Therefore, an aggregate particle size laser online monitoring system based on an AI model is proposed. Summary of the invention

[0006] The present invention proposes an aggregate particle size laser online monitoring system based on an AI model in order to solve the problem that the model occupies a large storage space, has high computational complexity, and is difficult to respond in a short time, resulting in reduced detection efficiency.

[0007] In order to achieve the above object, the present invention adopts the following technical solutions:

[0008] An image processing module, used for obtaining digital images of aggregates;

[0009] A deep learning module, used for predicting the particle size of aggregate digital images through a low-rank CNN model; the low-rank CNN model includes: a model compression unit and an incremental learning unit;

[0010] Adaptive acquisition module, used to obtain digital images of bone particles at multiple angles and times;

[0011] A real-time feedback adjustment module is used to broadcast and label according to the output of the deep learning module.

[0012] The above technical solution further includes:

[0013] The image processing module captures the density characteristics and structure of the aggregate surface through the high-resolution camera and laser scanning in the image acquisition module, sets the image acquisition parameters according to the actual situation, calibrates and corrects the acquired images, eliminates image aberrations and distortions, and stores and manages the acquired image data.

[0014] The image processing module pre-processes the collected original image and improves the quality and clarity of the image. At the same time, based on the extracted feature information, the particle size distribution and average particle size of the aggregate are measured, and according to the extracted feature information, the aggregate is classified and identified, and divided into different categories according to color, shape and size characteristics.

[0015] The deep learning module uses the CNN model to predict the particle size of the aggregate surface laser scanning image, and uses the mean square error as the loss function to measure the difference between the model prediction result and the true particle size label. The model prediction result is y_{pred}, and the true particle size label is y_{true}, then the mean square error is:

[0016] ;

[0017] Where N is the number of samples, is the prediction result of the i-th sample, is the true label of the i-th sample. In the optimization algorithm, gradient descent is used to update the model parameters. The update rule of gradient descent is:

[0018] ;

[0019] Where θ is the parameter of the model, α is the learning rate, and L is the loss function. For the mean square error loss function, the model parameters are updated by calculating its partial derivative with respect to the model parameters and using the gradient descent algorithm.

[0020] The model compression unit reduces the number of parameters in the deep neural network model by rank approximation, performs matrix decomposition on the weight matrix, decomposes it into the product of two low-rank matrices and focuses on the weight matrix of one of the convolutional layers. The weight matrix is ​​W, and its dimension is m×n, where m is the number of input channels and n is the number of output channels. The weight matrix is ​​low-rank approximated, and the approximate matrices U and V of the matrix W are iteratively solved using the alternating least squares algorithm. The iteration of the ALS algorithm first initializes the approximate matrices U and V, fixes V, and updates U by minimizing the loss function:

[0021] ;

[0022] in, represents the Frobenius norm, λ is the regularization parameter, U is fixed, and V is updated by minimizing the loss function:

[0023] ;

[0024] Repeat the steps until the convergence condition is reached. In each iteration, the ALS algorithm updates the approximate matrices U and V by minimizing the loss function, gradually decomposing the original weight matrix W into the product of the approximate matrices U and V.

[0025] The incremental learning unit obtains the gradient of the parameter W by deriving the loss function, and then updates the parameter using the gradient descent method:

[0026] ;

[0027] in, The updated parameter vector, The original parameter vector, Loss function L for parameters The gradient of α is the learning rate, which controls the step size of parameter update. It continuously receives new sample data and uses the data to update the parameters of the model. When new data is received, the new parameters are calculated using the above formula.

[0028] The adaptive acquisition module adjusts the sampling frequency and scanning range parameters of the laser scanning. When the model predicts the particle size with an error greater than the required standard, the sampling frequency or the number of scans is increased to increase the data volume, and the acquisition parameters are dynamically adjusted according to the current load conditions.

[0029] The real-time feedback adjustment module timely adjusts the parameters of the model according to the real-time collected data and monitoring results. When there is a deviation between the monitoring system and the actual scene, the parameters are corrected through real-time calibration.

[0030] The present invention has the following beneficial effects:

[0031] 1. In the present invention, the size, computational complexity or number of parameters of the deep neural network model are reduced through model compression, and lower computational complexity of aggregate data is utilized, and less computing resources are used for reasoning, thereby reducing the time and energy consumption of the reasoning process and improving the practicality and deployability of the aggregate model.

[0032] The number of parameters in the model can be reduced through low-rank approximation. Matrix decomposition is performed on certain weight matrices, which are decomposed into the product of two low-rank matrices to reduce the number of parameters. The original CNN model is compressed with low-rank approximation to obtain an approximate model with fewer parameters, thereby reducing computing resources, reducing the time and energy consumption of the inference process, and improving the practicality and deployability of the aggregate model.

[0033] 2. In the present invention, the sampling frequency and scanning angle of the laser scan are automatically adjusted according to the current status and needs, and the parameters of the model are updated. After receiving new data, the system continuously adapts to new data and environmental changes, so that the system can better adapt to different working environments and needs, and improve monitoring efficiency and accuracy. BRIEF DESCRIPTION OF THE DRAWINGS

[0034] Figure 1 This is an overall system block diagram of an aggregate particle size laser online monitoring system based on an AI model proposed by the present invention;

[0035] Figure 2 This is a partial block diagram of an aggregate particle size laser online monitoring system based on an AI model proposed in the present invention. DETAILED DESCRIPTION

[0036] Embodiment 1

[0037] The following will be combined with the drawings in the embodiments of the present invention to clearly and completely describe the technical solutions in the embodiments of the present invention. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without creative work are within the scope of protection of the present invention.

[0038] like Figure 1-2 As shown, the present invention proposes an aggregate particle size laser online monitoring system based on an AI model, comprising: an image processing module for acquiring digital images of aggregates; a deep learning module for predicting particle size of digital images of aggregates through a low-rank CNN model; the low-rank CNN model comprises: a model compression unit and an incremental learning unit; an adaptive acquisition module for acquiring digital images of bone particles at multiple angles and multiple times; a real-time feedback adjustment module for reporting and marking according to the output of the deep learning module;

[0039] The image processing module captures the density characteristics and structure of the aggregate surface through the high-resolution camera and laser scanning in the image acquisition module, sets the image acquisition parameters according to the actual situation, calibrates and corrects the acquired images, eliminates image distortion, and stores and manages the acquired image data;

[0040] The image processing module preprocesses the collected original images and improves the image quality and clarity. At the same time, based on the extracted feature information, the particle size distribution and average particle size of the aggregate are measured. According to the extracted feature information, the aggregate is classified and identified, and divided into different categories according to color, shape and size characteristics. The deep learning module uses the CNN model to predict the particle size of the laser scanning image of the aggregate surface, and uses the mean square error as the loss function to measure the difference between the model prediction result and the true particle size label. The model prediction result is y_{pred}, and the true particle size label is y_{true}, then the mean square error is:

[0041] ;

[0042] Where N is the number of samples, is the prediction result of the i-th sample, is the true label of the i-th sample. In the optimization algorithm, gradient descent is used to update the model parameters. The update rule of gradient descent is:

[0043] ;

[0044] Where θ is the model parameter, α is the learning rate, L is the loss function, and for the mean square error loss function, the model parameters are updated by calculating its partial derivative with respect to the model parameters and using the gradient descent algorithm;

[0045] The model compression unit reduces the number of parameters in the deep neural network model by rank approximation. The weight matrix is ​​decomposed into the product of two low-rank matrices and focuses on the weight matrix of one of the convolutional layers. The weight matrix is ​​W, and its dimension is m×n, where m is the number of input channels and n is the number of output channels. The weight matrix is ​​low-rank approximated and the approximate matrices U and V of the matrix W are iteratively solved using the alternating least squares algorithm. The iteration of the ALS algorithm first initializes the approximate matrices U and V, fixes V, and updates U by minimizing the loss function:

[0046] ;

[0047] in, represents the Frobenius norm, λ is the regularization parameter, U is fixed, and V is updated by minimizing the loss function:

[0048] ;

[0049] Repeat the steps until the convergence condition is reached. In each iteration, the ALS algorithm updates the approximate matrices U and V by minimizing the loss function, gradually decomposing the original weight matrix W into the product of the approximate matrices U and V.

[0050] In this embodiment, the particle size information of the aggregate surface is obtained through the image acquisition module, the data obtained by laser scanning is converted into a digital image, and the collected image data is stored and managed, and the data is transmitted to the image processing module to pre-process the collected original image, and improve the quality and clarity of the image, extract and identify the shape and size feature information of the aggregate particles, classify and identify the aggregate, and classify it into different categories according to color, shape and size features, including multiple convolutional layers, pooling layers and fully connected layers;

[0051] By extracting image features and predicting particle size, accurate monitoring and prediction of aggregate particle size can be achieved. Low-rank approximation is used to reduce the number of parameters in the model. Matrix decomposition is performed on certain weight matrices, which are decomposed into the product of two low-rank matrices to reduce the number of parameters. The original CNN model is compressed with low-rank approximation to obtain an approximate model with fewer parameters, thereby reducing computing resources, reducing the time and energy consumption of the inference process, and improving the practicality and deployability of the aggregate model.

[0052] Embodiment 2

[0053] like Figure 1-2 As shown, in the embodiment of the present invention, the incremental learning unit obtains the gradient of the parameter W by deriving the loss function, and then updates the parameter using the gradient descent method:

[0054] ;

[0055] in, The updated parameter vector, The original parameter vector, Loss function L for parameters The gradient of α is the learning rate, which controls the step size of parameter update. It continuously receives new sample data and uses the data to update the parameters of the model. When new data is received, the new parameters are calculated using the above formula. The adaptive acquisition module adjusts the sampling frequency and scanning range parameters of the laser scanning. When the model predicts the particle size and the error is greater than the required standard, the sampling frequency or the number of scans is increased to increase the data volume. At the same time, the acquisition parameters are dynamically adjusted according to the current load conditions.

[0056] The real-time feedback adjustment module adjusts the parameters of the model in a timely manner according to the real-time collected data and monitoring results. When there is a deviation between the monitoring system and the actual scene, the parameters are corrected through real-time calibration.

[0057] In this embodiment, through the adaptive acquisition module, when a large error occurs when predicting the particle size, the adaptive acquisition module can increase the amount of data by reducing the sampling frequency or increasing the number of scans to improve the training effect of the model, realize intelligent adjustment of the data acquisition parameters, and transmit the data to the incremental learning unit, continuously receive new aggregate sample data, and use the new aggregate sample data to update the parameters of the model;

[0058] When new data is received, new parameters are calculated and applied to the model to continuously adapt to new data and environmental changes, realize dynamic updating and adaptability of the model, and transmit the data to the real-time feedback adjustment module. Through the real-time feedback adjustment module, the aggregate particle size laser online monitoring system can respond to changing environments and demands in a timely manner during operation, maintain efficient and stable operation, and continuously improve monitoring effects and performance.

[0059] Although embodiments of the present invention have been shown and described, it will be appreciated by those skilled in the art that various changes, modifications, substitutions and variations may be made to the embodiments without departing from the principles and spirit of the present invention, and that the scope of the present invention is defined by the appended claims and their equivalents.

Claims

1. An aggregate particle size laser online monitoring system based on AI model, characterized in that: include: An image processing module, used for obtaining digital images of aggregates; A deep learning module for particle size prediction of aggregate digital images using a low-rank CNN model; The low-rank CNN model includes: a model compression unit and an incremental learning unit; Adaptive acquisition module, used to obtain digital images of bone particles at multiple angles and times; A real-time feedback adjustment module for broadcasting and labeling based on the output of the deep learning module; The model compression unit reduces the number of parameters in the deep neural network model by rank approximation, performs matrix decomposition on the weight matrix, decomposes it into the product of two low-rank matrices and focuses on the weight matrix of one of the convolutional layers. The weight matrix is ​​W, and its dimension is m×n, where m is the number of input channels and n is the number of output channels. The weight matrix is ​​low-rank approximated, and the approximate matrices U and V of the matrix W are iteratively solved using the alternating least squares algorithm. The iteration of the ALS algorithm first initializes the approximate matrices U and V, fixes V, and updates U by minimizing the loss function: ； Among them, |·|F represents the Frobenius norm, λ is the regularization parameter, U is fixed, and V is updated by minimizing the loss function: ； Repeat the steps until the convergence condition is reached. In each iteration, the ALS algorithm updates the approximate matrices U and V by minimizing the loss function, gradually decomposing the original weight matrix W into the product of the approximate matrices U and V. The incremental learning unit obtains the gradient of the parameter W by deriving the loss function, and then updates the parameter using the gradient descent method: ； in, The updated parameter vector, The original parameter vector, Loss function L for parameters The gradient of α is the learning rate, which controls the step size of parameter update. It continuously receives new sample data and uses the data to update the parameters of the model. When new data is received, the new parameters are calculated using the above formula.

2. According to the AI ​​model-based aggregate particle size laser online monitoring system of claim 1, it is characterized in that: The image processing module captures the density characteristics and structure of the aggregate surface through the high-resolution camera and laser scanning in the image acquisition module, sets the image acquisition parameters according to the actual situation, calibrates and corrects the acquired images, eliminates image aberrations and distortions, and stores and manages the acquired image data.

3. The aggregate particle size laser online monitoring system based on AI model according to claim 1 is characterized in that: The image processing module pre-processes the collected original image and improves the quality and clarity of the image. At the same time, based on the extracted feature information, the particle size distribution and average particle size of the aggregate are measured, and according to the extracted feature information, the aggregate is classified and identified, and divided into different categories according to color, shape and size characteristics.

4. The aggregate particle size laser online monitoring system based on AI model according to claim 1 is characterized in that: The deep learning module uses the CNN model to predict the particle size of the aggregate surface laser scanning image, and uses the mean square error as the loss function to measure the difference between the model prediction result and the true particle size label. The model prediction result is y_{pred}, and the true particle size label is y_{true}, then the mean square error is: ； Where N is the number of samples, is the prediction result of the i-th sample, is the true label of the i-th sample. In the optimization algorithm, gradient descent is used to update the model parameters. The update rule of gradient descent is: ； Where θ is the parameter of the model, α is the learning rate, and L is the loss function. For the mean square error loss function, the model parameters are updated by calculating its partial derivative with respect to the model parameters and using the gradient descent algorithm.

5. The AI ​​model-based aggregate particle size laser online monitoring system according to claim 1 is characterized in that: The adaptive acquisition module adjusts the sampling frequency and scanning range parameters of the laser scanning. When the model predicts the particle size with an error greater than the required standard, the sampling frequency or the number of scans is increased to increase the data volume, and the acquisition parameters are dynamically adjusted according to the current load conditions.

6. The AI ​​model-based aggregate particle size laser online monitoring system according to claim 1 is characterized in that: The real-time feedback adjustment module timely adjusts the parameters of the model according to the real-time collected data and monitoring results. When there is a deviation between the monitoring system and the actual scene, the parameters are corrected through real-time calibration.

Images