CN-121982388-A - Composite seasoning quality monitoring system based on artificial intelligence

CN121982388ACN 121982388 ACN121982388 ACN 121982388ACN-121982388-A

Abstract

The invention discloses a composite seasoning quality monitoring system based on artificial intelligence, which comprises an image data acquisition module, a composite seasoning quality monitoring model building module, a model performance optimization module and a composite seasoning quality intelligent monitoring module. The invention relates to the technical field of image recognition, in particular to a composite seasoning quality monitoring system based on artificial intelligence, which creatively provides an innovative model training improvement method by combining self-adaptive gradient calibration, weight dynamic range correction and weight overturning inhibition, improves model training convergence efficiency and improves model output result accuracy; the optimization algorithm is innovatively improved by dynamically calculating the nonlinear convergence factor based on the hyperbolic tangent function and combining the fusion iteration self-adaptive nonlinear weight adjustment mechanism, so that the efficiency and the precision of the super-parameter search of the composite seasoning quality monitoring model are improved, the accuracy of the model output result is improved, and the intelligent monitoring of the composite seasoning quality is realized.

Inventors

SONG PENG

Assignees

天津康宏盛代实业有限公司

Dates

Publication Date: 20260505
Application Date: 20260119

Claims (8)

1. The composite seasoning quality monitoring system based on artificial intelligence is characterized by comprising an image data acquisition module, a composite seasoning quality monitoring model building module, a model performance optimization module and a composite seasoning quality intelligent monitoring module; the image data acquisition module is used for acquiring image optimization data of the composite condiment through image data acquisition and image data preprocessing operation; The composite seasoning quality monitoring model building module is characterized in that an initial framework of a composite seasoning quality monitoring model is built based on a binary neural network, then a model training process is improved through self-adaptive gradient calibration, weight dynamic range correction and weight overturning inhibition, and the quality monitoring model is subjected to iterative training to obtain a composite seasoning quality monitoring model after preliminary training; the model performance optimization module is used for improving an optimization algorithm by dynamically calculating a nonlinear convergence factor based on a hyperbolic tangent function and combining a fusion iteration self-adaptive nonlinear weight adjustment mechanism, performing model super-parameter self-adaptive optimization by adopting the improved optimization algorithm to obtain an optimal super-parameter combination of a composite seasoning quality monitoring model, and then performing model super-parameter replacement and final improvement of model performance to finally obtain the optimal composite seasoning quality monitoring model; The intelligent monitoring module for the quality of the compound seasoning is characterized in that the image data of the real-time compound seasoning is input into a monitoring model for the quality of the compound seasoning with optimal performance, a monitoring result of the quality of the real-time compound seasoning is obtained, and the real-time quality monitoring of the compound seasoning production is realized according to the result.
2. The artificial intelligence-based composite seasoning quality monitoring system according to claim 1, wherein the construction of the composite seasoning quality monitoring model module specifically comprises the following steps: the quality monitoring model structure is initialized, and specifically a composite seasoning quality monitoring model is established based on a binary neural network, wherein the composite seasoning quality monitoring model comprises an input layer, an image multi-dimensional feature extraction layer, a binary full-connection layer and an output layer; the quality monitoring model iterative training is carried out by taking historical composite condiment image data as training data, and comprises the following steps: initializing model training parameters, specifically initializing training basic parameters and composite seasoning quality monitoring model super parameters respectively; the model forward propagation is specifically that the composite condiment image optimization data is input into a composite condiment quality monitoring model to obtain a composite condiment quality monitoring result, and the weights and the activation values of 3 serially connected binary convolution layers and binary full-connection layers are binary quantized by adopting a symbol function to obtain binary weights and binary activation values; the loss function value and the gradient solution are calculated, specifically, the current iteration loss function value is calculated by adopting a cross entropy loss function, and the gradient value corresponding to the binary weight is solved through back propagation, wherein the following formula is adopted: The self-adaptive gradient calibration is specifically based on the original gradient of each layer of binary weight of a composite condiment quality monitoring model, firstly, the mean absolute value of the gradient value of the first layer of binary weight is calculated, the mean absolute value is smoothed by adopting an exponential smoothing algorithm to obtain a gradient smoothed value, and finally, the gradient smoothed value is subjected to deviation correction to obtain the self-adaptive gradient calibration value of the scene of the composite condiment, wherein the formula is as follows: Correcting the dynamic range of the weight; weight overturn inhibition; Weight updating, specifically, weight updating is performed based on the weight of the first layer and the binary weight gradient value in the t-th iteration after suppressing overturning, so as to obtain the first layer The weight of the first layer in the next iteration is calculated according to a cosine annealing strategy; And stopping the training iteration, namely, recalculating the loss function value of the next iteration when the iteration training is finished every time, stopping the training iteration when the loss function value is lower than a preset loss threshold value or the maximum training iteration number is reached, outputting the final weight of each layer of the model, replacing the final weight of each layer of the model to the corresponding layer of the composite seasoning quality monitoring model, and obtaining the composite seasoning quality monitoring model after preliminary training as the final weight of the composite seasoning quality monitoring model.
3. The artificial intelligence-based composite seasoning quality monitoring system according to claim 2, wherein the weight dynamic range is corrected by calculating a weight dynamic adjustment threshold adapted to a current iteration based on an adaptive gradient calibration value, judging whether the absolute value of the weight of the first layer in the t-th iteration exceeds the weight dynamic adjustment threshold, cutting the weight exceeding the weight dynamic adjustment threshold, and pulling the weight back to the weight dynamic adjustment threshold range to obtain the weight of the first layer in the t-th iteration after correction, wherein the formula is as follows: ; ; In the formula, Representing the dynamic adjustment threshold of the weight of the first layer in the t-th iteration, Representing the weight adjustment strength super-parameter, Indicating the weight of the first layer in the t-th iteration after correction, Representing the adaptive gradient calibration value of the first layer in the t-th iteration, Representing the weight of the first layer in the t-th iteration.
4. The artificial intelligence-based composite seasoning quality monitoring system according to claim 2, wherein the weight overturn suppression is specifically configured to calculate a weight overturn suppression threshold based on an adaptive gradient calibration value and a current iteration learning rate, compare the weight of a first layer in a t-th iteration after correction with the weight of the first layer in the t-th iteration after the previous correction, determine whether weight overturn occurs, and obtain the weight of the first layer in the t-th iteration after the overturn suppression, and the formula is as follows: ; ; In the formula, Representing the weight-flip suppression threshold for the first layer in the t-th iteration, Indicating the over-parameter of the forced intensity of the turning, The learning rate of the t-th iteration is represented, the learning rate is attenuated according to a cosine annealing strategy, Indicating post-correction item The weight of the first layer in a number of iterations, Indicating the weight of the first layer in the t-th iteration after the suppression flip, The representation is a signed function.
5. The artificial intelligence-based composite seasoning quality monitoring system of claim 1, wherein the model performance optimization module comprises the following steps: The model super-parameter self-adaptive optimization is realized, in particular to an optimal super-parameter combination of a composite seasoning quality monitoring model is obtained through an improved optimization algorithm; The model super-parameter replacement is specifically to adjust the super-parameters of the composite seasoning quality monitoring model after preliminary training based on the optimal super-parameter combination of the composite seasoning quality monitoring model to obtain a composite seasoning quality monitoring model with the optimal super-parameters; And finally improving the performance of the model, namely adopting historical composite condiment image data as training data, and performing quality monitoring model iterative training on the composite condiment quality monitoring model configured with optimal super parameters to obtain the composite condiment quality monitoring model with optimal performance.
6. The artificial intelligence based composite seasoning quality monitoring system according to claim 5, wherein the model super-parameter adaptive optimization, in particular, obtaining the optimal super-parameter combination of the composite seasoning quality monitoring model by improving an optimization algorithm, comprises the following steps: Initializing a search population, namely mapping the optimal super-parameter combination of the composite seasoning quality monitoring model into search individual position vectors in an optimization algorithm, and generating N search individual position vectors through a random initialization method to obtain an initial search population; calculating individual fitness values, namely calculating fitness values of searched individuals in a population, taking the performance of a quality monitoring model based on composite seasoning quality as the fitness values of the searched individuals, sorting all the searched individuals in the population in descending order according to the fitness values, defining individuals with optimal fitness and positions thereof, sub-optimal fitness individuals and positions thereof, and individuals with third optimal fitness and positions thereof; The nonlinear convergence factor is calculated, in particular to the nonlinear convergence factor is calculated dynamically based on a hyperbolic tangent function, and the formula is as follows: ; In the formula, Represent the first The nonlinear convergence factor is iterated a number of times, Representing curvature control parameters, value ranges , Representing the number of current iterations and, Representing a maximum number of iterations; The core search individual position is dynamically updated, in particular to the individual position of the core search individual is dynamically updated by fusing an iterative self-adaptive nonlinear weight adjustment mechanism, and the used formula is as follows: ; In the formula, Representing adaptive nonlinear weights; updating the positions of the search individuals, specifically, carrying out collaborative guiding updating on the rest search individuals according to the updated core search individuals; Updating the optimal position of the searching individual, namely reevaluating fitness values of all updated searching individuals, comparing the fitness values with the global optimal position of the current searching individual based on the fitness values of the current searching individual, and updating the global optimal position of the searching individual if the fitness values of the current searching individual are better; And stopping searching iteration, namely stopping searching and obtaining the global optimal position of the searched individual when the fitness value of the global optimal individual is higher than the fitness threshold value or the maximum iteration number is reached, wherein the global optimal position of the searched individual is specifically the optimal super-parameter combination of the composite seasoning quality monitoring model.
7. The artificial intelligence-based composite seasoning quality monitoring system according to claim 1, wherein the composite seasoning quality intelligent monitoring module is used for inputting real-time composite seasoning image data into a performance-optimal composite seasoning quality monitoring model to obtain a real-time composite seasoning quality monitoring result, and performing quality grading control on the composite seasoning according to the result to realize real-time quality monitoring of composite seasoning production.
8. The artificial intelligence-based composite seasoning quality monitoring system is characterized in that the image data acquisition module is used for acquiring composite seasoning original image data through image data acquisition operation, and performing image data preprocessing on the composite seasoning original image data to obtain composite seasoning image optimization data, wherein the composite seasoning original image data comprises historical composite seasoning image data and real-time composite seasoning image data, and the image data preprocessing comprises image de-duplication, image de-noising, image enhancement and image normalization to obtain the composite seasoning image optimization data.

Description

Composite seasoning quality monitoring system based on artificial intelligence Technical Field The invention relates to the technical field of image recognition, in particular to a composite seasoning quality monitoring system based on artificial intelligence. Background The composite seasoning quality monitoring system is an intelligent detection system based on an artificial intelligence technology, and can automatically perform image recognition, analysis and quality characteristic marking in an image by collecting image data on a composite seasoning production line, comprehensively detect and evaluate the quality of the composite seasoning, and automatically judge whether a product meets quality standards or not by identifying the problems of color difference, uneven morphology, foreign matter pollution and the like of the composite seasoning in the image in real time, thereby realizing accurate and efficient composite seasoning quality monitoring. However, the traditional composite seasoning quality monitoring model has the technical problems that the gradient is easily influenced by image noise and product batch difference to be fluctuated in the training process, the effective constraint is insufficient in weight dynamic range to cause updating difficulty, the weight is frequently overturned to cause slow convergence speed of model training and low accuracy of model output results, the existing super-parameter setting suitable for the composite seasoning quality monitoring model is unreasonable, and the super-parameter searching space is insufficient in the super-parameter optimizing process to easily fall into local optimum, so that the model output results are insufficient in accuracy. Disclosure of Invention Aiming at the technical problems that the traditional composite seasoning quality monitoring model is easy to be severely fluctuated due to image noise and product batch difference in the training process, the effective constraint is lacked to cause difficult updating and frequent weight overturning, so that the model training convergence speed is low and the model output result accuracy is low, the invention creatively provides an innovative model training improvement method combining self-adaptive gradient calibration, weight dynamic range correction and weight overturning inhibition, the gradient amplitude is dynamically regulated through self-adaptive gradient calibration, gradient contributions of different feature dimensions are balanced, weight updating boundaries are limited through weight dynamic range correction, excessive or excessively small weight absolute values are avoided, invalid overturning when the potential weight is close to 0 is reduced through weight overturning inhibition, the continuity of feature learning is ensured, training gradient is stabilized, invalid weight updating logic is reduced, model training convergence efficiency is improved, the learning capacity of the model on the core quality characteristics of the composite seasoning is enhanced, the model output stability is improved, the final quality is fully optimized, the composite seasoning quality is fully optimized due to the fact that the super-adaptive gradient calibration is not suitable for being set up in the prior art, the super-optimal seasoning quality monitoring parameter is not well-optimized due to the fact that the super-optimal dynamic parameter is not covered by the self-adaptive dynamic function, the existing intelligent seasoning quality monitoring method is not well-optimized, and the super-optimal performance is well-optimized, the optimization algorithm is improved by combining the fusion iteration self-adaptive nonlinear weight adjustment mechanism, the optimal superparameter combination of the composite seasoning quality monitoring model is obtained, the superparameter space can be explored more comprehensively, the local optimum is avoided, the efficiency and the precision of superparameter searching of the composite seasoning quality monitoring model are obviously improved, the adaptability of the composite seasoning quality monitoring model to quality monitoring of different types of composite seasonings is enhanced, the accuracy of the model output result is improved, the application performance and the stability of the composite seasoning quality monitoring model on a large-scale production line are finally realized, and the requirements of intelligent, efficient and accurate quality monitoring are met. The invention adopts the technical scheme that the artificial intelligence-based composite seasoning quality monitoring system comprises an image data acquisition module, a composite seasoning quality monitoring model building module, a model performance optimization module and a composite seasoning quality intelligent monitoring module; the image data acquisition module is used for acquiring image optimization data of the composite condiment through image data acquisition and image