Dried beancurd stick production line full-process management system based on AI

Abstract

The invention relates to an AI-based full-process management system for a dried beancurd stick production line, which belongs to the technical field of intelligent production management and industrial process control, and comprises a real-time material characteristic sensing module used for detecting biophysical characteristics of soybean milk in the production process in real time and generating biophysical characteristic vectors; the quality evolution prediction module is used for calculating an expected quality score and a corresponding prediction confidence coefficient; the dynamic process optimization module is used for resolving dynamic compensation process parameters; the process instruction execution module is used for adjusting an execution unit on the production line; the finished product quality detection module is used for detecting the quality of a final finished product and generating the actual quality of the final product; and the data flywheel self-learning module is used for carrying out iterative training on the quality prediction model and the dynamic optimization model. According to the invention, a lagging management mode of postmortem detection in traditional production is converted into a prospective management mode of beforehand early warning and in-process regulation and control, so that potential quality reduction risks can be foreseen in advance and actively avoided.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of intelligent production management and industrial process control, in particular to an AI-based whole-process management system for a dried beancurd stick production line. BACKGROUND

[0002] Dried beancurd stick production mainly uses soybeans as raw materials. However, as agricultural products, the biological and physical properties of soybeans, such as protein and fat, can vary significantly due to differences in production location, batch, and storage conditions. However, traditional dried beancurd stick production lines generally use fixed and unchanging production process parameters. This rigid production mode conflicts fundamentally with the instability of raw material properties. The main drawback is that when the properties of raw materials change, the fixed process cannot be adjusted accordingly, resulting in fluctuations in the quality of the final product, making it difficult to ensure the consistent and stable production of high-quality products. The existing management method is a lagging reaction mode, that is, quality problems are detected only after the production is completed by testing the finished products. At this time, it is impossible to make up for the loss, which not only wastes raw materials and energy, but also reduces overall production efficiency.

[0003] Therefore, there is an urgent need in the industry for an intelligent solution that can actively sense changes in raw materials and adjust production processes in real time and proactively to address the challenges posed by raw material uncertainty.

[0004] The above information disclosed in the background section is only used to enhance the understanding of the background of the present disclosure, and therefore it can include information that does not constitute prior art known to those of ordinary skill in the art. SUMMARY

[0005] The purpose of the present application is to provide an AI-based whole-process management system for a dried beancurd stick production line to solve the problems raised in the background.

[0006] The technical solution of the present application is an AI-based whole-process management system for a dried beancurd stick production line, comprising:

[0007] A real-time material property sensing module for real-time detection of the biological and physical properties of soy milk in the production process to generate a biological and physical property vector;

[0008] A quality evolution prediction module for receiving the biological and physical property vector and calculating the expected quality score and corresponding prediction confidence through a pre-trained quality prediction model;

[0009] A dynamic process optimization module for calculating dynamic compensation process parameters based on the biological and physical property vector when the expected quality score is lower than the preset target quality score, with the goal of maximizing a preset comprehensive benefit function;

[0010] a process instruction execution module, configured to receive the dynamic compensation process parameters, convert them into control signals, and adjust the execution units on the production line;

[0011] a finished product quality detection module, configured to detect the quality of the final product, and generate the actual quality of the final product;

[0012] a data flywheel self-learning module, configured to record the biophysical property vector, the dynamic compensation process parameters, and the actual quality of the final product as a data closed loop, and iteratively train the quality evolution prediction model and the dynamic optimization model based on the data closed loop.

[0013] Preferably, the real-time material property perception module comprises:

[0014] The near-infrared spectroscopy technology is adopted to detect the soy milk in real time, and obtain spectral data;

[0015] The spectral data is parsed into a structured biophysical property vector;

[0016] The biophysical property vector comprises real-time protein concentration, real-time fat concentration, total solid content, and protein denaturation index.

[0017] Preferably, the system further comprises a quality risk assessment and grading decision step, which is executed after the quality evolution prediction module and before the dynamic process optimization module, and is specifically configured to:

[0018] calculate a quality risk factor based on the expected quality score and the target quality score;

[0019] determine whether the quality risk factor is greater than a preset risk activation threshold;

[0020] in response to a determination that the quality risk factor is greater than the risk activation threshold, activate the dynamic process optimization module;

[0021] in response to a determination that the quality risk factor is not greater than the risk activation threshold, do not activate the dynamic process optimization module.

[0022] Preferably, the quality risk factor is calculated based on the prediction confidence output by the quality evolution prediction module.

[0023] Preferably, the comprehensive benefit function is constructed by weighted algebraic summation of the following: normalized quality gain; normalized cost increment; normalized time increment;

[0024] The normalized quality gain is characterized by the ratio of the quality improvement obtained after applying the dynamic compensation process parameter to the difference between the target quality score and the expected quality score;

[0025] The normalized cost increment is characterized by the ratio of the additional cost required to perform the dynamic compensation process parameter to the preset reference cost;

[0026] The normalized time increment is characterized by the ratio of the additional time required to perform the dynamic compensation process parameter to a preset reference time.

[0027] Preferably, the reference cost is determined as follows:

[0028] The reference cost is calculated based on the expected total output value of the current batch of products and a preset standard gross profit margin.

[0029] Preferably, the reference time is determined as follows:

[0030] The reference time is calculated based on the ratio of the standard process time of the current production process to the preset maximum allowable delay.

[0031] Preferably, the data flywheel self-learning module is specifically used to:

[0032] Associating the biophysical characteristic vector, the applied dynamic compensation process parameters, and the corresponding actual quality of the final product and storing them as a data set;

[0033] The data set is used to perform periodic iterative optimization on the internal functions in the quality prediction model and the dynamic optimization model.

[0034] The present invention provides an AI-based full-process management system for a yuba production line through improvements. Compared with the existing technology, it has the following improvements and advantages:

[0035] 1. This invention transforms the traditional post-production management model of lagging after-the-fact detection into a proactive management model of pre-emptive warning and in-process regulation by sensing raw material characteristics in real time and predicting finished product quality. This enables the foreseeing and proactive avoidance of potential quality degradation risks.

[0036] 2. This invention solves the core contradiction between raw material property fluctuations and the curing production process. The system can calculate the optimal compensation process parameters in real time based on the actual conditions of each batch of raw materials and automatically execute them, making the production line highly flexible and adaptable, thereby ensuring the stability of the final product quality.

[0037] 3. When performing process optimization, the application not only aims to improve quality, but also creatively establishes a comprehensive benefit function, which integrates quality, cost and time into a unified decision-making framework. By balancing the relationship between the three, the system can maximize overall production efficiency and economy while ensuring product quality;

[0038] 4. The application builds a closed-loop self-learning system from production practice to data deposition and model optimization through the data flywheel module. Each production process becomes a case for system learning, enabling the prediction and decision-making model to continuously iterate and evolve, maintaining long-term precision adaptation to various production changes. BRIEF DESCRIPTION OF DRAWINGS

[0039] The application will be further explained below in conjunction with the accompanying drawings and examples:

[0040] Figure 1 is a flowchart of the system of the application.

[0041] Figure 2 is a flowchart of the application for building a comprehensive benefit function. DETAILED DESCRIPTION

[0042] To make the purpose, technical scheme and advantages of the application clearer and more explicit, the application will be further described in detail below in conjunction with specific examples.

[0043] Example 1

[0044] Please refer to Figure 1 An AI-based bean curd stick production line whole-process management system, comprising:

[0045] A real-time material property sensing module for real-time detection of the biophysical properties of soy milk in the production process to generate a biophysical property vector;

[0046] A quality evolution prediction module for receiving the biophysical property vector and calculating the expected quality score and corresponding prediction confidence through a pre-trained quality prediction model;

[0047] A dynamic process optimization module for calculating dynamic compensation process parameters based on the biophysical property vector to maximize the preset comprehensive benefit function when the expected quality score is lower than the preset target quality score;

[0048] A process instruction execution module for receiving dynamic compensation process parameters and converting them into control signals to adjust the execution unit on the production line;

[0049] A finished product quality detection module for detecting the quality of the final product to generate the actual quality of the final product;

[0050] A data flywheel self-learning module is configured to record the biophysical property vector, the dynamic compensation process parameter, and the actual quality of the final product as a data closed loop, and iteratively train the quality prediction model and the dynamic optimization model based on the data closed loop.

[0051] An AI-based whole-process management system for a dried beancurd production line aims to solve the technical conflict between the batch-to-batch fluctuation of biophysical properties of soybean and other agricultural raw materials and the solidified production process.

[0052] The core of the architecture of the system is to establish a complete closed-loop regulation logic from raw material sensing to finished product feedback. Based on this architecture, the real-time material property sensing module actively and in real time quantifies the core properties of the raw material. Subsequently, the quality evolution prediction module makes a forward-looking deduction of the production results under the current process conditions based on the objective data of the previous sensing, which enables the production management to change from a lagging reaction mode to a predictive regulation mode. When the quality evolution prediction module predicts potential quality decline risks, the dynamic process optimization module is activated to seek the optimal process compensation strategy. This strategy is then accurately executed by the process instruction execution module. After production is completed, the finished product quality detection module quantitatively verifies the final results, which are fed together with the initial material properties and process parameters to the data flywheel self-learning module. The data flywheel self-learning module constitutes the closed-loop optimization path of the system.

[0053] Through deep analysis and modeling of the data of each production practice, the accuracy of the prediction and decision-making model is continuously optimized, thereby building an intelligent production system that can adapt to changes in raw materials and maximize production efficiency.

[0054] Embodiment 2

[0055] The real-time material property sensing module includes:

[0056] The near-infrared spectroscopy technology is used to detect the soybean milk in real time and obtain spectral data.

[0057] The spectral data is parsed into a structured biophysical property vector.

[0058] The biophysical property vector includes real-time protein concentration, real-time fat concentration, total solid content, and protein denaturation index.

[0059] The core of the real-time material property sensing module is the application of near-infrared spectroscopy technology, which can quickly and continuously detect the soybean milk sample online without damaging it, thereby obtaining spectral data reflecting its internal chemical composition and physical state. To use the original spectral data for process decision-making, the module further parses the spectral data into a structured biophysical property vector through a pre-set chemometrics model The chemometric model is preferably a partial least squares regression model, and the construction process thereof includes: collecting at least 100 batches of soy milk samples, determining the protein content of each sample by a national standard method such as the Kjeldahl method while obtaining the near-infrared spectrum data of each sample, accurately determining the true value of the biophysical property of each sample, and then calibrating and verifying the PLS model by using the spectrum data and the true value data until the prediction correlation coefficient (R2) of the model is greater than 0.95; the vector contains four key variables that play a decisive role in the film-forming mechanism of bean curd sticks: real-time protein concentration , real-time fat concentration , total solid content , and protein denaturation index ; ;

[0060] In this way, the system can accurately understand the real-time state of the soy milk at the molecular level, and provide quantitative input basis for subsequent quality prediction and process optimization calculation.

[0061] Embodiment 3

[0062] The system further includes a quality risk assessment and grading decision step, which is executed after the quality evolution prediction module and before the dynamic process optimization module, and is specifically used for:

[0063] calculating a quality risk factor based on the expected quality score and the target quality score;

[0064] determining whether the quality risk factor is greater than a preset risk activation threshold;

[0065] in response to the determination result that the quality risk factor is greater than the risk activation threshold, activating the dynamic process optimization module;

[0066] in response to the determination result that the quality risk factor is not greater than the risk activation threshold, not activating the dynamic process optimization module;

[0067] The quality risk factor is calculated based on the prediction confidence output by the quality evolution prediction module;

[0068] The quality risk assessment and grading decision step is a pre-decision unit of the dynamic process optimization module, which is used to avoid unnecessary process adjustment of the system for small quality fluctuations that have no significant impact, so as to realize reasonable allocation of computing resources and stability of the production process; the core of the step is the calculation of the quality risk factor , which not only considers the degree of quality deviation, but also incorporates the uncertainty of the prediction result itself;

[0069] Its inherent logic is to quantify the necessity of intervention. To achieve this goal, a comprehensive indicator is needed to uniformly evaluate the two dimensions of deviation size and prediction uncertainty to trigger the subsequent optimization module.

[0070] This factor is calculated using the following formula:

[0071]

[0072] in, It is a dimensionless quality risk factor, and its value comprehensively reflects the degree of quality deviation from the target and the uncertainty of the prediction results; A pre-set target quality score, which can be determined based on high-quality batches in historical production data or industry standards; is the expected quality score output by the quality evolution prediction module; The quality evolution prediction module outputs The standard deviation of the prediction results when represents the uncertainty measure of the model in its prediction results; It is an adjustable, dimensionless uncertainty weight coefficient, whose value can be optimized through offline simulation or based on historical data to balance the system's sensitivity to deviation and uncertainty. A specific optimization method is to use grid search on the historical verification data set to find a set of risk factors that make the risk factor The highest classification accuracy between the production batches that actually need intervention When the model uncertainty is high, When the risk factor is relatively large The value of will be magnified, achieving quantitative amplification of uncertainty risk;

[0073] In the application, the system will calculate the with a preset risk activation threshold For comparison, this threshold is also based on statistical analysis of historical data and engineering parameters set in combination with the production manager's tolerance for risk. To achieve objective setting, the quality risk factors of all batches of finished products with qualified quality and without manual intervention in history can be calculated. and set the 95th percentile of the statistical distribution as the initial risk activation threshold , and can be fine-tuned according to the actual operation effect; only when Greater than Only when the system determines that the current quality risk requires intervention and activates the dynamic process optimization module;

[0074] This decision-making mechanism ensures that computing resources are prioritized for processing quality deviations with significant impact, thereby improving the system's control efficiency and stability.

[0075] Example 4

[0076] Please refer to Figure 2 The comprehensive benefit function is constructed by weighted algebraic summation of the following: normalized quality gain; normalized cost increment; normalized time increment;

[0077] The normalized quality gain is characterized as the ratio of the quality improvement obtained after applying the dynamic compensation process parameters to the difference between the target quality score and the expected quality score;

[0078] The normalized cost increment is characterized as the ratio of the additional cost required to execute the dynamic compensation process parameters to the preset reference cost;

[0079] The normalized time increment is characterized as the ratio of the additional time required to execute the dynamic compensation process parameters to the preset reference time;

[0080] The reference cost is determined based on the expected total output value of the current batch of products and the preset standard gross profit margin;

[0081] The reference time is determined based on the standard process duration of the current production process and the preset maximum allowed delay ratio;

[0082] The construction and solution of the comprehensive benefit function convert the three optimization objectives of quality, cost, and time, which have different physical dimensions, into the same comparable framework through dimensionless processing, in order to solve a set of compensation process parameters that achieve the optimal comprehensive benefit in the three dimensions of quality, cost, and time The goal of optimization is to solve where the expression of the comprehensive benefit function is:

[0083]

[0084] This formula is a specific embodiment of weighted algebraic summation, where cost and time are negative benefit items, and is the output set of optimal compensation process parameters; is the target function to be maximized, which is a dimensionless comprehensive evaluation value; are the weight coefficients of quality, cost, and time, respectively, which are set by production managers according to different stages of business strategy, and the sum of the three is 1, and this function is composed of three core normalized sub-functions:

[0085] The normalized quality gain is expressed as:

[0086]

[0087] This function characterizes the completion rate of quality improvement after the compensation process is implemented; where is an internal quality simulation function that is used to quickly estimate the new expected quality score after the compensation process is applied. Specifically, this simulation function can be constructed as a surrogate model that is built by regression analysis on historical production data to establish a non-linear function that maps the biophysical property vector together with the dynamic compensation process parameters to the actual quality of the final product ; is a very small positive number to prevent the denominator from being zero;

[0088] Normalized cost increment , which is expressed as:

[0089]

[0090] It measures the proportion of additional cost for executing the compensation process relative to a benchmark; where is the additional cost required for executing , which is calculated based on the quantification model of additional energy consumption, material consumption, and labor cost corresponding to the compensation process parameters (e.g., increased heating temperature, extended soaking time, etc.); the reference cost is calculated based on ; here, refers to the expected total revenue of the current batch of products, is the standard gross profit margin set by the manager; by this method, the cost constraint is anchored to the commercial value of the batch of products, making the optimization decision directly related to the economic target;

[0091] Normalized time increment , which is expressed as:

[0092]

[0093] It measures the proportion of additional time caused by executing the compensation process relative to a benchmark; where is the additional time required for executing , which can also be built by statistical analysis on historical production cycle data to establish a relationship model between process parameter adjustment and production line delay; the reference time is calculated based on ; here, is the standard process duration of the current process, is the maximum delay ratio allowed for this process; this design ensures that the evaluation of time cost is closely related to the standard production cycle;

[0094] By maximizing this comprehensive benefit function, the system can calculate a set of optimal compensation process parameters; the final output process instructions are the optimal solution after comprehensively considering quality improvement, economy and time efficiency, greatly enhancing the flexibility and efficiency of production.

[0095] Example 5

[0096] The data flywheel self-learning module is specifically used for:

[0097] Associating and storing the biophysical property vector, the applied dynamic compensation process parameters, and the corresponding actual quality of the final product as a data set;

[0098] Using data sets, periodically iteratively optimize the internal functions in the quality prediction model and dynamic optimization model;

[0099] The Data Flywheel self-learning module builds a continuously optimized data closed loop to ensure the system can achieve long-term performance improvement. The core working mechanism of this module is divided into two levels:

[0100] First, at the data sedimentation level, this module converts the initial biophysical characteristic vector , dynamic compensation process parameters used in the process , and the actual quality of the final product measured by the finished product quality inspection module , the three are associated and stored as a structured data group as a whole, which constitutes the data basis required for subsequent model iterations;

[0101] Second, at the model evolution level, this module uses long-term accumulated data sets to periodically iteratively optimize the core models within the system; these data are used to retrain the quality prediction model. Internal parameters , and optimize the dynamic optimization model Key internal functions in , such as , for estimating additional costs Functions and for simulating compensation effects function;

[0102] This design forms a positive feedback loop, where more accurate models lead to better production decisions, better decisions generate higher-quality production data, and higher-quality data further trains more accurate models; thus, the system model can continuously improve the accuracy of its predictions and decisions by learning from historical production data to adapt to changes in production conditions.

[0103] It should be noted that the above examples are only used to illustrate the technical solutions of the present application but not limit the present application. Although the present application is described in detail with reference to the preferred embodiments, those skilled in the art should understand that the technical solutions of the present application can be modified or equivalently replaced, without departing from the spirit and scope of the technical solutions of the present application, which should be covered in the scope of the claims of the present application.

Claims

1. An AI-based full-process management system for a yuba production line, characterized in that: include: Real-time material property perception module, used to detect the biophysical properties of soy milk in real time during the production process and generate biophysical property vectors; A quality evolution prediction module, configured to receive the biophysical property vector and calculate an expected quality score and a corresponding prediction confidence using a pre-trained quality prediction model; A dynamic process optimization module is used to calculate dynamic compensation process parameters based on the biophysical characteristic vector and maximizing a preset comprehensive benefit function when the expected quality score is lower than a preset target quality score; A process instruction execution module is used to receive the dynamic compensation process parameters, convert them into control signals, and adjust the execution units on the production line; Finished product quality inspection module, used to inspect the quality of the final product and generate the actual quality of the final product; The data flywheel self-learning module is used to record the biophysical characteristic vector, dynamic compensation process parameters and the actual quality of the final product as a data closed loop, and iteratively train the quality prediction model and the dynamic optimization model based on the data closed loop.

2. The AI-based full-process management system for a yuba production line according to claim 1 is characterized in that: The real-time material property perception module includes: Using near-infrared spectroscopy analysis technology to perform real-time detection on the soy milk to obtain spectral data; parsing the spectral data into structured biophysical property vectors; The biophysical property vector includes real-time protein concentration, real-time fat concentration, total solid content and protein denaturation index.

3. The AI-based full-process management system for a yuba production line according to claim 1 is characterized in that: The system also includes a quality risk assessment and classification decision-making step, which is executed after the quality evolution prediction module and before the dynamic process optimization module, and is specifically used to: Calculating a quality risk factor based on the expected quality score and the target quality score; Determining whether the quality risk factor is greater than a preset risk activation threshold; In response to a determination that the quality risk factor is greater than the risk activation threshold, activating the dynamic process optimization module; In response to a determination that the quality risk factor is not greater than the risk activation threshold, the dynamic process optimization module is not activated.

4. The AI-based full-process management system for a yuba production line according to claim 3 is characterized in that: The quality risk factor is calculated based on the prediction confidence output by the quality evolution prediction module.

5. The AI-based full-process management system for a yuba production line according to claim 3 is characterized in that: The comprehensive benefit function is constructed by performing a weighted algebraic summation of the following items: normalized quality gain; normalized cost increment; Normalized time increment; The normalized quality gain is characterized by the ratio of the quality improvement obtained after applying the dynamic compensation process parameter to the difference between the target quality score and the expected quality score; The normalized cost increment is characterized by the ratio of the additional cost required to perform the dynamic compensation process parameter to the preset reference cost; The normalized time increment is characterized by the ratio of the additional time required to perform the dynamic compensation process parameter to a preset reference time.

6. The AI-based full-process management system for a yuba production line according to claim 5 is characterized in that: The reference cost is determined as follows: The reference cost is calculated based on the expected total output value of the current batch of products and a preset standard gross profit margin.

7. The AI-based full-process management system for a yuba production line according to claim 5 is characterized in that: The reference time is determined as follows: The reference time is calculated based on the ratio of the standard process time of the current production process to the preset maximum allowable delay.

8. The AI-based full-process management system for a yuba production line according to claim 1 is characterized in that: The data flywheel self-learning module is specifically used for: Associating the biophysical characteristic vector, the applied dynamic compensation process parameters, and the corresponding actual quality of the final product and storing them as a data set; The data set is used to perform periodic iterative optimization on the internal functions in the quality prediction model and the dynamic optimization model.

Images