A multi-process collaborative scheduling optimization method for a pomegranate peeling production line

Abstract

The present application relates to the field of manufacturing, in particular to a multi-process collaborative scheduling optimization method for pomegranate peeling production line, comprising collecting and processing first business data stream from Internet of Things interface, inputting historical sensor data into business prediction model to generate second business data stream; generating third business data stream from data of multi-modal sensor array; constructing multi-objective business cost function through business decision optimization engine; starting iterative optimization business simulation process to obtain optimized scheduling scheme; generating corresponding digital production work order from the optimized scheduling scheme. Through the three business data stream fusion and multi-objective cost function construction method, the device state, device health degree cost and quality index are analyzed, and the comprehensive business value evaluation is generated. When the conditions are met, the iterative optimization simulation and closed-loop optimization are started, the scheduling optimization of time effectiveness and value fluctuation in perishable material scheduling is realized, and the business value creation and adaptive scheduling ability of production are improved.

Description

TECHNICAL FIELD

[0001] The present application relates to the field of food manufacturing, in particular to a multi-process collaborative scheduling optimization method for a pomegranate peeling production line. BACKGROUND

[0002] In modern manufacturing, production scheduling is a core link, and its efficiency directly determines the production cost and market competitiveness of an enterprise. Currently, enterprises widely use manufacturing execution systems (MES) to manage production processes. One of the core functions of the system is job shop scheduling, that is, under limited resources, the processing order of each production task is reasonably arranged to achieve a specific optimization goal, such as minimizing the total completion time or maximizing the utilization of equipment. Existing technologies such as flow shop scheduling methods based on MES systems are suitable for production lines with fixed process flows, such as food processing, electronic assembly, etc. In these scenarios, the scheduling system usually uses heuristic algorithms based on fixed rules to generate production plans, which generally allocate tasks to single processes or local links according to pre-set static parameters, to a certain extent, achieving automated management of production.

[0003] Although the above scheduling methods based on MES and heuristic algorithms have been widely applied, their inherent defects are also very obvious when dealing with complex and dynamic production environments: most traditional scheduling methods generate a fixed plan based on a static model set before production, but dynamic events such as equipment breakdowns, fluctuations in raw material batch quality, and emergency order insertion frequently occur in actual production. Static plans cannot respond to these disturbances in real time, resulting in large-scale waiting or shutdown of subsequent processes and low overall efficiency of the production line. Heuristic algorithms often only focus on local optimization of a single process, for example, the shortest processing time first rule will prioritize the fastest task in the current process, but this will cause long waiting times for downstream processes, forming new bottlenecks. Moreover, scheduling decisions for each process are independent of each other, lacking a collaborative mechanism to consider the overall situation, making it difficult to achieve optimal total completion time for the entire production line. Existing scheduling models are basically designed for general manufacturing, without fully considering the process constraints of specific industries. For example, in the food processing industry, materials generally have perishability, and long waiting times not only affect efficiency, but also directly lead to decreased product quality and increased loss rate.

[0004] In summary, the existing technology generally lacks dynamic and collaborative optimization capabilities among multiple processes, resulting in low overall efficiency of the production line, slow response to abnormal events, and difficulty in balancing efficiency and product quality.

[0005] Therefore, a multi-process collaborative scheduling optimization method for a pomegranate peeling production line is proposed. SUMMARY

[0006] The purpose of the present application is to provide a multi-process collaborative scheduling optimization method for a pomegranate peeling production line to solve the problems raised in the background art.

[0007] To achieve the above purpose, the present application provides the following technical solutions: a multi-process collaborative scheduling optimization method for a pomegranate peeling production line, comprising:

[0008] Collecting and processing first business data streams reflecting the physical state of devices and materials in the production line in real time from the Internet of Things interface;

[0009] Inputting historical sensor data into a business prediction model to generate second business data streams containing expected performance degradation costs of devices and expected business value decay trajectories of material batches;

[0010] Generating third business data streams containing composite quality indicators of the material batches by evaluating product value from data derived from a multi-modal sensor array;

[0011] Inputting the first business data streams, second business data streams and third business data streams into a business decision optimization engine to build a multi-objective business cost function containing perishable business cost items; starting an iterative optimization business simulation process according to the multi-objective business cost function to obtain an optimized scheduling scheme;

[0012] Generating corresponding digital production work orders from the optimized scheduling scheme and issuing them to processing operation nodes through an interface with a manufacturing execution system to control the production line and perform closed-loop management and business performance tracking on the production line.

[0013] Preferably, the specific implementation process of collecting and processing the first business data streams includes:

[0014] Periodically obtaining device layer data containing device operating parameters and working state identifiers through sensor groups associated with each processing unit in the production line; at the same time, obtaining inventory data representing the number of in-process products between processes through non-contact sensors arranged on the material flow paths between processing units; at the data processing node, performing data cleaning, format standardization and time stamp alignment operations on the collected device layer data and inventory data to obtain the first business data streams.

[0015] Preferably, the specific implementation process of generating second business data streams includes:

[0016] extracting a feature vector strongly related to the health state of the equipment from historical sensor data, training a business prediction model using the feature vector, establishing a mapping relationship between equipment performance degradation and historical running time, and generating an expected performance degradation cost of the equipment; extracting records of initial material quality, environmental temperature and humidity, and waiting time of each process in the historical data, training a kinetic model for describing the decay law of material quality indicators with time and environmental changes, and obtaining an expected business value decay trajectory of the material batch; and combining the expected performance degradation cost and the expected business value decay trajectory to form the second business data stream.

[0017] Preferably, the specific implementation process of generating the third business data stream includes:

[0018] acquiring image data reflecting the appearance features of the material batch using an optical sensor; acquiring spectrogram data reflecting the internal composition of the material batch using a spectrum sensor; extracting appearance feature parameters from the image data and internal composition feature parameters from the spectrogram data; combining the appearance feature parameters and the internal composition feature parameters through a preset weighted fusion algorithm to obtain a composite quality index quantitatively representing the current comprehensive quality of the material batch; and packaging the composite quality index to form the third business data stream.

[0019] Preferably, the specific implementation process of constructing the multi-objective business cost function includes:

[0020] inputting the first business data stream, the second business data stream, and the third business data stream into a business decision optimization engine, constructing a first cost component for measuring the order delivery cycle based on the equipment running parameters and the working state identifier in the first business data stream; constructing a second cost component for measuring the inventory holding cost and the capital occupation based on the inventory data in the first business data stream; constructing a third cost component as a perishable business cost item by taking the real-time composite quality index provided by the third business data stream as an initial value, calculating the material quality decline caused by waiting and processing under different scheduling schemes according to the expected business value decay trajectory provided by the second business data stream, and converting the quality decline into direct sales revenue loss; constructing a fourth cost component for measuring the impact of the scheduling scheme on the long-term asset value of the equipment based on the expected performance degradation cost of the equipment in the second business data stream; and forming the multi-objective business cost function by weighted combination of the first cost component, the second cost component, the third cost component, and the fourth cost component.

[0021] Preferably, the specific implementation process of obtaining the optimized scheduling scheme includes:

[0022] An initial decision set containing production scheduling schemes is constructed, and the expected comprehensive business value of each production scheduling scheme in the set is quantitatively evaluated according to the multi-objective business cost function; an overall business profit-oriented iterative optimization business simulation process is started; the initial decision set is globally explored, macro operation strategies are generated and evaluated by adjusting the allocation timing of production resources and the processing priority of material batches, and a high-potential profit interval is identified; the production scheduling schemes in the profit interval are locally optimized, and the growth point of marginal profit is explored and locked by fine-tuning the processing time window and reordering adjacent processes; the optimized production scheduling scheme is fed back to the initial decision set, and the business value benchmark of the overall scheme is improved; the iterative process continues, and converges when the expected profit gain brought by adjustment is lower than the preset business sensitivity threshold, and the production scheduling scheme with the highest expected comprehensive business value is determined as the optimized scheduling scheme.

[0023] Preferably, the specific implementation process of generating the digitized production work order comprises:

[0024] The optimized scheduling scheme is analyzed, and the specific processing sequence, starting time and estimated completion time of each material batch at each processing operation node are extracted; according to the extracted information, a structured digitized production work order containing a unique job identifier, material batch information and control instructions is generated for each processing operation node; the digitized production work order is transmitted in real time to the manufacturing execution system through the interface, and is issued to the corresponding processing operation node by the system for automatic control; the structured digitized production work order contains data fields for recording and returning the actual completion time, material consumption and quality detection results of each node, which are used for subsequent closed-loop management and business performance tracking.

[0025] Compared with the prior art, the beneficial effects of the present application are:

[0026] 1. At the level of business decision-making, by integrating equipment expected performance degradation cost and inventory deterioration and other operating cost indicators into decision-making, scheduling is transformed from pure cost control to maximizing future overall profit. This dynamic business decision-making can avoid potential economic losses in advance, change risk management from after-the-fact remediation to pre-emptive prevention, and significantly enhance the business resilience and profitability of enterprises.

[0027] 2. At the production operation level, by integrating real-time data and multi-stage optimization simulation, a leap from static planning to dynamic adaptive scheduling is achieved. The system can intelligently respond to real-time changes in equipment, materials and orders, automatically generate optimal schemes, minimize invalid waiting and downtime, and significantly improve production line operation efficiency, flexibility and market responsiveness.

[0028] 3. At the product value management level, by introducing perishable commercial cost items, combined with real-time value assessment, and using the product's value window as the core scheduling basis, the system prioritizes materials with high value decay risk, ensuring they are processed and sold at peak value. This precise management of the value of perishables throughout their lifecycle reduces supply chain losses at the source. BRIEF DESCRIPTION OF THE DRAWINGS

[0029] Figure 1 This is a flow chart of a multi-process collaborative scheduling optimization method for a pomegranate peeling production line proposed in an embodiment of the present invention;

[0030] Figure 2 A flowchart for generating a multi-objective business cost function is proposed for an embodiment of the present invention;

[0031] Figure 3 A flowchart of an optimized scheduling solution is proposed for an embodiment of the present invention. DETAILED DESCRIPTION

[0032] The following will clearly and completely describe the technical solutions in the embodiments of the present invention in conjunction with the accompanying drawings. 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 making creative efforts are within the scope of protection of the present invention.

[0033] See also Figures 1 to 3 The present invention relates to a multi-process collaborative scheduling optimization method for a pomegranate peeling production line, and the specific implementation process is as follows:

[0034] Collect and process the first business data stream reflecting the physical status of equipment and materials in the production line in real time from the Internet of Things interface;

[0035] Inputting historical sensor data into a business prediction model to generate a second business data stream, wherein the second business data stream includes an expected performance degradation cost of the equipment and an expected commercial value decay trajectory of the material batch;

[0036] Generating a third business data stream comprising real-time composite quality indicators of the material batch by performing product value assessment on data from the multimodal sensor array;

[0037] inputting the first business data stream, the second business data stream, and the third business data stream into a business decision optimization engine to construct a multi-objective business cost function including a perishable business cost item; initiating an iterative optimization business simulation process based on the multi-objective business cost function to obtain an optimized scheduling solution;

[0038] The optimization scheduling scheme is used to generate corresponding digital production work orders, which are sent to processing operation nodes through an interface with a manufacturing execution system to control the production line and perform closed-loop management and business performance tracking on the production line.

[0039] The technical solutions of the present application will be further described in detail below in combination with specific embodiments.

[0040] Embodiment one

[0041] The embodiments of the present application disclose a multi-process collaborative scheduling optimization method for a pomegranate peeling production line. In the production line of a pomegranate peeling factory A, data streams collected by sensors are processed, and the method comprises the following steps: Figure 1 The specific steps of the method proposed by the present application include: S1, collecting and processing first business data streams from an Internet of Things interface; S2, inputting historical sensor data into a business prediction model to generate second business data streams; S3, generating third business data streams from data of a multi-modal sensor array; S4, constructing a multi-objective business cost function through a business decision optimization engine; S5, starting an iterative optimization business simulation process to obtain an optimization scheduling scheme; and S6, generating corresponding digital production work orders from the optimization scheduling scheme.

[0042] Further, first business data streams reflecting the physical state of equipment and materials in the production line are collected and processed from the Internet of Things interface in real time; the step corresponds to the above S1; and the specific implementation process includes:

[0043] Through a sensor group associated with each processing unit (such as a cleaning machine and a peeling machine) in the production line, equipment layer data and inventory data are periodically obtained. The sensor group includes a multi-axis vibration acceleration sensor (with a sampling frequency of 1 kHz) and a temperature sensor installed on the main motor of the peeling machine, which are used to monitor the health status of the equipment; and a high-precision weighing belt feeder and a non-contact ultrasonic liquid level sensor on the top of the hopper, which are arranged on the conveying belt between each processing unit, are used to obtain inventory data representing the number of work-in-process between processes.

[0044] All sensor data are aggregated to an edge computing node through industrial Ethernet. At this node, the collected data are cleaned, formatted and time-stamped in real time. Data cleaning adopts 3σ criterion to eliminate outliers and uses the nearest point interpolation method to fill in missing values. Format standardization converts all data into JSON format containing device ID, data type, value and timestamp. For example, the vibration sensor data from the peeler is formatted as: {deviceID: Peeler_01, dataType: Vibration_RMS, value: 0.35, timestamp: 2025-07-10T08:30:01.123Z}; the inventory data from the raw material hopper is formatted as: {deviceID: Hopper_01, dataType: Inventory_Ton, value: 2.5, timestamp: 2025-07-10T08:30:05.456Z}. Time stamp alignment takes the central clock source as the reference to ensure the comparability of data from different sources in time, thus forming a structured and highly reliable first business data stream.

[0045] By defining the deployment strategy and data processing flow of the sensor group for the equipment layer and the material flow path respectively, the real-time, accuracy and consistency of the first business data stream are ensured. It overcomes the problem of data island and information delay caused by the diversity of data sources, different formats and different time in traditional production environment. It provides reliable and accurate input for upper-level optimization decision, thereby significantly improving the analysis quality of subsequent business decision optimization engine and the real executable of scheduling scheme.

[0046] Further, the historical sensor data are input into a business prediction model to generate a second business data stream containing the expected performance degradation cost of the equipment and the expected business value decay trajectory of the material batch; corresponding to the above S2 step; the specific implementation process includes:

[0047] First, the device health cost is quantified, which represents the monetized impact of different scheduling schemes on the physical life and long-term health of the device. A feature vector strongly related to the device health state is extracted from historical sensor data including vibration, temperature, and power. The specific implementation path is as follows: First, a preliminary feature set is extracted from the raw time series data of each sensor. This set includes time-domain features of the vibration signal (such as root mean square, peak-to-peak value, margin factor, kurtosis, and skewness), frequency-domain features (such as the energy proportion and spectral entropy near the device's natural frequency and its multiples obtained by fast Fourier transform), and real-time readings and statistical features of the temperature and power signals (such as mean and standard deviation). Subsequently, to reduce data redundancy and improve model performance, a two-stage feature selection strategy is adopted: In the first stage, weakly correlated features with a historical device failure record or remaining useful life (RUL) label below a preset threshold (e.g., 0.4) are removed through Pearson correlation coefficient analysis. In the second stage, the RFE method is used to select the optimal feature subset (e.g., 10 features) that maximizes the contribution to the prediction target, combining a benchmark model (such as random forest), to finally form an N-dimensional feature vector for input into the subsequent model. A commercial prediction model, long short-term memory network (LSTM), is trained to predict the device health index using the feature vector. The specific network topology is as follows: an input layer, two hidden LSTM layers, and a fully connected output layer. Each LSTM layer contains 128 neurons and uses the hyperbolic tangent function as the activation function. The hyperbolic tangent function is chosen because of its inherent S-shaped curve characteristics, which can effectively fit the observed nonlinear decay law of the device health state throughout its life cycle. By analyzing a large amount of device failure history data, it is found that the device health decay follows a three-stage pattern of slow-accelerating-sharp. The S-shaped curve characteristics of the tanh function can approximate this physical process: the flat region of the function corresponds to the stable operation period of the device, the inflection point region corresponds to the accelerated performance decay period, and the steep region corresponds to the near-failure period. This makes the learning process of the neural network highly consistent with the real physical decay law of the device, improving the physical interpretability of RUL prediction. Not only does it improve the prediction accuracy, but more importantly, it makes the prediction results have a clear physical meaning, which is convenient for maintenance personnel to understand and make decisions. Compared with activation functions such as ReLU, it improves the accuracy of device life prediction. To prevent overfitting, a Dropout layer with a dropout rate of 0.2 is set between the two LSTM layers. The full life cycle historical data of the device from operation to failure are used for model training, and the key hyperparameters are set as follows: learning rate is 0.001, batch size is 64, and training rounds are 100.The RUL prediction value outputted by the model is converted into a monetary unit quantified expected performance degradation cost, which is used to quantify the potential impact of different scheduling schemes on the long-term asset value of the equipment in business decision-making, while recording the confidence interval (e.g. ±5%) of its prediction, which will be used for subsequent more complex risk sensitivity analysis or robustness optimization, which is estimated by multiplying the equipment reset cost by the health function based on the exponential function, and then multiplying the risk adjustment coefficient, where the health function is equal to one minus the negative λt power of the natural constant e, λ is the health decay rate parameter, which is equal to the negative residual value rate natural logarithm divided by the rated service life, the residual value rate is generally 5% to 10%, and t is the current running time; the risk adjustment coefficient is equal to one plus the product of the sensitivity parameter α and the actual failure rate and the expected failure rate minus one, which can be dynamically determined through historical data analysis according to factors such as equipment criticality, historical maintenance cost and availability of spare parts, in this embodiment, α takes a value of 0.2 to 0.5.

[0048] Subsequently, the expected business value decay trajectory of the material batch is generated, and the business value decay represents the quantitative decline of physical quality. Records containing the initial quality of the material, environmental temperature and humidity, and process waiting time in the historical data are extracted, and various dynamic or statistical models are used. When the quality decay is mainly dominated by temperature-sensitive chemical reactions, the Arrhenius equation is preferably used to describe the decay rate of key chemical components dominated by temperature. In other cases, for example, when the quality decay appears as a probabilistic failure process (such as package breakage, microbial contamination reaching a threshold, etc.), or a more complex decay model needs to be fitted, statistical models such as Weibull distribution can be used. The Arrhenius equation is used to model the decay rate of pomegranate quality (such as total soluble solids content), which describes the relationship between the quality decay rate constant k and the absolute temperature T of the environment, specifically: the rate constant k is equal to the pre-exponential factor A multiplied by the exponential of the natural constant e, which is the negative activation energy Ea divided by the product of the ideal gas constant R and the absolute temperature T. The pre-exponential factor A and the activation energy Ea in the model are obtained by nonlinear least squares fitting of historical experimental data. Based on the model, the quality decay curve of any batch of material under different waiting times and environmental temperatures corresponding to different scheduling schemes can be predicted, i.e. the expected business value decay trajectory. Finally, the expected performance degradation cost and the expected business value decay trajectory are combined to form a second business data stream containing explicit prediction basis and quantitative results.

[0049] By training business prediction models with historical data, the method quantifies future equipment and business risks into concrete, comparable cost items. In particular, by employing mature kinetic models such as Arrhenius equation to describe the decay trajectory of material value over time and environment, the system is able to foresee the direct impact of different scheduling decisions on the final product business value. This ability to translate future uncertainty into current decision variables enables the system to make more profound economic trade-offs across time dimension, thus achieving preventive, value-preservation-oriented intelligent scheduling.

[0050] Further, the data from the multi-modal sensor array is processed by the product value assessment to generate a third business data stream containing real-time composite quality indicators of the material batches; corresponding to the S3 step above; the detailed implementation process includes:

[0051] First, a first-level fast online assessment system is established. Low-cost, fast-responding sensors such as industrial cameras and weighing sensors are deployed at key nodes of the production line (e.g. after cleaning, before peeling). This system continuously and rapidly detects each passing material batch, obtains easily analyzable appearance feature parameters (such as color uniformity, average size, estimated bruise area ratio) and weight data, and generates preliminary, fast quality assessment results. This result is mainly used for online monitoring of production stability and quickly screening batches with obvious appearance defects.

[0052] Second, a second-level accurate sampling detection system is established. A dedicated bypass or detection station is set up on the production line, and the high-precision multi-modal sensor array (such as hyperspectral imaging instrument and visible / near-infrared spectrometer) is deployed. This system does not detect all materials, but performs accurate sampling detection according to a pre-set strategy. Sampling strategies can include: 1) periodic sampling, such as randomly selecting one batch for in-depth analysis every hour; 2) triggered sampling, when the first-level fast assessment system detects that the quality parameters of a batch exceed the pre-set threshold, the batch is automatically introduced into the second-level system for accurate diagnosis.

[0053] Finally, data fusion and model self-calibration are performed. The detailed appearance features (e.g. precise bruise area) and internal composition features (e.g. total soluble solids content, total acidity) obtained from the second-tier precise detection system are combined, and the precise composite quality index (CQI) of the sampling batch is calculated through a weighted fusion algorithm, which is preferably a linear weighted summation model. Each normalized quality feature parameter value is multiplied by its corresponding weight, and then all the products are added to obtain the final comprehensive quality index score. The weights are objectively determined by regression analysis of historical sales data, including the following steps: 1) Collecting the quality feature parameters and their corresponding final market sales unit prices of historical batches; 2) Establishing a multiple linear regression model that expresses the market sales unit price as the sum of the base price and the product of multiple quality features multiplied by their regression coefficients; 3) Fitting the model using the collected historical data to obtain the standardized regression coefficients of each quality feature parameter; 4) Normalizing the absolute values of the coefficients to obtain the final weights for CQI calculation. More importantly, the precise CQI value and its corresponding rapid evaluation parameters will serve as new, high-quality training samples to periodically verify and calibrate the prediction model of the first-tier rapid evaluation system. In this way, the system can continuously improve the accuracy of large-scale, low-cost detection using a small amount of high-precision data, forming an intelligent, self-adaptive closed-loop quality control system. Finally, the composite quality index output by the hierarchical system, which combines the results of rapid evaluation and precise sampling, constitutes the third business data stream. For example, a complete third business data stream JSON object is as follows: {batchID: PG20250710-A01, timestamp: 2025-07-10T08:45:10.500Z, rawFeatures: {bruiseAreaRatio: 0.02, colorUniformity: 0.85, TSS: 14.5, totalAcidity: 1.9}, normalizedFeatures: {bruiseAreaRatio: 0.13, colorUniformity: 0.81, TSS: 0.65, totalAcidity: 0.47}, activeWeights: {bruiseAreaRatio: 0.30, colorUniformity: 0.20, TSS: 0.35, totalAcidity: 0.15}, CQI_score: 0.785, qualityGrade: A}.

[0054] By deploying a multi-modal sensor array and combining a weighted fusion algorithm, the current comprehensive quality of the material batch is quantified in real time, objectively and multidimensionally. The appearance features are obtained by optical sensors and the internal component features are obtained by spectral sensors, which overcomes the subjectivity, one-sidedness and delay caused by traditional manual sampling or single sensor detection. By generating real-time composite quality indicators, the scheduling system can differentiate the treatment based on the real-time and true value of each batch of materials, such as prioritizing high-value batches, thereby maximizing the use of limited production resources and improving the value conversion rate from raw materials to finished products.

[0055] Further, the first service data stream, the second service data stream and the third service data stream are input into a business decision optimization engine to construct a multi-objective business cost function containing perishable business cost items; corresponding to the S4 step; refer to Figure 2 , the specific implementation process includes:

[0056] First, based on the equipment operating parameters and working state identifiers in the first service data stream, a first cost component for measuring the order delivery cycle is constructed. This component is a quantitative measure of the total cost required to complete all orders under a specific scheduling scheme. The calculation method is to sum the total energy cost and total labor cost of all equipment involved in the scheme. Second, based on the inventory data in the first service data stream, a second cost component for measuring the inventory holding cost and capital occupation is constructed. The calculation method is to integrate the product of the value and holding time of the work-in-process inventory at each buffer node, and then multiply by the capital holding rate per unit time.

[0057] The key of the application is to construct the perishable business cost item as the third cost component. The system provides the real-time composite quality indicator (CQI) provided by the third service data stream as the initial value, and calculates the material quality degradation caused by waiting and processing under different scheduling schemes according to the expected business value decay trajectory provided by the second service data stream, and converts the quality degradation into direct sales revenue loss. The conversion is realized by a preset quality-price function, which describes the business value of the material as an S-shaped curve relationship. Specifically, the business value of the material is equal to the market minimum price plus the additional value increased with the quality improvement. The additional value is obtained by multiplying the difference between the market maximum price and the minimum price by an adjustment coefficient, which is calculated by one plus the reciprocal of a specific power of the natural constant e, which is the product of the difference between the negative steepness coefficient k and the current composite quality indicator CQI and the inflection point CQI_0 most sensitive to the quality-price relationship. All parameters in the function, such as the maximum price, the minimum price, the inflection point and the steepness coefficient, are objectively determined by nonlinear fitting of historical sales data. At the same time, the expected performance degradation cost of the device is constructed as the fourth cost component, which directly uses the expected performance degradation cost of the device generated in the second service data stream to quantify the impact of different scheduling schemes on the long-term asset value of the device. Then, the first, second, third and fourth cost components are combined by weighting to form the multi-objective business cost function. The weights of each cost component are scientifically determined by the analytic hierarchy process. The process includes: first, establish a hierarchical structure model with total cost minimization as the target layer and the four cost components as the criterion layer; then construct a judgment matrix for pairwise comparison, and the decision maker compares the importance of each criterion in pairs according to the 1-9 scale method; then calculate the weight vector of each criterion; finally, perform consistency test by calculating the consistency ratio to ensure it is less than 0.1, ensuring the logical consistency of the decision-making judgment. The weight combination obtained is used to construct the final multi-objective business cost function.

[0058] A comprehensive cost function is constructed that can measure and balance multiple core business objectives. By weighting the order delivery cycle, inventory holding cost and perishable business cost of innovation, a unified optimization target is created. This makes the decision of the optimization engine no longer isolatedly pursuing the optimization of a certain technical index, but intelligently finding a balance point that can minimize the overall business cost of the enterprise within the overall business framework. The unified quantification of multiple objectives is a key technical prerequisite for realizing real business decision optimization.

[0059] Further, according to the multi-objective business cost function, an iterative optimization business simulation process is started to obtain an optimized scheduling scheme; corresponding to the above S5 step; refer to Figure 3 , the specific implementation process includes:

[0060] To solve this scheduling problem, the method initiates an overall business profit-oriented iterative optimization business simulation process, which uses a hybrid meta-heuristic algorithm combining genetic algorithm (GA) and simulated annealing (SA). The system first constructs an initial decision set containing multiple production scheduling schemes, which is the initial population of the genetic algorithm. Each scheduling scheme is represented as a global sequence of all to-be-processed operations using an operation-based encoding method. Then, the system uses the genetic algorithm for global exploration. In each generation, the offspring schemes are generated by performing priority-preserving crossover and swap mutation operations. After each iteration of the genetic algorithm in each generation, a simulated annealing local search is initiated for the optimal individual (elite) in the population. To avoid the algorithm converging to a local optimal solution too early, in addition to the main basis of the comprehensive cost function value, an auxiliary evaluation mechanism aimed at promoting the diversity of solutions is introduced. For example, the Hamming distance between individual solutions in the elite population can be periodically calculated, or a penalty term can be introduced to moderately reduce the probability of selecting solutions that are too similar to existing elite individuals. This mechanism helps the algorithm strike a better balance between exploration and exploitation, thereby improving the likelihood of finding high-quality solutions while maintaining the diversity of solutions to address different production scenarios. This improves the robustness and practicality of the algorithm. In the face of sudden production changes, there is always an adaptive scheme in the elite pool, which shortens the decision response time and improves the stability of the scheme quality. SA generates neighborhood solutions by reordering two adjacent operations on key equipment, etc., to explore and lock in marginal profit growth points. If the result of the SA search is better than the original elite individual, it is replaced. The optimized elite individual then participates in the crossover and mutation process of the next generation. This iterative process continues until the expected profit gain from adjustment is lower than the preset business sensitivity threshold. This threshold is not a fixed value, but a multi-factor dynamic calculation based on current production costs, market volatility, and order urgency. According to historical optimization data, a marginal revenue diminishing model is established. When the profit gain of three consecutive iterations is less than 80% of the current threshold, the sensitivity threshold is automatically reduced for further optimization. When market demand surges, the threshold is increased to respond quickly. This avoids excessive iteration in low-value areas while ensuring that optimization space is not missed when high-value opportunities arise. It achieves intelligent allocation of computing resources, reduces some invalid calculation time compared to the fixed threshold method, and improves the profit optimization effect of key orders, truly focusing the algorithm's resources on business value and intelligently converging. The production scheduling scheme with the highest expected comprehensive business value is determined as the optimized scheduling scheme. In addition, to ensure the applicability of the algorithm in different production scenarios, the method does not use fixed algorithm parameters, but introduces experimental design techniques such as Taguchi method. Through a series of simulation experiments, the optimal algorithm parameter combinations such as crossover probability and cooling rate are systematically found under different production constraints, changing the one-time parameter selection to a repeatable optimization method.

[0061] By distinguishing between the two stages of global exploration and local optimization, it can effectively avoid falling into the trap of local optimal solution; the global exploration stage quickly identifies the high potential profit interval by adjusting the macro operation strategy; while the local optimization stage carries out fine-tuning in these high potential intervals, and explores and locks the growth point of marginal profit. This combination of rough and fine tuning strategy ensures that the algorithm has both breadth and depth, and can converge to a scheduling scheme with significantly better business value at a reasonable computational cost.

[0062] Further, the optimized scheduling scheme generates corresponding digital production work orders, which are transmitted to the processing operation nodes through the interface with the manufacturing execution system to control the production line and perform closed-loop management and business performance tracking on the production line; corresponding to the above S6 step; the specific implementation process includes:

[0063] By analyzing the optimized scheduling scheme, the specific processing sequence, starting time and estimated completion time of each material batch at each processing operation node are extracted. A structured digital production work order is generated for each processing operation node. The work order is transmitted in real time to the manufacturing execution system through the OPC UA standardized interface and is issued to the corresponding processing node. An example of a JSON digital production work order is as follows: {workOrderID: WO-2025-A-001, batchID: PG20250710-A01, nodeID: Peeler_01, startTime: 2025-07-10T09:00:00Z, estimatedEndTime: 2025-07-10T09:25:00Z, controlCommands: {speed: 1200, temperature: 25}, feedbackFields: {actualEndTime: null, materialConsumed_kg: null, qualityCheckResult: null}}. The feedbackFields field in the work order is used to record and return the actual completion time, material consumption and quality detection results of each node. These completed work orders with actual execution results form a closed-loop data flow, which is used for business performance tracking on the one hand, and as new training samples input into the business prediction model and the dynamics model on the other hand. Through this periodic model retraining mechanism, the system can adaptively correct and optimize its prediction accuracy, for example, by adjusting the equipment reliability parameters according to the actual equipment failure frequency, and updating the value decay coefficient according to the actual material loss, so that the decision quality of the entire system is continuously improved over time, forming a benign and self-adaptive intelligent manufacturing ecosystem.

[0064] By parsing the optimized scheduling scheme into structured digital production work orders and issuing them through the interface with the manufacturing execution system, it is ensured that the optimized decisions can be accurately and automatically executed. More importantly, by including data fields for returning actual execution results in the work orders, a continuous learning and self-improving mechanism is established. The returned actual data is not only used for performance tracking, but also used as new training samples to iteratively update and optimize the prediction model, so that the prediction accuracy and decision quality of the entire system are continuously improved over time, forming a benign and self-adaptive intelligent manufacturing ecosystem.

[0065] By integrating real-time physical data streams, predictive business data streams, and real-time product quality data streams and inputting them into a multi-objective optimization engine that includes perishable business cost items, the present application creates a decision-making framework that can dynamically balance multiple, even conflicting business objectives such as production efficiency, equipment depreciation, inventory cost, and product business value decay. By converting the optimization scheme into digital work orders and implementing closed-loop management, this method not only improves the intelligence and adaptability of production scheduling, but more importantly, it directly links production operations to the core financial objectives of the enterprise, thereby maximizing overall business profits in the production environment of perishable products.

[0066] Embodiment Two

[0067] In the production line of Pomegranate Peeling Factory B, the specific implementation of the three business data streams is as follows:

[0068] First, the system collects and processes the first business data stream in real time from the Internet of Things interface. The system periodically acquires device layer data and inventory data through the sensor groups associated with each processing unit in the production line. For example, the sensor group of the cleaning processing unit collects operating parameters such as motor power 52kW and rotational speed 1200rpm, as well as working state identifiers; the sensor group of the peeling processing unit collects data such as vibration acceleration. At the same time, the ultrasonic level sensor of the raw material buffer zone monitors that the pomegranate raw material inventory is 2.8 tons, and the photoelectric sensor of the clean fruit buffer zone monitors that the clean fruit inventory is 2.1 tons. At the data processing node, the system cleans, standardizes the format, and aligns the timestamps of these data to form the first business data stream.

[0069] Next, the system inputs the historical sensor data into the business prediction model to generate a second business data stream. To generate the expected performance degradation cost of the equipment, the system extracts a feature vector from the historical data, trains a long short-term memory (LSTM) model, and predicts the remaining useful life of the threshing machine to be 2847.5 hours. According to the equipment health cost quantification method, when dealing with a threshing machine with a reset cost of 20000 yuan, a rated total useful life of 10000 hours, and a current running time of 7152.5 hours, the following calculations can be made: first, set the residual value rate to 8% and calculate the health decay rate parameter λ to be about 0.00025257; based on this rate and the running time, the health function is obtained to be 0.8358. At the same time, in the case that the current actual failure rate is basically consistent with the expected failure rate, the risk adjustment coefficient is 1. Finally, the equipment reset cost, the health function, and the risk adjustment coefficient are multiplied to finally estimate the expected performance degradation cost of the equipment to be about 16716 yuan. To generate the expected business value decay trajectory of the material batch, the system extracts historical data to train a kinetic model for describing the quality decay law. The model uses the Arrhenius equation, and by fitting the historical experimental data with the nonlinear least squares method, the key parameters of the model are determined: the activation energy (Ea) is 75 kJ / mol, and the pre-exponential factor (A) is 1.87 x 10 11 h -1 . Thus, it is calculated that under the current 24.5 degrees Celsius environment, the quality decay rate of the material batch PG20250709_B_002847 is 0.0142 per hour. This decay trajectory will be used for subsequent cost calculation. Finally, the calculated expected performance degradation cost and the expected business value decay trajectory are combined to form the second business data stream.

[0070] First, the material batch PG20250709_B_002847 flows through the first level of the fast online evaluation system. The industrial camera preliminarily evaluates its color uniformity as 0.75 and the estimated bruise area ratio as 0.04, slightly higher than the average level of normal batches. Then, since the estimated bruise area ratio triggers the preset threshold, the system initiates the second level of the precise sampling detection. The batch is guided to the detection station for in-depth analysis by the hyperspectral imager and the visible / near-infrared spectrometer. The analysis results precisely identify the bruise area ratio of the batch as 0.038, calculate the color uniformity index as 0.762, and non-destructively predict the soluble solid content as 13.8% and the total acidity as 2.14 g / 100 g. Subsequently, the system normalizes these precise feature parameters. For example, the bruise area ratio is mapped to the interval [0, 0.15], the color uniformity is mapped to the interval [0.3, 1.0], the soluble solid is mapped to the interval [8.0, 18.0], and the total acidity is mapped to the interval [0.5, 3.5], respectively obtaining the normalized values as 0.253, 0.660, 0.580, and 0.547. According to the weight configuration determined by the regression of historical sales prices (where the bruise area weight is 0.25, the color uniformity weight is 0.20, the soluble solid weight is 0.15, the acidity weight is 0.10, and the weights of other features are 0.30 in total), the final precise composite quality index (CQI) of the material batch is calculated as 0.695 through the weighted fusion algorithm. Finally, the precise CQI value 0.695 is encapsulated to form the third business data stream. At the same time, this set of data (fast evaluation parameters and precise CQI value) is stored in the training database for subsequent iteration optimization of the first level evaluation model to improve its estimation accuracy.

[0071] Example Three

[0072] In the production line of the pomegranate de-seeding factory B, referring to Figure 2 The first business data stream, the second business data stream, and the third business data stream are input into the commercial decision optimization engine to construct a specific implementation of a multi-objective business cost function containing perishable commercial cost items as follows:

[0073] First, the system builds the first cost component based on the data in the first business data stream. According to the equipment operating parameters, the cleaning equipment runs for 6 hours at a power of 52 kW, the threshing equipment runs for 5.765 hours at a power of 43 kW, and the packaging equipment runs for 4.5 hours at a power of 26 kW. According to the industrial electricity price of 0.65 yuan per kilowatt-hour, the total energy consumption cost is 439.98 yuan. Adding the known labor cost of 2456 yuan, the first cost component (order delivery cycle cost) is 2895.98 yuan. Next, the system builds the second cost component based on the real-time inventory data in the first business data stream. The raw material inventory is 2.8 tons, with a unit price of 9.2 yuan per kilogram; the inventory after cleaning is 2.1 tons, with a unit price of 10.7 yuan per kilogram; and the inventory after threshing is 0.7 tons, with a unit price of 48.3 yuan per kilogram. According to the annual holding cost of 8.5%, the system calculates the hourly holding cost and integrates the work-in-process inventory value during the entire scheduling period to obtain the second cost component (inventory holding cost) of 182.5 yuan.

[0074] Subsequently, the system builds the third cost component, which is the perishable business cost, which is the core of the present application. Using an initial composite quality index (CQI) of 0.695 and a quality decay rate of 0.0142 per hour, the system adjusts the quality-price S-shaped function, which is determined by historical data fitting to have parameters: a market maximum price of 65 yuan per kilogram, a market minimum price of 10 yuan per kilogram, a quality-price relationship most sensitive inflection point CQI_0 of 0.75, and a curve steepness coefficient k of 10. According to this function, the system first calculates the initial business value, which is equal to the market minimum price of 10 yuan plus an additional value, which is the difference between the maximum and minimum prices (55 yuan) multiplied by an adjustment coefficient. The adjustment coefficient is calculated by taking the reciprocal of one plus the natural constant e raised to a certain power, which is the product of the difference between the negative steepness coefficient 10 and the initial CQI value 0.695 and the inflection point 0.75. The calculation results in an initial business value of 30.12 yuan per kilogram. After a total processing time of 5.25 hours, the system predicts the decay amount of each key internal component affecting CQI based on the comprehensive decay model built around the Arrhenius equation. Subsequently, based on the predicted values of these components, the system recalculates the final CQI prediction value by a pre-set weighted fusion algorithm, resulting in a decrease to 0.646. Again using the quality-price function to calculate the final business value, the system obtains 14.36 yuan per kilogram. Therefore, the expected direct sales revenue loss due to quality degradation is 15.76 yuan per kilogram. According to the weight of this batch of 680 kilograms, the system obtains an expected value of 10716.8 yuan for the third cost component (perishable business cost).

[0075] Next, the system constructs a fourth cost component: the expected performance degradation cost of the equipment. This cost component quantifies the impact of different scheduling options on the long-term asset value of the equipment. Its value is directly derived from the expected performance degradation cost in the second service data stream and is determined to be 16,716 yuan.

[0076] Finally, the system forms a multi-objective business cost function by weighting the four cost components. The weights are determined using the Analytic Hierarchy Process (AHP). In this example, to balance short-term production benefits with long-term asset maintenance, the weights are set as follows: order delivery weighted at 0.3, inventory holding weighted at 0.2, perishability cost weighted at 0.3, and expected equipment performance degradation cost weighted at 0.2. Each cost component is multiplied by its corresponding weight and summed: (2,895.98 yuan multiplied by 0.3) plus (182.5 yuan multiplied by 0.2) plus (10,716.8 yuan multiplied by 0.3) plus (16,716 yuan multiplied by 0.2). The final calculated multi-objective business cost function value for this scheduling scheme is 7,463.534 yuan.

[0077] Example 4

[0078] In the production line of pomegranate peeling factory B, see Figure 3 , based on the multi-objective business cost function, the iterative optimization business simulation process is started to obtain the specific implementation method of the optimized scheduling solution as follows:

[0079] The system initiates an overall business profit-oriented iterative optimization business simulation process, which uses a hybrid meta-heuristic algorithm combining global exploration with genetic algorithm (GA) and local optimization with simulated annealing (SA). First, an initial decision set containing production scheduling schemes is constructed, which is the initial population of the genetic algorithm. The population size is set to 50 individuals, and each scheduling scheme uses an operation-based encoding method to represent the global sequence of all operations to be processed. In each iteration of the genetic algorithm, the system evaluates the fitness of each scheduling scheme according to the multi-objective business cost function. The lower the cost, the higher the fitness. The system uses a priority-preserving crossover operator (crossover probability set to 0.8) and a swap mutation operator (mutation probability set to 0.05) to generate offspring schemes and explore new, high-potential macro-operation strategies. After each iteration of the genetic algorithm, the system initiates a simulated annealing local search for the optimal individual (elite) in the population. The initial temperature of the SA algorithm is set to 1000, and the cooling rate is set to 0.95. It generates neighborhood solutions by reordering two adjacent processes on key equipment, etc., to explore and lock the marginal profit growth point. The better scheme obtained through local optimization will be fed back to the genetic algorithm population, replacing the original elite individual. This iterative process continues until the optimal solution improvement rate of consecutive generations is less than the pre-set business sensitivity threshold of 0.1%, or the maximum number of iterations is reached, which is 500 generations. When the algorithm converges, the system determines the production scheduling scheme with the lowest total business cost found during the entire optimization process as the optimal scheduling scheme.

[0080] Example Five

[0081] In the production line of Pomegranate Peeling Factory B, the specific implementation of generating a corresponding digital production work order for the optimal scheduling scheme is as follows:

[0082] The system first parses the optimized scheduling scheme and extracts the specific processing sequence, starting time and estimated completion time of each material batch at each processing operation node. For example, for batch PG20250709_B_002847, the system parses that its sequence in the cleaning process is the 3rd, the starting time is 09:30, and the estimated completion time is 09:58. According to this information, the system generates a structured digital production work order containing unique job identification, material batch information and control instructions for the corresponding processing operation node. The work order is transmitted in real time to the manufacturing execution system through the OPC UA standardized interface, and is issued by the system to the corresponding processing operation node to perform automatic control. The JSON example of the digital production work order is as follows: {workOrderID: WO-2025-B-1847, batchID: PG20250709_B_002847, nodeID: Cleaner_01, startTime: 2025-07-10T09:30:00Z, estimatedEndTime: 2025-07-10T09:58:00Z, controlCommands: {speed: 520, temperature: 38}, feedbackFields: {actualEndTime: null, materialConsumed_kg: null, qualityCheckResult: null}}.

[0083] The structured digital production work order contains data fields for recording and returning the actual execution results of each node, i.e. feedbackFields. After the work order is executed, these fields are filled with actual data, such as actual completion time, actual material consumption and final quality detection result. These completed work orders with actual execution results form a closed-loop feedback data stream. On the one hand, the data stream is transmitted back to the business decision optimization engine for business performance tracking and accurate cost accounting analysis; on the other hand, and more importantly, it is input into the business prediction model as new and high-quality training samples. The system uses these real-time feedback data to establish a periodic model retraining mechanism, for example, to update the mapping relationship of equipment performance degradation according to the actual equipment failure frequency and maintenance records, or to correct the parameters of the value decay trajectory according to the actual material loss. Through this mechanism, the prediction accuracy of the system is continuously optimized.

[0084] Although embodiments of the present application have been shown and described, it is to be understood that various modifications, substitutions, replacements and changes can be made to these embodiments without departing from the principles and spirit of the present application, and the scope of the present application is defined by the appended claims and their equivalents.

Claims

1. A multi-process collaborative scheduling optimization method for a pomegranate peeling production line, characterized in that: include: The IoT interface collects real-time equipment-level data, including equipment operating parameters and working status indicators, through sensor groups associated with each processing unit in the production line. At the same time, non-contact sensors deployed along the material flow paths between processing units acquire inventory data representing the number of work-in-progress between processes. and processing the equipment layer data and the inventory data to obtain a first business data stream reflecting the physical status of equipment and materials in the production line; Extract feature vectors related to the health status of the equipment from historical sensor data and input them into a business prediction model to establish a mapping relationship between equipment performance degradation and historical operating time, thereby generating the expected performance degradation cost of the equipment due to operation. Extract historical data containing records of the initial quality of the material, ambient temperature and humidity, and waiting time for each process, and train a kinetic model to describe the decay law of material quality indicators over time and environmental changes. Develop an expected commercial value decay trajectory for the material batch, and generate a second business data stream containing the expected performance degradation cost of the equipment and the expected commercial value decay trajectory for the material batch. Performing a product value assessment on image data including appearance features of the material batch and spectral data of internal components of the material batch from a multimodal sensor array, extracting appearance feature parameters from the image data and extracting internal component feature parameters from the spectral data; combining the appearance feature parameters with the internal component feature parameters using a preset weighted fusion algorithm to generate a third business data stream containing a composite quality indicator of the material batch; Input the first business data stream, the second business data stream and the third business data stream into the business decision optimization engine, construct a first cost component for measuring the order delivery cycle, construct a second cost component for measuring inventory holding costs and capital occupation, calculate the material quality reduction caused by waiting and processing under different scheduling schemes, and convert the quality reduction into direct sales revenue loss, construct a perishable business cost item as the third cost component, and construct a fourth cost component for measuring the impact of the scheduling scheme on the long-term asset value of the equipment; construct a multi-objective business cost function including perishable business cost items by weighted combination of the first cost component, the second cost component, the third cost component and the fourth cost component; construct an initial operation cost function including the production scheduling scheme An initial decision set is generated, and the expected comprehensive business value of each production scheduling solution in the set is quantitatively evaluated based on the multi-objective business cost function; an iterative optimization business simulation process is initiated, and by adjusting the allocation sequence of production resources and the processing priority of material batches, a macro-operation strategy is generated and evaluated to identify high-potential profit ranges; local optimization is performed on the production scheduling solutions within the profit range, and by fine-tuning the processing time window and reordering adjacent processes, marginal profit growth points are discovered and locked in; the optimized production scheduling solution is fed back to the initial decision set to improve the business value benchmark of the overall solution; the iterative process continues, and converges when the expected profit gain brought about by the adjustment is lower than a preset business sensitivity threshold, thereby obtaining an optimized scheduling solution; The optimized scheduling plan extracts the specific processing sequence, start time and estimated completion time of each material batch at each processing operation node; based on the extracted information, a digital production work order containing a unique job identifier, material batch information and control instructions is generated for each processing operation node, and is sent to the processing operation node through an interface with the manufacturing execution system to control the production line. The digital production work order contains data fields for recording and returning the actual completion time, material consumption and quality inspection results of each node, so as to implement closed-loop management and business performance tracking of the production line.

Images