Method for collaborative optimization of kitchen waste grease separation process based on dynamic game theory

Abstract

The present application relates to the technical field of big data analysis, in particular to a kitchen waste oil separation process collaborative optimization method based on dynamic game theory; initial state data of a to-be-processed resource flow is acquired, impact factors in the to-be-processed resource flow are identified and quantified through an asset risk assessment model; based on the identification results of the impact factors, the to-be-processed resource flow is sorted to obtain an optimized resource flow; before the optimized resource flow enters a processing device, component data of the optimized resource flow is acquired; after the optimized resource flow is processed by the processing device, process characteristic data of the optimized resource flow is collected; a dynamic game optimization decision model is established to solve the acquired component data, to determine optimal operation data; and the dynamic game optimization decision model is fed back and calibrated using the collected process characteristic data. Through analysis and processing of the to-be-processed resource flow data, adaptive regulation and control of the processing device is realized.

Description

A Collaborative Optimization Method for Food Waste Oil Separation Based on Dynamic Game Theory Technical Field

[0001] This invention relates to the field of big data analytics, specifically to a collaborative optimization method for the separation of grease from kitchen waste based on dynamic game theory. Background Technology

[0002] In manufacturing companies that deal with dynamic and variable resources, the core operational management objective is to optimize workflows to maximize overall business value.

[0003] Current management paradigms typically rely on static, standard operating procedures (SOPs) based on historical averages. This model fails to provide effective data analysis of the real-time status of resource flows, leading to delayed adjustments in operational strategies. Furthermore, it lacks the ability to dynamically analyze costs and benefits to determine optimal resource allocation and workload at specific points in time. Decision-makers lack the quantitative tools to predict the combined impact of an operational adjustment on short-term productivity and long-term operating costs.

[0004] A company's profitability depends on balancing multiple conflicting key performance indicators (KPIs). Existing systems lack built-in predictive analytics and mathematical modeling capabilities, failing to provide managers with forward-looking decision-making insights. Operational decisions often remain at a reactive level rather than proactive, data-driven predictive planning.

[0005] To address this, a collaborative optimization method for the separation of grease from kitchen waste based on dynamic game theory is proposed. Summary of the Invention

[0006] The purpose of this invention is to provide a collaborative optimization method for the separation process of kitchen waste oil based on dynamic game theory, which achieves adaptive control of the processing equipment by dynamically balancing the crushing benefits and emulsification costs during the oil separation process.

[0007] To achieve the above objectives, the present invention provides the following technical solution:

[0008] A collaborative optimization method for the separation of grease from kitchen waste based on dynamic game theory includes:

[0009] The initial state data of the resource stream to be processed is obtained through sensing devices.

[0010] Based on initial state data, the asset risk assessment model is used to identify and quantify the factors that pose operational risks to the processing equipment in the resource stream to be processed.

[0011] Based on the identification results of influencing factors, the resource flow to be processed is sorted to obtain an optimized resource flow;

[0012] Before the optimized resource stream enters the processing equipment, the component data of the optimized resource stream is acquired; after the optimized resource stream is processed by the processing equipment, the process characteristic data of the optimized resource stream is collected.

[0013] A dynamic game optimization decision model is pre-established, using the operating parameters of the processing equipment as decision variables and incorporating a mathematical function that balances crushing benefits and emulsification costs. During real-time operation, the dynamic game optimization decision model is used to solve the acquired component data to determine the optimal operating data. The collected process characteristic data is then used to provide feedback calibration to the dynamic game optimization decision model, enabling adaptive control of the processing equipment.

[0014] By utilizing deployed image sensors, image data of the resource stream to be processed is continuously captured, forming an image data stream;

[0015] By using a dynamic weighing sensor installed below the resource stream to be processed, the mass data of the resource stream to be processed is acquired in real time, and the density distribution characteristics of the resource stream to be processed are calculated by combining the volume measured by the laser sensor.

[0016] The image data stream, quality data, and density distribution features are timestamped and fused to construct the initial state data of the resource stream to be processed.

[0017] The asset risk assessment model includes image data analysis, anomaly area identification, and influencing factor identification;

[0018] The image data analysis is based on a convolutional neural network target detection model to identify the image data stream in the initial state data, analyze the stacking structure, color distribution and texture information of the resource stream to be processed, and obtain image feature data.

[0019] The abnormal region judgment layer identifies unprocessed resource stream regions with target detection anomalies and density anomalies based on image feature data and corresponding density distribution features, and these regions are designated as abnormal regions.

[0020] The influencing factor identification identifies the influencing factors in abnormal areas. The identification results are obtained based on the size, shape, and density of the influencing factors. Combined with the operational interference caused by similar objects in historical data to the processing equipment, an operational risk index is output.

[0021] A near-infrared spectrometer is deployed at the feed inlet of the processing equipment to scan the optimized resource stream entering the processing equipment and obtain the composition data of the optimized resource stream; the composition data includes moisture content, oil content, protein content and cellulose content.

[0022] Online viscometers and laser particle size analyzers are deployed at the outlet of the processing equipment to continuously monitor and optimize the physical properties of the resource flow after processing, including viscosity data obtained from online viscometer monitoring and particle size data obtained from laser particle size analyzer monitoring.

[0023] The construction process of the dynamic game optimization decision model includes:

[0024] Acquire the component data, process characteristic data, equipment operating parameters, crushing revenue, and emulsification cost of the complete processing of the processing equipment; the operating parameters include equipment rotation speed and equipment spacing;

[0025] Construct a breakage revenue function and an emulsification cost function based on a neural network model;

[0026] The crushing revenue function takes component data and decision variables as input and crushing revenue as output; the emulsification cost function takes component data and decision variables as input and emulsification cost as output.

[0027] The fragmentation payoff function and the emulsification cost function are embedded within the dynamic game optimization decision model; wherein, the difference between the fragmentation payoff function and the emulsification cost function is defined as the net payoff objective function used for the final decision.

[0028] The adaptive control is a command-based closed-loop execution process;

[0029] The optimal operational data calculated by the dynamic game optimization decision model is formatted into control commands and sent to the underlying programmable logic controller of the processing equipment via the industrial Ethernet protocol.

[0030] The programmable logic controller (PLC) precisely adjusts the equipment speed and equipment spacing through the frequency converter based on the received instructions.

[0031] Meanwhile, the effects of equipment processing are continuously monitored through process characteristic data, and feedback is given to the dynamic game optimization decision-making model.

[0032] Dynamic game optimization decision-making models use feedback as input for the next round of decision-making, which is used to modify their internal functions and the policy output at the next time step.

[0033] Compared with the prior art, the beneficial effects of the present invention are as follows:

[0034] 1. This invention establishes a quantifiable risk assessment model, based on the feature identification and analysis of kitchen waste, to transform vague operational risks into precise values; it realizes automated, high-precision identification and removal of harmful impurities, ensuring the safe and stable operation of subsequent crushing equipment from the source, and significantly reducing downtime maintenance costs caused by stalling or wear.

[0035] 2. This invention transforms a complex production trade-off problem into a clear mathematical optimization problem. By simulating the rise and fall of the two sides in a game, it calculates in real time the optimal equipment operating parameters for energy efficiency and output under the current material composition, realizing a leap from experience-based operation to data-driven refined decision-making.

[0036] 3. This invention constitutes a complete "decision-execution-feedback-learning" closed-loop control system; the online calibration mechanism enables the model to continuously learn and automatically adapt to model drift caused by seasonal changes in raw materials, batch differences, and equipment wear, ensuring the long-term accuracy and robustness of optimization decisions and realizing adaptive control of the entire processing process. Attached Figure Description

[0037] Figure 1 is a flowchart illustrating the collaborative optimization method for the separation of grease from kitchen waste based on dynamic game theory according to the present invention.

[0038] Figure 2 is a schematic diagram of the asset risk assessment model of the present invention;

[0039] Figure 3 is a schematic diagram of the adaptive control of the oil separation process of the present invention. Detailed Implementation

[0040] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0041] Example 1:

[0042] This invention proposes a collaborative optimization method for the separation of grease from kitchen waste based on dynamic game theory. The process is shown in Figure 1, and includes:

[0043] The initial state data of the resource stream to be processed is obtained through sensing devices.

[0044] Based on initial state data, the asset risk assessment model is used to identify and quantify the factors that pose operational risks to the processing equipment in the resource stream to be processed.

[0045] Based on the identification results of influencing factors, the resource flow to be processed is sorted to obtain an optimized resource flow;

[0046] Before the optimized resource stream enters the processing equipment, the component data of the optimized resource stream is acquired; after the optimized resource stream is processed by the processing equipment, the process characteristic data of the optimized resource stream is collected.

[0047] A dynamic game optimization decision model is pre-established, using the operating parameters of the processing equipment as decision variables and incorporating a mathematical function that balances crushing benefits and emulsification costs. During real-time operation, the dynamic game optimization decision model is used to solve the acquired component data to determine the optimal operating data. The collected process characteristic data is then used to provide feedback calibration to the dynamic game optimization decision model, enabling adaptive control of the processing equipment.

[0048] The processing equipment is a food waste crushing device; the resource stream to be processed is food waste to be processed.

[0049] Preferably, image data of the resource stream to be processed is continuously captured using deployed image sensors to form an image data stream;

[0050] By using a dynamic weighing sensor installed below the resource stream to be processed, the mass data of the resource stream to be processed is acquired in real time, and the density distribution characteristics of the resource stream to be processed are calculated by combining the volume measured by the laser sensor.

[0051] The image data stream, quality data, and density distribution features are timestamped and fused to construct the initial state data of the resource stream to be processed.

[0052] This invention uses an image sensor to identify and acquire image data of kitchen waste to be processed, and uses the sensor to acquire real-time quality and volume data to obtain density distribution characteristics. The image data stream, quality data and density distribution characteristics are timestamped and fused to accurately reflect the initial state of the kitchen waste.

[0053] Preferably, the structure of the asset risk assessment model is shown in Figure 2, including image data analysis, abnormal area judgment, and influencing factor identification;

[0054] The image data analysis is based on a convolutional neural network target detection model to identify the image data stream in the initial state data, analyze the stacking structure, color distribution and texture information of the resource stream to be processed, and obtain image feature data.

[0055] The abnormal region judgment layer identifies unprocessed resource stream regions with target detection anomalies and density anomalies based on image feature data and corresponding density distribution features, and these regions are designated as abnormal regions.

[0056] The influencing factor identification identifies the influencing factors in abnormal areas. The identification results are obtained based on the size, shape, and density of the influencing factors. Combined with the operational interference caused by similar objects in historical data to the processing equipment, an operational risk index is output.

[0057] The image data analysis incorporates a convolutional neural network model based on the YOLOv5 architecture. This model has been trained on a large number of food waste images and can accurately identify several specific influencing factors, including: "large bones", "metal cutlery (knives, forks, spoons)", "ceramic fragments", "plastic packaging bags" and "long strips of fabric (rags, ropes)".

[0058] Furthermore, multimodal feature fusion can be used for judgment, that is, instead of looking at density values ​​or image features in isolation, density features and image features are fused more deeply. By allowing features from different sources to corroborate each other, the false judgment rate can be significantly reduced, making the judgment more accurate.

[0059] For example, the model can learn that regions with "high density" and exhibiting "metallic reflection" or "regular edges" in an image have a much higher risk weight than regions with "high density" but an image texture resembling "meat fibers." In this way, by allowing features from different sources to corroborate each other, the false positive rate can be significantly reduced, leading to more accurate judgments.

[0060] The logic for judging abnormal regions is as follows: if the YOLOv5 model identifies any of the above-mentioned specific influencing factors in the image, the region is marked as a target detection anomaly; at the same time, the average density of the material has been calculated, and if the density of a certain region calculated by the current laser sensor and weighing sensor fluctuates compared with the average value, it is marked as a density abnormal region.

[0061] The process of determining the operational risk index is as follows: a risk knowledge base of influencing factors is constructed; the risk knowledge base is collected based on historical data and includes the size, shape, density, and risk labels of the influencing factors; the risk labels are determined based on the difference between the equipment state of the processing equipment and the standard equipment state when processing the influencing factors; wherein, the standard equipment state is the equipment state of the processing equipment when there are no influencing factors; the equipment state is determined based on the vibration frequency of the processing equipment, or it can be determined based on the degree of wear of the equipment.

[0062] The operational risk index is determined based on the correlation between the size, shape, density, and other factors of influencing factors and risk labels. This determination can be based on a neural network model. Specifically, it involves constructing a three-layer feedforward neural network model. The input layer of this model receives three normalized parameters: the size, density, and shape of the influencing factor. The output layer is a single neuron, outputting a risk index ranging from 0 to 1. The model is trained using historical data. Each sample in the database contains its size, density, and shape, as well as the risk label corresponding to the peak equipment vibration it causes during processing. This label is a value between 0 and 1 obtained by normalizing or classifying the vibration peak value.

[0063] After identifying areas containing factors influencing high operational risk, the location coordinates and timestamp information of these areas are sent to a sorting device deployed downstream of the conveyor belt. Based on the received instructions, the sorting device converts the target location in the image sensor coordinate system to a gripping point in the robot coordinate system, removing the influencing factors.

[0064] This invention establishes a quantifiable risk assessment model, based on the feature identification and analysis of kitchen waste, to transform vague operational risks into precise values; it achieves automated, high-precision identification and removal of harmful impurities, ensuring the safe and stable operation of subsequent crushing equipment from the source, and significantly reducing downtime maintenance costs caused by stalling or wear.

[0065] Furthermore, an unsupervised or semi-supervised anomaly detection model is introduced into the asset risk assessment model. This model specifically learns the visual and density characteristics of "normal" food waste. During real-time processing, any area that significantly deviates from the learned "normal" pattern, regardless of its specific object, is marked as an anomaly. This approach can work in parallel with the existing YOLOv5 model, forming a complementary relationship: YOLOv5 is responsible for efficiently and accurately detecting known high-frequency risk objects, while the anomaly detection model serves as a supplement, discovering all unknown "abnormal" risks, thereby significantly improving the comprehensiveness of the risk assessment.

[0066] Preferably, a near-infrared spectrometer is deployed at the feed inlet of the processing equipment to scan the optimized resource stream entering the processing equipment and obtain the composition data of the optimized resource stream; the composition data includes water content, oil content, protein content and cellulose content data;

[0067] Online viscometers and laser particle size analyzers are deployed at the outlet of the processing equipment to continuously monitor and optimize the physical properties of the resource flow after processing, including viscosity data obtained from online viscometer monitoring and particle size data obtained from laser particle size analyzer monitoring.

[0068] This invention obtains viscosity and particle size data after optimizing resource flow processing by continuously monitoring and optimizing the physical properties of the processed resources. It enables online and real-time monitoring of materials entering the processing equipment, providing key feedforward information for subsequent optimization decision-making models. This allows the system to proactively adjust its strategy based on real-time changes in the materials, rather than responding with a lag.

[0069] Preferably, the construction process of the dynamic game optimization decision model includes:

[0070] Acquire component data, process characteristic data, equipment operating parameters, crushing revenue, and emulsification cost of the complete processing process of the processing equipment; the operating parameters include equipment rotation speed and equipment spacing; the equipment rotation speed is the rotation speed of the processing equipment cutters; the equipment spacing is the aperture size of the crusher screen.

[0071] A fragmentation revenue function and an emulsification cost function are constructed based on a neural network model. The fragmentation revenue function takes component data and decision variables as input and fragmentation revenue as output. The emulsification cost function takes component data and decision variables as input and emulsification cost as output.

[0072] The fragmentation payoff function and the emulsification cost function are embedded within the dynamic game optimization decision model; wherein, the difference between the fragmentation payoff function and the emulsification cost function is defined as the net payoff objective function used for the final decision.

[0073] This invention views the process of separating grease from kitchen waste as a game, with the two sides being the crushing revenue and emulsification cost of the kitchen waste after crushing treatment.

[0074] The aforementioned crushing benefit refers to the finer crushing of food waste by increasing rotation speed and reducing spacing, thereby releasing as much oil as possible. The aforementioned emulsification cost is due to the fact that excessive crushing leads to the formation of a stable emulsion of oil, water, and protein, resulting in a significant increase in viscosity. This not only makes subsequent oil-water separation extremely difficult and costly but also increases equipment energy consumption. Specifically, the crushing benefit refers to the economic efficiency of the separated oil during the complete oil separation process; the emulsification cost refers to the cost incurred in handling the emulsification phenomenon during oil separation.

[0075] Both the crushing revenue function and the emulsification cost function are constructed using a feedforward neural network with three hidden layers (128 neurons per layer, using the ReLU activation function). Experiments were conducted using kitchen waste of different compositions under various combinations of operating parameters to collect corresponding input vectors, actual oil release rates, and outlet viscosities, forming a training dataset. Mean squared error was used as the loss function, and the network was trained using the Adam optimizer until the model converged.

[0076] In real-time operation, the acquired component data is solved using a dynamic game optimization decision model. A particle swarm optimization algorithm is employed to search for the maximum value of the objective function within the range of rotational speed and spacing. This includes generating combinations of multiple equipment rotational speeds and equipment spacings. For each combination, the expected crushing revenue and emulsification cost are calculated, and then the net revenue is obtained by subtracting the cost from the revenue. Ultimately, the rotational speed-spacing combination that produces the highest net revenue is the optimal operating data under the current conditions.

[0077] This invention transforms a complex production trade-off problem into a clear mathematical optimization problem. By simulating the rise and fall of the two sides in a game, it calculates in real time the optimal equipment operating parameters for energy efficiency and output under the current material composition, achieving a leap from experience-based operation to data-driven refined decision-making.

[0078] Preferably, the adaptive control is a command-based closed-loop execution process, the logic of which is shown in Figure 3:

[0079] The optimal operational data calculated by the dynamic game optimization decision model is formatted into control commands and sent to the underlying programmable logic controller of the processing equipment via the industrial Ethernet protocol.

[0080] The programmable logic controller (PLC) precisely adjusts the equipment speed and equipment spacing through the frequency converter based on the received instructions.

[0081] Meanwhile, the effects of equipment processing are continuously monitored through process characteristic data, and feedback is given to the dynamic game optimization decision-making model.

[0082] Dynamic game optimization decision-making models use feedback as input for the next round of decision-making, which is used to modify their internal functions and the policy output at the next time step.

[0083] Feedback calibration is achieved through an online learning mechanism. At the end of each decision cycle, a new training sample is constructed from the optimized resource flow composition data, the actual operating parameters used, and the processed process feature data. This sample is used to perform a small-step gradient descent update on the neural network model for crushing revenue and emulsification cost. This continuous fine-tuning allows the model to adapt to gradual changes in raw materials and equipment wear, achieving adaptive control.

[0084] Furthermore, the residuals between the predicted values ​​of the two core functions and the actual values ​​collected by the outlet sensor are continuously compared. When the processing residuals of a certain material or a certain type of material continuously and systematically deviate from the normal range over a period of time (for example, the model always overestimates or underestimates the viscosity under a certain condition), the system will trigger a deep retraining process of the model.

[0085] Data from the period before and after the triggering event is cached as a high-value set of challenging samples. Using these challenging sample sets, specific neurons are fine-tuned or added to the original neural network through transfer learning, specifically for handling such challenging working conditions. This enables the model to autonomously identify knowledge gaps beyond the initial model's understanding, thereby greatly improving the model's adaptability and robustness to rare working conditions and changes in the nature of raw materials.

[0086] This invention constitutes a complete "decision-execution-feedback-learning" closed-loop control system; the online calibration mechanism enables the model to continuously learn and automatically adapt to model drift caused by seasonal changes in raw materials, batch differences, and equipment wear, ensuring the long-term accuracy and robustness of optimization decisions and realizing adaptive control of the entire processing process.

[0087] Example 2:

[0088] This invention proposes a collaborative optimization method for the separation of grease from kitchen waste based on dynamic game theory, including:

[0089] The initial state data of the resource stream to be processed is obtained through sensing devices.

[0090] Based on initial state data, the asset risk assessment model is used to identify and quantify the factors that pose operational risks to the processing equipment in the resource stream to be processed.

[0091] Based on the identification results of influencing factors, the resource flow to be processed is sorted to obtain an optimized resource flow;

[0092] Before the optimized resource stream enters the processing equipment, the component data of the optimized resource stream is acquired; after the optimized resource stream is processed by the processing equipment, the process characteristic data of the optimized resource stream is collected.

[0093] A dynamic game optimization decision model is pre-established, using the operating parameters of the processing equipment as decision variables and incorporating a mathematical function that balances crushing benefits and emulsification costs. During real-time operation, the dynamic game optimization decision model is used to solve the acquired component data to determine the optimal operating data. The collected process characteristic data is then used to provide feedback calibration to the dynamic game optimization decision model, enabling adaptive control of the processing equipment.

[0094] By utilizing deployed image sensors, image data of the resource stream to be processed is continuously captured, forming an image data stream;

[0095] By using a dynamic weighing sensor installed below the resource stream to be processed, the mass data of the resource stream to be processed is acquired in real time, and the density distribution characteristics of the resource stream to be processed are calculated by combining the volume measured by the laser sensor.

[0096] The image data stream, quality data, and density distribution features are timestamped and fused to construct the initial state data of the resource stream to be processed.

[0097] In this embodiment, the specific data acquisition includes:

[0098] The food waste to be processed is conveyed at a speed of 0.5 meters per second via a horizontal conveyor belt with a width of 1.2 meters. The data acquisition system is deployed in the middle section of the conveyor belt, and its specific structure is as follows:

[0099] Image data acquisition: An industrial camera is vertically mounted 1.5 meters directly above the conveyor belt, with a resolution of 1920x1080 pixels and a frame rate of 30fps, to continuously capture overhead images of the waste stream.

[0100] 3D contour and density data acquisition: A line laser contour sensor is installed 20 cm downstream of the camera. Its scanning line is perpendicular to the direction of conveyor belt movement, and the scanning frequency is 200 Hz. This sensor is used to obtain the precise height and volume distribution of the material. Directly below the laser sensor, at the bottom of the conveyor belt, a dynamic weighing sensor array is integrated. The weighing area is 0.5 meters long and has an accuracy of ±10 grams. This array is used to measure the mass of the corresponding volume of material in real time.

[0101] Data Synchronization and Fusion: The entire acquisition system is time-synchronized to ensure that the timestamp error of image, contour, and quality data is within 1 millisecond. The system fuses image, volume, and quality data within a time window at a cycle of 0.1 seconds to construct initial state data with precise spatiotemporal labels, accurately reflecting the accumulation morphology, color texture, and density distribution characteristics of food waste.

[0102] By integrating multiple sensors and achieving precise timing synchronization, a comprehensive and accurate quantification of the physical state of heterogeneous and dynamically changing food waste streams was achieved, providing high-fidelity raw data for subsequent risk assessment and component analysis.

[0103] The asset risk assessment model includes image data analysis, anomaly area identification, and influencing factor identification;

[0104] The image data analysis is based on a convolutional neural network target detection model to identify the image data stream in the initial state data, analyze the stacking structure, color distribution and texture information of the resource stream to be processed, and obtain image feature data.

[0105] The abnormal region judgment layer identifies unprocessed resource stream regions with target detection anomalies and density anomalies based on image feature data and corresponding density distribution features, and these regions are designated as abnormal regions.

[0106] The influencing factor identification identifies the influencing factors in abnormal areas. The identification results are obtained based on the size, shape, and density of the influencing factors. Combined with the operational interference caused by similar objects in historical data to the processing equipment, an operational risk index is output.

[0107] A near-infrared spectrometer is deployed at the feed inlet of the processing equipment to scan the optimized resource stream entering the processing equipment and obtain the composition data of the optimized resource stream; the composition data includes moisture content, oil content, protein content and cellulose content.

[0108] Online viscometers and laser particle size analyzers are deployed at the outlet of the processing equipment to continuously monitor and optimize the physical properties of the resource flow after processing, including viscosity data obtained from online viscometer monitoring and particle size data obtained from laser particle size analyzer monitoring.

[0109] Before entering the crushing equipment, the optimized resource stream, after sorting, passes through a near-infrared spectrometer deployed above the feed inlet. This analyzer scans the material in real time, analyzing the spectral absorption characteristics in the 900-1700nm wavelength band, and combining this with a pre-established chemometric model (partial least squares regression) to calculate the core component data of the optimized resource stream in real time, including moisture content (%), oil content (%), protein content (%), and cellulose content (%), with a data refresh rate of 1Hz.

[0110] The construction process of the dynamic game optimization decision model includes:

[0111] Acquire the component data, process characteristic data, equipment operating parameters, crushing revenue, and emulsification cost of the complete processing of the processing equipment; the operating parameters include equipment rotation speed and equipment spacing;

[0112] Construct a breakage revenue function and an emulsification cost function based on a neural network model;

[0113] The crushing revenue function takes component data and decision variables as input and crushing revenue as output; the emulsification cost function takes component data and decision variables as input and emulsification cost as output.

[0114] The fragmentation payoff function and the emulsification cost function are embedded within the dynamic game optimization decision model; wherein, the difference between the fragmentation payoff function and the emulsification cost function is defined as the net payoff objective function used for the final decision.

[0115] The adaptive control is a command-based closed-loop execution process;

[0116] The optimal operational data calculated by the dynamic game optimization decision model is formatted into control commands and sent to the underlying programmable logic controller of the processing equipment via the industrial Ethernet protocol.

[0117] The programmable logic controller (PLC) precisely adjusts the equipment speed and equipment spacing through the frequency converter based on the received instructions.

[0118] Meanwhile, the effects of equipment processing are continuously monitored through process characteristic data, and feedback is given to the dynamic game optimization decision-making model.

[0119] Dynamic game optimization decision-making models use feedback as input for the next round of decision-making, which is used to modify their internal functions and the policy output at the next time step.

[0120] Although embodiments of the invention have been shown and described, it will be understood by those skilled in the art that various changes, modifications, substitutions and alterations can be made to these embodiments without departing from the principles and spirit of the invention, the scope of which is defined by the appended claims and their equivalents.

Claims

1. A collaborative optimization method for the separation of grease from kitchen waste based on dynamic game theory, characterized in that, include: The initial state data of the resource stream to be processed is obtained through sensing devices. Based on initial state data, an asset risk assessment model is used to identify and quantify the factors posing operational risks to the processing equipment in the resource stream to be processed. The asset risk assessment model includes image data analysis, anomaly region identification, and influencing factor identification. Image data analysis uses a convolutional neural network-based target detection model to identify the image data stream in the initial state data, analyzing the stacking structure, color distribution, and texture information of the resource stream to be processed to obtain image feature data. Anomaly region identification uses the image feature data and corresponding density distribution features to identify regions in the resource stream where target detection and density anomalies exist, designating them as anomaly regions. Influencing factor identification identifies the influencing factors in the anomaly regions, obtaining identification results based on the size, shape, and density of the influencing factors. Combined with historical data showing similar objects causing operational interference to the processing equipment, an operational risk index is output, and the resource stream to be processed is sorted according to this index. Based on the identification results of the influencing factors, the resource stream to be processed is sorted to obtain an optimized resource stream. Before the optimized resource stream enters the processing equipment, its component data is acquired. After the optimized resource stream is processed by the processing equipment, its process feature data is collected. The process characteristic data includes viscosity data obtained from online viscometer monitoring and particle size data obtained from laser particle size analyzer monitoring; a dynamic game optimization decision model is pre-established, using the operating parameters of the processing equipment as decision variables and incorporating a mathematical function that balances crushing benefits and emulsification costs; during real-time operation, the dynamic game optimization decision model is used to solve for the acquired component data to determine the optimal operating data; and the collected process characteristic data is used to provide feedback calibration to the dynamic game optimization decision model, enabling adaptive control of the processing equipment; The construction process of the dynamic game optimization decision model includes: acquiring component data, process characteristic data, equipment operating parameters, and crushing revenue and emulsification cost of the complete processing of the processing equipment; the operating parameters include equipment rotation speed and equipment spacing; constructing a crushing revenue function and an emulsification cost function based on a neural network model; the crushing revenue function takes component data and decision variables as input and crushing revenue as output; the emulsification cost function takes component data and decision variables as input and emulsification cost as output; embedding the crushing revenue function and emulsification cost function within the dynamic game optimization decision model; wherein, the difference between the crushing revenue function and the emulsification cost function is defined as the net revenue objective function used for the final decision; wherein, the crushing revenue is the economic benefit of the separated oil in the complete oil separation processing process; the emulsification cost is the cost incurred in handling the emulsification phenomenon during the oil separation process.

2. The collaborative optimization method for the separation of grease from kitchen waste based on dynamic game theory according to claim 1, characterized in that: By utilizing deployed image sensors, image data of the resource stream to be processed is continuously captured, forming an image data stream; By using a dynamic weighing sensor installed below the resource stream to be processed, the mass data of the resource stream to be processed is acquired in real time, and the density distribution characteristics of the resource stream to be processed are calculated by combining the volume measured by the laser sensor. The image data stream, quality data, and density distribution features are timestamped and fused to construct the initial state data of the resource stream to be processed.

3. The collaborative optimization method for the separation of grease from kitchen waste based on dynamic game theory according to claim 1, characterized in that: A near-infrared spectrometer is deployed at the feed inlet of the processing equipment to scan the optimized resource stream entering the processing equipment and obtain the composition data of the optimized resource stream. The composition data includes water content, oil content, protein content and cellulose content. The process characteristic data of the optimized resource stream after processing are continuously monitored by an online viscometer and a laser particle size analyzer.

4. The collaborative optimization method for the separation of grease from kitchen waste based on dynamic game theory according to claim 1, characterized in that: The adaptive control is a command-driven closed-loop execution process. The optimal operating data calculated by the dynamic game optimization decision model is formatted into control commands and sent to the underlying programmable logic controller (PLC) of the processing equipment via the industrial Ethernet protocol. Based on the received commands, the PLC precisely adjusts the equipment speed and equipment spacing through the frequency converter. Simultaneously, it continuously monitors the state of the optimized resource flow after equipment processing through process characteristic data and feeds it back to the dynamic game optimization decision model. The dynamic game optimization decision model uses the feedback as the input for the next round of decision-making, correcting its internal functions and the strategy output for the next time step.

Images