Campus nutrition catering scene-oriented ai agent system and method

Abstract

The invention belongs to the technical field of smart catering, and discloses a campus nutrition catering scene-oriented ai agent system and method, and the method comprises the steps: obtaining canteen consumption data and kitchen residue monitoring data of a student group, and carrying out the correlation analysis to generate a group taste preference weight; in combination with the nutrition intake standard parameters and the diet taboo constraint set, constructing diet demand feature vectors; performing multi-target optimization in the standardized dish knowledge graph by using the vector to generate a target nutrition menu set; analyzing the preparation process of the dishes to construct a directed acyclic graph model, and generating a cooking sequence scheduling scheme based on kitchen resource constraints; in a scheme execution process, performing dynamic rearrangement optimization according to real-time time deviation data; the problem of data islands of all links of a traditional campus canteen is solved, catering scientificity is improved, and food waste is reduced.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of intelligent catering, more specifically, the present application relates to an ai agent system and method for campus nutrition catering scenario. BACKGROUND

[0002] As the main place for students to dine daily, the catering service quality of campus canteen is directly related to the physical health of the student group and the operation efficiency of the campus management. An ideal campus catering system needs to meet multiple goals such as nutritional balance, taste adaptation, efficient production, and cost control. However, the current campus nutrition catering management still has significant technical bottlenecks in the aspects of intelligent decision-making and refined execution, which are specifically manifested in the following two aspects: In terms of demand perception and menu recommendation, the existing technology has a serious data island phenomenon, and the optimization goal is single, making it difficult to achieve dynamic balance of nutrition and taste. Traditional solutions can be roughly divided into two categories: one is a static nutrition catering system based on student physiological data, which calculates the standard nutritional requirements according to age, gender, and other basic information, and selects dishes from a fixed menu library based on this. This system focuses on nutritional balance, but completely ignores the dynamic taste preferences of the student group, which can lead to recommended dishes that meet nutritional standards but are not popular, resulting in food waste. The other is a simple recommendation model based on historical sales, which only sorts and recommends dishes based on sales, without strict nutritional constraints. Long-term catering based on this may lead to an unbalanced student diet. In addition, existing kitchen waste data is often not effectively utilized, and cannot be analyzed in conjunction with consumption records, making it difficult to accurately identify real negative feedback from students who have ordered but not finished eating, leading to distorted taste preference models and an inability to accurately reflect the group's dining satisfaction.

[0003] Secondly, in the kitchen production scheduling aspect, existing methods have two limitations: one is the disconnection with the intelligent recommendation system in the front end, making it difficult to efficiently execute daily changing personalized menus; the second is the lack of ability to deal with uncertainties such as time delays in the production process. Currently, the work plan of the kitchen in the canteen relies on experienced head chefs for manual scheduling, and the decision-making process is difficult to quantify, making it difficult to efficiently handle the complex process arrangement brought by the daily changing personalized menus generated by the intelligent recommendation system. Although tools such as the critical path method (CPM) and Gantt chart in the project management field can be used for task sorting, they usually assume that resources are unlimited and task times are fixed, making it difficult to directly apply to the kitchen environment where resources such as equipment and manpower are limited. Moreover, they cannot quickly re-optimize the scheduling of all subsequent tasks when a delay occurs in a certain link (such as the cooking of a dish taking too long), thus affecting the overall meal delivery efficiency and dish quality.

[0004] In view of the above, the present application proposes an ai agent system and method for the campus nutrition catering scenario to solve the above problems. SUMMARY

[0005] In order to overcome the above-mentioned defects of the prior art, in order to achieve the above-mentioned purpose, the present application provides the following technical scheme: An ai agent method for a campus nutrition catering scene, comprising: Obtain dining hall consumption data and kitchen residue monitoring data of a student group, and perform correlation analysis on the obtained data to generate a group taste preference weight; Obtain a nutrition intake standard parameter and a dietary taboo constraint set of the student group, vectorize and splice the group taste preference weight, the nutrition intake standard parameter and the dietary taboo constraint set to construct a dietary demand feature vector; Obtain a standardized dish knowledge graph, perform multi-objective optimization in the search space of the standardized dish knowledge graph based on the dietary demand feature vector, and generate a target nutrition menu set containing a plurality of recommended dishes and a material ratio of each recommended dish; Analyze the manufacturing procedure of each recommended dish in the target nutrition menu set, generate an atomic operation step sequence, and construct a directed acyclic graph model; Obtain a back-of-house resource constraint condition, perform topological sorting and time parameter calculation on the directed acyclic graph model based on the critical path algorithm and the back-of-house resource constraint condition, and generate a cooking sequence scheduling scheme; Execute the cooking sequence scheduling scheme, and obtain time deviation data in the execution process; when the time deviation data exceeds a preset time threshold, rearrange and optimize the atomic operation step sequence that has not been executed based on the directed acyclic graph model.

[0006] Further, the way of performing correlation analysis on the dining hall consumption data and the kitchen residue monitoring data to generate the group taste preference weight comprises: Perform time series analysis on the historical sales data of each dish in the dining hall consumption data to obtain sales trend values of each dish; Analyze the kitchen residue monitoring data using image recognition technology, identify and correlate to each dish, and calculate the average residual rate of each dish; According to the sales trend values and the average residual rate, obtain dish preference scores of each dish; Obtain a dish flavor tag library, determine the flavor tags to which each dish belongs based on the dish flavor tag library, and perform mean value calculation on the dish preference scores of all dishes belonging to the same flavor tag to obtain an initial weight of the flavor tag; Normalize the initial weights of all flavor tags to generate the group taste preference weight.

[0007] Further, the way of constructing the dietary demand feature vector comprises: Obtaining basic physiological profile data of the student group, calculating basic metabolic requirements of the student group based on the basic physiological profile data through a growth and development model, generating a nutritional intake standard parameter including heat, protein and vitamin indicators, and encoding the parameter into a continuous numerical sub-vector; Obtaining allergen registration information of the student group, establishing a dietary taboo constraint set including a food blacklist, and encoding the set into a discrete binary sub-vector; Converting the group taste preference weight into a distribution type weight sub-vector; Performing dimension alignment and normalization splicing processing on the continuous numerical sub-vector, the distribution type weight sub-vector and the discrete binary sub-vector to generate a dietary demand feature vector.

[0008] Further, the way of obtaining the standardized dish knowledge graph comprises: Obtaining recipe text data, extracting dish name, food material, nutritional component, cost price and making process information from the recipe text data; Constructing an initial knowledge graph with dish name and food material as nodes and inclusion relationship as edges, and adding corresponding nutritional component attribute to the food material node, and adding corresponding cost price attribute and making process attribute to the dish name node; Obtaining seasonal information and food safety specifications, traversing the dish nodes in the initial knowledge graph for judgment, marking the dish nodes that do not match the seasonal information or violate the food safety specifications as unavailable state, and obtaining the standardized dish knowledge graph.

[0009] Further, the way of multi-objective optimization in the search space of the standardized dish knowledge graph comprises: Selecting dish nodes marked as available state from the standardized dish knowledge graph to form a candidate dish set; Based on the dietary demand feature vector, extracting several dish combinations from the candidate dish set to form an initial population; Calculating the fitness score of each dish combination in the initial population, and selecting superior combination individuals according to the score results; Performing crossover and mutation operations on the superior combination individuals to generate a new generation population until a preset convergence condition is reached, and outputting a recommended dish combination; Based on the dish-food material association relationship in the standardized dish knowledge graph, performing reverse deduction on the recommended dish combination to generate a target nutritional menu set including food material ratio.

[0010] Further, the way of calculating the fitness score of each dish combination in the initial population comprises: extracting cumulative values of each nutritional component of the dish combination, and calculating Euclidean distances between the cumulative values of each nutritional component and nutritional intake standard parameters in the dietary demand feature vector to generate a nutritional balance score; extracting flavor tag distribution of the dish combination, calculating cosine similarity between the flavor tag distribution and group taste preference weights in the dietary demand feature vector to generate a taste matching score; calculating the total cost of the dish combination based on the cost price attribute of each dish node in the knowledge graph; obtaining a preset total budget cost, calculating the difference between the total cost and the total budget cost to generate a cost control score; weighting and summing the nutritional balance score, the taste matching score and the cost control score to obtain the fitness score.

[0011] Further, the way of constructing the directed acyclic graph model comprises: Based on the manufacturing process attribute of each dish node in the knowledge graph, each recommended dish in the target nutritional recipe set is parsed into its corresponding cleaning, cutting and cooking steps to form an atomic operation step sequence; determining the standard operation time and required equipment type of each step in the atomic operation step sequence; regarding each step in the atomic operation step sequence as a graph node, regarding the process dependency relationship between the steps as a directed edge, regarding the standard operation time as the time attribute of the graph node, constructing a directed acyclic graph model, and marking the required equipment type as the resource demand attribute of the graph node.

[0012] Further, the way of generating the cooking sequence scheduling scheme comprises: topologically sorting the graph nodes in the directed acyclic graph model to obtain a topological sequence; based on the topological sequence and the time attribute of the graph node, calculating the earliest start time and the latest end time of each graph node in the directed acyclic graph model, and taking the earliest start time and the latest end time as the time parameters of the graph node; identifying the critical path in the directed acyclic graph model, marking the graph nodes located on the critical path as high-priority tasks, and calculating the total time difference of the graph nodes located on the non-critical path; based on the time parameters of each graph node, generating an initial scheduling arrangement, and checking whether the total amount of resources required by the graph nodes concurrently executed under the initial scheduling arrangement exceeds the back-of-house resource constraint condition along the time axis; If it exceeds, the scheduling time of the high-priority task is kept unchanged, and the non-critical path graph nodes causing resource conflict are time-shifted based on the total time difference to generate a task execution schedule without resource conflict; Converting the task execution schedule into a cooking sequence scheduling scheme containing dish preparation time window and equipment occupation planning.

[0013] Further, the way of rearranging and optimizing the sequence of atomic operation steps that have not been executed based on the directed acyclic graph model includes: receiving actual completion times of each atomic operation step uploaded by the back-of-house equipment; calculating the difference between the actual completion time and the completion time planned for each atomic operation step in the cooking sequence scheduling scheme to obtain time deviation data; if the time deviation data exceeds a preset time threshold, performing dynamic rearrangement processing on the atomic operation steps that have not been executed, which includes: taking the actual completion time as the time reference, recalculating the time parameters of all unexecuted graph nodes in the directed acyclic graph model; under the back-of-house resource constraint condition, checking whether the total amount of resources required for concurrent execution of unexecuted graph nodes based on the recalculated time parameters exceeds the back-of-house resource constraint condition along the time axis; if it exceeds, performing time translation adjustment on the unexecuted graph nodes that cause resource conflicts, and generating an updated cooking sequence scheduling scheme according to the adjusted time arrangement.

[0014] An AI agent system for a campus nutrition catering scenario is used to implement the AI agent method for the campus nutrition catering scenario, and includes: a preference analysis module configured to obtain dining hall consumption data and kitchen surplus monitoring data of a student group, and perform correlation analysis on the obtained data to generate group taste preference weights; a feature construction module configured to obtain nutrition intake standard parameters and dietary taboo constraint sets of the student group, and vectorize and splice the group taste preference weights, the nutrition intake standard parameters, and the dietary taboo constraint sets to construct a dietary demand feature vector; a menu generation module configured to obtain a standardized dish knowledge graph, and perform multi-objective optimization in a search space of the standardized dish knowledge graph based on the dietary demand feature vector to generate a target nutrition menu set including a plurality of recommended dishes and ingredient ratios of each recommended dish; a graph model construction module configured to analyze cooking procedures of each recommended dish in the target nutrition menu set, generate a sequence of atomic operation steps, and construct a directed acyclic graph model; a cooking scheduling module configured to obtain back-of-house resource constraint conditions, perform topological sorting and time parameter calculation on the directed acyclic graph model based on a critical path algorithm and the back-of-house resource constraint conditions, and generate a cooking sequence scheduling scheme; A dynamic optimization module is configured to issue a cooking sequence scheduling scheme for execution and acquire time deviation data in the execution process, and when the time deviation data exceeds a preset time threshold, rearrange and optimize a sequence of atomic operation steps that have not been executed based on a directed acyclic graph model.

[0015] Technical effects and advantages of the present application: 1. Kitchen residue monitoring data is introduced as a negative feedback mechanism, which is analyzed in association with canteen consumption data (positive sales), enabling accurate identification of high-waste dishes that are not popular, and correcting the bias of traditional recommendations that rely solely on sales. Combined with the group taste preference weights generated by the flavor tag library, the AI agent can not only meet the static nutritional intake standard, but also dynamically adapt to the real taste needs of the student group, reducing campus food waste caused by taste mismatch from the source.

[0016] 2. A standardized menu knowledge graph is used to build a compliance search space that includes seasonal and safety specifications, and a multi-objective evolutionary algorithm is used to optimize in a large menu combination. This mechanism can simultaneously handle the three mutually restrictive goals of nutritional balance, taste matching, and cost control, quickly generating a menu that meets strict nutritional standards, matches student tastes, and is cost-effective, significantly improving the scientificity of meal preparation.

[0017] 3. By parsing the recommended menu into atomic operation steps and building a directed acyclic graph (DAG) model, the traditional experience-based kitchen management is transformed into a computable digital process. Combined with the critical path algorithm and resource smoothing technology, the system can automatically identify critical bottleneck processes and stagger non-critical tasks under the constraint of limited kitchen resources, maximizing parallel job efficiency and effectively solving the meal delay problem caused by equipment conflicts during peak periods. Through the real-time feedback mechanism of "monitoring-judgment-rearrangement", when unforeseen time deviations occur during kitchen execution and exceed the preset threshold, the system can immediately globally rearrange and optimize the task sequence that has not been executed based on the DAG model. This dynamic correction capability ensures that the system can still provide the best emergency scheduling scheme under the current conditions when facing sudden disturbances, greatly ensuring the reliability of on-time meal service in campus canteens.

[0018] 4. By using a unified dietary demand feature vector and a standardized graph model, consumption data, kitchen residue monitoring data, and kitchen resource data are effectively combined, achieving seamless integration of demand-menu-process-execution-feedback-reoptimization, solving the problem of data silos in traditional campus canteens, and promoting the intelligentization and standardization upgrade of campus catering management. BRIEF DESCRIPTION OF DRAWINGS

[0019] Figure 1This is a schematic diagram of an AI agent method for campus nutrition meal planning according to the present invention; Figure 2 This is a schematic diagram of an AI intelligent agent system for campus nutrition meal planning according to the present invention. Detailed Implementation

[0020] 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.

[0021] Example 1 Please see Figure 1 As shown in this embodiment, an AI agent method for campus nutrition meal planning includes: Acquire student cafeteria consumption data and food waste monitoring data, and perform correlation analysis on the acquired data to generate group taste preference weights; Obtain the standard parameters of nutritional intake and the set of dietary restrictions for the student group, and then concatenate the group's taste preference weights, standard parameters of nutritional intake, and set of dietary restrictions into vectors to construct a dietary demand feature vector. Obtain a standardized recipe knowledge graph, and based on dietary demand feature vectors, perform multi-objective optimization within the search space of the standardized recipe knowledge graph to generate a target nutritional recipe set containing several recommended dishes and the ingredient ratios of each recommended dish. The preparation process of each recommended dish in the target nutritional recipe set is analyzed, the atomic operation step sequence is generated, and a directed acyclic graph model is constructed. Obtain the kitchen resource constraints, and perform topological sorting and time parameter calculation on the directed acyclic graph model based on the critical path algorithm and kitchen resource constraints to generate a cooking sequence scheduling scheme; The cooking sequence scheduling scheme is issued and executed, and the time deviation data during the execution process is obtained. When the time deviation data exceeds the preset time threshold, the sequence of atomic operation steps that have not yet been executed is rearranged and optimized based on the directed acyclic graph model.

[0022] Furthermore, correlation analysis was conducted between canteen consumption data and food waste monitoring data to generate group taste preference weights, including the following methods: By obtaining student consumption data in the cafeteria through API interfaces, historical sales data of each dish over the past one to two months are extracted. To eliminate accidental factors (such as sudden sales due to events on a particular day) and periodic fluctuations (such as natural sales declines on weekends), time series analysis algorithms such as Autoregressive Integrated Moving Average (ARIMA) models or exponential smoothing are used to process the data and calculate the sales trend value of each dish that reflects the current popularity trend. The higher the sales trend value, the stronger the students' willingness to purchase.

[0023] Smart cameras installed at the dish return area collect data on food waste. Using image recognition technology, semantic segmentation, or object detection algorithms, the system automatically identifies the types and volumes of leftover food on each dish and associates them with the specific dishes served. The average leftover rate for each dish is then calculated. This indicator directly reflects the extent to which students have consumed the food; a higher average leftover rate indicates lower actual taste satisfaction. The average leftover rate is calculated by adding the amount of food left on all dishes for a particular dish and dividing the sum by the total amount of food served for that dish.

[0024] Based on sales trend values ​​and average remaining rates, a dish preference score is calculated for each dish. This score reflects students' level of liking for a specific dish. The formula for calculating the dish preference score is as follows: In the formula, Indicates the first The dish preference score; Indicates the first The sales trend value of this dish; This indicates that the sales trend value is logarithmically smoothed to reduce the huge numerical difference between top-selling dishes and other dishes, and to prevent sales data from dominating the rating. Indicates the first The average leftover rate of each dish; The preset waste sensitivity coefficient has a value of 2 to 5.

[0025] To enable the agent to reason about food flavor preferences, it is also necessary to obtain a food flavor tag library, which includes tags such as spicy, sweet and sour, light, fried, and braised. Then, based on the food attributes (such as cuisine, cooking method, and seasoning recipe), the flavor tag to which each food belongs is determined; for example, Kung Pao Chicken and Mapo Tofu are classified as spicy.

[0026] The average of the individual dish preference scores of all dishes belonging to the same flavor label is calculated to eliminate noise caused by cooking errors in individual dishes, extract the common evaluation of a certain type of flavor by the group, and thus obtain the initial weight of the flavor label. To facilitate the subsequent construction and calculation of feature vectors, the initial weights of all flavor tags must be normalized (i.e., the sum of all weights is 1 or mapped to the 0-1 range), and finally the group taste preference weights are generated, thus providing a quantitative basis for flavor dimensions for subsequent dish recommendations or menu optimization.

[0027] By jointly analyzing canteen consumption data and image recognition-based food waste monitoring data, the problem of misjudging high sales volume as high satisfaction was solved by relying solely on sales data. By introducing dual evaluation indicators of sales trend value and average waste rate, high-waste dishes that are popular but not well received were eliminated, restoring the true dietary satisfaction of the student population. By using a dish flavor tag library to cluster and abstract discrete individual dish ratings into generalized group taste preference weights, the AI ​​agent has the ability to generalize flavor inference beyond specific dishes. This can not only reduce food waste caused by unpleasant tastes at the source, but also guide the accurate development and recommendation of new dishes not yet available based on abstract flavor weights, significantly improving the scientific and personalized level of meal planning.

[0028] Furthermore, methods for constructing dietary requirement feature vectors include: Basic physiological profile data of the student population is obtained from the campus health management database, including average age, gender ratio, average height, weight, and daily exercise volume. A pre-defined growth and development model (such as the Chinese Dietary Reference Intakes (DRIs) model or the Schofield formula) is used to calculate the daily basal metabolic requirements of the student population, generating standard parameters for nutrient intake including calories, protein, and vitamin indicators (such as vitamins A, C, calcium, and iron). To eliminate the influence of different nutrient units (e.g., calories are in kcal, with values ​​up to 2000; while trace elements are in mg, with values ​​only 10), these parameters also need to be standardized (e.g., Z-Score standardization), and the processed values ​​are arranged in a predetermined order and encoded as continuous numerical subvectors. Obtain allergen registration information (such as allergy records for peanuts, seafood, milk, etc.) from the student population, maintain a standard food ingredient database in the background, and establish a set of dietary restrictions containing a blacklist of ingredients based on the aforementioned allergen registration information. Use one-hot or many-hot coding techniques to map this set into a fixed-length vector space. For example, if the standard food ingredient database has N types of easily allergenic ingredients, construct a vector of length N. If peanuts are on the blacklist, the corresponding dimension has a value of 1; otherwise, it has a value of 0. This encodes dietary restrictions as discrete binary subvectors.

[0029] Read the generated group taste preference weights, since these weights are essentially a probability distribution (e.g., spicy 0.4, mild 0.3, sweet and sour 0.3), and thus the group taste preference weights can be directly converted into distributed weight subvectors; The data structures of continuous numerical subvectors, distributed weight subvectors, and discrete binary subvectors are examined and uniformly adjusted into low-dimensional dense vectors with consistent orientation (e.g., uniformly one-dimensional row or column vectors). To eliminate the influence of dimensions, normalization concatenation is also required. The max-min normalization algorithm is used to map each element in the continuous numerical subvector to the [0,1] interval, so that its numerical range is consistent with that of the distributed weight subvector and the discrete binary subvector. Finally, the normalized subvectors are concatenated on the feature channels in a fixed order of "nutrition-taboo-preference" to generate a dietary demand feature vector.

[0030] Furthermore, methods for obtaining standardized food knowledge graphs include: Obtain recipe text data from internet recipe databases or historical menus within school cafeterias; use natural language processing technologies (such as Named Entity Recognition, NER) to extract dish names, ingredients, nutritional components, cost prices, and preparation process information from the recipe text data.

[0031] An initial knowledge graph is constructed using a graph database (such as Neo4j), with dish names and ingredients as nodes and inclusion relationships as edges. Specifically, two types of nodes are created: dish nodes and ingredient nodes. Directed edges indicating inclusion relationships are established between them (e.g., "braised pork" -> "pork belly", indicating that braised pork contains pork belly). Then, corresponding nutritional component attributes are added to the ingredient nodes (e.g., xx grams of fat per 100g of pork belly), and corresponding cost price and cooking process attributes are added to the dish name nodes (e.g., stewing, cooking time 45 minutes). This provides a data foundation for subsequent cost control and kitchen scheduling.

[0032] The system obtains seasonal information and food safety regulations through an API interface. The seasonal information includes a list of in-season ingredients (such as bamboo shoots in spring and watermelons in summer) and a list of out-of-season ingredients. The food safety regulations include warnings issued by the education bureau about prohibited ingredients for schools (such as raw seafood and sprouted potatoes) and high-risk ingredients that may cause food poisoning. Then, the system iterates through the dish nodes in the initial knowledge graph to make judgments. If a dish contains ingredients that violate safety regulations or contains expensive out-of-season ingredients, the dish node is marked as unavailable. After the previous state update process, a standardized dish knowledge graph is finally obtained.

[0033] The above process uses natural language processing technology to transform unstructured recipe text into a structured graph network, clarifying the hierarchical relationship between dishes and ingredients, and achieving accurate cumulative calculation of nutritional components. Furthermore, the introduction of seasonal information and food safety regulations as a dynamic filtering layer of the graph can automatically filter out non-compliant, unsafe, or inappropriate dish nodes, ensuring that the subsequent intelligent meal matching algorithm operates within a compliant and safe search space, and significantly improving the feasibility and security of the recommendation results.

[0034] Campus meal planning is essentially a complex multi-objective combinatorial optimization problem. It requires selecting a combination (e.g., four dishes and one soup per day) from a vast menu, ensuring that the combination simultaneously meets nutritional standards, appeals to group preferences, stays within budget, and does not contain prohibited ingredients. Traditional brute-force methods would result in exponentially increased computational costs, failing to produce results quickly. Simple greedy algorithms (selecting only the best option each time) are prone to getting stuck in local optima (e.g., selecting four delicious but nutritionally deficient meat dishes). Therefore, a heuristic-based, efficient optimization mechanism is needed to avoid lengthy search times. This embodiment avoids this problem by performing multi-objective optimization within the search space of a standardized menu knowledge graph, thereby accelerating the search process. The specific steps are as follows: From the standardized food knowledge graph, food nodes marked as available are selected to form a set of candidate foods. This step uses the graph's filtering function to directly exclude non-compliant and unsafe foods, thus narrowing the search scope. Based on dietary demand feature vectors, several dishes are randomly selected from the candidate dish set to form an initial population. It's important to note that random selection means, during the random generation process, eliminating combinations containing severely prohibited ingredients based on the feature vectors, or slightly increasing the probability of selecting dishes that align with the group's taste preferences. This guided randomness not only ensures the diversity of the initial population but also makes the overall quality superior to completely blind random combinations. For example, 50 recipe schemes are randomly generated, each scheme representing an individual containing several dishes. If the dietary demand feature vectors indicate that the student group dislikes seafood and prefers spicy food, then during the initial population generation, the probability of selecting seafood dishes (such as steamed sea bass) will be reduced, while the probability of selecting spicy dishes (such as Mapo Tofu) will be increased. This generates an initial recipe combination that is both safe and initially matches the taste preferences, avoiding wasted search space on invalid solutions (such as recipes containing allergens).

[0035] Each individual in the population is evaluated, that is, the fitness score of each dish combination in the initial population is calculated, and the dominant combinations are selected based on the score results. The fitness score is a comprehensive indicator that measures the performance of the dish combination in terms of nutritional balance, taste matching, and cost control. For example, the top 20% of individuals with the highest scores are retained as dominant combinations, while individuals with low scores are eliminated.

[0036] To generate better menu recommendations, crossover and mutation operations need to be performed on the individuals with superior combinations. The crossover operation involves swapping some dishes from two high-quality recipes to generate a new recipe (e.g., lunch from recipe A + dinner from recipe B). The mutation operation involves randomly replacing a dish from a recipe with another dish from the candidate set (e.g., replacing "shredded potatoes" with "stir-fried lotus root slices") to increase population diversity and prevent the algorithm from converging to a local optimum too early.

[0037] After generating a new generation of the population, the evaluation, selection, and evolution steps need to be repeated until the preset convergence condition is met. The dish combination with the highest fitness score is then selected as the recommended dish combination and output. The preset convergence condition is that the number of iterations reaches 100, or the fitness score of the best individual changes by less than 5% over 3 to 5 consecutive generations. After obtaining the recommended dish combination (such as braised pork belly + stir-fried vegetables), the association capability of the graph is used to reverse-engineer the recommended dish combination based on the dish-ingredient association relationship in the standardized dish knowledge graph, calculate the total amount of ingredients required (such as 50kg of pork belly and 80kg of vegetables), and finally generate a target nutritional recipe set containing the ingredient ratios.

[0038] It should be noted that the methods for calculating the fitness scores of each dish combination in the initial population include: Extract the cumulative values ​​of each nutrient component from the dish combination, and calculate the Euclidean distance between the cumulative values ​​of each nutrient component and the standard nutrient intake parameters in the dietary requirement feature vector. The smaller the distance, the closer the nutrition is to the standard, thus generating a nutritional balance score; the formula for calculating the nutritional balance score is: In the formula, This represents the calculated score for nutritional balance. This represents a vector composed of the cumulative values ​​of calculated nutrients. The Euclidean norm representing the difference between the two.

[0039] The flavor tags for each dish in the dish combination are extracted, and a flavor tag distribution is generated statistically. The group taste preference weights in the dietary demand feature vector are obtained, and the cosine similarity between the flavor tag distribution and the group taste preference weights in the dietary demand feature vector is calculated to generate a taste matching score. For example, if the generated flavor tag distribution is [spicy: 0.5, sweet: 0.2, salty: 0.3], and the taste preference weights are [spicy: 0.6, sweet: 0.1, salty: 0.3], then the taste matching score is 0.981. Based on the cost and price attributes of each dish node in the knowledge graph, the total cost of the dish combination is calculated cumulatively. Simultaneously, a preset total budget cost is obtained, and the difference between the total cost and the total budget cost is calculated to obtain a cost control score. The formula for calculating the cost control score is as follows: In the formula, This indicates the cost control score; This represents the smoothing parameter, with a value range of 0.01 to 0.3; This represents the total cost of the calculated dish combination; Indicates the total budget cost; Indicates An exponential function with base 0.

[0040] Based on actual management needs (e.g., whether schools prioritize nutrition or taste), weighting coefficients are set for the three scores. The scores for nutritional balance, taste matching, and cost control are weighted and summed to obtain the adaptability score.

[0041] The above steps construct a secure search space through a standardized menu knowledge graph and solve the multi-objective combination optimization problem in meal planning using evolutionary algorithms. By constructing a three-dimensional fitness evaluation system that includes nutritional Euclidean distance, taste cosine similarity, and cost threshold penalty, qualitative factors such as health, taste, and affordability are transformed into quantifiable mathematical indicators. This enables the system to not only search for the globally optimal recipe that balances strict nutritional standards, group taste preferences, and rigid cost budgets from a massive number of combinations in a very short time, but also to automatically generate a precise ingredient ratio list using the reverse derivation capability of the graph. This achieves a closed loop from intelligent decision-making to refined procurement, significantly improving the scientific nature and operational efficiency of campus meal planning.

[0042] In traditional campus canteen kitchen management, the conversion from menu to execution relies heavily on human experience. However, menus generated by intelligent meal preparation systems often include multiple dishes and involve dozens of processes. These processes have complex logical dependencies and resource competition relationships (e.g., multiple dishes require the use of a limited number of woks). If only manual scheduling is used, logical confusion (e.g., starting cooking before all ingredients are prepared) or resource deadlock (e.g., everyone vying for the cutting and preparation station) can easily occur, leading to low food preparation efficiency. Therefore, this embodiment avoids this problem by modeling the kitchen using a directed acyclic graph (DAG) model. The methods for constructing the DAG model include: Based on the cooking process attributes of each dish node in the knowledge graph, each recommended dish in the target nutritional recipe set is parsed into its corresponding cleaning, cutting, and cooking steps, forming a sequence of atomic operation steps; for example, shredded green peppers and pork is parsed into the steps of cleaning green peppers, cutting green peppers into shreds, cutting pork into shreds, marinating pork into shreds, and stir-frying and mixing.

[0043] For each atomic operation step identified, the standard operating time and required equipment type for each step in the atomic operation step sequence are determined by consulting the Standard Operating Procedure (SOP) database. For example, the standard operating time for slicing green peppers is 5 minutes, the corresponding preparation table is for slicing green peppers, and the stir-frying corresponds to a Chinese-style wok stove.

[0044] By treating each step in the atomic operation sequence as a graph node, the process dependencies between steps as directed edges, and the standard operating time as the time attribute of the graph node, a directed acyclic graph (DAG) model is constructed. The required equipment type is then labeled as the resource requirement attribute of the graph node. This DAG model intuitively displays the flow direction of all tasks in the kitchen and guarantees no circular dependencies (i.e., no deadlocks), laying the mathematical foundation for subsequent critical path calculation and parallel scheduling.

[0045] The above method transforms traditional experience-based kitchen management into a computable digital process by constructing a directed acyclic graph (DAG) model. This DAG model not only clearly defines the logical dependencies between all atomic operation steps, avoiding process errors, but also binds time attributes and resource requirement attributes to graph nodes. This enables the computer to accurately identify the parallel possibilities between tasks (i.e., which tasks can be performed simultaneously), thus providing standardized data structure support for subsequent intelligent scheduling that minimizes time under limited resource constraints.

[0046] In a resource-constrained real-world kitchen (e.g., only 3 woks, but 5 dishes to be cooked), scheduling a specific execution timeline is a typical resource-constrained project scheduling problem. Simply executing in parallel leads to resource conflicts (wok grabbing); executing sequentially results in significant time waste (chef A is cooking while chef B is idle). Therefore, this embodiment proposes a scheduling strategy combining the critical path method and resource smoothing techniques. The specific steps are as follows: First, the graph nodes in the directed acyclic graph model are topologically sorted to transform the complex network structure into a linear sequence of operations (topological sequence). In this sequence, the parent task of any task is placed before its parent task, thus satisfying the process logic constraints.

[0047] Starting from the initial node, calculate the earliest start time and earliest end time for each node based on its time attribute (work hours). Starting from the final node, calculate the latest start time and latest end time for each node by working backwards from the total project duration. Store the earliest start time and latest end time as the time parameters of the node.

[0048] Calculate the total float for each graph node (total float = latest start time - earliest start time); if the total float for a graph node is 0, it means that there is no room for delay in this task, and any delay will directly lead to a postponement of the overall meal time.

[0049] Identify the critical path (i.e. the path consisting of all nodes with a total float of 0) in the directed acyclic graph model and mark the graph nodes on the critical path as high-priority tasks.

[0050] Calculate the total float of graph nodes located on non-critical paths; this time serves as a buffer for subsequent resource balancing.

[0051] An initial scheduling arrangement is generated based on the time parameters of each graph node (usually the earliest start time).

[0052] The system checks the total resources required for concurrent execution of graph nodes under the initial scheduling arrangement along the timeline to see if they exceed the kitchen resource constraints (e.g., the number of various kitchen utensils, the number of stoves, etc.). If they do, the scheduling time of high-priority tasks remains unchanged (prioritizing the critical path and ensuring the completion date). Using the total float of non-critical tasks, the non-critical path graph nodes that cause resource conflicts are time-shifted (execution is delayed) until the resource demand at that moment is less than the resource limit. After multiple iterations, a task execution schedule without resource conflicts is finally generated. The optimized schedule is then formatted and transformed into a cooking sequence scheduling scheme that includes the preparation time windows of each dish and equipment occupancy planning, and then distributed to the kitchen display screen or smart terminal.

[0053] The above solution solves the problem of multi-task concurrent scheduling under limited resources by introducing critical path algorithm and resource smoothing technology. It not only identifies the key bottleneck process that determines the overall meal preparation speed and ensures that it obtains resources first, but also uses the time buffer of non-critical processes to eliminate equipment conflicts through dynamic shifting and staggered execution, thereby maximizing the efficiency of parallel operations in the kitchen (such as parallel cutting and cooking), thus improving the utilization rate of kitchen resources while ensuring on-time meal preparation.

[0054] The kitchen of a school cafeteria is a dynamic environment full of uncertainties. Although the preceding steps generate a theoretically optimal static scheduling scheme, various unexpected disturbances often occur during actual execution: for example, a novice chef might be 10 minutes slower than the standard time for chopping vegetables (task delay), or an oven might suddenly malfunction and become unusable (resource loss). If these deviations are not detected and responded to, and subsequent tasks continue as planned, the resources required for subsequent tasks may be locked before the preceding tasks are completed, or chefs may be waiting idly, ultimately leading to a significant delay in overall food preparation. Therefore, this embodiment proposes a dynamic rescheduling strategy based on time window deviation triggering, with the specific steps as follows: By connecting sensors to smart kitchen appliances or chef operating terminals (such as wall-mounted tablets), the actual completion time of each atomic operation step uploaded by the kitchen equipment is received; the difference between the actual completion time and the completion time planned for each atomic operation step in the cooking sequence scheduling scheme is calculated to obtain time deviation data. To avoid excessively frequent scheduling jitter (such as reordering after only 1 minute of delay), a preset time threshold (e.g., 5 minutes) needs to be set for reordering. If the time deviation exceeds the preset time threshold, it is determined that the current progress has seriously deviated from the plan, and the atomic operation steps that have not yet been executed need to be dynamically reordered. The specific steps are as follows: Using the actual completion time as the time base (i.e., the current moment), all completed graph nodes are marked as completed, and the resources occupied by currently executing nodes are locked. Then, starting from the current moment, the time parameters of all unexecuted graph nodes in the directed acyclic graph model are recalculated. During this process, the earliest start time of subsequent subtasks affected by the preceding delayed tasks will be automatically postponed according to their dependencies. Simple time postponement may cause resource contention within the new time window (for example, the postponed cooking task may overlap with another cooking task originally scheduled for a later time, resulting in insufficient cooking stove resources). Therefore, a secondary check is required: under the constraints of kitchen resources, based on the recalculated time parameters, the total resources required for the concurrent execution of unexecuted graph nodes are checked moment by moment along the time axis to see if they exceed the kitchen resource constraints. If the check finds that the total resources exceed the constraints, resource smoothing is performed again, and the unexecuted graph nodes that cause resource conflicts are time-shifted. In terms of adjustment strategy, priority is still given to ensuring the time scheduling of tasks on the critical path, while tasks on non-critical paths are postponed and shifted using their total float to resolve resource conflicts. Finally, an updated cooking sequence scheduling plan is generated based on the adjusted schedule and pushed to the kitchen terminal in real time to prompt chefs to adjust the order of their work.

[0055] The above steps, by constructing a dynamic closed-loop mechanism of monitoring, judgment, and rescheduling, endow the AI ​​agent with robust capabilities to cope with uncertainty. This mechanism can quickly identify the scope of the deviation when operational delays or unexpected interference occur in the kitchen, and replan the optimal execution path for subsequent steps; this not only avoids systemic paralysis caused by single-point failures, but also ensures that in case of emergencies, the optimal emergency dispatch plan under the current conditions can be provided in a timely manner, greatly guaranteeing on-time meal service.

[0056] Example 2 Please see Figure 2 As shown, parts not described in detail in this embodiment are described in Embodiment 1. This embodiment provides an AI intelligent agent system for campus nutrition meal planning, including: The preference analysis module is used to acquire canteen consumption data and food waste monitoring data of students, and to perform correlation analysis on the acquired data to generate group taste preference weights. The feature construction module is used to obtain the standard parameters of nutritional intake and the set of dietary restrictions for the student group. The group's taste preference weights, standard parameters of nutritional intake and the set of dietary restrictions are vectorized and concatenated to construct a dietary demand feature vector. The recipe generation module is used to obtain a standardized recipe knowledge graph. Based on dietary demand feature vectors, it performs multi-objective optimization within the search space of the standardized recipe knowledge graph to generate a target nutritional recipe set containing several recommended dishes and the ingredient ratios of each recommended dish. The graph model building module is used to parse the preparation process of each recommended dish in the target nutritional recipe set, generate a sequence of atomic operation steps, and build a directed acyclic graph model. The cooking scheduling module is used to obtain the kitchen resource constraints, perform topological sorting and time parameter calculation on the directed acyclic graph model based on the critical path algorithm and the kitchen resource constraints, and generate a cooking sequence scheduling scheme. The dynamic optimization module is used to issue and execute the cooking sequence scheduling plan and obtain the time deviation data during the execution process. When the time deviation data exceeds the preset time threshold, the atomic operation step sequence that has not yet been executed is rearranged and optimized based on the directed acyclic graph model.

[0057] The modules are connected via wired and / or wireless means to enable data transmission between them.

[0058] The above description is merely a preferred embodiment of the present invention and is not intended to limit the present invention. Although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art can still modify the technical solutions described in the foregoing embodiments or make equivalent substitutions for some of the technical features. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the protection scope of the present invention.

[0059] It should be noted that, in this document, the terms "comprising," "including," or any other variations thereof are intended to cover non-exclusive inclusion, such that a process, method, article, or apparatus that comprises a list of elements includes not only those elements but also other elements not expressly listed, or elements inherent to such a process, method, article, or apparatus. Unless otherwise specified, an element defined by the phrase "comprising one..." does not exclude the presence of other identical elements in the process, method, article, or apparatus that includes said element.

[0060] In the description of this invention, it should be understood that the terms "first," "second," etc., are used only for distinguishing descriptions and should not be construed as indicating or implying relative importance.

[0061] In the description of this invention, unless otherwise stated, "a plurality of" means two or more.

[0062] In the description of this invention, "several" means one or more, and "a large number" means two or more.

[0063] In the description of this specification, references to terms such as "one embodiment," "some embodiments," "example," "specific example," or "some examples," etc., indicate that a specific feature, structure, material, or characteristic described in connection with that embodiment or example is included in at least one embodiment or example of the invention. In this specification, the illustrative expressions of the above terms do not necessarily refer to the same embodiment or example. Furthermore, the specific features, structures, materials, or characteristics described may be combined in any suitable manner in one or more embodiments or examples.

[0064] All formulas in this manual are dimensionless and calculated numerically. The formulas are derived from software simulations based on a large amount of collected data to obtain the most recent real-world results. The preset parameters and thresholds in the formulas are set by those skilled in the art according to the actual situation.

[0065] Although embodiments of the invention have been shown and described, those skilled in the art will understand 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 claims and their equivalents.

Claims

1. An AI agent method for campus nutrition meal planning, characterized in that, include: Acquire student cafeteria consumption data and food waste monitoring data, and perform correlation analysis on the acquired data to generate group taste preference weights; Obtain the standard parameters of nutritional intake and the set of dietary restrictions for the student group, and then concatenate the group's taste preference weights, standard parameters of nutritional intake, and set of dietary restrictions into vectors to construct a dietary demand feature vector. Obtain a standardized recipe knowledge graph, and based on dietary demand feature vectors, perform multi-objective optimization within the search space of the standardized recipe knowledge graph to generate a target nutritional recipe set containing several recommended dishes and the ingredient ratios of each recommended dish. The preparation process of each recommended dish in the target nutritional recipe set is analyzed, the atomic operation step sequence is generated, and a directed acyclic graph model is constructed. Obtain the kitchen resource constraints, and perform topological sorting and time parameter calculation on the directed acyclic graph model based on the critical path algorithm and kitchen resource constraints to generate a cooking sequence scheduling scheme; The cooking sequence scheduling scheme is issued and executed, and the time deviation data during the execution process is obtained. When the time deviation data exceeds the preset time threshold, the sequence of atomic operation steps that have not yet been executed is rearranged and optimized based on the directed acyclic graph model.

2. The method according to claim 1, characterized in that, Methods for generating group taste preference weights by correlating canteen consumption data with food waste monitoring data include: A time-series analysis was conducted on the historical sales data of each dish in the canteen consumption data to obtain the sales trend value of each dish; Image recognition technology is used to analyze food waste monitoring data, identify and associate it with each dish, and calculate the average waste rate of each dish; Based on sales trend values ​​and average remaining rates, a dish preference score is obtained for each dish. Obtain the flavor tag library for dishes, determine the flavor tag to which each dish belongs based on the flavor tag library, and calculate the average of the dish preference scores of all dishes belonging to the same flavor tag to obtain the initial weight of the flavor tag. The initial weights of all flavor labels are normalized to generate group taste preference weights.

3. The method according to claim 2, characterized in that, Methods for constructing dietary requirement feature vectors include: Obtain basic physiological profile data of the student population, calculate the basal metabolic needs of the student population based on the basic physiological profile data and the growth and development model, generate nutritional intake standard parameters including calorie, protein and vitamin indicators, and encode the parameters as continuous numerical subvectors. Obtain allergen registration information from the student population, establish a set of dietary restrictions that includes a blacklist of ingredients, and encode this set as a discrete binary subvector; Convert the group taste preference weights into distributed weight sub-vectors; The continuous numerical subvectors, distributed weight subvectors, and discrete binary subvectors are dimensionally aligned and normalized before being concatenated to generate dietary demand feature vectors.

4. The method according to claim 3, characterized in that, Methods for obtaining standardized food knowledge graphs include: Obtain recipe text data and extract information such as dish name, ingredients, nutritional components, cost price, and preparation process from the recipe text data; Construct an initial knowledge graph with dish names and ingredients as nodes and inclusion relationships as edges, and add corresponding nutritional component attributes to ingredient nodes, and add corresponding cost price attributes and cooking process attributes to dish name nodes; Obtain seasonal information and food safety standards, traverse the dish nodes in the initial knowledge graph for judgment, and mark the dish nodes that do not match the seasonal information or violate food safety standards as unusable to obtain a standardized dish knowledge graph.

5. The method according to claim 4, characterized in that, Multi-objective optimization methods within the search space of a standardized recipe knowledge graph include: From the standardized recipe knowledge graph, select recipe nodes that are marked as available to form a set of candidate recipes; Based on dietary demand feature vectors, several dishes are extracted from the candidate dish set and combined into an individual to form an initial population; Calculate the fitness score of each dish combination in the initial population, and select the dominant combination individuals based on the score results; Perform crossover and mutation operations on the dominant combination individuals to generate a new generation of population until the preset convergence condition is met, and output recommended dish combinations; Based on the dish-ingredient relationships in the standardized dish knowledge graph, the recommended dish combinations are derived in reverse to generate a set of target nutritional recipes containing ingredient ratios.

6. The method according to claim 5, characterized in that, Methods for calculating the fitness score of each dish combination in the initial population include: Extract the cumulative values ​​of each nutrient component from the dish combination, and calculate the Euclidean distance between the cumulative values ​​of each nutrient component and the nutrient intake standard parameters in the dietary demand feature vector to generate a nutritional balance score. Extract the flavor tag distribution of the dish combination, calculate the cosine similarity between the flavor tag distribution and the group taste preference weight in the dietary demand feature vector, and generate a taste matching score. Calculate the total cost of the dish combination based on the cost and price attributes of each dish node in the knowledge graph; Obtain the preset total budget cost, calculate the difference between the total cost and the total budget cost, and generate a cost control score; The fitness score is obtained by weighting and summing the scores for nutritional balance, taste matching, and cost control.

7. The method according to claim 6, characterized in that, Methods for constructing directed acyclic graph models include: Based on the cooking process attributes of each dish node in the knowledge graph, each recommended dish in the target nutritional recipe set is parsed into its corresponding cleaning, cutting and cooking steps, forming a sequence of atomic operation steps; Determine the standard operating time and required equipment type for each step in the atomic operation sequence; By treating each step in the atomic operation sequence as a graph node, the process dependencies between steps as directed edges, and the standard operating time as the time attribute of the graph node, a directed acyclic graph model is constructed, and the required equipment type is marked as the resource requirement attribute of the graph node.

8. The method according to claim 7, characterized in that, Methods for generating cooking sequence scheduling schemes include: Perform topological sorting on the graph nodes in the directed acyclic graph model to obtain a topological sequence; Based on the topological sequence and the time attributes of graph nodes, the earliest start time and latest end time of each graph node in the directed acyclic graph model are calculated, and the earliest start time and latest end time are used as the time parameters of the graph node. Identify the critical path in the directed acyclic graph model, mark the graph nodes on the critical path as high-priority tasks, and calculate the total float of the graph nodes on the non-critical paths. Based on the time parameters of each graph node, an initial scheduling arrangement is generated, and the total amount of resources required for concurrent execution of graph nodes under the initial scheduling arrangement is checked along the time axis to see if it exceeds the kitchen resource constraints. If the time limit is exceeded, the scheduling time of high-priority tasks will remain unchanged, and the non-critical path graph nodes that cause resource conflicts will be shifted based on the total float to generate a task execution schedule without resource conflicts. The task execution schedule is transformed into a cooking sequence scheduling scheme that includes the preparation time window for each dish and equipment occupancy planning.

9. The method according to claim 8, characterized in that, Methods for rearranging and optimizing the sequence of atomic operation steps that have not yet been executed based on the directed acyclic graph model include: Receive the actual completion time of each atomic operation step uploaded by the kitchen equipment; The time deviation data is obtained by calculating the difference between the actual completion time and the planned completion time for each atomic operation step in the cooking sequence scheduling scheme. If the time deviation data exceeds a preset time threshold, the atomic operation steps that have not yet been executed will be dynamically rearranged. The dynamic rearrangement process includes: Using the actual completion time as the time benchmark, recalculate the time parameters of all unexecuted graph nodes in the directed acyclic graph model; Under the constraints of kitchen resources, based on the recalculated time parameters, check along the time axis whether the total amount of resources required for the concurrent execution of unexecuted graph nodes exceeds the constraints of kitchen resources. If the time limit is exceeded, the unexecuted graph nodes that caused resource conflicts will be time-shifted and adjusted, and an updated cooking order scheduling scheme will be generated based on the adjusted time schedule.

10. An AI agent system for campus nutrition meal planning scenarios, used to implement the AI ​​agent method for campus nutrition meal planning scenarios as described in any one of claims 1 to 9, characterized in that, include: The preference analysis module is used to acquire canteen consumption data and food waste monitoring data of students, and to perform correlation analysis on the acquired data to generate group taste preference weights. The feature construction module is used to obtain the standard parameters of nutritional intake and the set of dietary restrictions for the student group. The group's taste preference weights, standard parameters of nutritional intake and the set of dietary restrictions are vectorized and concatenated to construct a dietary demand feature vector. The recipe generation module is used to obtain a standardized recipe knowledge graph. Based on dietary demand feature vectors, it performs multi-objective optimization within the search space of the standardized recipe knowledge graph to generate a target nutritional recipe set containing several recommended dishes and the ingredient ratios of each recommended dish. The graph model building module is used to parse the preparation process of each recommended dish in the target nutritional recipe set, generate a sequence of atomic operation steps, and build a directed acyclic graph model. The cooking scheduling module is used to obtain the kitchen resource constraints, perform topological sorting and time parameter calculation on the directed acyclic graph model based on the critical path algorithm and the kitchen resource constraints, and generate a cooking sequence scheduling scheme. The dynamic optimization module is used to issue and execute the cooking sequence scheduling plan and obtain the time deviation data during the execution process. When the time deviation data exceeds the preset time threshold, the atomic operation step sequence that has not yet been executed is rearranged and optimized based on the directed acyclic graph model.

Images