Campus intelligent nutrition catering system based on deep semantic understanding and hybrid optimization

Abstract

The application discloses a campus intelligent nutrition meal system based on deep semantic understanding and hybrid optimization, relates to the field of intelligent catering data processing, and comprises a voucher generation module, a demand purification module, a chain graph construction module and a hybrid optimization module.The voucher generation module collects system data containing associated text and transaction records, separates original intention and actual substitute dishes through a language model, and generates a will-violating meal replacement counterfactual voucher.The demand purification module calculates the amount of money to be credited back, the amount of money to be deducted, and the amount of missing supply to be supplemented according to the voucher and transaction data, and obtains the purified demand amount by fusion.The chain graph construction module establishes a directed substitution chain edge based on the voucher, calculates the substitution equivalent weight, and generates a substitution chain account graph.The hybrid optimization module establishes constraint conditions and a global objective function based on the purified demand amount and the substitution chain account graph, executes a multi-stage solving logic, and generates an anti-factual backtracking meal operation instruction.The application effectively corrects the underlying flow data polluted by the will-violating meal replacement behavior, breaks the supply-demand mismatch chain, and realizes accurate resource scheduling.

Description

Technical Field

[0001] This invention relates to the field of smart catering data processing, specifically to a campus intelligent nutrition meal planning system based on deep semantic understanding and hybrid optimization. Background Technology

[0002] With the popularization of smart campuses, intelligent meal distribution systems in school canteens play an important role in improving kitchen operational efficiency and dining experience. In conventional multi-window parallel meal service scenarios, meal service generally faces the objective situation of short-term high concurrency and highly concentrated resources. To achieve a reasonable allocation of catering processing resources and window service resources, existing intelligent meal distribution and scheduling systems usually use the structured transaction flow generated by the cashier terminal and historical apparent sales as the direct calculation benchmark for menu configuration, meal preparation decisions, and ingredient procurement. The underlying control logic of existing systems generally assumes that the terminal transaction flow can completely and equivalently reflect the actual preferences of the dining group. If a certain dish shows high sales in historical sales data, the system's data processing center will directly increase the target number of meals prepared, the amount of ingredients purchased, and the scheduling priority of the corresponding dish in the scheduling algorithm for the next meal supply cycle.

[0003] However, in the actual operation environment of multi-window catering, using purely structured apparent transaction data as the sole calculation benchmark can easily lead to the underlying business data deviating from the true intent. During peak dining hours, when popular nutritional meals sell out in advance or the queues at the target catering window become too congested, diners are often constrained by the objective environment and abandon their initial desired dishes, instead purchasing alternative dishes from adjacent windows or of similar types, thus triggering large-scale compromise and substitution behavior. Existing data collection and processing systems only have the ability to capture the final settlement results and cannot observe the implicit lost demand for initially desired dishes, nor can they effectively utilize the causal logic contained in unstructured text such as search records generated by interactive terminals and customer complaints to deeply correct the underlying ledger. Due to the aforementioned blind spots, existing systems inevitably misidentify the inflated sales of alternative dishes as the group's true preference, and at the same time misjudge the unmet demand for initially desired dishes as natural demand contraction. As time goes on, unfiltered biased data will produce a significant error propagation effect when it enters subsequent optimization algorithms. This will eventually cause the menu configuration instructions, meal preparation quantity calculations, ingredient procurement plans, and window supply scheduling schemes automatically issued by the system to continuously deviate from the actual business needs, thus leading to a serious supply and demand mismatch problem. Summary of the Invention

[0004] To address the shortcomings of existing technologies, this invention proposes a campus intelligent nutrition meal planning system based on deep semantic understanding and hybrid optimization. This system solves the problem that it relies solely on the apparent transaction flow polluted by compromised meal substitution behavior for scheduling, and fails to effectively extract counterfactual causal information from unstructured text for demand correction, ultimately leading to the system's scheduling and menu configuration continuously deviating from actual dining needs.

[0005] To achieve the above objectives, the present invention provides the following technical solution: a campus intelligent nutrition meal planning system based on deep semantic understanding and hybrid optimization, comprising: The voucher generation module is used to collect system data containing associated text records and transaction records, extract the text to be parsed, separate the original intended dish and the actual substitute dish through a pre-trained language model, calculate the comprehensive confidence of the voucher, and generate a counterfactual voucher for a dish that was not intended to be substituted. The demand purification module is used to calculate the apparent sales volume based on transaction record data, calculate the original intention to make up the amount of revenue and the amount of revenue deducted for substitution pollution based on the counterfactual vouchers of unwilling substitution, call the associated text records in the system data to calculate the demand for replenishment of supply without transactions, and integrate the apparent sales volume, the original intention to make up the amount of revenue, the amount of revenue deducted for substitution pollution and the demand for replenishment of supply without transactions to obtain the purified demand volume. The chain graph construction module is used to establish directed substitution chain edges based on counterfactual evidence of unwilling meal substitution, calculate substitution equivalence weights by combining the comprehensive confidence of the evidence, and generate substitution chain ledger records. The hybrid optimization module is used to calculate and predict the purification demand for the next cycle based on the purification demand. It combines the alternative chain ledger records to establish system constraints and a hybrid optimization global objective function that includes menu selection status and ingredient procurement quantity. It solves and generates a set of counterfactual callback-type meal preparation operation instructions and sends them to the terminal device.

[0006] Compared with existing technologies, it has the following advantages: This proposed intelligent campus nutrition meal planning system, based on deep semantic understanding and hybrid optimization, aligns and parses discrete related text and transaction records using a pre-trained language model. This accurately separates the original intended dishes and actual substitute dishes for diners, and generates counterfactual evidence of involuntary substitutions by combining multi-dimensional confidence calculations. Based on this evidence, the system performs deep purification of the underlying transaction flow, calculating the amount of revenue recovered and the amount of revenue deducted, and reconstructing the true objective demand boundary by integrating the demand for supplementary meals in case of supply shortages. This mechanism effectively overcomes the shortcomings of traditional meal planning systems that rely solely on apparent settlement flow, leading to loss of intent and data pollution. It fundamentally restores the implicit dining needs compromised due to supply constraints, providing an absolutely objective and accurate data foundation for subsequent production scheduling, and significantly improving the system's perception accuracy and decision-making scientific rigor.

[0007] Based on the obtained purification demand, this solution further establishes directed substitution chain edges and constructs a substitution chain ledger graph, which is deeply integrated into the multi-stage solution logic as a core constraint. The hybrid optimization global objective function constructed by the system forcibly introduces a substitution chain ledger graph correction penalty mechanism. When the amount of food prepared for the original intention dish allocated by the algorithm is lower than its purification demand, a significant product penalty will be triggered. This algorithm architecture directly blocks the erroneous transmission chain of supply and demand mismatch at the mathematical optimization level, ensuring that the generated counterfactual callback-type meal preparation instructions can accurately redirect the overflow supply capacity back to the true intention source. Through the above-mentioned closed-loop process from bottom-level data cleaning to top-level operation and scheduling, this invention achieves the ultimate matching of processing resources and window service resources, completely solving the industry problem of long-term shortage of popular dishes and waste of unsold dishes in the campus catering scenario. Attached Figure Description

[0008] Figure 1 This is a schematic diagram of the system framework of the present invention.

[0009] Figure 2 This is a schematic diagram of the system execution flow of the present invention. Detailed Implementation

[0010] 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.

[0011] Please see Figures 1 to 2 This application provides a campus intelligent nutrition meal distribution system based on deep semantic understanding and hybrid optimization, including a voucher generation module, a demand purification module, a chain graph construction module and a hybrid optimization module; Among them, the voucher generation module is used to extract counterfactual causal information from discrete and multi-source underlying data in the scenario of multi-window meal service in campus canteens, and transform it into structured vouchers that can represent the original true intentions of students, providing a data foundation for subsequent system requirement reconstruction and resource scheduling.

[0012] This module synchronously collects multi-dimensional system data from the school canteen according to the set target meal service cycle. Within the target meal service cycle, the module continuously acquires transaction record data, associated text records, dish and supply status data, and window operation data through the application programming interface. By parsing the database logs and calling the basic dish table records, it obtains the category and recipe quantity corresponding to each standard dish. It collects the availability status of dishes at preset time points in real time, as well as the system timestamp when a dish sells out. This module synchronously monitors the estimated queuing time, maximum supply capacity limit, and opening status of each meal service window at preset times.

[0013] Specifically, the target meal service cycle, as the smallest unit of meal service that can be independently statistically analyzed and decided upon, has a parameter set that includes at least the date, meal type, and cafeteria number. Transaction record data is obtained from structured transaction logs generated by the POS terminal, which includes the transaction index, anonymous student ID, actual purchased dishes, equivalent number of portions purchased, transaction amount, and transaction time accurate to the second. Related text records are obtained from unstructured text generated by students on the interactive terminal, which includes the text index, anonymous student ID, text source category, text content string, and text generation time.

[0014] It should be noted that the purchase equivalent quantity refers to the floating-point relative share obtained by dividing the weight of non-standardized weighed dishes by the system's preset standard benchmark weight. By aligning isolated POS hardware terminal data with kitchen service status monitoring data on the same timeline through application programming interfaces and database log parsing, it is possible to uncover abnormal deviations hidden behind transaction flows, thereby constructing an objective data foundation.

[0015] This module performs an underlying identity and time-based alignment mapping on the collected associated text records and transaction records. It determines whether the anonymous student ID in the associated text record matches the anonymous student ID in the transaction record and whether they are within the same target meal service period. Subsequently, it executes time window verification logic, determining whether the text generation time falls within the preceding and following association window based on the transaction time. When multiple associated text records fall within the association mapping range of the same transaction record, this module extracts the highest-priority text and maps it to the text to be parsed for that specific transaction, based on the priority rules of customer complaints over ratings, ratings over notes, and notes over searches.

[0016] Specifically, the time constraints are that the text generation time must not be earlier than the preset pre-search time before the transaction time, and must not be later than the preset post-evaluation time after the transaction time. The context window is a benchmark that extends bidirectionally to both the past and future time dimensions, with the transaction time as the baseline.

[0017] It should be noted that the preset search pre-set time and preset evaluation post-set time are constant parameters set based on the objective laws of canteen operation. For example, the preset search pre-set time can be set to 30 minutes, and the preset evaluation post-set time can be set to 120 minutes. By utilizing the proximity of time and the uniqueness of feedback paths, unstructured evaluations that are substantially relevant to the current transaction can be accurately extracted, thereby filtering out underlying data noise interference from cross-category or irrelevant historical text.

[0018] A finely tuned pre-trained Chinese language model is used as the semantic encoder. The text to be parsed is input into a multi-layer attention mechanism network to generate a contextual semantic vector. This module performs joint inference through three parallel, independent task output heads. First, the module extracts dish entities from the text using a sequence labeling network and extracts the local semantic feature vectors of these entities. A spatial distance algorithm is then used to compare these local semantic feature vectors with standard name vectors pre-stored in a basic dish table, accurately separating the original intention dish with an initial intention expression from the actual alternative dish with a compromise expression. Subsequently, a classification network identifies the objective reasons hindering the alternative purchase and categorizes them as alternative purchase reason labels. Finally, the module outputs the model probability value that the semantics of the text support the validity of the unintended alternative behavior, which is used as the text semantic confidence score.

[0019] Specifically, the standard name vectors pre-stored in the basic dish table are obtained by the system using the same semantic encoder to encode and map the basic dish names, ensuring that the spatial distance algorithm performs cosine similarity or inner product operations within a unified feature space. The original intended dish refers to the restaurant entity identifier that the user initially hoped to obtain but failed to acquire due to objective environmental limitations. The actual substitute dish refers to the restaurant entity identifier that the user actually paid for at the cashier after compromising due to objective conditions. The substitute purchase reason label refers to the operational label that abstracts and summarizes natural language into machine-readable discrete states such as sold out or queue congestion. Text semantic confidence refers to the dimensionless floating-point number output by the model used to quantify the reliability of the extracted conclusions.

[0020] It should be noted that this module uses historical evaluation corpora of campus catering to fine-tune the feature layer of the pre-trained model, reducing the dimensionality of complex multidimensional vectors of natural language into structured intent control labels that can directly participate in machine logic operations. Taking the actual operation process as an example, if the input text is such as "I wanted to eat sweet and sour pork ribs but they were sold out, so I could only get scrambled eggs with tomatoes," the output of this module will parse and map the original intent dish as sweet and sour pork ribs, the actual substitute dish as scrambled eggs with tomatoes, and the substitute purchase reason label as "sold out." This mechanism realizes cross-modal conversion from the natural semantic space to the computer resource scheduling instruction space.

[0021] This module uses objective operational status data to verify subjective semantics. It checks whether the availability of the original intended dish was zero at the transaction time, or whether its sold-out system timestamp was earlier than the transaction time. It extracts the estimated queuing time for the window containing the original intended dish at the transaction time and determines if it exceeds a preset excessive queuing time threshold. It verifies whether the window is closed. When any of the above objective conditions are met, the module assigns an objective confidence level of one to the operational evidence; otherwise, it assigns a value of zero. Simultaneously, the module calculates the time-matching confidence level.

[0022] Specifically, the time-match confidence score is obtained using the following normalization formula: In the formula, The confidence score for time matching is a dimensionless floating-point number that represents the quantitative assessment of the time correlation distance. A system timestamp indicating the time the text was generated; The system timestamp for the transaction time; The preset maximum tolerance time is a physical measure of the maximum allowable traceability or delay time of the system. In practice, the value can be 120 minutes.

[0023] It should be noted that the preset threshold for excessively long queue times is configured as the maximum of two factors: a preset fixed lower limit and a preset multiple of the historical median waiting time for the same meal type within that window. For example, this preset multiple could be 1.5 times. The objective confidence level of operational evidence is a Boolean verification identifier generated solely from kitchen sensors and the status of the cash register, completely eliminating subjective factors. Introducing objective state data and a time operator with linear decay characteristics effectively constrains the large-scale subjective semantic model. By using hard data verification to reduce the risk of semantic illusion, the stability of system decisions is improved.

[0024] This module performs comprehensive logical decision-making on multi-dimensional features. It extracts textual semantic confidence, objective confidence of business evidence, and time-matching confidence, performs algebraic weighted summation to obtain the overall confidence of the transaction's documentation. Subsequently, it performs strict interception verification; the module determines that the unauthorized substitution purchase actually occurred only if the preset interception conditions are met. After deduplication logic filtering based on the same anonymous student ID and short time intervals, the module generates standardized counterfactual evidence of unauthorized meal substitution and outputs it to the next-level processing module of the system.

[0025] Specifically, the weighted summation formula for the overall confidence level of the vouchers is as follows: In the formula, CF is the overall confidence level of the voucher, which is a dimensionless quantitative indicator used to assess the authenticity of alternative behaviors. For text semantic confidence; To ensure the objective confidence level of the evidence in business operations; Match confidence levels to time; As the first weight, As the second weight, The first weight is set to 0.55, the second weight to 0.35, and the third weight to 0.10. These three weights are normalized control parameters used to fuse multidimensional confidence features. In a preferred embodiment of the present invention, the first weight can be configured to 0.55, the second weight can be configured to 0.35, and the third weight can be configured to 0.10.

[0026] The preset interception conditions must simultaneously meet three independent data conditions: the first condition is that the referential data of the original intended dish and the actual substitute dish are inconsistent; the second condition is that the actual substitute dish is numerically equal to the actual purchased dish in the cashier transaction record; and the third condition is that the calculated overall confidence level of the voucher is greater than or equal to the preset effective threshold, the preferred value of which can be 0.65.

[0027] It should be noted that the aforementioned effective threshold refers to a numerical watershed used to filter low-value noise. The counterfactual credential for voluntary meal substitution is a read-only structured data packet defined at the underlying level. Internally, it encapsulates at least the original intended dish, the actual substitute dish, the equivalent number of portions purchased, and the overall confidence level of the credential. This credential declares and solidifies the fact that the abnormal catering transaction occurred, providing a precise and confidence-based trigger for intercepting tainted apparent sales and initiating data ledger reconstruction.

[0028] The demand purification module is used in the scenario of multi-window meal service in campus canteens to receive evidence of voluntary meal substitution, correct the data logic of the initial sales recorded by the transaction system, build a sales clawback ledger and output the purification demand to remove contamination, so as to provide a real and objective basis for the subsequent resource optimization of the system.

[0029] This module calculates the actual apparent sales volume of each standard dish within the target meal service period based on the transaction record data of the POS system.

[0030] Specifically, this module iterates through all transaction records within the target meal supply cycle, filters out records where the actual purchased dishes match the currently iterated standard dishes, sums up the equivalent purchase quantities in the corresponding records, and obtains the apparent sales volume of the standard dish.

[0031] It should be noted that apparent sales volume represents the actual number of units sold directly recorded by the transaction system. Due to the possibility of substitute purchases caused by popular dishes selling out or long queues, apparent sales volume does not equate to the actual objective demand from students. Therefore, it must be used as the basis for subsequent data logic correction.

[0032] Based on the counterfactual evidence of unwilling meal substitution output by the previous level, this module calculates the original intention to recover the amount of the meal and the amount to be deducted for substitution pollution for each standard dish, and implements anti-over-deduction restrictions on the amount to be deducted for substitution pollution.

[0033] Specifically, for calculating the amount of original intention to be reimbursed, this module filters out the set of dishes in the vouchers whose original intention matches the current standard dishes. It then multiplies the equivalent number of purchases in the vouchers by the corresponding voucher's overall confidence level and sums all the products. For calculating the amount of deduction due to substitution pollution, this module filters out the set of actual substitution dishes in the vouchers whose actual substitution dishes match the current standard dishes, and performs the same multiplication and summation operations. Simultaneously, this module compares the calculated amount of deduction due to substitution pollution with the apparent sales volume of the standard dish, taking the smaller of the two as the final amount of deduction due to substitution pollution.

[0034] It should be noted that the comprehensive confidence level of the voucher is used as a multiplicative operator in the calculation, transforming the replenishment and deduction amounts into equivalent shares calculated based on reliability, rather than simple absolute counts. The objective basis for the anti-over-deduction limit is that, logically, the artificially inflated sales volume deducted due to the substitution of any dish cannot exceed the number of servings of that dish actually recorded in the transaction system. This mechanism effectively avoids the risk of negative inventory due to data overflow in the system.

[0035] This module extracts records of unsold or out-of-stock items in the system that have not generated purchase transactions, and calculates the replenishment demand for each standard dish within the target meal supply cycle.

[0036] Specifically, when a student's text contains the original intention of listing a dish and the semantics of "sold out" or "queue congestion," and there are no alternative purchase transactions for the same anonymous student ID within a preset time span before and after the text's generation, this module generates a record of unsold items. This record includes the dish ID, the confidence level of the unsold item, and the indicated number of unsold portions. The module filters out records where the dish ID matches the current standard dish, multiplies the confidence level by the number of unsold portions, and performs a global summation to obtain the unsold replenishment demand for that standard dish.

[0037] It should be noted that the preset time span mentioned above is a constant parameter set based on the objective laws of dining decisions, such as 30 minutes; the default effective threshold for the lack of supply confidence is 0.65; when the number of unavailable portions is not specified, the system defaults to treating it as 1 portion. The aforementioned lack of supply confidence is the semantic confidence of the text output by the pre-trained language model for the segment of text with no transaction feedback. The basis for this calculation is that after popular dishes are sold out, some students will directly give up purchasing them and no longer generate alternative consumption. If only the alternative purchase behavior is corrected, this lost demand will be missed. Introducing this value can fully restore the true objective demand boundary of popular dishes.

[0038] Based on the initial apparent sales volume, this module integrates data from replenishment, deduction, and supply shortages to calculate the core indicator purification demand for each standard dish and generates a sales clawback ledger containing ledger reliability data.

[0039] Specifically, the purification requirements and ledger reliability data for this module are obtained through the following normalization and algebraic summation formulas: In the formula, The amount of purification required is the equivalent number of parts that represent the actual objective demand; Apparent sales figures; To replenish the amount recorded to the original intention; The amount deducted to replace pollution is a deduction from the account. This is to supplement demand to address the lack of supply due to a lack of transactions.

[0040] In the formula, The reliability data for the return of the accounting records for this dish is a normalized parameter used to assess the credibility of the purification demand. This represents the number of valid vouchers related to the dish; the constant 10 is a preferred embodiment value of the preset threshold for the number of sufficient vouchers. This module encapsulates apparent sales, replenished accounting entries, deducted accounting entries, supply shortage replenishment entries, purification demand entries, and reliability data into a sales clawback ledger for the standard dish.

[0041] It should be noted that the purification demand is a true operational demand indicator after completely eliminating pollution from alternative purchases and making up for lost demand due to supply shortages. This indicator is not an actual settlement record, but rather objective basic data directly used to drive the downstream hybrid optimization module. The ledger reliability data is set according to the objective statistical law of large numbers. Due to the randomness of individual student feedback, when the number of associated valid vouchers reaches the preset sufficient voucher quantity threshold, the sales pollution phenomenon can be considered to have group stability. This mechanism ensures that the system's underlying data ledger has both extremely high accuracy and robustness.

[0042] Among them, the chain graph construction module is used to transform discrete counterfactual vouchers into alternative chain edges with directed connections in the multi-window food service scenario of campus canteens, filter out occasional noise data, construct and output an alternative chain ledger graph that reflects the direction and scale of demand spillover, and provide a correlation basis for subsequent hybrid optimization.

[0043] This module constructs the basic topology of the substitution chain graph based on the received counterfactual vouchers of unwilling meal substitutions, extracts the original intended dish number and the actual substitute dish number from the voucher to establish directed substitution chain edges, and calculates the substitution equivalence weight of each directed substitution chain edge.

[0044] Specifically, this module maps each standard dish to an independent node in the substitution chain ledger. If there is a substitution record in the counterfactual evidence of a dish substitution that points from the original intended dish to the target actual substitute dish, the module establishes a directed connection between the corresponding starting node and the target node. For any directed substitution chain edge, the module extracts all evidence belonging to that edge, multiplies the overall confidence level of each evidence by the number of equivalent purchases, and globally sums the product results to obtain the substitution equivalence weight of that directed substitution chain edge. The calculation formula is: In the formula, The substitution equivalent weight is the number of equivalent portions that represent the scale of demand shift when the starting dish is replaced by the target dish. The overall confidence level of the corresponding voucher; This represents the number of equivalent purchases of the corresponding voucher.

[0045] It should be noted that transforming discrete vouchers into directed edges in a graph structure clearly depicts the transfer paths of demand between different windows or dishes. The substitution equivalent weight is not simply an additive calculation of occurrence frequencies, but incorporates a weighted calculation of the overall confidence level of the vouchers, ensuring the objectivity and accuracy of the transfer scale statistics.

[0046] Based on the calculated substitution equivalent weights, this module further calculates the demand substitution transfer rate from each starting dish to each target dish, and combines the demand for supplementing supply shortages due to lack of transactions with the demand for purification to calculate the supply shortage impact of each starting dish.

[0047] Specifically, the demand substitution shift rate and the impact of supply shortages are obtained through the following formula: In the formula, The demand substitution transfer rate from the starting dish to the target dish is a dimensionless parameter that characterizes the proportion of substitution distribution. The sum of the equivalent substitution weights for transferring the starting dish to all associated target dishes. To prevent the use of minimum constants with a denominator of zero.

[0048] In the formula, The impact of the shortage of the initial dish is a dimensionless indicator that measures the severity of the loss of demand. The sum of the equivalent substitution weights for the spillover of the initial dishes; To supplement the demand for the initial dishes that have not sold; This is the amount of purification required for the initial dishes.

[0049] It should be noted that the zero-minimum constant is... The preferred embodiment value can be 0.001. The demand substitution transfer rate reveals the spillover consumption preferences of the student group when a single dish is unavailable; the supply shortage impact combines the two situations of substitution purchase and direct loss, accurately quantifying the negative impact of dish shortage on the overall canteen operation, and providing a direct quantitative indicator for subsequent optimization algorithms to identify key bottleneck dishes.

[0050] This module performs multi-dimensional validity screening logic on the initial directed alternative chain edges, removes noisy edges caused by occasional text, and finally generates and outputs cleaned and denoised alternative chain ledger graph records.

[0051] Specifically, this module retains only directed substitution chain edges that meet preset valid edge determination conditions. These conditions consist of three independent logical branches; satisfying any one of them is sufficient to retain the edge: the first branch is that the substitution equivalent weight of the directed substitution chain edge is higher than or equal to a preset weight scale threshold; the second branch is that the demand substitution transfer rate is higher than or equal to a preset transfer ratio threshold, and the number of valid vouchers associated with the directed substitution chain edge is not less than a preset voucher counting threshold; the third branch is that the arithmetic mean of the overall confidence levels of all vouchers associated with the directed substitution chain edge is higher than or equal to a preset average confidence level threshold. This module structures and encapsulates the retained directed substitution chain edges and their associated weights, transfer rates, supply shortage impact, and the distribution ratio of substitution purchase reason tags, outputting a substitution chain ledger record.

[0052] It should be noted that the preset weight scale threshold can be configured to three equivalent parts, the preset transfer ratio threshold can be configured to 5%, the preset voucher piece-rate threshold can be configured to two, and the preset average confidence threshold can be configured to 0.70. School canteen meal service is characterized by high frequency, small amounts, and short cycles; a single, occasional customer complaint should not directly induce a drastic shift in overall meal preparation decisions. Through the joint screening mechanism of the above thresholds, it is possible to effectively distinguish between occasional student complaints and systemic supply-demand mismatch chains, making the final output alternative chain diagram possess extremely high operational guidance value and resistance to interference.

[0053] Among them, the hybrid optimization module is used in the scenario of multi-window catering in the campus canteen. Based on the purification demand output by the demand purification module and the alternative chain ledger output by the chain graph construction module, it executes rule filtering, branch and bound algorithm and swarm intelligence heuristic allocation to generate control instructions for catering menu, food preparation, procurement and window supply, and completes the transformation of the underlying data ledger into an entity resource scheduling scheme.

[0054] Based on the purification demand of historical meal service cycles and the semantic intent advance demand before the target meal service cycle, this module calculates the predicted purification demand for the next cycle for each standard dish.

[0055] Specifically, if the target meal supply cycle is a replenishment decision for the current cycle, the already generated purification demand is directly used; if it is a planned decision, the average purification demand for the same meal type over the last four weeks, the purification demand for the same meal type over the last meal supply cycle, and the semantic intent advance demand before the target meal supply cycle are extracted. These three data points are then combined using a weighted formula. The weighted formula is: In the formula, The predicted purification demand for the next cycle is the equivalent number of servings used to guide the production schedule for the target meal supply cycle. The average purification requirement for the same meal type and the same week in history; The purification requirement for the most recent meal service cycle of the same type; The semantic intent is the advance demand, which represents the advance demand expressed by students using search and note functions before the target meal supply cycle; As the first time weight, As the second time weight, The third time weight is used, and all three are normalized adjustment parameters.

[0056] Specifically, the demand for semantic intent in advance The acquisition method is as follows: the system extracts terminal search logs and inactive order remarks text within a preset time period (such as 2 hours before the start of the target meal supply cycle), uses the aforementioned pre-trained language model to identify intent entities that match the current standard dishes, and performs deduplication based on anonymous student IDs; the number of times the deduplicated independent and valid demand expressions are mapped to the equivalent number of advance demand portions of the dish in a one-to-one ratio.

[0057] It should be noted that the preferred embodiments for the first, second, and third time weights can be configured with values ​​of 0.60, 0.25, and 0.15, respectively. The objective basis for these weight settings lies in the strong cyclical nature of campus catering demand; therefore, the average value for the same week dominates, while the semantic intent-based advance demand reflects immediate, incidental preferences. Through the above weighted calculation, abnormal fluctuations in demand can be smoothed out, providing an objective data benchmark for subsequent meal preparation. For new dishes lacking historical data, the system automatically uses the average data of standard dishes of the same category as the initial benchmark for predicting the next cycle's purification demand.

[0058] Based on the predicted purification demand for the next cycle, this module establishes system constraints covering multi-dimensional decision variables such as menu selection status, target number of dishes to be prepared, amount of ingredients purchased, and amount allocated to designated windows. It also constructs a hybrid optimization global objective function containing multiple sub-objective functions.

[0059] Specifically, the system constraints include upper and lower limits for food preparation, deviation prevention constraints, inventory balance constraints, and window capacity constraints. The target number of servings of a dish must be between the preset minimum preparation batch size and the maximum preparation limit, and is restricted within the preset fluctuation range of the predicted purification demand for the next cycle. The total planned consumption of ingredients, after deducting the current available inventory, must not be lower than the preset safety reserve inventory, and the ingredient procurement quantity must comply with the supplier's preset packaging unit limits. The total amount of dishes allocated to a designated window must equal the target number of servings of the dish, and the total allocation to a single serving window must not exceed the maximum supply capacity limit of that window. The hybrid optimization global objective function and core sub-objective functions are constructed using the following formulas: In the formula, The global objective function for hybrid optimization is a dimensionless index that measures the overall operating cost and execution deviation of the system resource scheduling scheme. Each is an independent sub-objective function; These are the weight coefficients corresponding to each sub-objective function. The demand matching deviation sub-objective function is used to quantify the degree of supply-demand matching and to prevent erroneous substitution chains. Correcting the penalty sub-objective function with alternative chained ledger graph Expanded separately as follows: In the formula, The target number of servings to be prepared for the dishes is the decision variable. The target number of servings to be prepared for the original intended dish; The predicted purification demand for the next cycle of the original intended dishes; To retrieve the substitution equivalence weight of the starting dish that is replaced by the target dish from the substitution chain ledger graph; To prevent local minima with a denominator of zero; in the formula This represents a global traversal and summation of all retained directed alternative chain edges in the alternative chain ledger graph and their associated initial original intention dishes.

[0060] It should be noted that the preset floating range can be configured to be 70% to 130% of the predicted purification demand for the next cycle. The minimum batch size for meal preparation varies from 20 to 50 portions depending on the characteristics of the kitchen processing technology. The formula logic of the core sub-objective function is as follows: When the algorithm seeks the optimal solution, it must take into account the equivalent weights that were previously overflowed. Once the system provides the amount of food to prepare for the original intended dish... Below its actual purification requirements This will generate a huge product penalty value. This mathematical structure ensures that the final production scheduling instructions will prioritize increasing the supply of severely substituted dishes, completely blocking the continued transmission of the erroneous substitution chain. Other sub-objective functions, such as total procurement cost, are constructed using industry-standard normalized linear cost aggregation and deviation measurement methods, and will not be elaborated here.

[0061] This module performs a three-stage solution logic, which includes a combination of preset algorithms, on the constructed system constraints and the global objective function of hybrid optimization. It sequentially filters the candidate set through rules, establishes the meal preparation and procurement plan through the branch and bound algorithm, and completes the window supply scheduling through the particle swarm algorithm.

[0062] Specifically, in the first stage, dishes with no available inventory, whose respective serving windows are not open, or whose purification demand has been zero for a preset historical observation period (e.g., the most recent 7 consecutive days) are screened out, while standard dishes with a supply shortage impact exceeding a preset loss threshold are forcibly retained in the candidate menu set. In the second stage, a branch and bound algorithm is used on the candidate menu set to solve for the menu selection status, the target number of dishes to be prepared, and the amount of ingredients to be purchased, in order to minimize the set of preceding sub-objective functions (e.g., demand matching deviation, sell-out risk, leftover risk, and total procurement cost) included in the global objective function of the mixed optimization. to In the third stage, based on the determined target number of dishes to be prepared, the particle swarm optimization algorithm is used to solve the problem of allocating the dishes to a specified window. The configuration state of individual particles in the particle swarm optimization algorithm is updated through multiple rounds of iteration until the preset algorithm convergence or timeout stopping conditions are met.

[0063] It should be noted that the preset churn threshold value can be 0.20, the preset solution time limit for the branch-and-bound algorithm can be set to 300 seconds, and the preset number of particles for the particle swarm optimization algorithm can be set to 30 and the maximum number of iterations can be set to 80. In the particle swarm optimization algorithm, the configuration state of a single particle is strictly mapped to a multi-dimensional allocation matrix in the mathematical model. The rows of this matrix represent the standard dish set, the columns represent the serving window set, and the element values ​​within the matrix represent the allocation amount of the corresponding dish to the specified window. The above-mentioned three-stage solution mechanism, which combines rule filtering, discrete integer programming, and heuristic swarm intelligence algorithms, is a preferred embodiment of the multi-stage solution logic of this invention. It can avoid system timeouts and crashes caused by a single large mathematical model, ensuring stable output of suboptimal or optimal resource scheduling solutions that meet strict constraints within a limited time window.

[0064] Based on the final decision variable results output from the three-stage solution, this module packages and generates a set of counterfactual callback-type meal delivery operation instructions, and then sends the instruction set to the corresponding terminal devices in the school canteen.

[0065] Specifically, the meal preparation operation instruction set includes: a menu restoration instruction generated based on the menu selection status, used to re-include dishes whose actual demand was missed in the production queue; a meal preparation correction instruction generated based on the target number of servings of a dish; a procurement correction instruction generated based on the amount of ingredients purchased; and a window supply callback instruction generated based on the amount allocated to a designated window, used to reallocate supply and demand overflowing due to congestion to the original intended dish's serving window mapped to the starting node. When the load rate of a serving window exceeds the preset warning limit or the impact of a shortage of standard dishes is abnormal, a risk warning instruction is triggered simultaneously.

[0066] It should be noted that the preset warning upper limit can be set to 90% in this embodiment. The aforementioned terminal devices include, but are not limited to, electronic dashboards in the kitchen, resource planning system terminals in the purchasing department, and displays at the food service windows. By issuing a complete set of control commands, the system forms a closed-loop data feedback chain to prevent contamination, correcting the supply mismatch problem caused by blindly preparing meals based solely on apparent flow rates in the traditional meal preparation model, and achieving precise allocation of kitchen processing resources and window service resources to students' true intentions.

[0067] The above embodiments are only used to illustrate the technical methods of the present invention and are not intended to limit it. Although the present invention has been described in detail with reference to preferred embodiments, those skilled in the art should understand that modifications or equivalent substitutions can be made to the technical methods of the present invention without departing from the spirit and scope of the technical methods of the present invention.

Claims

1. A campus intelligent nutrition catering system based on deep semantic understanding and hybrid optimization, characterized in that, include: The voucher generation module is used to collect system data containing associated text records and transaction records, extract the text to be parsed, separate the original intended dish and the actual substitute dish through a pre-trained language model, calculate the comprehensive confidence of the voucher, and generate a counterfactual voucher for a dish that was not intended to be substituted. The demand purification module is used to calculate the apparent sales volume based on transaction record data, calculate the original intention to make up the amount of revenue and the amount of revenue deducted for substitution pollution based on the counterfactual vouchers of unwilling substitution, call the associated text records in the system data to calculate the demand for replenishment of supply without transactions, and integrate the apparent sales volume, the original intention to make up the amount of revenue, the amount of revenue deducted for substitution pollution and the demand for replenishment of supply without transactions to obtain the purified demand volume. The chain graph construction module is used to establish directed substitution chain edges based on counterfactual evidence of unwilling meal substitution, calculate substitution equivalence weights by combining the comprehensive confidence of the evidence, and generate substitution chain ledger records. The hybrid optimization module is used to calculate and predict the purification demand for the next cycle based on the purification demand. It combines the alternative chain ledger records to establish system constraints and a hybrid optimization global objective function that includes menu selection status and ingredient procurement quantity. It solves and generates a set of counterfactual callback-type meal preparation operation instructions and sends them to the terminal device. 2.The campus intelligent nutrition catering system based on deep semantic understanding and hybrid optimization according to claim 1, wherein, include: It was determined that the anonymous student ID in the associated text record was consistent with the anonymous student ID in the transaction record data and that they were within the same target meal supply cycle; The execution time window verification logic determines whether the text generation time is within the pre- and post-review correlation window constructed based on the transaction time and the pre- and post-review time. When multiple related text records fall into the association mapping range of the same transaction record data, the highest priority related text record is extracted and mapped to the text to be parsed, based on the priority rules that customer complaints are higher than evaluations, evaluations are higher than remarks, and remarks are higher than searches. 3.The campus intelligent nutrition catering system based on deep semantic understanding and hybrid optimization according to claim 2, characterized in that, include: The collected system data also includes operational status data; Obtain the text semantic confidence of the pre-trained language model when separating the original intended dish from the actual substitute dish; Extract the text generation time of the associated text records and calculate the time matching confidence level with the transaction time of the transaction record data; By accessing operational status data, we can verify the availability of the original dishes, the estimated waiting time, and the openness of the food service windows to obtain objective confidence in operational evidence. The overall confidence level of the voucher is obtained by weighting the textual semantic confidence level, the time matching confidence level, and the objective confidence level of the business evidence. When the following preset interception conditions are met simultaneously: the reference data of the original intended dish and the actual substitute dish are inconsistent; the actual substitute dish is strictly equal to the actual purchased dish in the transaction record data; and the overall confidence level of the voucher is greater than or equal to the preset effective threshold, a counterfactual voucher for a dish with a breach of consent is generated.

4. The campus intelligent nutrition meal distribution system based on deep semantic understanding and hybrid optimization according to claim 1, characterized in that, include: Extract the equivalent number of purchases from the transaction record data, multiply the equivalent number of purchases by the overall confidence level of the vouchers contained in the counterfactual voucher of the unwilling meal substitution, and sum them up to obtain the calculated amount of deduction for substitution pollution. The calculated amount of alternative pollution deduction is compared with the apparent sales volume, and the smaller value is taken as the final amount of alternative pollution deduction to implement the over-deduction limit.

5. The campus intelligent nutrition meal distribution system based on deep semantic understanding and hybrid optimization according to claim 3, characterized in that, include: If the text to be parsed contains the original intent of the dish entity and the semantics of sold out or queue congestion, and there is no alternative purchase transaction with the same anonymous student ID within a preset time span before and after the text generation time, a record of unsold dish containing the dish ID is generated. The semantic confidence of the text output by the pre-trained language model in the non-sold supply records is used as the supply shortage confidence. The supply shortage confidence is multiplied by the number of supply shortages contained in the non-sold supply records and then summed globally to obtain the supply shortage replenishment demand corresponding to the non-sold supply dish number.

6. The campus intelligent nutrition meal distribution system based on deep semantic understanding and hybrid optimization according to claim 1, characterized in that, include: Standard dishes are mapped to independent nodes. A directed connection is established between the starting node of the original intention dish mapping contained in the counterfactual evidence of voluntary substitution and the target node of the actual substitution dish mapping to establish a directed substitution chain edge. The substitution equivalence weight is obtained by multiplying the overall confidence level of the vouchers included in the counterfactual evidence of unwilling meal substitution by the number of equivalent purchases and summing them globally. A minimum constant to prevent the denominator from being zero is introduced in conjunction with the equivalent weight of substitution to calculate the demand substitution transfer rate; The impact of supply shortage is calculated by introducing a minimum constant to prevent the denominator from being zero, combined with the equivalent weights of substitutes, the demand for supplementing supply shortages without transactions, and the demand for purification.

7. The campus intelligent nutrition meal distribution system based on deep semantic understanding and hybrid optimization according to claim 6, characterized in that, include: Retain directed alternative chain edges that meet the preset valid edge determination conditions to generate alternative chain ledger graph records; The preset valid edge determination condition includes a logical branch that satisfies any one of the following: The equivalent weight is higher than or equal to the preset weight size threshold; The demand substitution transfer rate is higher than or equal to the preset transfer ratio threshold and the number of valid vouchers associated with the directed substitution chain edge is not less than the preset voucher counting threshold. The arithmetic mean of the overall confidence levels of all credentials associated with the directed substitution chain is higher than or equal to the preset average confidence level threshold.

8. The campus intelligent nutrition meal distribution system based on deep semantic understanding and hybrid optimization according to claim 1, characterized in that, include: Obtain the semantic intent and advance demand before the target meal supply cycle; Extract the average purification demand of the same week and the same meal type in history, the purification demand of the most recent meal type in the same meal type, and the semantic intent advance demand, and perform a weighted formula to calculate and calculate the predicted purification demand of the next period. For new dishes lacking historical data, the average data of standard dishes in the same category is used as the initial benchmark for predicting the purification demand in the next cycle.

9. The campus intelligent nutrition meal distribution system based on deep semantic understanding and hybrid optimization according to claim 6, characterized in that, include: The hybrid optimization global objective function includes an alternative ledger graph correction penalty sub-objective function; The execution logic of the alternative chained ledger graph correction penalty sub-objective function includes: For the decision variable, the target number of servings of dishes, in the global objective function of hybrid optimization, a global traversal is performed on all retained directed substitution chain edges and the original intention dishes associated with the directed substitution chain edges in the substitution chain ledger graph. When it is determined that the target number of servings of dishes allocated to the original intention dishes is lower than the predicted purification demand of the original intention dishes in the next cycle, the difference between the predicted purification demand in the next cycle and the target number of servings of dishes is extracted. The difference is multiplied by the substitution equivalent weight to obtain the product penalty value, which is then included in the substitution chain ledger graph correction penalty sub-objective function for hybrid optimization solution.

10. The campus intelligent nutrition meal distribution system based on deep semantic understanding and hybrid optimization according to claim 9, characterized in that, include: The configuration state of a single particle in the particle swarm optimization algorithm is mapped to a multi-dimensional allocation matrix where rows represent the standard dish set and columns represent the food window set. The element values ​​inside the matrix represent the amount of the corresponding dish allocated to the specified window. The counterfactual callback type of catering operation instruction set includes: Based on the menu selection status obtained from the hybrid optimization solution, a menu restoration instruction is generated to re-include dishes whose actual demand has been missed in the production queue; a meal preparation correction instruction is generated based on the target number of meals prepared for the dishes; a procurement correction instruction is generated based on the ingredient procurement quantity obtained from the hybrid optimization solution; and a window supply reallocation instruction is generated based on the allocation quantity to a specified window to reallocate the overflow supply demand to the original intention dish to which the original intention dish is mapped as the starting node.

Images