Smart park management method and system based on park informatization demand

Abstract

The invention discloses a smart park management method and system based on park informatization demands, and relates to the technical field of data management and information communication, and the method comprises the steps: obtaining the demand information of various users and devices from a mobile terminal device and an edge gateway; identifying the demand information to generate a corresponding label, screening the demand information according to a label confidence coefficient, and merging the demand information into standard demand data by means of a park domain knowledge graph; constructing a bipartite graph according to the standard demand data, and weighting and calculating the matching degree according to the semantic similarity, the demand priority and the module processing capability; and respectively inputting the trigger event characteristics into a rule engine and an AI model, and fusing and outputting a final execution strategy according to a smooth function. According to the method, through a unified corpus form, the embedding quality is improved, the false similarity and error entity matching rate is remarkably reduced, the matching degree, the emergency degree and the system bearing capacity can be balanced and considered, and a global optimal service calling set is automatically generated.

Description

Technical Field

[0001] The present invention relates to the field of data management and information communication technology, and specifically to a smart park management method and system based on the park's informatization needs. Background Art

[0002] With the rapid development of the Internet of Things (IoT), cloud computing, and mobile internet technologies, campus management has gradually evolved from traditional manual inspections and single-system monitoring to a comprehensive "smart campus" operation model. Early solutions relied on centralized PLCs or DCSs to isolatedly control energy consumption, access, and security equipment, making it difficult to achieve cross-subsystem data connectivity and unified scheduling. Multi-tiered architectures based on edge computing are emerging, offloading some latency-sensitive tasks to the campus gateway for local pre-processing and preliminary decision-making. Meanwhile, cloud-based big data platforms handle unified data aggregation, in-depth analysis, and visualization. In recent years, the convergence of knowledge graphs, digital twins, and artificial intelligence has become a research hotspot. By constructing semantic networks and virtual simulation models of multi-dimensional elements such as campus facilities, personnel, and energy consumption, they provide higher-dimensional support for operations and decision-making. Furthermore, the synergy of rule engines and machine learning has shown promising results in intelligent scheduling, security alerts, and energy optimization. However, existing solutions still rely primarily on single algorithms or pre-set processes, lacking a holistic framework based on dynamic demand matching and multi-strategy integration.

[0003] However, existing smart campus management technologies still face the following shortcomings: Multi-source demands are often presented in structured or semi-structured forms, lacking unified semantic annotation and confidence screening, making it difficult to balance data quality and real-time performance. Second, traditional demand-service mapping often relies on preset rules or simple similarity matching, ignoring demand priorities and service capability differences, resulting in inefficient resource allocation. Third, at the event decision-making level, pure rule engines are prone to decision-making blind spots, while pure machine learning models struggle to meet the compliance and explainability requirements of security-sensitive scenarios. These issues directly impact the flexibility, reliability, and scalability of campus management systems, making it difficult to meet dynamic service requests from multiple roles and scenarios. Summary of the Invention

[0004] In view of the above-mentioned problems, the present invention is proposed.

[0005] Therefore, the technical problem solved by the present invention is that the existing intelligent park management technology has problems of inaccurate labels and imprecise scheduling in terms of demand identification and service matching.

[0006] To solve the above technical problems, the present invention provides the following technical solutions: a smart park management method based on the park informationization needs, comprising:

[0007] Obtain demand information of various users and devices from mobile devices and edge gateways;

[0008] Identify demand information and generate corresponding labels. Filter demand information based on label confidence and merge it into standard demand data using the park domain knowledge graph.

[0009] A bipartite graph is constructed based on standard demand data. The matching degree is calculated based on the weighted semantic similarity, demand priority, and module processing capability. The maximum weight matching algorithm is used to determine the call service set.

[0010] The trigger event features are input into the rule engine and AI model respectively, and the AI ​​prediction results are given priority. If the rule confidence is higher than the set threshold, the rule decision is adopted, and the final execution strategy is output by smooth function fusion.

[0011] As a preferred solution of the smart park management method based on the park informationization demand of the present invention, wherein: the identification demand information generating corresponding tags includes: i Converted into a fixed-dimensional vector through the model mapping function Each demand information s i Input the pre-trained BERT-CRF entity recognition model;

[0012] In each sentence i Accurately identify the demand-related entity e and assign labels to the demand-related entity e words; the hidden layer dimension h, the number of label categories C, the ID sequence after input word segmentation Output hidden state sequence h t ;

[0013] Construct a linear mapping, perform CRF decoding, define the transfer matrix, construct a score function, and use the Viterbi algorithm to find the optimal label sequence y t ; Set the confidence threshold, the confidence threshold is ε NER ;

[0014] For each label position t i Calculate the marginal probability P(y t =k|s i ), as the confidence of the label; when the label confidence is higher than the confidence threshold, the recognition result is retained; when the label confidence is higher than the confidence threshold, it is determined to be noise and discarded.

[0015] As a preferred solution of the smart park management method based on park informationization needs of the present invention, wherein: the merging into standard demand data includes comparing the identified keywords with entries in the park domain knowledge graph, and mapping synonyms and near-synonyms to the same standard entry;

[0016] From yt Extract consecutive same-label intervals [t i ,t i+1 ], the corresponding original text substring is an entity e j ; If there are overlapping entities in the same sentence, the longer interval takes precedence; j Enter f KG Perform synonymous merging to obtain standard entities Forming final demand unit where q j Determined by the original data source; each standardized requirement is labeled with high, medium, or low priority based on the score and history of the requirement source.

[0017] As a preferred solution of the smart park management method based on the park informationization demand of the present invention, wherein: the weighted calculation matching degree includes: i ) is used for subsequent similarity matching; standardization requirement entity set {r j}, each containing entity text and numerical priority, expressed as φ(s i ) and {r j} is used as the basis for weighted comparison and bipartite graph construction;

[0018] For each standardized requirement r j , take out its sentence vector φ from the vector storage table j =φ(r j );

[0019] For each service module s k , take out its sentence vector ψ k =ψ(s k );

[0020] Perform L2 normalization on each vector, take all standardized requirements as the left node and callable functional modules as the right node; convert the requirement text and module description into vectors respectively, and perform L2 normalization on each pair of requirement-service (r j ,s k )calculate Use cosine similarity to measure their semantic fit, cosine similarity is the dot product cos j,k =dot j,k ;

[0021] For each pair of standardized requirements and service modules (r j ,s k) Calculate the weighted sum of the matching degrees of the similarity component, priority component, and ability component, map high, medium, and low into numerical values to reflect the priority of the requirements; at the same time, set three weight coefficients to control the contribution ratio of the three items to the final matching degree; in the bipartite graph, perform maximum weight matching, and only retain the edges with a matching degree higher than the minimum screening threshold;

[0022] Output a set of requirements and perform service module mapping to form a set of services to be called.

[0023] As a preferred solution of the intelligent park management method based on park informatization requirements described in the present invention, wherein: the determination of the service set to be called by using the maximum weight matching algorithm includes marking all rows and columns as uncovered at the beginning, and repeating the iterative operation until all zero elements are covered; finding the uncovered zero elements, if the row or column where it is located is not marked, then preferentially mark the row and column where the zero element is located;

[0024] Mark the row or column where the zero element is located as covered, and count the number of lines used L; if L = n, directly construct the matching; if L < n, then the cost matrix needs to be adjusted, and find the minimum element δ = minc at the intersection of all uncovered rows and uncovered columns j,k where j represents uncovered and k represents covered;

[0025] Perform subtraction operations on the elements on all uncovered rows and uncovered columns, c j,k ← c j,k - δ;

[0026] Perform addition operations on the intersections of all covered rows and covered columns, c j,k ← c j,k + δ;

[0027] Perform coverage counting until L = n. In the matrix C, select a set of zero positions of size n {(j, k)}, requiring exactly one zero to be selected in each row and each column;

[0028] Use the Hungarian growth path method. For each row, preferentially select the position with only one zero in that row; mark the row and the corresponding column as matched, and remove the other zeros in that column; repeat until all rows are matched;

[0029] Output the selected matching pairs corresponding to the decision variable x j,k .

[0030] As a preferred solution of the intelligent park management method based on park informatization requirements described in the present invention, wherein: the input of the trigger event characteristics into the rule engine and the AI model respectively includes, together with the service module s k matched in the previous stage and the corresponding requirement r j, demand text vector φ(r j ), service module vector ψ(s k ), numerical priority q j , concatenate the above three parts in sequence to form a column vector x with a length of 2d+1;

[0031] A set of predefined conditions covers high-priority scenarios such as safety, compliance, and emergency dispatch. Given an input x and its corresponding context label, it outputs a list of candidate actions and their corresponding confidence scores {(a1, p1), (a2, p2), …}.

[0032] Take the maximum confidence p max , evaluate the rule priority and preset the rule trigger threshold τ; if p max ≥τ, then directly select the rule action π=a rule Output to the final decision execution process, otherwise, enter the AI ​​prediction fusion output process.

[0033] As a preferred solution of the smart park management method based on the park informationization demand of the present invention, wherein: the outputting the final execution strategy by fusion of smooth functions includes constructing a lightweight multi-layer perceptron, the output dimension of which is equal to the number of action categories C;

[0034] Fusion selects the final action, based on p max Adjust the hard switch π and calculate the fusion weight γ. When p max When it is much higher than τ, the rule is dominant; when p max When it is far lower than τ, AI is mainly used; κ is used to control the steepness of the switching curve, and the final strategy π is sent to the execution unit through the message bus.

[0035] As a preferred solution of the smart park management system based on the park informationization demand of the present invention, it includes: a demand collection module, a semantic standardization module, a service matching module, and a decision fusion execution module;

[0036] The demand collection module is used to collect original demand information from various terminals and edge devices in real time and perform preliminary structuring;

[0037] The semantic standardization module is used to convert the original requirements into standardized requirement entities and assign priority labels;

[0038] The service matching module is used to determine the optimal service call set based on weighted matching of semantic similarity, priority and module capabilities;

[0039] The decision fusion execution module is used to simultaneously input the trigger event features into the rule engine and the AI ​​model, and generate the final execution strategy based on confidence gating and smooth fusion.

[0040] A computer device includes a memory and a processor, wherein the memory stores a computer program, and the processor executes the computer program as a step to implement a smart park management method based on the park's information needs.

[0041] A computer-readable storage medium stores a computer program, which, when executed by a processor, implements the steps of a smart park management method based on the park's information needs.

[0042] Beneficial effects of the present invention: The smart park management method based on the park informationization needs provided by the present invention unifies the corpus form, improves the embedding quality, and significantly reduces the matching rate of "false similarity" and "wrong entity" through text preprocessing and sentence segmentation, denoising, and stop word elimination, ensuring that the short text vector is stable and reliable. Based on the synonymous merging of BERT-CRF and domain knowledge graphs, entity annotations above the confidence threshold can capture the association between professional terms and context, eliminate noise and ambiguity, realize standardized demand units, and bridge the semantic gap between multiple sources. Combining the weighted matrix of semantic similarity, demand priority and module processing capacity, and adopting the maximum weight matching algorithm, it can balance the matching degree, urgency and system carrying capacity, automatically generate the globally optimal service call set, and improve system efficiency and user satisfaction. BRIEF DESCRIPTION OF THE DRAWINGS

[0043] In order to more clearly illustrate the technical solutions of the embodiments of the present invention, the following briefly introduces the drawings required for use in the description of the embodiments. Obviously, the drawings described below are only some embodiments of the present invention. For ordinary technicians in this field, other drawings can be obtained based on these drawings without paying any creative work.

[0044] Figure 1 An overall flow chart of a smart park management method based on the park informationization needs provided in the first embodiment of the present invention. DETAILED DESCRIPTION

[0045] To make the above-mentioned objects, features, and advantages of the present invention more clearly understood, the following detailed description of the specific embodiments of the present invention is given in conjunction with the accompanying drawings. It is obvious that the described embodiments are only part of the embodiments of the present invention, not all of them. Based on the embodiments of the present invention, all other embodiments obtained by ordinary persons in this field without creative work should fall within the scope of protection of the present invention.

[0046] Example 1, reference Figure 1 , which is an embodiment of the present invention, provides a smart park management method based on the park informationization needs, including:

[0047] S1: Obtain the demand information of various users and devices from mobile devices and edge gateways.

[0048] Furthermore, preprocess the original text, and input the collection D raw ={d1, d2, …, d N} of the original demand text collected from multiple sources, perform cleaning to remove HTML tags, URLs, and special symbols; segment long paragraphs into sentences according to punctuation marks. Remove stop words. Use the Chinese stop word list to remove high-frequency non-informative words such as "de", "le", "he", etc.

[0049] Output the processed sentence list {s1, s2, …, s M}.

[0050] Sentence vectorization maps each sentence s i to a fixed-dimensional vector for easy similarity calculation. d = 768: the output dimension of SBERT, BatchSize B = 32. MaxLen L max = 128 (truncate or divide very long sentences into chunks).

[0051] After segmenting s i into words, convert it into an ID sequence where T i ≤ L max .

[0052] It should be noted that the quality of the corpus is significantly improved, reducing the probability of "false similarity" or "wrong entity" matching; short texts at the single-sentence level are easier to be embedded into the model to generate stable vector representations, avoiding semantic ambiguity caused by the mixing of multiple topics in long paragraphs; the input scale is controllable after data cleaning, providing guarantee for the subsequent SBERT algorithm to maintain high throughput under fixed BatchSize and MaxLen.

[0053] S2: Identify the demand information to generate corresponding tags, screen the demand information according to the tag confidence, and merge it into standard demand data with the help of the knowledge graph in the park domain.

[0054] Furthermore, input the pre-trained SBERT model with Embedding to obtain the last-layer hidden state H = [h1, ..., hTi], where each ht ∈ Rd.

[0055] Vector aggregation, perform meanpooling on H:

[0056]

[0057] Output the set of sentence vectors For each sentence s iAccurately identify relevant entities and assign labels. Hidden layer dimension h = 768, number of label categories C, confidence threshold εNER = 0.90, input Output hidden state sequence {ht∈Rh}.

[0058] Linear mapping:

[0059]

[0060] CRF decoding, defining the transfer matrix The scoring function formula is expressed as:

[0061]

[0062] Finding the optimal tag sequence using Viterbi algorithm Confidence calculation, for each position t, calculate the edge probability

[0063]

[0064] Where is the edge score obtained by the CRF forward-backward algorithm. Only those annotations are retained as entity boundaries.

[0065] Output standardized demand fragments. Extract continuous same-label intervals [t i ,t i+1 ], the corresponding original text substring is an entity e j If there are overlapping entities in the same sentence, the longer interval takes precedence. Synonymize each input and get the standard entity The formula for forming the final demand unit is expressed as:

[0066]

[0067] where q j It can be determined by the original data source (such as "operation and maintenance work order" is given high priority).

[0068] Sentence vector φ(s i ) is used for subsequent similarity matching; standardization requirement entity set {r j}, each containing entity text and numerical priority. In this way, the subsequent "matching service" can directly use φ(s i ) and {r j} is used as the basis for weighted comparison and bipartite graph construction.

[0069] The identification of demand information to generate corresponding tags includes: i Converted into a fixed-dimensional vector through the model mapping function Each demand information s iInput the pre-trained BERT-CRF entity recognition model.

[0070] In each sentence i Accurately identify the demand-related entity e and assign labels to the demand-related entity e words; hidden layer dimension h, number of label categories C, input ID sequence after word segmentation Output hidden state sequence h t .

[0071] Construct a linear mapping, perform CRF decoding, define the transfer matrix, construct a score function, and use the Viterbi algorithm to find the optimal label sequence y t ; Set the confidence threshold, the confidence threshold is ε NER .

[0072] For each label position t i Calculate the marginal probability P(y t =k|s i ), as the confidence of the label; when the label confidence is higher than the confidence threshold, the recognition result is retained; when the label confidence is higher than the confidence threshold, it is determined to be noise and discarded.

[0073] It should be noted that the requirement entities in the sentence are accurately identified (a confidence threshold of 0.90 ensures reliability), and then synonyms are merged through the park domain knowledge graph to form standardized requirement units. Model-level entity extraction captures complex professional terms and contextual associations, avoiding missed detections that are difficult to cover with manual rules. Setting a high confidence threshold eliminates weakly correlated noise to ensure the accuracy of standard requirements. Knowledge graph mapping further eliminates multiple expressions of the same concept, bridging the semantic gap between work orders, questionnaires, and perception data from different sources, laying a solid foundation for unified scheduling.

[0074] S3: Construct a bipartite graph based on standard demand data, calculate the matching degree based on the weighted semantic similarity, demand priority and module processing capability, and use the maximum weight matching algorithm to determine the call service set.

[0075] Furthermore, the merging into standard demand data includes comparing the identified keywords with entries in the park domain knowledge graph, and mapping synonyms and antonyms to the same standard entry.

[0076] From y t Extract consecutive same-label intervals [t i ,t i+1 ], the corresponding original text substring is an entity e j If there are overlapping entities in the same sentence, the longer interval takes precedence; j Enter f KG Perform synonymous merging to obtain standard entities Forming final demand unit where q j Determined by the original data source; each standardized requirement is labeled with high, medium, or low priority based on the score and history of the requirement source.

[0077] The weighted calculation of matching degree includes: i ) is used for subsequent similarity matching; standardization requirement entity set {r j}, each containing entity text and numerical priority, expressed as φ(s i ) and {r j} is used as the basis for weighted comparison and bipartite graph construction.

[0078] For each standardized requirement r j , take out its sentence vector φ from the vector storage table j =φ(r j ).

[0079] For each service module s k , take out its sentence vector ψ k =ψ(s k ).

[0080] Perform L2 normalization on each vector, take all standardized requirements as the left node and callable functional modules as the right node; convert the requirement text and module description into vectors respectively, and perform L2 normalization on each pair of requirement-service (r j ,s k )calculate Use cosine similarity to measure their semantic fit, cosine similarity is the dot product cos j,k =dot j,k .

[0081] The triplet is determined in advance in the system configuration:

[0082] λ1, λ2, λ3≥0, λ1+λ2+λ3=1

[0083] λ1 = 0.6 (largest similarity), λ2 = 0.3, λ3 = 0.1. Calculate the matching weight by performing a weighted sum of the three components for each pair. Similarity component:

[0084] A j,k =λ1×cos j,k .

[0085] Priority component:

[0086] B j,k =λ2×q j .

[0087] Ability component:

[0088] C j,k =λ3×cap k .

[0089] Total weight:

[0090] w j,k =A j,k +B j,k +C j,k .

[0091] If cos j,k =0.8,q j =1.0,cap k =0.7, and take λ1=0.6, λ2=0.3, λ3=0.1.

[0092] Threshold setting, the system administrator sets the minimum screening threshold θ in the configuration match , requiring that only candidate edges with weights not less than this value be retained.

[0093] The specific executable process of the Hungarian algorithm in solving the maximum weight matching problem converts the weight matrix into a "minimize cost" form that can be used by the Hungarian method, and finally obtains a unique matching pair {(j,k)|x j,k =1}. Assume that the number of demands and the number of services are both n (if they are not equal, zeros can be added to the larger one):

[0094] Weight matrix:

[0095] W=[w j,k ] j＝1..n,k＝1..n

[0096] Take the maximum weight:

[0097]

[0098] Converted to cost, the formula is:

[0099] C=[c j,k ],c j,k =w max -w j,k

[0100] In this way, the higher the weight, the lower the cost, which is easier to minimize. j,k <θ match , then discard the pair; otherwise, record it in the candidate list.

[0101] All retained ones are formed into a sparse weight matrix for use by the "maximum matching" algorithm.

[0102] Through the above steps, accurate calculation is performed to obtain a matching weight in the interval [0,1] for each pair of demand and service module;i , θ match Flexibly adjust the matching preference. Each step corresponds to the system configuration or the model output, without code dependencies, and only requires common vector operations and threshold comparisons.

[0103] For each pair of standardized requirements and service modules (r j , s k ), calculate the weighted sum of the matching degrees of the similarity component, the priority component, and the capability component, map high, medium, and low to numerical values to reflect the priority of the requirements; at the same time, set three weight coefficients to control the contribution ratio of the three items to the final matching degree; in the bipartite graph, perform the maximum weight matching and only retain the edges with a matching degree higher than the minimum screening threshold.

[0104] Output a set of requirements and perform service module mapping to form a set of services to be invoked.

[0105] The determination of the service set to be invoked by using the maximum weight matching algorithm includes: initially marking all rows and columns as uncovered, and repeating the iterative operation until all zero elements are covered; finding the uncovered zero elements, and if the row or column where it is located is not marked, preferentially mark the row and column where the zero element is located.

[0106] Mark the row or column where the zero element is located as covered, and count the number of lines used L; if L = n, directly construct the matching; if L < n, then the cost matrix needs to be adjusted, and find the minimum element δ = minc j,k , where j represents uncovered and k represents covered.

[0107] Perform subtraction operations on all elements in the uncovered rows and uncovered columns, c j,k ← c j,k - δ.

[0108] Perform addition operations on all elements at the intersections of the covered rows and covered columns, c j,k ← c j,k + δ.

[0109] Preliminary normalization of rows and columns, row subtraction. For each row j, find the minimum cost of the row, which is expressed by the formula:

[0110]

[0111] And let:

[0112]

[0113] Column subtraction. For each column k after normalization, find the minimum cost:

[0114]

[0115] And order:

[0116]

[0117] At this time, each row and column has at least one zero element. Perform cover counting until L=n and select independent zeros.

[0118] In the matrix, select a set of zero locations {(j,k)} of size , requiring that exactly one zero is selected in each row and column.

[0119] The Hungarian growth path method is usually used. For each row, the position with only one zero in the row is preferentially selected; the row is marked as matched with the corresponding column, and the other zeros in the column are removed.

[0120] Repeat until all rows are matched. The output matching pair formula is expressed as:

[0121]

[0122] The corresponding decision variable setting formula is expressed as:

[0123]

[0124] It should be noted that a "globally optimal" match is constructed by weighting the sentence vector, requirement priority, and module capability score for each standardized requirement and each available service module. This is then distributed fairly using a maximum-weight matching algorithm, which considers both semantic fit and business priorities and system processing capabilities. The algorithm automatically strikes a balance between compatibility, urgency, and resource capacity, avoiding the unfairness associated with "first-come, first-served" or "human-first" policies. Maximum-weight matching ensures a globally optimal solution for the entire call set, improving system efficiency and user satisfaction. Configurable similarity thresholds and weight coefficients allow administrators to adjust matching preferences at any time based on campus load and policies, enabling flexible, on-demand scheduling.

[0125] S4: Input the trigger event features into the rule engine and AI model respectively, give priority to the AI ​​prediction results, and adopt the rule decision if the rule confidence is higher than the set threshold. The final execution strategy is output by smooth function fusion.

[0126] Furthermore, the trigger event features are input into the rule engine and AI model respectively, including the service module s matched in the previous stage. k and corresponding demand r j , demand text vector φ(r j ), service module vector ψ(s k ), numerical priority q j , concatenate the above three parts in sequence to form a column vector x with a length of 2d+1.

[0127] A set of predefined conditions covers high-priority scenarios such as safety, compliance, and emergency dispatch. Given an input x and its corresponding context label, the system outputs a list of candidate actions and their corresponding confidence levels {(a1, p1), (a2, p2), …}.

[0128] Take the maximum confidence p max , evaluate the rule priority and preset the rule trigger threshold τ; if p max ≥τ, then directly select the rule action π=a rule Output to the final decision execution process, otherwise, enter the AI ​​prediction fusion output process.

[0129] The engine type is a business rule management system; the rule base has a predefined set of "IF-THEN" conditions, covering high-priority scenarios such as security, compliance, and emergency dispatch.

[0130] Input x and its corresponding context label (such as "night mode", "high temperature warning")

[0131] Output a set of candidate actions and corresponding confidence lists:

[0132] {(a1,p1),(a2,p2),…},p i ∈[0,1]

[0133] Post-processing, take the maximum confidence, the formula is expressed as:

[0134]

[0135] Evaluate rule priority and preset rule trigger threshold τ (e.g. 0.8)

[0136] Determine if p max ≥τ, then directly select the rule action π=a rule and skip to the final policy execution step.

[0137] Otherwise, the AI ​​prediction process begins. The AI ​​model predicts the action, using a lightweight multi-layer perceptron (MLP) as the model type.

[0138] Compute the hidden representation:

[0139] h hid =σ(W1x+b1)

[0140] Calculate the output score:

[0141] z=W2h hid +b2

[0142] Convert to probability distribution:

[0143]

[0144] The action formula corresponding to the maximum probability is expressed as:

[0145]

[0146] Fusion or selection of the final action, the hard switching formula is expressed as:

[0147]

[0148] Smooth fusion, calculate fusion weight:

[0149]

[0150] Combining rules and AI actions, the final decision formula is output as:

[0151] π=γ×a rule +(1-γ)×a AI

[0152] The final execution strategy of outputting the fusion output according to the smooth function includes constructing a lightweight multi-layer perceptron, and the output dimension is equal to the number C of action categories.

[0153] Fusion selects the final action, based on p max Adjust the hard switch π and calculate the fusion weight γ. When p max When it is much higher than τ, the rule is dominant; when p max When it is far lower than τ, AI is mainly used; κ is used to control the steepness of the switching curve, and the final strategy π is sent to the execution unit through the message bus.

[0154] It should be noted that in high-risk or compliance scenarios, interpretable business rules (confidence threshold τ) are used to make rapid decisions. When rules are insufficient or confidence is low, lightweight MLP models are used for prediction. Finally, a smoothing function is used to balance the two, ensuring both transparency and adaptability. Decisions in critical scenarios can be fully traced back to predefined rules, meeting audit and compliance requirements. The AI ​​model has greater generalization capabilities for new scenarios, enabling timely responses to complex and changing campus operational needs. Smooth integration ensures smooth switching without jitter, achieving an optimal compromise between system stability and flexibility.

[0155] Example 2 is an embodiment of the present invention, which provides a smart park management method based on the park's information needs. In order to verify the beneficial effects of the present invention, scientific demonstration is carried out through economic benefit calculation and simulation experiments.

[0156] First, for the multi-source demand information processing and decision-making within the park, three scenarios were selected as experimental areas: Office Building A, Research Building B, and Dormitory Building C. The following preparations were required in the early stages of the experiment:

[0157] Data collection preparation: Deploy a unified collection agent in the mobile application and the edge gateway to ensure coverage of three types of demand sources, namely employee repair requests, visitor appointments, and environmental sensing. Each type of source generates approximately 1000 original demand texts per week, including redundant information such as HTML tags, URL links, and special symbols.

[0158] Preprocessing and annotation toolchain construction: Build a text cleaning pipeline based on OpenNLP to perform regular expression removal of HTML and URLs; split long paragraphs by "。!?;"; use an industry stop word list to remove invalid words such as "的、了、和"; call the SBERT model (BatchSize = 32, MaxLen = 128) to complete sentence vectorization and save it to the vector library.

[0159] Integration of knowledge graph and CRF model: Import a self-built KG in the park domain, including more than 5000 entities such as equipment, positions, and services; load a pre-trained BERT-CRF entity recognition model (hidden layer dimension 768, 8 label classes), and set the confidence threshold to 0.90 to accurately extract demand keywords and map them to KG standard entities.

[0160] Bipartite graph and matching configuration: Register 6 service modules that can be called in the system, including air conditioning scheduling, lighting control, security inspection, visitor management, energy consumption analysis, and equipment maintenance. Generate a description vector (with the same dimension as the demand, 768) for each module and set the ability score cap ∈ [0.6, 0.9]. Match the weight coefficient, and the minimum retention threshold θ_match = 0.25.

[0161] The specific experimental process is as follows: (1) S1 preprocessing: Clean and split sentences for 3000 collected original texts, and finally obtain 2800 cleaned short sentences; the average text length at the short sentence level is 35 characters, avoiding interference from multiple topics, and the model throughput is increased from the previous 300 sentences / second to 650 sentences / second. (2) S2 entity recognition and merging: Call BERT-CRF for 2800 short sentences, identify 3200 entities, and retain 2750 entity fragments after excluding candidates with confidence <P = 0.90; after mapping to KG, 2500 standardized demand units are formed, including classifications such as system log services, temporary security patrols, and face recognition door opening. (3) S3 matching and scheduling: Construct a bipartite graph for 2500 standardized demands and 6 service modules, calculate the weighted scores of cosine(demand, module), priority q_j, and cap; after screening w_{j,k} ≥ 0.25, a total of 3200 edges are retained, and the Hungarian algorithm is used to complete the maximum weight matching on a 6×6 matrix (padding with zeros if insufficient), and finally 1020 service call tasks are scheduled and allocated.

[0162] Through the above process, a closed-loop connection from demand to service is achieved, ensuring that different demand sources can obtain corresponding module execution accurately and efficiently in a distributed environment.

[0163] Comparing four existing control solutions, the number of valid sentences after cleaning was less than 2200, the entity recognition accuracy was below 85%, and the matching accuracy did not exceed 80%. In contrast, in each scenario of the embodiment, the number of valid sentences after cleaning was close to 2800, and the number of incoming entities was greater and more complete, indicating that the effective regularization and sentence segmentation strategies in the preprocessing stage significantly improved the short text coverage. BERT-CRF entity recognition accuracy jumped to over 91%, an increase of approximately 6 percentage points over the highest control value (85.1%), demonstrating that the high confidence threshold (0.90) and the park domain KG merging strategy can significantly reduce false positives and missed detections.

[0164] In the matching module phase, existing solutions achieved matching accuracy rates between 75% and 78%, while the implementation achieved nearly 90%, highlighting the rationale behind the ternary fusion strategy of cosine similarity, priority, and capability weighting. The weight distribution of λ1:λ2:λ3 = 0.6:0.2:0.2 ensures the dominance of semantic fit while also balancing business urgency and system capacity, achieving a balance between overall efficiency and fairness. The control solution averaged around 120ms in rule decision response time, while the implementation achieved a minimum of 92ms in scenario 3, with an average of less than 100ms. This represents a 20% improvement in response speed, meeting the real-time requirements of the park's real-time scheduling.

[0165] The AI ​​decision-making accuracy of the implementation example exceeded 87%, while the control solutions were all below 72%. This demonstrates that lightweight MLP combined with high-quality feature inputs can achieve stronger generalization and judgment capabilities in complex scenarios. In terms of system throughput, the implementation example can maintain a processing capacity of 630–660 items per second, an increase of over 100% compared to the control solution's peak of 320 items per second, meeting the requirements of large-scale concurrent processing.

[0166] Example 3 is an embodiment of the present invention, which provides a smart park management system based on the park's information needs, including a demand collection module, a semantic standardization module, a service matching module, and a decision fusion execution module.

[0167] The demand collection module is used to collect original demand information from various terminals and edge devices in real time and perform preliminary structuring.

[0168] The semantic standardization module is used to convert the original requirements into standardized requirement entities and assign priority labels.

[0169] The service matching module is used to determine the optimal service call set based on weighted matching of semantic similarity, priority and module capabilities.

[0170] The decision fusion execution module is used to simultaneously input the trigger event features into the rule engine and the AI ​​model, and generate the final execution strategy based on confidence gating and smooth fusion.

[0171] If the function is implemented in the form of a software functional unit and sold or used as an independent product, it can be stored in a computer-readable storage medium. Based on this understanding, the technical solution of the present invention, or the part that contributes to the prior art, or the part of the technical solution, can be embodied in the form of a software product. The computer software product is stored in a storage medium and includes several instructions for enabling a computer device (which can be a personal computer, server, or network device, etc.) to execute all or part of the steps of the various embodiments of the present invention. The aforementioned storage medium includes: U disk, mobile hard disk, read-only memory (ROM, Read-Only Memory), random access memory (RAM, Random Access Memory), disk or optical disk, and other media that can store program code.

[0172] The logic and / or steps represented in the flowcharts or otherwise described herein, for example, can be considered as an ordered list of executable instructions for implementing the logical functions, and can be embodied in any computer-readable medium for use by, or in conjunction with, an instruction execution system, apparatus, or device (e.g., a computer-based system, a system including a processor, or other system that can fetch and execute instructions from an instruction execution system, apparatus, or device). For purposes of this specification, a "computer-readable medium" can be any device that can contain, store, communicate, propagate, or transport a program for use by, or in conjunction with, an instruction execution system, apparatus, or device.

[0173] More specific examples (a non-exhaustive list) of computer-readable media include the following: an electrical connection with one or more wires (electronic devices), a portable computer disk cartridge (magnetic devices), a random access memory (RAM), a read-only memory (ROM), an erasable and programmable read-only memory (EPROM or flash memory), a fiber optic device, and a portable compact disc read-only memory (CDROM). In addition, the computer-readable medium may even be paper or other suitable medium on which the program is printed, since the program may be obtained electronically, for example, by optically scanning the paper or other medium, followed by editing, deciphering, or processing in another suitable manner as necessary, and then stored in a computer memory.

[0174] It should be understood that various parts of the present invention can be implemented using hardware, software, firmware, or a combination thereof. In the above-described embodiments, multiple steps or methods can be implemented using software or firmware stored in a memory and executed by a suitable instruction execution system. For example, if implemented using hardware, as in another embodiment, any one of the following technologies known in the art or a combination thereof can be used: a discrete logic circuit having logic gate circuits for implementing logical functions on data signals, an application-specific integrated circuit having suitable combinational logic gate circuits, a programmable gate array (PGA), a field-programmable gate array (FPGA), etc. It should be noted that the above embodiments are merely illustrative of the technical solutions of the present invention and are not intended to be limiting. Although the present invention has been described in detail with reference to preferred embodiments, those skilled in the art should understand that the technical solutions of the present invention may be modified or replaced with equivalents without departing from the spirit and scope of the technical solutions of the present invention, and all such modifications should be encompassed by the claims of the present invention.

[0175] It should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention and are not intended to limit the present invention. Although the present invention has been described in detail with reference to the preferred embodiments, those skilled in the art should understand that the technical solutions of the present invention may be modified or replaced by equivalents without departing from the spirit and scope of the technical solutions of the present invention, which should all be included in the scope of the claims of the present invention.

Claims

1. A smart park management method based on the park's information needs is characterized by: include: Obtain demand information of various users and devices from mobile devices and edge gateways; Identify demand information and generate corresponding labels. Filter demand information based on label confidence and merge it into standard demand data using the park domain knowledge graph. A bipartite graph is constructed based on standard demand data. The matching degree is calculated based on the weighted semantic similarity, demand priority, and module processing capability. The maximum weight matching algorithm is used to determine the call service set. The trigger event features are input into the rule engine and AI model respectively, and the AI ​​prediction results are given priority. If the rule confidence is higher than the set threshold, the rule decision is adopted, and the final execution strategy is output by smooth function fusion.

2. The smart park management method based on the park informationization needs according to claim 1, characterized in that: The identification of demand information to generate corresponding tags includes: i Converted into a fixed-dimensional vector through the model mapping function Each demand information s i Input the pre-trained BERT-CRF entity recognition model; In each sentence i Accurately identify the demand-related entity e and assign labels to the demand-related entity e words; hidden layer dimension h, number of label categories C, input ID sequence after word segmentation Output hidden state sequence h t ; Where d represents the dimension of the embedding vector; Construct a linear mapping, perform CRF decoding, define the transfer matrix, construct a score function, and use the Viterbi algorithm to find the optimal label sequence y t ; Set the confidence threshold, the confidence threshold is ε NER ; For each label position t i Calculate the marginal probability P(y t =k|s i ), as the confidence of the label; when the label confidence is higher than the confidence threshold, the recognition result is retained; when the label confidence is higher than the confidence threshold, it is determined to be noise and discarded.

3. The smart park management method based on the park informationization needs according to claim 2, characterized in that: The merging into standard demand data includes comparing the identified keywords with entries in the park domain knowledge graph and mapping synonyms and near-synonyms to the same standard entry; From y t Extract consecutive same-label intervals [t i ,t i+1 ], the corresponding original text substring is an entity e j ; If there are overlapping entities in the same sentence, the longer interval takes precedence; j Input synonym mapping function f KG Perform synonymous merging to obtain standard entities Forming final demand unit where q j Determined by the original data source, j represents the demand index; based on the score and historical records of the demand source, each standardized demand is labeled with high, medium, and low priority.

4. The smart park management method based on the park informationization needs according to claim 3, characterized in that: The weighted calculation of matching degree includes: i ) is used for subsequent similarity matching; Standardized requirement entity set {r j }, each containing entity text and numerical priority, expressed as φ(s i ) and {r j } is used as the basis for weighted comparison and bipartite graph construction; For each standardized requirement r j , take out its sentence vector φ from the vector storage table j =φ(r j ); For each service module s k , take out its sentence vector ψ k =ψ(s k ), k represents the service module index; Perform L2 normalization on each vector, with all normalization requirements as the left nodes and the callable function modules as the right nodes; Convert the requirement text and module description into vectors respectively, and for each pair of requirement-service (r j ,s k )calculate Use cosine similarity to measure their semantic fit, cosine similarity is the dot product cos j,k =dot j,k ; represents the corresponding unit vector; For each pair of standardized requirements and service modules (r j ,s k ), calculate the weighted sum of the similarity component, priority component, and capability component, and map high, medium, and low to numerical values ​​to reflect the priority of the requirements; at the same time, set three weight coefficients to control the contribution ratio of the three items to the final matching degree; in the bipartite graph, perform maximum weight matching and only retain edges with matching degrees higher than the minimum screening threshold; Output a set of requirements and perform service module mapping to form a set of services that need to be called.

5. The smart park management method based on the park informationization needs according to claim 4, characterized in that: The method of using the maximum weight matching algorithm to determine the call service set includes marking all rows and columns as uncovered initially, repeating the iterative operation until all zero elements are covered; finding an uncovered zero element, and if the row or column where the zero element is located is not marked, then marking the row and column where the zero element is located first; Mark the row or column where the zero element is located as covered, and count the number of lines L used; if L = n, directly construct the final match; if L < n, then the cost matrix C needs to be adjusted to find the smallest element δ = minc at the intersection of all uncovered rows and uncovered columns j,k ; Perform subtraction on all uncovered rows and columns, c j,k ←c j,k -δ; Perform addition operations on all intersections of covered rows and covered columns, c j,k ←c j,k +δ; Perform coverage counting until L=n; To construct the final matching, in the matrix C, select a set of n zero positions {(j, k)}, requiring that exactly one zero be selected in each row and column. For each row, use the Hungarian growth path method to preferentially select the position with only one zero in that row. Mark the row and the corresponding column as matched, and remove all other zeros in the column. Repeat until all rows are matched. Output the selected matching pairs Corresponding to the final output matching service subset S * .

6. The smart park management method based on the park informationization needs according to claim 5, characterized in that: The trigger event features are input into the rule engine and AI model respectively, including the service module s matched in the previous stage. k and corresponding demand r j , demand text vector φ(r j ), service module vector ψ(s k ), numerical priority q j , concatenate the three parts in sequence to form a column vector x with a length of 2d+1; A set of predefined conditions covers high-priority scenarios such as safety, compliance, and emergency dispatch. It takes as input x and its corresponding context label and outputs a list of candidate actions and their corresponding confidence levels. Take the maximum confidence p max , evaluate the rule priority and preset the rule trigger threshold τ; if p max ≥τ, then directly select the rule action π=a rule Output to the final decision execution process, otherwise, enter the AI ​​prediction fusion output process.

7. The smart park management method based on the park informationization needs according to claim 6, characterized in that: The final execution strategy of the fusion output according to the smooth function includes constructing a lightweight multi-layer perceptron with an output dimension equal to the number of action categories; Fusion selects the final action, according to p max Adjust the hard switch π and calculate the fusion weight γ. When p max When it is much higher than τ, the rule is dominant; when p max When it is far lower than τ, AI is mainly used; κ is used to control the steepness of the switching curve, and the final strategy π is sent to the execution unit through the message bus.

8. A system using the smart park management method based on park informationization needs as described in any one of claims 1 to 7, characterized in that: It includes demand collection module, semantic standardization module, service matching module, and decision fusion execution module; The demand collection module is used to collect original demand information from various terminals and edge devices in real time and perform preliminary structuring; The semantic standardization module is used to convert the original requirements into standardized requirement entities and assign priority labels; The service matching module is used to determine the optimal service call set based on semantic weighted matching; The decision fusion execution module is used to simultaneously input the trigger event features into the rule engine and the AI ​​model, and generate the final execution strategy based on confidence gating and smooth fusion.

9. A computer device comprising a memory and a processor, wherein the memory stores a computer program, wherein: When the processor executes the computer program, the steps of the smart park management method based on the park informationization needs described in any one of claims 1 to 7 are implemented.

10. A computer-readable storage medium having a computer program stored thereon, characterized in that: When the computer program is executed by the processor, the steps of the smart park management method based on the park informationization needs described in any one of claims 1 to 7 are implemented.

Images