A Smart Park Management Method and System Based on Multi-System Integration

Abstract

The present invention discloses a smart park management method and system based on multi-system integration, which relates to the technical field of smart parks. When the actual behavior of a certain subsystem deviates from the normal model, it is determined as a problematic subsystem, providing a basis for subsequent abnormal propagation analysis. After detecting the problematic subsystem, a problem event matrix is constructed to depict the relationship of abnormal occurrences among subsystems. By analyzing the matrix, the foresight ability of each subsystem can be calculated, that is, to determine whether the abnormality of a certain subsystem can predict the abnormality of other subsystems in advance. The recognition value is calculated through statistical methods. If it is significantly higher than the threshold, it indicates that this subsystem may be a key link in abnormal propagation. Based on the calculated recognition value, the system further identifies the abnormal propagation path, thereby constructing a data-driven optimization model. This model adjusts the control parameters of the subsystem through reinforcement learning to minimize the overall abnormal impact.

Description

Technical Field

[0001] The present invention relates to the technical field of smart campuses, and particularly to a smart campus management method and system based on multi-system integration. Background Art

[0002] With the development of smart cities and digital campus management, the management of smart campuses based on multi-system integration has gradually become an important trend in campus operation. Smart campus management involves multiple subsystems such as energy management, security monitoring, environmental monitoring, facility operation and maintenance, smart office, and parking management. While these subsystems operate independently, there are complex interaction relationships among them. In traditional campus management, each subsystem is often managed separately, lacking systematic global coordination and data sharing, resulting in limited intelligent level of campus management. When an abnormality occurs in a certain subsystem, due to the data barriers and the fragmentation of management methods among systems, the operating states of other relevant subsystems may not be analyzed in real time, affecting the overall management efficiency and stability.

[0003] In the current smart campus management system, due to the lack of intelligent analysis ability for multi-system integration, the abnormality management of each subsystem relies on single-point monitoring and isolated analysis. Once an abnormality occurs in a certain subsystem, its impact may gradually spread to other subsystems, but the existing methods often have difficulty in accurately identifying the abnormal subsystem, resulting in the lag of abnormality handling and the problem of local optimization. Summary of the Invention

[0004] In view of the above problems existing in the prior art, the present application provides a smart campus management method and system based on multi-system integration.

[0005] On the one hand, embodiments of the present disclosure provide a smart campus management method and system based on multi-system integration, including the following steps:

[0006] S1: Using sensor readings, collect the operation data of each subsystem in the campus in real time. After data preprocessing, construct the feature vector X of each subsystem, and combine with the historical data set to construct a normal behavior model , to identify the problem subsystem;

[0007] S2: Construct a problem event matrix E, and based on the problem event matrix E, analyze the foreseeability existing among each subsystem. After verification, obtain the identification value ZS;

[0008] S3: Based on the identification value ZS, identify the abnormal propagation path, and construct a data-driven optimization model to minimize the impact of system abnormalities.

[0009] Optionally, the specific steps of S1 include:

[0010] S11: The subsystems within the park include an energy management subsystem, a security monitoring and security protection subsystem, an environmental monitoring subsystem, a vehicle and intelligent parking management subsystem, a facility operation and maintenance and equipment management subsystem, an intelligent office and personnel management subsystem, and a data analysis and AI decision support subsystem. After cleaning, removing outliers, and standardizing the operation data of each subsystem within the park, a feature vector X of each subsystem is generated. The expression of the feature vector X of each subsystem is: , where n is the number of subsystems, are respectively the feature vector of the first subsystem, the feature vector of the second subsystem,..., the feature vector of the nth subsystem.

[0011] Optionally, the specific steps of S1 further include:

[0012] S12: Based on the real-time acquisition of the operation data of each subsystem within the park in S11, the operation data of each subsystem in the past day is stored in a NoSQL database, and the data in the NoSQL database is used as a historical data group;

[0013] S13: Extract the historical data of each subsystem in the normal operation state from the historical data group to generate a normal feature vector Y. The expression of the normal feature vector Y is: , where N is the number of feature components, i is the subsystem number, t is the historical time number, are respectively the first feature component in the normal feature vector of the ith subsystem at the tth historical time, the second feature component in the normal feature vector of the ith subsystem at the tth historical time,..., the Nth feature component in the normal feature vector of the ith subsystem at the tth historical time;

[0014] S14: Use machine learning techniques and train the model using the normal feature vector Y to construct a normal behavior model , specifically:

[0015] ;

[0016] In the formula, is the output value of the normal behavior model constructed according to the normal feature vector of the ith subsystem, is the intercept term, , ,..., are the regression coefficients, , ,..., are the respective feature components in the normal feature vector of the ith subsystem.

[0017] Optionally, the specific steps of S1 further include:

[0018] S15: Based on the historical data set and the normal behavior model , calculate and obtain the residuals , specifically: ; where is the actual value, and an abnormal critical value V is preset in advance, and by comparing the residual with the abnormal critical value V to identify the problem subsystem, specifically: if the absolute value of the residual > the abnormal critical value V, it is determined that the corresponding subsystem is abnormal, and at this time, the corresponding subsystem is marked as the problem subsystem.

[0019] Optionally, the specific steps of S2 include:

[0020] S21: Based on the method of the problem subsystem determined in S15, obtain several groups of problem subsystems corresponding to historical moments, and construct several groups of problem event matrices E for historical moments according to the problem subsystems corresponding to several groups of historical moments. Each row and each column of the problem event matrix E represents a subsystem; fill the problem event matrix E through the problem subsystems corresponding to several groups of historical moments, specifically: if the subsystems are not all marked as problem subsystems at the same moment, the value at the corresponding position in the problem event matrix E is 1, otherwise, the value is 0;

[0021] S22: Extract features from the positions corresponding to the value 1 in the problem event matrix E to obtain several groups of pairs of anticipations to be analyzed.

[0022] Optionally, the specific steps of S2 also include:

[0023] S23: Mark the problem subsystems in several groups of pairs of anticipations to be analyzed as subsystems to be analyzed, and obtain several groups of normal behavior models of the subsystems to be analyzed at different historical moments by the method of obtaining the normal behavior model in S14 , and in combination with the pairs of anticipations to be analyzed, expand the two groups of subsystems to be analyzed involved in the pairs of anticipations to be analyzed through the corresponding normal behavior models to obtain the expanded model , specifically: :

[0024] ;

[0025] In the formula is an autoregressive model that only includes its own lag terms;

[0026] ;

[0027] In the formula is the output value of the extended model constructed based on the normal eigenvector of the i-th subsystem at the t-th historical moment, is the intercept term, , , and are all regression coefficients, ,..., are respectively the output values of the normal behavior model constructed based on the normal eigenvectors of the i-th subsystem at the (t - 1)-th historical moment,..., the (t - q)-th historical moment; is the error term, q is the lag order of, p is the lag order of, ,..., are respectively the output values of the normal behavior model constructed based on the normal eigenvectors of the h-th subsystem at the (t - 1)-th historical moment,..., the (t - p)-th historical moment; h and i are both the numbers of subsystems.

[0028] Optionally, the specific steps of S2 further include:

[0029] S24: By comparing the sum of squared residuals of the autoregressive model that only contains its own lag terms and the output value of the extended model constructed based on the normal eigenvector of the i-th subsystem at the t-th historical moment, determine whether adding the lag term of improves the prediction ability of , specifically:

[0030] ;

[0031] In the formula, ZS is the identification value, represents the sum of squared residuals of the autoregressive model that only contains its own lag terms , represents the sum of squared residuals of the output value of the extended model constructed based on the normal eigenvector of the i-th subsystem at the t-th historical moment, T represents the number of samples;

[0032] S25: Preset an identification threshold. By comparing it with the identification value ZS, if the identification value ZS > the identification threshold, it means that adding the lag term of improves the prediction ability of , and mark the corresponding pair to be analyzed as a foresight pair and extract it to obtain a feature set.

[0033] Optionally, S3: The specific steps include:

[0034] S31: Based on the extraction set, reconstruct and populate a set of problem event matrices E. Each row and each column of the reconstructed problem event matrix E represents a subsystem to be analyzed. The specific content to be populated is as follows: Search for each pair of foresights in the reconstructed problem event matrix E and mark the corresponding positions with the value 1, and the values of the remaining positions are 0.

[0035] S32: Based on the content of S31, using the subsystems to be analyzed as nodes, if the position in the reconstructed problem event matrix E is , then draw a directed edge from to . After traversing, generate a set of directed graphs. Among them, and are different subsystems to be analyzed, and s and c are the numbers of the subsystems to be analyzed.

[0036] S33: According to the directed graph, identify the abnormal propagation paths of the corresponding subsystems to be analyzed and determine the abnormal source. According to the abnormal propagation paths of the corresponding subsystems to be analyzed, count the abnormal propagation time delays of each propagation segment in the abnormal propagation paths of the corresponding subsystems to be analyzed, and extract the minimum abnormal propagation time delay of each propagation segment.

[0037] Optionally, the specific steps of S3 include:

[0038] S34: Based on the minimum abnormal propagation time delay of each propagation segment, determine that when has an abnormality, predict the cascading abnormalities in each system, so as to count the probability that appears abnormal within the minimum abnormal propagation time of the corresponding propagation segment after has an abnormality .

[0039] S35: Use reinforcement learning to train and optimize the strategy based on historical data to construct a data-driven optimization model. Specifically:

[0040] ;

[0041] In the formula, L is the optimization objective function, is the decision variable, is the economic loss of the park caused by the abnormality of the jth subsystem to be analyzed, is the probability that appears abnormal within the minimum abnormal propagation time of the corresponding propagation segment after an abnormality occurs in the subsystems to be analyzed other than ; J is the number of subsystems to be analyzed; j is the number of the subsystem to be analyzed.

[0042] On the other hand, an embodiment of the present disclosure provides a smart campus management system based on multi-system integration, including: a primary identification module, a secondary identification module, and an optimization module;

[0043] The primary identification module uses sensor readings to collect the operation data of each subsystem in the campus in real time. After data preprocessing, it constructs the feature vector X of each subsystem, and combines it with the historical data set to construct a normal behavior model to identify problem subsystems;

[0044] The secondary identification module constructs a problem event matrix E, and based on the problem event matrix E, analyzes the foreseeability between each subsystem. After verification, it obtains the identification value ZS;

[0045] The optimization module, based on the identification value ZS, identifies the abnormal propagation path and constructs a data-driven optimization model to minimize the impact of system anomalies.

[0046] The present invention provides a smart campus management method and system based on multi-system integration, having the following beneficial effects:

[0047] (1) This method first uses the historical data set to train the normal behavior model of each subsystem using a machine learning model. When the actual behavior of a certain subsystem deviates from the normal model, it is determined as a problem subsystem, providing a basis for subsequent abnormal propagation analysis. After detecting the problem subsystem, a problem event matrix is constructed to describe the relationship of abnormal occurrences of each subsystem. By analyzing the matrix, the foreseeability of each subsystem can be calculated, that is, to judge whether the anomaly of a certain subsystem can predict the anomalies of other subsystems in advance. The identification value is calculated by statistical methods. If it is significantly higher than the threshold, it indicates that this subsystem may be a key link in abnormal propagation. Based on the calculated identification value, the system further identifies the abnormal propagation path, thereby constructing a data-driven optimization model. This model adjusts the control parameters of the subsystem through reinforcement learning to minimize the overall impact of anomalies. For example, in a smart campus, if the power load of the energy management system is abnormal, resulting in server overload, this method can identify that the energy management anomaly is the main cause, and by optimizing the energy distribution strategy, reduce the risk of server crash and improve the overall stability of the campus.

[0048] (2) Through the abnormal records in the historical data set, a problem event matrix E is constructed. The rows and columns of the matrix represent different subsystems, and historical abnormal information is filled. If two subsystems are not both abnormal at the same time, the matrix value is set to 1, indicating that there may be an abnormal association; otherwise, it is set to 0, indicating that the corresponding subsystems may be independently abnormal. Matrix analysis can discover the potential correlation between subsystems and avoid misjudgment of single-point anomalies. Through matrix analysis, pairs of subsystems with higher abnormal correlations are extracted to form pairs to be analyzed for foreseeability, and further analyze the foreseeability between systems to predict the abnormal propagation path.

[0049] (3) Through step S23, calculate the normal behavior models of the subsystem to be analyzed at different historical moments, and construct an extended model. Compared with the traditional autoregressive model based only on a single subsystem, this method can more comprehensively reflect the dynamic influence relationship between systems by introducing lag terms across subsystems, improving the accuracy of anomaly analysis. In the extended model, not only the historical characteristics of the subsystem itself are considered, but also the historical data of other subsystems are introduced as input variables, thus establishing a cross-system anomaly prediction model.

[0050] (4) By reconstructing the problem event matrix E, the system can re-analyze the anomaly correlation relationships of each subsystem to dynamically identify the linkage relationships between systems, thereby predicting the diffusion trend of anomalies, and by foreseeing the influence relationships between marked subsystems, draw a directed graph of anomaly propagation. In the directed graph, it is possible to clearly identify which subsystems are the "driving factors" of the anomaly (i.e., the source of anomaly propagation), and which are the "response factors" of the anomaly (affected subsystems), thus more accurately locking the root cause of the system anomaly. By calculating the anomaly propagation time delay, the present invention can analyze the shortest time required for an anomaly to propagate from one subsystem to another, and use this information to predict the occurrence probability of cascading anomalies. BRIEF DESCRIPTION OF THE DRAWINGS

[0051] In order to more clearly illustrate the technical solutions in the present application or the prior art, the following will briefly introduce the drawings required for use in the description of the embodiments or the prior art. Obviously, the drawings in the following description are some embodiments of the present application.

[0052] Figure 1 It is a schematic flowchart of a smart park management method based on multi-system integration according to the present invention;

[0053] Figure 2 It is a block diagram of a smart park management system based on multi-system integration according to the present invention. DETAILED DESCRIPTION OF THE EMBODIMENTS

[0054] To make the objectives, technical solutions, and advantages of the present application clearer, the following will clearly and completely describe the technical solutions in the present application with reference to the drawings in the present application. Obviously, the described embodiments are some, but not all, of the embodiments of the present application. Based on the embodiments in the present application, all other embodiments obtained by those of ordinary skill in the art without creative efforts

[0055] fall within the scope of protection of the present application.

[0056] Embodiment 1

[0057] Please refer to Figure 1, the present invention provides a smart park management method based on multi-system integration, including the following steps:

[0058] S1: Utilize sensor readings to collect the operation data of each subsystem in the park in real time. After data preprocessing, construct the feature vector X of each subsystem, and combine with the historical data set to construct a normal behavior model , to identify the problem subsystems;

[0059] S2: Construct a problem event matrix E, and based on the problem event matrix E, analyze the foreseeability existing between each subsystem. After verification, obtain the identification value ZS;

[0060] S3: Based on the identification value ZS, identify the abnormal propagation path and construct a data-driven optimization model to minimize the impact of system anomalies.

[0061] In the embodiment of this method, the smart park management method based on multi-system integration provided by the present invention can realize real-time anomaly monitoring, anomaly correlation analysis and intelligent optimization decision-making of the subsystems in the park by constructing a feature vector, a problem event matrix and a data-driven optimization model, thereby improving the intelligent level of park management. Compared with the traditional independent anomaly detection method, this method has the following significant advantages: Through step S1, the system collects the operation data of the subsystems in real time, combines with the historical data set, and constructs a normal behavior model. When the behavior of a certain subsystem deviates from the normal model, anomalies can be quickly detected, which avoids the false alarms and missed alarms caused by the traditional monitoring method based on fixed thresholds and improves the accuracy of anomaly identification. Through step S2, the system constructs a problem event matrix and analyzes the anomaly correlation between each subsystem. Traditional methods can often only detect anomalies in a single subsystem, while this method can reveal the causal relationship between different subsystems through matrix calculation and identification value, determine which anomalies have foreseeability, and provide data support for further optimization. In step S3, this method calculates the abnormal propagation path based on the identification value, constructs a data-driven optimization model, and dynamically adjusts the system parameters to minimize the impact of anomalies. For example, in a certain smart park, if the abnormal power load of the energy management subsystem causes the server of the equipment management subsystem to be overloaded, the traditional method may only handle the server anomaly alone and cannot identify the real source of the problem. However, the method of the present invention can accurately identify the energy management system as the anomaly source, and adjust the power supply plan through optimization strategies to avoid server failures in advance and improve the overall stability of the park. In summary, the present invention can not only improve the accuracy of anomaly detection, but also optimize the linkage management between subsystems, reduce the impact of park anomalies, and improve the operation and maintenance efficiency.

[0062] Embodiment 2

[0063] Please refer to Figure 1 , specifically: The specific steps of S1 include:

[0064] S11: The subsystems within the park include an energy management subsystem, a security monitoring and security protection subsystem, an environmental monitoring subsystem, a vehicle and intelligent parking management subsystem, a facility operation and maintenance and equipment management subsystem, an intelligent office and personnel management subsystem, and a data analysis and AI decision support subsystem. After cleaning, removing outliers, and standardizing the operation data of each subsystem within the park, feature vectors X of each subsystem are generated. The expression of the feature vector X of each subsystem is: , where n is the number of subsystems, are respectively the feature vector of the first subsystem, the feature vector of the second subsystem,..., the feature vector of the nth subsystem.

[0065] The specific steps of S1 also include:

[0066] S12: Based on the real-time acquisition of the operation data of each subsystem within the park in S11, the operation data of each subsystem in the past day is stored in a NoSQL database, and the data in the NoSQL database is used as a historical data group;

[0067] The data storage and management in S12 ensure that the operation data of each subsystem within the park can be stored for a long time and used for subsequent analysis; due to the large amount of data and complex data formats (such as time series data, log data, sensor data, etc.) generated by the multiple subsystems in the park, choosing a NoSQL database (such as MongoDB, HBase) for storage can provide efficient data reading and expansion capabilities. The stored data is divided into time windows (such as the past 24 hours) to form a historical data group to support data analysis and model training.

[0068] S13: The historical data of each subsystem in the normal operation state is respectively extracted from the historical data group to generate a normal feature vector Y. The expression of the normal feature vector Y is: , where N is the number of feature components, i is the subsystem number, and t is the historical moment number, are respectively the first feature component in the normal feature vector of the ith subsystem at the tth historical moment, the second feature component in the normal feature vector of the ith subsystem at the tth historical moment,..., the Nth feature component in the normal feature vector of the ith subsystem at the tth historical moment;

[0069] Among them, the normal operation state data of each subsystem is extracted from the historical data to construct a feature vector, providing input for subsequent model training.

[0070] S14: Use machine learning techniques and train the model using the normal feature vector Y to construct a normal behavior model , specifically:

[0071] ;

[0072] Wherein, is the output value of the normal behavior model constructed according to the normal eigenvector of the i-th subsystem, is the intercept term, , ,..., are the regression coefficients, representing the influence of each feature on the target variable ( ), , ,..., are the respective feature components within the normal eigenvector of the i-th subsystem.

[0073] The output value of the normal behavior model constructed according to the normal eigenvector of the i-th subsystem is for preparing for subsequent anomaly analysis. It represents the normal behavior model of the subsystem and is used to predict the target variable;

[0074] It should be noted that in the intelligent park management system, the dependent variable is the target variable we hope to predict, usually representing the key performance indicator (KPI) of a certain subsystem. This indicator is used to measure whether the operation status of the subsystem is normal.

[0075] Specifically, the main purpose of this paragraph is to construct a normal behavior model based on historical data to support subsequent anomaly detection and analysis. Specifically, this paragraph covers three core steps: data storage, feature extraction, and model training, aiming to establish an accurate normal operation baseline, thereby realizing intelligent anomaly monitoring. When the actual operation status of the subsystem deviates from the normal model, anomalies can be quickly identified, accurate early warnings can be achieved, and the intelligent level of park management can be improved.

[0076] Embodiment 3

[0077] Please refer to Figure 1 , specifically: The specific steps of S1 further include:

[0078] S15: Based on the historical data set and the normal behavior model , calculate and obtain the residual , specifically: ; wherein, is the actual value. An anomaly critical value V is preset, and by comparing the residual with the anomaly critical value V, the problem subsystem is identified. Specifically: If the absolute value of the residual > the anomaly critical value V, it is determined that the corresponding subsystem has an anomaly, and at this time, the corresponding subsystem is marked as a problem subsystem.

[0079] By establishing a normal behavior model and monitoring the residuals, the intelligent park management system can effectively detect abnormal behaviors of subsystems and ensure the stable operation of the system.

[0080] Specifically, the Grubbs test method can be used to set the threshold.

[0081] The specific steps of S2 include:

[0082] S21: Based on the way of determining the problem subsystems in S15, obtain several groups of problem subsystems corresponding to historical moments, and construct several groups of problem event matrices E for historical moments according to the problem subsystems corresponding to several groups of historical moments. Each row and each column of the problem event matrix E represents a subsystem; fill the problem event matrix E through the problem subsystems corresponding to several groups of historical moments. Specifically: if the subsystems are not all marked as problem subsystems at the same moment, the value at the corresponding position in the problem event matrix E is 1, otherwise, the value is 0.

[0083] S22: Extract features from the positions corresponding to the value 1 in the problem event matrix E to obtain several groups of pairs of anticipations to be analyzed.

[0084] The abnormal event matrix E is used to depict the simultaneous occurrence of abnormal events between subsystems.

[0085] Specifically, through residual analysis + problem event matrix, this method can effectively identify isolated anomalies and system-level anomalies, and improve the accuracy of anomaly detection. For example, in a certain intelligent park, if the air conditioning system has a temperature anomaly (S15 identifies the problem subsystem), and at the same time, it is found that the server cooling system is also abnormal at multiple historical moments (the matrix constructed in S21), then the system can automatically identify that the air conditioning failure may affect the server cooling (extraction of pairs of anticipations in S22). Traditional methods may only alarm when the server temperature exceeds the limit, while this method can discover associated anomalies in advance, adjust the air conditioning load in advance, avoid the server crashing due to high temperature, and improve the intelligent level of park operation and maintenance. In summary, this method can accurately analyze the internal relationship of subsystem anomalies, optimize anomaly management, and improve system stability and operation and maintenance efficiency.

[0086] Embodiment 4

[0087] Please refer to Figure 1 , specifically: The specific steps of S2 also include:

[0088] S23: Mark the problem subsystems in several groups of pairs of anticipations to be analyzed as subsystems to be analyzed, and obtain several groups of normal behavior models of the subsystems to be analyzed at different historical moments by the method of obtaining the normal behavior model in S14 ​, and in combination with the prediction pairs to be analyzed, the two subsystems to be analyzed involved in the prediction pairs to be analyzed are passed through the corresponding normal behavior model , after expansion, an expanded model is obtained , specifically:

[0089] ;

[0090] In the formula, is an autoregressive model that only contains its own lag term;

[0091] ;

[0092] In the formula, is the output value of the expanded model constructed according to the normal feature vector of the i-th subsystem at the t-th historical moment, is the intercept term, representing the intercept of the model;

[0093] , , and are all regression coefficients, respectively representing the influence degree of the first lag term of the q-th lag term of the influence degree of the first lag term of the p-th lag term of

[0094] ,..., are respectively the output values of the normal behavior model constructed according to the normal feature vectors of the i-th subsystem at the t - 1 historical moment,..., the t - q historical moment;

[0095] is the error term, q is the lag order of the lag order of ,..., are respectively the output values of the normal behavior model constructed according to the normal feature vectors of the h-th subsystem at the t - 1 historical moment,..., the t - p historical moment; h and i are both subsystem numbers. For the determination of the lag order, the Akaike Information Criterion (AIC) or the Bayesian Information Criterion (BIC) can be used to select the optimal lag order.

[0096] Specifically, traditional anomaly detection methods often only focus on the historical states of individual subsystems and are difficult to identify potential influence relationships between subsystems. This method first marks the problem subsystems based on the foresight pairs to be analyzed, and uses historical data to train a normal behavior model to construct a prediction model based on the autoregressive model (AR). Then, by introducing multiple lag orders, the model not only depends on the historical data of the subsystem itself but also considers the state changes of other subsystems, thereby optimizing the accuracy of anomaly detection. By expanding the behavior model, the normal behavior models of the two subsystems in the foresight pair to be analyzed are combined to establish a system of equations containing mutual influence relationships. For example, if an anomaly in the energy management subsystem may affect the facility operation and maintenance subsystem, the extended model will simultaneously consider the influence of the lag state of the energy management subsystem on the facility operation and maintenance subsystem, thereby improving the ability to identify the anomaly propagation path.

[0097] Embodiment 5

[0098] Please refer to Figure 1 , specifically: The specific steps of S2 also include:

[0099] S24: By comparing the sum of squared residuals of the autoregressive model that only contains its own lag terms and the output value of the extended model constructed according to the normal feature vector of the i-th subsystem at the t-th historical moment , determine whether the added lag term improves the prediction ability of , specifically:

[0100] ;

[0101] In the formula, ZS is the identification value, represents the sum of squared residuals of the autoregressive model that only contains its own lag terms (i.e., without the lag term), represents the sum of squared residuals of the output value of the extended model constructed according to the normal feature vector of the i-th subsystem at the t-th historical moment (i.e., with the lag term), and T represents the number of samples; The numerator part represents the degree of reduction in error after introducing the lag term. If is large, it indicates that after adding the lag term, the reduction in error is obvious, indicating that may have a foresight effect on ; The denominator part is used to standardize the numerator part so that the identification value ZS follows an F distribution;

[0102] Through the acquisition of the recognition value ZS, the potential associations and causal relationships between abnormal events can be effectively recognized, thereby revealing deep-seated problems.

[0103] Among them, the F-distribution is a probability distribution used for hypothesis testing, and is often used to compare two variances or the goodness of fit of a model;

[0104] S25: Preset a recognition threshold. By comparing it with the recognition value ZS, if the recognition value ZS > the recognition threshold, it means that the added lag term will improve the prediction ability for and mark the corresponding pair to be analyzed as a foresight pair and extract it to obtain a feature set. Otherwise, it cannot;

[0105] Specifically, this method first constructs an autoregressive model that only contains its own lag terms, and then expands it into a model that contains the lag terms of other subsystems, and calculates their residual sum of squares respectively. If the error of the expanded model is significantly reduced, it indicates that introducing the lag terms of other subsystems can improve the prediction effect, indicating that there may be an abnormal association between these two subsystems, thereby improving the accuracy of anomaly detection. By calculating the recognition value and performing an F-distribution test, it is judged whether the newly added lag term has truly improved the prediction ability. This statistical test method can effectively filter out irrelevant variables, only retain the lag terms of subsystems that contribute to abnormal prediction, avoid the increase in model complexity or misjudgment caused by too many parameters, and improve the stability of the model. If the recognition value exceeds the set threshold, the system marks the subsystem pair as a foresight pair and extracts its features for in-depth analysis. For example, in a certain smart park, there may be an implicit relationship between the voltage fluctuation of the energy management subsystem and the abnormal heat dissipation of the server. Traditional methods may only alarm when the server temperature is abnormal, but this method can identify the lag effect of voltage fluctuation on server heat dissipation, adjust the power supply strategy in advance, and reduce the risk of server failure. In summary, this method accurately identifies the abnormal associations between subsystems through residual comparison, F-distribution test, and foresight pair extraction, improves the anomaly warning ability, thereby optimizing the management of the smart park, reducing sudden failures, and improving the system stability.

[0106] Example 6

[0107] Please refer to Figure 1 Specifically: S3: The specific steps include:

[0108] S31: Based on the extraction set, re-establish and fill a set of problem event matrices E. Each row and each column of the re-established problem event matrix E represents a subsystem to be analyzed. The specific content filled is: Search for each foresight pair in the re-established problem event matrix E, and mark the corresponding positions with the value 1, and the values of the remaining positions are 0;

[0109] S32: Based on the content of S31, taking the subsystem to be analyzed as the node, if the position in the re-established problem event matrix E , then draw from point to After traversing the directed edges, a set of directed graphs is generated; among them, and are different subsystems to be analyzed, s and c are the numbers of the subsystems to be analyzed;

[0110] This process can help us clarify which subsystems are "driving factors" and which are "response factors", providing guidance for subsequent decision-making.

[0111] S33: According to the directed graph, the abnormal propagation path of the corresponding subsystem to be analyzed is identified, and the source of the abnormality is determined. According to the abnormal propagation path of the corresponding subsystem to be analyzed, the abnormal propagation time delay of each propagation segment in the abnormal propagation path of the corresponding subsystem to be analyzed is counted, and the minimum abnormal propagation time delay of each propagation segment is extracted.

[0112] The specific steps of S3 include:

[0113] S34: Determine the minimum abnormal propagation time delay of each propagation segment. When an abnormality occurs, predict the cascading abnormality in each system to statistically After an exception occurs, within the minimum exception propagation time of the corresponding propagation segment, Probability of abnormality ;

[0114] S35: Use reinforcement learning (RL) to train optimization strategies based on historical data to build a data-driven optimization model, specifically:

[0115] ;

[0116] Where L is the optimization objective function (the total amount of abnormal impact of the entire park), are decision variables (control strategy parameters), is the economic loss of the park caused by the abnormality of the jth subsystem to be analyzed, To remove the After an abnormality occurs in the subsystem to be analyzed other than The probability of abnormality; J is the number of subsystems to be analyzed; j is the number of subsystems to be analyzed. The purpose of this formula is to minimize the sum of the product of abnormal probability and economic loss of all subsystems to be analyzed by adjusting the control strategy θ, that is, to reduce the overall risk of the system.

[0117] The economic loss of the park caused by the abnormality of the jth subsystem to be analyzed It can be determined by business data or expert experience, such as production losses caused by equipment outages, additional costs due to abnormal energy consumption, etc.

[0118] The purpose of adjusting θ is to optimize the system control strategy to reduce the abnormal probability or mitigate the abnormal impact. The abnormal risks of different subsystems are different. Adjusting the decision variable θ can optimize resource allocation to minimize the overall abnormal impact. And some abnormal risks are difficult to completely eliminate, but by adjusting θ, subsystems with high loss risks can be preferentially protected.

[0119] Specifically, through steps S31 - S33, the system reconstructs the problem event matrix E based on the extraction set, identifies the foresight pairs, and generates a directed graph of abnormal propagation between subsystems. This method can reveal the abnormal transmission relationship between subsystems. By traversing the directed graph, it can identify which subsystems are abnormal driving factors and which are response factors. Compared with traditional independent anomaly detection, this method can analyze the abnormal propagation path more comprehensively to ensure that the anomaly management strategy is more targeted. In step S34, the system estimates the probability of cascading impact in different subsystems after an anomaly occurs based on the minimum abnormal propagation time delay of each propagation segment. For example, in a certain smart park, if there is a voltage fluctuation in the energy management subsystem (X1), it usually causes abnormal heat dissipation of the server in the equipment management subsystem (X2) after 5 minutes, and may affect the office automation system (X3) after 10 minutes. Traditional methods may trigger an alarm only after the server temperature is abnormal, while this method can predict the probability of future anomalies in X2 and X3 in advance to provide support for preventive maintenance. In step S35, the system trains and optimizes the strategy based on historical data, and uses reinforcement learning (RL) to dynamically adjust the subsystem control parameter θ to minimize the overall abnormal impact in the park. For example, if it is detected that the energy management subsystem may cause abnormal heat dissipation of the server, the system can automatically adjust the air - conditioning load or optimize the power distribution to prevent the temperature from being too high and avoid server downtime. In summary, this method can not only accurately identify the abnormal propagation path, but also perform cascading anomaly prediction based on the propagation time delay, and use reinforcement learning to dynamically optimize the control strategy, ultimately reducing the economic losses caused by anomalies and improving the stability, prediction ability and automated management level of the smart park.

[0120] Embodiment 7

[0121] Please refer to Figure 2 , specifically: A smart park management system based on multi - system integration, including a primary identification module, a secondary identification module, and an optimization module;

[0122] The primary identification module uses sensor readings to collect the operation data of each subsystem in the park in real - time. After data pre - processing, it constructs the feature vector X of each subsystem and combines it with the historical data set to construct a normal behavior model , to identify the problem subsystems;

[0123] The secondary recognition module constructs a problem event matrix E, and based on the problem event matrix E, analyzes the predictive capabilities existing between subsystems. After verification, the recognition value ZS is obtained.

[0124] The optimization module, based on the recognition value ZS, identifies the abnormal propagation path and constructs a data-driven optimization model to minimize the impact of system anomalies.

[0125] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present application, and are not intended to limit them; although the present application has been described in detail with reference to the foregoing embodiments, those of ordinary skill in the art should understand that they can still modify the technical solutions described in the foregoing embodiments, or perform equivalent replacements for some of the technical features; and these modifications or replacements do not cause the essence of the corresponding technical solutions to deviate from the spirit and scope of the technical solutions of the embodiments of the present application.

Claims

1. A smart park management method based on multi-system integration, characterized by: The following steps are involved: S1: Use sensor readings to collect real-time operating data of each subsystem in the park. After data preprocessing, construct the feature vector X of each subsystem, and combine it with the historical data set to build a normal behavior model , to identify the problem subsystem; S2: Construct the problem event matrix E, and analyze the foresight capabilities between the subsystems based on the problem event matrix E, and obtain the identification value ZS after verification; The specific steps of S2 include: S21: Based on the problem subsystem determined, the problem subsystems corresponding to several groups of historical moments are obtained, and according to the problem subsystems corresponding to several groups of historical moments, the problem event matrix E of several groups of historical moments is constructed, and each row and each column of the problem event matrix E represents a subsystem; the problem event matrix E is filled with the problem subsystems corresponding to several groups of historical moments, specifically: if the subsystems are not marked as problem subsystems at the same time, the value of the corresponding position in the problem event matrix E is 1, otherwise, the value is 0; S22: extracting features from the positions corresponding to the values ​​of 1 in the problem event matrix E to obtain several groups of foresight pairs to be analyzed; S23: Mark the problem subsystems in several groups of foreseen pairs to be analyzed as subsystems to be analyzed, and obtain the normal behavior model By using the method of , and combined with the predicted pair to be analyzed, the two groups of subsystems to be analyzed involved in the predicted pair to be analyzed are analyzed through the corresponding normal behavior model , after expansion, obtain the expanded model , specifically: ； In the formula, To contain only Autoregressive model with own lag; ； In the formula, is the output value of the extended model constructed based on the normal feature vector of the ith subsystem at the tth historical moment, is the intercept term, , , and are regression coefficients, ,..., are the output values ​​of the normal behavior model constructed according to the normal feature vectors of the ith subsystem at the t-1th historical moment, ..., tqth historical moment respectively; is the error term, q is The lag order, p, is The lag order of ,..., are the output values ​​of the normal behavior model constructed according to the normal feature vectors of the h-th subsystem at the t-1-th historical moment, ..., and the tp-th historical moment, respectively; h and i are the numbers of the subsystems; S3: Based on the identification value ZS, identify the abnormal propagation path and build a data-driven optimization model to minimize the impact of system abnormalities; S31: Based on the feature set, a set of problem event matrices E is re-established and filled, where each row and each column of the re-established problem event matrix E represents a subsystem to be analyzed, and the specific content of the filling is: each predicted pair is found in the re-established problem event matrix E, and the position corresponding to each predicted pair is marked as a value 1, and the values ​​of the remaining positions are 0; S32: Based on the content of S31, taking the subsystem to be analyzed as the node, if the position in the re-established problem event matrix E , then draw from point to After traversing the directed edges, a set of directed graphs is generated; among them, and are different subsystems to be analyzed, s and c are the numbers of the subsystems to be analyzed; S33: According to the directed graph, the abnormal propagation path of the corresponding subsystem to be analyzed is identified, and the source of the abnormality is determined. According to the abnormal propagation path of the corresponding subsystem to be analyzed, the abnormal propagation time delay of each propagation segment in the abnormal propagation path of the corresponding subsystem to be analyzed is counted, and the minimum abnormal propagation time delay of each propagation segment is extracted.

2. According to claim 1, a smart park management method based on multi-system integration is characterized in that: The specific steps of S1 include: S11: The subsystems in the park include energy management subsystem, safety monitoring and security subsystem, environmental monitoring subsystem, vehicle and smart parking management subsystem, facility operation and equipment management subsystem, smart office and personnel management subsystem, and data analysis and AI decision support subsystem. The operating data of each subsystem in the park are cleaned, outliers are eliminated, and data is standardized to generate the characteristic vector X of each subsystem. The expression of the characteristic vector X of each subsystem is: , n is the number of subsystems, They are respectively the eigenvector of the first subsystem, the eigenvector of the second subsystem, ..., the eigenvector of the nth subsystem.

3. According to claim 2, a smart park management method based on multi-system integration is characterized in that: The specific steps of S1 also include: S12: Based on the real-time acquisition of the operating data of each subsystem in the park by S11, the operating data of each subsystem in the past day is stored in the NoSQL database, and the data in the NoSQL database is used as the historical data group; S13: extracting historical data of each subsystem under normal operation from the historical data group to generate a normal feature vector Y. The expression of the normal feature vector Y is: , N is the number of characteristic components, i is the subsystem number, t is the historical moment number, are respectively the first characteristic component in the normal characteristic vector of the ith subsystem at the tth historical moment, the second characteristic component in the normal characteristic vector of the ith subsystem at the tth historical moment, ..., the Nth characteristic component in the normal characteristic vector of the ith subsystem at the tth historical moment; S14: Use machine learning techniques and train the model using the normal feature vector Y to build a normal behavior model , specifically: ； In the formula, is the output value of the normal behavior model constructed according to the normal feature vector of the i-th subsystem, is the intercept term, , ,..., is the regression coefficient, , ,..., are the eigencomponents in the normal eigenvector of the i-th subsystem.

4. The smart park management method based on multi-system integration according to claim 3 is characterized in that: The specific steps of S1 also include: S15: Based on historical data sets and normal behavior models , calculate the residual , specifically: ;in, is the actual value, and the abnormal critical value V is pre-set, and the residual Compared with the abnormal critical value V, to identify the problem subsystem, specifically: if the residual If the absolute value of is greater than the abnormal critical value V, the corresponding subsystem is judged to be abnormal, and the corresponding subsystem is marked as a problem subsystem.

5. The smart park management method based on multi-system integration according to claim 4 is characterized in that: The specific steps of S2 also include: S24: By comparing only Autoregressive Model with Its Own Lagged Term and the output value of the expanded model constructed based on the normal feature vector of the i-th subsystem at the t-th historical moment The residual sum of squares is used to determine the addition Does the lag term improve the The predictive power is: ； Where ZS is the identification value, Indicates that only Autoregressive Model with Its Own Lagged Term The residual sum of squares, Represents the output value of the expanded model constructed based on the normal feature vector of the ith subsystem at the tth historical moment The residual sum of squares, T represents the number of samples; S25: preset the recognition threshold, and compare it with the recognition value ZS. If the recognition value ZS> the recognition threshold, it means adding The lag term will increase the The prediction ability of the feature set is obtained by marking the corresponding predicted pairs to be analyzed as predicted pairs and extracting them to obtain the feature set.

6. The smart park management method based on multi-system integration according to claim 5 is characterized by: The specific steps of S3 include: S34: Determine the minimum abnormal propagation time delay of each propagation segment. When an abnormality occurs, predict the cascading abnormality in each system to statistically After an exception occurs, within the minimum exception propagation time of the corresponding propagation segment, Probability of abnormality ; S35: Use reinforcement learning to train optimization strategies based on historical data to build a data-driven optimization model, specifically: ； Where L is the optimization objective function, is the decision variable, is the economic loss of the park caused by the abnormality of the jth subsystem to be analyzed, To remove the After an abnormality occurs in the subsystem to be analyzed other than The probability of anomaly; J is the number of subsystems to be analyzed; j is the number of subsystems to be analyzed.

7. A smart park management system based on multi-system integration, used to implement a smart park management method based on multi-system integration as described in any one of claims 1 to 6 above, characterized in that: It includes a primary recognition module, a secondary recognition module and an optimization module; The primary recognition module uses sensor readings to collect real-time operating data of each subsystem in the park. After data preprocessing, it constructs the feature vector X of each subsystem and combines it with the historical data set to build a normal behavior model. , to identify the problem subsystem; The secondary identification module constructs the problem event matrix E, and analyzes the foresight capabilities between each subsystem based on the problem event matrix E. After testing, the identification value ZS is obtained; The optimization module, based on the identification value ZS, identifies the abnormal propagation path and builds a data-driven optimization model to minimize the impact of system abnormalities.

Images