Intelligent park operation and maintenance management method and system based on AI autonomous learning

Abstract

The invention provides a smart park operation and maintenance management method and system based on AI autonomous learning, and the method comprises the steps: obtaining a real-time operation and maintenance data set of a plurality of operation and maintenance data sources in a smart park, covering equipment operation logs, environment monitoring texts and user interaction data, carrying out the semantic feature extraction of the real-time operation and maintenance data set, and carrying out the semantic feature extraction of the real-time operation and maintenance data set; the method comprises the following steps: generating a set containing semantic features of an equipment operation state, environment association and user interaction intention, then performing dynamic anomaly analysis on the semantic feature set by using a preset AI autonomous learning model to obtain an anomaly feature set, matching the anomaly feature set with a preset optimization strategy library to generate a target operation and maintenance optimization strategy set, and finally performing optimization on the target operation and maintenance optimization strategy set. And finally, executing an operation and maintenance optimization strategy and feeding back a result to the operation and maintenance management platform, thereby improving the intelligence and automation level of operation and maintenance management of the park, and ensuring stable operation of the park.

Description

Technical Field

[0001] This invention relates to the field of smart park operation and maintenance technology, and more specifically, to a smart park operation and maintenance management method and system based on AI autonomous learning. Background Technology

[0002] With the development of technology, the construction of smart parks has become an important direction for urban development. Smart parks encompass numerous complex systems and equipment, such as power systems, security systems, and environmental monitoring equipment, and also involve a large number of user interactions and uses, which poses enormous challenges to the operation and maintenance management of the parks.

[0003] Currently, traditional smart park operation and maintenance management methods lack dynamic adaptation to data, failing to adjust analysis strategies in a timely manner based on real-time data changes, and making it difficult to detect potential anomalies and trends in the data. Moreover, they typically only perform simple statistical analysis of the data without delving into the semantic information behind it, failing to accurately grasp the inherent relationship between equipment operating status, environmental conditions, and user needs.

[0004] When anomalies are detected, traditional methods rely on human experience to formulate operational strategies. This approach is not only inefficient but also susceptible to human influence, resulting in strategies that lack scientific rigor and specificity. Different operations personnel may offer different solutions based on their own experience, making it difficult to guarantee the consistency and effectiveness of the strategies. Summary of the Invention

[0005] In view of the aforementioned problems, and in conjunction with the first aspect of the present invention, embodiments of the present invention provide a smart park operation and maintenance management method based on AI autonomous learning, the method comprising: Acquire a real-time operation and maintenance data set from multiple operation and maintenance data sources within the smart park. The real-time operation and maintenance data set includes equipment operation log data, environmental monitoring text data, and user interaction data. Semantic features are extracted from the real-time operation and maintenance data set to generate a semantic feature set, which includes semantic features of device operating status, semantic features of environment association, and semantic features of user interaction intent. Based on a preset AI autonomous learning model, dynamic anomaly analysis is performed on the semantic feature set to obtain an anomaly feature set. Based on the set of abnormal features, a strategy matching process is performed with a preset optimization strategy library to generate a target set of operation and maintenance optimization strategies; Execute each operation and maintenance optimization strategy in the target operation and maintenance optimization strategy set, and feed back the execution results to the operation and maintenance management platform of the smart park.

[0006] In one possible implementation of the first aspect, the step of extracting semantic features from the real-time operation and maintenance data set to generate a semantic feature set includes: The equipment operation log data is cleaned by removing redundant symbols and unstructured data segments to obtain standardized equipment operation text data. The standardized equipment operation text data is subjected to contextual semantic encoding processing, the temporal correlation semantic features in the equipment operation log data are extracted, and the temporal correlation semantic features are matched with a preset equipment operation status keyword library to generate the equipment operation status semantic features. The environmental monitoring text data is segmented to obtain multiple environmental monitoring word units. Based on a preset environmental association rule library, the multiple environmental monitoring word units are semantically clustered to generate the environmental association semantic features. A preset intent recognition model is invoked to perform intent classification processing on the user interaction data, generating semantic features of the user interaction intent.

[0007] In one possible implementation of the first aspect, the step of performing contextual semantic encoding processing on the standardized device operation text data and extracting time-series related semantic features from the device operation log data includes: The standardized equipment operation text data is divided into multiple equipment operation text fragments in chronological order; The pre-trained context semantic encoder is invoked to encode the text segments running on each device, generating a context semantic vector for each text segment running on each device. The context semantic vectors of adjacent time periods are aggregated using a sliding window to obtain the temporal association semantic features; wherein, the sliding window aggregation process includes: weighted summation of multiple context semantic vectors within the current time window, with the weight values ​​dynamically adjusted according to the temporal distance between each context semantic vector and the center point of the current time window.

[0008] In one possible implementation of the first aspect, the step of performing semantic clustering processing on the plurality of environmental monitoring word units based on a preset environmental association rule base to generate the environmental association semantic features includes: Match the environmental entity category and environmental attribute category corresponding to each environmental monitoring term unit from the environmental association rule base; Environmental monitoring terminology units belonging to the same environmental entity category are aggregated into environmental entity sets, and the distribution characteristics of environmental attribute values ​​for each environmental attribute category in each environmental entity set are extracted. The distribution characteristics of environmental attribute values ​​for each environmental attribute category in each set of environmental entities are standardized. Based on a preset environmental parameter weighting table, the standardized environmental attribute value distribution characteristics are weighted and summed to generate a comprehensive impact score. The set of environmental entities whose comprehensive impact score is greater than a preset threshold is marked as key environmental entities, and the environmental attribute distribution features corresponding to the key environmental entities are combined to generate the environmental association semantic features.

[0009] In one possible implementation of the first aspect, the dynamic anomaly analysis processing of the semantic feature set based on a preset AI autonomous learning model to obtain an anomaly feature set includes: The semantic features of the device's operating status, the semantic features of the environment association, and the semantic features of the user's interaction intent are input into the multi-branch feature extraction layer of a preset AI autonomous learning model, wherein: The first branch feature extraction layer uses a temporal convolutional network to perform sliding window feature extraction on the temporal dimension of the semantic features of the device's operating state, generating a device state feature vector; The second branch feature extraction layer uses a graph convolutional network to perform neighborhood propagation processing on the environmental associated semantic features based on the spatial topology of environmental monitoring nodes, generating an environmental spatial feature vector. The third branch feature extraction layer uses a multi-head attention mechanism to perform cross-dimensional correlation analysis on the semantic features of the user interaction intent, and generates an intent-focused feature vector. The device state feature vector, environmental space feature vector, and intent focus feature vector are mapped to a unified feature space according to a preset feature projection matrix and then fused to generate a fused feature vector. The feature projection matrix is ​​optimized during the model training phase by minimizing the cosine similarity difference of the feature vectors output by each branch feature extraction layer. The fused feature vector is input into the fully connected classification layer of the AI ​​autonomous learning model, and a comprehensive anomaly score sequence is output. The training data of the fully connected classification layer includes historical anomaly event annotation data and normalized vectors of corresponding semantic features. During the training process, the classification weight parameters of the focus loss function are dynamically adjusted according to the distribution ratio of anomaly categories in the current batch of training data to optimize the classification boundary. After aligning the timestamps of the comprehensive anomaly score sequence with the real-time judgment threshold output by the dynamic threshold generation module, point-by-point comparison is performed to determine anomalies, and feature segments whose comprehensive anomaly scores exceed the real-time judgment threshold are extracted as candidate anomaly features. Contextual verification is performed on the candidate abnormal features: based on the temporal continuity of the semantic features of the device operating status, the spatial distribution consistency of the semantic features of the environment association, and the logical correlation of the semantic features of the user interaction intent, false detection feature fragments are eliminated to generate the final abnormal feature set.

[0010] In one possible implementation of the first aspect, the output processing of the dynamic threshold generation module includes: The comprehensive anomaly score sequence output by the AI ​​autonomous learning model within a preset time window is collected in real time, and the mean and standard deviation of the comprehensive anomaly score sequence are calculated. After converting the equipment operating load rate into a normalized value in the range of 0 to 1, a benchmark threshold is generated based on the linear combination of the mean and standard deviation, wherein the coefficient of the standard deviation term is positively correlated with the normalized equipment operating load rate. An initial real-time judgment threshold is formed by superimposing a preset baseline offset, wherein the baseline offset is adjusted exponentially based on the frequency of anomalies occurring in the same historical period. When a sudden change in the parameters of the environmental associated semantic features is detected, the threshold dynamic compensation mechanism is activated: based on the ratio of the rate of change of environmental parameters to the safety threshold, the threshold offset is calculated using a preset logarithmic function coefficient, and the initial real-time judgment threshold is offset downward to form the final real-time judgment threshold for application.

[0011] In one possible implementation of the first aspect, the step of performing strategy matching processing based on the abnormal feature set and a preset optimization strategy library to generate a target operation and maintenance optimization strategy set includes: Extract the preset anomaly trigger feature vector corresponding to each operation and maintenance optimization strategy from the optimization strategy library, wherein the preset anomaly trigger feature vector is generated in the unified feature space of the AI ​​autonomous learning model; For each abnormal feature vector in the abnormal feature set, calculate its similarity with the preset abnormal trigger feature vectors of all operation and maintenance optimization strategies. When the similarity value between the preset abnormal trigger feature vector of any operation and maintenance optimization strategy and the corresponding type of abnormal feature vector exceeds the preset strategy trigger threshold, add the operation and maintenance optimization strategy to the candidate strategy set. The operation and maintenance optimization strategies in the candidate strategy set are prioritized and sorted. The top N operation and maintenance optimization strategies are selected to generate the target operation and maintenance optimization strategy set, where N is dynamically set according to the concurrent processing capability of the current operation and maintenance management platform.

[0012] In one possible implementation of the first aspect, the priority ranking process for each operation and maintenance optimization strategy in the candidate strategy set includes: Obtain the historical execution effect score for each operation and maintenance optimization strategy. The historical execution effect score is calculated and normalized to the range of 0 to 1 by weighting the deviation reduction rate in the equipment status response data after the operation and maintenance optimization strategy is executed, the safety threshold achievement rate of the environmental parameter change data, and the improvement of user satisfaction feedback data. Extract the dynamic weight coefficients of each abnormal feature vector in the current abnormal feature set. The dynamic weight coefficients are normalized according to the proportion of the abnormal score value of the corresponding feature segment in the comprehensive abnormal score sequence in the total abnormal score. The historical execution performance score is multiplied by the dynamic weight coefficient of the current abnormal feature vector of the corresponding operation and maintenance optimization strategy to generate a strategy priority index. The candidate strategy set is sorted by priority index from high to low.

[0013] In one possible implementation of the first aspect, executing each operation and maintenance optimization strategy in the target operation and maintenance optimization strategy set and feeding back the execution results to the operation and maintenance management platform of the smart park includes: The operation instruction sequence in the equipment operation optimization strategy is parsed to generate a set of equipment control instructions, which is then sent to the corresponding target equipment controller. Monitor the device status response data returned by the target device controller. If the device status response data matches the expected optimization effect characteristics, mark the device operation optimization strategy as successfully executed. The set of environmental adjustment instructions in the environmental monitoring optimization strategy is analyzed to activate environmental control equipment and collect environmental parameter change data in real time. When the environmental parameter change data reaches the preset safety threshold range, the environmental monitoring optimization strategy is marked as completed. The interaction logic update instructions in the user interaction optimization strategy are parsed, the response logic of the user interface is modified, and user satisfaction feedback data is collected. When the user satisfaction feedback data exceeds a preset satisfaction threshold, the user interaction optimization strategy is marked as effective. In one possible implementation of the first aspect, monitoring the device status response data returned by the target device controller includes: After extracting real-time equipment operation characteristics from the equipment status response data and performing standardization processing, the deviation between these characteristics and the expected equipment operation characteristics is calculated. If the deviation remains below the dynamic threshold for more than a preset time window, the device operation optimization strategy is deemed to have been successfully executed; wherein, the dynamic threshold is adaptively adjusted based on the device type and historical optimization effect data.

[0014] In another aspect, embodiments of the present invention also provide a smart park operation and maintenance management system based on AI autonomous learning, including a processor and a machine-readable storage medium. The machine-readable storage medium is connected to the processor. The machine-readable storage medium is used to store programs, instructions, or code. The processor is used to execute the programs, instructions, or code in the machine-readable storage medium to implement the above-mentioned method.

[0015] Based on the above, this embodiment of the invention acquires a real-time operation and maintenance data set from multiple operation and maintenance data sources within a smart park, covering equipment operation log data, environmental monitoring text data, and user interaction data. Semantic feature extraction is performed on this real-time operation and maintenance data set to generate a semantic feature set containing semantic features of equipment operating status, environmental association, and user interaction intent. This effectively mines the potential semantic information behind the data, achieving the transformation from raw data to valuable features. Based on a preset AI autonomous learning model, dynamic anomaly analysis is performed on the semantic feature set to obtain an anomaly feature set. Leveraging the AI's autonomous learning capabilities, the system can adaptively and accurately monitor and identify anomalies in the park's operation and maintenance status. It can promptly identify potential problems, and generate a target set of operation and maintenance optimization strategies by matching the set of abnormal features with a preset optimization strategy library. It can quickly and accurately match appropriate optimization strategies according to specific abnormal situations, improving the scientific nature and effectiveness of operation and maintenance decisions. It executes each operation and maintenance optimization strategy in the target set of operation and maintenance optimization strategies and feeds the execution results back to the operation and maintenance management platform of the smart park. It can realize dynamic adjustment and continuous optimization of smart park operation and maintenance management, thereby significantly improving the efficiency and quality of smart park operation and maintenance management, ensuring the stable operation of park equipment, environmental suitability, and a good user experience, and improving the overall intelligence and automation level of smart park operation and maintenance management. Attached Figure Description

[0016] Figure 1 This is a schematic diagram of the execution flow of the smart park operation and maintenance management method based on AI autonomous learning provided in the embodiments of the present invention.

[0017] Figure 2 This is a schematic diagram of exemplary hardware and software components of the AI-based autonomous learning-based smart park operation and maintenance management system provided in an embodiment of the present invention. Detailed Implementation

[0018] The present invention will now be described in detail with reference to the accompanying drawings. Figure 1 This is a flowchart illustrating a smart park operation and maintenance management method based on AI autonomous learning, provided in one embodiment of the present invention. The following is a detailed description of this smart park operation and maintenance management method based on AI autonomous learning.

[0019] Step S110: Obtain a real-time operation and maintenance data set from multiple operation and maintenance data sources within the smart park. The real-time operation and maintenance data set includes equipment operation log data, environmental monitoring text data, and user interaction data.

[0020] Within a smart park, there are various types of operation and maintenance data sources, each providing specific information related to park operation and maintenance.

[0021] Regarding equipment operation log data, numerous devices are distributed throughout the park, such as electrical equipment, water supply and drainage equipment, elevator equipment, and ventilation and air conditioning equipment. Taking elevator equipment as an example, its operation log data includes information such as elevator start time, stop time, operating speed, load status, and floor stop records. This data is collected through sensors and data recording modules installed in the elevator control system. Sensors monitor various operating parameters of the elevator in real time and transmit the data to the recording module for storage. For electrical equipment, operation log data may include information such as voltage, current, power, and electricity consumption, which is collected and recorded by devices such as meters and sensors.

[0022] Environmental monitoring data is acquired through environmental monitoring equipment deployed in various locations within the park. Outdoors, weather stations collect meteorological information such as temperature, humidity, wind speed, wind direction, and light intensity. Indoors, air quality monitors are installed in different areas, such as office areas, public areas, and computer rooms, to monitor in real time the concentration of harmful gases in the air, such as formaldehyde, benzene, and carbon dioxide, as well as particulate matter concentrations, such as PM2.5 and PM10. This environmental monitoring equipment converts the collected data into text information for storage and transmission.

[0023] User interaction data primarily originates from the park's user interaction platforms, including the park's official website and mobile app. Various interaction data are generated when users interact with these platforms. For example, a user searching for parking space availability on the app will generate a query record; a user evaluating and providing feedback on a park service will leave behind evaluation text and rating data; and a user submitting a repair request on the website will record information such as the equipment being repaired, the description of the fault, and the submission time.

[0024] To acquire this real-time operational data, a data acquisition system needs to be established. This system connects to various operational data sources via network interfaces, acquiring data periodically or in real-time. For equipment operation log data, the data acquisition system can interface with the equipment's monitoring system, using data interface protocols (such as Modbus, OPC, etc.) to acquire the data. For environmental monitoring text data, the data acquisition system can communicate with environmental monitoring equipment via wireless communication technologies (such as ZigBee, WiFi, etc.) to receive data sent by the equipment. For user interaction data, the data acquisition system can extract data from the user interaction platform's database. The acquired data is then uniformly stored in a data warehouse, forming a real-time operational data set.

[0025] Step S120: Extract semantic features from the real-time operation and maintenance data set to generate a semantic feature set, which includes semantic features of device operating status, semantic features of environment association, and semantic features of user interaction intent.

[0026] Semantic feature extraction from real-time operations and maintenance (O&M) datasets aims to extract valuable information from the raw data for subsequent anomaly analysis and strategy formulation. Different methods are employed for semantic feature extraction depending on the type of real-time O&M data.

[0027] Step S121: Perform text cleaning processing on the equipment operation log data to remove redundant symbols and unstructured data segments, and obtain standardized equipment operation text data.

[0028] Equipment operation log data typically contains a large amount of redundant information and unstructured content, such as special symbols, useless comments, and formatting tags. This information can interfere with subsequent semantic analysis, thus requiring text cleaning. Taking elevator operation log data as an example, the raw data may contain special symbols used to separate fields, such as commas and semicolons, as well as comments recording equipment debugging information. During text cleaning, a set of cleaning rules must first be defined, containing the symbols to be removed and the patterns of data segments. For example, for special symbols, regular expressions can be used to match and replace them with spaces or empty strings. For unstructured data segments, judgments can be made based on the data's format and content, deleting segments that do not conform to the rules. The cleaned equipment operation log data will become more standardized and cleaner, facilitating subsequent semantic encoding processing.

[0029] Step S122: Perform contextual semantic encoding on the standardized equipment operation text data, extract the temporal correlation semantic features from the equipment operation log data, and match the temporal correlation semantic features with a preset equipment operation status keyword library to generate the equipment operation status semantic features.

[0030] Step S1221: Divide the standardized equipment operation text data into multiple equipment operation text segments in chronological order.

[0031] Standardized equipment operation text data is recorded in chronological order. To better analyze the temporal correlation information within the data, it needs to be divided into multiple equipment operation text segments. The division can be based on time intervals; for example, hourly equipment operation log data can be divided into one text segment. Taking power equipment operation log data as an example, dividing it according to hourly time intervals, each text segment contains the changes in various operating parameters of the power equipment within that hour, such as voltage, current, and power. This division allows each text segment to have relatively independent temporal information, facilitating subsequent semantic encoding processing.

[0032] Step S1222: Call the pre-trained context semantic encoder to encode the text segments running on each device, and generate the context semantic vector of each text segment running on the device.

[0033] A pre-trained contextual semantic encoder is a deep learning-based model that transforms text data into vector representations while considering the contextual information of the text. In this embodiment, a pre-trained BERT (Bidirectional Encoder Representations from Transformers) model is used as the contextual semantic encoder. The BERT model is pre-trained on large-scale text data, enabling it to learn rich linguistic knowledge and semantic information. Each device operation text segment is input into the BERT model, which encodes the text and outputs a fixed-length contextual semantic vector. This contextual semantic vector represents the semantic information of the text segment and takes into account the contextual relationships within the text. For example, for a text segment in elevator operation log data, the BERT model transforms it into a high-dimensional vector, where each dimension of the contextual semantic vector represents a semantic feature of that text segment.

[0034] Step S1223: Perform sliding window aggregation processing on the context semantic vectors of adjacent time periods to obtain the temporal association semantic features; wherein, the sliding window aggregation processing includes: performing weighted summation on multiple context semantic vectors within the current time window, and dynamically adjusting the weight values ​​according to the temporal distance between each context semantic vector and the center point of the current time window.

[0035] To extract time-series semantic features from device operation log data, a sliding window aggregation process is needed to aggregate the context semantic vectors of adjacent time periods. The sliding window is a fixed-size time window that slides along the time axis, covering a set number of context semantic vectors each time. For example, setting the sliding window size to 3 time units means that each window will cover the context semantic vectors of 3 adjacent time periods. Within each time window, the multiple context semantic vectors within the window are summed using weights. The weights are dynamically adjusted based on the temporal distance of each context semantic vector from the center point of the current time window. Context semantic vectors closer to the center point have larger weights, while those farther away have smaller weights. This highlights key information within the current time window while also considering the correlation information between adjacent time periods. Through sliding window aggregation, the resulting time-series semantic features reflect the changes in the device's operating status over time.

[0036] Step S1224: Match the temporal correlation semantic features with a preset device operating status keyword library to generate the device operating status semantic features.

[0037] The pre-defined keyword library for equipment operating status is a collection of keywords related to various equipment operating states. For example, for elevator equipment, the keyword library might include keywords such as "normal operation," "fault alarm," and "maintenance"; for electrical equipment, it might include keywords such as "overload," "undervoltage," and "short circuit." The extracted temporal-related semantic features are matched against the keyword library to determine the equipment operating state corresponding to each temporal-related semantic feature. Similarity calculation methods can be used for matching; for example, the similarity between the temporal-related semantic feature vector and the semantic vector of each keyword in the keyword library is calculated, and the keyword with the highest similarity is selected as the equipment operating state corresponding to that temporal-related semantic feature. Through this matching operation, the temporal-related semantic features are transformed into equipment operating state semantic features, facilitating subsequent anomaly analysis and strategy formulation.

[0038] Step S123: Perform word segmentation on the environmental monitoring text data to obtain multiple environmental monitoring word units. Perform semantic clustering on the multiple environmental monitoring word units based on a preset environmental association rule library to generate the environmental association semantic features.

[0039] Step S1231: Perform word segmentation on the environmental monitoring text data to obtain multiple environmental monitoring word units.

[0040] Environmental monitoring text data typically consists of words and sentences describing environmental parameters. To perform semantic analysis on this data, word segmentation is necessary, breaking the text down into multiple independent word units. In this embodiment, a Chinese word segmentation tool (such as jieba) is used to segment the environmental monitoring text data. For example, for the environmental monitoring text "Today the temperature in the park is 25 degrees Celsius, and the air quality is good," word segmentation yields environmental monitoring word units such as "today," "inside the park," "temperature," "25 degrees Celsius," "air quality," and "good." These environmental monitoring word units form the basis for subsequent semantic clustering processing.

[0041] Step S1232: Match the environmental entity category and environmental attribute category corresponding to each environmental monitoring word unit from the environmental association rule base.

[0042] The pre-defined environmental association rule base is a collection of rules that maps environmental monitoring terms to environmental entity categories and environmental attribute categories. Environmental entity categories refer to various entities in the environment, such as parks, buildings, and rooms; environmental attribute categories refer to various attributes of environmental entities, such as temperature, humidity, and air quality. For example, the rule base specifies that the environmental entity category corresponding to "temperature" is "park," and the environmental attribute category is "temperature"; similarly, the environmental entity category corresponding to "air quality" is "park," and the environmental attribute category is "air quality." By matching within the environmental association rule base, the corresponding environmental entity category and environmental attribute category can be determined for each environmental monitoring term.

[0043] Step S1233: Aggregate environmental monitoring word units belonging to the same environmental entity category into an environmental entity set, and extract the environmental attribute numerical distribution features of the environmental attribute category in each environmental entity set.

[0044] Based on the matched environmental entity categories, environmental monitoring terminology belonging to the same category is aggregated into a single environmental entity set. For example, all environmental monitoring terminology related to the park environment is aggregated into a single park environmental entity set. For each environmental entity set, the distribution characteristics of environmental attribute values ​​within its categories are extracted. Taking the park environmental entity set as an example, if the set contains multiple monitoring terminology related to temperature, such as "25 degrees Celsius," "26 degrees Celsius," and "24 degrees Celsius," the distribution characteristics of these temperature values, such as mean, standard deviation, maximum, and minimum, can be calculated. These distribution characteristics reflect the changes and stability of environmental attributes.

[0045] Step S1234: Standardize the distribution characteristics of environmental attribute values ​​for each environmental attribute category in each set of environmental entities. Based on the preset environmental parameter weight allocation table, perform weighted summation on the standardized environmental attribute value distribution characteristics to generate a comprehensive impact score.

[0046] To comprehensively assess the environmental attributes of different sets of environmental entities, it is necessary to standardize the distribution characteristics of environmental attribute values. Standardization transforms environmental attribute values ​​with different dimensions and ranges into a unified standard range, facilitating subsequent weighted summation calculations. Commonly used standardization methods include Z-score standardization and Min-Max standardization. Taking Z-score standardization as an example, for each environmental attribute value distribution characteristic, its mean and standard deviation are calculated. Then, the mean is subtracted from each characteristic value and divided by the standard deviation to obtain the standardized characteristic value. A preset environmental parameter weighting table contains weight values ​​corresponding to different environmental attribute categories. For example, for the park environment, the weight value for temperature might be 0.4, and the weight value for air quality might be 0.6. Based on the weighting table, the standardized environmental attribute value distribution characteristics are weighted and summed to obtain the comprehensive impact score for each set of environmental entities. This comprehensive impact score reflects the degree of impact of the set of environmental entities on the overall environment of the park.

[0047] Step S1235: Mark the set of environmental entities whose comprehensive impact score is greater than a preset threshold as key environmental entities, and combine the environmental attribute distribution features corresponding to the key environmental entities to generate the environmental association semantic features.

[0048] The preset threshold is a pre-defined scoring standard used to determine whether a set of environmental entities has a significant impact on the park's environment. Environmental entities with a comprehensive impact score greater than the preset threshold are marked as key environmental entities. For example, if the preset threshold is 0.7, and a set of park environmental entities has a comprehensive impact score greater than 0.7, it is marked as a key environmental entity. The environmental attribute distribution features corresponding to these key environmental entities are combined to generate environmental association semantic features. These environmental association semantic features can reflect the key factors within the park that have a significant impact on the environment and the relationships between them.

[0049] Step S124: Call the preset intent recognition model to perform intent classification processing on the user interaction data and generate the semantic features of the user interaction intent.

[0050] The pre-defined intent recognition model is a machine learning or deep learning-based model used to classify user interaction data into intents. In this embodiment, a convolutional neural network (CNN)-based intent recognition model is used. This intent recognition model has been trained on a large amount of labeled data (user interaction data and intent labels) and is able to learn the feature patterns of different user interaction intents.

[0051] User interaction data is input into an intent recognition model, which processes and analyzes the data, outputting classification results of user interaction intents. For example, for a user's query request on the park's app, the intent recognition model can determine whether the user's intent is to query parking space availability, park event information, or property service information. For user ratings and feedback data, the intent recognition model can determine whether the user's intent is to express satisfaction, dissatisfaction, or to offer suggestions. These classification results are then transformed into semantic features of user interaction intents for subsequent anomaly analysis and strategy formulation.

[0052] Step S130: Perform dynamic anomaly analysis on the semantic feature set based on the preset AI autonomous learning model to obtain an anomaly feature set.

[0053] The pre-defined AI self-learning model is a model used for dynamic anomaly analysis of semantic feature sets. This model can automatically learn and adapt to different anomaly patterns, and promptly detect anomalies in park operations and maintenance.

[0054] Step S131: Input the device operating status semantic features, environment association semantic features, and user interaction intent semantic features into the multi-branch feature extraction layer of the preset AI autonomous learning model, wherein: Step S1311: The first branch feature extraction layer uses a temporal convolutional network to perform sliding window feature extraction on the temporal dimension of the semantic features of the device's operating state, generating a device state feature vector.

[0055] A Temporal Convolutional Network (TCN) is a convolutional neural network specifically designed for processing temporal data. In the first branch, the feature extraction layer, TCN uses a sliding window to extract temporal features of the device's operational state semantics. The sliding window moves along the temporal dimension, covering a predetermined number of time steps each time. Within each window, TCN extracts local features of the device's operational state semantics through convolution operations. These local features reflect the changes in the device's operational state at different points in time. By concatenating and aggregating the features from all windows, a device state feature vector is generated. This device state feature vector reflects the overall temporal characteristics of the device's operational state.

[0056] Step S1312: The second branch feature extraction layer uses a graph convolutional network to perform neighborhood propagation processing on the environmental associated semantic features based on the spatial topology of the environmental monitoring nodes, generating an environmental spatial feature vector.

[0057] Graph Convolutional Networks (GCNs) are a type of neural network used to process graph-structured data. In a park environmental monitoring system, environmental monitoring nodes (such as sensors) have a certain spatial topology, which can be represented by a graph. In the second branch feature extraction layer, GCN is used to perform neighborhood propagation processing on the environmental association semantic features based on the spatial topology of the environmental monitoring nodes. GCN fuses the environmental association semantic features of each node with the features of its neighboring nodes through convolution operations, thereby capturing the spatial correlation information of environmental parameters. Through multiple neighborhood propagations, an environmental spatial feature vector is generated. This environmental spatial feature vector can reflect the overall spatial characteristics of the park environment.

[0058] Step S1313: The third branch feature extraction layer uses a multi-head attention mechanism to perform cross-dimensional correlation analysis on the semantic features of the user interaction intent, and generates an intent-focused feature vector.

[0059] Multi-head attention (MFA) is a mechanism used to capture the correlations between different positions in sequential data. In the third branch feature extraction layer, MFA is used to perform cross-dimensional correlation analysis on the semantic features of user interaction intent. MFA captures the relationships between different intents by simultaneously focusing on different dimensions of the semantic features of user interaction intent with multiple attention heads. For example, for user query requests and evaluation feedback data, MFA can analyze the potential connections between them. By concatenating and aggregating the outputs of MFA, an intent-focused feature vector is generated. This intent-focused feature vector reflects the overall correlation features of the user interaction intent.

[0060] Step S132: The device state feature vector, environmental space feature vector, and intent focus feature vector are mapped to a unified feature space according to a preset feature projection matrix and then fused to generate a fused feature vector; wherein, the feature projection matrix is ​​optimized by minimizing the cosine similarity difference of the feature vectors output by each branch feature extraction layer during the model training phase.

[0061] The pre-defined feature projection matrix is ​​a matrix used to map the feature vectors output by different branch feature extraction layers to a unified feature space. During model training, the parameters of the feature projection matrix are optimized by minimizing the cosine similarity difference between the feature vectors output by each branch feature extraction layer. Specifically, the cosine similarity between the feature vectors of each branch is calculated, and then the parameters of the feature projection matrix are adjusted to make these similarities as close as possible. This ensures that feature vectors of different types have similar distributions and representational capabilities in the unified feature space.

[0062] The device status feature vector, environmental space feature vector, and intent focus feature vector are each multiplied by the feature projection matrix to map them to a unified feature space. Then, a fusion operation is performed on the mapped feature vectors. This can be achieved by concatenating them into a longer vector, generating a fused feature vector. This fused feature vector integrates information from multiple aspects, including device operating status, environmental space, and user interaction intent, facilitating subsequent anomaly classification.

[0063] Step S133: Input the fused feature vector into the fully connected classification layer of the AI ​​autonomous learning model and output a comprehensive anomaly score sequence; wherein, the training data of the fully connected classification layer includes historical anomaly event annotation data and normalized vectors of corresponding semantic features. During the training process, the classification weight parameters of the focus loss function are dynamically adjusted according to the distribution ratio of anomaly categories in the current batch of training data to optimize the classification boundary.

[0064] The fully connected classification layer is the final layer of the AI ​​autonomous learning model, used to classify the fused feature vectors and output a comprehensive anomaly score sequence. The training data for this fully connected classification layer includes historical anomaly event annotation data and normalized vectors of corresponding semantic features. Historical anomaly event annotation data refers to records of past anomaly events and their anomaly type labels. The normalized vectors of corresponding semantic features are vectors obtained by normalizing the semantic features corresponding to the historical anomaly events.

[0065] During training, a focal loss function is used to optimize the parameters of the fully connected classification layer. The focal loss function addresses the data imbalance problem, where the number of samples in the anomaly category is typically much smaller than the number of samples in the normal category. The classification weights of the focal loss function are dynamically adjusted based on the distribution ratio of anomaly categories in the current batch of training data. When the number of anomaly samples is small, the weights for anomaly categories are increased, causing the model to focus more on classifying anomaly samples. By continuously adjusting the classification weights, the classification boundary is optimized, improving the model's accuracy in classifying anomalies.

[0066] After the fused feature vector is input into the fully connected classification layer, the neurons within the layer perform linear transformations and non-linear activation operations on the feature vector. The linear transformation multiplies the input feature vector by the neuron's weight matrix and adds a bias term. The non-linear activation operation applies an activation function, such as the ReLU function, to the result of the linear transformation to introduce non-linearity and enhance the model's expressive power. After multiple such transformations and activation operations, a comprehensive anomaly score sequence is finally output. Each score value in this comprehensive anomaly score sequence represents the probability of an anomaly for the corresponding time point or sample.

[0067] Step S134: After aligning the timestamps of the comprehensive anomaly score sequence with the real-time judgment threshold output by the dynamic threshold generation module, perform point-by-point comparison for anomaly judgment, and extract feature segments whose comprehensive anomaly scores exceed the real-time judgment threshold as candidate anomaly features.

[0068] The dynamic threshold generation module is designed to dynamically generate real-time judgment thresholds based on the park's real-time operational status. This module considers multiple factors, such as equipment operating load rate, historical anomaly frequency, and changes in environmental parameters.

[0069] First, the comprehensive anomaly score sequence output by the AI ​​self-learning model is collected in real time within a preset time window. The preset time window can be set according to actual conditions, such as one day or one week. Then, the mean and standard deviation of the comprehensive anomaly score sequence are calculated. The mean reflects the average level of the score sequence, while the standard deviation reflects the degree of fluctuation in the score sequence.

[0070] Next, the equipment operating load rate is converted into a normalized value within the range of 0 to 1. The equipment operating load rate refers to the ratio of the actual operating load of the equipment to its rated load. Normalization unifies the load rates of different devices to the same range, facilitating subsequent calculations. A baseline threshold is generated based on a linear combination of the mean and standard deviation, where the coefficient of the standard deviation term is positively correlated with the normalized equipment operating load rate. That is, when the equipment operating load rate is high, the coefficient of the standard deviation term increases, making the baseline threshold more flexible in adapting to the equipment's operating status.

[0071] Then, a preset baseline offset is superimposed to form the initial real-time judgment threshold. The baseline offset is a preset offset value used to adjust the initial position of the threshold. This offset is adjusted exponentially based on the frequency of anomalies occurring in the same historical period. If the frequency of anomalies occurring in the same historical period is high, the baseline offset will be increased accordingly to improve the sensitivity of the threshold; conversely, if the frequency of anomalies is low, the baseline offset will be decreased to reduce the sensitivity of the threshold.

[0072] When a sudden change in a parameter within the semantic features associated with the environment is detected, a dynamic threshold compensation mechanism is activated. The threshold offset is calculated using a preset logarithmic function coefficient based on the ratio of the rate of change of the environmental parameter to a safety threshold. The rate of change of the environmental parameter refers to the magnitude of change of the environmental parameter within a certain time period, while the safety threshold is the normal range of the environmental parameter. By calculating the threshold offset, the initial real-time judgment threshold is adjusted downwards to form the final real-time judgment threshold for application. This allows for timely adjustment of the threshold when sudden changes occur in environmental parameters, improving the accuracy of anomaly detection.

[0073] The timestamps of the comprehensive anomaly scoring sequence are time-aligned with the real-time judgment thresholds output by the dynamic threshold generation module to ensure temporal correspondence between the scoring sequence and the thresholds. Then, a point-by-point anomaly judgment is performed, comparing each score value in the scoring sequence with its corresponding real-time judgment threshold. If a score value exceeds the real-time judgment threshold, the feature segment corresponding to that score value is extracted as a candidate anomaly feature.

[0074] Step S135: Perform contextual verification on the candidate abnormal features: Based on the temporal continuity of the device operating status semantic features, the spatial distribution consistency of the environment-related semantic features, and the logical correlation of the user interaction intent semantic features, eliminate false detection feature fragments and generate the final abnormal feature set.

[0075] For the extracted candidate abnormal features, contextual verification is required to eliminate possible false detection feature fragments.

[0076] Verification is based on the temporal continuity of semantic features of equipment operating status. Equipment operating status typically exhibits a certain degree of temporal continuity, meaning that the operating status at adjacent time points should show similarities or gradual changes. If a candidate anomaly feature is discontinuous in time with the semantic features of the equipment operating status before and after it, showing a sudden jump or abnormal change without a reasonable explanation, then that feature may be a false positive. For example, the normal operating status of equipment gradually increases or decreases, but if a candidate anomaly feature shows that the equipment status suddenly changes from normal to abnormal, and there is no corresponding equipment fault record or operation record to support it, then that candidate anomaly feature may be a false positive.

[0077] Verification is based on the spatial distribution consistency of environmental semantic features. Environmental parameters within the park exhibit certain spatial distribution patterns and correlations. For example, environmental parameters such as temperature and humidity in adjacent areas should be similar. If the environmental parameter corresponding to a candidate anomaly differs significantly from the surrounding environmental parameters spatially and does not conform to normal environmental change patterns, then this feature may be a false positive. For instance, in a relatively enclosed park, if the air quality at a certain monitoring point suddenly becomes abnormal, while the air quality at other surrounding monitoring points is normal and there is no obvious source of pollution, then this candidate anomaly feature may be a false positive.

[0078] Validation is based on the logical correlation of semantic features of user interaction intent. User interaction intents typically exhibit a certain degree of logical coherence. If the user interaction intent corresponding to a candidate anomaly feature is unrelated to or logically contradictory to the preceding and following interaction intents, then that feature may be a false positive. For example, if a user submits a request to inquire about park activities within a short period, and then immediately submits an anomaly feature expressing dissatisfaction with the park's parking services, but there is no obvious logical connection between the two, then this candidate anomaly feature may be a false positive.

[0079] Through the above contextual verification, false positive feature fragments are eliminated, and the remaining candidate anomaly features are combined to generate the final anomaly feature set. The anomaly features in this set have undergone rigorous screening and verification, and can accurately reflect the real anomalies in park operations and maintenance.

[0080] Step S140: Perform strategy matching processing based on the set of abnormal features and the preset optimization strategy library to generate a target operation and maintenance optimization strategy set.

[0081] The pre-defined optimization strategy library is a collection of various operation and maintenance optimization strategies for different anomaly situations. Each operation and maintenance optimization strategy has a corresponding pre-defined anomaly trigger feature vector, which is generated within the unified feature space of the AI ​​self-learning model.

[0082] Step S141: Extract the preset anomaly trigger feature vector corresponding to each operation and maintenance optimization strategy from the optimization strategy library, wherein the preset anomaly trigger feature vector is generated in the unified feature space of the AI ​​autonomous learning model.

[0083] During the model training phase, based on historical anomaly events and corresponding optimization strategies, anomaly triggering features corresponding to each optimization strategy are extracted. These features are then mapped into the unified feature space of the AI ​​autonomous learning model to generate pre-defined anomaly triggering feature vectors. These vectors have the same dimension and feature representation, facilitating subsequent similarity calculations.

[0084] Step S142: For each abnormal feature vector in the abnormal feature set, calculate its similarity with the preset abnormal trigger feature vectors of all operation and maintenance optimization strategies. When the similarity value between the preset abnormal trigger feature vector of any operation and maintenance optimization strategy and the corresponding type of abnormal feature vector exceeds the preset strategy trigger threshold, add the operation and maintenance optimization strategy to the candidate strategy set.

[0085] For each anomalous feature vector in the anomalous feature set, a similarity calculation method (such as cosine similarity) is used to calculate its similarity with each preset anomalous trigger feature vector in the optimization strategy library. Cosine similarity measures the similarity between two vectors by calculating the cosine of the angle between them; the closer the value is to 1, the higher the similarity.

[0086] The preset strategy trigger threshold is a pre-defined similarity standard used to determine whether a certain operation and maintenance optimization strategy is applicable to the current abnormal situation. When the similarity value between the preset abnormality trigger feature vector of a certain operation and maintenance optimization strategy and the feature vector of the corresponding type of abnormality exceeds the preset strategy trigger threshold, it indicates that the strategy may be effective in resolving the current abnormal situation, and the operation and maintenance optimization strategy is added to the candidate strategy set.

[0087] Step S143: Prioritize each operation and maintenance optimization strategy in the candidate strategy set, and select the top N operation and maintenance optimization strategies to generate the target operation and maintenance optimization strategy set, wherein N is dynamically set according to the concurrent processing capability of the current operation and maintenance management platform.

[0088] Prioritize the various operation and maintenance optimization strategies in the candidate strategy set to determine which strategies should be executed first.

[0089] Step S1431: Obtain the historical execution effect score for each operation and maintenance optimization strategy. The historical execution effect score is calculated by weighting the deviation reduction rate in the equipment status response data after the operation and maintenance optimization strategy is executed, the safety threshold achievement rate of the environmental parameter change data, and the improvement of user satisfaction feedback data, and normalized to the range of 0 to 1.

[0090] Each operation and maintenance optimization strategy generates corresponding effect data after execution, including equipment status response data, environmental parameter change data, and user satisfaction feedback data. Equipment status response data reflects changes in the equipment's operational status after strategy execution; the improvement effect of the strategy on equipment operation is evaluated by calculating the rate of reduction in equipment status deviation. Environmental parameter change data reflects changes in environmental parameters after strategy execution; the improvement effect of the strategy on the environment is evaluated by calculating the rate of achievement of safe thresholds for environmental parameters. User satisfaction feedback data reflects user evaluation of the strategy's effectiveness; the improvement effect of the strategy on user experience is evaluated by calculating the increase in user satisfaction.

[0091] Based on a preset weighting, the deviation reduction rate in equipment status response data, the safety threshold achievement rate in environmental parameter change data, and the improvement in user satisfaction feedback data are weighted and calculated to obtain a historical execution effect score for each operation and maintenance optimization strategy. These scores are then normalized to a range of 0 to 1 for subsequent comparison and ranking.

[0092] Step S1432: Extract the dynamic weight coefficients of each abnormal feature vector in the current abnormal feature set. The dynamic weight coefficients are normalized according to the proportion of the abnormal score value of the corresponding feature segment in the comprehensive abnormal score sequence in the total abnormal score.

[0093] Each anomalous feature vector in the current set of anomalous features corresponds to a feature segment in the comprehensive anomalous score sequence, and its anomalous score reflects the severity of the anomaly. By calculating the proportion of each anomalous score in the total anomalous score and performing normalization, the dynamic weight coefficient of each anomalous feature vector is obtained. The larger the dynamic weight coefficient, the more severe the anomalous situation corresponding to that feature, and the higher its priority for processing.

[0094] Step S1433: Multiply the historical execution effect score with the dynamic weight coefficient of the current abnormal feature vector of the corresponding operation and maintenance optimization strategy to generate the strategy priority index.

[0095] The strategy priority index is obtained by multiplying the historical performance score of each operation and maintenance optimization strategy by the dynamic weight coefficient of the corresponding anomaly feature vector. The strategy priority index comprehensively considers the historical performance of the strategy and the severity of the current anomaly; the higher the index, the higher the priority of the strategy.

[0096] Step S1434: Sort the candidate strategy set according to the strategy priority index from high to low.

[0097] The operation and maintenance optimization strategies in the candidate strategy set are sorted from highest to lowest according to the strategy priority index. After sorting, the top N operation and maintenance optimization strategies are selected to generate the target operation and maintenance optimization strategy set. The value of N is dynamically set according to the concurrent processing capability of the current operation and maintenance management platform. The concurrent processing capability of the operation and maintenance management platform refers to the maximum number of strategies that the platform can execute at the same time. If the platform has a strong concurrent processing capability, more strategies can be selected; if the concurrent processing capability is weak, fewer strategies are selected to ensure that the platform can execute strategies stably.

[0098] Step S150: Execute each operation and maintenance optimization strategy in the target operation and maintenance optimization strategy set, and feed back the execution results to the operation and maintenance management platform of the smart park.

[0099] Step S151: Parse the operation instruction sequence in the device operation optimization strategy, generate a set of device control instructions, and send them to the corresponding target device controller.

[0100] The equipment operation optimization strategy contains a series of operation instruction sequences used to control the equipment's operating state. These operation instruction sequences are parsed and converted into specific equipment control instructions. For example, if the strategy requires adjusting the elevator's operating speed to a certain value, the parsed result will generate corresponding speed control instructions. These sets of equipment control instructions are then sent to the corresponding target equipment controller, which will control the equipment according to the instructions.

[0101] Step S152: Monitor the device status response data returned by the target device controller. If the device status response data matches the expected optimization effect characteristics, mark the device operation optimization strategy as successfully executed. Step S1521: Extract real-time equipment operation characteristics from the equipment status response data, perform standardization processing, and calculate the deviation between the characteristics and the expected equipment operation characteristics; After executing equipment control commands, the target equipment controller returns equipment status response data. Real-time equipment operating characteristics, such as operating speed, temperature, and pressure, are extracted from this data. These real-time operating characteristics are then standardized, transforming them into a unified standard range for easier subsequent comparisons and calculations.

[0102] Expected equipment operating characteristics are target characteristics set when formulating equipment operation optimization strategies, representing the ideal operating state that the equipment should achieve after executing the strategies. The deviation between the standardized real-time equipment operating characteristics and the expected equipment operating characteristics is calculated. The deviation can be measured by calculating the distance between two feature vectors (such as Euclidean distance).

[0103] Step S1522: If the deviation is continuously lower than the dynamic threshold for more than a preset time window, the device operation optimization strategy is determined to be successfully executed; wherein, the dynamic threshold is adaptively adjusted according to the device type and historical optimization effect data.

[0104] A dynamic threshold is a threshold that adaptively adjusts based on device type and historical optimization performance data. Different types of devices have different operating characteristics and optimization requirements, so the dynamic threshold is set according to the device type. Meanwhile, historical optimization performance data reflects the device's performance when executing optimization strategies in the past; this data can be used to adjust the dynamic threshold to better reflect the actual situation of the device.

[0105] If the deviation between the real-time device operating characteristics and the expected device operating characteristics remains below the dynamic threshold for a period exceeding a preset time window, it indicates that the device's operating state has stably approached the expected target, and the device operation optimization strategy is deemed to have been successfully executed. The preset time window is a pre-defined time range used to ensure the stability of the device's operating state and avoid misjudgments due to short-term fluctuations.

[0106] Step S153: Analyze the set of environmental adjustment instructions in the environmental monitoring optimization strategy, start the environmental control equipment and collect environmental parameter change data in real time; The environmental monitoring optimization strategy includes a set of environmental adjustment instructions. These instructions control environmental control equipment, such as air conditioners, humidifiers, and air purifiers, to adjust the park's environmental parameters. This set of environmental adjustment instructions is parsed and translated into specific equipment control commands. The corresponding environmental control equipment is then activated, and real-time data on changes in environmental parameters, such as temperature, humidity, and air quality, are collected.

[0107] Step S154: When the environmental parameter change data reaches the preset safety threshold range, mark the environmental monitoring optimization strategy as completed; The preset safety threshold range represents the normal range of environmental parameters, indicating a suitable state for the park environment. When the real-time collected environmental parameter change data reaches the preset safety threshold range, it indicates that the environmental control equipment has successfully adjusted the environmental parameters to a suitable range, marking the environmental monitoring optimization strategy as completed.

[0108] Step S155: Analyze the interaction logic update instructions in the user interaction optimization strategy, modify the response logic of the user interaction interface, and collect user satisfaction feedback data. User interaction optimization strategies include interaction logic update instructions, which modify the response logic of the user interface to improve user experience. These instructions are parsed to modify the user interface code and update its response logic. After the update, user satisfaction feedback data is collected to understand user evaluation of the new interaction logic.

[0109] Step S156: When the user satisfaction feedback data exceeds the preset satisfaction threshold, mark the user interaction optimization strategy as effective. The preset satisfaction threshold is a pre-defined standard used to determine whether users are satisfied with the new interaction logic. When the collected user satisfaction feedback data exceeds the preset satisfaction threshold, it indicates that users are relatively satisfied with the new interaction logic, and the user interaction optimization strategy is marked as effective.

[0110] Finally, the execution results of the equipment operation optimization strategy, environmental monitoring optimization strategy, and user interaction optimization strategy are fed back to the smart park's operation and maintenance management platform. The platform can monitor and evaluate the park's operation and maintenance status based on this feedback, promptly identifying and adjusting any issues. Simultaneously, this feedback can also be used to update and optimize the optimization strategy library, improving the effectiveness of subsequent strategy implementation.

[0111] It is worth noting that in the above embodiments, The data collection and processing process involves sensitive user data, such as personal information and interaction logs. To protect user privacy, a series of privacy protection and leak prevention measures are necessary.

[0112] For example, during the data collection phase, users' privacy-sensitive data can be encrypted. Symmetric encryption algorithms (such as AES) or asymmetric encryption algorithms (such as RSA) can be used to encrypt the data. The encrypted data remains in ciphertext form during transmission and storage, and only authorized personnel can decrypt and use it.

[0113] For example, establish a strict access control mechanism to manage data access permissions. Only authorized personnel should be able to access and process users' sensitive privacy data. A role-based access control system can be used to assign different access permissions to different user roles.

[0114] For example, during data processing, sensitive user data can be anonymized. Data anonymization techniques can be used to replace users' personal information (such as names, ID numbers, etc.) with anonymous identifiers, making it impossible to directly associate the data with specific users during processing. For instance, a user's name can be replaced with a randomly generated number. When user information is needed, a lookup table mapping the number to the real information can be used, but this mapping table also needs to be strictly encrypted and access controlled.

[0115] Furthermore, users' privacy-sensitive data can be backed up regularly and stored in a secure location. Backup data also needs to be encrypted to prevent data leakage. Simultaneously, a data recovery mechanism should be established to promptly recover data in case of loss or corruption, ensuring data security and integrity. A security audit system should also be established to audit the data access and processing processes. All access operations to privacy-sensitive data should be recorded, including access time, access personnel, and access content. Analysis of audit logs can promptly identify abnormal access behavior and take appropriate measures, such as restricting access permissions or conducting security investigations. This ensures that the data collection, storage, and processing processes comply with relevant laws, regulations, and industry standards.

[0116] Figure 2 The illustration shows exemplary hardware and software components of an AI-based autonomous learning smart park operation and maintenance management system 100, which can implement the ideas of this application, according to some embodiments of this application. For example, a processor 120 can be used in the AI-based autonomous learning smart park operation and maintenance management system 100 and to perform the functions in this application.

[0117] The AI-based autonomous learning-based smart park operation and maintenance management system 100 can be a general-purpose server or a special-purpose server; both can be used to implement the AI-based autonomous learning-based smart park operation and maintenance management method of this application. Although only one server is shown in this application, for convenience, the functions described in this application can be implemented in a distributed manner on multiple similar platforms to balance the load.

[0118] For example, the AI-based autonomous learning smart park operation and maintenance management system 100 may include a network port 110 connected to a network, one or more processors 120 for executing program instructions, a communication bus 130, and various forms of storage media 140, such as a disk, ROM, or RAM, or any combination thereof. Exemplarily, the AI-based autonomous learning smart park operation and maintenance management system 100 may also include program instructions stored in ROM, RAM, or other types of non-transitory storage media, or any combination thereof. The methods of this application can be implemented according to these program instructions. The AI-based autonomous learning smart park operation and maintenance management system 100 also includes an I / O interface 150 between the computer and other input / output devices.

[0119] For ease of explanation, only one processor is described in the AI-based autonomous learning smart park operation and maintenance management system 100. However, it should be noted that the AI-based autonomous learning smart park operation and maintenance management system 100 of this application may also include multiple processors. Therefore, the steps executed by one processor as described in this application may also be executed jointly by multiple processors or individually. For example, if the processor of the AI-based autonomous learning smart park operation and maintenance management system 100 executes steps A and B, it should be understood that steps A and B may also be executed jointly by two different processors or individually by one processor. For example, the first processor executes step A, the second processor executes step B, or the first processor and the second processor jointly execute steps A and B.

[0120] Furthermore, this embodiment of the invention also provides a readable storage medium, wherein computer-executable instructions are preset in the readable storage medium, and when the processor executes the computer-executable instructions, the above-mentioned smart park operation and maintenance management method based on AI autonomous learning is realized.

[0121] It should be noted that, in order to simplify the description of the present invention and thus help to understand one or more embodiments of the invention, multiple features may sometimes be grouped into one embodiment, drawing or description thereof in the foregoing description of the embodiments of the present invention.

Claims

1. A smart park operation and maintenance management method based on AI autonomous learning, characterized in that, The method includes: Acquire a real-time operation and maintenance data set from multiple operation and maintenance data sources within the smart park. The real-time operation and maintenance data set includes equipment operation log data, environmental monitoring text data, and user interaction data. Semantic features are extracted from the real-time operation and maintenance data set to generate a semantic feature set, which includes semantic features of device operating status, semantic features of environment association, and semantic features of user interaction intent. Based on a preset AI autonomous learning model, dynamic anomaly analysis is performed on the semantic feature set to obtain an anomaly feature set. Based on the set of abnormal features, a strategy matching process is performed with a preset optimization strategy library to generate a target set of operation and maintenance optimization strategies; Execute each operation and maintenance optimization strategy in the target operation and maintenance optimization strategy set, and feed back the execution results to the operation and maintenance management platform of the smart park.

2. The smart park operation and maintenance management method based on AI autonomous learning according to claim 1, characterized in that, The step of extracting semantic features from the real-time operation and maintenance data set to generate a semantic feature set includes: The equipment operation log data is cleaned by removing redundant symbols and unstructured data segments to obtain standardized equipment operation text data. The standardized equipment operation text data is subjected to contextual semantic encoding processing, the temporal correlation semantic features in the equipment operation log data are extracted, and the temporal correlation semantic features are matched with a preset equipment operation status keyword library to generate the equipment operation status semantic features. The environmental monitoring text data is segmented to obtain multiple environmental monitoring word units. Based on a preset environmental association rule library, the multiple environmental monitoring word units are semantically clustered to generate the environmental association semantic features. A preset intent recognition model is invoked to perform intent classification processing on the user interaction data, generating semantic features of the user interaction intent.

3. The smart park operation and maintenance management method based on AI autonomous learning according to claim 2, characterized in that, The step of performing contextual semantic encoding on the standardized equipment operation text data and extracting time-series related semantic features from the equipment operation log data includes: The standardized equipment operation text data is divided into multiple equipment operation text fragments in chronological order; The pre-trained context semantic encoder is invoked to encode the text segments running on each device, generating a context semantic vector for each text segment running on each device. The context semantic vectors of adjacent time periods are aggregated using a sliding window to obtain the temporal association semantic features; wherein, the sliding window aggregation process includes: weighted summation of multiple context semantic vectors within the current time window, with the weight values ​​dynamically adjusted according to the temporal distance between each context semantic vector and the center point of the current time window.

4. The smart park operation and maintenance management method based on AI autonomous learning according to claim 2, characterized in that, The process of semantic clustering of the multiple environmental monitoring word units based on a preset environmental association rule base to generate the environmental association semantic features includes: Match the environmental entity category and environmental attribute category corresponding to each environmental monitoring term unit from the environmental association rule base; Environmental monitoring terminology units belonging to the same environmental entity category are aggregated into environmental entity sets, and the distribution characteristics of environmental attribute values ​​for each environmental attribute category in each environmental entity set are extracted. The distribution characteristics of environmental attribute values ​​for each environmental attribute category in each set of environmental entities are standardized. Based on a preset environmental parameter weighting table, the standardized environmental attribute value distribution characteristics are weighted and summed to generate a comprehensive impact score. The set of environmental entities whose comprehensive impact score is greater than a preset threshold is marked as key environmental entities, and the environmental attribute distribution features corresponding to the key environmental entities are combined to generate the environmental association semantic features.

5. The smart park operation and maintenance management method based on AI autonomous learning according to claim 1, characterized in that, The set of semantic features is dynamically analyzed and processed based on a preset AI autonomous learning model to obtain an abnormal feature set, including: The semantic features of the device's operating status, the semantic features of the environment association, and the semantic features of the user's interaction intent are input into the multi-branch feature extraction layer of a preset AI autonomous learning model, wherein: The first branch feature extraction layer uses a temporal convolutional network to perform sliding window feature extraction on the temporal dimension of the semantic features of the device's operating state, generating a device state feature vector; The second branch feature extraction layer uses a graph convolutional network to perform neighborhood propagation processing on the environmental associated semantic features based on the spatial topology of environmental monitoring nodes, generating an environmental spatial feature vector. The third branch feature extraction layer uses a multi-head attention mechanism to perform cross-dimensional correlation analysis on the semantic features of the user interaction intent, and generates an intent-focused feature vector. The device state feature vector, environmental space feature vector, and intent focus feature vector are mapped to a unified feature space according to a preset feature projection matrix and then fused to generate a fused feature vector. The feature projection matrix is ​​optimized during the model training phase by minimizing the cosine similarity difference of the feature vectors output by each branch feature extraction layer. The fused feature vector is input into the fully connected classification layer of the AI ​​autonomous learning model, and a comprehensive anomaly score sequence is output. The training data of the fully connected classification layer includes historical anomaly event annotation data and normalized vectors of corresponding semantic features. During the training process, the classification weight parameters of the focus loss function are dynamically adjusted according to the distribution ratio of anomaly categories in the current batch of training data to optimize the classification boundary. After aligning the timestamps of the comprehensive anomaly score sequence with the real-time judgment threshold output by the dynamic threshold generation module, point-by-point comparison is performed to determine anomalies, and feature segments whose comprehensive anomaly scores exceed the real-time judgment threshold are extracted as candidate anomaly features. Contextual verification is performed on the candidate abnormal features: based on the temporal continuity of the semantic features of the device operating status, the spatial distribution consistency of the semantic features of the environment association, and the logical correlation of the semantic features of the user interaction intent, false detection feature fragments are eliminated to generate the final abnormal feature set.

6. The smart park operation and maintenance management method based on AI autonomous learning according to claim 5, characterized in that, The output processing of the dynamic threshold generation module includes: The comprehensive anomaly score sequence output by the AI ​​autonomous learning model within a preset time window is collected in real time, and the mean and standard deviation of the comprehensive anomaly score sequence are calculated. After converting the equipment operating load rate into a normalized value in the range of 0 to 1, a benchmark threshold is generated based on the linear combination of the mean and standard deviation, wherein the coefficient of the standard deviation term is positively correlated with the normalized equipment operating load rate. An initial real-time judgment threshold is formed by superimposing a preset baseline offset, wherein the baseline offset is adjusted exponentially based on the frequency of anomalies occurring in the same historical period. When a sudden change in the parameters of the environmental associated semantic features is detected, the threshold dynamic compensation mechanism is activated: based on the ratio of the rate of change of environmental parameters to the safety threshold, the threshold offset is calculated using a preset logarithmic function coefficient, and the initial real-time judgment threshold is offset downward to form the final real-time judgment threshold for application.

7. The smart park operation and maintenance management method based on AI autonomous learning according to claim 1, characterized in that, The step of performing strategy matching processing based on the abnormal feature set and a preset optimization strategy library to generate a target operation and maintenance optimization strategy set includes: Extract the preset anomaly trigger feature vector corresponding to each operation and maintenance optimization strategy from the optimization strategy library, wherein the preset anomaly trigger feature vector is generated in the unified feature space of the AI ​​autonomous learning model; For each abnormal feature vector in the abnormal feature set, calculate its similarity with the preset abnormal trigger feature vectors of all operation and maintenance optimization strategies. When the similarity value between the preset abnormal trigger feature vector of any operation and maintenance optimization strategy and the corresponding type of abnormal feature vector exceeds the preset strategy trigger threshold, add the operation and maintenance optimization strategy to the candidate strategy set. The operation and maintenance optimization strategies in the candidate strategy set are prioritized and sorted. The top N operation and maintenance optimization strategies are selected to generate the target operation and maintenance optimization strategy set, where N is dynamically set according to the concurrent processing capability of the current operation and maintenance management platform.

8. The smart park operation and maintenance management method based on AI autonomous learning according to claim 7, characterized in that, The priority ranking process for each operation and maintenance optimization strategy in the candidate strategy set includes: Obtain the historical execution effect score for each operation and maintenance optimization strategy. The historical execution effect score is calculated and normalized to the range of 0 to 1 by weighting the deviation reduction rate in the equipment status response data after the operation and maintenance optimization strategy is executed, the safety threshold achievement rate of the environmental parameter change data, and the improvement of user satisfaction feedback data. Extract the dynamic weight coefficients of each abnormal feature vector in the current abnormal feature set. The dynamic weight coefficients are normalized according to the proportion of the abnormal score value of the corresponding feature segment in the comprehensive abnormal score sequence in the total abnormal score. The historical execution performance score is multiplied by the dynamic weight coefficient of the current abnormal feature vector of the corresponding operation and maintenance optimization strategy to generate a strategy priority index. The candidate strategy set is sorted by priority index from high to low.

9. The smart park operation and maintenance management method based on AI autonomous learning according to claim 1, characterized in that, The execution of each operation and maintenance optimization strategy in the target operation and maintenance optimization strategy set, and the feedback of the execution results to the operation and maintenance management platform of the smart park, includes: The operation instruction sequence in the equipment operation optimization strategy is parsed to generate a set of equipment control instructions, which is then sent to the corresponding target equipment controller. Monitor the device status response data returned by the target device controller. If the device status response data matches the expected optimization effect characteristics, mark the device operation optimization strategy as successfully executed. The set of environmental adjustment instructions in the environmental monitoring optimization strategy is analyzed to activate environmental control equipment and collect environmental parameter change data in real time. When the environmental parameter change data reaches the preset safety threshold range, the environmental monitoring optimization strategy is marked as completed. The interaction logic update instructions in the user interaction optimization strategy are parsed, the response logic of the user interface is modified, and user satisfaction feedback data is collected. When the user satisfaction feedback data exceeds a preset satisfaction threshold, the user interaction optimization strategy is marked as effective. The monitoring of the device status response data returned by the target device controller includes: After extracting real-time equipment operation characteristics from the equipment status response data and performing standardization processing, the deviation between these characteristics and the expected equipment operation characteristics is calculated. If the deviation remains below the dynamic threshold for more than a preset time window, the device operation optimization strategy is deemed to have been successfully executed; wherein, the dynamic threshold is adaptively adjusted based on the device type and historical optimization effect data.

10. A smart park operation and maintenance management system based on AI autonomous learning, characterized in that, The system includes a processor and a memory, the memory being connected to the processor. The memory is used to store programs, instructions, or code, and the processor is used to execute the programs, instructions, or code in the memory to implement the AI-based autonomous learning-based smart park operation and maintenance management method as described in any one of claims 1-9.

Images