Intelligent space operation and maintenance optimization method and device based on data elements

Abstract

This invention relates to the field of smart space operation and maintenance technology, and discloses a smart space operation and maintenance optimization method and device based on data elements. The method includes: collecting operational data from IoT terminals, edge nodes, and a cloud platform within the smart space, and fusing the operational data to obtain fused data; identifying and encrypting sensitive information in the fused data to obtain encrypted data; inputting the encrypted data into the local models of each edge node to extract device operational features, and integrating the device operational feature set in a cloud aggregation server; identifying anomalies based on the device operational feature set to obtain device abnormal status information; and conducting anomaly cause tracing analysis based on the device abnormal status information to obtain a smart space abnormal situation report. This invention provides operation and maintenance personnel with an interpretable anomaly analysis path and a comprehensive situational awareness report, significantly improving fault handling efficiency and realizing intelligent operation and maintenance management of smart spaces.

Description

Technical Field

[0001] This invention relates to the field of smart space operation and maintenance technology, and in particular to a smart space operation and maintenance optimization method and apparatus based on data elements. Background Technology

[0002] With the rapid development of IoT, AI, and cloud computing technologies, smart spaces, as an important component of smart city construction, have been widely applied in smart buildings, smart communities, and smart parks. Smart spaces utilize a large number of IoT terminal devices, edge computing nodes, and cloud platform services to achieve real-time collection and monitoring of multi-dimensional information such as environmental parameters, equipment status, and energy consumption data. However, existing smart space operation and maintenance management technologies generally suffer from data silos, lacking effective data association and fusion mechanisms between various devices and systems, making it difficult to form a unified operational situation awareness capability. Traditional smart space operation and maintenance methods mainly rely on simple threshold comparisons and rule matching based on a single data source, failing to effectively handle the complex relationships between multi-source heterogeneous data, resulting in low anomaly detection accuracy and high false alarm rates. Summary of the Invention

[0003] This invention provides a smart space operation and maintenance optimization method and device based on data elements. This invention provides operation and maintenance personnel with an interpretable anomaly analysis path and a comprehensive situational awareness report, which significantly improves fault handling efficiency and realizes intelligent operation and maintenance management of smart spaces.

[0004] In a first aspect, the present invention provides a smart space operation and maintenance optimization method based on data elements, the smart space operation and maintenance optimization method based on data elements comprising:

[0005] The system collects operational data from IoT terminals, edge nodes, and cloud platforms within the smart space, and then fuses this operational data to obtain fused data.

[0006] Sensitive information is identified and encrypted in the fused data to obtain encrypted data;

[0007] The encrypted data is input into the local model of each edge node to extract device operation features, and the device operation feature set is integrated in the cloud aggregation server;

[0008] Anomaly identification is performed based on the device operation feature set to obtain device abnormal status information;

[0009] Based on the abnormal status information of the equipment, the cause of the abnormality is traced and analyzed to obtain an abnormal situation report of the smart space.

[0010] In conjunction with the first aspect, in a first implementation of the first aspect of the present invention, the step of collecting operational data from IoT terminals, edge nodes, and cloud platforms within the smart space, and fusing the operational data to obtain fused data, includes:

[0011] The system collects operational data from IoT terminals, edge nodes, and cloud platforms within the smart space.

[0012] Calculate the terminal edge similarity between IoT terminal data and edge node data in the operational data;

[0013] Calculate the edge-cloud similarity and device time coordination between edge node data and cloud platform data in the operational data;

[0014] Based on the terminal edge similarity, the edge cloud similarity, and the device time coordination, the data is correlated and fused to obtain fused data.

[0015] In conjunction with the first aspect, in a second implementation of the first aspect of the present invention, the calculation of the edge-cloud similarity and device time coordination between the edge node data and the cloud platform data in the running data includes:

[0016] Extract the numerical range, data type, and trend characteristics of edge node data from the operational data to construct edge node feature vectors. Simultaneously, extract the corresponding feature parameters of cloud platform data to construct cloud platform feature vectors.

[0017] The cosine similarity between the feature vectors of the edge nodes and the feature vectors of the cloud platform is calculated to obtain the edge-cloud similarity.

[0018] A sliding time window is established based on the timestamp information of the running data. The arrival time difference and data update frequency of the edge node data and the cloud platform data within the same sliding time window are analyzed to obtain the time synchronization analysis results.

[0019] Based on the time synchronization analysis results, the latency variance and frequency matching degree of data transmission between edge nodes and cloud platforms are calculated, and the device time coordination is calculated based on the latency variance and frequency matching degree.

[0020] In conjunction with the first aspect, in a third implementation of the first aspect of the present invention, the step of identifying and encrypting sensitive information in the fused data to obtain encrypted data includes:

[0021] A sensitive data labeling system is established based on the fused data, and the sensitive data labeling system includes public level, internal level, confidential level and top secret level;

[0022] Based on the aforementioned sensitive data labeling system, the fused data is input into a large language model fine-tuned by LoRA for intelligent classification, resulting in intelligently classified data, which includes public-level data, internal-level data, confidential-level data, and top-secret-level data.

[0023] The public-level data and the internal-level data are encrypted using the SM4 symmetric encryption algorithm, and the confidential-level data and the top-secret-level data are encrypted using a combination of SM2 digital signature, SM3 integrity check and SM4 symmetric encryption.

[0024] In conjunction with the first aspect, in the fourth implementation of the first aspect of the present invention, the step of extracting device operating features from the local models of each edge node by inputting the encrypted data, and integrating the device operating feature set in the cloud aggregation server, includes:

[0025] The encrypted data is distributed to the local models of each edge node, and the device operating characteristics are extracted from the encrypted temperature, humidity, current, voltage and power data based on the local models.

[0026] The device operation characteristics and weight parameters of the local model are encrypted and protected by a fully homomorphic encryption algorithm to ensure that the feature data remains in ciphertext state during network transmission, thus obtaining homomorphically encrypted feature data.

[0027] Upload the homomorphic encryption feature data of each edge node to the cloud aggregation server, perform homomorphic addition and homomorphic multiplication operations to achieve weighted aggregation calculation in the ciphertext state, and obtain the global aggregation feature;

[0028] The global model parameters of federated learning are updated based on the global aggregated features to obtain the updated model parameters. The updated model parameters are then distributed to each edge node and combined with the device operation features of each edge node to obtain the device operation feature set.

[0029] In conjunction with the first aspect, in the fifth implementation of the first aspect of the present invention, the global model parameters of the federated learning are updated based on the global aggregated features to obtain updated model parameters, and the updated model parameters are distributed to each edge node and the device operation features of each edge node are combined to obtain a device operation feature set, including:

[0030] The gradient of the loss function of the federated learning model is calculated based on the global aggregated features, and the weight matrix and bias vector of the local model are updated using the federated averaging algorithm based on the gradient of the loss function to obtain the updated model parameters.

[0031] The updated model parameters are sent from the cloud to each edge node through an encrypted communication protocol. Each edge node synchronously updates the weight matrix and bias vector of its local model to obtain a parameter-synchronized edge node model.

[0032] The edge node model that runs the parameter synchronization is used to re-extract the device operation features from the encrypted data to obtain the synchronized device operation features.

[0033] The synchronized device operation characteristics of each edge node are aggregated to obtain a device operation characteristic set.

[0034] In conjunction with the first aspect, in the sixth implementation of the first aspect of the present invention, the step of identifying anomalies based on the device operating feature set to obtain device abnormal state information includes:

[0035] The device operation feature set is converted into a device state feature vector that includes temperature status, power status, network status, and time status.

[0036] The isolated forest algorithm is used to process the device state feature vector. A binary tree isolation structure is constructed by random sampling, and the path depth of the device state feature vector is calculated to obtain the device operation anomaly score.

[0037] The reinforcement learning model is trained to analyze the historical distribution pattern of the abnormal equipment operation scores. Combined with expert experience-annotated data, the normal state boundary and the abnormal judgment threshold are dynamically optimized to obtain an adaptive abnormal threshold.

[0038] By comparing the device operation anomaly score with the adaptive anomaly threshold, devices that exceed the normal boundary are identified and labeled with anomaly type and severity level to obtain device anomaly status information.

[0039] In conjunction with the first aspect, in the seventh implementation of the first aspect of the present invention, the step of performing anomaly cause tracing analysis based on the abnormal device status information to obtain a smart space abnormal situation report includes:

[0040] Retrieve the knowledge graph of abnormal events of smart space equipment, match the equipment type and abnormal pattern in the abnormal status information of the equipment, and create basic data for abnormal source tracing based on the equipment type and the abnormal pattern;

[0041] The thought chain reasoning engine is run to analyze the basic data of anomaly tracing step by step to obtain the reasoning and tracing results;

[0042] Access the LSM tree log storage system to query the historical operation records of smart space devices, identify historical anomaly patterns that are correlated with the current anomaly event in time and space, and obtain spatiotemporal correlation data of the anomaly event;

[0043] By integrating the inference and tracing results with the spatiotemporal correlation data of the abnormal events, a smart space abnormal situation report is generated.

[0044] In conjunction with the first aspect, in the eighth implementation of the first aspect of the present invention, the intelligent space operation and maintenance optimization method based on data elements further includes:

[0045] Extract the location, fault type, and severity information of abnormal devices from the abnormal situation report of the smart space, construct a three-dimensional model of the physical environment of the smart space, and mark the abnormal areas and device nodes to obtain a three-dimensional abnormal situation map;

[0046] Assess the risk propagation path and impact range of each abnormal area in the three-dimensional abnormal situation map, and calculate the regional risk level data;

[0047] Match the regional risk level data with candidate operation and maintenance handling strategies in the historical operation and maintenance case library;

[0048] The candidate operation and maintenance handling strategies are rendered onto an interactive 3D interface, and risk levels are displayed using color coding, handling steps are dynamically labeled, and mobile terminal control functions are integrated to obtain a comprehensive operation and maintenance decision scheme for the smart space.

[0049] Secondly, the present invention provides a smart space operation and maintenance optimization device based on data elements, the smart space operation and maintenance optimization device based on data elements comprising:

[0050] The data acquisition module is used to collect operational data from IoT terminals, edge nodes, and cloud platforms within the smart space, and to fuse the operational data to obtain fused data.

[0051] The encryption module is used to identify and encrypt sensitive information in the fused data to obtain encrypted data;

[0052] The extraction module is used to input the encrypted data into the local model of each edge node to extract device operation features, and integrate the device operation feature set in the cloud aggregation server;

[0053] Anomaly identification module is used to identify anomalies based on the device operation feature set and obtain device anomaly status information;

[0054] The source tracing analysis module is used to perform source tracing analysis based on the abnormal status information of the equipment and obtain an abnormal situation report of the smart space.

[0055] The technical solution provided by this invention achieves precise association and fusion of data from three layers of IoT terminals, edge nodes, and cloud platforms by constructing a device data correlation calculation model and a time coordination analysis mechanism. This effectively solves the technical limitations of data silos and simple aggregation in traditional methods. A four-level security labeling system is established by combining a sensitive data intelligent hierarchical classification algorithm with AI large-scale model LoRA fine-tuning technology. Differential encryption protection is achieved using national cryptographic algorithms SM2 / SM3 / SM4, ensuring personalized protection for data at different security levels and significantly improving the security protection capabilities of smart space data. Federated learning feature extraction and aggregation in encrypted state are achieved through fully homomorphic encryption technology. Distributed model training is completed while protecting the data privacy of each node, ensuring both data security and collaborative intelligence, thus solving the data leakage risk problem in traditional centralized learning. An adaptive anomaly detection mechanism is constructed by combining the isolated forest algorithm and reinforcement learning human feedback methods. This mechanism can dynamically optimize the detection threshold and adapt to changes in the operating mode of smart space devices, achieving higher detection accuracy and lower false alarm rate compared to traditional fixed threshold methods. By leveraging a thought-chain reasoning mechanism and a knowledge graph of abnormal events in smart space equipment, a technological leap has been achieved from simple anomaly alarms to in-depth root cause tracing. This provides maintenance personnel with interpretable anomaly analysis paths and comprehensive situational awareness reports, significantly improving fault handling efficiency. Through a 3D spatial situational map and a multi-dimensional risk assessment model, abstract anomaly information is transformed into intuitive visualizations. Combined with an intelligent maintenance decision recommendation engine, specific handling suggestions are generated, realizing a complete technical closed loop from anomaly detection to maintenance decision-making, and achieving intelligent maintenance management of smart spaces.

[0056] Other features and advantages of the invention will be set forth in the description which follows, and will be apparent in part from the description, or may be learned by practicing the invention. The objects and other advantages of the invention are realized and obtained in accordance with the structures particularly pointed out in the description, claims and drawings.

[0057] To make the above-mentioned objects, features and advantages of the present invention more apparent and understandable, preferred embodiments are described below in detail with reference to the accompanying drawings. Attached Figure Description

[0058] Figure 1 This is a schematic diagram of an embodiment of the intelligent space operation and maintenance optimization method based on data elements in this invention.

[0059] Figure 2 This is a schematic diagram of one embodiment of the intelligent space operation and maintenance optimization device based on data elements in this invention.

[0060] Figure 3 This is a diagram illustrating the interaction process of the smart space component in an embodiment of the present invention. Detailed Implementation

[0061] To make the objectives, technical solutions, and advantages of the embodiments of the present invention clearer, the technical solutions of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0062] The terms "comprising" and "having," and any variations thereof, used in the embodiments of this invention are intended to cover non-exclusive inclusion. For example, a process, method, system, product, or device that includes a series of steps or units is not limited to the steps or units listed, but may optionally include other steps or units not listed, or may optionally include other steps or units inherent to these processes, methods, products, or devices.

[0063] To facilitate understanding of this embodiment, a detailed description of a data-based intelligent space operation and maintenance optimization method disclosed in this embodiment of the invention will be provided first. For example... Figure 1 As shown, the intelligent space operation and maintenance optimization method based on data elements includes the following steps:

[0064] 101. Collect operational data from IoT terminals, edge nodes, and cloud platforms within the smart space, and fuse the operational data to obtain fused data;

[0065] Specifically, data acquisition channels are deployed at various levels of the smart space. IoT terminals upload real-time terminal data, including device operating status, environmental monitoring information, and control feedback. Edge nodes collect locally aggregated operating status information and task execution logs. The cloud platform acquires global scheduling records, cross-regional device interaction information, and operation and maintenance strategy execution data. To eliminate the isolation of data from different sources, the similarity between IoT terminal data and edge node data is calculated by analyzing the consistency of numerical distribution, trends, and business semantics between the two types of data. This reflects the degree of coupling between terminal devices and edge nodes in terms of operating status. The edge-cloud similarity between edge node data and cloud platform data is calculated, and device time coordination is simultaneously evaluated. Edge-cloud similarity measures the consistency between the edge layer and the cloud in global control strategy execution and status feedback, while device time coordination reflects the synchronization and collaboration of devices at different levels in the time dimension through timestamp alignment and event sequence matching. Terminal-edge similarity, edge-cloud similarity, and device time coordination are used as multi-dimensional inputs. A cross-level data association matrix is ​​constructed using a fusion algorithm, and weighted fusion processing is performed to obtain fused data.

[0066] 102. Sensitive information is identified and encrypted from the fused data to obtain encrypted data;

[0067] Specifically, a sensitive data labeling system is constructed based on the structure and business semantic features of the fused data. This system covers four security levels: public, internal, confidential, and top secret, with clearly defined criteria, data scope, and security policies for each level. Based on this system, the fused data is input into a large language model fine-tuned using LoRA. The model, pre-trained, undergoes domain-adaptive optimization incorporating the scene characteristics and security requirements of the smart space, resulting in higher semantic understanding capabilities and data sensitivity identification accuracy. It can automatically classify the fused data, outputting intelligently graded data corresponding to public, internal, confidential, and top secret levels according to their security. For public and internal data in the classification results, the national standard SM4 symmetric encryption algorithm is used for encryption to ensure high-efficiency encryption performance at a lower security level. For confidential and top-secret data, a combined encryption strategy is adopted: the SM2 algorithm is used for digital signature to ensure the authenticity of the data source, the SM3 algorithm is used to perform data integrity verification to prevent tampering, and the SM4 symmetric encryption algorithm is used to encrypt the data content with high strength, forming a multi-layer protection mechanism covering identity verification, integrity protection and content encryption.

[0068] 103. Input encrypted data into the local model of each edge node to extract device operation features, and integrate the device operation feature set in the cloud aggregation server;

[0069] Specifically, encrypted data is distributed via a secure transmission channel to local feature extraction models deployed on various edge nodes. These local models are built on convolutional neural network (CNN) or Transformer architectures and have been structurally optimized for the operational characteristics of smart space devices. They can directly perform feature calculations on encrypted data under fully homomorphic encryption protection. CNNs, when processing continuous data such as temperature, humidity, current, voltage, and power, extract local change patterns through multi-layer convolutional kernels, capturing the spatial distribution characteristics of abnormal trends. The Transformer architecture, on the other hand, utilizes a multi-head self-attention mechanism to globally model the dependencies of parameters from different time points and sensor sources, uncovering implicit correlations between multiple dimensions. Based on these architectures, the local models extract multi-dimensional feature vectors of device operation under encrypted conditions. A fully homomorphic encryption algorithm is used to further encrypt and protect the device operation features and the weight parameters of the local models, ensuring that these feature data remain encrypted throughout the entire process of network transmission and aggregation calculation, generating homomorphically encrypted feature data. Each edge node uploads its encrypted feature data to the cloud aggregation server. The cloud server directly performs homomorphic addition and multiplication operations in the encrypted state to achieve weighted aggregation calculation of multi-node features, obtaining a global aggregated feature representing the overall operational status of the smart space. Based on the global aggregated feature, the cloud server updates the global model parameters in the federated learning framework to adapt to the latest multi-node operational status and distributes the updated model parameters to each edge node. After receiving the updated model parameters, each edge node combines them with locally stored device operational features to generate a device operational feature set covering multi-dimensional operational information of the smart space.

[0070] 104. Based on the equipment operation feature set, perform anomaly identification to obtain equipment abnormal status information;

[0071] Specifically, the equipment operation feature set is structured and mapped according to parameter categories. Multidimensional data related to temperature, humidity, current, voltage, power, network latency, and time synchronization are normalized and feature-filtered, transforming them into equipment state feature vectors containing four main dimensions: temperature, power, network, and time. Anomaly patterns are mined from these feature vectors using the Isolation Forest algorithm. This algorithm selects feature dimensions through multiple rounds of random sampling and constructs a binary tree isolation structure based on the sample distribution. It utilizes the randomness of node splitting to gradually separate data points in the feature space and generates corresponding equipment operation anomaly scores after calculating the average path depth of each feature vector. A reinforcement learning model is introduced to continuously analyze the distribution patterns of historical anomaly scores and, combined with positive and negative sample data labeled by maintenance experts, dynamically optimizes the normal state boundary and the critical value for anomaly detection, resulting in an adaptive anomaly threshold that can be adjusted in real time according to changes in equipment operation modes. During training, the reinforcement learning model incorporates anomaly detection accuracy and false alarm rate as part of the reward function, updating the strategy parameters through multiple iterations to ensure that the threshold setting balances stability and sensitivity. The system compares the abnormal equipment score with an adaptive abnormal threshold. When the abnormal score exceeds the normal boundary, the corresponding equipment is automatically identified as abnormal. Based on the magnitude of the abnormal score, the degree to which it exceeds the threshold, and the concentration of abnormalities in related features, the system determines the type and severity level of the equipment abnormality, forming equipment abnormality status information that includes information such as abnormality category, severity, occurrence time, and scope of impact.

[0072] 105. Based on the abnormal equipment status information, conduct source tracing analysis to obtain an abnormal situation report for the smart space.

[0073] Specifically, a structured retrieval of the knowledge graph of abnormal events in smart space equipment is performed. The equipment type and abnormal patterns contained in the current abnormal status information are matched with entities and relationships in the knowledge graph to determine fault modes, influencing factors, and historical handling solutions highly relevant to the current event, thus constructing the foundational data for anomaly tracing. A thought chain reasoning engine is then activated to perform step-by-step reasoning analysis on the foundational data. The reasoning process follows a chain logic of "phenomenon—cause—deduction—conclusion," performing multi-stage correlation calculations between the matched abnormal patterns and the equipment operating environment, historical maintenance records, and related influencing conditions to obtain reasoning tracing results with causal explanatory power. Simultaneously with the reasoning analysis, a high-performance log storage system using an LSM tree architecture is accessed to query and compare historical operating records of the smart space equipment. By leveraging the hierarchical structure advantages of LSM tree's in-memory tables, immutable in-memory tables, and disk storage, historical abnormal patterns that are temporally continuous or periodic and spatially adjacent or system-cascaded with the current abnormal event are quickly retrieved under high concurrency conditions, obtaining the spatiotemporal correlation data of the abnormal event. By integrating the results of reasoning and tracing with spatiotemporal correlation data of abnormal events, a smart space anomaly situation report is formed, covering the anomaly triggering mechanism, propagation path, scope of impact, and prediction of potential risks.

[0074] In one specific embodiment, the process of performing step 101 may specifically include the following steps:

[0075] The system collects operational data from IoT terminals, edge nodes, and cloud platforms within the smart space.

[0076] Calculate the terminal edge similarity between IoT terminal data and edge node data in the operational data;

[0077] Calculate the edge-cloud similarity and device time coordination between edge node data and cloud platform data in the operational data;

[0078] Based on the terminal edge similarity, the edge cloud similarity, and the device time coordination, the data is correlated and fused to obtain fused data.

[0079] Specifically, a data acquisition channel is constructed within the multi-layered data acquisition and transmission architecture of the smart space. IoT terminals collect fundamental parameters directly reflecting device operating status, such as temperature, humidity, current, voltage, power, vibration, and network latency. Edge nodes integrate multiple terminal data sources and record task scheduling and device interaction status, possessing local caching, preprocessing, and analysis capabilities. The cloud platform, acting as the top-level operation and management center, stores global scheduling records, cross-node operation logs, system health status, and security event information, forming a vertically layered and horizontally interconnected multi-dimensional operational data network. During the data acquisition phase, operational data is simultaneously acquired from different nodes in the three-layer structure. The acquired data is categorized and identified using a device type coding table, and standardized processing of multi-source data is achieved using a unified timestamp standard and field mapping rules. The terminal-edge similarity between IoT terminal data and edge node data is calculated. Using the device data association quantification calculation model in the Multi-Source Data Acquisition Association Analysis (MDCA) method, similarity is defined as a weighted combination of data semantic similarity and temporal coordination. Semantic similarity is calculated using a cosine similarity algorithm to measure the angle between the feature vectors of the terminal and the edge within the same time window, reflecting their consistency in numerical trends. Temporal coordination is quantitatively assessed by calculating the timestamp difference of corresponding data records, event trigger synchronization rate, and acquisition delay distribution to evaluate the degree of matching between terminal and edge data in the temporal dimension. The weighting coefficients α and β are dynamically adjusted by the system based on historical data stability and business importance. For example, when the terminal and edge show high temporal consistency in historical data, the weight of temporal coordination in the overall similarity calculation is increased; conversely, the weight of semantic similarity is increased. The system also calculates edge-cloud similarity and device temporal coordination between edge node data and cloud platform data. The calculation method for edge-cloud similarity is similar to that for terminal-edge similarity, but an additional control strategy matching analysis is introduced in the semantic similarity part. By comparing the scheduling commands and policy parameters issued by the cloud with the actual execution records of the edge nodes, the consistency between the execution of the edge layer and the cloud instructions is evaluated. Time coordination, on the other hand, is identified by quantitatively analyzing the differences in timestamps between the edge and cloud at key nodes such as task scheduling, data reporting, and event response, thus identifying cross-level synchronization delays or scheduling deviations. Terminal-edge similarity, edge-cloud similarity, and device time coordination are constructed into a cross-level data association matrix, and a weighted fusion operation is performed based on MDCA. In the fusion algorithm, the weights of each indicator are updated in real time by an adaptive adjustment mechanism. This mechanism uses short-term data volatility, long-term trend stability, and business impact as input variables, and dynamically corrects the weight allocation using sliding window statistics and exponential smoothing methods. Simultaneously with weighted fusion, data semantic association analysis is introduced to match structured numerical data with semi-structured log records and event descriptions, achieving comprehensive utilization of cross-type information and ultimately outputting fused data.

[0080] (1) DINO (2021): This embodiment also includes feature extraction processing of smart space video surveillance data: deploying a video feature extraction model based on the DINO self-supervised learning architecture, constructing a teacher network and a student network using the Vision Transformer structure, updating the teacher network weights through exponential moving average, and obtaining a smart space video analysis model; collecting real-time video stream data from security cameras, equipment monitoring cameras, and environmental monitoring cameras in the smart space, converting video frames into image block sequences and performing multi-cropping preprocessing to obtain multi-scale video surveillance data; training the smart space video analysis model based on the multi-scale video surveillance data, with the student network processing video image blocks at all scales and the teacher network processing global video frames, learning the device status and abnormal behavior patterns in the video through a self-distillation mechanism to obtain a video feature extraction model; running the video feature extraction model to perform unsupervised feature learning on the smart space surveillance video, automatically identifying abnormal device status, abnormal personnel behavior, and abnormal environmental changes in the video to obtain video abnormal features; fusing the video abnormal features with sensor data features in a multimodal manner to generate comprehensive smart space operation status features containing visual information and sensor information, and obtaining enhanced operation data features.

[0081] (2)DINOv2 (2023): This embodiment also includes high-quality feature extraction processing of smart space video surveillance data: a large-scale video feature extraction model based on the improved DINOv2 architecture is constructed, pre-trained on the smart space video dataset, and KoLeo regularization and L2 normalization techniques are used to improve the quality of video features to obtain the DINOv2 video analysis model; high-resolution video data of multiple types of camera devices in the smart space are collected, and a high-quality smart space video dataset is constructed through deduplication and quality screening to support multiple application scenarios such as equipment monitoring, security protection and environmental monitoring to obtain a high-quality video dataset; the DINOv2 video analysis model is pre-trained on a large scale based on the high-quality video dataset, and the knowledge of the large-parameter video model is transferred to the lightweight model deployed on the edge node using model distillation technology to obtain the edge video feature model; the edge video feature model is run to perform multi-task feature extraction on the real-time video stream of the smart space, which can simultaneously support tasks such as video classification, target detection, behavior recognition and anomaly localization without fine-tuning to obtain multi-task video features; the multi-task video features are combined with the specific video analysis tasks of the smart space for feature optimization to generate video feature representations specifically for smart space operation and maintenance monitoring to obtain dedicated video operation data features.

[0082] (3) DINO-WM (2024): This embodiment also includes predictive feature extraction processing of smart space video surveillance data: a smart space video prediction model based on the DINO-WM world model is constructed, and the video feature extraction capability is used to train on the historical video trajectory data of the smart space to learn the spatiotemporal evolution law of the video scene and obtain the video world model; video surveillance data of continuous time periods in the smart space is collected, and video time-series trajectory data containing changes in equipment operating status, personnel activity trajectory and environmental state evolution is constructed to obtain the video trajectory dataset; the video world model is trained based on the video trajectory dataset to learn the physical change law and state transition mechanism in the smart space video scene, and to achieve accurate prediction of future video frames and scene states, and obtain the video prediction model; the video prediction model is run to perform zero-sample video scene prediction, and potential equipment failures and security risks are identified in advance by predicting abnormal events and state changes in future video frames, and video prediction features are obtained; the video prediction features are fused with the current video features in a time series to generate a comprehensive video feature representation containing historical video states, current video states and future predicted states, and predictive video operation data features are obtained.

[0083] In one specific embodiment, the process of calculating the edge-cloud similarity and device time coordination between edge node data and cloud platform data in the runtime data may specifically include the following steps:

[0084] Extract the numerical range, data type, and trend characteristics of edge node data from the operational data to construct edge node feature vectors. At the same time, extract the corresponding feature parameters of cloud platform data to construct cloud platform feature vectors.

[0085] The cosine similarity between the feature vectors of edge nodes and the feature vectors of the cloud platform is calculated to obtain the edge-cloud similarity.

[0086] A sliding time window is established based on the timestamp information of the running data. The arrival time difference and data update frequency of edge node data and cloud platform data within the same sliding time window are analyzed to obtain time synchronization analysis results.

[0087] Based on the time synchronization analysis results, the latency variance and frequency matching degree of data transmission between edge nodes and cloud platforms are calculated, and the device time coordination is calculated based on the latency variance and frequency matching degree.

[0088] Specifically, the collected operational data undergoes structured analysis, extracting data belonging to edge nodes separately. Numerical statistics and type identification are performed on each data field to obtain the numerical range, data type, and trend characteristics of the edge node data. The numerical range is obtained by analyzing the maximum, minimum, and upper / lower limit fluctuations of the data over a period of time. Data types are classified based on field definitions and actual value characteristics; for example, temperature and humidity are continuous numerical data, device on / off status is discrete or Boolean data, and alarm information is enumerated data. Trend characteristics are quantified using time series analysis methods to quantify the rate of increase / decrease, periodicity, and abrupt changes of data at continuous time points, thereby capturing the behavioral patterns of edge node data during dynamic operation. Based on the above numerical range, data type, and trend characteristics, a high-dimensional vector, namely the edge node feature vector, is combined and encoded. Simultaneously, a feature extraction process is performed on the operational data collected from the cloud platform. Since the cloud platform data includes global scheduling records, cross-node status summaries, and integrated monitoring parameters, consistency with the edge node feature extraction process is maintained when extracting corresponding feature parameters to ensure the correspondence and comparability of the two types of feature vectors in terms of dimensions. By extracting the numerical range, data type, and trend characteristics of the cloud platform data, a cloud platform feature vector with the same structure as the edge node feature vector is constructed. Cosine similarity is calculated between the edge node feature vector and the cloud platform feature vector to measure the similarity between the two sets of features in multidimensional space. Cosine similarity represents the consistency of feature distribution by calculating the cosine value of the angle between the two vectors. The result ranges from 0 to 1; the closer the value is to 1, the higher the similarity between the two types of features, meaning that the operating status of the edge node and the global monitoring results of the cloud platform have higher consistency in numerical patterns and trends. Conversely, a lower value indicates data deviation, information lag, or loss and distortion during the aggregation process. A sliding time window mechanism is established based on the timestamp information of the operating data to analyze the difference in arrival time and data update frequency between edge node data and cloud platform data within the same time window. The length of the sliding time window is set based on the statistical characteristics of the device acquisition cycle and network transmission latency. Within the time window, the time difference between the arrival of each edge node's data and the corresponding cloud platform data record is statistically analyzed to form an arrival time difference distribution. This distribution is then used to analyze the latency characteristics of the edge-to-cloud transmission link within the window. Simultaneously, the data update frequency is statistically analyzed, i.e., the number of data updates generated by the edge node and the cloud platform within the window, to measure the consistency of their data refresh rates. Based on the statistical results of the time difference and update frequency, the latency variance and frequency matching degree of the data transmission between the edge node and the cloud platform are calculated.Delay variance reflects the stability of transmission delay; a smaller variance indicates less delay fluctuation and a more stable transmission process. Frequency matching degree is measured by comparing the difference and proportion of update counts between edge nodes and the cloud platform within a window; a higher matching degree indicates greater consistency in their data refresh rhythms. Using delay variance and frequency matching degree as inputs, a weighted calculation is performed to obtain device time coordination, comprehensively reflecting the synchronization level and collaborative capability of edge nodes and the cloud platform in the time dimension.

[0089] The process involves calculating the latency variance and frequency matching degree of data transmission between edge nodes and the cloud platform based on the time synchronization analysis results. Then, it calculates the device time coordination based on these variances. This includes: extracting the arrival time difference sequence between edge node data and cloud platform data within each time window from the time synchronization analysis results; calculating the mean and standard deviation of the time difference sequence to obtain data transmission latency statistical parameters; calculating the latency variance of data transmission between the edge nodes and the cloud platform based on these statistical parameters; constructing a device correlation quantification model using a weighting coefficient α and a time coordination coefficient β to obtain a latency variance weighting factor; analyzing the ratio between the update frequency of edge node data and the update frequency of cloud platform data in the time synchronization analysis results; calculating the stability index and periodic matching degree of the frequency ratio to obtain a frequency matching degree score; and inputting the latency variance weighting factor and the frequency matching degree score into the device time coordination calculation formula. A weighted summation operation is then performed to generate a quantitative index reflecting the degree of time consistency between edge node and cloud platform data, thus obtaining the device time coordination.

[0090] In one specific embodiment, the process of performing step 102 may specifically include the following steps:

[0091] A sensitive data labeling system is established based on the integrated data. The sensitive data labeling system includes public level, internal level, confidential level and top secret level.

[0092] Based on the sensitive data labeling system, the fused data is input into a large language model that has been fine-tuned by LoRA for intelligent classification, resulting in intelligently classified data, which includes public-level data, internal-level data, confidential-level data, and top-secret-level data.

[0093] The SM4 symmetric encryption algorithm is used for public and internal data, while the SM2 digital signature, SM3 integrity check and SM4 symmetric encryption are combined for confidential and top-secret data to obtain encrypted data.

[0094] Specifically, a sensitive data labeling system is established based on fused data. This system combines various factors such as the data sensitivity distribution patterns in the smart space operating environment, business privacy protection levels, and national and industry data security management standards to construct a sensitive data labeling system, which is divided into four levels: public, internal, confidential, and top secret. Public data refers to information that can be freely disseminated within or outside the system, such as basic information like equipment models and room numbers. Internal data includes information that circulates only within the system but does not involve critical control, such as intermediate calculation results and general operating parameters. Confidential data involves equipment control commands, operating strategies, or data with potential security risks, such as voltage and power thresholds for critical equipment. Top secret data includes core elements that seriously impact system security, such as authentication keys, policy configuration parameters, and equipment command interfaces. The labeling system clearly defines the data attributes, access permissions, processing requirements, and encryption strategies corresponding to each level during its construction. The fused multi-source data is input into a large language model fine-tuned using LoRA for intelligent hierarchical processing. The large language model incorporates LoRA (Low-Rank Adaptation) technology on top of pre-training. By fine-tuning the low-rank subspace of the parameter matrix, it effectively reduces the amount of parameter updates, improves the model's adaptation efficiency in specific domains, and avoids the decline in generalization ability caused by full parameter updates. During fine-tuning, the model is guided to learn the sensitivity features of various data fields in a smart space scenario, including field semantics, historical leakage risk, and contextual dimensions. This is achieved through a LoRA insertion module, injecting a structure in the form of a low-rank matrix A=UV^T, forming an intelligent grading module with semantic recognition and sensitivity judgment capabilities. The large language model automatically determines the sensitivity level of each field or combination of fields based on the training-derived classification weights and discrimination boundaries, and outputs structured intelligently graded data. The intelligently graded data uses data blocks as basic units, labeled as public, internal, confidential, and top-secret data. Each type of data is assigned different encryption logic and access permission strategies. Differentiated encryption strategies are executed based on the grading results, with the SM4 symmetric encryption algorithm used for public and internal data. SM4 is a national commercial cryptographic algorithm with advantages such as fast encryption and decryption speed, efficient hardware implementation, and high encryption strength. It is suitable for data encryption scenarios with large data volumes but relatively low security requirements. When processing such data, a corresponding symmetric key is generated, the SM4 encryption module is called to complete the field encryption, and the key is stored in a controlled manner in an intermediate-level key management unit to ensure that it is used only within authorized scope.For confidential and top-secret data, a combined encryption mechanism is used. The SM2 algorithm is used for digital signature to ensure that the data has a verifiable source identity during subsequent transmission and use. The SM3 algorithm is used to calculate the data digest and embed an integrity verification field to ensure that the data has not been tampered with during transmission. The SM4 symmetric encryption algorithm is used to encrypt the data body, forming a three-layer security mechanism covering identity authentication, integrity verification and content confidentiality, and outputting encrypted data.

[0095] The process involves several steps. First, based on a sensitive data labeling system, the fused data is input into a LoRA-tuned large language model for intelligent classification, resulting in intelligent classification data. This includes: constructing a sensitive data identification training sample set specifically for the smart space domain; manually labeling device identifiers, location information, and operating parameters in the fused data according to public, internal, confidential, and top-secret levels to obtain a labeled training dataset; performing LoRA fine-tuning training on the pre-trained large language model based on the labeled training dataset; decomposing the weight update matrix into the product of two low-rank matrices using low-rank matrix factorization; freezing the original model parameters and updating only the low-rank adapter parameters to obtain a smart space domain-adapted model; converting the fused data into a natural language description format and inputting it into the smart space domain-adapted model; the model outputs the security level prediction probability distribution for each data item based on contextual semantic understanding and domain knowledge reasoning to obtain the classification probability prediction result; setting a security level judgment threshold based on the classification probability prediction result; using the level with the highest prediction probability exceeding the judgment threshold as the final classification result; and using expert rules for secondary judgment on data items below the threshold to obtain intelligent classification data.

[0096] In one specific embodiment, the process of performing step 103 may specifically include the following steps:

[0097] The encrypted data is distributed to the local models of each edge node, and the device operation characteristics are extracted from the encrypted temperature, humidity, current, voltage and power data based on the local models.

[0098] A fully homomorphic encryption algorithm is used to encrypt and protect the device operation characteristics and the weight parameters of the local model, ensuring that the feature data always remains in ciphertext state during network transmission, thus obtaining homomorphically encrypted feature data;

[0099] Upload the homomorphic encrypted feature data of each edge node to the cloud aggregation server, perform homomorphic addition and homomorphic multiplication operations to achieve weighted aggregation calculation in the ciphertext state, and obtain the global aggregation feature;

[0100] The global model parameters of federated learning are updated based on the global aggregated features to obtain the updated model parameters. The updated model parameters are then distributed to each edge node and combined with the device operation features of each edge node to obtain the device operation feature set.

[0101] Specifically, the fused data, after hierarchical encryption processing, is securely distributed to the local model processing units of each edge node through a controlled channel. Lightweight local models built with convolutional neural networks or Transformer architectures are pre-deployed in each edge node. The convolutional neural network extracts local trends and amplitude variations in the time-series data, while the Transformer model captures global dependencies between device parameters through a multi-head attention mechanism, making feature representations more expressive. The local model structure is compatible with a fully homomorphic encryption environment, allowing for non-decryption calculations on encrypted data such as temperature, humidity, current, voltage, and power in an encrypted state, achieving privacy-preserving feature extraction. The extracted device operating features and the weight parameters of the local model are further encrypted and encapsulated using a fully homomorphic encryption algorithm to generate homomorphically encrypted feature data, ensuring that the encrypted state is maintained throughout network transmission and preventing data interception or reverse inference. Each edge node uploads the homomorphically encrypted feature data to a cloud aggregation server. The cloud performs homomorphic addition and multiplication operations on the feature data uploaded by all nodes in an encrypted state, completing weighted aggregation in the encrypted space to obtain global aggregated features representing the overall device operating state. The aggregation weights between nodes are comprehensively allocated based on the historical data volume, collection completeness, and model stability of each edge node. After aggregation, the parameters of the federated learning model are updated based on the global aggregated features. The global model completes iteration without decryption by optimizing the loss function in encrypted state, and the updated model parameters are sent back to each edge node. After receiving the parameters, each node quickly verifies and fuses them with its local running features, starts a new round of feature extraction tasks, and records the model accuracy index and convergence status as feedback parameters and synchronizes them to the cloud. This forms a federated collaborative learning process with closed-loop control, dynamic iteration, and privacy protection capabilities, resulting in a device running feature set.

[0102] In this embodiment, the updated model parameters are distributed from the cloud to each edge node via an encrypted communication protocol. Each edge node synchronously updates the weight matrix and bias vector of its local model, resulting in a parameter-synchronized edge node model. This includes: establishing a federated learning node authentication mechanism based on CA digital certificate authentication, issuing digital certificates to the cloud aggregation server and each edge node, and verifying the legitimacy of the node identity participating in federated learning through certificate chain verification, thus obtaining a trusted node authentication system; establishing an end-to-end encrypted communication channel based on the trusted node authentication system, using the TLS protocol and national cryptographic algorithms to encrypt and protect the model parameter transmission process, ensuring that the updated model parameters are not stolen or tampered with during network transmission, thus obtaining a secure parameter transmission channel; distributing the updated model parameters to each edge node through the secure parameter transmission channel, where each edge node verifies the integrity and authenticity of the received parameters, and uses digital signatures to verify that the parameters have not been maliciously modified, thus obtaining verified model parameters; and replacing the corresponding weight matrix and bias vector of the local model with the verified model parameters, performing a model parameter synchronization update operation, and returning a synchronization confirmation message to the cloud, thus obtaining a parameter-synchronized edge node model.

[0103] Principal Component Analysis (PCA): In this embodiment, the encrypted data is distributed to the local models of each edge node, and equipment operation features are extracted from the encrypted temperature, humidity, current, voltage, and power data based on the local models. This includes: standardizing the high-dimensional equipment operation parameters such as temperature, humidity, current, voltage, and power in the encrypted data; eliminating the dimensional differences of each parameter through centralization and standardization operations to obtain standardized equipment operation data; calculating the covariance matrix between equipment parameters based on the standardized equipment operation data; analyzing the correlation between parameters such as temperature, humidity, current, voltage, and power to obtain the equipment parameter covariance matrix; and further processing the encrypted data. The covariance matrix of the prepared parameters is subjected to eigenvalue decomposition. The principal component directions are obtained by arranging the eigenvalues ​​in descending order. The top k eigenvectors whose cumulative variance contribution rate reaches a set threshold are selected as principal components to obtain the principal component basis of the smart space equipment. The standardized equipment operation data is projected onto the principal component basis of the smart space equipment. A linear transformation is used to map the high-dimensional equipment operation parameters to the low-dimensional principal component space, retaining the main variation information of the equipment operation state to obtain the dimensionality-reduced equipment features. Based on the dimensionality-reduced equipment features, the explained variance ratio and cumulative contribution rate of each principal component are calculated. The feature dimensionality reduction effect is evaluated, and a compact feature representation containing the key modes of equipment operation is generated to obtain the equipment operation features.

[0104] In one specific embodiment, the process of updating the global model parameters of federated learning based on global aggregated features to obtain updated model parameters, and then distributing the updated model parameters to each edge node and combining the device operation features of each edge node to obtain a device operation feature set can specifically include the following steps:

[0105] The gradient of the loss function of the federated learning model is calculated based on the global aggregated features, and the weight matrix and bias vector of the local model are updated using the federated averaging algorithm based on the gradient of the loss function to obtain the updated model parameters.

[0106] The updated model parameters are sent from the cloud to each edge node through an encrypted communication protocol. Each edge node synchronously updates the weight matrix and bias vector of its local model, resulting in a parameter-synchronized edge node model.

[0107] The edge node model for synchronized operation parameters re-extracts device operation features from encrypted data to obtain synchronized device operation features.

[0108] By aggregating the synchronized device operation characteristics of each edge node, a device operation characteristic set is obtained.

[0109] Specifically, using homomorphic encrypted features aggregated from multiple edge nodes as input, the gradient of the loss function of the current federated learning model is calculated in the cloud. The gradient reflects the degree of deviation between the current parameter configuration of the model and the ideal predicted output. The loss function is constructed based on the prediction error of the smart space device state, such as defining mean squared error or cross-entropy loss based on regression or classification objectives, and the error signal of the weighted contribution of each node is accumulated in the ciphertext space using encryption operations. After the loss gradient is calculated, the weight matrix and bias vector of the local model of all edge nodes are updated using a federated averaging algorithm. While maintaining the consistency of the model structure of each node, the global loss gradient is mapped to each weight parameter dimension, and a weighted average update is performed according to the node weights to obtain a unified updated model parameter set. The updated model parameters are securely distributed from the cloud to each edge node through an encrypted communication protocol. The protocol supports end-to-end encryption, key negotiation, and integrity verification to ensure the tamper resistance and consistency verification of the parameters during transmission. After receiving the model parameters from the cloud, each edge node synchronously updates the weight matrix and bias vector of its local model and reconstructs the synchronized local model structure based on the new parameters, thus maintaining consistency between the edge model and the global model at the structural and parameter levels. The synchronized edge node model then performs feature extraction on the locally cached encrypted device operation data. Based on the improved feature representation capabilities of the updated model parameters, it extracts more accurate, stable, and representative device operation features from key parameters such as temperature, humidity, current, voltage, and power under encrypted conditions. The feature extraction process is still completed in a fully homomorphic encrypted environment, ensuring that data privacy is not leaked on the local nodes and guaranteeing that the model has stronger sensitivity to abnormal state changes after parameter iterations. After all edge nodes complete the synchronized feature extraction task, each node uploads its generated synchronized device operation features to the cloud in an encrypted manner. The cloud server then performs structural alignment, dimensional standardization, and redundancy removal on the features from all nodes, integrating the features from multiple nodes to form a device operation feature set.

[0110] like Figure 3 As shown, Figure 3This invention constructs a clear three-layer symmetrical architecture interaction system to illustrate the interaction process of smart space components. The top-level IoT terminal layer includes IoT terminal 1, IoT terminal 2, and IoT terminal N, corresponding to sensor devices, actuator devices, and gateway devices, respectively, responsible for collecting operational data of the smart space. The middle-level edge node layer includes edge node 1 and edge node 2, deploying local model A and local model B respectively, performing feature extraction processing. The bottom-level cloud platform serves as the core processing center, containing four functional modules: a cloud aggregation server for federated learning aggregation and fully homomorphic encrypted computation; an anomaly detection module using the isolated forest algorithm; a source tracing analysis module using chain-of-thought reasoning; and a visualization decision-making module using a 3D situation map. A clear bidirectional data flow is achieved: operational data flows from top to bottom to the edge nodes, while encrypted feature data continues to be transmitted to the cloud aggregation server; model parameter updates are returned to each edge node from bottom to top via dashed arrows. The modules within the cloud platform are sequentially connected by horizontal arrows, ultimately outputting a smart space anomaly situation report and an operation and maintenance optimization decision plan. The entire interaction process embodies the core technological features of multi-source data element fusion, privacy-preserving federated learning, and intelligent situational awareness, forming a complete intelligent space operation and maintenance optimization technology chain, ensuring the system's security, intelligence, and practicality.

[0111] In one specific embodiment, the process of performing step 104 may specifically include the following steps:

[0112] The equipment operation feature set is converted into an equipment state feature vector that includes temperature status, power status, network status, and time status.

[0113] The isolated forest algorithm is used to process the device state feature vector. A binary tree isolation structure is constructed by random sampling and the path depth of the device state feature vector is calculated to obtain the device operation anomaly score.

[0114] The reinforcement learning model is trained to analyze the historical distribution pattern of equipment operation anomaly scores. Combined with expert experience-annotated data, the normal state boundary and the anomaly judgment threshold are dynamically optimized to obtain an adaptive anomaly threshold.

[0115] By comparing the equipment operation anomaly score with the adaptive anomaly threshold, the equipment that exceeds the normal boundary is identified and the anomaly type and severity level are labeled to obtain the equipment anomaly status information.

[0116] Specifically, semantic structure mapping is performed on the equipment operation feature set to construct an equipment state feature vector for anomaly detection. This vector includes temperature status reflecting the stability and heat dissipation performance of temperature sensors, power status measuring the equipment's electrical load and fluctuations, network status assessing data communication latency and packet loss, and time status characterizing the rhythm, periodicity, and sequence of events. During construction, representative values ​​are extracted from the original equipment operation features and normalized, such as calculating the rate of temperature change, maximum current fluctuation amplitude, mean and variance of network latency, and time offset of periodic tasks. These parameters are then arranged in an ordered manner to form a high-dimensional equipment state feature vector. Unsupervised anomaly detection is then performed on the equipment state feature vector using the Isolation Forest algorithm. The Isolation Forest algorithm constructs multiple independent binary tree isolation structures, performing random feature selection and interval segmentation operations on the input state feature vector. Each segmentation is equivalent to an anomaly sample localization attempt. The core idea of ​​the Isolation Forest algorithm is that, compared to normal samples in dense regions, anomaly samples are more easily isolated by fewer segmentations, thus resulting in a shorter path depth within the tree structure. After training, the average path depth of each feature vector across multiple trees is traversed and converted into an anomaly score based on statistical regularity. Lower scores indicate easier isolation and a higher probability of anomaly. A reinforcement learning model is constructed to model the historical distribution of anomaly scores and, combined with expert-annotated normal and anomaly samples, dynamically learns and optimizes the boundary conditions for anomaly identification. The reinforcement learning model employs an iterative training mechanism with feedback rewards, using recognition accuracy, false positive rate, and false negative rate as primary evaluation metrics to dynamically adjust the position of the normal state boundary and the anomaly judgment threshold. An initial judgment threshold is set based on the distribution density, fluctuation range, and historical anomaly hotspots of the anomaly scores. The model is then guided to migrate the judgment boundary to the optimal range through strategy training on historical data, obtaining the most suitable adaptive anomaly threshold under the current environment and data model. After comparing the anomaly scores with the adaptive anomaly threshold, all devices with anomaly scores exceeding the threshold are identified as potential anomalies. Anomaly type and severity level are then determined based on the degree of anomaly and the type of indicator. For example, if the temperature status anomaly score is prominent and the power status fluctuation is slight, it is judged as a heat dissipation system failure; if the voltage and current fluctuations are significant and the network status is synchronized well, it is judged as an electrical system anomaly. Through a preset anomaly type matching table and threshold level hierarchical mechanism, each abnormal device is assigned a specific anomaly label and severity level, and the device anomaly status information is output, including device identifier, anomaly type, severity level, timestamp, and location data.

[0117] In one specific embodiment, the process of performing step 105 may specifically include the following steps:

[0118] Retrieve the knowledge graph of abnormal events of smart space equipment, match the equipment type and abnormal pattern in the abnormal status information of the equipment, and create basic data for abnormal source tracing based on the equipment type and abnormal pattern;

[0119] The thought chain reasoning engine is run to analyze the basic data of anomaly tracing step by step, and the reasoning and tracing results are obtained.

[0120] Access the LSM tree log storage system to query the historical operation records of smart space devices, identify historical anomaly patterns that are correlated with the current anomaly event in time and space, and obtain spatiotemporal correlation data of the anomaly event;

[0121] By integrating the results of reasoning and tracing and the spatiotemporal correlation data of abnormal events, a report on abnormal situations in smart space is generated.

[0122] Specifically, a knowledge graph retrieval mechanism for operation and maintenance reasoning is initiated to perform structured analysis on the currently identified equipment anomaly status information. Key fields such as equipment type, anomaly mode, severity level, and spatial location are extracted. Equipment type and anomaly mode are then used as entities to perform semantic matching and relationship queries on the pre-constructed smart space equipment anomaly event knowledge graph. The smart space equipment anomaly event knowledge graph consists of a multi-relationship network composed of equipment operation categories, typical fault manifestations, known causal chains, impact scope, intervention methods, and historical handling cases. Each anomaly event node in the graph is labeled with its corresponding type, occurrence conditions, and evolution path. Through high-precision matching, the corresponding mapping position of the current anomaly status in the graph is determined. Based on its equipment type and matched anomaly mode, related upstream and downstream causal paths and semantic context relationships are quickly retrieved, creating basic anomaly tracing data for source analysis, including multi-dimensional content such as event preconditions, concurrent phenomena, known causes, and intervention results. The cloud-based MindChain inference engine is used to progressively analyze the foundational data for anomaly tracing. Leveraging its large-model inference capabilities and incorporating MindChain methodology, the engine constructs a reasoning process, starting with "observed phenomena" and sequentially guiding the raw data through a reasoning path of "attribution analysis—mechanism modeling—possible evolution—intervention recommendations." During operation, the engine semantically compares the anomaly with historical cases and automatically constructs a causal path linked by multiple conditional relationships and event nodes. This identifies the possible triggers and underlying logic behind the current anomaly, generating structured inference tracing results. Outputs include inferred causes, potential root cause devices, impact chains, and prioritized intervention strategies. It also provides differentiated inference models for different device types to enhance explanatory power and decision-making accuracy. To corroborate the inference results and construct a temporal evolution chain, a smart space log database with an LSM tree storage structure is accessed, performing high-concurrency queries on historical records. The LSM tree architecture utilizes write optimization features, caching all write operations in a memory table and periodically merging them to disk, forming a persistent data organization method that supports both high-speed writing and controllable compression. During the retrieval process, the timestamp and spatial coordinates of the current abnormal event are used as the search criteria window. Historical logs are searched for operational anomaly records that are temporally adjacent and spatially close. Simultaneously, other nodes with the same device type or system coupling relationships are focused on to identify historical anomaly patterns that are significantly correlated with the current abnormal event in terms of spatiotemporal distribution. By comparing and analyzing the triggering conditions, duration, intervention feedback, and final handling status of these historical anomalies, spatiotemporal correlation data of abnormal events is constructed.By integrating and interpreting the inference and tracing results with the spatiotemporal correlation data of abnormal events, a structured intelligent space abnormal situation report is formed. The report displays the diagnostic results and inferred causes of the current abnormal equipment, as well as spatiotemporal information such as whether there are duplicates in historical data, whether it is in a high-risk area of ​​the system, and whether there are abnormal linkages between upstream and downstream. The report output includes an anomaly overview, causal chain diagram, risk level assessment, scope of impact definition, handling suggestions and auxiliary charts.

[0123] The process involves several steps: First, the thought chain reasoning engine analyzes the underlying data for anomaly tracing step by step to obtain the reasoning results. This includes: 1) Activating the observation and analysis module of the thought chain reasoning engine to extract the current state parameters, anomaly characteristics, and environmental conditions of the abnormal equipment from the anomaly tracing data, generating a structured description of the anomaly, and obtaining the observation results; 2) Running the causal reasoning module, based on the anomaly observation results and combined with the working principle and fault propagation mechanism of the smart space equipment, using reverse reasoning to trace the anomaly back to the possible root cause, generating multiple candidate fault cause hypotheses, and obtaining a set of cause hypotheses; 3) Executing the impact assessment reasoning module to perform forward reasoning verification on each hypothesis in the set of cause hypotheses, analyzing the possible chain reactions and impact range of each hypothesis, calculating the matching score between the hypothesis and the actual anomaly, and obtaining the hypothesis verification score; 4) Activating the decision suggestion generation module, based on the hypothesis verification score, selecting the most likely fault cause, and generating targeted handling strategy suggestions by combining historical handling experience and expert knowledge base, forming a complete reasoning chain including fault location, impact analysis, and handling suggestions, and obtaining the reasoning results.

[0124] In this embodiment, the LSM tree log storage system is accessed to query the historical operation records of smart space devices, identify historical anomaly patterns that are correlated with the current anomaly event in time and space, and obtain spatiotemporal correlation data of the anomaly event. This includes: constructing a hierarchical log storage structure based on the LSM tree architecture, organizing the smart space device operation records according to a three-level structure of memory table, immutable memory table, and disk storage, supporting high-concurrency log writing operations of more than 100,000 records per second, resulting in a high-performance log storage system; establishing a composite query mechanism of time index and spatial index based on the high-performance log storage system, constructing query conditions according to the occurrence time of the current anomaly event and device location information, retrieving historical anomaly records within a specified time window and spatial range, and obtaining a candidate set of correlated anomaly events; analyzing the time interval, spatial distance, and anomaly type similarity between each historical anomaly event and the current anomaly event in the candidate set of correlated anomaly events, calculating the spatiotemporal correlation score, and filtering out historical anomaly events with correlation scores exceeding a set threshold, obtaining strongly correlated historical anomaly events; extracting the anomaly pattern features, propagation path information, and processing result data of the strongly correlated historical anomaly events, constructing a multidimensional correlation graph containing time series correlation, spatial propagation correlation, and causal logic correlation, and obtaining spatiotemporal correlation data of the anomaly event.

[0125] In one specific embodiment, the method for optimizing smart space operation and maintenance based on data elements further includes the following steps:

[0126] Extract the location, fault type, and severity information of abnormal equipment from the abnormal situation report of the smart space, construct a three-dimensional model of the physical environment of the smart space, and mark the abnormal areas and equipment nodes to obtain a three-dimensional abnormal situation map;

[0127] Assess the risk propagation path and impact range of each abnormal area in the three-dimensional abnormal situation map, and calculate the regional risk level data;

[0128] Match regional risk level data with candidate operation and maintenance handling strategies in the historical operation and maintenance case library;

[0129] The candidate operation and maintenance handling strategies are rendered onto an interactive 3D interface. The risk level is displayed using color coding, the handling steps are dynamically marked, and mobile terminal control functions are integrated to obtain a comprehensive operation and maintenance decision-making scheme for the smart space.

[0130] Specifically, key structured information is extracted from the abnormal situation report, including the location coordinates of the abnormal equipment, the fault type label corresponding to the equipment type, and the severity level assessed based on the abnormal score and rule engine. The 3D physical model construction module of the smart space environment is invoked. This module unifies the modeling of building structure data, spatial layout diagrams, equipment deployment layers, and sensor distribution coordinates to form an interactive 3D spatial infrastructure. The abnormal equipment extracted from the anomaly report is mapped to its corresponding physical coordinates. Abnormal equipment nodes and their respective areas are labeled in the 3D model, and nodes are color-coded according to the semantic classification labels of fault types (e.g., red for high-temperature faults, orange for power anomalies, purple for network outages, etc.). Visual enhancements are also provided based on the severity grading criteria, using node volume, border thickness, or dynamic flashing effects, generating a 3D abnormal situation map covering both equipment and spatial levels. The risk propagation path and impact range of each abnormal area in the 3D abnormal situation map are then assessed. By introducing a spatial propagation modeling algorithm, each device node marked as abnormal is taken as a risk source. A risk propagation map is constructed by combining the building's internal spatial channel structure, equipment pipeline connections, thermal field, power network topology, and data communication links. The set of paths that the risk may propagate to other areas along the time axis and their propagation probabilities are calculated. Risk propagation paths are weighted based on structural connectivity, fault category coupling, and equipment state dependence, and parameters are corrected using actual diffusion cases from historical events to assess the potential impact range of each abnormal area. Based on this, multi-dimensional risk indicators are overlaid on each abnormal area, including risk source intensity, propagation density, number of affected nodes, and distribution density of critical equipment. Standardized regional risk level data is calculated comprehensively and mapped back to the 3D model as attribute layers attached to each corresponding area. Regional risk level data is used as a key search condition and matched with candidate operation and maintenance (O&M) handling strategies archived in the historical O&M case library. The historical O&M case library covers O&M response strategies under various equipment types, abnormal modes, severity levels, and spatial distribution scenarios, including automated recovery commands, manual inspection tasks, parameter adjustment suggestions, and multi-level handling solutions such as emergency power outages / isolation operations. The matching process is carried out through semantic vector matching and case index comparison algorithms. Based on a combination of conditions such as device type, anomaly type, risk level and spatial location, historical strategies are prioritized and filtered, and a set of candidate operation and maintenance handling strategies covering the entire region is output.Candidate operation and maintenance (O&M) strategies are rendered onto an interactive 3D visualization interface. This interface is based on a pre-constructed 3D anomaly map, dynamically loading strategy layers and highlighting areas of varying risk levels using color coding. For example, red indicates severe / high-risk areas, yellow indicates medium-risk areas, and green indicates controllable areas. Simultaneously, dynamic icons or sequence animations display the corresponding processing steps and intervention order for each area, such as equipment restart, power supply switching, manual inspection, or alarm transfer. The interface integrates a mobile control interface, allowing O&M personnel to select strategy nodes in real-time via tablets or mobile phones, remotely issue processing commands, view processing progress, and receive execution status feedback. This achieves integrated closed-loop control of anomaly identification, risk assessment, strategy selection, interactive decision-making, and remote response, forming a comprehensive O&M decision-making solution for the smart space.

[0131] The above describes the smart space operation and maintenance optimization method based on data elements in the embodiments of the present invention. The following describes the smart space operation and maintenance optimization device based on data elements in the embodiments of the present invention. Please refer to [link / reference]. Figure 2 One embodiment of the intelligent space operation and maintenance optimization device based on data elements in this invention includes:

[0132] The data acquisition module 201 is used to collect operational data from IoT terminals, edge nodes, and cloud platforms within the smart space, and to fuse the operational data to obtain fused data.

[0133] Encryption module 202 is used to identify and encrypt sensitive information in the fused data to obtain encrypted data;

[0134] The extraction module 203 is used to input encrypted data into the local model of each edge node to extract device operation features, and integrate the device operation feature set in the cloud aggregation server;

[0135] The anomaly identification module 204 is used to identify anomalies based on the equipment operation feature set and obtain equipment anomaly status information;

[0136] The source tracing analysis module 205 is used to perform source tracing analysis on the abnormal status information of the equipment and obtain an abnormal situation report of the smart space.

[0137] Through the collaborative efforts of the aforementioned components, and by constructing a device data correlation calculation model and a time coordination analysis mechanism, precise correlation and fusion of data across the three-layer architecture of IoT terminals, edge nodes, and cloud platforms were achieved. This effectively addresses the technical limitations of data silos and simple aggregation in traditional methods. A four-level security labeling system was established using a sensitive data intelligent hierarchical classification algorithm combined with AI large-scale model LoRA fine-tuning technology. Differential encryption protection using national cryptographic algorithms SM2 / SM3 / SM4 ensures personalized protection for data at different security levels, significantly improving the security protection capabilities of smart space data. Federated learning feature extraction and aggregation in encrypted state were achieved through fully homomorphic encryption technology. Distributed model training was completed while protecting the data privacy of each node, ensuring both data security and collaborative intelligence, thus solving the data leakage risk problem in traditional centralized learning. An adaptive anomaly detection mechanism was constructed by combining the isolated forest algorithm and reinforcement learning human feedback methods. This mechanism can dynamically optimize detection thresholds and adapt to changes in the operating modes of smart space devices, achieving higher detection accuracy and lower false alarm rate compared to traditional fixed threshold methods. By leveraging a thought-chain reasoning mechanism and a knowledge graph of abnormal events in smart space equipment, a technological leap has been achieved from simple anomaly alarms to in-depth root cause tracing. This provides maintenance personnel with interpretable anomaly analysis paths and comprehensive situational awareness reports, significantly improving fault handling efficiency. Through a 3D spatial situational map and a multi-dimensional risk assessment model, abstract anomaly information is transformed into intuitive visualizations. Combined with an intelligent maintenance decision recommendation engine to generate specific handling suggestions, a complete technical closed loop from anomaly detection to maintenance decision-making has been realized, greatly enhancing the intelligence level of smart space maintenance management.

[0138] Those skilled in the art will clearly understand that, for the sake of convenience and brevity, the specific working processes of the systems and units described above can be referred to the corresponding processes in the foregoing method embodiments, and will not be repeated here.

[0139] If the integrated unit is implemented as a software functional unit and sold or used as an independent product, it can be stored in a computer-readable storage medium. Based on this understanding, the technical solution of the present invention, in essence, or the part that contributes to the prior art, or all or part of the technical solution, can be embodied in the form of a software product. This computer software product is stored in a storage medium and includes several instructions to cause a computer device (which may be a personal computer, server, or network device, etc.) to execute all or part of the steps of the methods described in the various embodiments of the present invention. The aforementioned storage medium includes various media capable of storing program code, such as USB flash drives, portable hard drives, read-only memory (ROM), random access memory (RAM), magnetic disks, or optical disks.

[0140] The above-described embodiments are only used to illustrate the technical solutions of the present invention, and are not intended to limit it. Although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some of the technical features. Such modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the spirit and scope of the technical solutions of the embodiments of the present invention.

Claims

1. A smart space operation and maintenance optimization method based on data elements, characterized in that, include: The process involves collecting operational data from IoT terminals, edge nodes, and cloud platforms within a smart space, and then fusing this operational data to obtain fused data. Specifically, this includes: collecting operational data from each of the IoT terminals, edge nodes, and cloud platforms within the smart space; calculating the terminal-edge similarity between IoT terminal data and edge node data in the operational data; calculating the edge-cloud similarity and device time coordination between edge node data and cloud platform data in the operational data; and performing correlation and fusion based on the terminal-edge similarity, the edge-cloud similarity, and the device time coordination to obtain fused data. Sensitive information is identified and encrypted in the fused data to obtain encrypted data; The encrypted data is input into the local model of each edge node to extract device operation features, and the device operation feature set is integrated in the cloud aggregation server; Anomaly identification is performed based on the equipment operation feature set to obtain equipment abnormal state information. Specifically, this includes: converting the equipment operation feature set into equipment state feature vectors containing temperature, power, network, and time status; processing the equipment state feature vectors using the isolated forest algorithm, constructing a binary tree isolation structure through random sampling, and calculating the path depth of the equipment state feature vectors to obtain an equipment operation anomaly score; training a reinforcement learning model to analyze the historical distribution pattern of the equipment operation anomaly score, and dynamically optimizing the normal state boundary and anomaly judgment threshold by combining expert experience-annotated data to obtain an adaptive anomaly threshold; comparing the equipment operation anomaly score with the adaptive anomaly threshold, identifying equipment exceeding the normal boundary, and labeling the anomaly type and severity level to obtain equipment abnormal state information. Based on the abnormal device status information, an abnormality cause tracing analysis is performed to obtain a smart space abnormal situation report. Specifically, this includes: retrieving the smart space device abnormal event knowledge graph, matching the device type and abnormality pattern in the abnormal device status information, and creating basic data for abnormality tracing based on the device type and the abnormality pattern; running the thought chain reasoning engine to progressively analyze the basic data for abnormality tracing to obtain reasoning tracing results; accessing the LSM tree log storage system to query the historical operation records of the smart space devices, identifying historical abnormality patterns that are correlated with the current abnormal event in time and space, and obtaining spatiotemporal correlation data for the abnormal event; and integrating the reasoning tracing results and the spatiotemporal correlation data for the abnormal event to generate a smart space abnormal situation report.

2. The intelligent space operation and maintenance optimization method based on data elements according to claim 1, characterized in that, The calculation of edge-cloud similarity and device time coordination between edge node data and cloud platform data in the operational data includes: Extract the numerical range, data type, and trend characteristics of edge node data from the operational data to construct edge node feature vectors. Simultaneously, extract the corresponding feature parameters of cloud platform data to construct cloud platform feature vectors. The cosine similarity between the feature vectors of the edge nodes and the feature vectors of the cloud platform is calculated to obtain the edge-cloud similarity. A sliding time window is established based on the timestamp information of the running data. The arrival time difference and data update frequency of the edge node data and the cloud platform data within the same sliding time window are analyzed to obtain the time synchronization analysis results. Based on the time synchronization analysis results, the latency variance and frequency matching degree of data transmission between edge nodes and cloud platforms are calculated, and the device time coordination is calculated based on the latency variance and frequency matching degree.

3. The intelligent space operation and maintenance optimization method based on data elements according to claim 1, characterized in that, The process of identifying and encrypting sensitive information in the fused data to obtain encrypted data includes: A sensitive data labeling system is established based on the fused data, and the sensitive data labeling system includes public level, internal level, confidential level and top secret level; Based on the aforementioned sensitive data labeling system, the fused data is input into a large language model fine-tuned by LoRA for intelligent classification, resulting in intelligently classified data, which includes public-level data, internal-level data, confidential-level data, and top-secret-level data. The public-level data and the internal-level data are encrypted using the SM4 symmetric encryption algorithm, and the confidential-level data and the top-secret-level data are encrypted using a combination of SM2 digital signature, SM3 integrity check and SM4 symmetric encryption.

4. The intelligent space operation and maintenance optimization method based on data elements according to claim 1, characterized in that, The step of inputting the encrypted data into the local model of each edge node to extract device operating features, and integrating the device operating feature set in the cloud aggregation server, includes: The encrypted data is distributed to the local models of each edge node, and the device operating characteristics are extracted from the encrypted temperature, humidity, current, voltage and power data based on the local models. The device operation characteristics and weight parameters of the local model are encrypted and protected by a fully homomorphic encryption algorithm to ensure that the feature data remains in ciphertext state during network transmission, thus obtaining homomorphically encrypted feature data. Upload the homomorphic encryption feature data of each edge node to the cloud aggregation server, perform homomorphic addition and homomorphic multiplication operations to achieve weighted aggregation calculation in the ciphertext state, and obtain the global aggregation feature; The global model parameters of federated learning are updated based on the global aggregated features to obtain the updated model parameters. The updated model parameters are then distributed to each edge node and combined with the device operation features of each edge node to obtain the device operation feature set.

5. The intelligent space operation and maintenance optimization method based on data elements according to claim 4, characterized in that, The global model parameters of the federated learning are updated based on the global aggregated features to obtain the updated model parameters. These updated model parameters are then distributed to each edge node, and the device operation features of each edge node are combined to obtain a device operation feature set, including: The gradient of the loss function of the federated learning model is calculated based on the global aggregated features, and the weight matrix and bias vector of the local model are updated using the federated averaging algorithm based on the gradient of the loss function to obtain the updated model parameters. The updated model parameters are sent from the cloud to each edge node through an encrypted communication protocol. Each edge node synchronously updates the weight matrix and bias vector of its local model to obtain a parameter-synchronized edge node model. The edge node model that runs the parameter synchronization is used to re-extract the device operation features from the encrypted data to obtain the synchronized device operation features. The synchronized device operation characteristics of each edge node are aggregated to obtain a device operation characteristic set.

6. The intelligent space operation and maintenance optimization method based on data elements according to claim 1, characterized in that, The intelligent space operation and maintenance optimization method based on data elements also includes: Extract the location, fault type, and severity information of abnormal devices from the abnormal situation report of the smart space, construct a three-dimensional model of the physical environment of the smart space, and mark the abnormal areas and device nodes to obtain a three-dimensional abnormal situation map; Assess the risk propagation path and impact range of each abnormal area in the three-dimensional abnormal situation map, and calculate the regional risk level data; Match the regional risk level data with candidate operation and maintenance handling strategies in the historical operation and maintenance case library; The candidate operation and maintenance handling strategies are rendered onto an interactive 3D interface, and risk levels are displayed using color coding, handling steps are dynamically labeled, and mobile terminal control functions are integrated to obtain a comprehensive operation and maintenance decision scheme for the smart space.

7. A smart space operation and maintenance optimization device based on data elements, characterized in that, For executing the data-feature-based smart space operation and maintenance optimization method as described in any one of claims 1-6, the data-feature-based smart space operation and maintenance optimization device comprises: The data acquisition module is used to collect operational data from IoT terminals, edge nodes, and cloud platforms within the smart space, and to fuse the operational data to obtain fused data. The encryption module is used to identify and encrypt sensitive information in the fused data to obtain encrypted data; The extraction module is used to input the encrypted data into the local model of each edge node to extract device operation features, and integrate the device operation feature set in the cloud aggregation server; Anomaly identification module is used to identify anomalies based on the device operation feature set and obtain device anomaly status information; The source tracing analysis module is used to perform source tracing analysis based on the abnormal status information of the equipment and obtain an abnormal situation report of the smart space.

Images