Building intelligent inspection method and system based on AI algorithm

Abstract

The application provides a building intelligent inspection method and system based on an AI algorithm, which comprises: acquiring first information, second information, third information and fourth information; performing sensing data feature extraction according to the first information and the second information to generate a sensing data feature matrix; performing equipment state recognition according to the first information and the third information to generate an equipment state feature vector; performing anomaly detection according to the sensing data feature matrix and the equipment state feature vector to obtain an anomaly detection result; constructing a fault diagnosis model according to the fourth information to generate a fault diagnosis report; and performing path planning processing according to the first information and the fault diagnosis report to generate an inspection scheme. Through wavelet transform, multi-scale feature extraction, graph neural network and a dynamic clustering algorithm, the application can accurately identify the abnormal state of the equipment and the environment in the building, and can improve the comprehensiveness and accuracy of anomaly detection by combining cross-layer information propagation and correlation modeling.

Description

Technical Field

[0001] This invention relates to the field of intelligent inspection technology, and more specifically, to an intelligent building inspection method and system based on AI algorithms. Background Technology

[0002] In the current field of building management and inspection, with the continuous development of intelligence and digitalization, the demand for automated management of multi-story complex buildings is increasing. Modern buildings are equipped with a large number of sensors, cameras, and other devices that can collect environmental data, equipment status, and energy consumption information in real time. This data provides the foundation for intelligent building inspection. However, existing building inspection technologies typically rely on rule-driven manual inspection and simple static data analysis. These methods have the following problems: First, rule-driven analysis cannot dynamically adapt to the complex spatial structure and environmental changes of buildings, making it difficult to detect potential anomalies and faults in a timely manner; second, existing methods lack in-depth mining of the correlation between multimodal data (such as sensor data and image data), resulting in low anomaly detection accuracy; finally, for inspection path planning in multi-story buildings, most methods fail to consider the priority and energy consumption constraints of cross-floor equipment, resulting in low inspection efficiency and high maintenance costs. Existing technologies attempt to solve the above problems through fixed rules or simple statistical methods, but their effectiveness is limited in complex multi-story building scenarios, failing to meet the requirements of high efficiency and accuracy for intelligent inspection.

[0003] Given the shortcomings of existing technologies, there is an urgent need for an intelligent building inspection method and system based on AI algorithms. Summary of the Invention

[0004] The purpose of this invention is to provide an intelligent building inspection method and system based on AI algorithms to improve the aforementioned problems. To achieve the above objective, the technical solution adopted by this invention is as follows:

[0005] Firstly, this application provides a building intelligent inspection method based on AI algorithms, including:

[0006] Acquire first information, second information, third information and fourth information. The first information is a building topology map, which includes the building's spatial structure, hierarchical connections and equipment deployment locations. The second information includes sensor deployment locations and sensor data. The third information includes camera deployment locations and image information. The fourth information includes historical log data of the equipment.

[0007] Based on the first and second information, sensor data features are extracted, wavelet transform is used to perform multi-scale analysis of sensor data, and sensor data from different areas are aggregated in combination with the spatial location relationship of buildings to generate a sensor data feature matrix.

[0008] Based on the first information and the third information, the device status is identified using a preset deep learning model to identify the device's operating status and appearance changes, and to generate a device status feature vector.

[0009] Anomaly detection is performed based on the sensor data feature matrix and the device state feature vector. The spatial correlation between devices is modeled using a graph neural network to identify the anomaly propagation path and the scope of influence, and the anomaly detection result is obtained.

[0010] A fault diagnosis model is constructed based on the fourth information, and the anomaly detection results are used as the input values ​​of the fault diagnosis model. By analyzing the time series and spatial correlation between anomaly points, a fault diagnosis report is generated.

[0011] Based on the first information and the fault diagnosis report, path planning is performed. Combining the priority of abnormal equipment, the connectivity between floors, and energy consumption constraints, an inspection plan is generated. The inspection plan includes an optimal inspection path map and corresponding maintenance suggestions.

[0012] Secondly, this application also provides an intelligent building inspection system based on AI algorithms, including:

[0013] The acquisition module is used to acquire first information, second information, third information and fourth information. The first information is a building topology map, which includes the building's spatial structure, hierarchical connection relationship and equipment deployment location. The second information includes sensor deployment location and sensing data. The third information includes camera deployment location and image information. The fourth information includes historical log data of the equipment.

[0014] The extraction module is used to extract sensor data features based on the first information and the second information, perform multi-scale analysis of sensor data using wavelet transform, and combine the spatial location relationship of buildings to perform feature aggregation of sensor data in different areas to generate a sensor data feature matrix.

[0015] The identification module is used to identify the device status based on the first information and the third information, and to identify the device's operating status and appearance changes using a preset deep learning model, thereby generating a device status feature vector.

[0016] The detection module is used to perform anomaly detection based on the sensor data feature matrix and the device state feature vector, and to use graph neural networks to model the spatial correlation between devices, identify the anomaly propagation path and the scope of influence, and obtain the anomaly detection result.

[0017] The diagnostic module is used to construct a fault diagnosis model based on the fourth information, and to use the anomaly detection results as the input values ​​of the fault diagnosis model. By analyzing the time series and spatial correlation between anomaly points, a fault diagnosis report is generated.

[0018] The planning module is used to perform path planning based on the first information and the fault diagnosis report, and generate an inspection plan by combining the priority of abnormal equipment, the connection relationship between floors and energy consumption constraints. The inspection plan includes an optimal inspection path map and corresponding maintenance suggestions.

[0019] The beneficial effects of this invention are as follows:

[0020] This invention employs wavelet transform, multi-scale feature extraction, graph neural networks, and dynamic clustering algorithms to accurately identify abnormal states of equipment and the environment within buildings. By combining cross-floor information propagation and correlation modeling, it enhances the comprehensiveness and accuracy of anomaly detection. Regarding inspection path planning, this invention combines building topology, equipment priority, and energy consumption constraints, utilizing reinforcement learning and multi-objective optimization algorithms to generate energy-optimized inspection paths, significantly improving inspection efficiency and reducing maintenance costs. Furthermore, this invention addresses the uneven distribution of equipment and the complexity of floor connections in multi-story buildings by designing hierarchical feature aggregation and dynamic clustering algorithms, exhibiting good scenario adaptability and flexibly meeting the needs of buildings of different sizes and layouts. Attached Figure Description

[0021] To more clearly illustrate the technical solutions of the embodiments of the present invention, the accompanying drawings used in the embodiments will be briefly introduced below. It should be understood that the following drawings only show some embodiments of the present invention and should not be regarded as a limitation on the scope. For those skilled in the art, other related drawings can be obtained based on these drawings without creative effort.

[0022] Figure 1 This is a schematic diagram of the building intelligent inspection method described in the embodiments of the present invention;

[0023] Figure 2 This is a schematic diagram of the building intelligent inspection system described in an embodiment of the present invention.

[0024] Figure 3 This is a schematic diagram of the building intelligent inspection equipment described in an embodiment of the present invention. Detailed Implementation

[0025] To make the objectives, technical solutions, and advantages of the embodiments of the present invention clearer, the technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some, not all, of the embodiments of the present invention. The components of the embodiments of the present invention described and shown in the accompanying drawings can generally be arranged and designed in various different configurations. Therefore, the following detailed description of the embodiments of the present invention provided in the accompanying drawings is not intended to limit the scope of the claimed invention, but merely to illustrate selected embodiments of the invention. All other embodiments obtained by those skilled in the art based on the embodiments of the present invention without inventive effort are within the scope of protection of the present invention.

[0026] It should be noted that similar reference numerals and letters in the following figures indicate similar items; therefore, once an item is defined in one figure, it does not need to be further defined and explained in subsequent figures. Furthermore, in the description of this invention, terms such as "first," "second," etc., are used only to distinguish descriptions and should not be construed as indicating or implying relative importance.

[0027] Example 1:

[0028] This embodiment provides a building intelligent inspection method based on AI algorithms. (See also...) Figure 1 The figure shows that the method includes steps S100 to S600.

[0029] Step S100: Obtain first information, second information, third information and fourth information. The first information is a building topology map, which includes the building's spatial structure, hierarchical connection relationship and equipment deployment location. The second information includes sensor deployment location and sensing data. The third information includes camera deployment location and image information. The fourth information includes historical log data of the equipment.

[0030] Understandably, building topology maps, generated through spatial modeling techniques and topology analysis algorithms, not only provide the physical layout of the building but also clarify the specific distribution of equipment in three-dimensional space and the connection methods between floors, such as stairs, elevators, and ventilation ducts. These topology maps allow for the efficient construction of relationships between devices, providing precise spatial references for data integration. Sensor data includes the deployment locations of sensors and their real-time collected information, covering various physical parameters such as temperature, humidity, gas concentration, and energy consumption. Preferably, to ensure data validity and usability, the steps combine heterogeneous data integration technology, time alignment algorithms, and signal correction mechanisms to process different types of sensor data, addressing potential timestamp errors, signal noise, and data loss. Camera information, through the combination of deployment locations and image data, provides intuitive visual monitoring capabilities for the environment and equipment operating status. Field-of-view modeling and image coordinate mapping methods ensure that the camera's shooting area accurately corresponds to the building's physical space, enabling a one-to-one association between the video stream and the spatial layout, facilitating subsequent analysis. The equipment historical log provides data on the historical status and trends of equipment operation. This data is processed through log parsing and cleaning techniques to ensure the integrity and consistency of the information, and to extract key features such as equipment operation cycle and fault history to support a comprehensive understanding of building status.

[0031] Step S200: Extract sensor data features based on the first and second information, perform multi-scale analysis of sensor data using wavelet transform, and aggregate features of sensor data in different areas by combining the spatial location relationship of buildings to generate a sensor data feature matrix.

[0032] This step first uses initial information to clarify the distribution and regional division of sensors within the building space, ensuring that the data from each sensor can be correctly spatially assigned. Then, based on the second set of information, multi-scale analysis is performed on the data using wavelet transform. Wavelet decomposition is used to break down the time-series signal into components of different frequency bands, thereby capturing various change patterns in the sensor data, including low-frequency trends, periodic fluctuations, and high-frequency anomalies. This multi-scale feature extraction not only identifies long-term trends but also keenly captures short-term sudden changes, improving the understanding of the complex dynamics of sensor data. Building upon this, feature aggregation is performed on sensor data from different areas, considering the spatial relationships within the building. Feature aggregation introduces a spatial weighting mechanism, constructing a weighted feature matrix based on the physical adjacency and functional relevance of sensors in space. Specifically, the spatial correlation weights between sensors are calculated using a building topology map, and the feature values ​​of sensors within the same area are integrated using a weighted average method. The aggregated results generate spatially weighted comprehensive features for each floor or specific area, forming a sensor data feature matrix.

[0033] Step S300: Based on the first information and the third information, identify the device status, use a preset deep learning model to identify the device's operating status and appearance changes, and generate a device status feature vector;

[0034] The core of equipment status recognition lies in analyzing image data using a pre-defined deep learning model. Therefore, this step, targeting the equipment's operating status, extracts features from the equipment's display screen (e.g., numerical readings), indicator lights (e.g., color changes, flashing states), and overall appearance (e.g., surface damage, degree of contamination) using a Convolutional Neural Network (CNN) model. For dynamic changes in operating status, the frequency and stability of operating status switching are analyzed using models such as Long Short-Term Memory (LSTM) networks, combining continuous frame information from a time series. Furthermore, image differencing and edge detection techniques are used to capture subtle changes in the equipment's appearance, such as corrosion, cracks, or surface wear.

[0035] Step S400: Perform anomaly detection based on the sensor data feature matrix and device status feature vector, use graph neural network to model the spatial relationship between devices, identify the anomaly propagation path and the scope of influence, and obtain the anomaly detection result;

[0036] It should be noted that the spatial relationships between devices are defined by the first piece of information, and the physical distance, functional associations, and data similarities between devices are quantified by constructing an adjacency matrix. The construction of the spatial graph can capture potential anomaly associations between devices. For example, sensor anomalies between adjacent devices are propagating, and the connectivity between devices on different floors may amplify the impact of specific faults. Graph neural networks dynamically update the anomaly score of each device by aggregating node features layer by layer. Specifically, sensor data feature matrices and device state feature vectors serve as initial node features, which, combined with the adjacency relationships of the spatial graph, propagate and aggregate anomaly feature signals layer by layer to identify the associations between devices. This method can detect the propagation path of anomalies in the spatial dimension and quantify their impact range. The advantage of graph neural networks lies in their ability to capture nonlinear relationships between high-dimensional features, making them particularly suitable for the correlation analysis of equipment and environmental data in multi-story complex buildings.

[0037] Step S500: Construct a fault diagnosis model based on the fourth information, and use the anomaly detection results as the input value of the fault diagnosis model. By analyzing the time series and spatial correlation between anomaly points, generate a fault diagnosis report.

[0038] To deeply analyze the causal relationships and propagation logic among these anomalies, the diagnostic model constructs a temporal and spatial correlation structure for these anomalies using a dynamic Bayesian network. In the temporal dimension, the model utilizes causal inference techniques to analyze the time-series relationships between anomalies, uncovering temporal patterns and trigger sequences of fault propagation. In the spatial dimension, by combining the spatial correlation weights of equipment in the building topology map, the model identifies functional linkages and physical proximity between anomalies, thereby capturing the spatial characteristics of anomaly propagation. The fault diagnosis model employs deep learning classification models (such as Transformer or LSTM) to further enhance its fault type prediction capabilities. Based on the comprehensive feature vectors of anomalies, the model classifies potential fault modes and predicts possible fault types and their probability distributions. Combining time-series and spatial correlations, the model can also generate key cause chains for faults, clarifying the origin of the fault, the main propagation path, and possible secondary impacts.

[0039] Step S600: Based on the first information and the fault diagnosis report, perform path planning processing, and combine the priority of abnormal equipment, the connection relationship between floors and energy consumption constraints to generate an inspection plan. The inspection plan includes the optimal inspection path map and corresponding maintenance suggestions.

[0040] It's important to note that the first step in path planning is prioritizing inspection targets based on fault priority. Using a multi-attribute decision algorithm, the algorithm comprehensively considers the degree of equipment anomaly, the scope of fault propagation, and the equipment's criticality to building operation, assigning priority scores to target equipment and generating an optimized sequence of inspection targets. Next, using the spatial location and connectivity relationships in the building topology map, a shortest path algorithm is employed for path initialization, ensuring that the inspection path covers all target points while minimizing travel distance. To further optimize the path to meet energy consumption constraints, a reinforcement learning algorithm is used for dynamic path adjustment. The energy consumption constraint model, combined with inter-floor connectivity (such as elevator consumption and walking time), calculates the total energy consumption of the path and uses it as the optimization objective of reinforcement learning. In the reinforcement learning environment, each path selection adjusts the strategy according to the reward function, ultimately converging to the optimal inspection path that covers all target equipment and has the lowest energy consumption. Finally, maintenance recommendations are generated based on the inspection path and equipment status. These recommendations generate specific operational guidelines (such as replacing parts, cleaning equipment, or adjusting parameters) based on the fault type, equipment characteristics, and priority allocation in the diagnostic report. Meanwhile, the optimal route map presents the inspection route and key task nodes in a visual way, making it easier for inspection personnel to perform tasks efficiently.

[0041] Further, step S200 includes steps S210 to S240.

[0042] Step S210: Perform time series smoothing processing based on the second information. Obtain preliminary multi-scale features by multi-scale decomposition and noise reduction of the data.

[0043] Because time series data are often affected by various factors, such as momentary equipment failures, environmental fluctuations, or external interference, this noise can mask the true trend or periodic changes, thus requiring smoothing. Therefore, this step uses wavelet transform to decompose the time series into multiple frequency bands, identifying the variations in different frequency components. In this way, high-frequency noise and low-frequency trends can be separated, ensuring that the most important information in the signal is preserved. Specifically, wavelet decomposition can not only remove irregular fluctuations in the short term (such as occasional reading errors from sensors), but also highlight long-term trends and periodic characteristics (such as seasonal changes in temperature and humidity within buildings or the normal operating status of equipment).

[0044] Step S220: Perform hierarchical feature decomposition based on the preliminary multi-scale features and the first information, enhance the features of sensor data in different regions through spatial location correlation, and obtain a spatially weighted hierarchical feature matrix. Each row in the hierarchical feature matrix represents the comprehensive features of a level.

[0045] Therefore, in this step, hierarchical feature decomposition ensures that sensor data from adjacent areas or floors can be reasonably integrated in the feature matrix by considering the functional associations and physical proximity between floors, areas, or devices. Each row of the feature matrix is ​​enhanced spatially and hierarchically, thus making the data from each area exhibit more clearly defined spatial and functional characteristics. The hierarchical feature matrix is ​​represented as follows:

[0046]

[0047] Among them, S l,m T represents the comprehensive feature value of floor l in the m-th feature dimension; :,m W represents the multi-scale feature values ​​of all sensors in the m-th feature dimension; i,: P represents the spatial correlation weight between sensor i and other sensors; l,: This represents the cross-floor propagation weight between floor l and other floors; This represents the set of sensors on floor l; i represents the sensor number; l represents the floor number; m represents the feature dimension number. P represents the eigenvalue of floor k in the m-th feature dimension in the previous iteration; l,k denoted by α, which represents the information propagation weight between floor l and floor k; L represents the total number of floors; t-1 represents the previous iteration; α represents the cross-floor propagation adjustment coefficient; and σ is a nonlinear activation function.

[0048] Step S230: Perform feature aggregation based on the hierarchical feature matrix, capture feature distribution patterns and identify the feature distribution of similar regions through dynamic clustering algorithm, and obtain the regional feature aggregation result;

[0049] This step utilizes a clustering algorithm to intelligently group data from different regions, enabling the system to automatically identify and summarize areas with similar characteristics within a building. Compared to traditional manual region division or static analysis methods, dynamic clustering is more flexible and efficient, adaptively handling the complex and ever-changing environment and data patterns within building spaces. The dynamic clustering algorithm formula is as follows:

[0050]

[0051] Among them, R c C represents the combined feature vector of the c-th cluster; c S represents the set of floor indices belonging to the c-th cluster; c represents the cluster number; S l,: S represents the eigenvector of floor l; k,: Let λ represent the feature vector of floor k; λ is the feature balance factor; and K represents the total number of clusters.

[0052] Step S240: Based on the regional feature aggregation results, perform global modeling of cross-layer features, perform dimensionality reduction of cross-layer features through principal component analysis and extract key features to generate the final sensor data feature matrix.

[0053] While the regional feature aggregation results have already integrated sensor data in both spatial and hierarchical dimensions, these features still suffer from high-dimensional redundancy, especially across multiple floors or areas where some features are highly correlated. Principal component analysis (PCA) maps data from a high-dimensional space to a low-dimensional space through linear transformation, while preserving data variability (i.e., information content) to the greatest extent possible. In this step, PCA identifies the direction of maximum variance between cross-layer features, eliminates redundant information, and extracts the principal components that best represent data changes, thereby compressing the feature dimension from the original high-dimensional data space to a lower dimension. Through this dimensionality reduction process, the final sensor data feature matrix not only reduces the dimensionality of the data but also ensures that the key features contained in the feature matrix can most accurately describe the overall operating status and environmental changes of the building. These key features can reflect the main correlation patterns between equipment and the environment, improving the training efficiency and performance of subsequent models.

[0054] Example 2:

[0055] The only difference between this embodiment and Embodiment 1 is that step S300 includes steps S310 to S340.

[0056] Step S310: Perform image segmentation processing on the video image based on the third information. By segmenting the device region in the video frame and removing background and interference information, a device image sequence is obtained.

[0057] Preferably, to improve segmentation results, different image features, such as color, texture, shape, or motion information, are combined to accurately identify the location and boundaries of the device. For example, for HVAC (Heating, Ventilation, and Air Conditioning) systems, lighting equipment, or elevators in buildings, the shape and color features of the equipment are quite prominent. Utilizing this information can help the algorithm more accurately delineate the device area. After image segmentation, the resulting device image sequence can accurately represent the device's position and state changes within the video frame, eliminating interference from cluttered background information.

[0058] Step S320: Spatial feature calibration is performed based on the device image sequence and the first information. By spatially calibrating the device in the image and establishing a correspondence between the image features and the building spatial location of the device, the spatial positioning features of the device are obtained.

[0059] In this process, spatial calibration techniques from computer vision are used to extract key features from equipment images (such as equipment edges, corners, or specific identifiers) and match them with their actual coordinates in the building topology map to accurately calculate the spatial location of the equipment. Spatial positioning features help the system understand the exact location of equipment within the building and ensure that subsequent equipment status analysis (such as anomaly detection or fault diagnosis) can be accurately correlated with the equipment's location. For example, spatial calibration ensures that when a fault is detected in a piece of equipment, the system can accurately identify the floor, area, or even the specific room where the equipment is located, thereby optimizing inspection routes and maintenance tasks.

[0060] Step S330: Identify the operating status of the equipment based on its spatial positioning characteristics, model the state changes in consecutive frames using time series analysis, extract the switching frequency and stability information of the equipment, and obtain the state change information of the equipment.

[0061] Specifically, changes in equipment operating status typically manifest as a series of time-series data, especially in the complex environment of multi-story buildings where equipment status is influenced by various factors (such as load, environmental conditions, and equipment aging). By analyzing consecutive frames in a sequence of equipment images, these status changes can be captured. For example, equipment status might manifest as switching between on / off states (e.g., from off to on, or from normal operation to abnormal state), or changes in the equipment's appearance (e.g., changes in indicator light colors, or screen displays). These status changes are modeled using time-series data, and their switching frequency and stability are further analyzed. Long Short-Term Memory (LSTM) networks can effectively capture the temporal dependence of equipment status, not only identifying the current state but also extracting patterns and regularities of status changes from historical data. For example, frequent status switching may indicate potential faults or instability, while stable operation suggests the equipment is in a normal state. Furthermore, LTM models can learn the temporal characteristics of equipment status changes, predict potential state transitions, and promptly detect potential faults or anomalies.

[0062] Step S340: Identify changes in the appearance of the equipment based on the status change information and the equipment image sequence. Generate a device status feature vector by detecting changes in the appearance of the equipment and combining them with the operating status features.

[0063] Changes in equipment operating status and changes in appearance are often complementary. Changes in equipment status often leave traces on the appearance (such as deformation of the equipment casing due to increased temperature, or wear caused by abnormal operation of mechanical parts). Therefore, by combining changes in the appearance of equipment with its operating status characteristics (such as on / off status, temperature changes, energy consumption changes, etc.), a more comprehensive equipment condition assessment model can be formed.

[0064] Example 3:

[0065] The only difference between this embodiment and embodiment 1 or 2 is that step S400 includes steps S410 to S440.

[0066] Step S410: Perform feature fusion processing based on the sensor data feature matrix and device status feature vector. Through feature splicing and normalization, map the features of different modes to a unified space and dynamically enhance the mutation trend and operating status change features to obtain a multi-dimensional feature description.

[0067] First, the sensor data feature matrix and the device status feature vector originate from different data sources. Sensor data involves environmental parameters (such as temperature, humidity, and gas concentration), while device status features represent the device's operating status, appearance changes, and other operating parameters. Since these two types of data differ in feature dimensions, units, and value ranges, they need to be mapped to a unified feature space using methods such as feature concatenation and normalization.

[0068] Feature concatenation refers to merging feature vectors from sensor data and device status sequentially to form a unified feature vector. The concatenated vector can simultaneously contain feature information from both environmental and device data. Next, normalization is performed to standardize the features from different data sources, ensuring they have the same scale. This process ensures that data from different modalities are not affected by scale discrepancies, thus improving the model's training performance and facilitating subsequent processing by algorithms such as machine learning or deep learning models.

[0069] Specifically, the feature fusion process requires dynamically enhancing abrupt change trends and operational state change features. Abrupt change trends refer to sudden changes in the data, such as abnormal equipment startup or abrupt changes in environmental parameters; these changes are often signals of potential faults or anomalies. Enhancing these abrupt change features helps the system more accurately identify abnormal fluctuations in equipment or the environment, thereby improving fault early warning capabilities. Furthermore, operational state change features reflect the transition of equipment from one state to another, such as the switching between normal operation and fault states. Dynamically combining these two types of features provides richer input to subsequent models across multiple dimensions.

[0070] Step S420: Detect outliers based on multidimensional feature descriptions. Evaluate the isolation level of each device feature point by establishing a random segmentation tree for feature distribution. Set a dynamic threshold to filter outliers based on the outlier level of the device in the feature space. Obtain a preliminary set of outliers. The preliminary set of outliers includes the device ID and a preliminary outlier score.

[0071] Random segmentation trees are an anomaly detection method based on the isolation of data points. Their basic principle is to generate a tree-like structure by randomly segmenting the data space; the depth and structure of the tree reflect the degree of isolation of the data points. If a feature point (device status feature) is far from other points in the feature space, it means that the point is highly isolated in the feature space, usually indicating that the point is an anomaly. By constructing a random segmentation tree, device features that differ significantly from other data points can be effectively identified, and these points can be further marked as potential anomalies.

[0072] Building upon this, outlier degree refers to the deviation of a device's position in the feature space from the feature distribution of other devices. By analyzing the differences between device features and the overall data distribution, it is possible to further assess whether a device is an anomaly. Combining outlier degree with outlier degree analysis, dynamic thresholds can be set to filter out anomalies. The dynamic threshold setting is based on the device's feature distribution; that is, the criteria for judging anomalies are dynamically adjusted through learning from training data. For example, the threshold may be more lenient for some devices, while it may be more stringent for high-precision or high-load devices. This flexible dynamic threshold setting can adapt to the characteristics of different devices and environments, improving the accuracy of anomaly detection. This process generates a preliminary set of anomalies, including the ID of each device identified as an anomaly and its corresponding preliminary anomaly score. The anomaly score reflects the severity of the device anomaly; the higher the score, the greater the likelihood of the device being anomaly.

[0073] Step S430: Model the anomaly propagation based on the preliminary set of anomalies. By inputting the anomaly nodes into the graph neural network and combining the building topology, the anomaly signal is propagated layer by layer. After the feature update and aggregation mechanism of the network layer, the anomaly path graph of cross-layer propagation is obtained.

[0074] In practical applications, graph neural networks (Graph Neural Networks) can represent the physical proximity, functional correlation, and data similarity between devices through adjacency matrices, enabling accurate propagation modeling of abnormal signals. For example, when a critical device on a floor malfunctions, the signal may propagate to other floors via elevators or ventilation ducts. During propagation modeling, the Graph Neural Network can dynamically adjust the propagation path and signal strength based on the device type and its associated weights, thereby accurately locating the chain of abnormal propagation and the affected area. In this way, this step achieves abnormal correlation analysis in complex cross-floor environments and generates a visualized abnormal propagation path diagram, providing inspection personnel with an intuitive display of fault propagation and a clear basis for priority judgment.

[0075] Step S440: Classify anomalies based on the anomaly path graph, group device nodes in the path using spectral clustering algorithm and label anomaly priorities to obtain anomaly detection results.

[0076] It should be noted that the core of the spectral clustering algorithm lies in using the Laplacian matrix of the graph to perform feature decomposition on the correlation between device nodes. Clustering groups device nodes with similar behavioral patterns together, and the algorithm prioritizes anomalies based on their propagation characteristics. In this embodiment, spectral clustering can flexibly handle diverse anomaly patterns among devices in complex buildings, such as temperature and humidity anomalies on the same floor and energy consumption anomalies between devices across floors. By analyzing the node and edge weights of the path graph, the algorithm can identify the key device nodes most relevant to the source of the anomaly. Furthermore, this method's node grouping not only effectively narrows the scope of problem investigation but also determines the processing order of key devices or areas through priority labeling, avoiding resource waste and overlooking potential risks.

[0077] Example 4:

[0078] The only difference between this embodiment and any one of embodiments 1 to 3 is that step S500 includes steps S510 to S540.

[0079] Step S510: Based on the fourth information, perform historical data modeling, extract the evolution pattern of equipment operating status through time series analysis, and obtain the historical characteristics of equipment operating status;

[0080] Preferably, the analysis methods used include Autoregressive Moving Average (ARIMA), Long Short-Term Memory (LSTM) network, or wavelet transform. These methods can efficiently process multi-dimensional time series data and are especially suitable for complex scenarios where equipment states frequently switch or change abnormally.

[0081] Step S520: Perform time series anomaly correlation analysis based on anomaly detection results and historical characteristics. Use a dynamic Bayesian network to take the historical state of the device as state nodes, the anomaly propagation path as edge weights, and evaluate the impact of each anomaly point on the state changes of other devices to obtain the anomaly causal chain.

[0082] Dynamic Bayesian networks are a time-series analysis method based on probabilistic graphical models, which can effectively model the temporal and spatial relationships between devices. In this step, the historical state of a device is defined as a state node, and the anomaly propagation path is quantified as the edge weights in the network, thereby enabling a quantitative assessment of how abnormal signals are transmitted between devices and their impact range. For example, when a device experiences a high-temperature anomaly, this signal may affect the operating status of other devices through physical connections or functional associations. Dynamic Bayesian networks can track the energy consumption anomalies or alarm behaviors of other devices triggered by the high-temperature anomaly and clarify their causal relationships.

[0083] Step S530: Predict the fault type based on the abnormal causal chain, construct the fault feature vector using causal nodes and propagation relationships, and identify the fault mode through a deep learning classification model to obtain the prediction result, which includes the fault type and its probability distribution.

[0084] Understandably, deep learning classification models, preferably convolutional neural networks or long short-term memory networks, are used to learn the nonlinear mapping relationship between these high-dimensional features and failure modes. By training on a large amount of historical data, the model can accurately identify different types of failure modes (such as sensor failure, equipment aging, or network communication anomalies) and generate probability distributions of failure types based on input features. This method not only outputs the most likely failure type but also assesses the probability of other possible types, providing a comprehensive reference for maintenance decisions. For example, when a device displays an abnormally high temperature, and a propagation chain indicates a sharp increase in energy consumption in its adjacent devices, the classification model might predict the failure type as a cooling system failure with a probability of 85%, while also suggesting other possible causes, such as sensor failure (probability 10%) or external environmental interference (probability 5%).

[0085] Step S540: Based on the prediction results, conduct a fault impact assessment, quantify the scope of the fault's impact on building operation through correlation analysis, and generate a fault diagnosis report.

[0086] Specifically, the key steps in the assessment process include: 1. Determining the main fault types based on the probability distribution of fault prediction, and identifying potentially affected key equipment and areas by combining the abnormal propagation path; 2. Quantifying the scope of impact by considering the importance of equipment in building operation (e.g., core or non-core equipment) and the transmission effect of its failure on the overall system; 3. Identifying potential secondary problems or long-term risks caused by the fault through dynamic assessment of system performance indicators (e.g., temperature and humidity stability, energy efficiency, and operational reliability). For example, a fault in the HVAC system may directly lead to the failure of temperature and humidity control, while indirectly increasing the load on the power system, thereby affecting the overall energy consumption and comfort of the building.

[0087] Further, step S600 includes steps S610 to S640.

[0088] Step S610: Based on the first information and the fault diagnosis report, prioritize the inspection targets, calculate the relative proximity of each abnormal device to the positive and negative ideal solutions using the approximation ideal solution algorithm, and generate a priority ranking table.

[0089] It should be noted that this step, by combining building topology information and fault diagnosis reports, first extracts multiple evaluation indicators, including the severity, propagation range, importance, and maintenance complexity of equipment faults. These evaluation indicators are normalized to form a unified evaluation system, avoiding the influence of unit or magnitude differences on the ranking results. The core of the approximation-ideal-solution algorithm is to define a positive ideal solution (optimal state) and a negative ideal solution (worst-case state), representing the optimal and worst values ​​of each equipment indicator under ideal conditions, respectively. The algorithm determines the relative proximity of each equipment to the positive and negative ideal solutions by calculating the Euclidean distance between each device and the positive and negative ideal solutions; devices with higher proximity have higher priority. This method can scientifically and quantitatively assess the urgency and inspection value of each device, and the generated priority ranking table clearly marks the devices and areas that require priority inspection.

[0090] Step S620: Initialize the path according to the priority sorting table. By mapping the location of the inspection target equipment to the topology graph node, the shortest path algorithm is used to calculate the preliminary inspection path covering all target equipment.

[0091] This step combines priority sorting with spatial layout to ensure the adaptability of the inspection path in complex multi-story buildings, avoiding inefficiency caused by unnecessary repetition or long-distance movement.

[0092] Step S630: Optimize the path based on the preliminary inspection path. Dynamically adjust the path selection strategy by constructing a reinforcement learning model with the goal of minimizing the total energy consumption of the path, and generate the energy-optimal path.

[0093] Step S640: Based on the optimal energy consumption path, output the final inspection plan through path labeling and task allocation.

[0094] Specifically, firstly, the optimized energy-efficient path is used for path marking, clearly defining the specific location, priority, and operational requirements of each node (target equipment) in the inspection route. Path marking, combined with building topology information, visually displays the key areas and floors that inspection personnel need to traverse, marking the main nodes in the path and the methods of movement between them (such as stairs or elevators). Secondly, based on the task allocation of the inspection plan, the inspection tasks are rationally decomposed and assigned to different inspection personnel or teams to fully utilize human resources. Task allocation considers equipment priority, the coverage area of ​​the inspection path, and a balanced workload for inspection personnel to ensure scientific and rational resource allocation.

[0095] Example 5:

[0096] The only difference between this embodiment and any one of embodiments 1 to 4 is that step S630 includes steps S631 to S634.

[0097] Step S631: Based on the preliminary inspection path and the building topology map in the first information, construct a path energy consumption model by combining the energy consumption characteristics of different connection methods such as stair walking and elevator operation, and calculate the energy consumption of each path segment based on the path energy consumption model to generate a path energy consumption matrix. Each element in the path energy consumption matrix represents the unit energy consumption between any two abnormal device nodes.

[0098] This step first identifies the connectivity methods between floors based on the building topology map in the initial information, including walking up stairs and taking elevators. These connectivity methods differ significantly in energy and time consumption. The path energy consumption model quantifies these differences, assigning corresponding energy consumption weights to each connectivity method. For example, the energy consumption of walking up stairs is related to the physical exertion of inspection personnel, while the energy consumption of elevator operation depends on the elevator's load capacity and travel distance. The path energy consumption matrix comprehensively considers path length, connectivity method, and the physical location of equipment. This matrix provides a precise energy consumption evaluation standard for further optimization of inspection paths.

[0099] Step S632: Construct a reinforcement learning environment based on the path energy consumption matrix, define the reward function through energy consumption and priority, transform the path optimization problem into a path selection problem based on state-action reward, and obtain the environment model of the reinforcement learning optimization strategy.

[0100] The reinforcement learning environment is based on a building topology map and a path energy consumption matrix. It maps the spatial location of the inspection target equipment and the energy consumption relationship of the path segments into a state space, where each state represents the position of the inspection personnel at a certain node (equipment). The action space is defined as the possible choices from the current node to adjacent nodes, including the direction of the path and the connection method (such as elevator or stairs). The reward function is designed based on the energy consumption of the path selection and the equipment priority. The higher the equipment priority and the lower the path energy consumption, the larger the reward value, and vice versa.

[0101] Reinforcement learning environment models learn to select optimal paths in complex building scenarios by dynamically updating the relationship between states and actions. For example, when an inspection path passes through high-priority devices, the reward function tends to incentivize paths that reach these nodes faster; simultaneously, to reduce inspection costs, the reward function also penalizes the selection of high-energy-consuming paths. With the support of a path energy consumption matrix, the model can dynamically evaluate the energy consumption weight of each path segment, providing real-time optimization basis for inspection path selection.

[0102] Step S633: Train the path strategy according to the environment model. Initialize the Q table using the deep Q learning algorithm, randomly select the path as the initial strategy, and update the path selection strategy by repeatedly simulating and inspecting the path selection to obtain the initial energy-optimal path.

[0103] Specifically, starting with the path optimization problem, a Q-Table is initialized using the previously constructed environment model. The Q-Table records the reward value for each state-action pair, used to evaluate the merits of each path selection. In the initial stage, the path selection strategy adopts a random exploration approach, encouraging the model to try different paths to fully understand all possibilities in the environment. During training, the deep Q-learning algorithm continuously simulates the inspection path selection process, updating the state-action values ​​in the Q-Table using environmental feedback (i.e., energy consumption and priority reward function values). Specifically, in each iteration, the algorithm selects an action (path) based on the current state, then obtains the reward value from the environment, calculates the maximum Q-value for the next step, and updates the Q-value of the current state-action pair. This update follows the Bellman equation. By introducing techniques such as experience replay and target networks, deep Q-learning can effectively handle the high-dimensional state space and dynamic characteristics of path selection in complex building environments. Through repeated training, the model gradually converges to the optimal path selection strategy, i.e., the path selection scheme that maximizes the reward value given the inspection target and energy consumption constraints. The initial energy-optimal path covers all high-priority devices while minimizing the total energy consumption required for inspection.

[0104] Step S634: Perform global optimization based on the initial energy-optimal path, and balance energy consumption and path complexity through a multi-objective optimization algorithm to generate the final energy-optimal path.

[0105] The multi-objective optimization algorithm first constructs an objective function based on the energy consumption and complexity indices of each path segment in the preliminary path, while introducing weight coefficients to reflect the priority balance between the two. Through an iterative optimization process, the algorithm searches for the path scheme that optimizes the overall objective function from the candidate path set. For example, an energy-saving priority path may increase complexity during cross-layer switching, while a complexity-priority path may increase energy consumption. The algorithm dynamically adjusts among multiple path schemes, gradually approaching the optimal solution that satisfies both trade-offs. The final energy-optimal path not only covers all high-priority target devices but also significantly reduces redundant movement and cross-layer frequency during inspections, while maintaining a low energy consumption level. This path design better adapts to the actual inspection needs of multi-story complex buildings, improving the actual execution efficiency of inspections.

[0106] Example 6:

[0107] like Figure 2 As shown, this embodiment provides an AI-based intelligent building inspection system, which can implement the intelligent building inspection method described in any one of embodiments 1-5. Specifically, the intelligent building inspection system includes:

[0108] The acquisition module 901 is used to acquire first information, second information, third information and fourth information. The first information is a building topology map, which includes the building's spatial structure, hierarchical connection relationship and equipment deployment location. The second information includes sensor deployment location and sensing data. The third information includes camera deployment location and image information. The fourth information includes historical log data of the equipment.

[0109] The extraction module 902 is used to extract sensor data features based on the first information and the second information, perform multi-scale analysis of sensor data using wavelet transform, and combine the spatial location relationship of buildings to perform feature aggregation of sensor data in different areas to generate a sensor data feature matrix.

[0110] The identification module 903 is used to identify the device status based on the first information and the third information, and to identify the device's operating status and appearance changes using a preset deep learning model, thereby generating a device status feature vector.

[0111] The detection module 904 is used to perform anomaly detection based on the sensor data feature matrix and the device status feature vector, and to use graph neural network to model the spatial relationship between devices, identify the anomaly propagation path and the scope of influence, and obtain the anomaly detection result.

[0112] The diagnostic module 905 is used to construct a fault diagnosis model based on the fourth information, and to use the anomaly detection results as the input value of the fault diagnosis model. By analyzing the time series and spatial correlation between anomaly points, it generates a fault diagnosis report.

[0113] The planning module 906 is used to perform path planning based on the first information and the fault diagnosis report. It combines the priority of abnormal equipment, the connection relationship between floors and energy consumption constraints to generate an inspection plan. The inspection plan includes the optimal inspection path map and corresponding maintenance suggestions.

[0114] Example 6:

[0115] Corresponding to the above method embodiments, this embodiment also provides an AI-based intelligent building inspection device. The AI-based intelligent building inspection device described below and the AI-based intelligent building inspection method described above can be referred to and correspond to each other.

[0116] Figure 3 This is a block diagram illustrating an AI-based intelligent building inspection device 800 according to an exemplary embodiment. Figure 3As shown, the AI-based intelligent building inspection device 800 may include a processor 801 and a memory 802. The AI-based intelligent building inspection device 800 may also include one or more of a multimedia component 803, an I / O interface 804, and a communication component 805.

[0117] The processor 801 controls the overall operation of the AI-based intelligent building inspection device 800 to complete all or part of the steps in the AI-based intelligent building inspection method described above. The memory 802 stores various types of data to support the operation of the AI-based intelligent building inspection device 800. This data may include, for example, instructions for any application or method operating on the AI-based intelligent building inspection device 800, as well as application-related data such as contact data, sent and received messages, images, audio, video, etc. The memory 802 can be implemented using any type of volatile or non-volatile storage device or a combination thereof, such as Static Random Access Memory (SRAM), Electrically Erasable Programmable Read-Only Memory (EEPROM), Erasable Programmable Read-Only Memory (EPROM), Programmable Read-Only Memory (PROM), Read-Only Memory (ROM), magnetic storage, flash memory, magnetic disk, or optical disk. The multimedia component 803 may include a screen and an audio component. The screen may be, for example, a touchscreen, and the audio component is used to output and / or input audio signals. For example, the audio component may include a microphone for receiving external audio signals. The received audio signals may be further stored in the memory 802 or transmitted via the communication component 805. The audio component also includes at least one speaker for outputting audio signals. I / O interface 804 provides an interface between processor 801 and other interface modules, such as keyboards, mice, and buttons. These buttons can be virtual or physical. Communication component 805 is used for wired or wireless communication between the AI-based intelligent building inspection device 800 and other devices. Wireless communication includes Wi-Fi, Bluetooth, Near Field Communication (NFC), 2G, 3G, or 4G, or a combination thereof. Therefore, the corresponding communication component 805 may include a Wi-Fi module, a Bluetooth module, and an NFC module.

[0118] In an exemplary embodiment, an AI-based intelligent building inspection device 800 may be implemented by one or more application-specific integrated circuits (ASICs), digital signal processors (DSPs), digital signal processing devices (DSPDs), programmable logic devices (PLDs), field-programmable gate arrays (FPGAs), controllers, microcontrollers, microprocessors, or other electronic components to execute the aforementioned AI-based intelligent building inspection method.

[0119] In another exemplary embodiment, a computer-readable storage medium including program instructions is also provided. When executed by a processor, these program instructions implement the steps of the AI-based intelligent building inspection method described above. For example, the computer-readable storage medium may be the memory 802 including the program instructions, which may be executed by a processor 801 of an AI-based intelligent building inspection device 800 to complete the AI-based intelligent building inspection method described above.

[0120] The above description is merely a specific embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any changes or substitutions that can be easily conceived by those skilled in the art within the scope of the technology disclosed in the present invention should be included within the scope of protection of the present invention.

Claims

1. A building intelligent inspection method based on AI algorithms, characterized in that, include: Acquire first information, second information, third information and fourth information. The first information is a building topology map, which includes the building's spatial structure, hierarchical connectivity and equipment deployment locations. The second information includes sensor deployment locations and sensor data. The third information includes camera deployment locations and image information. The fourth information includes historical log data of the equipment. Based on the first and second information, sensor data features are extracted, wavelet transform is used to perform multi-scale analysis of sensor data, and sensor data from different areas are aggregated in combination with the spatial location relationship of buildings to generate a sensor data feature matrix. Based on the first information and the third information, the device status is identified, and a preset deep learning model is used to identify the device's operating status and appearance changes, generating a device status feature vector. Anomaly detection is performed based on the sensor data feature matrix and the device state feature vector. The spatial correlation between devices is modeled using a graph neural network to identify the anomaly propagation path and the scope of influence, and the anomaly detection result is obtained. A fault diagnosis model is constructed based on the fourth information, and the anomaly detection results are used as the input values ​​of the fault diagnosis model. By analyzing the time series and spatial correlation between anomaly points, a fault diagnosis report is generated. Based on the first information and the fault diagnosis report, path planning is performed, and an inspection plan is generated by combining the priority of abnormal equipment, the connectivity between floors and energy consumption constraints. Specifically, sensor data features are extracted based on the first and second information; wavelet transform is used to perform multi-scale analysis of the sensor data; and sensor data from different areas are aggregated by combining the spatial location relationships of buildings to generate a sensor data feature matrix, including: Based on the second information, time series smoothing is performed, and preliminary multi-scale features are obtained by multi-scale decomposition and denoising of the data. Based on the preliminary multi-scale features and the first information, hierarchical feature decomposition is performed. The features of sensor data in different regions are enhanced by spatial location correlation to obtain a spatially weighted hierarchical feature matrix. Each row in the hierarchical feature matrix represents a comprehensive feature of a level. Feature aggregation is performed based on the hierarchical feature matrix. A dynamic clustering algorithm is used to capture feature distribution patterns and identify the feature distribution of similar regions to obtain the regional feature aggregation result. Based on the aggregation results of the regional features, a global model of the cross-layer features is performed. Principal component analysis is used to reduce the dimensionality of the cross-layer features and extract key features to generate the final sensor data feature matrix.

2. The intelligent building inspection method according to claim 1, characterized in that, Based on the preliminary multi-scale features and the first information, hierarchical feature decomposition is performed. The features of sensor data from different regions are enhanced through spatial location correlation to obtain a spatially weighted hierarchical feature matrix. Each row in the hierarchical feature matrix represents a comprehensive feature of one level, including: The hierarchical feature matrix is ​​represented as follows: ； in, Indicates floor In the The comprehensive feature value across each feature dimension; Indicates that all sensors are in the first... Multi-scale feature values ​​across multiple feature dimensions; Indicates sensor Spatial correlation weights between sensors; Indicates floor Cross-floor propagation weights to other floors; Indicates floor The sensor set on; Indicates the sensor number; Indicates the floor number; Indicates the number of the feature dimension; Indicates the floor in the previous iteration In the Feature values ​​in the feature dimension; Indicates floor and floors The weight of information transmission between them; Indicates the total number of floors; Indicates the previous iteration; Indicates the cross-layer propagation modulation coefficient; It is a non-linear activation function.

3. The intelligent building inspection method according to claim 2, characterized in that, Feature aggregation is performed based on the hierarchical feature matrix. A dynamic clustering algorithm is used to capture feature distribution patterns and identify the feature distribution of similar regions, resulting in the region feature aggregation result, including: The formula for the dynamic clustering algorithm is as follows: ； in, Indicates the first The combined feature vector of each cluster; Indicates belonging to the first A set of floor indexes for each cluster; Indicates the cluster number; Indicates floor eigenvectors; Indicates floor eigenvectors; As a characteristic balance factor; This represents the total number of clusters.

4. The intelligent building inspection method according to claim 1, characterized in that, Based on the first information and the third information, device status identification is performed. A preset deep learning model is used to identify the device's operating status and appearance changes, generating a device status feature vector, including: Based on the third information, the video image is segmented. By segmenting the device region in the video frame and removing background and interference information, a device image sequence is obtained. Spatial feature calibration is performed based on the device image sequence and the first information. By spatially calibrating the device in the image and establishing a correspondence between the image features and the building spatial location of the device, the spatial positioning features of the device are obtained. The device's operating status is identified based on its spatial positioning characteristics. The state changes in consecutive frames are modeled using time series analysis methods. The switching frequency and stability information of the device are extracted to obtain the device's state change information. Based on the state change information and the device image sequence, the device appearance changes are identified. By detecting the device appearance changes and combining them with the operating state features, a device state feature vector is generated.

5. The intelligent building inspection method according to claim 1, characterized in that, Anomaly detection is performed based on the sensor data feature matrix and the device state feature vector. A graph neural network is used to model the spatial relationships between devices, identify the anomaly propagation path and its impact range, and obtain the anomaly detection results, including: Based on the sensor data feature matrix and the device state feature vector, feature fusion processing is performed. By feature splicing and normalization, features of different modes are mapped to a unified space and the mutation trend and operating state change features are dynamically enhanced to obtain a multi-dimensional feature description. Anomaly detection is performed based on the multidimensional feature description. The isolation degree of each device feature point is evaluated by establishing a random segmentation tree of the feature distribution. Anomalies are filtered by setting a dynamic threshold based on the outlier degree of the device in the feature space, resulting in a preliminary set of anomalies. The preliminary set of anomalies includes the device ID and a preliminary anomaly score. Anomaly propagation modeling is performed based on the preliminary set of anomalies. By inputting the anomaly nodes into a graph neural network and combining the building topology, the anomaly signal is propagated layer by layer. Through the feature update and aggregation mechanism of the network layer, the anomaly path graph propagating across layers is obtained. Anomalies are classified based on the anomaly path graph. The device nodes in the path are grouped using a spectral clustering algorithm, and the anomaly priority is marked to obtain the anomaly detection results.

6. The intelligent building inspection method according to claim 1, characterized in that, A fault diagnosis model is constructed based on the fourth piece of information, and the anomaly detection results are used as input values ​​for the fault diagnosis model. By analyzing the time series and spatial correlation between anomalies, a fault diagnosis report is generated, including: Based on the fourth information, historical data modeling is performed, and the evolution pattern of equipment operating status is extracted through time series analysis to obtain the historical characteristics of equipment operating status; Based on the anomaly detection results and the historical features, a time series anomaly correlation analysis is performed. A dynamic Bayesian network is used to take the historical state of the device as the state node, the anomaly propagation path as the edge weight, and the impact of each anomaly point on the state changes of other devices is evaluated to obtain the anomaly causal chain. Based on the abnormal causal chain, the fault type is predicted. A fault feature vector is constructed using causal nodes and propagation relationships. The fault mode is then identified through a deep learning classification model to obtain the prediction result, which includes the fault type and its probability distribution. Based on the predicted results, an impact assessment of the fault is conducted, and the scope of the fault's impact on building operation is quantified through correlation analysis to generate a fault diagnosis report.

7. The intelligent building inspection method according to claim 1, characterized in that, Based on the first information and the fault diagnosis report, path planning is performed. Combining the priority of abnormal equipment, inter-floor connectivity, and energy consumption constraints, an inspection plan is generated, including: Based on the first information and the fault diagnosis report, the inspection targets are prioritized and sorted. The approximation ideal solution algorithm is used to calculate the relative proximity of each abnormal device to the positive ideal solution and the negative ideal solution, and a priority sorting table is generated. The path is initialized according to the priority sorting table. By mapping the location of the inspection target equipment to the topology graph node, the shortest path algorithm is used to calculate the initial inspection path covering all target equipment. Based on the preliminary inspection path, the path is optimized by constructing a reinforcement learning model with the goal of minimizing the total energy consumption of the path, and the path selection strategy is dynamically adjusted to generate the energy-optimal path. Based on the energy-optimal path, the final inspection plan is output through path labeling and task allocation.

8. The intelligent building inspection method according to claim 7, characterized in that, Based on the preliminary inspection path, path optimization is performed. A reinforcement learning model is constructed with the goal of minimizing total path energy consumption to dynamically adjust the path selection strategy and generate the energy-optimal path, including: Based on the preliminary inspection path and the building topology map in the first information, a path energy consumption model is constructed by combining the energy consumption characteristics of different connection methods such as stair walking and elevator operation. Based on the path energy consumption model, the energy consumption of each path segment is cumulatively calculated to generate a path energy consumption matrix. Each element in the path energy consumption matrix represents the unit energy consumption between any two abnormal device nodes. A reinforcement learning environment is constructed based on the path energy consumption matrix. The reward function is defined by energy consumption and priority, and the path optimization problem is transformed into a path selection problem based on state-action reward, thus obtaining an environment model for the reinforcement learning optimization strategy. Based on the environmental model, path strategy training is performed. The Q-table is initialized using a deep Q-learning algorithm with random path selection as the initial strategy. The path selection strategy is updated by repeatedly simulating and inspecting path selection to obtain the initial energy-optimal path. Based on the initial energy-optimal path, global optimization is performed, and a multi-objective optimization algorithm is used to balance energy consumption and path complexity to generate the final energy-optimal path.

9. A building intelligent inspection system based on AI algorithms, characterized in that, include: The acquisition module is used to acquire first information, second information, third information and fourth information. The first information is a building topology map, which includes the building's spatial structure, hierarchical connectivity and equipment deployment locations. The second information includes sensor deployment locations and sensor data. The third information includes camera deployment locations and image information. The fourth information includes historical log data of the equipment. The extraction module is used to extract sensor data features based on the first information and the second information, perform multi-scale analysis of sensor data using wavelet transform, and combine the spatial location relationship of buildings to perform feature aggregation of sensor data in different areas to generate a sensor data feature matrix. The identification module is used to identify the device status based on the first information and the third information, and to identify the device's operating status and appearance changes using a preset deep learning model, thereby generating a device status feature vector. The detection module is used to perform anomaly detection based on the sensor data feature matrix and the device state feature vector, and to use graph neural networks to model the spatial correlation between devices, identify the anomaly propagation path and the scope of influence, and obtain the anomaly detection result. The diagnostic module is used to construct a fault diagnosis model based on the fourth information, and to use the anomaly detection results as the input values ​​of the fault diagnosis model. By analyzing the time series and spatial correlation between anomaly points, a fault diagnosis report is generated. The planning module is used to perform path planning processing based on the first information and the fault diagnosis report, and generate an inspection plan by combining the priority of abnormal equipment, the connectivity between floors and energy consumption constraints. The inspection plan includes an optimal inspection path map and corresponding maintenance suggestions. Specifically, sensor data features are extracted based on the first and second information; wavelet transform is used to perform multi-scale analysis of the sensor data; and sensor data from different areas are aggregated by combining the spatial location relationships of buildings to generate a sensor data feature matrix, including: Based on the second information, time series smoothing is performed, and preliminary multi-scale features are obtained by multi-scale decomposition and denoising of the data. Based on the preliminary multi-scale features and the first information, hierarchical feature decomposition is performed. The features of sensor data in different regions are enhanced by spatial location correlation to obtain a spatially weighted hierarchical feature matrix. Each row in the hierarchical feature matrix represents a comprehensive feature of a level. Feature aggregation is performed based on the hierarchical feature matrix. A dynamic clustering algorithm is used to capture feature distribution patterns and identify the feature distribution of similar regions to obtain the regional feature aggregation result. Based on the aggregation results of the regional features, a global model of the cross-layer features is performed. Principal component analysis is used to reduce the dimensionality of the cross-layer features and extract key features to generate the final sensor data feature matrix.

Images