Inspection system based on edge device gateway

Abstract

The invention relates to the technical field of artificial intelligence, provides an edge device gateway-based inspection system, and aims to improve the operation perception capability and intelligent inspection efficiency of a heating ventilation air conditioning system. The system comprises a data acquisition module, an edge gateway module, an inspection decision module and an inspection task management module, the data acquisition module is used for acquiring multi-source operation data, the edge gateway module is responsible for data preprocessing and homology feature extraction, the inspection decision module generates an optimization strategy based on multi-scale dynamic modeling and graph convolution reasoning, and the inspection task management module realizes task scheduling and execution feedback; the system has the capabilities of high-precision state recognition, abnormity early warning and energy efficiency optimization, and is suitable for intelligent routing inspection application of the heating ventilation air conditioner in a complex building environment.

Description

Technical Field

[0001] The present invention relates to the technical field of artificial intelligence, and in particular to an inspection system based on an edge device gateway. Background Art

[0002] With the intelligent development of urban infrastructure, the heating, ventilation, and air conditioning (HVAC) system undertakes crucial heating, cooling, and ventilation tasks in environments such as large buildings, data centers, hospitals, and industrial plants. To ensure the efficient, safe, and stable operation of the HVAC system, regular inspections and monitoring of the operating status have become an important part of operation and maintenance management. However, the traditional HVAC inspection system still has many deficiencies: First, during the inspection process, it involves various data types such as temperature, pressure, water quality, and electricity, and these data come from different protocols and equipment manufacturers. The traditional system lacks a unified specification and efficient data structure design in data integration and processing, resulting in data isolation and affecting subsequent unified analysis and strategy formulation. Second, the traditional system mostly uses static rules, empirical models, or simple statistical analysis methods to identify abnormalities or formulate strategies, and it is difficult to capture the non-stationary characteristics, mutation behaviors, and dynamic coupling relationships between multiple variables during equipment operation, resulting in low accuracy of anomaly detection and limited energy efficiency optimization capabilities. Summary of the Invention

[0003] The present invention provides an inspection system based on an edge device gateway. Aiming at problems such as the lagging response, weak data fusion ability, and rough strategy formulation of the existing intelligent inspection system, a data modeling method with both local sensitivity and global abstraction ability is proposed, and an inspection decision-making mechanism with non-stationary modeling and dynamic reasoning capabilities is constructed to achieve high-precision understanding of the operating status of the HVAC system and efficient strategy generation. The core innovation of the present invention lies in two aspects: First, the persistent homology analysis and local homology tensor modeling technology are introduced to construct a local-global homology feature screening method, perform topological scale analysis on the multi-variable inspection data collected by the system, and extract a structured and tensored homology feature set to enhance the system's perception ability of complex dynamic states. Second, a hybrid inference network integrating multi-scale dynamic perception and non-stationary modeling is proposed. Non-stationary modeling of data is performed through multi-time-scale sliding slices and gated fusion mechanisms, time series features are extracted by combining local perception convolution and multi-scale Transformer structures, and graph convolution mechanisms are used to achieve structured reasoning and strategy generation of the inspection status. The finally output inspection strategy has high precision, strong structure, and good interpretability, and can be widely applied to the intelligent inspection tasks of HVAC in high-standard and complex building scenarios, significantly improving the system's safety, response efficiency, and energy efficiency management level.

[0004] The present invention provides an inspection system based on an edge device gateway, which includes a data collection module, an edge gateway module, an inspection decision module, and an inspection task management module;

[0005] The data acquisition module is used to collect the operation parameter data, environmental state data and equipment state data of the heat source, heat network, heat exchange station and their supporting high-voltage and low-voltage electrical rooms in the HVAC system, and integrate and process the collected data to generate HVAC inspection data; the data acquisition module establishes a data communication connection with the edge gateway module through a standard industrial protocol to continuously and real-time upload the HVAC inspection data;

[0006] The edge gateway module receives the HVAC inspection data from the data acquisition module, cleans the noise data, eliminates the data abnormal points, unifies the continuous data sampling rate and standardizes the HVAC inspection data to obtain a multi-variable inspection data set; the inspection decision module called by the edge gateway module processes it;

[0007] The inspection decision module constructs a local-global homology feature screening method by introducing persistent homology analysis and local homology tensor modeling techniques, processes the multi-variable inspection data set through the local-global homology feature screening method to generate a homology screening feature set; constructs a non-stationary hybrid inference network by jointly introducing multi-scale dynamic perception and non-stationary modeling technique features; processes the homology screening feature set and the multi-variable inspection data set through the non-stationary hybrid inference network to generate an inspection strategy;

[0008] The inspection task management module formulates an inspection plan in combination with the inspection strategy, including regularly inspecting the heat source unit, real-time monitoring the water quality status of the heat exchange station, and periodically verifying the integrity of the monitoring camera screen coverage to ensure that all types of inspection tasks are executed in a timely and efficient manner.

[0009] Furthermore, the process of the inspection decision module generating the homology screening feature set specifically includes the following steps:

[0010] Step S1: Calculate the variable correlation matrix: Calculate the Pearson correlation coefficient of the multi-variable inspection data set to generate a symmetric correlation matrix;

[0011] Step S2: Construct a distance matrix: Map the symmetric correlation matrix into a distance matrix to generate an inspection variable distance matrix; this mapping ensures that the distance matrix satisfies the basic properties of distance metrics, including non-negativity, symmetry, and the triangle inequality;

[0012] Step S3: Persistent homology: Set the topological scale threshold. Based on the inspection variable distance matrix, gradually construct the corresponding Vietoris-Rips complex. During the construction of the Vietoris-Rips complex, continuously track the birth time and death time of the topological features in the Vietoris-Rips complex, count the death times of all topological features, and calculate the median of their death times as the optimal topological scale threshold; reconstruct the Vietoris-Rips complex based on the optimal topological scale threshold, and perform pruning processing on the multivariate inspection data set to obtain the pruned inspection feature set.

[0013] Step S4: Local homology: Based on the pruned inspection feature set, to further explore the local topological properties and local consistency, use the optimal topological scale threshold as the neighborhood radius to determine the local neighborhood, extract the local subgraph, construct the local Vietoris-Rips complex, and capture the local topological structure; calculate the local homology group for each local Vietoris-Rips complex, extract the local topological feature vectors of each homology dimension, and generate the homology screening feature set.

[0014] Further, Step S3 specifically includes: setting the topological scale threshold, based on the inspection variable distance matrix, gradually constructing the corresponding Vietoris-Rips complex. During the construction of the Vietoris-Rips complex, continuously track the birth time and death time of the topological features in the Vietoris-Rips complex, record them as a persistence bar chart to visualize the life cycle of the features, count the death times of all topological features, and calculate the median of their death times as the optimal topological scale threshold; reconstruct the Vietoris-Rips complex based on the optimal topological scale threshold, and perform pruning processing on the multivariate inspection data set to remove the feature nodes that are still isolated and have no connection relationship at this scale, and obtain the pruned inspection feature set.

[0015] The specific process of gradually constructing the corresponding Vietoris-Rips complex: At each topological scale threshold, if the distance between any two feature points in the inspection variable distance matrix is less than the topological scale threshold, establish a connection, and the connections between multiple feature points form higher-order simplices, including triangular complex structures and tetrahedral complex structures; as the topological scale threshold continues to increase, the overall topological structure continuously evolves, and the originally isolated features gradually become connected.

[0016] Further, the process of the inspection decision module generating the inspection strategy specifically includes the following steps:

[0017] Step B1: Align the homology screening feature set and the multivariate inspection data set in the time dimension and splice them in the feature dimension to obtain the fused feature sequence.

[0018] Step B2: Perform sliding slicing of the fused feature sequence at different time scales to obtain multi-scale time slices;

[0019] Step B3: Calculate the mean and standard deviation of each time slice of the multi-scale time slices, and calculate the mean and standard deviation after gated fusion by combining the gated fusion mechanism;

[0020] Step B4: Perform normalization processing on the fused feature sequence to obtain a normalized feature sequence, and perform inverse normalization by combining the mean and standard deviation after gated fusion to complete non-stationarity modeling and obtain a non-stationary feature representation;

[0021] Step B5: Construct a multi-scale hybrid Transformer-Graph network through local perception convolutional feature extraction, multi-scale Transformer time series modeling, and dynamic feature map inference mechanism, and process the non-stationary feature representation through the multi-scale hybrid Transformer-Graph network to generate an inspection strategy.

[0022] Furthermore, Step B5 specifically includes the following steps:

[0023] Step B51: Use sliding local convolution on the non-stationary feature representation to extract short-term time features and generate a local convolution feature sequence;

[0024] Step B52: Establish three-scale Transformer modules, each of which includes a local self-attention module and a multi-scale position encoding module; use the local convolution feature sequence as the input to generate a multi-scale feature set;

[0025] Step B53: Calculate the feature node matrix and the adjacency matrix based on the multi-scale feature set;

[0026] Step B54: Establish a graph convolution model, and process the feature node matrix and the adjacency matrix through the graph convolution model to generate a policy feature representation;

[0027] Step B55: Combine the policy feature representation and design 4 parallel regression heads, each regression head being a multi-layer perceptron, to generate an inspection strategy.

[0028] Adopting the above solution, the beneficial effects obtained by the present invention are as follows:

[0029] By introducing persistent homology analysis and local homology tensor modeling techniques, the present invention constructs a local-global homology feature screening method for multivariate inspection data of HVAC systems, achieving an in-depth understanding of the system's topological structure and efficient elimination of redundant information. Compared with traditional methods that rely on single physical or statistical features, the present invention effectively extracts structural features that are more sensitive to changes in operating states through a homology screening mechanism, improving the accuracy and expressiveness of inspection data modeling, and providing an input basis with more physical meaning and dynamic identification capabilities for subsequent strategy generation, thus solving the problem of "abundant data but low information utilization rate" in existing systems.

[0030] The present invention further constructs a hybrid inference network that integrates multi-scale dynamic perception and non-stationary modeling capabilities, which can fully identify the characteristics of changes in the operating states of HVAC systems at different time scales, and realizes dynamic weighting and modeling standardization between feature scales through a gating mechanism, enhancing the system's adaptability to complex non-linear and non-stationary behaviors. On this basis, short-term dependencies are extracted through local convolution, long-time series features are modeled by multi-scale Transformer, and the correlation map between states is inferred using a graph convolution structure, achieving multi-dimensional prediction of the system's operating trends and dynamic optimization of task strategies, thus effectively improving the accuracy and intelligence level of strategy output.

[0031] Through the integration and implementation of the above key technologies, the present invention realizes a comprehensive intelligent upgrade of the inspection tasks of HVAC systems, significantly improving the inspection efficiency and energy efficiency management level, and solving core problems such as task response lag, poor strategy generalization ability, and lack of systematic structural understanding in existing systems. At the same time, the inspection strategies generated by the present invention have multi-dimensional output capabilities such as priority ranking, anomaly warning, energy-saving suggestions, and scheduling order, greatly enhancing the stability, reliability, and controllability of system operation, and providing important technical support for intelligent building operation and maintenance. BRIEF DESCRIPTION OF THE DRAWINGS

[0032] Figure 1 It is a schematic diagram of the modules of an inspection system based on an edge device gateway provided by the present invention;

[0033] Figure 2 It is a schematic diagram of the connected component topological features and loop topological features provided in the second embodiment;

[0034] Figure 3 It is a schematic diagram of the cavity topological features provided in the second embodiment. DETAILED DESCRIPTION OF THE INVENTION

[0035] Next, the technical solutions in the embodiments of the present invention will be clearly and completely described in conjunction with the accompanying drawings in the embodiments of the present invention. Obviously, the described embodiments are only a part of the embodiments of the present invention, rather than all the embodiments. All other embodiments obtained by those of ordinary skill in the art based on the embodiments of the present invention without creative efforts belong to the scope of protection of the present invention.

[0036] Embodiment 1. According to Figure 1 , the present invention provides an inspection system based on an edge device gateway. The system includes a data acquisition module, an edge gateway module, an inspection decision-making module, and an inspection task management module;

[0037] The data acquisition module is used to collect the operation parameter data, environmental state data, and equipment state data of the heat source, heat network, heat exchange station, and its supporting high-voltage and low-voltage electrical rooms in the HVAC system, and integrate and process the collected data to generate HVAC inspection data. The data acquisition module establishes a data communication connection with the edge gateway module through a standard industrial protocol to continuously and real-time upload the HVAC inspection data;

[0038] The edge gateway module receives the HVAC inspection data from the data acquisition module, performs noise data cleaning, data anomaly point elimination, continuous data sampling rate unification, and standardization processing on the HVAC inspection data to obtain a multi-variable inspection data set. The edge gateway module calls the inspection decision-making module for processing;

[0039] The inspection decision-making module constructs a local-global homology feature screening method by introducing persistent homology analysis and local homology tensor modeling techniques, processes the multi-variable inspection data set through the local-global homology feature screening method to generate a homology screening feature set; constructs a non-stationary hybrid inference network by jointly introducing multi-scale dynamic perception and non-stationary modeling technique features; processes the homology screening feature set and the multi-variable inspection data set through the non-stationary hybrid inference network to generate an inspection strategy;

[0040] The inspection task management module formulates an inspection plan in combination with the inspection strategy, including regularly inspecting the heat source unit, real-time monitoring the water quality status of the heat exchange station, and periodically verifying the integrity of the monitoring camera screen coverage to ensure that various inspection tasks are executed in a timely and efficient manner.

[0041] Embodiment 2. According to Figure 2 , Figure 3 , based on Embodiment 1, in this embodiment, the process of the inspection decision-making module generating the homology screening feature set specifically includes the following steps:

[0042] Step S1: Calculate the variable correlation matrix: Calculate the Pearson correlation coefficient of the multi-variable inspection data set to generate a symmetric correlation matrix;

[0043] Step S2: Construct a distance matrix: Map the symmetric correlation matrix into a distance matrix to generate a distance matrix for inspection variables; this mapping ensures that the distance matrix satisfies the basic properties of distance metrics, including non-negativity, symmetry, and the triangle inequality.

[0044] Step S3: Persistent homology: Set a topological scale threshold. Based on the distance matrix of inspection variables, gradually construct the corresponding Vietoris-Rips complex. During the construction of the Vietoris-Rips complex, continuously track the birth time and death time of topological features in the Vietoris-Rips complex, count the death times of all topological features, calculate the median of their death times as the optimal topological scale threshold; reconstruct the Vietoris-Rips complex based on the optimal topological scale threshold, and perform pruning on the multivariate inspection data set to obtain a pruned set of inspection features.

[0045] Step S4: Local homology: Based on the pruned set of inspection features, to further explore local topological properties and local consistency, use the optimal topological scale threshold as the neighborhood radius to determine the local neighborhood, extract the local subgraph, construct the local Vietoris-Rips complex, and capture the local topological structure; calculate the local homology group for each local Vietoris-Rips complex, extract the local topological feature vectors of each homology dimension, and generate a set of homology screening features.

[0046] The local homology groups include:

[0047] 1.0-dimensional homology group: Connected component topological features.

[0048] 2.1-dimensional homology group: Loop topological features.

[0049] 3.2-dimensional homology group: Cavity topological features.

[0050] Example 3: This example is based on Example 2. In this example, Step S3 specifically includes: setting a topological scale threshold, based on the distance matrix of inspection variables, gradually constructing the corresponding Vietoris-Rips complex. During the construction of the Vietoris-Rips complex, continuously track the birth time and death time of topological features in the Vietoris-Rips complex, record them as a persistent bar chart to visualize the life cycle of the features, count the death times of all topological features, calculate the median of their death times as the optimal topological scale threshold; reconstruct the Vietoris-Rips complex based on the optimal topological scale threshold, and perform pruning on the multivariate inspection data set to remove feature nodes that are still isolated and have no connection relationship at this scale, to obtain a pruned set of inspection features.

[0051] The specific process of gradually constructing the corresponding Vietoris-Rips complex: At each topological scale threshold, if the distance between any two feature points in the inspection variable distance matrix is less than the topological scale threshold, a connection is established, and the connections between multiple feature points form higher-order simplices, including triangular complex structures and tetrahedral complex structures; as the topological scale threshold continues to increase, the overall topological structure continuously evolves, and originally isolated features gradually become connected.

[0052] Example 4. This example is based on Example 3. In this example, the process of the inspection decision module generating the inspection strategy specifically includes the following steps:

[0053] Step B1: Align the homology screening feature set and the multivariate inspection data set in the time dimension and splice them in the feature dimension to obtain a fused feature sequence.

[0054] Step B2: Perform sliding slicing of the fused feature sequence at different time scales to obtain multi-scale time slices.

[0055] Step B3: Calculate the mean and standard deviation of each time slice of the multi-scale time slices, and calculate the mean and standard deviation after gated fusion in combination with the gated fusion mechanism. The formulas used are as follows:

[0056] ;

[0057] ;

[0058] Among them, represents the mean after gated fusion, represents the standard deviation after gated fusion, represents the time scale index, represents the gated attention weight of the th time scale, represents the mean at the th time scale, represents the standard deviation at the

[0059] Step B4: Standardize the fused feature sequence to obtain a standardized feature sequence, and perform inverse normalization in combination with the mean and standard deviation after gated fusion to complete non-stationarity modeling and obtain a non-stationary feature representation. The formulas used are as follows:

[0060] ;

[0061] Among them, represents the non-stationary feature representation, represents the standardized feature sequence;

[0062] Step B5: Construct a multi-scale hybrid Transformer-Graph network through local perception convolutional feature extraction, multi-scale Transformer temporal modeling, and dynamic feature map inference mechanism. Process the non-stationary feature representation through the multi-scale hybrid Transformer-Graph network to generate an inspection strategy.

[0063] Embodiment 5. This embodiment is based on Embodiment 3. In this embodiment, the process of the inspection decision module generating an inspection strategy specifically includes the following steps:

[0064] Step R1: Align the homology screening feature set and the multi-variable inspection data set in the time dimension and splice them in the feature dimension to obtain a fused feature sequence.

[0065] Step R2: Perform sliding slicing of the fused feature sequence at different time scales to obtain multi-scale time slices.

[0066] Step R3: Calculate the mean and standard deviation of each time slice of the multi-scale time slices.

[0067] Step R4: Standardize the fused feature sequence to obtain a standardized feature sequence, and perform inverse normalization in combination with the mean and standard deviation to complete non-stationarity modeling and obtain a non-stationary feature representation.

[0068] Step R5: Process the non-stationary feature representation through a Transformer network to generate an inspection strategy.

[0069] Embodiment 6. This embodiment is based on Embodiment 4. In this embodiment, Step B5 specifically includes the following steps:

[0070] Step B51: Use sliding local convolution on the non-stationary feature representation to extract short-term time features and generate a local convolution feature sequence.

[0071] Step B52: Establish three-scale Transformer modules, each of which includes a local self-attention module and a multi-scale position encoding module; use the local convolution feature sequence as the input to generate a multi-scale feature set.

[0072] Step B53: Calculate the feature node matrix and the adjacency matrix based on the multi-scale feature set.

[0073] Step B54: Establish a graph convolution model and process the feature node matrix and the adjacency matrix through the graph convolution model to generate a policy feature representation.

[0074] Step B55: Combine the policy feature representation and design 4 parallel regression heads, each regression head being a multi-layer perceptron, to generate an inspection strategy.

[0075] Head 1: Patrol priority prediction;

[0076] Head 2: Equipment anomaly warning;

[0077] Head 3: Energy efficiency optimization suggestions;

[0078] Head 4: Patrol task scheduling order.

[0079] The above describes the present invention and its implementation manners. Such description is not restrictive. What is shown in the drawings is only one of the implementation manners of the present invention, and the actual structure is not limited thereto. Generally speaking, if those of ordinary skill in the art are inspired by it and, without departing from the purpose of the present invention, design similar structural manners and embodiments to this technical solution without creative efforts, they shall fall within the protection scope of the present invention.

Claims

1. An inspection system based on an edge device gateway, the system comprising an edge gateway module, the edge gateway module generating a multivariate inspection data set; characterized in that: The system also includes an inspection decision module; The inspection decision module constructs a local-global coherence feature screening method, processes a multivariate inspection data set through the local-global coherence feature screening method, and generates a coherence screening feature set; constructs a non-stationary hybrid inference network; processes the coherence screening feature set and the multivariate inspection data set through the non-stationary hybrid inference network to generate an inspection strategy.

2. According to claim 1, a patrol inspection system based on an edge device gateway is characterized in that: The construction method of the local-global homology feature screening method is: constructed by introducing persistent homology analysis and local homology tensor modeling technology.

3. The inspection system based on edge device gateway according to claim 1, characterized in that: The construction method of the non-stationary hybrid inference network is: jointly constructed by introducing multi-scale dynamic perception and non-stationary modeling technical features.

4. The inspection system based on edge device gateway according to claim 2 is characterized in that: The inspection decision module generates a process of generating a homology screening feature set, specifically comprising the following steps: Step S1: Calculate the Pearson correlation coefficient of the multivariate inspection data set to generate a symmetric correlation matrix; Step S2: Map the symmetric correlation matrix to a distance matrix to generate an inspection variable distance matrix; Step S3: Based on the inspection variable distance matrix, the corresponding Vietoris-Rips complex is gradually constructed, and the multivariate inspection data set is pruned to obtain the pruned inspection feature set; Step S4: Based on the pruned inspection feature set, a local Vietoris-Rips complex is constructed, and the local homology group of each local Vietoris-Rips complex is calculated to generate a homology screening feature set.

5. The inspection system based on edge device gateway according to claim 4 is characterized in that: Step S3 specifically includes: setting a topological scale threshold, and gradually constructing a Vietoris-Rips complex based on the inspection variable distance matrix; tracking the birth time and death time of the topological features in the Vietoris-Rips complex during the construction process, and calculating the median of its death time as the optimal topological scale threshold; reconstructing the Vietoris-Rips complex based on the optimal topological scale threshold, and pruning the multivariate inspection data set to obtain a pruned inspection feature set.

6. The inspection system based on edge device gateway according to claim 3 is characterized in that: The inspection decision module generates an inspection strategy, which specifically includes the following steps: Step B1: align and concatenate the homology screening feature set and the multivariate inspection data set to obtain a fused feature sequence; Step B2: Sliding slice the fused feature sequence to obtain multi-scale time slices; Step B3: Calculate the mean and standard deviation of each time slice in the multi-scale time slice, and combine the gated fusion mechanism to obtain the mean and standard deviation after gated fusion; Step B4: normalize the fused feature sequence, and perform denormalization based on the mean and standard deviation after gated fusion to obtain a non-stationary feature representation; Step B5: Construct a multi-scale hybrid Transformer-Graph network, process non-stationary feature representation, and generate inspection strategies.

7. The inspection system based on edge device gateway according to claim 6 is characterized in that: The construction method of the multi-scale hybrid Transformer-Graph network is: jointly constructed by introducing local-aware convolutional feature extraction, multi-scale Transformer temporal modeling and dynamic feature graph inference mechanism.

Images