Intelligent inspection data analysis method and system based on remote interaction and multi-agent

Abstract

The present application relates to the technical field of industrial intelligent inspection and multi-agent collaboration, and provides an intelligent inspection data analysis method and system based on remote interaction and multi-agent. The method collects device vibration, infrared thermal imaging, image defects, and environmental temperature and humidity multi-source data through multi-agent inspection terminals, and transmits the data remotely through Mesh adaptive networking. After data preprocessing, reliability evaluation, and scene feature extraction, the improved OVO attention algorithm is used to realize scene adaptive weighted fusion. Based on the fusion feature extraction time series change quantity, dynamic threshold is used to complete anomaly detection, and combined with multi-agent collaborative scheduling, the anomaly is accurately located. The detection result feedback optimizes the fusion weight, and generates a remote inspection instruction to adjust the collection strategy. The system is composed of a multi-agent inspection terminal, a remote interaction communication module, an inspection scene adaptive multi-source data fusion module, a multi-agent collaborative anomaly positioning module, a data storage module, and a remote terminal interaction interface.

Description

Technical Field

[0001] This invention relates to the field of intelligent inspection and data analysis technology, specifically to an intelligent inspection data analysis method and system based on remote interaction and multi-agent systems. Background Technology

[0002] In the field of industrial intelligent inspection, remote interaction and multi-agent collaborative inspection have become the mainstream development direction. The core requirement is to collect on-site inspection data through multiple agents (such as inspection robots, drones, and portable testing equipment), transmit this data remotely to the backend, and then use data analysis to identify equipment anomalies and provide early warnings of faults. However, current technologies suffer from the following technical problems, which seriously affect the accuracy, efficiency, and collaborative nature of inspection data analysis: The inspection data collected by multi-agent systems are multi-source heterogeneous data (such as equipment vibration data, infrared thermal imaging data, image defect data, and environmental temperature and humidity data). The dimensions, scales, and reliability of different types of data vary greatly. Existing fusion algorithms mostly adopt fixed-weight fusion or simple splicing methods, without considering the dynamic changes in the inspection scenario (such as changes in equipment operating load and changes in environmental interference intensity) and the dynamic fluctuations in data reliability, which leads to the distortion of fused data and affects the accuracy of anomaly identification. Especially in complex industrial scenarios, the false negative rate and false positive rate are both high. During remote interaction, communication between multiple agents and the backend is subject to delays and packet loss. Furthermore, there is a lack of collaborative data analysis mechanisms among the multiple agents. In existing technologies, each agent independently completes data collection and preliminary analysis, and the backend only processes the data of a single agent in isolation. This makes it impossible to leverage the spatial distribution advantages of multiple agents to achieve accurate anomaly location and root cause tracing. At the same time, the anomaly detection uses fixed thresholds, which cannot adapt to the dynamic changes in the device's operating status, further exacerbating the problems of misjudgment and missed anomaly detection. The existing system's remote interaction only realizes data transmission and simple command issuance, without establishing a closed-loop mechanism of "data acquisition-analysis-anomaly feedback-command adjustment-secondary acquisition". As a result, the backend cannot adjust the inspection strategy of the multi-agent in real time based on the data analysis results. The inspection resources of the multi-agent are not allocated reasonably, and there are duplicate inspections or inspection blind spots, which reduces the inspection efficiency.

[0003] In existing technologies, most intelligent inspection systems adopt a modular division of "data acquisition module - data preprocessing module - anomaly detection module - remote interaction module". The algorithms mostly use traditional general algorithms such as CNN, LSTM, and K-means, without making targeted improvements to address the above-mentioned problems. Furthermore, the correlation between the modules is weak, making it impossible to form an efficient and collaborative data analysis system, which is difficult to meet the high-precision and high-efficiency intelligent inspection requirements in industrial scenarios. Summary of the Invention

[0004] The purpose of this invention is to provide an intelligent inspection data analysis method and system based on remote interaction and multi-agent technology to solve the problems mentioned in the background art.

[0005] To achieve the above objectives, the present invention provides the following technical solution: an intelligent inspection data analysis method based on remote interaction and multi-agent systems, comprising the following steps: Step S1: The multi-agent inspection terminal collects multi-source heterogeneous inspection data on site. The multi-source heterogeneous inspection data includes equipment vibration data, infrared thermal imaging data, image defect data, and environmental temperature and humidity data. The multi-source heterogeneous inspection data is transmitted to the inspection scene adaptive multi-source data fusion module through the remote interactive communication module. Step S2: The inspection scenario adaptive multi-source data fusion module receives multi-source heterogeneous inspection data, performs data preprocessing, data reliability assessment, scene feature extraction, and improved OVO attention fusion algorithm calculation, and outputs fused feature data; Step S3: The feature-driven multi-agent collaborative anomaly localization and feedback module receives the fused feature data, performs temporal feature extraction, improved dynamic threshold anomaly detection, multi-agent collaborative anomaly localization, and identifies the anomaly type, anomaly location, and anomaly confidence. Step S4: The feature-driven multi-agent collaborative anomaly localization and feedback module feeds back the anomaly detection results to the inspection scenario adaptive multi-source data fusion module, adjusts the fusion weight of the inspection scenario adaptive multi-source data fusion module, and generates anomaly feedback information and remote inspection instructions, which are transmitted to the remote terminal interaction interface and multi-agent inspection terminal through the remote interactive communication module. Step S5: The multi-agent inspection terminal receives remote instructions, adjusts the inspection strategy, performs secondary data collection, and repeats steps S2-S4 to form a closed loop; the data storage module stores all inspection data, fusion results, abnormal information, and instruction logs. Step S6: Staff members view relevant information through the remote terminal interactive interface, issue fault handling instructions, and complete intelligent inspection data analysis and anomaly handling.

[0006] Preferably, the specific implementation logic of step S2 is as follows: Step S2.1: Data Reception and Preprocessing: Receive multi-source heterogeneous inspection data transmitted by the multi-agent inspection terminal, including equipment vibration data. Infrared thermal imaging data Image defect data Ambient temperature and humidity data Preprocessing of various data types to eliminate noise and outliers: Step S2.2: Data reliability assessment and scene feature extraction: S2.2.1: Calculate the reliability index for each type of preprocessed data. , This reflects the reliability of the data and avoids the impact of unreliable data on the fusion results. ,in, For the first Signal-to-noise ratio of the data type For the first Consistency of class data (the deviation coefficient of data collected by the same agent in three consecutive data collections). , These are the weighting coefficients. , (Based on the priority settings of the inspection data, the signal-to-noise ratio has a higher weight.) The range of values ​​is , The closer the value is to 1, the higher the data reliability. S2.2.2: Extracting dynamic features of the inspection scene Including equipment operating load (Value range 0-1, 1 represents full load), environmental disturbance intensity (Value range 0-1, 1 indicates the strongest interference), calculated from auxiliary data (such as device current data and environmental noise data) collected by multiple agents, and used to dynamically adjust the fusion weights; Step S2.3: Calculation of the improved adaptive OVO attention fusion algorithm for the inspection scenario: Abandoning the shortcomings of traditional OVO attention mechanisms that only consider inter-modal interactions and do not take into account scene features and data reliability, we have improved the design of the fusion weight calculation method to achieve adaptive fusion of multi-source data: S2.3.1: Calculate the OVO attention weights for each type of data in step S2.1. , indicating the first Importance of interaction between class data and all other classes of data: ,in For the similarity calculation function, an improved cosine similarity is adopted, which is adjusted in combination with the feature differences of the inspection data to avoid the problem of poor adaptability of traditional cosine similarity to heterogeneous data. Indicates except the first A collection of all other data types besides Class Data; S2.3.2: Combining data reliability indicators Dynamic characteristics of inspection scenarios Calculate the adaptive adjustment coefficient for the scenario : When the equipment is running at full load And environmental interference is weak hour, Take the maximum value, as this corresponds to the highest data fusion weight; when environmental interference is strong... And the equipment is operating at low load. hour, Reduce, minimize the impact of unreliable data; S2.3.3: Calculate the final fusion weights This enables dynamic adaptive adjustment of weights. ; S2.3.4: Based on the final fusion weights, the preprocessed multi-source data is weighted and fused to obtain fused feature data. : ,in For the first Preprocessing results of class data, It is a one-dimensional fusion feature vector containing core information from multiple source data, used for subsequent anomaly detection; Step S2.4: Output of fused features: The fused feature data... The data is transmitted to the feature-driven multi-agent collaborative anomaly localization and feedback module, while the fusion weights are also... Data reliability indicators The data is stored in the data storage module to support subsequent anomaly tracing and weight optimization.

[0007] Preferably, the specific details of preprocessing various types of data to eliminate noise and outliers in step S2.1 are as follows: An improved adaptive wavelet denoising algorithm is used to denoise the vibration data. Noise reduction processing is performed to remove environmental interference noise during equipment operation. The formula is as follows: ,in The vibration data after noise reduction. The number of wavelet basis functions. These are wavelet coefficients. For improved wavelet basis functions, It is a time series. The translation parameters of the wavelet basis functions are: For scale parameters, The value is adaptively adjusted based on the frequency range of the vibration data (the normal vibration frequency of the inspection equipment is 50-100Hz). When the vibration frequency is higher than 80Hz, When the value is 0.8, below 50Hz, Take 1.2, and the middle range is 1.0; Infrared thermal imaging data Perform grayscale normalization processing to map pixel values ​​to Range, eliminating grayscale differences caused by different shooting distances and ambient temperatures; for image defect data Contour extraction and feature quantization are performed, extracting feature parameters including but not limited to the area, perimeter, and gray-level mean of defects, and converting them into one-dimensional feature vectors; environmental temperature and humidity data are also analyzed. Standardize the data to eliminate the influence of dimensions.

[0008] Preferably, the specific implementation steps of step S3 are as follows: Step S3.1: Fusion of feature reception and temporal feature extraction: Receive fused feature data output by the adaptive multi-source data fusion module for inspection scenarios Simultaneously, historical fusion feature data (fusion feature sequences within the past hour) is retrieved from the data storage module to extract the temporal variation features of the fusion features, including the feature change rate. Characteristic fluctuation variance Dynamic threshold calculation for anomaly detection: ; ,in For the fused feature data at the current moment, The fused feature data from the previous time step. Number of historical data ( (corresponding to 1 hour, 1 data point per minute). The mean of the historical fusion feature data; Step S3.2: Calculation of the improved dynamic threshold anomaly detection algorithm: Overcoming the limitations of traditional fixed thresholds in adapting to dynamic changes in equipment operating status, a dynamic threshold anomaly detection algorithm is designed based on fused temporal features and historical anomaly data to achieve accurate anomaly identification. S3.2.1: Calculate the dynamic anomaly threshold This threshold is dynamically adjusted according to the device's operating status and fusion characteristics. ,in For adaptive threshold coefficients, Calibrate based on historical anomaly data. The rate of change weighting coefficient ( (To avoid excessive threshold fluctuations caused by feature mutations) When fusing features If the condition is met, it is considered abnormal; otherwise, it is considered normal. S3.2.2: Calculate the anomaly confidence level This is used to determine the severity of the anomaly and avoid misjudgment. ,in The fusion weights output by the adaptive multi-source data fusion module for inspection scenarios. As a data reliability indicator; The range of values ​​is , If the condition is deemed serious, it requires immediate attention. At that time, it was determined to be a minor abnormality, requiring close monitoring; If the alarm is false, the anomaly is eliminated. S3.2.3: Anomaly type identification, combining the fusion weights and feature differences of multi-source data to trace the root cause of the anomaly: if the fusion weights of vibration data... And abnormal confidence level The problem is determined to be a mechanical failure of the equipment; if the fusion weight of the infrared data is... and The fault was determined to be an overheating device; if the image data fusion weights and This was determined to be a defect in the appearance of the equipment. Step S3.3: Multi-agent collaborative anomaly localization: Based on anomaly type and anomaly confidence, a multi-agent cooperative scheduling algorithm is designed to leverage the spatial distribution advantages of multiple agents to achieve accurate anomaly localization. Step S3.4: Remote interactive feedback and closed-loop optimization: S3.4.1: Generate abnormal feedback information, including abnormal location, abnormal type, abnormal confidence level, and secondary detection results, and transmit it to the remote terminal interactive interface through the remote interactive communication module for staff to view. At the same time, generate remote instructions (such as key inspection instructions and fault handling instructions). S3.4.2: The anomaly detection results (anomaly type, anomaly confidence level) Feedback is sent to the inspection scenario adaptive multi-source data fusion module to adjust the scenario adaptive adjustment coefficient of the inspection scenario adaptive multi-source data fusion module. If it is a serious abnormality Data source weight corresponding to the anomaly type Increase by 20%, that is Multiply by 1.2 to increase the fusion priority of this type of data; if it is a false alarm. The corresponding data source Multiply by 0.8 to reduce its weight and avoid misjudgment again; S3.4.3: Based on the anomaly location results, the remote interactive communication module sends an inspection strategy adjustment instruction to the multi-agent inspection terminal, such as adjusting the inspection frequency of the abnormal area (from once every hour to once every 10 minutes). Idle agents go to the abnormal area for continuous monitoring, realizing a remote interactive closed loop of "data collection-fusion-detection-feedback-strategy adjustment".

[0009] Preferably, the specific implementation steps of step S3.3, multi-agent collaborative anomaly localization, are as follows: S3.3.1: Retrieve real-time location information of each agent Current inspection task status , Indicates free time. Indicates being busy. , Given the number of agents, calculate the distance from each agent to the suspected location of the anomaly. : ,in, For the first The real-time location of each agent. This is a suspected location of anomaly (preliminarily determined by the location collected by the agent corresponding to the fused features); S3.3.2: Calculate the scheduling priority of each agent. The higher the priority, the more likely it is to be scheduled to the suspected anomaly location for secondary detection: ,in, Weights for agent types (matched to anomaly types, such as vibration detection agents in the event of mechanical failure). Other intelligent agents ); S3.3.3: Select the two highest priority agents (to avoid detection errors by a single agent), and dispatch them to the suspected anomaly location to collect secondary data. The collected data is transmitted to the inspection scenario adaptive multi-source data fusion module for re-fusion and detection to confirm the anomaly location and type, and achieve accurate anomaly localization.

[0010] Preferably, an intelligent inspection data analysis system based on remote interaction and multi-agent technology is used to implement an intelligent inspection data analysis method based on remote interaction and multi-agent technology. The system includes: Multi-agent inspection terminal: It consists of multiple intelligent inspection devices of different types (such as vibration detection agent, infrared inspection agent, and image acquisition agent), which are used to collect on-site inspection data, receive remote interactive commands, execute inspection tasks, and transmit the collected data to the inspection scene adaptive multi-source data fusion module through the remote interactive communication module. Remote interactive communication module: It adopts Mesh adaptive networking technology to realize data transmission and command interaction between multiple agents and the backend, and between multiple agents, reducing transmission latency and packet loss rate, and ensuring the real-time performance and stability of data transmission; The inspection scenario adaptive multi-source data fusion module receives multi-source heterogeneous inspection data transmitted by multi-agent inspection terminals, and adopts an improved inspection scenario adaptive OVO attention fusion algorithm to solve the problems of distortion in multi-source heterogeneous data fusion and difficulty in dynamic weight adjustment, and outputs high-precision fusion feature data. The feature-driven multi-agent collaborative anomaly localization and feedback module receives fused feature data output from the inspection scenario adaptive multi-source data fusion module. It adopts an improved dynamic threshold anomaly detection and multi-agent collaborative scheduling algorithm to achieve accurate anomaly identification, localization and root cause tracing. At the same time, it generates remote interaction commands and feeds them back to the multi-agent inspection terminal through the remote interaction communication module to realize real-time adjustment of the inspection strategy and form a closed loop. Data storage module: Used to store raw inspection data, fused feature data, anomaly detection results, remote interaction commands and inspection logs collected by multiple agents, providing data support for subsequent data analysis and traceability; Remote terminal interaction interface: Allows staff to view inspection data, fusion results, and abnormal information, issue remote inspection commands, and realize remote interaction with multiple intelligent agents.

[0011] Compared with the prior art, the beneficial effects of the present invention are: This invention utilizes an improved OVO attention fusion algorithm for adaptive inspection scenarios in an adaptive multi-source data fusion module. By combining data reliability with the dynamic characteristics of inspection scenarios, it achieves adaptive adjustment of fusion weights. Compared with traditional fusion algorithms, this improves fusion accuracy and effectively reduces the false negative and false positive rates of anomaly detection. This invention utilizes an improved dynamic threshold anomaly detection and multi-agent collaborative scheduling algorithm based on a feature-driven multi-agent collaborative anomaly localization and feedback module. It achieves accurate anomaly identification and localization based on the fused feature data of the inspection scenario adaptive multi-source data fusion module. Simultaneously, it establishes a feedback closed loop with the inspection scenario adaptive multi-source data fusion module to further optimize fusion accuracy. The multi-agent collaborative localization error is controlled within 0.5 meters, and the anomaly detection response time is shortened. This invention establishes a remote interactive closed loop of "data acquisition-fusion-detection-feedback-strategy adjustment", which solves the problem that the remote interaction in the prior art only realizes data transmission and has no closed loop optimization. The multi-agent inspection resource allocation is more reasonable, the inspection efficiency is improved, and the duplicate inspection and blind spots are reduced. Attached Figure Description

[0012] Figure 1 This is a schematic diagram of the method flow structure of the present invention; Figure 2 This is a schematic diagram of the workflow of the inspection scenario adaptive multi-source data fusion module of the present invention; Figure 3 This is a schematic diagram of the workflow of the multi-agent collaborative anomaly localization and feedback module driven by fusion features of the present invention. Detailed Implementation

[0013] 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 embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0014] Example 1: Please see Figure 1-3 This invention provides a technical solution: an intelligent inspection data analysis method based on remote interaction and multi-agent systems, comprising the following steps: Step S1: The multi-agent inspection terminal collects multi-source heterogeneous inspection data on site. The multi-source heterogeneous inspection data includes equipment vibration data, infrared thermal imaging data, image defect data, and environmental temperature and humidity data. The multi-source heterogeneous inspection data is transmitted to the inspection scene adaptive multi-source data fusion module through the remote interactive communication module. Step S2: The inspection scenario adaptive multi-source data fusion module receives multi-source heterogeneous inspection data, performs data preprocessing, data reliability assessment, scene feature extraction, and calculation using an improved OVO attention fusion algorithm, and outputs fused feature data; the specific implementation logic is as follows: Step S2.1: Data Reception and Preprocessing: Receive multi-source heterogeneous inspection data transmitted by the multi-agent inspection terminal, including equipment vibration data. Infrared thermal imaging data Image defect data Ambient temperature and humidity data Preprocessing of various data types to eliminate noise and outliers: Details are as follows: An improved adaptive wavelet denoising algorithm is used to denoise the vibration data. Noise reduction processing is performed to remove environmental interference noise during equipment operation. The formula is as follows: ,in The vibration data after noise reduction. The number of wavelet basis functions. These are wavelet coefficients. For improved wavelet basis functions, It is a time series. The translation parameters of the wavelet basis functions are: For scale parameters, The value is adaptively adjusted based on the frequency range of the vibration data (the normal vibration frequency of the inspection equipment is 50-100Hz). When the vibration frequency is higher than 80Hz, When the value is 0.8, below 50Hz, Take 1.2, and the middle range is 1.0; The following section discusses the improved wavelet basis function. Explanation: Step a: Obtain the basic wavelet basis function. Combined with the conventional frequency range (50-100Hz) of the vibration data of the inspection equipment, select the db4 wavelet basis as the basic wavelet basis. This wavelet basis has good time-domain locality and frequency-domain resolution, which is suitable for the non-stationary characteristics of industrial equipment vibration signals and can initially capture the effective features in the vibration data. Step b: Collect initial vibration data (without past noise processing) transmitted by the multi-agent vibration detection agent, perform frequency domain analysis on the initial vibration data through Fast Fourier Transform (FFT), extract the actual frequency distribution of the vibration data, and determine the dominant vibration frequency range of the current inspection equipment (such as 50-60Hz, 70-80Hz, etc.). Step c: Adjust the parameters of the basic db4 wavelet basis based on the dominant vibration frequency range. Adjust the scaling factor and translation factor of the wavelet basis to make the frequency domain response of the improved wavelet basis function accurately match the dominant vibration frequency range of the current inspection equipment: if the dominant vibration frequency is higher than 80Hz, reduce the scaling factor to 0.8 to enhance the capture capability of high-frequency signals; if the dominant vibration frequency is lower than 50Hz, increase the scaling factor to 1.2 to improve the resolution accuracy of low-frequency signals; if the dominant vibration frequency is in the 50-80Hz range, keep the scaling factor at 1.0 to ensure the integrity of the intermediate frequency signal. Step d: Apply the adjusted wavelet basis function to the pre-denoising test of a small amount of vibration data, calculate the signal-to-noise ratio (SNR) of the denoised data. If the SNR does not reach the preset threshold (the preset SNR in this invention is ≥30dB), repeat steps b and c, fine-tuning the scaling factor and translation factor until the SNR requirement is met, and finally determine the final value. The specific parameters are determined to avoid the problem of poor adaptability of traditional wavelet basis functions.

[0015] Infrared thermal imaging data Perform grayscale normalization processing to map pixel values ​​to Range, eliminating grayscale differences caused by different shooting distances and ambient temperatures; for image defect data Contour extraction and feature quantization are performed, extracting feature parameters including but not limited to the area, perimeter, and gray-level mean of defects, and converting them into one-dimensional feature vectors; environmental temperature and humidity data are also analyzed. Standardize the data to eliminate the influence of dimensions.

[0016] Step S2.2: Data reliability assessment and scene feature extraction: S2.2.1: Calculate the reliability index for each type of preprocessed data. , This reflects the reliability of the data and avoids the impact of unreliable data on the fusion results. ,in, For the first Signal-to-noise ratio of the data type For the first Consistency of class data (the deviation coefficient of data collected by the same agent in three consecutive data collections). , These are the weighting coefficients. , (Based on the priority settings of the inspection data, the signal-to-noise ratio has a higher weight.) The range of values ​​is , The closer the value is to 1, the higher the data reliability. S2.2.2: Extracting dynamic features of the inspection scene Including equipment operating load (Value range 0-1, 1 represents full load), environmental disturbance intensity (Values ​​range from 0 to 1, with 1 indicating the strongest interference), calculated using auxiliary data (such as device current data and environmental noise data) collected by multiple agents, and used to dynamically adjust the fusion weights; this step is explained in detail below: Step S2.2.2.1: Auxiliary data acquisition and preprocessing Relying on the dedicated auxiliary acquisition unit in the multi-agent inspection terminal (which works in collaboration with intelligent agents such as vibration detection and infrared inspection, and is a standard extended function of the multi-agent inspection terminal), two types of core auxiliary data are collected simultaneously: one is equipment operation auxiliary data, which is collected by the current acquisition intelligent sub-unit deployed on the inspection equipment itself to obtain the real-time operating current value of the equipment. (Unit: A), the sampling frequency is consistent with the vibration data (100Hz) to ensure data timing synchronization; secondly, environmental interference auxiliary data is collected by the environmental monitoring intelligent agent to obtain the on-site environmental noise value. (Unit: dB), the sampling frequency is 1 time / minute, which matches the sampling frequency of ambient temperature and humidity data to avoid data redundancy.

[0017] Preprocess the collected auxiliary data to eliminate acquisition errors: This includes processing the equipment current data. A moving average filtering method (with a window size of 5 data points) is used to remove errors caused by instantaneous current fluctuations, resulting in a smoothed current value. ; Environmental noise data Outliers exceeding the reasonable range (30-80dB) of the industrial inspection environment were removed, and missing data were supplemented using linear interpolation to obtain effective noise values. The preprocessed data is stored synchronously with the preprocessing results of various inspection data in step S2.1 to ensure time sequence consistency.

[0018] Step S2.2.2.2: Equipment operating load Calculation: Equipment operating load This reflects the real-time operating status of the inspection equipment. The calculation is based on the ratio of the equipment's rated current to the real-time smoothed current, and is normalized according to the equipment's operating characteristics to ensure the value strictly falls within the range of 0-1. The specific calculation process is as follows: Retrieve the rated current value of the currently inspected equipment pre-stored in the data storage module. (This value is the equipment's factory parameter, which is pre-entered by staff through a remote terminal interface and corresponds one-to-one with the model of the equipment being inspected, such as the rated current of the main transformer in the substation being 1000A). The formula for calculating the real-time current load ratio is as follows: ; Normalize the load percentage to avoid exceeding the 0-1 range: when When this happens, the equipment is deemed to be shut down. ;when At that time, it is determined that the equipment is operating at full load. ;when At that time, take directly This calculation method is consistent with the scene adaptive adjustment coefficient in subsequent step S2.3.2. The calculation logic is adapted to ensure that changes in equipment load can be accurately reflected in the adjustment of fusion weights.

[0019] Step S2.2.2.3: Environmental Interference Intensity Calculation: Environmental disturbance intensity This reflects the degree of interference from the on-site environment on the collection of inspection data. It is mainly based on the calculation of environmental noise values, and a reasonable threshold is set in combination with the interference characteristics of industrial inspection scenarios to achieve normalization in the range of 0-1. The specific calculation process is as follows: Setting an environmental noise baseline threshold for industrial inspection scenarios: Based on the applicable scenarios of this invention, such as power and chemical industries, an interference baseline threshold is set. (No obvious interference) (Strong interference can cause the inspection data to be distorted.) The formula for calculating the proportion of noise interference is as follows: ; Normalize and calibrate the interference ratio: when At that time, it was determined to be without interference. ;when At that time, it was determined to be strong interference. ;when At that time, take directly The calculation results correspond to the vibration data denoising in step S2.1.1 and the data reliability assessment in step S2.2.1; the higher the interference intensity, the lower the data reliability. The lower, and thus through Reduce the fusion weight of the corresponding data.

[0020] Step S2.2.2.4: Scene Dynamic Features Integration: Scene dynamic features Equipment operating load Intensity of environmental disturbance The comprehensive representation, combined with the priority of the impact of both on the fusion of inspection data, is integrated using a weighted summation method, as shown in the following formula: The weighting coefficients are set to 0.6 and 0.4 because the impact of equipment operating load on the characteristics of inspection data (especially vibration data and infrared data) is far greater than that of environmental interference, which aligns with the data reliability index in step S2.2.1. The weight setting logic is consistent to ensure scene features It can accurately reflect the core scenario factors affecting the fusion of multi-source data. (Integrated) The value range remains 0-1, and it is directly used as the scene adaptive adjustment coefficient in subsequent step S2.3.2. The calculation enables the dynamic correlation between scene features and fusion weights.

[0021] Step S2.3: Calculation of the improved adaptive OVO attention fusion algorithm for the inspection scenario: Abandoning the shortcomings of traditional OVO attention mechanisms that only consider inter-modal interactions and do not take into account scene features and data reliability, we have improved the design of the fusion weight calculation method to achieve adaptive fusion of multi-source data: S2.3.1: Calculate the OVO attention weights for each type of data in step S2.1. , indicating the first Importance of interaction between class data and all other classes of data: ,in For the similarity calculation function, an improved cosine similarity is adopted, which is adjusted in combination with the feature differences of the inspection data to avoid the problem of poor adaptability of traditional cosine similarity to heterogeneous data. Indicates except the first A collection of all other data types besides Class Data; S2.3.2: Combining data reliability indicators Dynamic characteristics of inspection scenarios Calculate the adaptive adjustment coefficient for the scenario : When the equipment is running at full load And environmental interference is weak hour, Take the maximum value, as this corresponds to the highest data fusion weight; when environmental interference is strong... And the equipment is operating at low load. hour, Reduce, minimize the impact of unreliable data; S2.3.3: Calculate the final fusion weights This enables dynamic adaptive adjustment of weights. ; S2.3.4: Based on the final fusion weights, the preprocessed multi-source data is weighted and fused to obtain fused feature data. : ,in For the first Preprocessing results of class data, It is a one-dimensional fusion feature vector containing core information from multiple source data, used for subsequent anomaly detection; Step S2.4: Output of fused features: The fused feature data... The data is transmitted to the feature-driven multi-agent collaborative anomaly localization and feedback module, while the fusion weights are also... Data reliability indicators The data is stored in the data storage module to support subsequent anomaly tracing and weight optimization.

[0022] Step S3: The feature-driven multi-agent collaborative anomaly localization and feedback module receives the fused feature data, performs temporal feature extraction, improved dynamic threshold anomaly detection, and multi-agent collaborative anomaly localization, identifying the anomaly type, anomaly location, and anomaly confidence level; the specific implementation steps are as follows: Step S3.1: Fusion of feature reception and temporal feature extraction: Receive fused feature data output by the adaptive multi-source data fusion module for inspection scenarios Simultaneously, historical fusion feature data (fusion feature sequences within the past hour) is retrieved from the data storage module to extract the temporal variation features of the fusion features, including the feature change rate. Characteristic fluctuation variance Dynamic threshold calculation for anomaly detection: ; ,in For the fused feature data at the current moment, The fused feature data from the previous time step. Number of historical data ( (corresponding to 1 hour, 1 data point per minute). The mean of the historical fusion feature data; Step S3.2: Calculation of the improved dynamic threshold anomaly detection algorithm: Overcoming the limitations of traditional fixed thresholds in adapting to dynamic changes in equipment operating status, a dynamic threshold anomaly detection algorithm is designed based on fused temporal features and historical anomaly data to achieve accurate anomaly identification. S3.2.1: Calculate the dynamic anomaly threshold This threshold is dynamically adjusted according to the device's operating status and fusion characteristics. ,in For adaptive threshold coefficients, Calibrate based on historical anomaly data. The rate of change weighting coefficient ( (To avoid excessive threshold fluctuations caused by feature mutations) When fusing features If the condition is met, it is considered abnormal; otherwise, it is considered normal. S3.2.2: Calculate the anomaly confidence level This is used to determine the severity of the anomaly and avoid misjudgment. ,in The fusion weights output by the adaptive multi-source data fusion module for inspection scenarios. As a data reliability indicator; The range of values ​​is , If the condition is deemed serious, it requires immediate attention. At that time, it was determined to be a minor abnormality, requiring close monitoring; If the alarm is false, the anomaly is eliminated. S3.2.3: Anomaly type identification, combining the fusion weights and feature differences of multi-source data to trace the root cause of the anomaly: if the fusion weights of vibration data... And abnormal confidence level The problem is determined to be a mechanical failure of the equipment; if the fusion weight of the infrared data is... and The fault was determined to be an overheating device; if the image data fusion weights and This was determined to be a defect in the appearance of the equipment. Step S3.3: Multi-agent collaborative anomaly localization: Based on anomaly type and anomaly confidence, a multi-agent cooperative scheduling algorithm is designed to leverage the spatial distribution advantages of multiple agents to achieve accurate anomaly localization. The specific implementation steps are as follows: S3.3.1: Retrieve real-time location information of each agent Current inspection task status , Indicates free time. Indicates being busy. , Given the number of agents, calculate the distance from each agent to the suspected location of the anomaly. : ,in, For the first The real-time location of each agent. This is a suspected location of anomaly (preliminarily determined by the location collected by the agent corresponding to the fused features); S3.3.2: Calculate the scheduling priority of each agent. The higher the priority, the more likely it is to be scheduled to the suspected anomaly location for secondary detection: ,in, Weights for agent types (matched to anomaly types, such as vibration detection agents in the event of mechanical failure). Other intelligent agents ); S3.3.3: Select the two highest priority agents (to avoid detection errors by a single agent), and dispatch them to the suspected anomaly location to collect secondary data. The collected data is transmitted to the inspection scenario adaptive multi-source data fusion module for re-fusion and detection to confirm the anomaly location and type, and achieve accurate anomaly localization. Step S3.4: Remote interactive feedback and closed-loop optimization: S3.4.1: Generate abnormal feedback information, including abnormal location, abnormal type, abnormal confidence level, and secondary detection results, and transmit it to the remote terminal interactive interface through the remote interactive communication module for staff to view. At the same time, generate remote instructions (such as key inspection instructions and fault handling instructions). S3.4.2: The anomaly detection results (anomaly type, anomaly confidence level) Feedback is sent to the inspection scenario adaptive multi-source data fusion module to adjust the scenario adaptive adjustment coefficient of the inspection scenario adaptive multi-source data fusion module. If it is a serious abnormality Data source weight corresponding to the anomaly type Increase by 20%, that is Multiply by 1.2 to increase the fusion priority of this type of data; if it is a false alarm. The corresponding data source Multiply by 0.8 to reduce its weight and avoid misjudgment again; S3.4.3: Based on the anomaly location results, the remote interactive communication module sends an inspection strategy adjustment instruction to the multi-agent inspection terminal, such as adjusting the inspection frequency of the abnormal area (from once every hour to once every 10 minutes). Idle agents go to the abnormal area for continuous monitoring, realizing a remote interactive closed loop of "data collection-fusion-detection-feedback-strategy adjustment".

[0023] Step S4: The feature-driven multi-agent collaborative anomaly localization and feedback module feeds back the anomaly detection results to the inspection scenario adaptive multi-source data fusion module, adjusts the fusion weight of the inspection scenario adaptive multi-source data fusion module, and generates anomaly feedback information and remote inspection instructions, which are transmitted to the remote terminal interaction interface and multi-agent inspection terminal through the remote interactive communication module. Step S5: The multi-agent inspection terminal receives remote instructions, adjusts the inspection strategy, performs secondary data collection, and repeats steps S2-S4 to form a closed loop; the data storage module stores all inspection data, fusion results, abnormal information, and instruction logs. Step S6: Staff members view relevant information through the remote terminal interactive interface, issue fault handling instructions, and complete intelligent inspection data analysis and anomaly handling.

[0024] Example 2: A smart inspection data analysis system based on remote interaction and multi-agent technology is used to implement a smart inspection data analysis method based on remote interaction and multi-agent technology. The system includes: Multi-agent inspection terminal: Composed of multiple intelligent inspection devices of different types (such as vibration detection agent, infrared inspection agent, and image acquisition agent), it is used to collect on-site inspection data, receive remote interaction commands, execute inspection tasks, and transmit the collected data to the inspection scene adaptive multi-source data fusion module through the remote interaction communication module; the following is a detailed explanation of this module: 1.1: Composition and Layout of Multi-Agent Inspection Terminal: The multi-agent inspection terminal adopts a distributed collaborative layout, adapting to the large-scale, multi-equipment inspection needs of industrial scenarios such as power and chemical industries. The number and deployment location of each type of intelligent agent can be flexibly configured according to the scale of the inspection area and the characteristics of equipment distribution (e.g., deploying 3 vibration detection intelligent agents, 2 infrared inspection intelligent agents, 2 image acquisition intelligent agents, and 1 environmental monitoring intelligent agent in a power substation scenario). All intelligent agents have built-in positioning modules (such as GPS / BeiDou positioning), data acquisition modules, Mesh communication modules, and command execution modules, which can realize integrated operation of autonomous inspection, data acquisition, command reception, and data transmission. Moreover, the intelligent agents can realize real-time interaction of location information and task status through the collaborative transmission channel of the remote interactive communication module, providing a foundation for subsequent multi-agent collaborative anomaly location (step S3.3).

[0025] 1.2: Specific functions and workflows of various types of intelligent agents: Each intelligent agent collects corresponding types of inspection data according to its own functional positioning. The collected parameters are precisely matched with the requirements of subsequent data preprocessing (step S2.1) and data reliability assessment (step S2.2.1), as follows: 1.2.1: Vibration Detection Intelligent Agent: The core is used to collect equipment vibration data (corresponding to the adaptive multi-source data fusion module in the inspection scenario). This system is designed to meet the mechanical fault detection needs of industrial equipment (such as transformers, pumps, and motors). Each vibration detection agent is equipped with a high-precision vibration sensor and deployed at key parts of the inspected equipment (such as bearings and machine bases). The acquisition frequency is set to 100Hz (consistent with the frequency requirement for vibration data preprocessing in the adaptive multi-source data fusion module for the inspection scenario). The acquired data includes core parameters such as vibration acceleration and vibration velocity. During the acquisition process, the data is first initially denoised (removing high-frequency interference noise) through a built-in filtering module, then temporarily stored in a local cache, and simultaneously transmitted in real time to a remote interactive communication module via a Mesh communication module. Finally, it is sent to the adaptive multi-source data fusion module for further denoising and preprocessing. In addition, the agent can receive secondary acquisition commands (such as key detection commands during anomaly location) from a multi-agent collaborative anomaly localization and feedback module driven by fusion features, adjusting the acquisition frequency (up to 200Hz) to accurately acquire vibration data in abnormal areas and support the tracing of the root cause of anomalies.

[0026] 1.2.2: Infrared Inspection Intelligent Agent: The core is used to collect infrared thermal imaging data (corresponding to the adaptive multi-source data fusion module in the inspection scenario). This system is designed to meet the needs of equipment overheating fault detection. Equipped with an infrared thermal imaging camera with a resolution of 640×480, it captures images of the surface temperature distribution of equipment at a rate of 1 time per minute, matching the ambient temperature and humidity data acquisition frequency to avoid data redundancy. During acquisition, it automatically corrects for the influence of ambient temperature on the infrared image, converts the image data into grayscale images for temporary storage, and then transmits it to the backend via a remote interactive communication module for grayscale normalization processing by the adaptive multi-source data fusion module for inspection scenarios. Simultaneously, it can receive key area inspection commands from a remote terminal interface, focusing on suspected abnormal areas and improving the accuracy and efficiency of infrared image acquisition.

[0027] 1.2.3: Image Acquisition Intelligent Agent: The core is used to acquire defect data of equipment appearance images (corresponding to the adaptive multi-source data fusion module in the inspection scenario). This system is designed to detect defects such as damage, rust, and foreign matter attachment on equipment. It features a high-definition camera (1920×1080 resolution) and employs multi-angle acquisition modes (front, side, top) to ensure no blind spots. The acquisition frequency is adjusted according to the equipment type: critical equipment acquires data every 30 minutes, while ordinary equipment acquires data every hour. After acquisition, the images undergo preliminary preprocessing (size normalization, noise reduction) to extract preliminary features of the image contours. These features are then transmitted to the inspection scenario adaptive multi-source data fusion module for further feature quantization and processing. When the feature-driven multi-agent collaborative anomaly localization and feedback module determines that an appearance defect exists, the agent can receive a secondary acquisition command, travel to the anomaly location, and acquire high-definition detailed images to assist staff in confirming the degree of defect.

[0028] 1.2.4: Environmental Monitoring Intelligent Agent: The core component is used to collect environmental temperature and humidity data (corresponding to the adaptive multi-source data fusion module in the inspection scenario). The system collects environmental interference auxiliary data (environmental noise data from step S2.2.2) to provide environmental parameter support for data preprocessing and scene feature extraction. It is equipped with temperature and humidity sensors and a noise sensor. The temperature and humidity data is collected once per minute, and the noise data is collected at the same frequency. The collected data includes ambient temperature (°C), relative humidity (%RH), and ambient noise (dB). After collection, the data undergoes preliminary verification, and outliers exceeding reasonable ranges (such as temperature and humidity exceeding the operating environment range of industrial equipment) are removed. The data is then transmitted to the backend for standardization processing (temperature and humidity data) and scene feature extraction (noise data) by the adaptive multi-source data fusion module for the inspection scenario. Furthermore, the intelligent agent can monitor the intensity of environmental interference in real time. When the noise intensity is too high (≥80dB), it automatically sends an interference warning to the backend, reminding the adaptive multi-source data fusion module to reduce the reliability weight of the corresponding data source.

[0029] 1.3: Multi-agent collaborative working mechanism: The multi-agent inspection terminal adopts a "autonomous inspection + collaborative scheduling" working mode to ensure that the inspection coverage is without blind spots and that the data collection time sequence is synchronized. Autonomous Inspection: Each agent autonomously executes inspection tasks according to the inspection route pre-entered by the data storage module (set by staff through a remote terminal interface), records its own location information and task status (idle / busy) in real time, and feeds it back to the background through the remote interactive communication module for coordinated scheduling by the feature-driven multi-agent collaborative anomaly localization and feedback module (step S3.3); if an obstacle is encountered during the inspection, the agent autonomously adjusts the route through the built-in obstacle avoidance module to ensure the smooth progress of the inspection task.

[0030] Collaborative Scheduling: When the feature-driven multi-agent collaborative anomaly localization and feedback module detects an anomaly and requires secondary data collection, it sends scheduling instructions to each agent through the remote interactive communication module. Based on the agent's location, task status, and type weight (e.g., prioritizing vibration detection agents in case of mechanical failure), the optimal agent is selected to perform the secondary data collection task. After receiving the instructions, each agent collaboratively heads to the suspected anomaly location and simultaneously collects multiple types of data to ensure the accuracy of anomaly localization. At the same time, idle agents can assist busy agents in completing key area inspections, optimizing inspection resource allocation, and avoiding duplicate inspections and blind spots.

[0031] 1.4: Command Reception and Data Transmission Integration: The multi-agent inspection terminal establishes a stable connection with the remote interactive communication module through its built-in Mesh communication module, enabling bidirectional interaction of command reception and data transmission, and connecting the preceding and following modules. Command reception: It receives two types of core commands: one is manual commands issued by the remote terminal interactive interface (such as patrol route adjustment, data collection parameter modification, and fault handling assistance commands); the other is automatic commands issued by the feature-driven multi-agent collaborative anomaly localization and feedback module (such as secondary data collection commands, patrol frequency adjustment commands, and anomaly area key monitoring commands). The command reception delay is controlled within 50ms (consistent with the first-level priority transmission requirements of the remote interactive communication module) to ensure that commands are executed in real time.

[0032] Data transmission: The collected raw inspection data (vibration, infrared, image, temperature and humidity) and auxiliary data (current, noise) are transmitted in real time to the inspection scenario adaptive multi-source data fusion module (for data fusion) and the data storage module (for data backup and traceability) through the uplink transmission channel of the remote interactive communication module. During the transmission, a data priority scheduling strategy is followed (the instructions of the multi-agent collaborative anomaly localization and feedback module driven by fusion features take precedence over the raw data) to ensure the real-time performance and stability of core data transmission. At the same time, all transmitted data is marked with an identity tag to facilitate data source matching during subsequent data traceability and anomaly localization.

[0033] Remote interactive communication module: Employing Mesh adaptive networking technology, this module enables data transmission and command interaction between multiple agents and the backend, as well as among multiple agents themselves, reducing transmission latency and packet loss rate, and ensuring the real-time performance and stability of data transmission. The following is an explanation of this module: Step 2.1: Mesh networking initialization configuration: After the remote interactive communication module is started, it first completes the initial deployment of the Mesh network to adapt to the distributed layout of multi-agent inspection terminals (such as large-scale inspection scenarios such as power substations and chemical plant areas). Determine the network nodes: Each multi-agent inspection terminal (vibration detection agent, infrared inspection agent, etc.) is designated as the terminal node of the Mesh network, and the backend server is designated as the core control node. Each terminal node has a built-in Mesh communication module that supports self-organizing network functionality. At the same time, 3-5 relay nodes are evenly deployed in the inspection area to expand the network coverage and avoid communication interruptions caused by blind spots in the inspection. The relay nodes establish initial communication links with the core control node and the terminal nodes.

[0034] Node authentication and parameter configuration: The core control node authenticates all terminal nodes and relay nodes using a preset encryption key (using the AES-128 encryption algorithm, the key is pre-entered by staff through a remote terminal interface and stored in the data storage module) to verify the legitimacy of the nodes and prevent unauthorized nodes from accessing the network and causing data leakage or interference. After successful authentication, the communication parameters of each node are configured, including communication frequency (set to the 2.4GHz industrial band to avoid interference with other equipment signals on site), transmission bandwidth (dynamically allocated according to the amount of inspection data, with higher bandwidth allocated to vibration data and image data), and heartbeat detection cycle (set to 100ms for real-time monitoring of link status).

[0035] Step 2.2: Adaptive Channel Selection and Link Establishment: To avoid transmission instability caused by environmental interference (such as electromagnetic interference from industrial equipment and environmental noise), an improved channel quality assessment algorithm is adopted to achieve adaptive channel selection and link optimization, ensuring the stability of data transmission. Real-time channel quality monitoring: Each terminal node and relay node monitors the signal strength of the currently available communication channels in real time. ), bit error rate ( The system collects detection data every 50ms for three core indicators: signal strength, interference strength, and signal intensity. The data is then transmitted to the core control node for aggregation and analysis. Bit error rate When the interference intensity is ≤0.3 (normalized value), the channel is considered available.

[0036] Adaptive channel selection: The core control node calculates the channel quality score using a weighted summation method based on the detection indicators of each channel. The formula is as follows: ,in This is the normalized value of the signal strength (mapped to the 0-1 interval). The normalized value for interference intensity; select the score. The highest-rated channel is used as the current communication channel. If the current channel quality score is below 0.6, channel switching is immediately triggered to reselect the optimal channel and avoid transmission interruption.

[0037] Dynamic link establishment: Each terminal node automatically selects the optimal communication link based on its own location and signal strength (prioritizing direct communication with the core control node; if the signal is weak, it forwards through a relay node). The core control node records the link connection status of each node in real time, establishes a link topology map, and stores it in the data storage module to support subsequent link maintenance. At the same time, multiple agents establish collaborative communication channels through direct links between terminal nodes for real-time data interaction during multi-agent collaborative inspections (such as location information during anomaly location and coordination of data collection tasks).

[0038] Step 2.3: Data Transmission Optimization and Priority Scheduling: To address the varying transmission requirements of multi-source inspection data (vibration data, infrared data, image data, etc.), and considering the data interaction requirements of the creative module, an improved data packet fragmentation and priority scheduling strategy is adopted to reduce transmission latency and packet loss rate, ensuring real-time data transmission. Data packet fragmentation and encoding: Different types of inspection data are processed differently. Large-capacity data such as image data and infrared thermal imaging data are transmitted in fragments, dividing the data packets into fixed-size segments (1024 bytes / segment). Each segment is assigned a checksum (using the CRC-32 checksum algorithm) to prevent fragment loss or errors. Small-capacity real-time data such as vibration data and environmental temperature and humidity data are transmitted as a whole to reduce fragmentation overhead. All data packets are encoded using LDPC (Low-Density Parity Check) for error correction, reducing the bit error rate during transmission and further improving data transmission reliability.

[0039] Data Priority Scheduling: Based on the importance and real-time requirements of the data, a three-level priority system is implemented to ensure priority transmission of core data: Level 1 priority is given to remote inspection commands issued by the feature-driven multi-agent collaborative anomaly localization and feedback module (such as secondary acquisition commands for abnormal areas and inspection strategy adjustment commands), which prioritizes communication bandwidth and keeps transmission latency within 50ms; Level 2 priority is given to fused feature data and anomaly detection-related data from the adaptive multi-source data fusion module for inspection scenarios, with transmission latency kept within 100ms; Level 3 priority is given to raw inspection data (such as unprocessed vibration data and image data), which is transmitted when bandwidth is sufficient to ensure that the transmission efficiency of core data is not affected; the core control node sorts data packets of different priorities using a queue scheduling algorithm, prioritizing the transmission of high-priority data to avoid data congestion.

[0040] Step 2.4: Link Maintenance and Fault Self-Healing To address link interruptions caused by multi-agent movement and increased environmental interference during inspections, a dynamic link maintenance and fault self-healing mechanism is established to ensure communication continuity and support the closed-loop operation of multi-agent collaborative inspections and creative modules. Real-time link status monitoring: Each node sends a heartbeat packet to the core control node every 100ms through a heartbeat detection mechanism. The core control node determines whether the link is normal based on the reception of the heartbeat packet. If a node fails to receive a heartbeat packet for 3 consecutive times, it is determined that the link is interrupted and the fault investigation process is immediately initiated.

[0041] Fault self-healing: For nodes with broken links, the core control node automatically reassigns the optimal communication link based on the link topology (such as switching to other relay nodes or adjusting the communication channel), and sends a link reconnection command to the node. After receiving the command, the node quickly completes the link reconnection, with the reconnection time controlled within 500ms. If the node cannot reconnect (e.g., due to agent failure), the core control node immediately transmits the fault information to the remote terminal interface to remind the staff to handle it. At the same time, it adjusts the inspection tasks of other adjacent agents to make up for the inspection blind spots of the faulty node and ensure that the inspection work is not interrupted.

[0042] Step 2.5: Implementation of closed-loop data interaction: Combining the closed-loop requirements of "data acquisition-fusion-detection-feedback-strategy adjustment" in this invention, the remote interactive communication module realizes closed-loop transmission of multi-directional data interaction: Uplink transmission: The raw inspection data and auxiliary data collected by the multi-agent inspection terminal are transmitted to the backend through the Mesh network and sent to the inspection scenario adaptive multi-source data fusion module (multi-source data fusion) and the data storage module (data backup), respectively. The fused feature data output by the inspection scenario adaptive multi-source data fusion module is transmitted to the fused feature-driven multi-agent collaborative anomaly localization and feedback module through this module to support anomaly detection and localization.

[0043] Downlink transmission: Remote inspection commands and anomaly feedback information generated by the feature-driven multi-agent collaborative anomaly localization and feedback module are transmitted to the multi-agent inspection terminal via the Mesh network to guide the agents to adjust their inspection strategies and perform secondary data collection; Fault handling commands and inspection parameter adjustment commands issued by staff through the remote terminal interactive interface are forwarded to the corresponding agents through the core control node to achieve remote interactive control.

[0044] Collaborative transmission: Collaborative data between multiple agents (such as location information and task allocation information during anomaly localization) is transmitted directly through terminal nodes to achieve collaborative inspection among multiple agents. This supports the implementation of multi-agent collaborative scheduling algorithms for feature-driven multi-agent collaborative anomaly localization and feedback modules. At the same time, logs of all transmitted data (transmission time, data type, and transmission status) are stored in real time in the data storage module to support subsequent data traceability and communication performance optimization.

[0045] The inspection scenario adaptive multi-source data fusion module receives multi-source heterogeneous inspection data transmitted by multi-agent inspection terminals, and adopts an improved inspection scenario adaptive OVO attention fusion algorithm to solve the problems of distortion in multi-source heterogeneous data fusion and difficulty in dynamic weight adjustment, and outputs high-precision fusion feature data. The feature-driven multi-agent collaborative anomaly localization and feedback module receives fused feature data output from the inspection scenario adaptive multi-source data fusion module. It adopts an improved dynamic threshold anomaly detection and multi-agent collaborative scheduling algorithm to achieve accurate anomaly identification, localization and root cause tracing. At the same time, it generates remote interaction commands and feeds them back to the multi-agent inspection terminal through the remote interaction communication module to realize real-time adjustment of the inspection strategy and form a closed loop. Data storage module: Used to store raw inspection data, fused feature data, anomaly detection results, remote interaction commands and inspection logs collected by multiple agents, providing data support for subsequent data analysis and traceability; Remote terminal interaction interface: This allows staff to view inspection data, fusion results, and anomaly information, issue remote inspection commands, and achieve remote interaction with multiple intelligent agents. The following is an explanation of this module: A: Interface architecture design: The remote terminal interface adopts a dual architecture of "PC + mobile terminal" to adapt to the mobile operation and fixed control needs of personnel in industrial inspection scenarios. Data is synchronized in real time between the two, and both support hierarchical permission management (three levels of permissions: administrator, maintenance personnel, and inspection personnel) to ensure operational security and clear responsibilities. The core of the interface is divided into five functional areas: data monitoring area, anomaly handling area, instruction issuance area, historical traceability area, and system settings area. The functions of each area are interconnected to form a complete human-computer interaction closed loop, supporting the entire inspection process.

[0046] B: Detailed explanation of core functions: The core functions of the remote terminal interface revolve around "data viewing - anomaly handling - command issuance - traceability and review," and each function is deeply integrated with other modules in the invention: Multi-dimensional data viewing function: Through this function, staff can view the entire inspection process data in real time. The data sources are the data storage module, the inspection scenario adaptive multi-source data fusion module, and the feature-driven multi-agent collaborative anomaly localization and feedback module, ensuring the authenticity and integrity of the data. Specific content to view includes: Raw inspection data: Real-time viewing of various raw data collected by the multi-agent inspection terminal, including vibration acceleration / velocity data of vibration detection agent, thermal imaging images of infrared inspection agent, appearance defect images of image acquisition agent, and temperature, humidity and noise data of environmental monitoring agent. It supports filtering and viewing by agent type, inspection time, and device number. At the same time, it can view the data collection location (synchronously displayed through GPS / BeiDou coordinates fed back by the agent positioning module), and is precisely connected with the data collection function of the multi-agent inspection terminal and the storage function of the data storage module. Data fusion results: View the fused feature data output by the adaptive multi-source data fusion module for inspection scenarios. The fusion weight of various types of data Data reliability indicators and scene dynamic features It supports viewing key parameters of the fusion process (such as OVO attention weights and scene adaptive adjustment coefficients). This facilitates staff in analyzing the rationality of the fusion results and allows for comparison of fusion weight changes at different times, providing a reference for optimizing the fusion algorithm. Anomaly detection related information: View the anomaly detection results output by the feature-driven multi-agent collaborative anomaly localization and feedback module, including anomaly location, anomaly type (mechanical fault, overheating fault, appearance defect), and anomaly confidence level. The system displays the secondary detection results and simultaneously shows the scheduling records of multiple agents during the anomaly localization process (scheduling priority, secondary data collection). It is closely linked with the anomaly localization and collaborative scheduling functions of the feature-driven multi-agent collaborative anomaly localization and feedback module, providing a basis for staff to judge the severity of the anomaly and formulate a handling plan.

[0047] Error handling and feedback functions: This function is the core of connecting staff with the anomaly handling process, enabling real-time reception of anomaly information and rapid issuance of handling instructions. Specifically, it includes: Anomaly warning alert: When the feature-driven multi-agent collaborative anomaly localization and feedback module detects anomalies (especially severe anomalies)... When an anomaly occurs, the remote terminal interface will immediately issue an audible and visual warning, and simultaneously display the core information of the anomaly (anomaly location, type, and confidence level) in a pop-up window, and mark the intelligent agent's collected data and fusion features corresponding to the anomaly, so as to avoid staff missing key anomaly information; at the same time, it supports anomaly classification reminders (red warning for serious anomalies, yellow warning for minor anomalies, and blue warning for false alarms), which is consistent with the anomaly confidence level classification logic of the multi-agent collaborative anomaly localization and feedback module driven by fusion features; Anomaly Handling Procedures: Staff can view anomaly details (including historical data comparison and anomaly root cause analysis) through the interface, fill in handling opinions, upload handling records (such as fault repair reports and on-site photos), and synchronously store the handling results to the data storage module to achieve full-process traceability of anomaly handling; for minor anomalies, key monitoring instructions can be set through the interface to adjust the inspection frequency of the abnormal area, and continuous monitoring can be performed by multiple agents; for serious anomalies, fault handling instructions can be issued immediately to notify maintenance personnel to go to the site for handling.

[0048] Remote command issuance function: This function, through a remote interactive communication module, enables staff to remotely control the multi-agent inspection terminal and two innovative modules. The command types are precisely matched with the functional requirements of each module in the invention, specifically including: Multi-agent inspection commands: These commands include: issuing commands to adjust inspection strategies (such as modifying inspection routes, adjusting collection frequencies, and issuing commands for key areas); secondary collection commands (issuing precise collection commands for suspected anomaly locations located by the multi-agent collaborative anomaly localization and feedback module driven by fusion features); and inspection task allocation commands (adjusting the inspection range of each agent to compensate for blind spots of faulty agents). Commands are transmitted to the multi-agent inspection terminal via the downlink transmission channel of the remote interactive communication module, with a reception delay controlled within 50ms, consistent with the first-priority transmission requirements of the remote interactive communication module. Simultaneously, the execution status of commands (received, executing, completed) can be viewed in real time to ensure command implementation. Data analysis parameter adjustment instructions: For the adaptive multi-source data fusion module and the fusion feature-driven multi-agent collaborative anomaly localization and feedback module in the inspection scenario, algorithm parameter adjustment instructions are issued, such as adjusting the adaptive wavelet denoising threshold (SNR threshold) of module 1 and the weight coefficients of the fusion weight calculation. ), adjust the dynamic threshold coefficient of the feature-driven multi-agent collaborative anomaly localization and feedback module ( Anomaly confidence level grading standards are implemented, and the corresponding modules are updated synchronously after the instructions are issued to achieve dynamic optimization of data analysis algorithms. System control commands: Issue data storage commands (such as data backup and data cleanup) and network parameter adjustment commands for remote interactive communication modules (such as communication frequency and relay node configuration) to ensure the stable operation of the entire system.

[0049] Historical data tracing and statistical analysis functions: This function relies on the data storage module to achieve traceability and statistics of data throughout the entire inspection process, providing data support for optimizing inspection work: Historical data traceability: Users can query historical inspection data, fusion results, anomaly detection records, instruction issuance records, and anomaly handling records by time range, equipment number, and anomaly type. Data export is supported (e.g., Excel, PDF format), which facilitates staff to review the inspection process and analyze equipment operation patterns. Statistical analysis: Automatically calculates indicators such as the number of inspected devices, the number of anomalies, the distribution of anomaly types, the inspection efficiency of multi-agent systems, and the data transmission success rate within a specified time period, and generates visual statistical charts (such as bar charts and line charts) to intuitively present the effectiveness of inspection work, providing a basis for staff to optimize inspection strategies and rationally allocate inspection resources.

[0050] C. Collaboration logic with other modules: The remote terminal interface, as the core of human-computer interaction, forms a close collaborative relationship with other modules of this invention to ensure a closed-loop inspection process. Data storage module: All data viewing and historical tracing functions on the interface are implemented by calling the stored data in the data storage module; the abnormal handling records and instruction issuance logs filled in by staff are synchronously stored in the data storage module to achieve two-way data interaction; Remote interactive communication module: Various remote commands issued by staff are transmitted to the corresponding nodes through the downlink transmission channel of the remote interactive communication module; the command execution status of the multi-agent and the running status of the two creative modules are fed back to the interface through the uplink transmission channel of the remote interactive communication module to ensure command closed loop; The interface receives anomaly detection results from the feature-driven multi-agent collaborative anomaly localization and feedback module, providing staff with a basis for handling anomalies; parameter adjustment instructions issued by staff are fed back to the inspection scenario adaptive multi-source data fusion module and the feature-driven multi-agent collaborative anomaly localization and feedback module to optimize data analysis algorithms; at the same time, the interface can view the operating parameters of the two modules, making it convenient for staff to monitor the module's operating status. Multi-agent inspection terminal: Through the interface, the location, task status, and data collection of each agent can be monitored in real time, and inspection commands can be issued to realize remote control of multiple agents and closely connect with the collaborative working mechanism of multiple agents.

[0051] D: Operating Procedure Instructions The operation process of staff through the remote terminal interactive interface conforms to the closed-loop logic of "data acquisition-fusion-detection-feedback-strategy adjustment" of this invention. Specifically, it is as follows: Log in to the interface → View real-time inspection data and fusion results → Receive anomaly warnings and view anomaly details → Issue anomaly handling instructions or secondary acquisition instructions → View instruction execution status and secondary acquisition data → Fill in anomaly handling records → Trace historical data and statistically analyze inspection effectiveness → Optimize inspection strategies and algorithm parameters, forming a complete human-computer interaction operation closed loop to ensure efficient and accurate progress of inspection work.

[0052] This invention relates to the field of industrial intelligent inspection and multi-agent collaboration technology. It addresses the problems of distortion in the fusion of multi-source heterogeneous inspection data, high false detection and false negative rates of static threshold for anomaly detection, insufficient multi-agent collaboration, and lack of closed-loop inspection mechanism by providing a remote interactive-driven closed-loop intelligent inspection data analysis method and system.

[0053] This method collects multi-source data on equipment vibration, infrared thermal imaging, image defects, and environmental temperature and humidity through a multi-agent inspection terminal, and transmits the data remotely via a Mesh adaptive network. After data preprocessing, reliability assessment, and scene feature extraction, an improved OVO attention algorithm is used to achieve scene adaptive weighted fusion. Based on the fusion features, temporal changes are extracted, and anomaly detection is completed with a dynamic threshold. Combined with multi-agent collaborative scheduling, anomaly precise location is achieved. The detection results are fed back to optimize the fusion weights and generate remote inspection instructions to adjust the acquisition strategy, forming a closed loop of acquisition-fusion-detection-feedback-reacquisition.

[0054] The system consists of a multi-agent inspection terminal, a remote interactive communication module, an inspection scenario adaptive multi-source data fusion module, a multi-agent collaborative anomaly localization module, a data storage module, and a remote terminal interactive interface. This invention achieves adaptive and accurate fusion of multi-source data and intelligent anomaly identification and localization, significantly reducing false positives and false negatives, improving inspection resource utilization and operational efficiency, and is suitable for intelligent inspection in complex industrial scenarios such as power and chemical industries.

[0055] Although embodiments of the invention have been shown and described, it will be understood by those skilled in the art that various changes, modifications, substitutions and alterations can be made to these embodiments without departing from the principles and spirit of the invention, the scope of which is defined by the appended claims and their equivalents.

Claims

1. A method for intelligent inspection data analysis based on remote interaction and multi-agent systems, characterized in that: Includes the following steps: Step S1: The multi-agent inspection terminal collects multi-source heterogeneous inspection data on site. The multi-source heterogeneous inspection data includes equipment vibration data, infrared thermal imaging data, image defect data, and environmental temperature and humidity data. The multi-source heterogeneous inspection data is transmitted to the inspection scene adaptive multi-source data fusion module through the remote interactive communication module. Step S2: The inspection scenario adaptive multi-source data fusion module receives multi-source heterogeneous inspection data, performs data preprocessing, data reliability assessment, scene feature extraction, and improved OVO attention fusion algorithm calculation, and outputs fused feature data; The specific implementation logic of step S2 is as follows: Step S2.1: Data Reception and Preprocessing: Receive multi-source heterogeneous inspection data transmitted by the multi-agent inspection terminal, including equipment vibration data. Infrared thermal imaging data Image defect data Ambient temperature and humidity data Preprocess various types of data to eliminate noise and outliers; Step S2.2: Data reliability assessment and scene feature extraction: S2.2.1: Calculate the reliability index for each type of preprocessed data. , This reflects the reliability of the data and avoids the impact of unreliable data on the fusion results. ,in, For the first Signal-to-noise ratio of the data type For the first Consistency of class data , These are the weighting coefficients. , ; The range of values ​​is , The closer the value is to 1, the higher the data reliability. S2.2.2: Extracting dynamic features of the inspection scene Including equipment operating load Environmental interference intensity It is calculated using auxiliary data collected by multiple agents and is used to dynamically adjust the fusion weights; Step S2.3: Calculation of the improved OVO attention fusion algorithm for the inspection scenario: S2.3.1: Calculate the OVO attention weights for each type of data in step S2.

1. , indicating the first Importance of interaction between class data and all other classes of data: ,in The similarity calculation function uses an improved cosine similarity, which is adjusted based on the feature differences in the inspection data. Indicates except the first A collection of all other data types besides Class Data; S2.3.2: Combining data reliability indicators Dynamic characteristics of inspection scenarios Calculate the adaptive adjustment coefficient for the scenario : When the equipment is running at full load And environmental interference is weak hour, Take the maximum value, as this corresponds to the highest data fusion weight; when environmental interference is strong... And the equipment is operating at low load. hour, Reduce, minimize the impact of unreliable data; S2.3.3: Calculate the final fusion weights This enables dynamic adaptive adjustment of weights. ; S2.3.4: Based on the final fusion weights, the preprocessed multi-source data is weighted and fused to obtain fused feature data. : ,in For the first Preprocessing results of class data, This is a one-dimensional fused feature vector used for subsequent anomaly detection; Step S2.4: Output of fused features: The fused feature data... The data is transmitted to the feature-driven multi-agent collaborative anomaly localization and feedback module, while the fusion weights are also processed. Data reliability indicators Stored to the data storage module; Step S3: The feature-driven multi-agent collaborative anomaly localization and feedback module receives the fused feature data, performs temporal feature extraction, improved dynamic threshold anomaly detection, multi-agent collaborative anomaly localization, and identifies the anomaly type, anomaly location, and anomaly confidence. Step S4: The feature-driven multi-agent collaborative anomaly localization and feedback module feeds back the anomaly detection results to the inspection scenario adaptive multi-source data fusion module, adjusts the fusion weight of the inspection scenario adaptive multi-source data fusion module, and generates anomaly feedback information and remote inspection instructions, which are transmitted to the remote terminal interaction interface and multi-agent inspection terminal through the remote interactive communication module. Step S5: The multi-agent inspection terminal receives remote instructions, adjusts the inspection strategy, performs secondary data collection, and repeats steps S2-S4 to form a closed loop; the data storage module stores all inspection data, fusion results, abnormal information, and instruction logs. Step S6: Staff members view relevant information through the remote terminal interactive interface, issue fault handling instructions, and complete intelligent inspection data analysis and anomaly handling.

2. The intelligent inspection data analysis method based on remote interaction and multi-agent as described in claim 1, characterized in that: The specific details of preprocessing various types of data and eliminating noise and outliers in step S2.1 are as follows: An improved adaptive wavelet denoising algorithm is used to denoise the vibration data. Noise reduction processing is performed to remove environmental interference noise during equipment operation. The formula is as follows: ,in The vibration data after noise reduction. The number of wavelet basis functions. These are wavelet coefficients. For improved wavelet basis functions, It is a time series. The translation parameters of the wavelet basis functions are: For scale parameters, The value is adaptively adjusted based on the frequency range of the vibration data. When the vibration frequency is higher than 80Hz... When the frequency is 0.8, below 50Hz, Take 1.2, and the middle range is 1.0; Infrared thermal imaging data Perform grayscale normalization processing to map pixel values ​​to Range, eliminating grayscale differences caused by different shooting distances and ambient temperatures; for image defect data Contour extraction and feature quantization are performed, extracting feature parameters including the area, perimeter, and gray-scale mean of the defect, and converting them into a one-dimensional feature vector; Environmental temperature and humidity data Standardize the data to eliminate the influence of dimensions.

3. The intelligent inspection data analysis method based on remote interaction and multi-agent as described in claim 1, characterized in that: The specific implementation steps of step S3 are as follows: Step S3.1: Fusion of feature reception and temporal feature extraction: Receive fused feature data output from the adaptive multi-source data fusion module for inspection scenarios Simultaneously, historical fusion feature data is retrieved from the data storage module to extract the temporal variation features of the fusion features. Including characteristic rate of change Characteristic fluctuation variance Dynamic threshold calculation for anomaly detection: ; ,in For the fused feature data at the current moment, The fused feature data from the previous time step. The number of historical data points. The mean of the historical fusion feature data; Step S3.2: Calculation of the improved dynamic threshold anomaly detection algorithm: S3.2.1: Calculate the dynamic anomaly threshold This threshold is dynamically adjusted according to the device's operating status and fusion characteristics. ,in For adaptive threshold coefficients, Calibrate based on historical anomaly data. The change rate weighting coefficient; when the fusion features If the condition is met, it is considered abnormal; otherwise, it is considered normal. S3.2.2: Calculate the anomaly confidence level This is used to determine the severity of the anomaly and avoid misjudgment. ,in The fusion weights output by the adaptive multi-source data fusion module for inspection scenarios. As a data reliability indicator; The range of values ​​is , If the condition is deemed serious, it requires immediate attention. At that time, it was determined to be a minor abnormality, requiring close monitoring; If the alarm is false, the anomaly is resolved. S3.2.3: Anomaly type identification, combining the fusion weights and feature differences of multi-source data to trace the root cause of the anomaly: if the fusion weights of vibration data And abnormal confidence level The problem is determined to be a mechanical failure of the equipment; if the fusion weight of the infrared data is... and The fault was determined to be an overheating device; if the image data fusion weights and This was determined to be a defect in the appearance of the equipment. Step S3.3: Multi-agent collaborative anomaly localization: Based on anomaly type and anomaly confidence, a multi-agent cooperative scheduling algorithm is designed to leverage the spatial distribution advantages of multiple agents to achieve accurate anomaly localization. Step S3.4: Remote interactive feedback and closed-loop optimization: S3.4.1: Generate anomaly feedback information, including anomaly location, anomaly type, anomaly confidence level, and secondary detection results, and transmit it to the remote terminal interactive interface through the remote interactive communication module for staff to view, while also generating remote commands; S3.4.2: Feedback the anomaly detection results to the inspection scenario adaptive multi-source data fusion module, and adjust the scenario adaptive adjustment coefficient of the inspection scenario adaptive multi-source data fusion module. If it is a serious abnormality Data source weight corresponding to the anomaly type Increase by 20%, that is Multiply by 1.2 to increase the fusion priority of this type of data; if it is a false alarm. The corresponding data source Multiply by 0.8 to reduce its weight and avoid misjudgment again; S3.4.3: Based on the anomaly location results, the remote interactive communication module sends an inspection strategy adjustment command to the multi-agent inspection terminal to adjust the inspection frequency of the abnormal area, and the idle agent goes to the abnormal area for continuous monitoring.

4. The intelligent inspection data analysis method based on remote interaction and multi-agent as described in claim 3, characterized in that: The specific implementation steps of step S3.3, multi-agent cooperative anomaly localization, are as follows: S3.3.1: Retrieve real-time location information of each agent Current inspection task status , Indicates free time. Indicates being busy. , Given the number of agents, calculate the distance from each agent to the suspected location of the anomaly. : ,in, For the first The real-time location of each agent. This is a suspected abnormal location; S3.3.2: Calculate the scheduling priority of each agent. The higher the priority, the more likely it is to be scheduled to the suspected anomaly location for secondary detection: ,in, Weights for agent type; S3.3.3: Select the two highest priority agents and dispatch them to the suspected anomaly location to collect secondary data. The collected data is transmitted to the inspection scenario adaptive multi-source data fusion module for re-fusion and detection to confirm the anomaly location and type, and achieve accurate anomaly localization.

5. An intelligent inspection data analysis system based on remote interaction and multi-agent systems, used to implement the method described in any one of claims 1-4, characterized in that, The system includes: Multi-agent inspection terminal: It consists of multiple intelligent inspection devices of different types, used to collect on-site inspection data, receive remote interactive commands, execute inspection tasks, and transmit the collected data to the inspection scene adaptive multi-source data fusion module through the remote interactive communication module. Remote interactive communication module: It adopts Mesh adaptive networking technology to realize data transmission and command interaction between multiple agents and the backend, and between multiple agents, reducing transmission latency and packet loss rate, and ensuring the real-time performance and stability of data transmission; The inspection scenario adaptive multi-source data fusion module receives multi-source heterogeneous inspection data transmitted by multi-agent inspection terminals, and adopts an improved inspection scenario adaptive OVO attention fusion algorithm to solve the problems of distortion in multi-source heterogeneous data fusion and difficulty in dynamic weight adjustment, and outputs high-precision fusion feature data. The feature-driven multi-agent collaborative anomaly localization and feedback module receives fused feature data output from the inspection scenario adaptive multi-source data fusion module. It adopts an improved dynamic threshold anomaly detection and multi-agent collaborative scheduling algorithm to achieve accurate anomaly identification, localization and root cause tracing. At the same time, it generates remote interaction commands and feeds them back to the multi-agent inspection terminal through the remote interaction communication module to realize real-time adjustment of the inspection strategy and form a closed loop. Data storage module: Used to store raw inspection data, fused feature data, anomaly detection results, remote interaction commands and inspection logs collected by multiple agents, providing data support for subsequent data analysis and traceability; Remote terminal interaction interface: Allows staff to view inspection data, fusion results, and abnormal information, issue remote inspection commands, and realize remote interaction with multiple intelligent agents.

Images