AI large model driven iot fault diagnosis system

Abstract

The application discloses an AI large model driven Internet of Things fault diagnosis system, which comprises a data acquisition module, a dynamic detection module, a positioning analysis module and a fault diagnosis module; the data acquisition module is used for acquiring original data of terminal equipment and performing feature classification to obtain initial data of the Internet of Things; the dynamic detection module is used for establishing a multi-modal fault alarm device according to the initial data of the Internet of Things; the positioning analysis module is used for analyzing an influence range of downstream equipment by using three-dimensional influence modeling and adopting a backstepping hypothesis method; and the fault diagnosis module is used for calling a large model driven interface, positioning a fault source, diagnosing a root cause of equipment fault, and realizing a complete closed loop from original signals to feature extraction, multi-modal alarm, influence analysis and root cause diagnosis.

Description

Technical Field

[0001] This invention relates to the field of anomaly detection and fault classification technology, and in particular to an IoT fault diagnosis system driven by AI large models. Background Technology

[0002] With the development of modern technology, enterprises have significantly increased their demand for equipment health correlation. Traditional fault diagnosis methods can no longer meet the needs of efficient operation and maintenance in modern industry. However, IoT-based intelligent fault diagnosis systems can achieve real-time monitoring, fault early warning, and predictive diagnosis, thereby improving production efficiency.

[0003] A Chinese invention with application number 202410624493.X discloses an IoT fault diagnosis method and system based on intelligent optimization algorithms. The method includes: collecting operating data of devices in a mobile IoT device and preprocessing the operating data to eliminate the influence of noise and outliers; processing the preprocessed operating data to identify abnormal data and potential faulty devices, determine the device operating status, and locate the faulty device; performing a detailed analysis of the faulty device based on the output results of the intelligent optimization algorithm to determine the cause of the fault, explore potential causes of the fault, and conduct a comprehensive analysis of the fault in combination with actual operating conditions and historical data.

[0004] The above methods failed to perform cross-dimensional data fusion based on nonlinear data, did not set up a spatial regression model, did not use the reverse hypothesis method to locate the fault source, lacked the establishment of topological relationships between devices, and lacked the establishment of an alarm system. Summary of the Invention

[0005] The technical problems solved by this invention are: failure to perform cross-dimensional data fusion based on nonlinear data, lack of spatial regression model, lack of fault source location using reverse hypothesis method, lack of establishment of topological relationship between devices, and lack of alarm system.

[0006] To solve the above-mentioned technical problems, the present invention provides the following technical solution:

[0007] The AI-driven IoT fault diagnosis system includes: a data acquisition module, a dynamic detection module, a location analysis module, and a fault diagnosis module.

[0008] The data acquisition module is used to collect raw data from terminal devices and perform feature classification to obtain initial IoT data.

[0009] The dynamic detection module is used to establish a multimodal fault alarm device based on the initial IoT data;

[0010] The positioning analysis module is used to analyze the impact range of downstream equipment by using three-dimensional impact modeling and the reverse assumption method.

[0011] The fault diagnosis module is used to call the large model driver interface to diagnose the root cause of equipment faults based on the fault source location.

[0012] Preferably, the specific method of the data acquisition module includes:

[0013] Edge computing is used to collect raw data from various terminals and sensors in real time, and the raw data is uploaded to the Internet of Things cloud platform. Based on the operating characteristics of the raw data, features are extracted from the raw data to obtain initial Internet of Things data, which includes first fault data, second fault data, third fault data and fourth fault data.

[0014] The raw data includes equipment status logs, running time, status signal data, physical monitoring data, environmental data, and positioning beacon data. The physical monitoring data includes vibration signals and current signals, and the environmental data includes temperature signals, air pressure signals, and air humidity.

[0015] The first fault data is used to store instantaneous sudden faults;

[0016] The second fault data is used to store long-term progressive faults;

[0017] The third fault data is used to store multi-factor composite faults;

[0018] The fourth fault data is used to store unknown new types of faults.

[0019] Preferably, the specific method for extracting features from the original data based on its operational characteristics is as follows:

[0020] Using the sliding window on the IoT platform, the original data in the sliding window within a unit time period is extracted, the peak value and RMS value of the state signal of the original data are calculated, wavelet packet decomposition is performed on the state signal data, and the Welch method is used to calculate the proportion of energy greater than the unit frequency. When the ratio of the peak value of the state signal to the RMS value is greater than a set threshold and the proportion of energy is greater than the average energy per unit frequency, the original data is saved as the first fault data.

[0021] Based on the running time of the original data, the running time is divided into 30-day units to plot the regular fluctuation chart of the original data, obtain the regular fluctuation trend of each week, calculate the slope of the regular fluctuation trend, and when the slope continues to increase within the unit period, the original data is saved as the second fault data.

[0022] The mean values ​​of the vibration signal, current signal, and temperature signal of the original data are calculated respectively. The original data that are equal to the corresponding mean values ​​of the vibration signal, current signal, and temperature signal are deleted. The cross-correlation function is used to calculate the cross-correlation function value of vibration, temperature, and current. When the cross-correlation function value is greater than the correlation threshold, the original data is saved as the third fault data.

[0023] The original data is used to construct a Vietoris-Rips complex topology based on single sample data points. The number of ring structures and connected components of the sample data points are obtained, and the life cycle of the topological features is recorded. The mean of the number of ring structures and connected components is calculated. When the number of connected components and the number of ring structures are greater than the mean, the sample data is saved as the fourth fault data.

[0024] Preferably, the dynamic detection module is used to establish a multimodal fault alarm device, which includes a pulse alarm unit, a progressive coloring unit, a reverse tracing unit, and a topological anomaly unit.

[0025] The pulse alarm unit includes: establishing a peak value preserver based on the status signal of the first fault data; detecting the status signal within the current sampling period at each clock rise; calculating the standard deviation and average value of the status signal within the current sampling period; and calculating a dynamic threshold within a set unit threshold clock period. The expression for calculating the dynamic threshold is as follows:

[0026] Dynamic threshold = Average state signal + 5 x Standard deviation of state signal;

[0027] When the status signal is greater than the dynamic threshold, the first alarm is triggered and the timing device is started.

[0028] Preferably, the progressive staining unit includes:

[0029] The device that acquires the data corresponding to the second fault data is independently simulated. Data interfaces are set for each device. The three-dimensional device model is tinted according to the slope fluctuation range of the second fault data. The initial model tint is set to blue, which means the device is in a healthy state. The hue is rendered from high to low according to the slope fluctuation range. Red is set to indicate the device is in a fault state. When the device is in a fault state, the second alarm is triggered and the long-term positioning timing device is started.

[0030] Preferably, the reverse tracing unit includes:

[0031] The average cross-correlation function of vibration, temperature and current under normal operating conditions is calculated per unit time. An adaptive adjustment mechanism is used to automatically update the average cross-correlation function weekly to obtain the average dynamic cross-correlation function. When any cross-correlation value is greater than the average dynamic cross-correlation function and the duration exceeds the average time of the third fault data, a third alarm is triggered and the reverse tracing and positioning device is activated.

[0032] Preferably, the topological anomaly unit includes:

[0033] A lightweight topology feature extraction module is deployed to calculate the number of connected components and the number of ring structures in real time. The mean of the number of connected components and the number of ring structures per unit time is calculated. An initial threshold is set based on the mean to obtain the fluctuation range of the number of connected components and the number of ring structures. The fluctuation range is used to obtain the unit time span period of the number of connected components and the number of ring structures. When the number of connected components and the number of ring structures are greater than the initial threshold and the fluctuation range exceeds three span periods, a fourth alarm is triggered and the terminal feedback device is activated.

[0034] Preferably, the positioning analysis module includes:

[0035] Based on a random algorithm, N possible fault sources are automatically generated for the device. Using these N potential fault source devices, a three-dimensional impact model is constructed. The three-dimensional impact model is built using mechanical vibration, heat conduction, and current disturbance. The mechanical vibration is calculated using an exponential decay model, the heat conduction uses a non-homogeneous thermal equation, and the current disturbance is calculated using an impedance network model. Based on digital twin technology, the initial IoT data is input into the three-dimensional impact model. The impact area is displayed using a color-temperature gradient surface. When the vibration signal, heat conduction, and current signal disturbance distributions exceed 1.5 times the threshold of the three-dimensional impact model, the device is set to the red core area. When the vibration signal, heat conduction, and current signal disturbance distributions are below 1.5 times the threshold but above 1.2 times the threshold, the device is set to the yellow transition area. When the vibration signal, heat conduction, and current signal disturbance distributions are below 1.2 times the threshold, the device is set to the green safe area.

[0036] Preferably, the fault diagnosis module includes:

[0037] Fault data of equipment in the red core area is extracted, and the fault data is fused with multimodal features. The multimodal feature fusion is used to sort the sensor data and maintenance log text of the fault data according to the time series, call the large model interface, and obtain diagnostic results and decision suggestions.

[0038] Preferably, the first alarm is used to set a countdown alarm at a set time;

[0039] The second alarm is used to activate the sensor's continuous recording mode and start the positioning device to obtain the location of the anomaly source;

[0040] The third alarm is used to reverse track the rendering propagation path and automatically lock the upstream and downstream three-level related devices;

[0041] The fourth alarm is used to compare the virtual model with the actual fault time in real time and adjust the fault stop time accordingly.

[0042] The beneficial effects of this invention are as follows: It constructs a complete equipment health management system through four core modules, realizing a complete closed loop from raw signals to feature extraction, multimodal alarms, impact analysis, and root cause diagnosis. At the same time, it integrates three-dimensional data of time-domain waveforms, frequency-domain features, and topological relationships to achieve cross-dimensional data fusion. Based on four alarm devices, it thoroughly classifies the root causes of faults, making it easier for staff to find the overall direction of the fault. It achieves full-process coverage from early warning to emergency braking, simulates the topological structure of the equipment, and establishes a three-dimensional visualized fault model, providing an effective solution for the full life cycle management of industrial equipment. Attached Figure Description

[0043] Figure 1 This is a basic flowchart of an AI-driven IoT fault diagnosis system provided in one embodiment of the present invention. Detailed Implementation

[0044] To make the above-mentioned objects, features and advantages of the present invention more apparent and understandable, the specific embodiments of the present invention will be described in detail below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments.

[0045] Example 1, referring to Figure 1 As one embodiment of the present invention, an AI-driven large model-based IoT fault diagnosis system is provided, comprising:

[0046] Data acquisition module, dynamic detection module, location analysis module, and fault diagnosis module;

[0047] The data acquisition module is used to collect raw data from terminal devices and perform feature classification to obtain initial IoT data;

[0048] The dynamic detection module is used to establish a multimodal fault alarm device based on initial IoT data;

[0049] The location analysis module is used to analyze the impact range of downstream equipment using three-dimensional impact modeling and a reverse assumption method.

[0050] The fault diagnosis module is used to call the large model driver interface, locate the fault source, and diagnose the root cause of the equipment fault.

[0051] In this embodiment, the system is divided into four modules, and the main operation content and implementation rules of each module are initially defined, which helps to refine the system construction method.

[0052] The specific methods of the data acquisition module include:

[0053] Edge computing is used to collect raw data from various terminals and sensors in real time and upload the raw data to the IoT cloud platform. Based on the operational characteristics of the raw data, features are extracted to obtain the initial IoT data, which includes the first fault data, the second fault data, the third fault data and the fourth fault data.

[0054] The raw data includes equipment status logs, running time, status signal data, physical monitoring data, environmental data, and positioning beacon data. The physical monitoring data includes vibration signals and current signals, and the environmental data includes temperature signals, air pressure signals, and air humidity.

[0055] The first fault data is used to store transient, sudden faults;

[0056] The second fault data is used to store long-term progressive faults;

[0057] The third fault data is used to store multi-factor composite faults;

[0058] The fourth fault data is used to store unknown new types of faults.

[0059] In this embodiment, edge computing is used to integrate multi-source heterogeneous data such as equipment status, vibration current and environmental parameters in real time to build a digital twin base for industrial sites. A four-level fault management mechanism is initially set up to provide a basic direction for subsequent accurate fault diagnosis.

[0060] Based on the operational characteristics of the original data, the specific method for feature extraction from the original data is as follows:

[0061] Using the sliding window on the IoT platform, the original data in the sliding window within a unit time is extracted, the peak value and RMS value of the state signal of the original data are calculated, wavelet packet decomposition is performed on the state signal data, and the Welch method is used to calculate the proportion of energy greater than the unit frequency. When the ratio of the peak value of the state signal to the RMS value is greater than the set threshold and the proportion of energy is greater than the average energy of the unit frequency, the original data is saved as the first fault data.

[0062] Based on the running time of the original data, the running time is divided into 30-day units to plot the regular fluctuation chart of the original data, obtain the regular fluctuation trend of each week, calculate the slope of the regular fluctuation trend, and when the slope continues to increase within the unit period, the original data is saved as the second fault data.

[0063] The mean values ​​of vibration, current and temperature signals of the original data are calculated separately. The original data that are equal to the corresponding mean values ​​of vibration, current and temperature signals are deleted. The cross-correlation function is used to calculate the cross-correlation function value of vibration, temperature and current. When the cross-correlation function value is greater than the correlation threshold, the original data is saved as the third fault data.

[0064] The original data is used to construct a Vietoris-Rips complex topology based on single sample data points. The number of ring structures and connected components of the sample data points are obtained, and the life cycle of the topological features is recorded. The mean of the number of ring structures and connected components is calculated. When the number of connected components and the number of ring structures are greater than the mean, the sample data is saved as the fourth fault data.

[0065] This embodiment uses four types of fault identification methods to cover the entire lifecycle of equipment faults. Secondly, it adopts adaptive signal processing technology and uses sliding window dynamic optimization to ensure the timeliness of feature extraction. It constructs a triple verification mechanism of time threshold peak value, frequency domain energy and topological features, and develops a dynamic threshold adjustment method to effectively adapt to environmental fluctuations of new data and reduce the false alarm rate.

[0066] The dynamic detection module is used to establish a multimodal fault alarm device, which includes a pulse alarm unit, a progressive coloring unit, a reverse tracing unit, and a topological anomaly unit.

[0067] The pulse alarm unit includes: establishing a peak value preserver based on the status signal of the first fault data; detecting the status signal within the current sampling period at each clock rise; calculating the standard deviation and average value of the status signal within the current sampling period; and calculating a dynamic threshold within a set unit threshold clock period. The expression for calculating the dynamic threshold is:

[0068] Dynamic threshold = Average state signal + 5 x Standard deviation of state signal;

[0069] When the status signal exceeds the dynamic threshold, the first alarm is triggered and the timing device is activated.

[0070] In this embodiment, a dynamic detection module and a multimodal fault alarm device are designed. The standard deviation of the state signal is set to 5 times. Based on Gaussian distribution operation, the dynamic threshold can be fully covered. An adaptive threshold model is adopted to effectively capture transient pulses of duration. An early warning strategy is set to realize accurate identification and graded early warning of industrial equipment faults. Based on multi-channel collaborative verification, pulse duration analysis is realized.

[0071] The progressive staining unit includes:

[0072] The device that acquires the data corresponding to the second fault data is independently simulated. The data interface corresponding to each device is set. The 3D device model is tinted according to the slope fluctuation range of the second fault data. The initial model tint is set to blue, which means the device is in a healthy state. The hue is rendered from high to low according to the slope fluctuation range. Red is set to the device status fault. When the device status is faulty, the second alarm is triggered and the long-term positioning timer is started.

[0073] This embodiment employs dynamic coloring technology to provide intuitive perception and accurate early warning based on device status, effectively enabling long-term advance warning and ensuring real-time monitoring of faults.

[0074] The reverse traceability unit includes:

[0075] The system calculates the mean cross-correlation function of vibration, temperature, and current under normal operating conditions based on raw data within a unit of time. It employs a sampling adaptive adjustment mechanism to automatically update the mean cross-correlation function weekly, obtaining a dynamic mean cross-correlation function. When any cross-correlation value exceeds the dynamic mean cross-correlation function and its duration exceeds the mean of the correlation time of the third fault data, a third alarm is triggered, and a reverse tracing and positioning device is activated.

[0076] In this embodiment, a multi-dimensional tracing engine is set up. Based on the mean of the cross-correlation function of the time and space dimensions and the similarity of historical cases, an intelligent learning mechanism is constructed to realize the automatic updating of the fault propagation mode.

[0077] Topological anomaly units include:

[0078] A lightweight topology feature extraction module is deployed to calculate the number of connected components and the number of ring structures in real time. The average number of connected components and the number of ring structures per unit time is calculated. An initial threshold is set based on the average value to obtain the fluctuation range of the number of connected components and the number of ring structures. The fluctuation range is used to obtain the unit time span period of the number of connected components and the number of ring structures. When the number of connected components and the number of ring structures are greater than the initial threshold and the fluctuation range exceeds three span periods, a fourth alarm is triggered and the terminal feedback device is activated.

[0079] In this embodiment, lightweight topology feature extraction is configured, and abnormal structures are highlighted and fault locations are marked according to the fluctuation range set by the topology device, thereby improving the characteristic status display of the fourth alarm and facilitating timely feedback from staff.

[0080] The location analysis module includes:

[0081] Based on a random algorithm, N possible fault sources are automatically generated for the device. Using these N potential fault source devices, a three-dimensional impact model is constructed. The three-dimensional impact model is constructed using mechanical vibration, heat conduction, and current disturbance. The mechanical vibration is calculated using an exponential decay model, the heat conduction uses a non-homogeneous thermal equation, and the current disturbance is calculated using an impedance network model. Based on digital twin technology, the initial IoT data is input into the three-dimensional impact model, and the impact area is displayed using a color-temperature gradient surface. When the vibration signal, heat conduction, and current signal disturbance distributions exceed 1.5 times the threshold of the three-dimensional impact model, the device is set to the red core area. When the vibration signal, heat conduction, and current signal disturbance distributions are below 1.5 times the threshold but above 1.2 times the threshold, the device is set to the yellow transition area. When the vibration signal, heat conduction, and current signal disturbance distributions are below 1.2 times the threshold, the device is set to the green safe area.

[0082] In this embodiment, the digital twin technology of multi-physics coupling is used to achieve accurate visualization and positioning of the fault impact domain and randomly generate N candidate sources, which effectively improves the emergency positioning response speed of the fault location and improves the economic effect of IoT fault positioning.

[0083] The fault diagnosis module includes:

[0084] Fault data of equipment in the red core area is extracted, and multimodal feature fusion is performed on the fault data. Multimodal feature fusion is used to sort the sensor data and maintenance log text of the fault data according to the time series, call the large model interface, and obtain diagnostic results and decision suggestions.

[0085] In this embodiment, the knowledge enhancement model is trained by fusing multimodal data, which effectively improves diagnostic accuracy, decision response speed, and cross-system compatibility, reduces system misjudgment losses, and extends equipment life.

[0086] The first alarm is used to set a countdown alarm at a set time.

[0087] The second alarm is used to activate the sensor's continuous recording mode and start the positioning device to obtain the location of the anomaly source;

[0088] The third alarm is used to reverse trace the rendering propagation path and automatically lock the upstream and downstream three-level related devices;

[0089] The fourth alarm is used to compare the virtual model with the actual fault time in real time and adjust the fault stop time accordingly.

[0090] In this embodiment, four types of alarm devices are set up to provide graded response efficiency. Log recording, location points and path visualization are used to effectively improve equipment safety, track faulty equipment in real time, and facilitate subsequent analysis of single-level fault trends.

[0091] Those skilled in the art will understand that embodiments of the present invention can be provided as methods, systems, or computer program products. Therefore, the present invention can take the form of a completely hardware embodiment, a completely software embodiment, or an embodiment combining software and hardware aspects. Furthermore, the present invention can take the form of a computer program product implemented on one or more computer-usable storage media containing computer-usable program code. The storage medium can be implemented by any type of volatile or non-volatile storage device or a combination thereof, such as Static Random Access Memory (SRAM), Electrically Erasable Programmable Read-Only Memory (EEPROM), Erasable Programmable Read Only Memory (EPROM), Programmable Red-Only Memory (PROM), Read-Only Memory (ROM), magnetic storage, flash memory, magnetic disk, or optical disk. These computer program instructions may also be stored in a computer-readable storage medium that can direct a computer or other programmable data processing device to function in a particular manner, such that the instructions stored in the computer-readable storage medium produce an article of manufacture including instruction means, which are implemented in a process Figure 1 One or more processes and / or boxes Figure 1 The function specified in one or more boxes.

[0092] It should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention and are not intended to limit it. Although the present invention has been described in detail with reference to preferred embodiments, those skilled in the art should understand that modifications or equivalent substitutions can be made to the technical solutions of the present invention without departing from the spirit and scope of the technical solutions of the present invention, and all such modifications or substitutions should be covered within the scope of the claims of the present invention.

Claims

1. An AI-driven IoT fault diagnosis system, characterized in that, include: Data acquisition module, dynamic detection module, location analysis module, and fault diagnosis module; The data acquisition module is used to collect raw data from terminal devices and perform feature classification to obtain initial IoT data. The dynamic detection module is used to establish a multimodal fault alarm device based on the initial IoT data; The positioning analysis module is used to analyze the impact range of downstream equipment by using three-dimensional impact modeling and the reverse assumption method. The fault diagnosis module is used to call the large model driver interface to diagnose the root cause of equipment faults based on the fault source location. The dynamic detection module is used to establish a multimodal fault alarm device, which includes a pulse alarm unit, a progressive coloring unit, a reverse tracing unit, and a topological anomaly unit. The pulse alarm unit includes: establishing a peak value preserver based on the status signal of the first fault data; detecting the status signal within the current sampling period at each clock rise; calculating the standard deviation and average value of the status signal within the current sampling period; and calculating a dynamic threshold within a set unit threshold clock period. The expression for calculating the dynamic threshold is as follows: Dynamic threshold = Average state signal + 5 x Standard deviation of state signal; When the status signal is greater than the dynamic threshold, a first alarm is triggered and a timing device is started; The progressive staining unit includes: The device that acquires the data corresponding to the second fault data is independently simulated. Data interfaces are set for each device. The three-dimensional device model is colored according to the slope fluctuation range of the second fault data. The initial model color is set to blue, which means the device is healthy. The hue is rendered from high to low according to the slope fluctuation range. Red is set to the device status fault. When the device status is faulty, the second alarm is triggered and the long-term positioning timer is started. The reverse tracing unit includes: The mean cross-correlation function of the vibration signal, current signal and temperature signal under normal operating conditions is calculated within a unit time. An adaptive adjustment mechanism is adopted to automatically update the mean cross-correlation function every seven days to obtain the mean dynamic cross-correlation function. When any cross-correlation value is greater than the mean dynamic cross-correlation function and the duration exceeds the mean correlation time of the third fault data, the third alarm is triggered and the reverse tracing and positioning device is activated. The topology anomaly unit includes: A lightweight topology feature extraction module is deployed to calculate the number of connected components and the number of ring structures in real time. The mean of the number of connected components and the number of ring structures per unit time is calculated. An initial threshold is set based on the mean to obtain the fluctuation range of the number of connected components and the number of ring structures. The fluctuation range is used to obtain the unit time span period of the number of connected components and the number of ring structures. When the number of connected components and the number of ring structures are greater than the initial threshold and the fluctuation range exceeds three span periods, a fourth alarm is triggered and the terminal feedback device is activated. The location analysis module includes: Based on a random algorithm, N possible fault sources are automatically generated for the device. Using these N possible fault source devices, a three-dimensional impact model is constructed. The three-dimensional impact model is constructed using mechanical vibration, heat conduction, and current disturbance. The mechanical vibration is calculated using an exponential decay model, the heat conduction is calculated using a non-homogeneous thermal equation, and the current disturbance is calculated using an impedance network model. Based on digital twin technology, the initial IoT data is input into the three-dimensional impact model. The impact area is displayed using a color-temperature gradient surface. When the vibration signal, heat conduction, and current signal disturbance distributions exceed 1.5 times the threshold of the three-dimensional impact model, the device is set to the red core area. When the vibration signal, heat conduction, and current signal disturbance distributions are below 1.5 times the threshold but above 1.2 times the threshold, the device is set to the yellow transition area. When the vibration signal, heat conduction, and current signal disturbance distributions are below 1.2 times the threshold, the device is set to the green safe area. The fault diagnosis module includes: Fault data of equipment in the red core area is extracted, and the fault data is subjected to multimodal feature fusion. The multimodal feature fusion is used to sort the sensor data and maintenance log text of the fault data according to the time series. The fault data is input into the large model for training, and historical fault data and online networking mechanism are fused to obtain a knowledge-enhanced large model. The knowledge-enhanced large model interface is called to obtain diagnostic results and decision suggestions.

2. The AI-driven large-scale IoT fault diagnosis system as described in claim 1, characterized in that: The specific methods of the data acquisition module include: Edge computing is used to collect raw data from various terminals and sensors in real time, and the raw data is uploaded to the Internet of Things cloud platform. Based on the operating characteristics of the raw data, features are extracted from the raw data to obtain initial Internet of Things data, which includes first fault data, second fault data, third fault data and fourth fault data. The raw data includes equipment status logs, running time, status signal data, physical monitoring data, environmental data, and positioning beacon data. The physical monitoring data includes vibration signals and current signals, and the environmental data includes temperature signals, air pressure signals, and air humidity. The first fault data is used to store instantaneous sudden faults; The second fault data is used to store long-term progressive faults; The third fault data is used to store multi-factor composite faults; The fourth fault data is used to store unknown new types of faults.

3. The AI-driven large-scale IoT fault diagnosis system as described in claim 2, characterized in that: The specific method for feature extraction from the original data based on its operational characteristics is as follows: Using the sliding window on the IoT cloud platform, the original data in the sliding window within a unit time period is extracted, the peak value and RMS value of the state signal of the original data are calculated, wavelet packet decomposition is performed on the state signal data, and the energy ratio of the extracted signal greater than the unit frequency is calculated using the Welch method. When the ratio of the peak value of the state signal to the RMS value is greater than a set threshold and the energy ratio is greater than the average energy ratio of the signal, the original data is saved as the first fault data. Based on the running time of the original data, the running time is divided into 30-day units to plot the regular fluctuation chart of the original data, obtain the regular fluctuation trend every seven days, calculate the slope of the regular fluctuation trend, and when the slope continues to increase within the unit period, the original data is saved as the second fault data. The mean values ​​of the vibration signal, current signal, and temperature signal of the original data are calculated respectively. The original data that are equal to the corresponding mean values ​​of the vibration signal, current signal, and temperature signal are deleted. The cross-correlation function is used to calculate the cross-correlation function value of the vibration signal, current signal, and temperature signal. When the cross-correlation function value is greater than the correlation threshold, the original data is saved as the third fault data. The original data is used to construct a Vietoris-Rips complex topology based on single sample data points. The number of ring structures and connected components of the sample data points are obtained, and the life cycle of the topological features is recorded. The mean of the number of ring structures and connected components is calculated. When the number of connected components and the number of ring structures are greater than the mean, the sample data is saved as the fourth fault data.

4. The AI-driven large-scale IoT fault diagnosis system as described in claim 3, characterized in that: The first alarm is used to set a countdown alarm at a set time. The second alarm is used to activate the sensor's continuous recording mode and start the positioning device to obtain the location of the anomaly source; The third alarm is used to reverse track the rendering propagation path and automatically lock the upstream and downstream three-level related devices; The fourth alarm is used to compare the virtual model with the actual fault time in real time and adjust the fault stop time accordingly.

Images