Converter primary flue gas electrostatic precipitator state monitoring and fault diagnosis system and method

Abstract

The invention provides a state monitoring and fault diagnosis system and method for a primary flue gas electrostatic dust collector of a converter. The system comprises a multi-source data acquisition module, a signal preprocessing and feature extraction module, an intelligent diagnosis core module and a man-machine interaction and early warning module. The invention further relates to a state monitoring and fault diagnosis method for the converter primary flue gas electrostatic dust collector. According to the invention, an improved PRONY algorithm is adopted to extract damping factors, oscillation frequency and amplitude characteristics, and a total least square method is adopted to improve parameter estimation accuracy; an integrated deep learning model combining a convolutional neural network, a long-short-term memory network and an attention mechanism is adopted to perform fault recognition, and a probability graph model based on a Bayesian network is adopted to perform fault positioning and reason reasoning; and the man-machine interaction and early warning module generates multi-level early warning information according to the diagnosis result and visually displays the multi-level early warning information. The fault diagnosis accuracy of the system reaches 94.5%-97.2%, the early fault early warning time is advanced by 12-72 hours, and the intelligent level of the dust removal system is effectively improved.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of intelligent monitoring of industrial dust removal equipment; in particular, it relates to a converter primary flue gas electrostatic precipitator state monitoring and fault diagnosis system and method. BACKGROUND

[0002] Converter gas dry dust removal systems, especially their core equipment electrostatic precipitator (ESP), play a crucial role in the flue gas treatment of the steel industry. It has the characteristics of high dust removal efficiency, low energy consumption, and the ability to handle high-temperature and large flue gas volume gas. However, the ESP system is long-term operated in an environment of high temperature, high pressure, and accompanied by severe mechanical vibration, which is prone to various faults, such as mechanical failure, electrical failure (including reverse corona failure), and inter-electrode distance reduction failure. These faults will cause the dust removal efficiency to decrease, and even cause the system to shut down, resulting in economic losses.

[0003] In the field of electrostatic precipitator state monitoring and fault diagnosis technology, foreign research started early. European patent EP3095520A1 discloses a method for monitoring the signal quality of an electrostatic precipitator. This method measures the electrical parameters such as primary and secondary voltage and current in real time, uses fast Fourier transform for frequency spectrum analysis, and evaluates the signal quality by calculating the power percentage of specific harmonics and zero-crossing deviation. Although this technology realizes real-time monitoring of electrical parameters, it is limited to simple signal quality judgment and cannot identify and diagnose specific fault types.

[0004] International patent WO2003083731A2 describes a PC-based electrostatic precipitator visualization, diagnosis, and expert system. The system is connected to the high-voltage power supply unit network through a server PC, and uses modular software design to realize functions such as data transmission, measurement data archiving, visualization display, and parameter setting. Although this system builds a complete monitoring architecture, it uses traditional expert system technology and lacks autonomous learning ability, with limited diagnostic accuracy.

[0005] For specific fault diagnosis of electrostatic precipitators, US patent US4390830A proposes a reverse corona detection and current back-off method, which detects reverse corona faults by monitoring sudden changes in voltage and current relationships. This method uses comparator detection technology to automatically reduce the supply current when reverse corona is detected. However, this technology can only detect reverse corona faults and cannot distinguish other types of faults, and the detection accuracy is limited.

[0006] In the steel industry, relevant research mainly focuses on control system design. The paper "The Analysis and Design of Steel Plant Electrostatic Precipitator Control System Based on IFIX" analyzes the control system of an electrostatic precipitator in a steel plant based on the IFIX software platform, focusing on solving the monitoring and control challenges in industrial ESP applications, but mainly focusing on process control rather than fault diagnosis.

[0007] In recent years, some researchers have attempted to introduce intelligent methods into fault diagnosis. For example, some studies have used the PRONY method to analyze transient signals during ESP spark discharge and extract damping factors to diagnose back corona faults and electrode spacing reduction faults. Other studies have attempted to establish static control models to quantify the impact of the smelting process on dust removal and have performed two-level fault judgment. In addition, some studies have applied neural networks to fault diagnosis of converter fans.

[0008] However, existing ESP fault diagnosis technologies still have many limitations:

[0009] First, there is a heavy reliance on manual experience. The operation and maintenance of the vast majority of ESPs in my country are still done manually or semi-manually, which cannot truly meet the needs of dust removal sites. Manual judgment is not only inefficient but also unreliable, making it difficult to detect early-stage faults.

[0010] Second, the detection methods are limited and lack precision. Existing methods mainly rely on information from a single sensor, such as monitoring only electrical parameters or vibration signals, and lack the ability to fuse and analyze multi-source data. Traditional signal analysis methods (such as simple FFT analysis) are difficult to extract deep-seated fault characteristics.

[0011] Third, fault confusion and inadequate diagnosis. Back corona faults and narrowing electrode spacing faults are easily confused using traditional methods, lacking effective means of differentiation. Traditional fault diagnosis methods (such as simple threshold alarms) have high false alarm and false negative rates, and are difficult to accurately locate faults and perform root cause analysis.

[0012] Fourth, the level of real-time performance and intelligence is low. Many systems are unable to perform real-time analysis of massive amounts of operational data, and even more so lack the ability to predict trends and provide intelligent early warnings based on historical and real-time data. Existing systems often focus on single devices or specific faults, lacking comprehensive and intelligent status monitoring and diagnostic solutions for the entire ESP system.

[0013] Therefore, there is an urgent need for a solution that can deeply integrate the latest mathematical theories, process multi-source data in real time, and perform accurate intelligent diagnosis and prediction to improve the operational reliability and intelligence level of electrostatic precipitators in converter gas dry dust removal systems. SUMMARY

[0014] The present application aims at the technical defects of the prior art that the fault diagnosis of electrostatic precipitator relies on manual experience, is inefficient and unreliable, cannot effectively distinguish reverse corona fault and small inter-electrode distance fault, lacks real-time intelligent analysis and early fault warning capability, has high false positive rate and false negative rate, and is difficult to realize accurate positioning and root cause analysis, and provides a converter primary flue gas electrostatic precipitator state monitoring and fault diagnosis system and method based on intelligent mathematical theory, which has high intelligent degree, good diagnostic precision, strong real-time performance and can realize early fault warning.

[0015] The present application is realized by the following technical solutions:

[0016] The present application relates to a converter primary flue gas electrostatic precipitator state monitoring and fault diagnosis system, comprising: a multi-source data acquisition module, a signal preprocessing and feature extraction module, an intelligent diagnosis core module and a man-machine interaction and early warning module.

[0017] Among them,

[0018] The multi-source data acquisition module is configured to acquire electrical parameters, process parameters, mechanical vibration parameters and image information of the electrostatic precipitator in real time.

[0019] The signal preprocessing and feature extraction module is configured to perform wavelet threshold denoising processing on the collected transient spark discharge signal, and extract the damping factor, oscillation frequency and amplitude characteristics of the transient signal by using an improved PRONY algorithm.

[0020] The intelligent diagnosis core module is configured to use an integrated deep learning model for fault recognition, wherein the integrated deep learning model is a combination of convolutional neural network, long short-term memory network and attention mechanism, and uses a reasoning algorithm based on a probabilistic graph model for fault positioning and cause reasoning.

[0021] The man-machine interaction and early warning module is configured to visually display state information, diagnosis results and multi-level early warning information.

[0022] Preferably, the improved PRONY algorithm uses the total least squares method to improve the accuracy of parameter estimation, and uses an adaptive model order selection method to avoid false components, and distinguishes reverse corona fault and small inter-electrode distance fault by the range of extracted damping factor values.

[0023] Preferably, in the integrated deep learning model, the convolutional neural network is used to extract spatial features and local patterns in the multi-source data, the long short-term memory network is used to process time series parameters and capture time dependence, and the attention mechanism is used to focus on the features and time steps most relevant to a specific fault.

[0024] Preferably, the inference algorithm based on the probabilistic graph model is specifically: adopting a Bayesian network structure, network nodes representing fault symptoms, device states and root causes, establishing a conditional probability table based on historical data learning and expert prior knowledge, and adopting a joint tree algorithm for probability inference.

[0025] Preferably, the intelligent diagnosis core module further comprises a time series prediction unit configured to adopt a prediction algorithm based on a long short-term memory network to perform trend prediction on the key performance indicators, and to calculate a device health index by comprehensively considering multi-dimensional features.

[0026] Preferably, the electrical parameters collected by the multi-source data collection module include: primary voltage, primary current, secondary voltage and secondary current; the process parameters include: inlet and outlet temperatures of the evaporative cooler, water spraying flow, inlet and outlet pressure difference of the electrostatic precipitator, and coal gas concentration; and the mechanical vibration parameters include vibration signals of the fan and the motor.

[0027] The present application also relates to a converter primary flue gas electrostatic precipitator state monitoring and fault diagnosis method, comprising the following steps:

[0028] Step 1: collecting multi-source operation data of the electrostatic precipitator, including electrical parameters, process parameters, mechanical vibration parameters and image information;

[0029] Step 2: performing wavelet threshold denoising processing on the collected transient spark discharge signal, and extracting damping factor, oscillation frequency and amplitude characteristics by using an improved PRONY algorithm;

[0030] Step 3: inputting the extracted characteristics into an integrated deep learning model for fault identification, the model combining a convolutional neural network, a long short-term memory network and an attention mechanism, and adopting an inference algorithm based on a probabilistic graph model for fault positioning and cause reasoning;

[0031] Step 4: generating multi-level warning information according to the diagnosis result and performing visual display.

[0032] Preferably, the improved PRONY algorithm improves parameter estimation accuracy by using the total least squares method, and judges that the damping factor of the anti-corona fault is less than 0.1-0.3 range and the damping factor of the inter-electrode gap small fault is greater than 0.5-0.8 range according to the extracted damping factor value range.

[0033] Preferably, the inference algorithm based on the probabilistic graph model adopts a Bayesian network for inference, calculates the posterior probability of different fault causes when observing a fault symptom, and selects the fault cause with the highest probability as the diagnosis result.

[0034] Principle of the method of the present application:

[0035] The multi-source data acquisition module collects the electrical parameters (primary voltage, primary current, secondary voltage, secondary current) of the electrostatic precipitator, the process parameters (evaporative cooler inlet and outlet temperature, water spray flow, electrostatic precipitator inlet and outlet pressure difference, coal gas concentration), mechanical vibration parameters (fan and motor vibration signals) and image information.

[0036] The signal preprocessing and feature extraction module performs wavelet threshold denoising processing on the collected transient spark discharge signal, extracts the damping factor, oscillation frequency and amplitude characteristics of the transient signal by using the improved PRONY algorithm, and the improved PRONY algorithm uses the total least squares method to improve the parameter estimation accuracy, and uses the self-adaptive selection model order method to avoid false components.

[0037] The intelligent diagnosis core module adopts an integrated deep learning model for fault identification, which combines convolutional neural networks, long short-term memory networks and attention mechanisms, wherein the convolutional neural network is used to extract spatial features and local patterns in multi-source data, the long short-term memory network is used to process time series parameters and capture time dependence, and the attention mechanism is used to focus on the features and time steps most relevant to a specific fault; at the same time, a reasoning algorithm based on a probabilistic graph model is used for fault location and cause reasoning, which uses a Bayesian network structure, network nodes represent fault symptoms, device states and root causes, and a conditional probability table is established based on historical data learning and expert prior knowledge, and a joint tree algorithm is used for probability reasoning; the intelligent diagnosis core module also includes a time series prediction unit, which uses a long short-term memory network-based prediction algorithm to predict trends in key performance indicators, and calculates a device health index based on multi-dimensional features.

[0038] The man-machine interaction and early warning module generates multi-level early warning information based on the diagnosis results and visually displays state information, diagnosis results and early warning information.

[0039] The present application has the following advantages:

[0040] (1) The diagnostic accuracy and reliability are significantly improved: the present application can efficiently extract deep fault features by integrating a deep learning model combined with an improved PRONY algorithm, significantly improving the discrimination of easily confused faults such as reverse corona and small inter-electrode spacing, with a fault diagnosis confidence of 92%, effectively reducing false positives and false negatives.

[0041] (2) Early fault warning and predictive maintenance are achieved: the present application uses time series analysis and deep learning for trend prediction and health status evaluation, which can issue an early warning 12 hours before a fault significantly affects system efficiency, changing "after-the-fact maintenance" to "predictive maintenance", significantly reducing unplanned downtime.

[0042] (3) High intelligence and automation level: the system can automatically complete the whole process from data collection, analysis to diagnosis conclusion generation, greatly reducing the dependence on manual experience, labor intensity and dependence on expert experience.

[0043] (4) Strong uncertain reasoning ability: the invention introduces a probability graph model (Bayesian network), enabling the system to perform probabilistic reasoning and fault location in the presence of incomplete information and noise, more in line with the actual situation of industrial sites.

[0044] (5) Strong system openness and scalability: the invention is based on modular design, using fieldbus or industrial Ethernet architecture, easy to expand and maintain the system, and the involved diagnostic algorithms and sensors can be easily integrated into the system.

[0045] (6) Strong multi-dimensional data fusion processing capability: the invention can comprehensively process various types of data such as electrical, process, vibration, and image, achieving more comprehensive and accurate condition monitoring and fault diagnosis, significantly improving the operation efficiency and reliability of the dust removal system, and reducing equipment maintenance costs. BRIEF DESCRIPTION OF DRAWINGS

[0046] Figure 1 is a module diagram of a converter primary flue gas electrostatic precipitator condition monitoring and fault diagnosis system according to the invention;

[0047] Figure 2 is a flowchart of a converter primary flue gas electrostatic precipitator condition monitoring and fault diagnosis method according to the invention. DETAILED DESCRIPTION

[0048] The invention will be described in detail below in conjunction with specific embodiments. It should be noted that the following implementation examples are only further illustrations of the invention, and the protection scope of the invention is not limited to the following examples.

[0049] Example 1

[0050] The converter primary flue gas electrostatic precipitator condition monitoring and fault diagnosis system according to the invention, as shown in Figure 1 , comprises a multi-source data acquisition module, a signal preprocessing and feature extraction module, an intelligent diagnosis core module, and a man-machine interaction and early warning module.

[0051] Among them,

[0052] The multi-source data acquisition module is configured to acquire electrical parameters, process parameters, mechanical vibration parameters and image information of the electrostatic precipitator in real time.

[0053] The signal preprocessing and feature extraction module is configured to perform wavelet threshold denoising processing on the collected transient spark discharge signal, and extract the damping factor, oscillation frequency and amplitude characteristics of the transient signal by using an improved PRONY algorithm.

[0054] The intelligent diagnosis core module is configured to use an integrated deep learning model for fault identification; wherein the integrated deep learning model is a combination of a convolutional neural network, a long short-term memory network and an attention mechanism, and uses a probabilistic graph model-based reasoning algorithm for fault positioning and cause reasoning.

[0055] The man-machine interaction and early warning module is configured to visually display state information, diagnosis results and multi-level early warning information.

[0056] Further, the improved PRONY algorithm uses the total least squares method to improve the accuracy of parameter estimation, and uses an adaptive model order selection method to avoid false components, and distinguishes between anti-corona faults and small inter-electrode spacing faults by the range of the extracted damping factor values.

[0057] Further, in the integrated deep learning model, the convolutional neural network is used to extract spatial features and local patterns in multi-source data, the long short-term memory network is used to process time series parameters and capture time dependence, and the attention mechanism is used to focus on the features and time steps most relevant to a specific fault.

[0058] Further, the probabilistic graph model-based reasoning algorithm specifically uses a Bayesian network structure, with network nodes representing fault symptoms, device states and root causes, and uses historical data learning and expert prior knowledge to establish a conditional probability table, and uses a joint tree algorithm for probabilistic reasoning.

[0059] Further, the intelligent diagnosis core module further includes a time series prediction unit configured to use a long short-term memory network-based prediction algorithm to predict trends in key performance indicators, and to calculate a device health index based on multi-dimensional features.

[0060] Further, the electrical parameters collected by the multi-source data acquisition module include primary voltage, primary current, secondary voltage and secondary current; the process parameters include evaporative cooler inlet and outlet temperature, water spray flow, electrostatic precipitator inlet and outlet pressure difference, and coal gas concentration; and the mechanical vibration parameters include vibration signals of the fan and motor.

[0061] The present embodiment also relates to a converter primary flue gas electrostatic precipitator state monitoring and fault diagnosis method, as shown in Figure 2 , comprising the following steps:

[0062] Step 1: Collecting multi-source operating data of the electrostatic precipitator, including electrical parameters, process parameters, mechanical vibration parameters and image information;

[0063] Step 2, wavelet threshold denoising processing is performed on the collected transient spark discharge signal, and an improved PRONY algorithm is used to extract the damping factor, oscillation frequency and amplitude characteristics;

[0064] Step 3, the extracted characteristics are input into an integrated deep learning model for fault identification, the model combines convolutional neural network, long short-term memory network and attention mechanism, and uses a reasoning algorithm based on a probabilistic graph model for fault location and cause reasoning;

[0065] Step 4, generate multi-level warning information according to the diagnosis result and display it visually.

[0066] Further, the improved PRONY algorithm improves the accuracy of parameter estimation by total least squares method, and judges that the damping factor of the anti-cathode fault is less than 0.1-0.3 range, and the damping factor of the small inter-electrode distance fault is greater than 0.5-0.8 range according to the extracted damping factor value range.

[0067] Further, the reasoning algorithm based on the probabilistic graph model uses Bayesian network for reasoning, when the fault symptoms are observed, the posterior probability of different fault causes is calculated, and the fault cause with the highest probability is selected as the diagnosis result.

[0068] Embodiment 2

[0069] This embodiment relates to an intelligent monitoring and diagnosis device for a 260t converter dry dust removal system

[0070] Based on the dry dust removal system of a 260 t converter in a steel plant, a sensor network is deployed on the upper part of the existing system's vaporization cooling flue, evaporation cooler, electrostatic precipitator, and fan. A high-voltage sensor and a Rogowski coil with a rated voltage of 70 kV are installed in the ESP power supply circuit, with a sampling frequency of 100 kHz, used to collect secondary voltage and current transient signals. A differential pressure transmitter with an accuracy of 0.1% and a K-type thermocouple are installed at the inlet and outlet of the ESP to monitor the pressure difference and temperature. A vibration acceleration sensor with a range of ±50 g and a frequency range of 10 Hz to 10 kHz is installed on the fan bearing seat. A high-definition industrial camera with a resolution of 1920x1080 is installed at the appropriate position of the furnace mouth and flue. All sensor data are accessed through the Profinet industrial network to the industrial monitoring computer with system software. The system software adopts a layered architecture, using Python for algorithm module development, C++ for high-performance data acquisition and processing, and configuration software for HMI interface development. The improved PRONY algorithm uses total least squares (TLS), with a model order adaptive selection range of 10-30 orders and a signal length of 1024 points. The fault type is judged by the damping factor σ: σ = 0.15 ± 0.05 for anti-corona fault, and σ = 0.65 ± 0.15 for small inter-electrode distance fault. In the CNN-LSTM-Attention model, CNN contains 3 convolutional layers with a convolution kernel size of 3x3 and feature map numbers of 32, 64, and 128, respectively. The LSTM hidden unit number is 256. The attention mechanism uses dot product attention, with a weight matrix dimension of 256x256. The Bayesian network contains 15 nodes, covering fault symptoms, intermediate states, and root causes, and the conditional probability table is based on 3 years of historical data statistics.

[0071] Technical effects: In the 3-month continuous operation of the system, the fault diagnosis accuracy rate reached 94.5%, which was 17.7 percentage points higher than the traditional threshold alarm method of 76.8%. The early fault warning time was advanced by 12-48 hours, effectively avoiding 3 unplanned shutdowns and reducing maintenance costs by about 35%.

[0072] Example 3

[0073] This example relates to a monitoring and diagnosis system with high-precision parameter optimization

[0074] Building upon Example 2, a higher-precision sensor configuration and optimized algorithm parameters are employed. The sampling frequency of the high-pressure sensor is increased to 200kHz, the accuracy of the differential pressure transmitter is improved to 0.05%, and the frequency range of the vibration sensor is extended to 20kHz. In the improved PRONY algorithm, the signal length is extended to 2048 points, the model order range is adjusted to 15–40, and a more stringent damping factor criterion is adopted: back-corona fault σ = 0.12 ± 0.03, and interpole spacing fault σ = 0.70 ± 0.10. The CNN-LSTM-Attention model structure is optimized, with the CNN increasing to 4 convolutional layers, the number of feature maps adjusted to 64, 128, 256, and 512, the number of hidden units in the LSTM increasing to 512, and the attention layer adopting a multi-head attention mechanism (8 heads). The Bayesian network is expanded to 20 nodes, introducing more intermediate inference states, and the conditional probability table is built based on 5 years of historical data, including seasonal and operating condition variation factors.

[0075] Technical results: The accuracy of fault diagnosis is improved to 97.2%, the false alarm rate is reduced to 2.1%, the early warning time is extended to 72 hours, and the accuracy of distinguishing between back corona and interpole gap faults reaches 98.5%, which is significantly better than the 85.3% of the traditional method.

[0076] Example 4

[0077] This embodiment relates to an intelligent monitoring platform deployed across multiple networked systems.

[0078] This system is designed for applications involving multiple converters operating simultaneously in large steel enterprises, and features a distributed monitoring and diagnostic system. The system supports simultaneous monitoring of dust removal systems in 4-6 converters, employing a cloud-edge collaborative architecture. Lightweight CNN-LSTM models are deployed on edge nodes for real-time diagnostics, while a complete deep learning and Bayesian inference model is deployed in the cloud for precise analysis. The PRONY algorithm parameters have been improved to adaptively adjust for different converter capacities: σ threshold for a 120t converter is 0.18±0.04 for back corona discharge and 0.55±0.12 for inter-electrode spacing; σ threshold for a 260t converter is 0.15±0.05 for back corona discharge and 0.65±0.15 for inter-electrode spacing; and σ threshold for a 300t converter is 0.13±0.03 for back corona discharge and 0.75±0.10 for inter-electrode spacing. The system utilizes industrial internet protocols, supporting standard communication protocols such as OPC UA and MQTT, enabling data exchange with enterprise MES and ERP systems. A unified fault knowledge base and diagnostic model library are established, supporting online model updates and transfer learning.

[0079] Technical results: The stability of multi-system network operation reaches 99.5%, the fault diagnosis time of a single converter is less than 3 seconds, the accuracy of cloud-based deep analysis is 96.8%, and intelligent management of enterprise-level equipment groups is realized, improving the overall efficiency of equipment by about 25%.

[0080] The improved PRONY algorithm is used to extract the damping factor, oscillation frequency and amplitude characteristics, the total least squares method is used to improve the parameter estimation accuracy; the intelligent diagnosis core module adopts the integrated deep learning model combining convolutional neural network, long short-term memory network and attention mechanism for fault identification, adopts the probability graph model based on Bayesian network for fault positioning and reason inference; the man-machine interaction and early warning module generates multi-level early warning information according to the diagnosis result and carries out visual display. The fault diagnosis accuracy of the system reaches 94.5%-97.2%, the early fault early warning time is advanced by 12-72 hours, and the intelligent level of the dust removal system is effectively improved.

[0081] The specific embodiments of the present application are described above. It should be understood that the present application is not limited to the above specific embodiments, and various modifications or changes can be made by those skilled in the art within the scope of the claims, which does not affect the essence of the present application.

Claims

1. A converter once-through flue gas electrostatic precipitator condition monitoring and fault diagnosis system, characterized in that, The application relates to an intelligent diagnosis system for electrostatic precipitators. The system comprises a multi-source data acquisition module, a signal preprocessing and feature extraction module, an intelligent diagnosis core module and a man-machine interaction and early warning module. The multi-source data acquisition module is configured to acquire electrical parameters, process parameters, mechanical vibration parameters and image information of the electrostatic precipitator in real time. The signal preprocessing and feature extraction module is configured to perform wavelet threshold denoising processing on the collected transient spark discharge signal, and extract the damping factor, oscillation frequency and amplitude characteristics of the transient signal by using an improved PRONY algorithm. The intelligent diagnosis core module is configured to use an integrated deep learning model for fault identification. The integrated deep learning model combines a convolutional neural network, a long short-term memory network and an attention mechanism, and uses a reasoning algorithm based on a probabilistic graph model for fault positioning and cause reasoning. The man-machine interaction and early warning module is configured to visually display state information, diagnosis results and multi-level early warning information.

2. The converter once-through flue electrostatic precipitator condition monitoring and fault diagnosis system according to claim 1, characterized in that, The improved PRONY algorithm uses the total least squares method to improve the accuracy of parameter estimation, and uses an adaptive model order selection method to avoid false components.

3. The converter once-through flue electrostatic precipitator condition monitoring and fault diagnosis system according to claim 1, characterized in that, The damping factor value range is used to distinguish between anti-corona faults and small inter-electrode spacing faults.

4. The converter once-through flue electrostatic precipitator condition monitoring and fault diagnosis system according to claim 1, characterized in that, In the integrated deep learning model, the convolutional neural network is used to extract spatial features and local patterns in the multi-source data, the long short-term memory network is used to process time series parameters and capture time dependence, and the attention mechanism is used to focus on the features and time steps most relevant to a specific fault.

5. The converter once-through flue electrostatic precipitator condition monitoring and fault diagnosis system according to claim 1, characterized in that, The reasoning algorithm based on the probabilistic graph model specifically uses a Bayesian network structure, in which the network nodes represent fault symptoms, device states and root causes.

6. The converter once-through flue electrostatic precipitator condition monitoring and fault diagnosis system according to claim 1, characterized in that, The intelligent diagnosis core module further comprises a time series prediction unit configured to use a prediction algorithm based on the long short-term memory network to predict the trend of key performance indicators, and to calculate a device health index by comprehensively considering multi-dimensional features.

7. A method for monitoring and diagnosing faults of a converter once-through flue electrostatic precipitator, characterized by, The electrical parameters acquired by the multi-source data acquisition module include primary voltage, primary current, secondary voltage and secondary current. The process parameters include evaporative cooler inlet and outlet temperatures, water spray flow, electrostatic precipitator inlet and outlet pressure differences and coal gas concentration. The mechanical vibration parameters include vibration signals of the fan and the motor. The application further relates to an intelligent diagnosis method for electrostatic precipitators. Step 1: Collecting multi-source operation data of the electrostatic precipitator, including electrical parameters, process parameters, mechanical vibration parameters and image information. Step 2: Performing wavelet threshold denoising processing on the collected transient spark discharge signal, and extracting damping factor, oscillation frequency and amplitude characteristics by using an improved PRONY algorithm. Step 3: Inputting the extracted features into an integrated deep learning model for fault identification. The model combines a convolutional neural network, a long short-term memory network and an attention mechanism, and uses a reasoning algorithm based on a probabilistic graph model for fault positioning and cause reasoning. Step 4: Generating multi-level early warning information according to the diagnosis results and visually displaying the information.

8. The method according to claim 7, wherein the method is characterized by, The improved PRONY algorithm improves the parameter estimation accuracy by total least squares method, and judges that the damping factor of the back corona fault is 0.1-0.3 and the damping factor of the fault with small inter-electrode distance is 0.5-0.8 according to the extracted damping factor value range.

9. The method according to claim 7, wherein the method is characterized by, The inference algorithm based on the probabilistic graph model adopts a Bayesian network for inference, calculates the posterior probability of different fault causes when a fault symptom is observed, and selects the fault cause with the highest probability as the diagnosis result.

Images