Fault prediction and diagnosis system and method for regenerative thermal oxidizer

Abstract

This invention relates to a fault prediction and diagnosis system and method for a regenerative thermal oxidizer (RTO), comprising a multi-parameter prediction module, an abnormal parameter monitoring module, and a causal graph inference module. The system predicts the changing trends of parameters at various points in the RTO using a multi-parameter prediction model, establishes an abnormal parameter monitoring model to achieve real-time anomaly monitoring of future parameter trends, and inputs potentially abnormal parameters into the causal graph inference module in real time. Combining an expert knowledge base and a data-driven causal graph, the system uses simplification rules and Bayesian formulas to simplify and calculate the graph, outputting possible fault locations and probabilities. This enables predictive fault tracing and guides maintenance personnel to intervene in advance, preventing equipment shutdowns and effectively reducing pollutant emissions during shutdowns and the large amount of fuel required for restarting.

Description

Technical Field

[0001] This invention belongs to the field of environmental diagnostic technology, specifically a fault prediction and diagnosis system and method for a regenerative thermal oxidizer. Background Technology

[0002] Regenerative thermal oxidizers (RTOs) are widely used in the treatment of organic waste gas due to their high purification efficiency and low cost. However, RTOs pose several safety hazards during operation. For example, fluctuations in the concentration or flow rate of the input waste gas can cause significant temperature changes in the furnace, posing a potential explosion risk. Several RTO explosion accidents have occurred in China in recent years.

[0003] RTOs pose safety hazards during operation, including the possibility of explosions that could cause significant loss of life and property. Furthermore, RTOs often experience issues such as pipe blockages, unstable gas pressure and concentrations, leading to equipment shutdowns. Each shutdown releases large amounts of unpurified harmful gases, and restarting the equipment requires substantial energy.

[0004] Currently, RTO equipment maintenance is often reactive, with repairs typically only undertaken after a malfunction occurs, sometimes even failing to detect the problem promptly and leading to greater risks. While RTO equipment is now equipped with DCS monitoring systems that monitor and record parameters for various critical components (such as combustion chamber temperature, pressure, and valve opening), triggering alarms when monitored parameters exceed thresholds, maintenance personnel must then analyze DCS parameter trends based on their experience, troubleshoot the fault location, and conduct on-site verification to pinpoint the cause of the abnormal parameter changes. However, this process requires extensive experience from maintenance personnel and takes considerable time to discern parameter trends, essentially constituting a reactive, post-fault troubleshooting approach.

[0005] Therefore, developing a predictive fault diagnosis system and method for RTOs that enables early warning before faults occur and allows computer-aided maintenance personnel to conduct fault inspections, reducing the experience requirements for personnel, is of great significance for achieving efficient and stable operation of RTOs, reducing major risks, and reducing emissions. Summary of the Invention

[0006] In order to overcome the shortcomings and deficiencies of the prior art, the present invention provides a fault prediction and diagnosis system and method for a regenerative thermal oxidation furnace.

[0007] This invention first uses the DCS system to monitor the upstream exhaust gas parameters of the RTO and the real-time operating parameters and historical operating parameters of each RTO point. It then uses a Long Short-Term Memory Neural Network (LSTM) to predict the changes in parameters at each point and establishes an abnormal parameter monitoring model to achieve real-time anomaly monitoring of the future trend of each parameter. Parameters that may be abnormal are input parameters in real time into the causal graph inference module. The causal graph is constructed using an expert knowledge base, graph recovery algorithm, and causal discovery algorithm. The simplification rules of the causal graph and Bayes' theorem are used to simplify and calculate the graph, outputting possible fault locations and probabilities, thus achieving predictive fault tracing.

[0008] To achieve the above objectives, the technical solution adopted by the present invention is as follows:

[0009] A fault prediction and diagnosis system for a regenerative thermal oxidizer includes a multi-parameter prediction module, an abnormal parameter monitoring module, and a cause-effect graph inference module.

[0010] The multi-parameter prediction module includes a multi-parameter prediction model based on the multi-parameters of the regenerative thermal oxidizer. The multi-parameter prediction model predicts the changes of each parameter in advance based on the DCS data acquisition system on the RTO, combined with historical operating data and online operating data.

[0011] The abnormal parameter monitoring module includes an abnormal parameter monitoring model when the RTO equipment may fail. It monitors and transmits abnormal parameters by comparing the abnormal score of the predicted parameters with a threshold.

[0012] The causal graph inference module includes establishing causal models of various faults and abnormal parameter changes through an artificial expert knowledge base, and a causal directed graph obtained by computer based on data graph recovery and causal discovery to assist in establishing a fault diagnosis causal graph; based on the input abnormal parameters and the causal probability relationship between each parameter and the fault, the fault diagnosis causal graph is simplified through a series of event operations and simplification principles, and the posterior probability is calculated through Bayes' formula, thereby outputting the fault location and probability.

[0013] Preferably, the regenerative thermal oxidizer includes multiple parameters such as the temperature, LEL concentration, gas pressure, and frequency of each fan in each chamber.

[0014] As a preferred method, predictive fault diagnosis is performed based on a multi-parameter prediction model and an abnormal parameter monitoring model.

[0015] The multi-parameter prediction model is derived from real-time and historical operating data collected by the DCS data acquisition system and trained by machine learning.

[0016] A hybrid model based on Long Short-Term Memory Neural Network (LSTM) and autoencoder is used to establish a multi-parameter prediction model from sequence to sequence. The input parameters of the multi-parameter prediction model include historical data of exhaust gas inlet temperature, LEL concentration, scrubber pressure difference, RTO chamber temperature, fan frequency, and air pressure, as well as DCS real-time monitoring data. The output is the predicted value of each parameter.

[0017] The network structure of the multi-parameter prediction model is an autoencoder structure based on ConvLSTM and BiGRU, consisting of an encoder and a decoder;

[0018] The encoder consists of a ConvLSTM layer, a tanh activation function layer, and a Dropout function layer stacked together. The ConvLSTM uses convolution operators instead of fully connected layer operators, enabling it to extract not only temporal features but also spatial information through convolution operations. Furthermore, it employs a gating mechanism to overcome the vanishing gradient problem. Its calculation formula is as follows:

[0019]

[0020]

[0021]

[0022]

[0023]

[0024] Among them, σ, tanh, × and These represent the Sigmoid, tanh, convolution, and Hadamard product, respectively; i t f t C t and o t These represent the input gate, forget gate, internal memory unit, and output gate in the current state, respectively; C t-1 X represents the internal memory unit in the previous state; t h represents the input data. t-1 h represents the loop state of the previous ConvLSTM unit output. t W represents the current loop state; xi The weight matrix from the input data to the input gate; W hi W represents the weight matrix from the output loop state to the input gate. ci W represents the weight matrix from the memory cell to the input gate. xf W represents the weight matrix from the input data to the forget gate. hf W represents the weight matrix from the output loop state to the forget gate. cfW represents the weight matrix from the memory unit to the forget gate. xc W represents the weight matrix from the input data to the memory unit. hc W represents the weight matrix from the output loop state to the memory unit. xo W represents the weight matrix from the input data to the output layer. ho W represents the weight matrix from the output loop state to the output layer. co This represents the weight matrix from the memory unit to the output layer; b i b represents the bias vector of the input gate; f b represents the bias vector of the forget gate; c b represents the bias vector of a memory cell; o This represents the bias vector of the output gate;

[0025] The decoder consists of a stacked BiGRU layer, a tanh activation function layer, and a Dropout function layer. BiGRU is a bidirectional recurrent neural network where the output layer contains both the forward and backward states for each point, allowing for the simultaneous extraction of past and future features, thus achieving bidirectional data reuse. Its calculation formula is as follows:

[0026] Z t =σ(W z [h t-1 X t ]+b z )

[0027] r t =σ(W r [h t-1 X t ]+b r )

[0028]

[0029]

[0030] Among them, Z t Indicates the update gate; r t Indicates that the door is being reset; Indicates a candidate state; W z W represents the weight matrix of the updated gate; r W represents the weight matrix of the reset gate; h The weight matrix representing the memory state; b z b represents the bias vector of the update gate; r b represents the bias vector of the reset gate; h A bias vector representing the memory state;

[0031] After the encoding layer, a repeating vector layer is used to copy the output vector of the encoding layer as the input vector of the decoding layer. After the decoding layer, a fully connected layer is used to output the predicted value sequence of each data.

[0032] The multi-parameter prediction module selects the mean squared error (MSE) as the loss function:

[0033]

[0034] Where Y i Y represents the predicted value at time i; i represents the true value at time i; n represents the total number of data samples; MSE represents the mean squared error.

[0035] As a preferred approach, the anomaly parameter monitoring module employs K-means clustering analysis combined with an autoencoder to establish an anomaly monitoring model. This model calculates the predicted parameter values ​​to determine whether they fall within the normal range. Since the autoencoder generates errors during the decoding and reconstruction of encoded features, and training minimizes these errors through backpropagation, learning the implicit features and patterns of normal data, values ​​falling within the normal range have smaller reconstruction errors. However, the reconstruction errors of anomalous data are larger. These reconstruction errors can be considered as anomaly scores. Clustering is used to assist in analyzing the reconstruction error ratio of the predicted data. If the anomaly score exceeds a set threshold, anomaly parameters are detected, and their related data are used as evidence parameters and passed to the causal graph inference module.

[0036] As a preferred approach, the causal graph inference module utilizes data mining, graph recovery algorithms, and causal discovery algorithms to perform full-graph modeling of historical operational data to construct a fault diagnosis causal graph. This is combined with expert experience to establish an expert knowledge base, which establishes the causal relationship between point faults and abnormal parameter changes. Subsequently, based on abnormal parameter-related data obtained from the abnormal parameter monitoring module, the fault diagnosis causal graph is simplified according to event logic expansion and simplification principles. Based on collected abnormal parameter evidence, probabilistic inference is performed on the root variables causing these outcome variable changes. Bayes' theorem is used to inversely calculate the posterior state probability caused by a point fault, and then the ranking probability is calculated. The module outputs each possible fault point and its probability, achieving predictive fault tracing.

[0037] This invention also provides a method for fault prediction and diagnosis of a regenerative thermal oxidizer, which, using the above-mentioned fault prediction and diagnosis system, includes the following steps:

[0038] S1: The key operational data of each point and parameter collected by the DCS data acquisition system and sensor array on the RTO device are transmitted to the cloud database using a network communication module;

[0039] S2: The prediction model is trained using historical operating data of various parameters collected by DCS, and the optimal prediction model is fitted. Based on historical operating data and data causal mining, an expert knowledge base is established, and a fault diagnosis causal graph is constructed accordingly.

[0040] S3: Based on the multi-parameter prediction model and the abnormal parameter monitoring module, the prediction parameters are monitored for anomalies. Once the parameters change, an anomaly will occur, and an early warning will be issued. The abnormal parameters and data will be passed to the causal graph inference module as evidence parameters.

[0041] S4: Based on the collected abnormal parameters, use the fault diagnosis cause-effect graph to expand the event logic, and simplify the fault diagnosis cause-effect graph according to the event operation and simplification rules.

[0042] S5: Calculate the posterior probability and ranking probability of each fault location using Bayes' theorem on the simplified cause-effect graph, and then output each possible fault location and its fault probability.

[0043] This invention trains a long short-term memory neural network using historical operational data of various parameters. It combines real-time online data monitored by the DCS monitoring system with the neural network to predict the changing trends of various parameters in advance over a period of time, monitors parameters that are likely to be abnormal, and inputs them as evidence parameters into the causal graph inference module. Through the fault diagnosis causal graph constructed by the expert knowledge base, it performs calculations and simplifications of events, and uses Bayes' theorem to calculate and output the possible fault locations and probabilities, guiding maintenance personnel to make adjustments in advance.

[0044] Compared with the prior art, the beneficial effects of the present invention are as follows:

[0045] This invention predicts the changing trends of parameters at various RTO points using a multi-parameter prediction model, establishes an abnormal parameter monitoring model to achieve real-time anomaly monitoring of the future changing trends of each parameter, and inputs parameters that may be abnormal into a graph theory inference algorithm in real time. Combining an expert knowledge base and a data-driven causal graph, the algorithm uses the simplification rules of the causal graph and Bayes' theorem to simplify and calculate the graph, outputting possible fault locations and probabilities, achieving predictive fault tracing, and guiding maintenance personnel to intervene in advance to avoid equipment shutdowns, effectively reducing pollutant gas emissions during shutdowns and the large amount of fuel required for restarting. Attached Figure Description

[0046] Figure 1 This is a flowchart of the present invention. Detailed Implementation

[0047] The technical solution of the present invention will be further described in detail below through embodiments. These embodiments are for illustrative purposes only and are not intended to limit the present invention. Based on the embodiments in this application, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of this application.

[0048] Reference Figure 1 A fault prediction and diagnosis system for a regenerative thermal oxidizer includes a multi-parameter prediction module, an abnormal parameter monitoring module, and a cause-effect graph inference module.

[0049] The multi-parameter prediction module includes a multi-parameter prediction model based on the temperature, air pressure, and frequency of each fan in each cavity of the regenerative thermal oxidizer. The multi-parameter prediction model predicts the changes of each parameter in advance based on the DCS data acquisition system on the RTO, combined with historical operating data and online operating data.

[0050] The abnormal parameter monitoring module includes an abnormal parameter monitoring model when the RTO equipment may fail. It monitors and transmits abnormal parameters by comparing the abnormal score of the predicted parameters with a threshold.

[0051] The causal graph inference module includes establishing causal models of various faults and abnormal parameter changes through an artificial expert knowledge base, and a causal directed graph obtained by computer based on data graph recovery and causal discovery to assist in establishing a fault diagnosis causal graph; based on the input abnormal parameters and the causal probability relationship between each parameter and the fault, the fault diagnosis causal graph is simplified through a series of event operations and simplification principles, and the posterior probability is calculated through Bayes' formula, thereby outputting the fault location and probability.

[0052] This embodiment describes the implementation of a waste gas treatment system for a pharmaceutical company. The waste gas originates from the pharmaceutical processing workshops producing vitamins, carotene, clarithromycin, and vancomycin hydrochloride formulations. During the production process, organic waste gases such as acetone, dichloromethane, xylene, and toluene are generated. The company's waste gas treatment RTO (Regenerative Thermal Oxidizer) equipment adopts a three-chamber structure with a designed air volume of 37,000 m³ / h. 3 The system is controlled directly by a PLC and uses methanol as fuel. It has online monitoring points for the temperature of the upper, middle, and lower layers of the heat storage medium, the combustion chamber temperature, the LEL concentration at the exhaust gas inlet, the exhaust gas inlet temperature, the exhaust gas inlet humidity, the exhaust gas inlet pressure, the exhaust gas inlet oxygen concentration, and the RTO outlet temperature. Gas flow sensors and VOCs concentration sensors can be added to the exhaust gas pipelines in various areas, low-concentration exhaust pipelines, and the main pipeline.

[0053] Faults at certain points in the company's RTO equipment can lead to abnormal parameters and cause the equipment to shut down. For example, a sudden increase in exhaust gas concentration, excessive fuel supply, severe blockage of the regenerator, malfunction of the hot bypass valve, or poor sealing or deformation of the outlet riser valve can cause the maximum outlet temperature to be higher than normal; a malfunction of the pressure reducing valve, excessively high pressure after the fuel pump, or a malfunction of the pressure switch can cause excessively high fuel oil pressure; a malfunction of the circulating water pump, blockage of the circulating pipeline filter, or blockage of the quenching nozzle can cause low flow rate of the quenching tower spray water, etc. All these abnormal parameter changes will result in a shutdown.

[0054] A method for fault prediction and diagnosis of a regenerative thermal oxidizer, employing the aforementioned fault prediction and diagnosis system, includes the following steps:

[0055] S1: The key operational data of each point and parameter collected by the DCS data acquisition system and sensor array on the RTO device are transmitted to the cloud database using a network communication module;

[0056] S2: The prediction model is trained using historical operating data of various parameters collected by DCS, and the optimal prediction model is fitted. Based on historical operating data and data causal mining, an expert knowledge base is established, and a fault diagnosis causal graph is constructed accordingly.

[0057] S3: Based on the multi-parameter prediction model and the abnormal parameter monitoring module, the prediction parameters are monitored for anomalies. Once the parameters change, an anomaly will occur, and an early warning will be issued. The abnormal parameters and data will be passed to the causal graph inference module as evidence parameters.

[0058] S4: Based on the collected abnormal parameters, use the fault diagnosis cause-effect graph to expand the event logic, and simplify the fault diagnosis cause-effect graph according to the event operation and simplification rules.

[0059] S5: Calculate the posterior probability and ranking probability of each fault location using Bayes' theorem on the simplified cause-effect graph, and then output each possible fault location and its fault probability.

[0060] As a preferred method, predictive fault diagnosis is performed based on a multi-parameter prediction model and an abnormal parameter monitoring model.

[0061] The multi-parameter prediction model is derived from real-time and historical operating data collected by the DCS data acquisition system and trained by machine learning.

[0062] A hybrid model based on Long Short-Term Memory Neural Network (LSTM) and autoencoder is used to establish a multi-parameter prediction model from sequence to sequence. The input parameters of the multi-parameter prediction model include historical data of exhaust gas inlet temperature, LEL concentration, scrubber pressure difference, RTO chamber temperature, fan frequency, and air pressure, as well as DCS real-time monitoring data. The output is the predicted value of each parameter.

[0063] The network structure of the multi-parameter prediction model is an autoencoder structure based on ConvLSTM and BiGRU, consisting of an encoder and a decoder;

[0064] The encoder consists of a ConvLSTM layer, a tanh activation function layer, and a Dropout function layer stacked together. The ConvLSTM uses convolution operators instead of fully connected layer operators, enabling it to extract not only temporal features but also spatial information through convolution operations. Furthermore, it employs a gating mechanism to overcome the vanishing gradient problem. Its calculation formula is as follows:

[0065]

[0066]

[0067]

[0068]

[0069]

[0070] Among them, σ, tanh, × and These represent the Sigmoid, tanh, convolution, and Hadamard product, respectively; i t f t C t and o t These represent the input gate, forget gate, internal memory unit, and output gate in the current state, respectively; C t-1 X represents the internal memory unit in the previous state; t h represents the input data. t-1 h represents the loop state of the previous ConvLSTM unit output. t W represents the current loop state. xi The weight matrix from the input data to the input gate; W hi W represents the weight matrix from the output loop state to the input gate. ci W represents the weight matrix from the memory cell to the input gate. xf W represents the weight matrix from the input data to the forget gate. hf W represents the weight matrix from the output loop state to the forget gate. cf W represents the weight matrix from the memory unit to the forget gate. xc W represents the weight matrix from the input data to the memory unit. hc W represents the weight matrix from the output loop state to the memory unit. xo W represents the weight matrix from the input data to the output layer. hoW represents the weight matrix from the output loop state to the output layer. co This represents the weight matrix from the memory unit to the output layer; b i b represents the bias vector of the input gate; f b represents the bias vector of the forget gate; c b represents the bias vector of a memory cell; o This represents the bias vector of the output gate.

[0071] The decoder consists of a stacked BiGRU layer, a tanh activation function layer, and a Dropout function layer. BiGRU is a bidirectional recurrent neural network where the output layer contains both the forward and backward states for each point, allowing for the simultaneous extraction of past and future features, thus achieving bidirectional data reuse. Its calculation formula is as follows:

[0072] Z t =σ(W z [h t-1 X t ]+b z )

[0073] r t =σ(W r [h t-1 X t ]+b r )

[0074]

[0075]

[0076] Among them, Z t Indicates the update gate; r t Indicates that the door is being reset; Indicates a candidate state; W z W represents the weight matrix of the updated gate; r W represents the weight matrix of the reset gate; h The weight matrix representing the memory state; b z b represents the bias vector of the update gate; r b represents the bias vector of the reset gate; h A bias vector representing the memory state;

[0077] After the encoding layer, a repeating vector layer is used to copy the output vector of the encoding layer as the input vector of the decoding layer. After the decoding layer, a fully connected layer is used to output the predicted value sequence of each data.

[0078] The multi-parameter prediction module selects the mean squared error (MSE) as the loss function:

[0079]

[0080] Where Y i Y represents the predicted value at time i; i represents the true value at time i; n represents the total number of data samples; MSE represents the mean squared error.

[0081] As a preferred approach, the anomaly parameter monitoring module employs K-means clustering analysis combined with an autoencoder to establish an anomaly monitoring model. This model calculates the predicted parameter values ​​to determine whether they fall within the normal range. Since the autoencoder generates errors during the decoding and reconstruction of encoded features, and training minimizes these errors through backpropagation, learning the implicit features and patterns of normal data, values ​​falling within the normal range have smaller reconstruction errors. However, the reconstruction errors of anomalous data are larger. These reconstruction errors can be considered as anomaly scores. Clustering is used to assist in analyzing the reconstruction error ratio of the predicted data. If the anomaly score exceeds a set threshold, anomaly parameters are detected, and their related data are used as evidence parameters and passed to the causal graph inference module.

[0082] As a preferred approach, the causal graph inference module utilizes data mining, graph recovery algorithms, and causal discovery algorithms to perform full-graph modeling of historical operational data to construct a fault diagnosis causal graph. This is combined with expert experience to establish an expert knowledge base, which establishes the causal relationship between point faults and abnormal parameter changes. Subsequently, based on abnormal parameter-related data obtained from the abnormal parameter monitoring module, the fault diagnosis causal graph is simplified according to event logic expansion and simplification principles. Based on collected abnormal parameter evidence, probabilistic inference is performed on the root variables causing these outcome variable changes. Bayes' theorem is used to inversely calculate the posterior state probability caused by a point fault, and then the ranking probability is calculated. The module outputs each possible fault point and its probability, achieving predictive fault tracing.

[0083] This invention trains a Long Short-Term Memory (LSTM) neural network using historical operational data of various parameters. It then uses real-time online data monitored by a DCS (Distributed Control System) to predict the future trends of each parameter, identifying parameters prone to anomalies and inputting these as evidence into a causal graph inference module. A fault diagnosis causal graph constructed from an expert knowledge base is used to perform event calculations and simplifications, and Bayes' theorem is used to calculate and output the possible fault locations and probabilities, guiding maintenance personnel to make proactive adjustments. Specifically, this invention first uses real-time monitoring data from the DCS system, including upstream exhaust gas parameters of the RTO (Regenerative Thermal Oxidizer) and real-time operational parameters at each RTO location, as well as historical operational parameters. Combined with a LSTM neural network to predict parameter changes at each location, an abnormal parameter monitoring model is established to achieve real-time anomaly monitoring of future parameter trends. Parameters potentially prone to anomalies are input into the causal graph inference module in real-time. The causal graph, constructed using an expert knowledge base, graph recovery algorithm, and causal discovery algorithm, is then simplified and calculated using causal graph simplification rules and Bayes' theorem to output possible fault locations and probabilities, achieving predictive fault tracing.

[0084] The present invention has been described in detail above with reference to the embodiments. However, the content described is only a specific implementation of the present invention and should not be construed as limiting the scope of the present invention. It should be noted that for those skilled in the art, any modifications and improvements made in accordance with the scope of the present invention without departing from the concept of the present invention should still fall within the patent coverage of the present invention.

Claims

1. A fault prediction and diagnosis system for a regenerative thermal oxidizer, characterized in that: It includes a multi-parameter prediction module, an anomaly parameter monitoring module, and a cause-effect graph inference module; The multi-parameter prediction module includes a multi-parameter prediction model based on the multi-parameters of the regenerative thermal oxidizer. The multi-parameter prediction model predicts the changes of each parameter in advance based on the DCS data acquisition system on the RTO, combined with historical operating data and online operating data. The abnormal parameter monitoring module includes an abnormal parameter monitoring model when the RTO equipment may fail. It monitors and transmits abnormal parameters by comparing the abnormal score of the predicted parameters with a threshold. The causal graph inference module includes establishing causal models of various faults and abnormal parameter changes through an expert knowledge base, and a causal directed graph obtained by computer based on data graph recovery and causal discovery to assist in establishing a fault diagnosis causal graph; based on the input abnormal parameters and the causal probability relationship between each parameter and the fault, the fault diagnosis causal graph is simplified through a series of event operations and simplification principles, and the posterior probability is calculated through Bayes' formula to output the fault location and probability; The regenerative thermal oxidizer has multiple parameters, including the temperature, LEL concentration, gas pressure, and frequency of each fan in each chamber. Predictive fault diagnosis is achieved based on multi-parameter prediction models and abnormal parameter monitoring models; The multi-parameter prediction model is derived from real-time and historical operating data collected by the DCS data acquisition system and trained by machine learning. A hybrid model based on long short-term memory neural network and autoencoder is used to establish a multi-parameter prediction model from sequence to sequence. The input parameters of the multi-parameter prediction model include historical data of exhaust gas inlet temperature, LEL concentration, scrubber pressure difference, RTO chamber temperature, fan frequency, and air pressure, as well as DCS real-time monitoring data. The output is the predicted value of each parameter. The network structure of the multi-parameter prediction model is an autoencoder structure based on ConvLSTM and BiGRU, consisting of an encoder and a decoder; The abnormal parameter monitoring module uses K-means clustering analysis combined with an autoencoder to establish an abnormal monitoring model, calculates the predicted parameter values, and determines whether they are within the normal range. By using clustering to assist in the analysis of the reconstruction error ratio of the predicted data, if the abnormal score exceeds the set threshold, the abnormal parameter is detected, and its related data is passed into the causal graph inference module as evidence parameters.

2. The fault prediction and diagnosis system for the regenerative thermal oxidizer according to claim 1, characterized in that: The encoder consists of a ConvLSTM layer, a tanh activation function layer, and a Dropout function layer stacked together. The ConvLSTM uses convolution operators instead of fully connected layer operators, and its calculation formula is as follows: ； in, σ tanh, × and These represent the Sigmoid, tanh, convolution, and Hadamard product, respectively. i t ，f t , C t and o t These represent the input gate, forget gate, internal memory unit, and output gate in the current state, respectively. C t-1 This represents the internal memory unit in the previous state; X t Indicates input data, h t-1 This indicates the loop state of the output of the previous ConvLSTM unit. h t This indicates the current loop state; W xi The table shows the weight matrix from the input data to the input gate; W hi This represents the weight matrix from the output loop state to the input gate; W ci This represents the weight matrix from the memory cell to the input gate; W xf This represents the weight matrix from the input data to the forget gate; W hf This represents the weight matrix from the output loop state to the forget gate; W cf The weight matrix representing the distance from the memory unit to the forget gate; W xc This represents the weight matrix from the input data to the memory unit; W hc The weight matrix representing the output loop state to the memory unit; W xo This represents the weight matrix from the input data to the output layer; W ho This represents the weight matrix from the output loop state to the output layer; W co This represents the weight matrix from the memory unit to the output layer; b i This represents the bias vector of the input gate; b f The bias vector representing the forget gate; b c The bias vector representing the memory cell; b o This represents the bias vector of the output gate; The decoder consists of a stacked BiGRU layer, a tanh activation function layer, and a Dropout function layer. BiGRU is a bidirectional recurrent neural network where the output layer contains both the forward and backward states for each point, enabling simultaneous extraction of past and future features, thus achieving bidirectional data reuse. Its calculation formula is as follows: ； in, Z t Indicates an update to the door; r t Indicates that the door is being reset; Indicates the candidate state; W z This represents the weight matrix of the updated gate; W r This represents the weight matrix of the reset gate; W h The weight matrix representing the memory state; b z This represents the bias vector of the updated gate; b r This represents the bias vector of the reset gate; b h A bias vector representing the memory state; After the encoding layer, a repeating vector layer is used to copy the output vector of the encoding layer as the input vector of the decoding layer. After the decoding layer, a fully connected layer is used to output the predicted value sequence of each data. The multi-parameter prediction module selects the mean squared error (MSE) as the loss function: ； in Indicates the first i The predicted value at any given time; Y i Indicates the first i The true value at time t; n represents the total number of data samples; MSE represents the mean squared error.

3. The fault prediction and diagnosis system for the regenerative thermal oxidizer according to claim 1, characterized in that: The causal graph inference module uses data mining, graph recovery algorithm and causal discovery algorithm to build a fault diagnosis causal graph by performing full graph modeling on historical operating data. It combines expert experience to help build an expert knowledge base, that is, the causal relationship between point faults and abnormal changes in a certain parameter is established through expert knowledge. Subsequently, based on the abnormal parameter data obtained by the abnormal parameter monitoring module, the cause-effect graph of fault diagnosis is simplified according to the event logic expansion and simplification principles. Based on the collected abnormal parameter evidence information, the root variables that cause these changes in outcome variables are probabilistically inferred. The Bayesian formula is used to inversely calculate the posterior state probability caused by a fault at a certain point, and then the ranking probability is calculated. Each possible fault point and its probability are output to achieve predictive fault tracing.

4. A method for fault prediction and diagnosis of a regenerative thermal oxidizer, characterized in that: The fault prediction and diagnosis system according to any one of claims 1-3 includes the following steps: S1: The key operational data of each point and parameter collected by the DCS data acquisition system and sensor array on the RTO device are transmitted to the cloud database using a network communication module; S2: The prediction model is trained using historical operating data of various parameters collected by DCS, and the optimal prediction model is fitted. Based on historical operating data and data causal mining, an expert knowledge base is established, and a fault diagnosis causal graph is constructed accordingly. S3: Based on the multi-parameter prediction model and the abnormal parameter monitoring module, the prediction parameters are monitored for anomalies. Once the parameters change, an anomaly will occur, and an early warning will be issued. The abnormal parameters and data will be passed to the causal graph inference module as evidence parameters. S4: Based on the collected abnormal parameters, use the fault diagnosis cause-effect graph to expand the event logic, and simplify the fault diagnosis cause-effect graph according to the event operation and simplification rules. S5: Calculate the posterior probability and ranking probability of each fault location using Bayes' theorem on the simplified cause-effect graph, and then output each possible fault location and its fault probability.

Images