Fluidized bed caking and attrition fault parallel prediction method and device based on gatt-cnn-lstm

Abstract

The application discloses a kind of based on GATT-CNN-LSTM's fluidized bed agglomeration and wear fault parallel prediction method and device, the method steps include: step S01, temperature and mass flow rate at different monitoring positions are collected to form fault data set;Step S02, according to fault data set respectively extract change rate to obtain fault feature;Step S03, according to each fault feature sequence, calculate fault degree index value, to judge corresponding fault degree, and according to the contribution state of each parameter, judge fault category, configure fault degree and fault type label;Step S04, using the fault feature data set with label trains fault prediction classification model;Step S05, extract real-time fault feature sequence, input into trained fault prediction classification model, obtain fault degree and fault type result.The application can realize the identification of agglomeration fault and wear fault in fluidized bed in parallel, and also can realize fault degree prediction.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of coal power equipment fault diagnosis, and particularly relates to a fluidized bed coking and wear fault parallel prediction method and device based on GATT-CNN-LSTM. BACKGROUND

[0002] The fluidized bed can realize efficient and stable combustion and material processing of solid fuels such as coal due to its unique fluidized state, and is widely used in the fields of power generation, chemical industry, metallurgy and the like. The fluidized bed is complex in the running process and is affected by multiple factors. On the one hand, the fluid mechanics characteristics in the fluidized bed are complex and changeable, and the interaction of gas-solid two-phase flow is difficult to accurately predict and control. For example, abnormal fluidization phenomena such as channeling and slugging may occur in the fluidization process, which can seriously affect the uniformity and stability of combustion or reaction. On the other hand, the changes of fuel properties, particle size distribution, fluidized wind speed, temperature and many other process parameters will interfere with the running state of the fluidized bed. The combined action of these complex factors can cause various faults in the running process of the fluidized bed, such as local coking, flameout, over-temperature and the like, which can not only cause equipment damage and production interruption, but also can cause serious safety hazards and environmental pollution problems. Therefore, it is particularly necessary to diagnose faults in time and accurately during the running process of the fluidized bed, which can not only ensure the stable operation of the fluidized bed equipment, improve the production efficiency and reduce the economic loss, but also is beneficial to environmental protection and safety production.

[0003] At present, there are mainly the following ways for the fault diagnosis of the fluidized bed:

[0004] 1. Fault diagnosis based on sensor monitoring data: sensors such as temperature sensors, vibration sensors, pressure sensors or acoustic sensors are installed at some key positions of the fluidized bed, and whether there is a fault is judged according to the data monitored by the sensors. This kind of method can only simply identify the over-limit state of the running parameters of the key positions, cannot identify the specific fault type, and cannot determine the fault degree, so it often can only monitor the parameter abnormality after a serious fault occurs, and cannot give an early warning in the early stage of the fault.

[0005] 2. Fault diagnosis based on physical model: a physical model is established based on the physical characteristics, chemical reaction mechanism and fluid mechanics principle of the fluidized bed, a fault mode library is established based on the physical model and historical data, and during the running process of the fluidized bed, the measured data are acquired in real time and compared with the predicted values of the model to identify the fault type. However, the physical and chemical processes of the fluidized bed itself are complex, and the internal flow, heat transfer, mass transfer and chemical reaction processes are interwoven, which are difficult to accurately describe by a simple physical model, so the actual prediction accuracy is not high.

[0006] 3. Fault diagnosis based on machine learning algorithms: Use machine learning algorithms (such as support vector machines, neural networks, random forests, etc.) to analyze and model the operation data of the fluidized bed, and realize the classification and prediction of faults.

[0007] However, on the one hand, the data measured by the sensor often has the characteristics of high dimension, nonlinearity and strong noise, different fault types may show similar characteristics on the sensor signal, and the same fault type may show different characteristic patterns due to changes in working conditions, and different fault types have different sensitivities to different types of sensors. The difficulty of feature extraction and fault classification is great, the fault classification accuracy of traditional machine learning algorithms is not high, and the fluidized bed system has great time-varying and uncertainty, its operating conditions may change significantly over time, raw material batches, etc., resulting in that the established fault diagnosis model is difficult to adapt to new working conditions, and needs to be frequently updated and optimized.

[0008] On the other hand, the existing technology based on machine learning algorithms can only realize the classification of a single fault type formed by the same mechanism, such as one of clogging fault or blockage fault, and cannot realize the classification of fault types formed by different mechanisms. Clogging and wear faults belong to faults formed by different mechanisms, and the existing technology cannot effectively classify them. The generation of faults in the fluidized bed is often not independent, and if only a single fault of a certain type is monitored, the risk of coordinated deterioration between different faults may be easily ignored. The traditional machine learning algorithm can only realize simple fault classification, and cannot judge the degree of fault occurrence. Especially, some early faults may not be obvious in the original sensor data set, and the weak abnormal signal caused by early faults is easily covered by background noise during normal operation and coupling fluctuations between parameters, making it difficult to accurately identify the early fault through simple operation monitoring parameters. SUMMARY

[0009] The technical problem to be solved by the present application is that in view of the technical problems existing in the prior art, the present application provides a GATT-CNN-LSTM-based fluidized bed clogging and wear fault parallel prediction method and device, which has the advantages of simple implementation method, low cost, high fault diagnosis efficiency and precision, and can realize the identification and classification of clogging faults and wear faults in the gasifier and combustion furnace of the fluidized bed in parallel, and can also realize fault degree prediction in parallel.

[0010] To solve the above technical problems, the technical solution provided by the present application is:

[0011] A GATT-CNN-LSTM-based fluidized bed clogging and wear fault parallel prediction method, comprising the following steps:

[0012] Step S01, fault data collection: collect state parameters of the gasifier and the combustor at different monitoring positions under different fault degree conditions of the hot-state double-circulation fluidized bed system to form a fault data set, the state parameters including temperature data and mass flow rate;

[0013] Step S02, fault feature extraction: extract variation values and deviation from mean values of the state parameters of each monitoring position at different measurement time points from the fault data set to obtain fault features, and obtain a group of fault feature sequences at each measurement time point;

[0014] Step S03, fault data classification: calculate a fault degree index value corresponding to each measurement time point according to the fault feature sequence, judge whether it is in a fault condition and judge the fault degree and the fault category according to the fault degree index value, configure a fault degree label and a fault category label corresponding to each fault feature sequence according to the judgment results of the fault degree and the fault category, and obtain a fault feature data set with labels;

[0015] Step S04, fault prediction model training: train a pre-constructed GATT-CNN-LSTM fault prediction classification model using the fault feature data set with labels;

[0016] Step S05, real-time fault parallel prediction: collect state parameters of the gasifier and the combustor at different positions in the running process of the measured fluidized bed system in real time, extract corresponding variation values and deviation from mean values to obtain real-time fault feature sequences, input them into the trained GATT-CNN-LSTM fault prediction classification model, and obtain fault degree prediction results and fault type classification results output.

[0017] Further, in step S01, the fault degree conditions include normal operation conditions, early fault conditions and severe fault conditions, the collected temperature data include cross-section temperatures at multiple different heights from the combustor wind distribution plate in the gasifier, temperature data changing over time in the horizontal flue of the gasifier, and temperature data changing over time in the horizontal flue of the combustor, the collected mass flow rates include mass flow rates changing over time in the horizontal flue of the combustor and mass flow rates changing over time in the horizontal flue of the gasifier, the early fault conditions include early heating surface wear faults of the gasifier and early coking faults of the combustor, and the severe fault conditions include severe heating surface wear faults of the gasifier and severe coking faults of the combustor.

[0018] Further, in step S03, calculating a fault degree index value corresponding to the fault feature sequence includes:

[0019] The fault degree index is calculated according to the temperature variation parameters of the gasifier and the combustion furnace at different positions in the fault feature sequence, the mass flow rate variation rates, and corresponding deviation mean degree values R The temperature variation parameters include temperature variation acceleration absolute values, and the deviation mean degree values include temperature deviation mean degree values and mass flow rate deviation mean degree values R The calculation model of the fault degree index includes a coking fault calculation item, a wear fault calculation item, and a cross feedback calculation item, the coking fault calculation item is calculated according to the temperature variation parameters and corresponding deviation mean degree values, the wear fault calculation item is calculated according to the mass flow rate variation rates and corresponding deviation mean degree values, and the cross feedback calculation item is calculated according to total temperature variation parameters of each position and total mass flow rate deviation mean degree values R The calculation model of the fault degree index is as follows:

[0020]

[0021]

[0022] Wherein, R is a fault degree index value, , are temperature variation acceleration absolute values of two different positions of the gasifier away from the air distribution plate of the combustion furnace, , are temperature variation acceleration absolute values of the horizontal flue position of the gasifier and the horizontal flue position of the combustion furnace, , are mass flow rate variation rates of the horizontal flue position of the gasifier and the horizontal flue position of the combustion furnace, ~ are temperature deviation mean degrees of the two different positions of the gasifier away from the air distribution plate of the combustion furnace, the horizontal flue position of the gasifier, and the horizontal flue position of the combustion furnace, , are mass flow rate deviation mean degrees of the horizontal flue position of the gasifier and the horizontal flue position of the combustion furnace, TD is a total temperature deviation mean degree of each position, TC is a total temperature variation rate of each position, FC is a total mass flow rate variation rate of the horizontal flue of the gasifier and the horizontal flue position of the combustion furnace, FD is a total mass flow rate deviation mean degree of the horizontal flue position of the gasifier and the horizontal flue position of the combustion furnace, , are respectively global basic weight coefficients, gamma The weights are used to reflect the mutual aggravation effect of different features on quantitative coking and wear.

[0023] Further, in step S03, determining whether the system is in a fault condition and determining the degree of fault based on the fault severity index value includes: if the fault severity index value R is less than a preset early fault threshold, it is determined to be a normal operating condition, and the data tag is configured as normal; if the fault severity index value R is greater than the preset early fault threshold and less than the preset severe fault threshold, it is determined to be an early fault condition, and the corresponding data tag is configured as early fault; if the fault severity index value R is greater than the preset severe fault threshold, it is determined to be a severe fault condition, and the corresponding data tag is configured as severe fault.

[0024] Furthermore, in step S03, when a fault condition is determined, the fault category is determined based on the contribution status of each parameter during the calculation of the fault severity index value, including:

[0025] Step S321. Determine the impact of the agglomeration fault calculation item and the wear fault calculation item on the fault severity index value. R The contribution ratios are calculated as follows: if the contribution ratio corresponding to the clumping fault calculation item is greater than the preset contribution ratio threshold, it is determined to be a clumping fault-dominated mode, and the fault label is set as clumping fault. If the contribution ratio corresponding to the wear fault calculation item is greater than the preset contribution ratio threshold, it is determined to be a wear fault-dominated mode, and the fault label is set as wear fault. If the contribution ratios corresponding to the clumping fault calculation item and the wear fault calculation item are equal, the process proceeds to step S322 for secondary judgment, where the temperature parameter includes the absolute value of temperature change acceleration. TC and the degree of temperature deviation from the mean TD ;

[0026] Step S322. Determine the fault severity index value R The proportion of cross-feedback terms in the calculation model, if the cross-feedback terms affect the fault severity index value R If the proportion is greater than the preset dual-fault judgment threshold, it is judged as a dual fault of agglomeration and wear and an early warning is triggered; if the cross-feedback item affects the fault severity index value R If the proportion is less than the preset single fault judgment threshold, it is judged as a single fault. The fault type is determined according to the larger contribution ratio of the clumping fault calculation item and the wear fault calculation item.

[0027] Furthermore, when the fault is determined to be a dual fault of agglomeration and wear, the weights in the cross-feedback term are dynamically adjusted based on the total temperature change rate and the degree of deviation of the total mass flow rate from the mean at each location. gamma The calculation expression is:

[0028]

[0029] in, Indicates to Adjusted weights, RTC R is the total temperature change rate for each position FD R is the total mass flow rate deviation from the mean for the gasifier horizontal flue position, the combustion furnace horizontal flue position R is the total temperature change rate R is the maximum value of the total temperature change rate R is the maximum value of the total mass flow rate deviation from the mean R is the maximum value of the total mass flow rate deviation from the mean

[0030] Further, the GATT-CNN-LSTM fault prediction classification model comprises a feature extraction module, a fault classification branch and a fault degree prediction branch, the feature extraction module comprises a Gaussian convolution kernel time sequence attention layer GATT, a CNN layer and an LSTM layer connected in sequence, the Gaussian convolution kernel time sequence attention layer GATT in the GATT-CNN-LSTM fault prediction classification model adopts a Gaussian convolution model as follows:

[0031]

[0032] wherein, is a weight coefficient in a calculation model of a fault degree index R , , σ is a standard deviation of a Gaussian distribution for controlling a smoothing degree;

[0033] The output of the Gaussian convolution kernel time sequence attention layer GATT is represented as follows:

[0034]

[0035]

[0036] wherein, is the output of the Gaussian kernel convolution layer, is a batch index, is a time step index, is a feature index, is a convolution kernel size, is a kernel center position, is a normalized Gaussian convolution kernel weight, and are learnable parameters of the feature .

[0037] Further, the fault classification branch includes at least two CNN layers and one fully connected layer, the features output by the feature extraction module are received by the input end of the CNN layer, and the fault type classification result is output after the features pass through each CNN layer and the fully connected layer; the fault degree prediction branch includes at least two LSTM layers and one fully connected layer, the features output by the feature extraction module are received by the input end of the LSTM layer, and the fault degree prediction result is output after the features pass through each LSTM layer and the fully connected layer.

[0038] A computer device includes a processor and a memory, the memory is used to store a computer program, and the processor is used to execute the computer program to perform the method as described above.

[0039] A computer readable storage medium storing a computer program, the computer program is executed by a processor to implement the method as described above.

[0040] Compared with the prior art, the beneficial effects of the present application are that:

[0041] 1、The present application forms a fault data set by collecting temperature data and mass flow rate of the gasification furnace and the combustion furnace at different monitoring positions under different fault degree conditions, extracts temperature change parameters, mass flow rate change rates and deviation from the mean degree as fault features, calculates a fault degree index value to judge the fault degree, and further identifies the fault type according to the contribution state of each parameter in the process of calculating the fault degree index value when it is judged to be in a fault condition, so that the correlation between different feature agglomeration faults and wear faults and between different features and different degree faults can be fully mined, and then a fault prediction classification model is trained by using the fault feature data set with two labels, so that the fluidized bed agglomeration and wear faults can be predicted in parallel, the fault degree can be predicted, the parallel prediction of fault type identification and fault degree prediction is realized, and thus the corresponding maintenance decision can be made for different fault degrees of different fault types, and the maintenance cost is reduced.

[0042] 2、In the present application, the GATT-CNN-LSTM architecture is adopted to construct a fault prediction classification model, the Gaussian convolution kernel time sequence attention layer GATT, the CNN layer and the LSTM layer are used for feature extraction, the requirements of shared feature extraction and classification and prediction tasks are retained, the advantages of the convolutional neural network in data feature extraction can be fully utilized, and the advantages of the recurrent neural network in time sequence feature capture can be utilized, so that the overall predictability of the model is improved. BRIEF DESCRIPTION OF DRAWINGS

[0043] Figure 1 is the implementation process schematic diagram of the fluidized bed agglomeration and wear fault parallel prediction method based on GATT-CNN-LSTM of the present embodiment.

[0044] Figure 2 Fig. 1 is a schematic diagram of the principle of sensor arrangement in the fluidized bed of the present embodiment.

[0045] Figure 3 Fig. 2 is a schematic diagram of the interaction between the wear failure and the agglomeration failure in the fluidized bed.

[0046] Figure 4 Fig. 3 is a schematic diagram of the process of failure type classification in the present embodiment.

[0047] Figure 5 Fig. 4 is a schematic diagram of the structure of the GATT-CNN-LSTM model in the present embodiment.

[0048] Figure 6 Fig. 5 is a schematic diagram of the model parameter configuration effect of the GATT-CNN-LSTM model in the specific application embodiment of the present application. DETAILED DESCRIPTION

[0049] The present application will be further described in conjunction with the accompanying drawings and specific preferred embodiments, but the protection scope of the present application is not limited thereby.

[0050] For the convenience of understanding, first, the related technical background involved in the present application is exemplarily introduced.

[0051] The internal structure of the fluidized bed is a multiphase flow system, which is relatively complex, and the internal state is also relatively complicated. Often, one failure will trigger a chain reaction and simultaneously trigger the abnormality of multiple parameters. For example, in the fluidized bed boiler, agglomeration will lead to: local high temperature → accelerated oxidation of the heating surface → accelerated peeling of the oxidation layer and accelerated wear → wear promotes agglomeration. Therefore, the occurrence of failure in the fluidized bed is often not independent. If only a single failure of a certain type is monitored, the risk of synergistic deterioration between different failures will be easily ignored. Moreover, the failure characterization of different failure types is often different. For a certain type of failure, multiple sensors may be needed for monitoring. However, when the failure occurs, the data change of some parameters may not be obvious, which leads to the fact that some sensors cannot monitor the abnormal state. Therefore, the sensitivity of different failure types to each type of sensor is also different. For example, for agglomeration failure, the main manifestation is temperature change. Therefore, in addition to temperature sensors, other types of sensors such as vibration and acoustic sensors may not be able to detect abnormal states. For wear failure, compared with agglomeration failure, the temperature change is smaller, and the mass flow rate change is larger. Therefore, a mass flow meter needs to be used for monitoring. In the gasification furnace and the combustion furnace, agglomeration failure and wear failure may exist simultaneously. Through the monitoring of a single type of sensor, the failure state in the gasification furnace and the combustion furnace cannot be comprehensively reflected. If multiple types of sensors are used, the sensitivity of the data monitored by different sensors to different failure types is different, and it is still difficult to determine the specific failure type and the failure degree.

[0052] Part of the early failure may not be obvious in the original sensor data set, on the one hand, the early failure tends to be hidden, its development degree is relatively slight, the disturbance to the physical and chemical process and the hydrodynamic state in the fluidized bed is small, it is difficult to form significant characteristic changes in the original data such as temperature, pressure, wind speed and other conventional monitoring parameters; on the other hand, the fluidized bed system itself is highly complex, each parameter is related and interferes with each other, the weak abnormal signal caused by early failure is easily covered by the background noise during normal operation and the coupling fluctuation between parameters, so that these fault features are difficult to be directly identified and extracted in the unprocessed original data set, and the limitation of the accuracy, resolution and data acquisition frequency of the sensor further aggravates the unobviousness of the early failure in the original sensor data set, so it is difficult to identify the early failure directly based on the original sensor data.

[0053] Fluidized bed agglomeration failure and heating surface wear are two common failure types of circulating fluidized bed, and fluidized bed agglomeration failure and heating surface wear are a mutually influencing and related process, both of which will induce each other in terms of cause and promote each other in development process. Fluidized bed agglomeration failure will destroy the uniform fluidization state of the bed material, form a high-speed airflow channel or jet zone around or above it, and the hard bed material particles (such as quartz sand, limestone, and ash particles) carried by high-speed airflow will strongly scour the nearby heating surface, increase the frequency and energy of particle impact on the heating surface, and in the agglomeration process, mineral particles (such as quartz) with higher hardness may be wrapped or adhered, when the agglomerate breaks or peels off, these hard particles are released and scoured the heating surface with the airflow, aggravating the wear, and the wear process will produce a large amount of fine metal particles and ash particles, these fine particles have larger specific surface area and stronger adhesion, fine particles are more likely to react with alkali metal, chlorine, sulfur and other low melting point ash components at relatively low temperature to form a viscous low-temperature eutectic, which becomes the "glue" of agglomeration, and continuous wear leads to the decrease of the average particle size of the bed material and the increase of the proportion of fine particles, which in turn facilitates the contact and adhesion between particles, thereby increasing the risk of agglomeration. The agglomerates cover on the heating surface or its vicinity, which will seriously hinder the heat transfer efficiency of the local area, leading to the increase of local flue gas temperature and wall temperature. That is, the formation of heating surface wear failure and agglomeration failure belongs to a gradual development process, and the symptoms presented in the initial stage of failure are generally not obvious, so such a gradual change is sometimes difficult to intuitively reflect through the operation monitoring parameters.

[0054] The application considers the interaction between the agglomeration and the wear failure and the risk of synergistic deterioration between the two failures, forms a fault data set by collecting the temperature data of the gasifier and the combustion furnace at different monitoring positions and the mass flow rate under different failure degree conditions (normal, early failure, severe failure) through simultaneous parallel prediction of the fluidized bed agglomeration and the wear failure, extracts the temperature change parameters, the mass flow rate change rate and the deviation from the mean value as fault features to form a fault feature sequence at different time points, calculates the failure degree index value of the fault feature sequence at each time point to judge the failure degree, and further identifies the failure type according to the contribution state of each parameter in the process of calculating the failure degree index value when it is judged to be in a failure condition, and then configures the failure degree label and the failure category label accordingly. The correlation between different characteristic agglomeration failures, wear failures and different characteristic and different degree failures can be fully mined, and then the fault prediction classification model is trained using the fault feature data set with two labels. Not only can the fault be classified, but also the failure degree can be predicted, realizing parallel prediction of fault type identification and failure degree prediction, so as to facilitate making corresponding maintenance decisions for different failure degrees of fluidized bed agglomeration failure and wear failure, reducing maintenance cost. At the same time, the fault prediction classification model adopts GATT-CNN-LSTM architecture, uses Gaussian convolution kernel time series attention layer GATT, CNN layer and LSTM layer for feature extraction, and through sharing feature extraction and retaining the needs of different tasks (classification and prediction tasks), the advantages of convolutional neural network in data feature extraction can be fully utilized, and the advantages of recurrent neural network in time series feature capture can be utilized, thereby improving the overall predictability of the model.

[0055] The application will be further described below in combination with specific embodiments.

[0056] As shown in Figure 1 , the steps of the fluidized bed agglomeration and wear failure parallel prediction method based on GATT-CNN-LSTM in the embodiment include:

[0057] Step S01, fault data acquisition: collect the state parameters of the gasifier and the combustion furnace at different monitoring positions under different failure degree conditions of the hot-state double-circulation fluidized bed system to form a fault data set, and the state parameters include temperature data and mass flow rate.

[0058] Boiler agglomeration can affect the heat transfer efficiency of the boiler, resulting in reduced thermal efficiency of the boiler, and can also cause leakage and separation. Agglomeration usually occurs in areas where gas distribution is uneven, temperature gradient is large, or material residence time is long, for example, it usually occurs at 1-5 meters above the connection between the water-cooled wall and the wear-resistant material and at the turning place of the flue gas at the furnace outlet. The temperature data collected in this embodiment can include cross-section temperature at different heights from the air distribution plate of the combustion furnace in the gasification furnace, temperature data changing over time in the horizontal flue of the gasification furnace, temperature data changing over time in the horizontal flue of the combustion furnace, and the like. The mass flow rate collected includes mass flow rate changing over time in the horizontal flue of the combustion furnace and mass flow rate changing over time in the horizontal flue of the gasification furnace, and the like. Early failure conditions include early heating surface wear failure of the gasification furnace and early agglomeration failure of the combustion furnace, and severe failure conditions include severe heating surface wear failure of the gasification furnace and severe agglomeration failure of the combustion furnace. In this embodiment, the state parameters of the gasification furnace and the combustion furnace at different monitoring positions are collected when the hot-state double-circulation fluidized bed system is in normal operation conditions, early failure conditions, and severe failure conditions.

[0059] As shown in Figure 2 , considering that agglomeration usually occurs in areas where gas distribution is uneven, temperature gradient is large, or material residence time is long, and the particle concentration in the dense phase zone of the fluidized bed combustion furnace is high, if the local temperature is too high (such as the high-temperature zone of the combustion furnace), the ash or low-melting-point substance will melt and bond the particles to form agglomeration. In this embodiment, a temperature sensor is arranged at a height of 3200 mm from the air distribution plate (cross-section modeling reference) in the gasification furnace, corresponding to position ①, to correlate with high-temperature coking failure. Considering that the gas distribution plate is the core component of the fluidized bed, responsible for uniformly distributing gas to the bed, if the distribution plate is poorly designed (such as uneven opening rate) or blocked, it can cause local gas velocity to be too low, forming a "dead zone", where particles deposit and stick together due to high temperature or the presence of sticky substances, which can trigger low-temperature coking of the gasification furnace. In this embodiment, a temperature sensor is arranged at a cross-section at a distance of 1450 mm from the air distribution plate in the combustion furnace, corresponding to position ②, to correlate with agglomeration failure. Considering that the particle concentration in the dilute phase zone is low, but if fine particles carried by the gas settle here due to reduced flow rate, or adhere to the reactor wall after collision, soft agglomeration may be formed, i.e., horizontal flue heating surface wear failure, including the combustion furnace and the gasification furnace. In this embodiment, mass flow rate sensors and temperature sensors are arranged in the horizontal flues of the combustion furnace and the gasification furnace, respectively, corresponding to positions ③ and ④, to correlate with failure of the horizontal flue heating surface.

[0060] In a specific application embodiment, for the fault classification and fault degree prediction of a hot-state double-circulation fluidized bed, the following data is collected at different measurement time points to form a data set, with the combustion furnace 1450 mm, the combustion furnace 3200 mm, the combustion furnace horizontal flue, and the gasifier horizontal flue as the key monitoring positions, as shown in Table 1: a gasifier section temperature data set f-g-3200-t at a height of 3200 mm from the combustion furnace air distribution plate (cross-section modeling reference) in the hot-state full cycle of a 1 MW double-circulation fluidized bed system; a combustion furnace section average temperature data set f-c-1450-t at a distance of 1450 mm from the air distribution plate in the hot-state full cycle of a 1 MW double-circulation fluidized bed system; a gasifier horizontal flue temperature data set f-g-out-t varying with time in the hot-state full cycle of a 1 MW double-circulation fluidized bed system; a combustion furnace horizontal flue temperature data set f-c-out-t varying with time in the hot-state full cycle of a 1 MW double-circulation fluidized bed system; a combustion furnace horizontal flue mass flow rate data set f-c-out varying with time in the hot-state full cycle of a double-circulation fluidized bed system; and a gasifier horizontal flue mass flow rate data set f-g-out varying with time in the result of the hot-state full cycle of a double-circulation fluidized bed system.

[0061] Table 1: Basic characteristics of data sets

[0062]

[0063] Step S02, fault feature extraction: according to the fault data set, the change value and the deviation from the mean value of the state parameters of each monitoring position at different measurement time points are extracted to obtain the fault features, wherein the fault features corresponding to each monitoring position at the same measurement time point constitute a group of fault feature sequences, and one group of fault feature sequences is obtained at each measurement time point.

[0064] Temperature and mass flow rate can effectively reflect the boiler combustion condition and heat transfer state, and are closely related to the coking phenomenon. Abnormal and too fast temperature rising or falling rate can indicate that there may be sudden deterioration of local heat transfer efficiency, which can mark the amplification of local thermal imbalance. And the wear failure is directly related to the impact frequency and particle kinetic energy and temperature: the higher the temperature → the number of particles impacting the heated surface per unit time increases → the number of particles impacting the heated surface per unit time rises → the gas flow rate increases → the particle impact speed increases → the single particle kinetic energy significantly increases, and the wear is more serious. In order to reflect the gradual development process of the coking failure and the wear failure, the temperature change parameter in this embodiment adopts the absolute value of temperature change acceleration, which is obtained by extracting the absolute value of temperature change acceleration TC and the mass flow rate change rate FC to obtain the temperature and mass flow rate change trend at each monitoring position in the gasifier and the combustion furnace. The absolute value of temperature change acceleration TCIt can indicate a sudden deterioration of local heat transfer efficiency, while wear is usually a gradual, slow process, which results in a generally smaller and relatively stable mass flow rate change rate. By detecting a continuous, small mass flow rate change rate at a steady state operating condition, it can be used as an early signal of wear expansion.

[0065] Specifically, the absolute value of the temperature change acceleration TC The calculation expression is:

[0066] (1)

[0067] Where: Δt is the time difference between the initial and final time, represents the temperature of the jth monitoring position at the ith measurement time point.

[0068] The mass flow rate change rate can represent the wear rate, and because it simultaneously affects the impact frequency and particle velocity, it is nonlinearly amplified. In this embodiment, the mass flow rate change rate is the mass of fluid or particles passing through a certain section within a certain period of time, and the calculation expression is:

[0069] (2)

[0070] Where: FC is the mass flow rate change rate, m i,j is the mass flow rate of the jth monitoring position at the ith strategy time point. j i

[0071] In this embodiment, the deviation from the mean degree value includes a temperature deviation from the mean degree value and a mass flow rate deviation from the mean degree value, which can represent the severity of the fault. The greater the deviation degree value, the higher the severity of the fault. For example, temperature deviation from the mean degree and mean deviation from too much data can indicate that the operating state of the fluidized bed is abnormal, which can indicate an early caking fault, and the mass flow rate deviation from the mean degree can indicate a wear fault.

[0072] Specifically, the calculation expression of the temperature deviation from the mean degree is:

[0073] (3)

[0074] Where, represents the temperature of the jth monitoring position at the ith measurement time point, M represents the number of monitoring positions, N represents the total number of measurement time points, represents the average value of the temperature of all measurement time points.

[0075] The calculation expression of the mass flow rate deviation from the mean degree is:

[0076] ​​ (4)

[0077] wherein, represents the mass flow rate of the jth monitoring position at the ith measurement time point, M represents the number of monitoring positions, N represents the total number of measurement time points, represents the average value of the mass flow rate of all measurement time points.

[0078] The embodiment identifies the main associated sensors with high sensitivity to specific fault types by analyzing the structure of the fluidized bed, and after arranging corresponding sensors at the gasifier, combustion furnaces ①, ②, ③, ④ positions, continuously monitors the absolute value of the temperature change acceleration TC , the temperature deviation from the mean value TD , the mass flow rate change rate FC , and the mass flow rate deviation from the mean value FD , forms the fault characteristics, as shown in Table 2, and further based on the sensor data, continuously monitors the development process of the gasifier, combustion furnace agglomeration fault and heating surface wear fault, to identify the fault degree of the gasifier, combustion furnace agglomeration fault and heating surface wear fault.

[0079] Table 2: Fault characteristic table

[0080]

[0081] Step S03, fault data classification: according to the fault characteristic sequence of each measurement time point, calculate the corresponding fault degree index value, according to the fault degree index value, judge whether it is in fault working condition and judge the fault degree and fault category, according to the judgment result of fault degree and fault category, configure the fault degree label and fault category label corresponding to each fault characteristic sequence, get the fault characteristic data set with label.

[0082] When agglomeration or heating surface wear fault occurs, the absolute value of temperature change acceleration, mass flow rate characteristics may change abnormally, and temperature and mass flow rate may also deviate significantly from the normal mean value, so by combining the absolute value of temperature change acceleration, mass flow rate and deviation from the mean value, the correlation between different sensors and different fault types can be fully explored. And the wear fault and the agglomeration fault interact with each other, as shown in Figure 3 , can be divided into the following three action chains:

[0083] Triggering of agglomeration fault (starting point): local agglomeration → heat transfer obstruction → TC abnormal rise (local temperature sudden rise); uneven temperature distribution → TD continuous increase (long-term deviation from the mean value) → FD abnormal;

[0084] Causes of wear failure: abnormal temperature ( TC / TD ↑) → Increased thermal stress on the heated surface → Material fatigue → Accelerated wear; expanded wear area → Turbulent flow within the flue → Abnormal fluctuations in FC (Fluid flow) → TC abnormal;

[0085] Regeneration of new clumps: mass flow rate anomaly ( FC / FD ↑) → Uneven particle flow → Localized material accumulation; High-temperature sintering in the accumulation area → New agglomerates form → Feedback to stage 1.

[0086] Based on the above characteristics, this embodiment uses a weighted summation of the absolute value of temperature change acceleration, the degree of temperature deviation from the mean, the rate of change of mass flow rate, and the degree of mass flow rate deviation from the mean to form a fault severity index value. R Based on the fault severity index value R Determining the degree of a fault can accurately characterize its extent and status during the development process.

[0087] In this embodiment, the calculation of the corresponding fault severity index value based on each fault feature sequence, and the determination of whether the system is in a fault condition based on the fault severity index value, and the determination of the fault severity if it is in a fault condition, specifically include:

[0088] Step S311. Fault severity index calculation: Calculate the fault severity index based on the temperature change parameters, mass flow rate change rate, and corresponding deviation from the mean value of the gasifier and combustion furnace at different locations in the fault characteristic sequence. R Fault severity index R The calculation model includes agglomeration fault calculation terms, wear fault calculation terms, and cross-feedback calculation terms. The agglomeration fault calculation term is calculated based on temperature transformation parameters and their corresponding deviations from the mean. The wear fault calculation term is calculated based on the mass flow rate change rate and its corresponding deviation from the mean. The cross-feedback calculation term is calculated based on the temperature transformation parameters and the deviations of the mass flow rate change rate from the mean, reflecting the cross-influence between agglomeration faults and wear faults. In other words, the agglomeration fault calculation term is derived from temperature parameters (including the absolute value of temperature change acceleration). TC and the degree of temperature deviation from the mean TD The calculation term is composed of mass flow rate parameters (including mass flow rate of change) to characterize the contribution of agglomeration faults to the overall fault. FC and the degree of deviation of mass flow rate from the mean FD) is used to represent the contribution of the abrasion failure to the failure, and the cross feedback calculation term is constructed by the total temperature transformation parameters (including the temperature change rate) and the mass flow rate deviation average value, to represent the contribution of the cross influence between the coking failure and the abrasion failure to the failure.

[0089] Specifically, the failure degree index R The calculation model can be represented as:

[0090] (5)

[0091] (6)

[0092] wherein, R is the failure degree index, ranging from (0~1), , are the temperature change acceleration absolute values of two different positions in the gasifier away from the air distribution plate of the combustor, , are the temperature change acceleration absolute values of the horizontal flue position of the gasifier and the horizontal flue position of the combustor, , are the mass flow rate change rates of the horizontal flue position of the gasifier and the horizontal flue position of the combustor, ~ are the temperature deviation average values of the two different positions in the gasifier away from the air distribution plate of the combustor, the horizontal flue position of the gasifier, and the horizontal flue position of the combustor, , are the mass flow rate deviation average values of the horizontal flue position of the gasifier and the horizontal flue position of the combustor, R TD is the total temperature deviation average value of each position, R TC is the total temperature change rate of each position, R FC is the total mass flow rate change rate of the horizontal flue position of the gasifier and the horizontal flue position of the combustor, R FD is the total mass flow rate deviation average value of the horizontal flue position of the gasifier and the horizontal flue position of the combustor, , are the global basic weight coefficients, which can be initially set as 0.4, for example, gamma is the weight used to reflect the mutual aggravation effect of different features on the evaluation of coking and abrasion, which can be initially set as 0.2, for example.

[0093] The failure degree index R constructed according to the above formula (5) is Ris a continuous value between 0 and 1, the closer to 0, the closer to normal state, the closer to 1, the further away from normal state, the higher the potential failure degree, and the continuous failure degree indicator R is used as the target variable of the subsequent failure classification prediction model. The first term represents the contribution of the caking failure, the second term represents the contribution of the wear failure, and the third term is a cross term, which can directly quantify: temperature anomaly (TC↑) and flow persistence damage (FD↑) jointly → accelerate new caking formation. For example: when and are both above the threshold, the γ term will significantly amplify the R value, and dynamic feedback capture can be achieved; when , it indicates that it is in the caking dominant mode, that is, caking failure is dominant, and the R characteristic is represented as: TC , TD rapidly rising, FC , FD , , FC , FD sudden increase, TC , TD slowly rising.

[0094] Step S312. Failure degree discrimination: according to the failure degree indicator R to determine whether it is a normal working condition or a failure working condition, and if it is a failure working condition, whether it is an early failure or a serious failure.

[0095] In this embodiment, if the failure degree indicator value R is less than the preset early failure threshold, it is determined to be a normal working condition, and the data label is configured as normal state, if the failure degree indicator value R is greater than the preset early failure threshold and less than the preset serious failure threshold, it is determined to be an early failure working condition, and the corresponding data label is configured as early failure, and if the failure degree indicator value R is greater than the preset serious failure threshold, it is determined to be a serious failure working condition, and the corresponding data label is configured as serious failure.

[0096] For example, fluidized beds are mostly in a normal state, with normal data accounting for about 90%. Early failures need to be increased to 20% (normal + early failures ≈ 90%), and severe failures account for 10%. Therefore, three key thresholds can be set to classify the failure levels. The specific value of R is related to each failure severity. When the R value is less than 0.8 (early failure threshold), it indicates normal data, and the failure severity label will be set to 0 to represent normal data. 0.9 is set as the severe failure threshold. When 0.8 ≤ R < 0.9, it indicates an early failure; when R ≥ 0.9, it indicates a severe failure, as shown in Table 3. Furthermore, when R is greater than 0.8, further failure classification judgment is required.

[0097] Table 3: Fault Severity Index Values ​​for Different Fault Levels R Range table

[0098]

[0099] In this embodiment, when the fault severity index value R When the value is less than the early fault threshold (e.g., 0.8), it indicates normal data, and the fault label will be set to 0, meaning it is normal data. When the fault severity index value... R Triggering early failure thresholds (e.g.) R When the value is ≥0.8, proceed to fault classification judgment, such as... Figure 4 As shown, the specific steps for determining the fault category based on the contribution status of each parameter during the calculation of the fault severity index value include:

[0100] Step S321. Dominant Pattern Recognition: Determine the impact of the agglomeration fault calculation item and the wear fault calculation item on the fault severity index value. R The contribution ratios are calculated as follows: if the contribution ratio corresponding to the clumping fault calculation item is greater than the preset contribution ratio threshold, it is determined to be a clumping fault-dominated mode, and the fault label is set as clumping fault. If the contribution ratio corresponding to the wear fault calculation item is greater than the preset contribution ratio threshold, it is determined to be a wear fault-dominated mode, and the fault label is set as wear fault. If the contribution ratios corresponding to the clumping fault calculation item and the wear fault calculation item are equal, the process proceeds to step S322 for secondary judgment, where the temperature parameter includes the absolute value of temperature change acceleration. TC and the degree of temperature deviation from the mean TD The mass flow rate parameter includes the mass flow rate of change. FC and the degree of deviation of mass flow rate from the mean FD .

[0101] Taking equation (5) as an example, if the first term Values ​​and Fault Severity Index Values R If the proportion exceeds a preset threshold (e.g., 70%), it is determined that the blockage fault is dominant, and the blockage fault is labeled with fault tags 1 and 2. If the second item Values ​​and Fault Severity Index Values R If the proportion exceeds a preset threshold (e.g., 70%), it is determined that wear failure is the dominant factor, and the fault label is set to wear failure at positions 3 and 4; if the first item The value, the second item The value of the fault severity index R If the values ​​are evenly distributed (e.g., all are 40%~60%), then a secondary judgment is required.

[0102] Step S322. Determine the fault severity index value R The proportion of cross-feedback terms in the calculation model, if the cross-feedback terms affect the fault severity index value R If the proportion is greater than the preset dual-fault judgment threshold, it is judged as a dual fault of agglomeration and wear and an early warning is triggered; if the cross-feedback item affects the fault severity index value R If the proportion is less than the preset single fault judgment threshold, it is judged as a single fault. The fault type is determined according to the larger contribution ratio of the clumping fault calculation item and the wear fault calculation item.

[0103] Taking equation (5) as an example, determine Fault severity indicators R If the percentage of [something] exceeds the preset dual-fault judgment threshold (e.g., 30%), it is judged as a dual fault of agglomeration and wear (i.e., agglomeration fault and wear fault coexist) and an early warning is triggered. Simultaneously, countermeasures are initiated, such as adjusting the fluidization velocity or injecting inert particles. If the cross-feedback item affects the fault severity index... R If the proportion is less than the preset single fault judgment threshold (e.g., 10%), the corresponding fault type is determined according to the contribution ratio of the blockage fault calculation item and the wear fault calculation item. For example, if the contribution ratio of the blockage fault calculation item is greater than the contribution ratio of the wear fault calculation item, it is judged as a blockage fault; otherwise, if the contribution ratio of the wear fault calculation item is greater than the contribution ratio of the blockage fault calculation item, it is judged as a wear fault.

[0104] Furthermore, in this embodiment, when a dual fault of agglomeration and wear is identified, the weights in the cross-feedback term are dynamically adjusted based on the degree of deviation of the total temperature change rate and the total mass flow rate from the mean at each location. gamma Both agglomeration and wear failures require long-term development processes, and these two types of failures can influence each other. Initially, the mutual influence is relatively small, but it gradually increases as the failure severity worsens. Therefore, the weights in the cross-feedback term in this embodiment... gammaThe initial value can be configured to a small value. When it is determined that a dual fault of agglomeration and wear has occurred, it indicates that there is a high probability of mutual influence between the two faults, and this mutual influence will gradually increase. Therefore, the weight in the cross-feedback term is increased accordingly. gamma The value can be dynamically updated based on real-time fault diagnosis results. gamma This allows for a more accurate representation of the real-time fault level and improves the accuracy of fault level calculation.

[0105] Specifically, the weights in the cross-feedback terms can be dynamically adjusted according to the following formula. gamma :

[0106] (7)

[0107] in, Indicates to Adjusted weights R TC This represents the total rate of temperature change at each location. R FD The degree to which the total mass flow rate deviates from the mean at the horizontal flue locations of the gasifier and the combustion furnace. Total temperature change rate The maximum value, The degree to which the total mass flow rate deviates from the mean The maximum value.

[0108] As shown in equation (7), by increasing Items can be increased The value is used to increase the importance of the cross-calculation term when both agglomeration and wear faults occur, and the total temperature change rate after the fault occurs. R TC and / or the degree of deviation of total mass flow rate from the mean R FD It will gradually increase accordingly, therefore The value of the item will gradually increase, allowing it to match increasingly deeper levels of fault and progressively increase its weight. gamma This value effectively improves the accuracy of fault severity calculation.

[0109] By following the above method, when Basic Items gamma When a certain percentage (e.g., 30%) is reached, the feedback loop will be activated, thereby increasing the warning priority. When temperature and mass flow anomalies persist, the base term γ will automatically increase, thus enhancing the warning priority. This strengthens the weight of the feedback item.

[0110] The fault type classification and fault degree judgment by the above method can further ensure that the diagnosis result conforms to the actual process state, and the continuous fault degree index can be constructed to realize fine quantization of the fault severity, instead of simple binary or multi-classification.

[0111] Step S04, fault prediction model training: training the pre-constructed GATT-CNN-LSTM fault prediction classification model using the labeled fault feature data set.

[0112] As shown in Figure 5 The GATT-CNN-LSTM fault prediction classification model in the embodiment includes a feature extraction module and a fault classification branch and a fault degree prediction branch connected with the feature extraction module, respectively. The feature extraction module includes a Gaussian convolution kernel time sequence attention layer GATT, a CNN layer and an LSTM layer connected in sequence. The fault type classification task is performed by the fault classification branch to output the fault type result, and the fault degree prediction task is performed by the fault degree prediction branch to output the fault degree prediction result. The fault category judgment signal (dominant mode signal) obtained each time is returned to the Gaussian convolution kernel time sequence attention layer GATT layer for input for the next fault diagnosis. The fault category judgment result of the last time is characterized by the GATT layer, and the characterized features are input into the LSTM module to predict the fault degree value of the current time, which is fed back to the fault classification decision module for fault type discrimination.

[0113] The GATT-CNN-LSTM fault prediction classification model is constructed in the embodiment to realize the prediction of the fault type and the fault degree of the agglomeration fault in the fluidized bed gasifier and the combustion furnace and the heating surface wear fault in parallel, as shown in Figure 6 After the Gaussian attention layer, the CNN layer and the LSTM are processed, the model extracts the key features, and then the data enters the classification task branch and the prediction task branch. In the classification branch, the model passes through two CNN layers and a fully connected layer, and finally outputs the classification result. In the prediction task branch, two LSTM layers and a fully connected layer are used, and finally the continuous fault index is output. The specific model parameters are shown in Figure 6

[0114] Input layer: The input layer is used to receive the feature data after normalization and serialization processing. Its main purpose is to format the input data into a shape that the model can process. The experimental data is merged to construct 12 features: temperature change acceleration absolute value TC 1- TC 4, temperature mean deviation TD 1- TD 4, mass flow rate change rate FC 1-​FC 2. Mass flow rate deviation from mean FD 1- FD 2. With 50 data points as a sample, the batch size of column processing is set to 50, the actual time is 0.2s, and the input data format is , where B is the batch size, B = 50, T is the time step, T = 0.2s, F is the number of features, F = 12. Two values (value as predicted value, as fault classification value) are constructed for this sample, which are the Fault value at the last time point plus 1 and the fault classification value. For example, the first sample is taken at time point 1-50, and the value of this sample is the Fault value at time point 51 and the fault classification at time point 51.

[0115]

[0116] Specifically, the GATT based on the Gaussian convolution kernel is a time series data processing by combining the Gaussian convolution kernel and the feature-level attention mechanism. Since there is strong time series in the sampling of the CFB data, the target data comes from the function mapping of the past data for a certain period of time, so the model hopes to select the time series of the features. The Gaussian convolution kernel is a one-dimensional filter based on Gaussian distribution, which is used for smoothing the signal. The variable is regarded as the representation of data at different time in the same high-dimensional feature space, and for each input feature of the time series attention layer, there is a corresponding Gaussian convolution model for calculation, and the models are independent of each other. The GATT layer is used to learn the weight coefficients of the fault degree index R , 、 , the mathematical definition is:

[0117] (8)

[0118] wherein is the weight coefficient of the calculation model of the fault degree index R , 、 , and σ is the standard deviation of the Gaussian distribution, which controls the smoothing degree. The data set is discretized into a one-dimensional array by the Gaussian kernel:

[0119] (9)

[0120] wherein i is the array index, and c is the center position of the kernel ( ), and then normalized to make the weight sum of the convolution kernel equal to 1.

[0121] The data set is discretized into a one-dimensional array by the Gaussian kernel, including 12 feature dimensions (temperature change acceleration absolute value TC1- TC 4. Temperature average deviation TD 1- TD 4. Mass flow rate change rate FC 1- FC 2. Mass flow rate deviates from the mean FD 1- FD 2) Learnable parameters, and incorporate the fault severity index values ​​as shown in equations (5) and (6) into the calculation model. , The initial weight coefficients for the global layer can be set to 0.4. The initial weights for γ, which reflect the mutual aggravation effect of different features on the evaluation of quantization of clumping and wear, are set to 0.2. These parameters are used as learnable initial values ​​for the attention layer to be dynamically optimized during training.

[0122] In the feature attention mechanism, for the input tensor, Where B is the batch size, T is the time step, and F is the number of features. This layer calculates attention weights separately for each feature dimension. For each feature f, the attention calculation process is as follows:

[0123] Feature extraction:

[0124] (10)

[0125] Applying Gaussian convolution:

[0126] (11)

[0127] in, This is a one-dimensional convolution operation.

[0128] Applying linear transformations and activation functions:

[0129] (12)

[0130] in, and These are learnable scaling and offset parameters.

[0131] The final output is the element-wise product of the original input and the attention weights:

[0132] (13)

[0133] Assume the input data sequence is: The output of the Gaussian convolution kernel temporal attention layer can be expressed as:

[0134] (14)

[0135] in It is the output of the Gaussian kernel convolution layer. It is a batch index. It is a time step index. It is a feature index. It is the kernel size. It is the central location of the core. These are the normalized Gaussian convolution kernel weights. and It is a feature Learnable parameters.

[0136] This embodiment sets up a temporal attention layer with the help of Gaussian convolution kernels, enabling the model to focus on data points that are adjacent in time and capture local temporal patterns. Each feature dimension has its own attention weight, which allows the model to dynamically adjust the degree of attention to different features. Moreover, the Gaussian convolution kernel attention layer is located after the input layer and before the CNN layer. By learning the attention weights, the model can automatically select the features that are more critical to the prediction task. At the same time, the Gaussian convolution takes into account the local correlation between time steps and can also highlight the key time steps and features.

[0137] Convolutional layer design CNN: In the GATT-CNN-LSTM fault prediction classification model in this embodiment, there are two parts of convolutional layer operations. One is in the general feature extraction part, which is composed of Gaussian convolutional layers, convolutional layers and LSTM layers, and is used to extract features together; the other convolutional layer operation is in the fault classification task of the model, which is used to complete the fault classification task.

[0138] In this embodiment, given that the input data is a time series, a small convolutional kernel is slid across the input sequence to detect short-term trends in the time series. This convolutional kernel can effectively capture translation-invariant features in the sequence data. The convolutional kernel moves along the sequence dimension, i.e., the time step, and calculates a weighted sum of the kernel and the corresponding input subsequence at each position. Assume the input is the first... The layer's output (ignoring batch dimension B, where L is the sequence length and C is the number of channels), and the convolutional kernel. ( For the size of the core, (Number of filters / Number of output channels) is used to obtain the final output feature map. .

[0139] For the first... One channel, at location Values ​​(step size S=1, padding='same' to adjust input boundaries):

[0140] (15)

[0141] in This offset is required to achieve 'same' padding, and padding will be performed according to the 'same' padding rule outside the boundary. The CNN in the first convolutional layer, which is after the Gaussian attention layer and before the LSTM layer, is used to extract the primary feature data after the Gaussian attention layer. It captures local features by using small convolutional kernels. The model structure and parameters of this CNN layer are shown in Table 4.

[0142] Table 4: Convolutional Layer Parameter Table for Feature Extraction Module

[0143]

[0144] Conv1D represents a one-dimensional convolutional layer. When the data passes through a Gaussian convolutional layer, the input data to the one-dimensional convolutional layer is... The mathematical formula for a one-dimensional convolutional layer is:

[0145] (16)

[0146] in, , These are the convolution kernel weights. It is a bias term. for In .

[0147] BatchNormalization is a batch normalization layer that accelerates the training process, improves model stability, reduces sensitivity to initialization parameters, and also has a certain regularization effect. It performs a standardization operation on the activation value of each channel within the current mini-batch, first calculating the mean of each feature channel in the mini-batch. and variance Then, using the calculated mean and variance, the activation values ​​of this channel are standardized to obtain the corresponding results. Then, using learnable scaling and offset parameters, a linear transformation is applied to the standardized values ​​to obtain the final output. .

[0148] The activation layer, or activation function, is a key component for introducing expressive power into the network. Introducing non-linearity allows the network to learn more complex functional relationships. In this embodiment, the ReLU activation function is used.

[0149] (17)

[0150] By using the activation function described above, computation can be made more efficient and gradient vanishing can be mitigated.

[0151] MaxPooling1D is a max pooling layer. Pooling layers achieve feature compression through spatial downsampling. Its core functions are twofold: reducing the feature map size and decreasing computational complexity. Max pooling improves the spatial invariance of features by selecting local maxima, while average pooling smooths local fluctuations and enhances the model's generalization ability. This embodiment uses a max pooling layer, namely:

[0152] (18)

[0153] in, It is the input feature map at the location The activation value.

[0154] The Dropout layer achieves regularization by randomly dropping a subset of neurons, reducing dependence on specific neurons and thus preventing overfitting.

[0155] (19)

[0156] After this layer, the original input data shape Become Assuming the input data... After passing through the CNN layer in the general feature extraction layer, the data will become , After BatchNormalization, Activation, MaxPooling1D, and DropOut, it becomes .

[0157] In the fault classification task, the CNN is placed in the classification task branch after the feature extraction module. It is used to further extract and refine patterns from the temporal features of the data processed by the feature extraction module, thereby improving the accuracy of fault classification in this branch. Specifically, the fault classification branch includes at least two CNN layers and one fully connected layer. Each layer has a similar model structure to the CNN layer in the feature extraction module, with five functional layers: Conv1D, BatchNormalization, Activation, MaxPooling1D, and DropOut. The CNN layer at the input receives features output from the feature extraction module, and after passing through each CNN layer and the fully connected layer, it outputs the fault type classification result. The fault severity prediction branch includes at least two LSTM layers and one fully connected layer. The LSTM layer at the input receives features output from the feature extraction module, and after passing through each LSTM layer and the fully connected layer, it outputs the fault severity prediction result. Specific model structures and parameters are shown in Tables 5 and 6. Table 6 represents a deeper CNN model that uses more convolutional kernels, enabling more in-depth feature extraction from the data.

[0158] Table 5: Parameters of the first convolutional layer in the fault classification task

[0159]

[0160] Table 6: Parameters of the second convolutional layer in the fault classification task

[0161]

[0162] LSTM Layer: After the convolutional layers in the feature extraction module extract relevant features, the data enters the LSTM layer, which captures long-term dependencies in the sequence data. Located after the CNN layer in the feature extraction module, the LSTM layer receives the local features extracted by the CNN, captures the temporal dependencies of these features, improves the spatial features extracted by the CNN through temporal modeling, and simultaneously remembers and processes long-term dependencies in the time series. After this layer, the input data... ( ) becomes ( The feature dimension increased from 64 to 128, while the batch size remained the same. After BatchNormalization and Dropout, it becomes , This will serve as input data for both the classification and regression tasks.

[0163] The LSTM layer in the fault severity prediction task branch consists of two layers. The first LSTM layer extracts more refined temporal features for the fault severity prediction regression task after the data input from the feature extraction module. This LSTM layer preserves sequence information and provides it to the next LSTM layer for further processing. The second LSTM layer is used to compress the entire sequence information into a fixed-length vector representation, extracting the final high-level temporal features.

[0164] In this embodiment, the GATT-CNN-LSTM fault prediction and classification model includes two branches: fault classification and fault severity prediction. Two fully connected layers and two output layers are set as the inputs for the classification and regression tasks, respectively. The input data for both branches is the data after feature extraction. The fully connected layers integrate and perform nonlinear transformations to help the model learn complex feature relationships. The output layer then maps these features to the final prediction result, enabling effective prediction of fault indicators. The fault classification task uses the softmax function to output classification label values, while the fault severity prediction task uses the linear activation function to predict a continuous value to represent the prediction of the fault severity indicator value R.

[0165] Step S05: Real-time parallel fault prediction: Real-time acquisition of state parameters of the gasifier and combustion furnace at different locations during the operation of the tested hot dual-circulation fluidized bed system, extraction of corresponding change values ​​and deviation from the mean values ​​to obtain real-time fault feature sequences, input into the trained GATT-CNN-LSTM fault prediction classification model, and output of fault degree prediction results and fault type classification results.

[0166] Specifically, the temperature data of the gasifier and combustion furnace at various monitoring locations (location 1 to location 4) and the mass flow rate of the gasifier and combustion furnace at designated locations are collected in real time during the operation of the fluidized bed system. The absolute value of temperature change acceleration and the rate of change of mass flow rate are then extracted to obtain a real-time fault feature sequence. This sequence is then input into the GATT-CNN-LSTM fault prediction and classification model trained in step S04 to perform fault degree prediction and fault type classification in real time and in parallel. It can not only determine whether a fault has occurred, but also determine the fault type, so as to provide timely warnings in the early stages of a fault. It can also determine the specific fault type, classify agglomeration faults and heated surface wear faults formed by different mechanisms, and accurately locate the specific location.

[0167] This invention employs a hybrid deep learning model that integrates convolutional neural networks, Gaussian convolutional kernel temporal attention layers, and long short-term memory networks for prediction. The Gaussian convolutional kernel temporal attention layer focuses on key time steps and features, while CNN extracts local patterns and spatial features from temperature sequences. LSTM captures long-term temporal dependencies. By constructing a shared feature extraction layer, the extracted features are input into the classification and prediction task branches respectively. This enables parallel classification of agglomeration faults and heated surface wear faults occurring during fluidized bed operation, and predicts the severity of normal, early, and severe faults. Simultaneously, it achieves the identification of fluidized bed fault types and continuous quantitative assessment of fault severity. This allows fluidized bed operators to promptly address early faults and make appropriate maintenance decisions based on different fault types and severity levels, reducing the operation and maintenance costs of fluidized bed equipment and ensuring equipment safety.

[0168] This embodiment further provides a computer device, including a processor and a memory, the memory for storing a computer program and the processor for executing the computer program to perform the method as described above.

[0169] It is understood that the method described in this embodiment can be executed by a single device, such as a computer or server, or it can be applied to a distributed scenario where multiple devices cooperate to complete the task. In a distributed scenario, one of the multiple devices may execute only one or more steps of the method described in this embodiment, and the multiple devices interact to complete the method. The processor can be implemented using a general-purpose CPU, microprocessor, application-specific integrated circuit, or one or more integrated circuits, and is used to execute relevant programs to implement the method described in this embodiment. The memory can be implemented using read-only memory (ROM), random access memory (RAM), static storage devices, and dynamic storage devices. The memory can store the operating system and other applications. When the method described in this embodiment is implemented through software or firmware, the relevant program code is stored in the memory and called and executed by the processor.

[0170] This embodiment further provides a computer-readable storage medium storing a computer program that, when executed by a processor, implements the method described above.

[0171] Those skilled in the art will understand that the above 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 embodied on one or more computer-readable storage media (including, but not limited to, disk storage, CD-ROM, optical storage, etc.) containing computer-usable program code. The present invention is described with reference to flowchart illustrations and / or block diagrams of methods, apparatus (systems), and computer program products according to embodiments of the present invention. It will be understood that each block of the flowchart illustrations and / or block diagrams, and combinations of blocks in the flowchart illustrations and / or block diagrams, can be implemented by computer program instructions. These computer program instructions can be provided to a processor of a general-purpose computer, special-purpose computer, embedded processor, or other programmable data processing apparatus to produce a machine, such that the instructions, which execute via the processor of the computer or other programmable data processing apparatus, produce implementations of the flowchart illustrations and / or block diagrams. Figure 1 One or more processes and / or boxes Figure 1 The 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 operate 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 functions specified in one or more boxes. These computer program instructions may also be loaded onto a computer or other programmable data processing apparatus to cause a series of operational steps to be performed on the computer or other programmable apparatus to produce a computer-implemented process, thereby providing instructions that execute on the computer or other programmable apparatus for implementing the process. Figure 1 One or more processes and / or boxes Figure 1 The steps of the function specified in one or more boxes.

[0172] The above description is merely a preferred embodiment of the present invention and is not intended to limit the invention in any way. Although the present invention has been disclosed above with reference to preferred embodiments, it is not intended to limit the invention. Therefore, any simple modifications, equivalent changes, and alterations made to the above embodiments based on the technical essence of the present invention without departing from the scope of the present invention should fall within the protection scope of the present invention.

Claims

1. A GATT-CNN-LSTM-based parallel prediction method for fluidized bed agglomeration and attrition faults, characterized by the steps of Comprise: Step S01, fault data acquisition: respectively in different fault degree of hot double circulating fluidized bed system under the condition of collecting gasification furnace, combustion furnace in different monitoring position state parameters form fault data set, the state parameters include temperature data and mass flow rate; Step S02, fault feature extraction: according to fault data set respectively extracted at different measurement time point of each monitoring position state parameter change value and deviation from mean value value get fault feature, each measurement time point get a group of fault feature sequence; Step S03, fault data classification: according to each measurement time point of fault feature sequence calculation corresponding fault degree index value, according to the fault degree index value judge whether in fault condition and judge fault degree and fault category, according to the judgment result of fault degree and fault category configuration each fault feature sequence corresponding fault degree label and fault category label, get fault feature data set with label; Step S04, fault prediction model training: using the fault feature data set with label trains pre-constructed GATT-CNN-LSTM fault prediction classification model; Step S05, real-time fault parallel prediction: real-time acquisition of fluidized bed system in the running process of gasification furnace, combustion furnace in different position state parameters, and extract the corresponding change value and deviation from mean value value get real-time fault feature sequence, input into the trained GATT-CNN-LSTM fault prediction classification model, get fault degree prediction result and fault type classification result output; In step S03, when judging as fault condition, according to the contribution state of each parameter in fault degree index value calculation process judges fault category, including: Step S321. Determine the impact of the agglomeration fault calculation item and the wear fault calculation item on the fault severity index value. R The contribution ratios are calculated as follows: if the contribution ratio corresponding to the clumping fault calculation item is greater than the preset contribution ratio threshold, it is determined to be a clumping fault-dominated mode, and the fault label is set as clumping fault. If the contribution ratio corresponding to the wear fault calculation item is greater than the preset contribution ratio threshold, it is determined to be a wear fault-dominated mode, and the fault label is set as wear fault. If the contribution ratios corresponding to the clumping fault calculation item and the wear fault calculation item are equal, the process proceeds to step S322 for secondary judgment, where the temperature parameter includes the absolute value of temperature change acceleration. TC and the degree of temperature deviation from the mean TD ; Step S322. Determine the fault degree index value R The proportion of the cross feedback item in the calculation model, if the proportion of the cross feedback item to the fault degree index value R is greater than the preset double fault judgment threshold, it is determined that there is a caking and wear double fault and a warning is triggered; if the proportion of the cross feedback item to the fault degree index value R is less than the preset single fault judgment threshold, it is determined that there is a single fault, and the fault type is determined according to the larger item corresponding to the contribution ratio of the caking fault calculation item and the wear fault calculation item. In step S03, according to the fault feature sequence calculation corresponding fault degree index value includes: The fault degree index is calculated according to the temperature variation parameters of the gasifier and the combustion furnace at different positions in the fault feature sequence, the mass flow rate variation rates, and corresponding deviation mean degree values R The temperature variation parameters include temperature variation acceleration absolute values, and the deviation mean degree values include temperature deviation mean degree values and mass flow rate deviation mean degree values R The calculation model of the fault degree index includes a coking fault calculation item, a wear fault calculation item, and a cross feedback calculation item. The coking fault calculation item is calculated according to the temperature variation parameters and corresponding deviation mean degree values. The wear fault calculation item is calculated according to the mass flow rate variation rates and corresponding deviation mean degree values. The cross feedback calculation item is calculated according to total temperature variation parameters of each position and total mass flow rate deviation mean degree values.

2. The GATT-CNN-LSTM based parallel prediction method of the fluidized bed agglomeration and attrition fault according to claim 1, characterized in that, In step S01, the fault degree condition includes normal operation condition, early fault condition and serious fault condition, the collected temperature data includes the cross section temperature at different heights from the combustion furnace air distribution plate in the gasification furnace, the temperature data changing with time in the gasification furnace horizontal flue, the temperature data changing with time in the combustion furnace horizontal flue, the collected mass flow rate includes the mass flow rate changing with time in the combustion furnace horizontal flue and the mass flow rate changing with time in the gasification furnace horizontal flue, the early fault condition includes early heating surface wear fault of gasification furnace and early coking fault of combustion furnace, the serious fault condition includes serious heating surface wear fault of gasification furnace and serious coking fault of combustion furnace.

3. The GATT-CNN-LSTM based parallel prediction method of the fluidized bed agglomeration and attrition fault according to claim 1, characterized in that, Fault level indicator R The computational model for the fault level indicator is: wherein, R is the fault degree index value, , are the absolute values of the temperature variation acceleration at two different positions in the gasifier, , are the absolute values of the temperature variation acceleration at the horizontal flue position of the gasifier and at the horizontal flue position of the burner, , are the mass flow rate variation rates at the horizontal flue position of the gasifier and at the horizontal flue position of the burner, ~ is the temperature deviation from the mean value at two different positions in the gasifier, at the horizontal flue position of the gasifier and at the horizontal flue position of the burner, , is the mass flow rate deviation from the mean value at the horizontal flue position of the gasifier and at the horizontal flue position of the burner, TD is the total temperature deviation from the mean value at each position, TC is the total temperature variation rate at each position, FC is the total mass flow rate variation rate at the horizontal flue in the gasifier and at the horizontal flue position of the burner, FD is the total mass flow rate deviation from the mean value at the horizontal flue position of the gasifier and at the horizontal flue position of the burner, , are the global base weight coefficients, Gamma The weights are used to reflect the mutual aggravation effect of different features on the assessment of the agglomeration and wear.​ 4. The GATT-CNN-LSTM based parallel prediction method of the fluidized bed agglomeration and attrition fault according to claim 3, characterized in that, In step S03, judging whether the working condition is in failure and judging the failure degree according to the failure degree index value includes: if the failure degree index value R is less than the preset early failure threshold, it is determined that the working condition is normal, and the data label is configured as normal state; if the failure degree index value R is greater than the preset early failure threshold and less than the preset serious failure threshold, it is determined that the working condition is early failure, and the corresponding data label is configured as early failure; and if the failure degree index value R is greater than the preset serious failure threshold, it is determined that the working condition is serious failure, and the corresponding data label is configured as serious failure.

5. The GATT-CNN-LSTM based parallelized prediction method of the fluidized bed agglomeration and attrition faults according to claim 1, characterized in that, comprising dynamically adjusting the weight in the cross feedback term according to the total temperature rate of change at each location and the degree of deviation of the total mass flow rate from the mean value when the coking and wear double fault is judged γ , the calculation expression is: wherein, represents the total temperature change rate adjusted weight, R TC is the total temperature change rate for each position, R FD is the total mass flow rate deviation from the mean for the gasifier horizontal flue position, the combustion furnace horizontal flue position, is the maximum value of the total temperature change rate , is the maximum value of the total mass flow rate deviation from the mean .

6. The GATT-CNN-LSTM based parallel prediction method of the fluidized bed agglomeration and attrition fault according to any one of claims 1-5, characterized in that, The GATT-CNN-LSTM fault prediction classification model comprises a feature extraction module, a fault classification branch and a failure degree prediction branch, the feature extraction module comprises a Gaussian convolution kernel time sequence attention layer GATT, a CNN layer and an LSTM layer connected in sequence, the Gaussian convolution model adopted by the Gaussian convolution kernel time sequence attention layer GATT in the GATT-CNN-LSTM fault prediction classification model is: wherein is a fault degree indicator R is a weight coefficient in a calculation model of , σ is a standard deviation of a Gaussian distribution for controlling a smoothing degree; The output of the Gaussian convolution kernel time sequence attention layer GATT is expressed as: where, is the output of the Gaussian kernel convolution layer, is the batch index, is the time step index, is the feature index, is the convolution kernel size, is the center position of the kernel, is the normalized Gaussian kernel weight, and are learnable parameters of the feature .

7. The GATT-CNN-LSTM based parallel prediction method of the fluidized bed agglomeration and attrition fault according to claim 6, characterized in that, The fault classification branch comprises at least two CNN layers and one fully connected layer, the features output by the feature extraction module are received by the CNN layer at the input end, and the fault type classification result is output after the features pass through the CNN layers and the fully connected layer; The failure degree prediction branch comprises at least two LSTM layers and one fully connected layer, the features output by the feature extraction module are received by the LSTM layer at the input end, and the failure degree prediction result is output after the features pass through the LSTM layers and the fully connected layer.

8. A computer apparatus comprising a processor and a memory for storing a computer program, characterised in that, The processor is configured to execute the computer program to perform the method of any one of claims 1-7.

9. A computer readable storage medium storing a computer program, characterized in that, The computer program is configured to be executed by the processor to perform the method of any one of claims 1-7.

Images