Fluidized bed caking and wear fault parallel prediction method and device based on GATT-CNN-LSTM

Abstract

The invention discloses a fluidized bed caking and wear fault parallel prediction method and device based on GATT-CNN-LSTM. The method comprises the following steps: step S01, acquiring temperatures and mass flow rates at different monitoring positions to form a fault data set; s02, respectively extracting change rates according to the fault data set to obtain fault features; s03, calculating a fault degree index value according to each fault feature sequence so as to judge a corresponding fault degree, judging a fault type according to the contribution state of each parameter, and configuring a fault degree and a fault type label; step S04, training a fault prediction classification model by using the fault feature data set with the label; and S05, extracting a real-time fault feature sequence, and inputting the real-time fault feature sequence into the trained fault prediction classification model to obtain a fault degree and a fault type result. According to the method, the caking fault and the wear fault in the fluidized bed can be identified in parallel, and the fault degree can be predicted.

Description

Technical Field

[0001] The present invention relates to the technical field of coal power equipment fault diagnosis, and in particular to a method and device for parallel prediction of fluidized bed agglomeration and wear faults based on GATT-CNN-LSTM. Background Art

[0002] Fluidized beds, with their unique fluidization state, enable efficient and stable combustion and material handling of solid fuels such as coal, making them widely used in power generation, chemical engineering, metallurgy, and other fields. Fluidized beds operate in a complex process and are affected by numerous factors. For one thing, the fluid dynamics within a fluidized bed are complex and variable, making the interaction between the gas-solid two-phase flow difficult to accurately predict and control. For example, abnormal fluidization phenomena such as channeling and surging can severely affect the uniformity and stability of combustion or reactions. Furthermore, variations in numerous process parameters, such as fuel properties, particle size distribution, fluidizing velocity, and temperature, can interfere with the fluidized bed's operation. The combined effects of these complex factors can make fluidized beds susceptible to various faults during operation, such as localized coking, flameout, and overheating. These can not only cause equipment damage and production interruptions, but can also pose serious safety risks and environmental pollution. Therefore, timely and accurate fault diagnosis during fluidized bed operation is essential. This not only ensures stable operation of fluidized bed equipment, improves production efficiency, and reduces economic losses, but also contributes to environmental protection and safe production.

[0003] At present, there are mainly the following methods for fluidized bed fault diagnosis: 1. Fault diagnosis based on sensor monitoring data: Sensors, such as temperature, vibration, pressure, or acoustic sensors, are installed at key locations on the fluidized bed. The data collected by these sensors is used to determine if a fault exists. This method can only simply identify operating parameter out-of-limit conditions at key locations, but cannot identify the specific fault type or severity. Consequently, parameter anomalies are often detected only after a serious fault has occurred, failing to provide early warning of a fault.

[0004] 2. Fault diagnosis based on physical models: A physical model is established based on the physical properties, chemical reaction mechanisms, and fluid mechanics of the fluidized bed. A fault pattern library is established based on the physical model and historical data. During the operation of the fluidized bed, real-time measurement data is acquired and compared with the model's predicted values ​​to identify the fault type. However, the physical and chemical processes of the fluidized bed itself are complex. The flow, heat transfer, mass transfer, and chemical reactions within it are intertwined, making it difficult to accurately describe them using simple physical models. As a result, the actual prediction accuracy is not high.

[0005] 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 operating data of the fluidized bed to achieve fault classification and prediction.

[0006] However, on the one hand, the data measured by sensors often have the characteristics of high dimensionality, nonlinearity and strong noise. Different fault types may show similar characteristics in the sensor signal, while the same fault type may present different characteristic patterns due to changes in operating conditions. Different fault types also have different sensitivities to different types of sensors. Feature extraction and fault classification are difficult. Traditional fault classification based on machine learning algorithms is not accurate. In addition, the fluidized bed system has large time variability and uncertainty. Its operating conditions may change significantly with time, raw material batches and other factors, resulting in the established fault diagnosis model being difficult to adapt to new operating conditions and requiring frequent updates and optimization.

[0007] On the other hand, the existing technology based on machine learning algorithms can usually only realize the classification of single fault types caused by the same mechanism, such as agglomeration faults or blockage faults, and cannot realize the classification of fault types caused by different mechanisms. Agglomeration and wear faults are faults caused by different mechanisms, and cannot be effectively classified based on the existing technical solutions. The occurrence of faults in the fluidized bed is often not independent. If only a single fault of a certain type is monitored, it is easy to ignore the risk of synergistic deterioration between different faults. In addition, the traditional machine learning algorithm can only realize simple fault classification, but cannot judge the extent of the fault. In particular, some early faults may not be obvious in the original sensor data set. The weak abnormal signals caused by early faults are easily masked by the background noise during normal operation and the coupling fluctuations between parameters. It is difficult to accurately identify the early fault simply by operating monitoring parameters. Summary of the Invention

[0008] The technical problem to be solved by the present invention is: in response to the technical problems existing in the prior art, the present invention provides a parallel prediction method and device for fluidized bed agglomeration and wear faults based on GATT-CNN-LSTM, which has a simple implementation method, low cost, high fault diagnosis efficiency and accuracy, and can parallelly realize the identification and classification of agglomeration faults and wear faults of gasification furnaces and combustion furnaces in the fluidized bed, and can also parallelly realize the prediction of fault degree.

[0009] In order to solve the above technical problems, the technical solution proposed by the present invention is: A parallel prediction method for agglomeration and wear failure in a fluidized bed based on GATT-CNN-LSTM, comprising the following steps: Step S01, fault data collection: collecting state parameters of the gasifier and the combustion furnace at different monitoring positions under different fault severity conditions of the hot dual circulating fluidized bed system to form a fault data set, wherein the state parameters include temperature data and mass flow rate; Step S02, fault feature extraction: extract the change value of the state parameter of each monitoring location at different measurement time points and the degree of deviation from the mean value to obtain the fault feature according to the fault data set, and obtain a set of fault feature sequences at each measurement time point; Step S03, fault data classification: Calculate the corresponding fault severity index value based on the fault feature sequence at each measurement time point, determine whether it is in a faulty working condition and determine the fault severity and fault category based on the fault severity index value, and assign a fault severity label and a fault category label corresponding to each fault feature sequence based on the fault severity and fault category judgment results to obtain a labeled fault feature dataset; Step S04, fault prediction model training: using the labeled fault feature dataset to train a pre-built GATT-CNN-LSTM fault prediction classification model; Step S05, real-time fault parallel prediction: The state parameters of the gasifier and combustion furnace at different positions of the fluidized bed system under test are collected in real time during operation, and the corresponding change values ​​and deviation from the mean are extracted to obtain a real-time fault feature sequence, which is input into the trained GATT-CNN-LSTM fault prediction and classification model to obtain the fault degree prediction result and the fault type classification result output.

[0010] Furthermore, in step S01, the fault degree conditions include normal operating conditions, early fault conditions and severe fault conditions, the collected temperature data include cross-sectional temperatures at multiple different heights from the combustion furnace air distribution plate in the gasifier, temperature data that changes with time in the horizontal flue of the gasifier, and temperature data that changes with time in the horizontal flue of the combustion furnace, the collected mass flow rate includes the mass flow rate that changes with time in the horizontal flue of the combustion furnace and the mass flow rate that changes with time in the horizontal flue of the gasifier, the early fault conditions include early heating surface wear faults in the gasifier and early agglomeration faults in the combustion furnace, and the severe fault conditions include severe heating surface wear faults in the gasifier and severe agglomeration faults in the combustion furnace.

[0011] Furthermore, in step S03, calculating the corresponding fault degree index value according to the fault feature sequence includes: The fault severity index is calculated 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 positions in the fault feature sequence. RThe temperature change parameter includes the absolute value of the temperature change acceleration, the deviation from the mean value includes the temperature deviation from the mean value and the mass flow rate deviation from the mean value, the fault degree index R The calculation model includes agglomeration fault calculation items, wear fault calculation items and cross feedback calculation items. The agglomeration fault calculation item is calculated based on the temperature conversion parameter and the corresponding deviation from the mean value. The wear fault calculation item is calculated based on the mass flow rate change rate and the corresponding deviation from the mean value. The cross feedback calculation item is calculated based on the total temperature conversion parameter of each position and the total mass flow rate deviation from the mean value. The fault degree index R The calculation model is:

[0012]

[0013] in, R is the fault degree index value, 、 are the absolute values ​​of temperature change acceleration at two different locations in the gasifier away from the combustion furnace air distribution plate, 、 are the absolute values ​​of temperature change acceleration at the horizontal flue position of the gasifier and the combustion furnace, respectively. 、 are the mass flow rate change rates at the gasifier horizontal flue position and the combustion furnace horizontal flue position, respectively. ~ is the degree of temperature deviation from the mean at two different locations in the gasifier, the gasifier horizontal flue position, and the combustion furnace horizontal flue position. 、 R is the degree of deviation of the mass flow rate at the gasifier horizontal flue position and the combustion furnace horizontal flue position from the mean, TD is the degree of deviation of the total temperature at each location from the mean, R TC is the total temperature change rate at each location, R FC R is the total mass flow rate change rate in the gasifier horizontal flue and the combustion furnace horizontal flue, FD is the degree of deviation of the total mass flow rate at the gasifier horizontal flue position and the combustion furnace horizontal flue position from the mean, 、 are the global basic weight coefficients, γ The weights are used to reflect the mutual aggravating effects of different features in evaluating and quantifying caking and wear.

[0014] Further, in step S03, judging whether the system is in a fault condition and judging the fault degree according to the fault degree index value include: if the fault degree index value R is less than the preset early fault threshold, it is judged as a normal condition and the data tag is configured as a normal state; if the fault degree index value R is greater than the preset early fault threshold and less than the preset severe fault threshold, it is judged as an early fault condition and the corresponding data tag is configured as an early fault; if the fault degree index value R is greater than the preset severe fault threshold, it is judged as a severe fault condition and the corresponding data tag is configured as a severe fault.

[0015] Furthermore, in step S03, when it is determined to be a fault condition, the fault type is determined according to the contribution status of each parameter in the fault degree index value calculation process, including: Step S321. Determine the fault degree index value of the agglomeration fault calculation item and the wear fault calculation item R If the contribution ratio of the agglomeration fault calculation item is greater than the preset contribution ratio threshold, it is determined to be the agglomeration fault dominant mode, and the fault label is set to agglomeration fault. If the contribution ratio of the wear fault calculation item is greater than the preset contribution ratio threshold, it is determined to be the wear fault dominant mode, and the fault label is set to wear fault. If the contribution ratios of the agglomeration fault calculation item and the wear fault calculation item are balanced, then the process goes to step S322 for secondary judgment, where the temperature parameter includes the absolute value of the 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 cross feedback items in the calculation model of R If the proportion of the cross feedback item is greater than the preset double fault judgment threshold, it is judged as a double fault of agglomeration and wear and triggers an early warning; if the cross feedback item has a negative impact on the fault degree index value R If the proportion is less than the preset single fault judgment threshold, it is determined to be a single fault, and the fault type is determined based on the fault corresponding to the larger item in the contribution ratio of the agglomeration fault calculation item and the wear fault calculation item.

[0016] Furthermore, when it is determined that there is a dual fault of agglomeration and wear, the weight of the cross feedback item is dynamically adjusted according to the total temperature change rate of each position and the degree of deviation of the total mass flow rate from the mean. γ , the calculation expression is:

[0017] in, Express Adjusted weight, R TC is the total temperature change rate at each location, R FDis the degree of deviation of the total mass flow rate at the gasifier horizontal flue position and the combustion furnace horizontal flue position from the mean, is the total temperature change rate The maximum value of is the degree of deviation of the total mass flow rate from the mean The maximum value of .

[0018] Furthermore, the GATT-CNN-LSTM fault prediction and classification model includes a feature extraction module, a fault classification branch, and a fault severity prediction branch. The feature extraction module includes a Gaussian convolution kernel temporal attention layer GATT, a CNN layer, and an LSTM layer connected in sequence. The Gaussian convolution model used in the Gaussian convolution kernel temporal attention layer GATT in the GATT-CNN-LSTM fault prediction and classification model is:

[0019] in, Is an indicator of the degree of failure R The weight coefficient in the calculation model 、 , σ is the standard deviation of the Gaussian distribution used to control the degree of smoothing; The output of the Gaussian convolution kernel temporal attention layer GATT is expressed as:

[0020]

[0021] in, 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 of the nucleus, is the normalized Gaussian convolution kernel weight, and It is a feature The learnable parameters.

[0022] Furthermore, the fault classification branch includes at least two CNN layers and one fully connected layer. The CNN layer at the input end receives the features output by the feature extraction module, and outputs the fault type classification result after passing 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 LSTM layer at the input end receives the features output by the feature extraction module, and outputs the fault degree prediction result after passing through each LSTM layer and the fully connected layer.

[0023] A computer device includes a processor and a memory, wherein the memory is used to store a computer program, and the processor is used to execute the computer program to perform the above method.

[0024] A computer-readable storage medium storing a computer program, wherein the computer program implements the above method when executed by a processor.

[0025] Compared with the prior art, the present invention has the following beneficial effects: 1. The present invention collects temperature data and mass flow rate at different monitoring positions of the gasifier and combustion furnace under different fault severity conditions to form a fault data set, extracts temperature change parameters, mass flow rate change rate and deviation from the mean from the fault data set as fault features, first calculates the fault severity index value to judge the fault severity, and when it is judged to be in a fault condition, further identifies the fault type according to the contribution state of each parameter in the process of calculating the fault severity index value, which can fully explore the correlation between different characteristic agglomeration faults and wear faults and between different characteristics and different severity faults, and then uses the fault feature data set with two labels to train the fault prediction classification model, which can not only simultaneously predict the fluidized bed agglomeration and wear faults, but also predict the fault severity, thereby realizing the parallel prediction of fault type identification and fault severity prediction, thereby facilitating making corresponding maintenance decisions for different fault types and fault severity, and reducing maintenance costs.

[0026] 2. In the present invention, a fault prediction classification model is constructed by adopting the GATT-CNN-LSTM architecture, and feature extraction is performed using the Gaussian convolution kernel temporal attention layer GATT, CNN layer and LSTM layer. By sharing feature extraction and retaining the requirements of classification and prediction tasks, the advantages of convolutional neural networks in data feature extraction and the advantages of recurrent neural networks in capturing temporal features can be fully utilized, thereby improving the overall predictability of the model. BRIEF DESCRIPTION OF THE DRAWINGS

[0027] Figure 1 This is a schematic diagram of the implementation flow of the parallel prediction method for fluidized bed agglomeration and wear failure based on GATT-CNN-LSTM in this embodiment.

[0028] Figure 2 It is a schematic diagram of the principle of sensor arrangement in the fluidized bed of this embodiment.

[0029] Figure 3 Schematic diagram of the interaction between attrition failure and agglomeration failure in a fluidized bed.

[0030] Figure 4 FIG. 4 is a flow chart of fault type classification in this embodiment.

[0031] Figure 5 Schematic diagram of the structural principle of the GATT-CNN-LSTM model in this embodiment.

[0032] Figure 6 It is a schematic diagram of the model parameter configuration effect of the GATT-CNN-LSTM model in a specific application embodiment of the present invention. DETAILED DESCRIPTION

[0033] The present invention will be further described below in conjunction with the accompanying drawings and specific preferred embodiments, but the scope of protection of the present invention is not limited thereby.

[0034] For ease of understanding, the relevant technical background of the present invention is first introduced by way of example.

[0035] Fluidized beds are multiphase flow systems with complex structures and diverse internal conditions. A single fault often triggers a chain reaction, triggering abnormalities in multiple parameters simultaneously. For example, in a fluidized bed boiler, caking can lead to: localized high temperatures → accelerated oxidation of the heated surface → oxide layer spalling, exacerbated wear → wear further promoting caking. Therefore, faults within a fluidized bed are often not isolated. Monitoring only for a single fault type can easily overlook the risk of synergistic deterioration among different faults. Furthermore, different fault types often exhibit distinct fault signatures. Multiple sensors may be required to monitor a given fault type, but changes in some parameters may not be obvious during a fault, preventing some sensors from detecting the abnormal condition. Consequently, the sensitivity of different sensor types to different fault types varies. For example, caking primarily manifests itself as temperature fluctuations, so sensors other than temperature sensors, such as vibration and acoustic sensors, may not be able to detect the abnormal condition. In contrast, wear failures may exhibit smaller temperature changes than caking, while larger mass flow rate changes, necessitating the use of mass flow meters for monitoring. Agglomeration faults and wear faults may exist simultaneously in gasifiers and combustion furnaces. Monitoring by a single type of sensor cannot fully reflect the fault status in the gasifiers and combustion furnaces. If multiple types of sensors are used, the data monitored by different sensors have different sensitivities to different fault types, and it is still difficult to determine the specific fault type and fault severity. Some early faults may not be obvious in the original sensor data set. On the one hand, early faults are often hidden and their development is relatively mild. They cause little disturbance to the physical and chemical processes and fluid dynamics in the fluidized bed, making it difficult to form significant characteristic changes in the original data of conventional monitoring parameters such as temperature, pressure, and wind speed. On the other hand, the fluidized bed system itself is highly complex, and the various parameters are interrelated and interfere with each other. The weak abnormal signals caused by early faults are easily masked by the background noise during normal operation and the coupled fluctuations between parameters, making it difficult to directly identify and extract these fault characteristics in the unprocessed original data set. The limitations of the sensor's own accuracy, resolution, and data acquisition frequency also further aggravate the inconspicuousness of early faults in the original sensor data set. Therefore, it is difficult to identify early faults directly based on the original sensor data.

[0036] Fluidized bed agglomeration failure and heating surface wear are two common types of failures in circulating fluidized beds. Fluidized bed agglomeration failure and heating surface wear are mutually influential and interrelated processes. The two will induce each other in causes and promote each other in the development process. Agglomeration in a fluidized bed disrupts the uniform fluidization of the bed material, forming high-velocity airflow channels or jet zones around or above it. These high-speed, hard bed material particles (such as quartz sand, limestone, and ash) carry strong erosion against the nearby heating surface, increasing the frequency and energy of particle impacts. During agglomerate formation, these particles may encapsulate or adhere to harder mineral particles (such as quartz). When the agglomerates break or flake off, these hard particles are released and scour the heating surface with the airflow, exacerbating wear. This wear process, in turn, produces a large number of fine metal and ash particles. These fine particles have a larger specific surface area and stronger cohesive properties. At relatively low temperatures, these fine particles react more readily with low-melting-point ash components, such as alkali metals, chlorine, and sulfur, forming a viscous low-temperature eutectic that acts as the "glue" for agglomerates. Continued abrasion reduces the average particle size of the bed material and increases the proportion of fine particles, which in turn facilitates contact and bonding between particles, thereby increasing the risk of agglomeration. Agglomerates covering or near the heating surface can severely hinder heat transfer efficiency in that area, leading to local increases in flue gas and wall temperatures. That is, the formation of heated surface wear failure and agglomeration failure is a gradual development and evolution process, and in the initial stage when the failure occurs, the symptoms are generally not obvious. Such a gradual change is sometimes difficult to be intuitively reflected through operating monitoring parameters. The present invention takes into account the interaction between agglomeration and wear failures and the risk of synergistic deterioration between the two failures. By simultaneously predicting agglomeration and wear failures in the fluidized bed, the temperature data and mass flow rate of the gasifier and the combustion furnace at different monitoring positions are collected under different fault degree conditions (normal, early failure, serious failure) to form a fault data set. The temperature change parameters, mass flow rate change rate and degree of deviation from the mean are extracted from the fault data set as fault characteristics to form a fault feature sequence at different time points. The fault degree index value of the fault feature sequence at each time point is first calculated to judge the fault degree. When it is judged to be in a fault condition, the fault type is further identified according to the contribution state of each parameter in the process of calculating the fault degree index value, and then the fault degree label and fault category label are configured accordingly, which can fully explore the relationship between different characteristic agglomeration failures, wear failures and non-characteristic The fault prediction classification model is trained using a fault feature dataset with two labels. This not only classifies faults but also predicts their severity, achieving parallel prediction of fault type identification and fault severity prediction. This facilitates making corresponding maintenance decisions for different fault severity levels of fluidized bed agglomeration and wear faults, thereby reducing maintenance costs. At the same time, the fault prediction classification model adopts a GATT-CNN-LSTM architecture and uses the Gaussian convolution kernel temporal attention layer GATT, CNN layer, and LSTM layer for feature extraction. By sharing feature extraction and retaining the requirements of different tasks (classification and prediction tasks), it can fully utilize the advantages of convolutional neural networks in data feature extraction and the advantages of recurrent neural networks in capturing temporal features, thereby improving the overall predictive ability of the model.

[0037] The present invention will be further described below with reference to specific embodiments.

[0038] like Figure 1 As shown, the steps of the parallel prediction method for fluidized bed agglomeration and wear failure based on GATT-CNN-LSTM in this embodiment include: Step S01, fault data collection: collecting state parameters of the gasifier and combustion furnace at different monitoring positions under different fault severity conditions of the hot dual circulating fluidized bed system to form a fault data set, the state parameters including temperature data and mass flow rate.

[0039] Boiler caking affects the boiler's heat transfer efficiency, reducing its thermal efficiency and causing leakage and separation. Caking typically occurs in areas with uneven gas distribution, large temperature gradients, or long material residence times. For example, it primarily occurs 1-5 meters above the connection between the water-cooled wall and the wear-resistant material, and at the flue gas bend at the furnace outlet. In this embodiment, the collected temperature data may include cross-sectional temperatures at multiple heights within the gasifier at the distance from the burner's air distribution plate, temperature data within the gasifier's horizontal flue over time, and temperature data within the burner's horizontal flue over time. The collected mass flow rate includes the mass flow rate within the burner's horizontal flue over time and the mass flow rate within the burner's horizontal flue over time. Early failure conditions include early heating surface wear failure in the gasifier and early caking failure in the burner. Severe failure conditions include severe heating surface wear failure in the gasifier and severe caking failure in the burner. In this embodiment, the state parameters of the gasifier and burner at different monitoring locations are collected while the hot dual circulating fluidized bed system is in normal operating conditions, early failure conditions, and severe failure conditions.

[0040] like Figure 2 As shown, considering that agglomeration usually occurs in areas with uneven gas distribution, large temperature gradient or long material residence time, and the particle concentration in the dense phase area of ​​the fluidized bed combustion furnace is high, if the local temperature is too high (such as the high temperature area of ​​the combustion furnace), the ash or low melting point material will melt and stick to the particles to form agglomerates. In this embodiment, a temperature sensor is arranged at a height of 3200mm from the combustion furnace air distribution plate (cross-section modeling reference) in the gasification furnace, corresponding to position ①, to associate high-temperature coking failure; considering that the gas distribution plate is the core component of the fluidized bed, it is responsible for evenly distributing the gas to the bed layer. If the distribution plate is poorly designed (such as uneven opening rate) or blocked, it will cause the local gas velocity to be too low, forming a "dead zone" and the particles Deposition here and agglomeration due to high temperature or the presence of sticky substances may trigger low-temperature coking in the gasifier. In this embodiment, a temperature sensor is arranged in the cross section 1450mm away from the air distribution plate in the combustion furnace, corresponding to position ②, to associate the agglomeration fault; considering that the particle concentration in the dilute phase zone is low, but if the gas carries fine particles and settles here due to the reduction in flow rate, or collides with the reactor wall and adheres, soft agglomerates may be formed, that is, the wear fault of the horizontal flue heating surface, including the combustion furnace and the gasifier. In this embodiment, mass flow sensors and temperature sensors are arranged in the horizontal flues of the combustion furnace and the gasifier, respectively, corresponding to positions ③ and ④, to associate the wear fault of the horizontal flue heating surface.

[0041] In a specific application embodiment, for the fault classification and fault degree prediction of a hot double circulating fluidized bed, 1450 mm from the combustion furnace, 3200 mm from the combustion furnace, the horizontal flue of the combustion furnace, and the horizontal flue of the gasifier are taken as key monitoring positions, and the following data are collected at different measurement time points to form a data set, as shown in Table 1: The temperature data set fg-3200-t of the section at a height of 3200 mm from the air distribution plate of the combustion furnace in the hot full cycle of a double circulating fluidized bed system with a thermal power of 1 MW; the temperature data set fg-3200-t of the section at a height of 1450 mm from the air distribution plate of the combustion furnace in the hot full cycle of a double circulating fluidized bed system with a thermal power of 1 MW; The data set fc-1450-t shows the average temperature of the cross section at 50mm; the data set fg-out-t shows the temperature changing over time in the horizontal flue of the gasifier during the full hot cycle of a double circulating fluidized bed system with a thermal power of 1MW; the data set fc-out-t shows the temperature changing over time in the horizontal flue of the combustion furnace during the full hot cycle of a double circulating fluidized bed system with a thermal power of 1MW; the data set fc-out shows the mass flow rate changing over time in the horizontal flue of the combustion furnace during the full hot cycle of the double circulating fluidized bed system; the data set fg-out shows the mass flow rate changing over time in the horizontal flue of the gasifier in the results of the full hot cycle of the double circulating fluidized bed system.

[0042] Table 1: Basic characteristics of the dataset

[0043] Step S02, fault feature extraction: extract the change value of the state parameter of each monitoring location at different measurement time points and the degree of deviation from the mean value to obtain the fault feature according to the fault data set. The fault features corresponding to each monitoring location at the same measurement time point constitute a set of fault feature sequences, and a set of fault feature sequences is obtained at each measurement time point.

[0044] Temperature and mass flow rate can effectively reflect the boiler combustion conditions and heat transfer status, and are closely related to the agglomeration phenomenon. Abnormal and excessively fast temperature rise or fall rates can indicate a possible sudden deterioration in local heat transfer efficiency, and can mark and amplify local thermal imbalances. Wear failures are directly related to the impact frequency, particle kinetic energy, and temperature: the higher the temperature → the number of particles hitting the heated surface per unit time increases → the number of particles hitting the heated surface per unit time increases → the air flow velocity increases → the particle impact velocity increases → the kinetic energy of a single particle increases significantly, and the wear becomes more severe. In order to reflect the gradual development and evolution of agglomeration and wear failures, the temperature change parameter in this embodiment uses the absolute value of the temperature change acceleration. By extracting the absolute value of the temperature change acceleration TC and the rate of change of mass flow rate FC , obtain the changing trend of temperature and mass flow rate at each monitoring position in the gasifier and combustion furnace. Using the absolute value of temperature change acceleration TCIt can indicate a sudden deterioration of local heat transfer efficiency, while wear is usually a gradual and slow-changing process, and the resulting mass flow rate change rate is usually small and relatively stable. By detecting a continuous and small mass flow rate change rate under steady-state conditions, it can be used as an early signal of wear expansion. Specifically, the absolute value of the temperature change acceleration TC The calculation expression is: (1) Where: ∆t is the time difference between the initial and final moments, represents the temperature of the jth monitoring location at the i-th measurement time point.

[0045] The mass flow rate change rate can represent the wear rate, and because it affects both the impact frequency and the particle velocity, it is a nonlinear amplification. In this embodiment, the mass flow rate change rate is the mass of fluid or particles passing through a certain cross section within a certain period of time, and the calculation expression is: (2) Where: FC is the mass flow rate change rate, m i,j For the j The monitoring location is i The mass flow rate at each strategic time point.

[0046] In this embodiment, the deviation from the mean values ​​include the temperature deviation from the mean value and the mass flow rate deviation from the mean value, which can indicate the severity of the fault, wherein the larger the deviation value, the greater the severity of the fault. For example, data with excessive deviation from the mean and mean values ​​for temperature may indicate abnormal operation of the fluidized bed, which may indicate an early stage of agglomeration, while the deviation from the mean value for mass flow rate may indicate a wear fault.

[0047] Specifically, the calculation expression for the degree of temperature deviation from the mean is: (3) in, represents the temperature of the jth monitoring location at the i-th measurement time point, M Indicates the number of monitoring locations, N Indicates the total number of measurement time points, It represents the average temperature of all measurement time points.

[0048] The calculation expression of the degree of deviation of mass flow rate from the mean is: (4) in, represents the mass flow rate of the jth monitoring position at the i-th measurement time point, M Indicates the number of monitoring locations, NIndicates the total number of measurement time points, It represents the average value of mass flow rate at all measurement time points.

[0049] This embodiment analyzes the structure of the fluidized bed to identify the main associated sensors with high sensitivity to specific fault types. After placing corresponding sensors at positions ①, ②, ③, and ④ of the gasifier and combustion furnace, the absolute value of the temperature change acceleration at each position is continuously monitored. TC , the degree of temperature deviation from the mean TD , mass flow rate change rate FC and the degree to which the mass flow rate deviates from the mean FD , forming fault characteristics, as shown in Table 2. Based on the sensor data, the development process of the gasifier and combustion furnace agglomeration faults and the heating surface wear faults can be continuously monitored to identify the fault extent of the gasifier and combustion furnace agglomeration faults and the heating surface wear faults.

[0050] Table 2: Fault characteristics table

[0051] Step S03, fault data classification: Calculate the corresponding fault severity index value based on the fault feature sequence at each measurement time point, determine whether it is in a faulty working condition and determine the fault severity and fault category based on the fault severity index value, and configure the fault severity label and fault category label corresponding to each fault feature sequence based on the judgment results of the fault severity and fault category to obtain a labeled fault feature data set.

[0052] When agglomeration or heated surface wear occurs, the absolute value of temperature change acceleration and mass flow rate characteristics may change abnormally, and the temperature and mass flow rate may also deviate significantly from the normal mean. Therefore, by combining the absolute value of temperature change acceleration, mass flow rate and the degree of deviation from the mean, we can fully explore the correlation between different sensors and different fault types. Wear faults and agglomeration faults interact with each other, such as Figure 3 As shown, it can be divided into the following three action chains: The triggering (starting point) of the agglomeration failure: local agglomeration → heat transfer is blocked → TC Abnormal temperature rise (sudden rise in local temperature); uneven temperature distribution → TD Sustained increase (long-term deviation from the mean) → FD abnormal; Induction of wear failure: abnormal temperature ( TC / TD ↑) → Intensified thermal stress on the heated surface → Material fatigue → Accelerated wear; Expansion of the wear area → Turbulent flow in the flue → Abnormal FC fluctuations → TC abnormal; Regeneration of new agglomerates: Abnormal mass flow rate ( FC / FD ↑) → uneven particle flow → local material accumulation; high-temperature sintering in the accumulation area → new agglomerates formed → feedback to stage 1.

[0053] Based on the above characteristics, this embodiment forms a fault degree index value by weighted summing the absolute value of temperature change acceleration, the degree of temperature deviation from the mean, the mass flow rate change rate, and the degree of mass flow rate deviation from the mean. R , according to the fault degree index value R Judging the extent of a fault can accurately characterize the extent of the fault in its development process.

[0054] In this embodiment, calculating a corresponding fault severity index value according to each fault feature sequence, and determining whether a fault condition exists according to the fault severity index value and, if a fault condition exists, determining the fault severity specifically includes: Step S311. Calculate the fault severity index: 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 positions in the fault feature sequence. R , fault severity index R The calculation model includes agglomeration fault calculation items, wear fault calculation items and cross-feedback calculation items. The agglomeration fault calculation item is calculated based on the temperature conversion parameter and the corresponding deviation from the mean value. The wear fault calculation item is calculated based on the mass flow rate change rate and the corresponding deviation from the mean value. The cross-feedback calculation item is calculated based on the temperature conversion parameter and the deviation from the mean value of the mass flow rate change rate to reflect the cross-influence between agglomeration fault and wear fault. That is, the agglomeration fault calculation item is calculated based on the temperature parameter (including the absolute value of the temperature change acceleration) TC and the degree of temperature deviation from the mean TD ) is used to characterize the contribution of agglomeration failure to failure, and the wear failure calculation item is composed of mass flow rate parameters (including mass flow rate change rate FC and the degree of deviation of mass flow rate from the mean FD ) is used to characterize the contribution of wear fault to the fault, and the cross-feedback calculation term is constructed by the total temperature transformation parameter (including the temperature change rate) and the degree of deviation of the mass flow rate from the mean value, which is used to characterize the contribution of the cross-influence between agglomeration fault and wear fault to the fault.

[0055] Specifically, the failure degree index R The calculation model can be expressed as: (5) (6) in, R is the fault degree index, ranging from (0 to 1), 、 are the absolute values ​​of temperature change acceleration at two different locations in the gasifier away from the combustion furnace air distribution plate, 、 are the absolute values ​​of temperature change acceleration at the horizontal flue position of the gasifier and the combustion furnace, respectively. 、 are the mass flow rate change rates at the gasifier horizontal flue position and the combustion furnace horizontal flue position, respectively. ~ is the degree of temperature deviation from the mean at two different locations in the gasifier, the gasifier horizontal flue position, and the combustion furnace horizontal flue position. 、 R is the degree of deviation of the mass flow rate at the gasifier horizontal flue position and the combustion furnace horizontal flue position from the mean, TD is the degree of deviation of the total temperature at each location from the mean, R TC is the total temperature change rate at each location, R FC R is the total mass flow rate change rate in the gasifier horizontal flue and the combustion furnace horizontal flue, FD is the degree of deviation of the total mass flow rate at the gasifier horizontal flue position and the combustion furnace horizontal flue position from the mean, 、 They are the global basic weight coefficients, for example, they can be initially set to 0.4, γ The weight is used to reflect the mutual aggravating effect of different features in evaluating and quantifying agglomeration and wear, and can be initially set to 0.2, for example.

[0056] The fault degree index constructed as above formula (5) R , the fault severity index R It is a continuous value between 0 and 1. The closer the value is to 0, the closer the device status is to normal. The closer the value is to 1, the further the device deviates from the normal status and the higher the potential fault level. This continuous fault level index R is used as the target variable of the subsequent fault classification prediction model. The contribution to characterizing agglomeration failure, the second term Used to characterize the wear failure sharing effect, the third term is a cross term that can be directly quantified: Temperature anomaly (TC↑) and flow continuity disruption (FD↑) work together to accelerate the formation of new agglomerates. For example: when and At the same time, when the threshold is exceeded, the γ term will significantly amplify the R value, which can realize dynamic feedback capture; When , it indicates that it is in the agglomeration-dominated mode, that is, agglomeration failure is dominant, and the R feature is characterized as: TC 、 TDRapid rise, FC 、 FD Delayed response; when When , it indicates that it is in the wear-dominated mode, that is, wear failure is dominant, and the corresponding R feature is characterized as: FC , FD Sudden increase, TC , TD Rising slowly.

[0057] Step S312. Fault degree determination: Determine whether it is a normal operating condition or a faulty operating condition based on the fault degree index R, and if it is a faulty operating condition, whether it is an early fault or a serious fault.

[0058] In this embodiment, if the fault degree index value R If the fault index value is less than the preset early fault threshold, it is judged as normal working condition and the data tag is configured as normal state. R If the fault severity index value 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 label is configured as an early fault. R If the value 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 a severe fault.

[0059] For example, a fluidized bed is typically in a normal state, with normal data accounting for approximately 90%. Early-stage faults need to be increased to 20% (normal + early-stage ≈ 90%), and severe faults account for 10%. Therefore, three key thresholds can be set to classify fault levels. The specific R value is related to the severity of each fault. When the R value is less than 0.8 (the early-stage fault threshold), it indicates normal data, and the fault severity label is set to 0 to represent normal data. A threshold of 0.9 is set as the severe fault threshold. When 0.8 ≤ R < 0.9, it indicates an early-stage fault, and when R ≥ 0.9, it indicates a severe fault, as shown in Table 3. Furthermore, when R is greater than 0.8, further fault classification is required.

[0060] Table 3: Fault severity index values ​​for different fault severity levels R Range Table

[0061] In this embodiment, when the fault degree index value R When the value is less than the early fault threshold (such as 0.8), it means it is normal data and the fault label will be set to 0, that is, normal data. R Triggering early failure thresholds (e.g. R ≥0.8), enter the fault classification judgment, such as Figure 4 As shown in FIG, the specific steps of determining the fault category according to the contribution status of each parameter in the fault degree index value calculation process include: Step S321. Dominant mode recognition: determine the fault degree index value of the agglomeration fault calculation item and the wear fault calculation item R If the contribution ratio of the agglomeration fault calculation item is greater than the preset contribution ratio threshold, it is determined to be the agglomeration fault dominant mode, and the fault label is set to agglomeration fault. If the contribution ratio of the wear fault calculation item is greater than the preset contribution ratio threshold, it is determined to be the wear fault dominant mode, and the fault label is set to wear fault. If the contribution ratios of the agglomeration fault calculation item and the wear fault calculation item are balanced, then the process goes to step S322 for secondary judgment. The temperature parameter includes the absolute value of the temperature change acceleration. TC and the degree of temperature deviation from the mean TD , mass flow rate parameters include mass flow rate change FC and the degree of deviation of mass flow rate from the mean FD .

[0062] Taking formula (5) as an example, if the first term The value of the fault degree index R If the proportion of exceeds the preset threshold (such as 70%), it is determined that the agglomeration fault is dominant, and the agglomeration fault with fault labels 1 and 2 is set. If the second The value of the fault degree index R If the proportion of exceeds the preset threshold (such as 70%), it is determined that the wear fault is dominant and the fault label is set as the wear fault of position 3 and 4; if the first item The value of the second The value of the fault severity index R If the proportion of the values ​​is balanced (e.g. 40% to 60%), a secondary judgment is required.

[0063] Step S322. Determine the fault degree index value R The proportion of cross feedback items in the calculation model of R If the proportion of the cross feedback item is greater than the preset double fault judgment threshold, it is judged as a double fault of agglomeration and wear and triggers an early warning; if the cross feedback item has a negative impact on the fault degree index value R If the proportion is less than the preset single fault judgment threshold, it is determined to be a single fault, and the fault type is determined based on the fault corresponding to the larger item in the contribution ratio of the agglomeration fault calculation item and the wear fault calculation item.

[0064] Taking formula (5) as an example, we can judge Fault severity index RIf the proportion is greater than the preset double fault judgment threshold (such as 30%), it is judged as a double fault of agglomeration and wear (i.e., agglomeration fault and wear fault exist at the same time) and an early warning is triggered. At the same time, countermeasures are initiated, such as adjusting the fluidization wind speed, injecting inert particles, etc.; if the cross feedback item has an impact on the fault degree index R If the proportion of is less than the preset single fault judgment threshold (such as 10%), the corresponding fault type is determined according to the contribution ratio of the agglomeration fault calculation item and the wear fault calculation item. For example, if the contribution ratio corresponding to the agglomeration fault calculation item is greater than the contribution ratio corresponding to the wear fault calculation item, it is determined to be agglomeration fault; otherwise, if the contribution ratio corresponding to the wear fault calculation item is greater than the contribution ratio corresponding to the agglomeration fault calculation item, it is determined to be a wear fault.

[0065] Furthermore, in this embodiment, when it is determined that there is a dual fault of agglomeration and wear, the weights in the cross-feedback items are dynamically adjusted according to the total temperature change rate of each position and the degree of deviation of the total mass flow rate from the mean. γ The development of agglomeration and wear failures requires a long process, and there will be mutual influence between the two failures. In the early stage, the mutual influence is small, and as the degree of the failure continues to deepen, the mutual influence will gradually deepen. Based on this, the weight of the cross feedback item in this embodiment is γ The initial value of can be configured as a small value. When it is judged that the dual fault of agglomeration and wear occurs, it indicates that there is a high probability that there will be a mutual influence between the two faults, and the mutual influence will gradually increase, so the weight of the cross feedback item is correspondingly increased. γ The value of can be dynamically updated based on the real-time fault diagnosis results. γ , thereby more accurately representing the real-time fault degree and improving the accuracy of fault degree calculation.

[0066] Specifically, the weights in the cross-feedback terms can be dynamically adjusted according to the following formula: γ : (7) in, Express The adjusted weights, R TC is the total temperature change rate at each location, R FD is the degree of deviation of the total mass flow rate at the gasifier horizontal flue position and the combustion furnace horizontal flue position from the mean, is the total temperature change rate The maximum value of is the degree of deviation of the total mass flow rate from the mean The maximum value of .

[0067] As shown in formula (7), by increasing Item, can be increased The value of is used to increase the importance of cross calculation items when agglomeration and wear faults occur. When a fault occurs, the total temperature change rate R TC and / or the total mass flow rate deviates from the mean R FD will gradually increase accordingly. The value of the item will also gradually increase, so that the weight can be gradually increased as the fault severity becomes increasingly serious. γ value, thereby effectively improving the accuracy of fault degree calculation.

[0068] By following the above method, when >Basic items γ When a certain proportion (such as 30%) of the temperature and mass flow anomalies continue to coexist, the basic term γ will automatically increase, thereby increasing the warning priority. Make the weight of feedback item strengthened.

[0069] This embodiment uses the above method to classify fault types and determine fault severity, which can further ensure that the diagnosis results are consistent with the actual process status. At the same time, by constructing a continuous fault severity index, it can achieve fine quantification of fault severity rather than simple binary or multivariate classification.

[0070] Step S04: Fault prediction model training: Use the labeled fault feature dataset to train the pre-built GATT-CNN-LSTM fault prediction classification model.

[0071] like Figure 5 As shown, the GATT-CNN-LSTM fault prediction and classification model in this embodiment includes a feature extraction module and a fault classification branch and a fault severity prediction branch respectively connected to the feature extraction module. The feature extraction module includes a Gaussian convolution kernel temporal attention layer GATT, a CNN layer, and an LSTM layer connected in sequence. The fault classification branch performs the fault type classification task and outputs the fault type result. The fault severity prediction branch performs the fault severity prediction task and outputs the fault severity prediction result. The fault category judgment signal (dominant mode signal) obtained from each fault classification is returned to the Gaussian convolution kernel temporal attention layer GATT layer for input into the next fault diagnosis. The previous fault category judgment result is enhanced by the GATT layer. The enhanced feature is then passed through the LSTM module to predict the current fault severity value, and then fed back to the fault classification decision module for fault type judgment.

[0072] This embodiment builds a GATT-CNN-LSTM fault prediction classification model to parallelly predict the fault type and fault severity of agglomeration faults and heating surface wear faults in fluidized bed gasifiers and combustion furnaces. Figure 6As shown in the figure, after the model is processed by the Gaussian attention layer, the CNN layer, and the LSTM layer, the key features are extracted. Then, these data enter the classification task branch and the prediction task branch. In the classification branch, the model passes through two layers of CNN and a fully connected layer, and finally outputs the classification result; in the prediction task branch, after passing through two layers of LSTM and one fully connected layer, the predicted continuous fault indicator is finally output. The specific model parameters are shown in the figure. Figure 6 As shown, the structure of each layer is as follows: Input layer: The input layer is used to receive the normalized and serialized feature data. Its main purpose is to format the input data into a shape that the model can process. The experimental data is merged into the dataset to construct 12 features: temperature change, absolute value of acceleration TC 1- TC 4. Temperature mean 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. Take 50 data points as a sample, set the batch size of column processing 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, and F is the number of features, F=12. Two values ​​are constructed for this sample (the value used as the prediction value and the value used as the fault classification value). These two values ​​are the fault value at the last time point plus 1, and the fault classification value. For example, if the first sample is obtained at time points 1-50, then the values ​​of this sample are the fault value at time point 51 and the fault classification value at time point 51.

[0073]

[0074] Specifically, the Gaussian convolution kernel-based temporal attention layer GATT processes temporal data by combining Gaussian convolution kernels and feature-level attention mechanisms. Since there is a strong temporal nature in the sampling of CFB data, and the target data comes from the function mapping of data from a certain period of time in the past, the model hopes to select the temporal segment of the feature. The Gaussian convolution kernel is a one-dimensional filter based on Gaussian distribution, which is used to smooth the signal. The variable is regarded as the representation of data at different times in the same high-dimensional feature space. For each input feature of the temporal attention layer, there is a corresponding Gaussian convolution model for calculation, and the models are independent of each other. This embodiment uses the GATT layer to learn the fault degree index. R The weight coefficient in 、 , which is mathematically defined as: (8) in, Is an indicator of the degree of failure R The weight coefficient in the calculation model 、 , σ is the standard deviation of the Gaussian distribution, which controls the degree of smoothing. The data set is discretized into a one-dimensional array using the Gaussian kernel: (9) Where i is the array index and c is the center position of the kernel ( ), and then normalize it so that the sum of the weights of the convolution kernel is 1.

[0075] In this embodiment, the data set is discretized into a one-dimensional array using a Gaussian kernel, including 12 feature dimensions (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 deviates from the mean FD 1- FD 2) and the calculation model of the fault degree index value as shown in formula (5) and (6) 、 As the global basic weight coefficient, it can be initially set to 0.4. The γ weight that reflects the mutual aggravation effect of different features in evaluating quantization agglomeration and wear is initially set to 0.2. The above parameters are used as the learnable initial values ​​of the attention layer for dynamic optimization during training.

[0076] 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 the attention weight for each feature dimension separately. For each feature f, the attention calculation process is, Extract features: (10) Apply Gaussian convolution: (11) in, It is a one-dimensional convolution operation.

[0077] Apply a linear transformation and an activation function: (12) in, and are learnable scale and offset parameters.

[0078] The final output is the element-wise product of the original input and the attention weights: (13) Assume that the input data input sequence is , the output of the Gaussian convolution kernel temporal attention layer can be expressed as: (14) in 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, It is the center of the nucleus. is the normalized Gaussian convolution kernel weight, and It is a feature The learnable parameters.

[0079] This embodiment sets a temporal attention layer with the help of Gaussian convolution kernel, so that the model can focus on temporally adjacent data points and capture local temporal patterns. Each feature dimension has its own attention weight, which enables the model to dynamically adjust the degree of attention to different features. The Gaussian convolution kernel attention layer is located after the input layer and before the CNN layer. By relying on learning attention weights, the model can automatically select features that are more critical to the prediction task. At the same time, Gaussian convolution takes into account the local correlation between time steps and can also highlight key time steps and features.

[0080] Convolutional layer design CNN: In the GATT-CNN-LSTM fault prediction and classification model of this embodiment, there are two parts of convolutional layer operations. One is in the general feature extraction part, which is composed of a Gaussian convolution layer, a convolution layer, and an LSTM layer for jointly extracting features; the other convolutional layer operation is in the fault classification task of the model, used to complete the fault classification task.

[0081] In this embodiment, since the input data is a time series, a small convolution kernel is slid over the input sequence to detect short-term trends in the time series. The convolution kernel can effectively capture the translation-invariant features in the sequence data. The convolution kernel moves along the sequence dimension, that is, the time step, and calculates the weighted sum of the kernel and the corresponding input subsequence at each position. Assume that the input is The output of the layer (ignoring the batch dimension B, L is the sequence length, C is the number of channels), the convolution kernel ( is the kernel size, is the number of filters / number of output channels), and finally the output feature map is obtained .

[0082] For the output feature map channels, at position (step size S=1, padding='same' adjusts input boundaries): (15) in This is the offset required to achieve "same" padding. Outside the boundaries, padding is applied according to the "same" padding rule. The first convolutional layer, located after the Gaussian attention layer and before the LSTM layer, is used to extract the primary feature data after the Gaussian attention layer has enhanced it. It uses small convolution kernels to capture local features. The model structure and parameters of this CNN layer are shown in Table 4.

[0083] Table 4: Convolutional layer parameters of feature extraction module

[0084] Conv1D represents a one-dimensional convolution layer. When passing through the Gaussian convolution layer, the data input to the one-dimensional convolution layer is , the mathematical formula after the one-dimensional convolution layer is: (16) in, , is the convolution kernel weight, is the bias term, for in .

[0085] BatchNormalization is a batch normalization layer that can accelerate the training process, improve model stability, reduce sensitivity to initialization parameters, and has a certain regularization effect. It will perform a standardization operation on the activation value of each channel within the current mini-batch range, first calculating the mean of each feature channel in the mini-batch. and variance Then, the activation value of the channel is standardized using the calculated mean and variance to obtain the corresponding result Then, with the help of learnable scaling parameters and offset parameters, a linear transformation is applied to the normalized values, and the output is .

[0086] Activation is the activation layer, and the nonlinear activation function is a key component for introducing expressiveness into the network. Introducing nonlinearity can enable the network to learn more complex functional relationships. The activation function used in this embodiment is the ReLU function: (17) By adopting the above activation function, the calculation can be made efficient and the gradient disappearance can be alleviated.

[0087] MaxPooling1D is the maximum pooling layer. The pooling layer achieves feature compression by spatial downsampling. Its core functions are two-fold: on the one hand, it reduces the size of the feature map; on the other hand, it reduces computational complexity. Maximum pooling improves the spatial invariance of features by selecting local maxima, while average pooling smoothes local fluctuations and improves the generalization ability of the model. This embodiment uses the maximum pooling layer, namely: (18) in, is the input feature map at position The activation value of .

[0088] The DropOut layer reduces the dependence on specific neurons by randomly discarding some neurons, thereby achieving regularization and preventing overfitting: (19) After this layer, the original input data Shape becomes Assume that 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 .

[0089] The CNN in the fault classification task is placed in the classification task branch after the feature extraction module. It is used to further extract and refine patterns in the time series features from the data passed through the feature extraction module, further improving the accuracy of the fault classification task in this task branch. Specifically, the fault classification branch includes at least two CNN layers and one fully connected layer. Each model structure is similar to the CNN layer in the feature extraction module, namely, it has five model function layers: Conv1D, BatchNormalization, Activation, MaxPooling1D, and DropOut. The CNN layer at the input receives the features output by the feature extraction module and passes through each CNN layer and the fully connected layer to output 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 the features output by the feature extraction module and passes through each LSTM layer and the fully connected layer to output the fault severity prediction result. The specific model structure and parameters are shown in Tables 5 and 6. Table 6 represents a deeper CNN model, using a larger number of convolutional kernels, which can more deeply extract feature patterns from the data.

[0090] Table 5: Parameters of the first convolutional layer in the fault classification task

[0091] Table 6: Parameters of the second convolutional layer in the fault classification task

[0092] LSTM layer: After the convolution layer completes the extraction of relevant features in the feature extraction module, it enters the LSTM layer, which captures the long-term dependencies in the sequence data. The LSTM layer in the feature extraction module is located after the CNN layer. It is used to receive the local features extracted by CNN, capture the dependencies of these features in the time dimension, and improve the spatial features extracted by CNN through time series modeling. At the same time, it remembers and processes the long-term dependencies in the time series. After passing through this layer, the input data ( ) becomes ( ), the feature dimension increases from 64 to 128, while the batch size remains unchanged, After BatchNormalization and Dropout, it becomes , It will serve as input data for both the classification task and the regression task.

[0093] The LSTM layer in the fault severity prediction task branch consists of two layers. The first LSTM layer, after receiving data from the feature extraction module, extracts more refined time series features for the fault severity prediction regression task. This LSTM layer retains the sequence information and feeds it to the next LSTM layer for further processing. The second LSTM layer compresses the entire sequence information into a fixed-length vector representation, extracting the final high-level time series features.

[0094] In this embodiment, the GATT-CNN-LSTM fault prediction classification model includes two branches: fault classification and fault severity prediction. Two fully connected layers and two output layers are set as the input of the classification task and regression task respectively. The input data of the two branches are the feature extraction module. , integrated and nonlinearly transformed through the fully connected layer to help the model learn complex feature relationships. The output layer maps these features into the final prediction results, which can effectively predict the fault indicators. The fault classification task uses the softmax function to output the classification label value, and the fault degree prediction task uses the linear activation function to predict a continuous value to represent the prediction of the fault degree indicator value R.

[0095] Step S05, real-time fault parallel prediction: The state parameters of the gasifier and combustion furnace at different positions of the hot dual circulating fluidized bed system under test are collected in real time during operation, and the corresponding change values ​​and deviation from the mean are extracted to obtain a real-time fault feature sequence, which is input into the trained GATT-CNN-LSTM fault prediction and classification model to obtain the fault degree prediction result and the fault type classification result output.

[0096] Specifically, the temperature data of the gasifier and combustion furnace at various monitoring positions (position 1 to position 4) and the mass flow rate of the gasifier and combustion furnace at the designated positions during the operation of the fluidized bed system under test are collected in real time, and then the absolute value of the temperature change acceleration and the mass flow rate change rate are extracted to obtain a real-time fault feature sequence. After inputting it into the GATT-CNN-LSTM fault prediction and classification model trained in step S04, the fault degree prediction and fault type classification can be performed in real time and in parallel. It can not only determine whether a fault has occurred, but also determine the fault type, so that timely warning can be given in advance when the fault state is in the early stage. At the same time, it can also determine the specific fault type, classify the agglomeration faults and heating surface wear faults formed by different mechanisms, and accurately locate the specific location.

[0097] The present invention makes predictions through a hybrid deep learning model that integrates convolutional neural networks, Gaussian convolution kernel temporal attention layers, and long short-term memory networks. The Gaussian convolution kernel temporal attention layer is used to focus on key time steps and features, CNN extracts local patterns and spatial features in the temperature sequence, and LSTM captures long-term temporal dependencies. By constructing a shared feature extraction layer, the extracted features are input into the classification task branch and the prediction task branch respectively. This can parallelly classify agglomeration faults and heated surface wear faults that occur during fluidized bed operation, and predict the degree of normal, early, and severe faults. At the same time, it can identify the type of fluidized bed faults and continuously quantify the severity of the faults. This can enable fluidized bed operators to promptly handle early faults and take appropriate maintenance decisions for different types and degrees of faults, thereby reducing the operation and maintenance costs of fluidized bed equipment and ensuring equipment safety.

[0098] This embodiment further provides a computer device, including a processor and a memory, wherein the memory is used to store a computer program, and the processor is used to execute the computer program to perform the above method.

[0099] It is understandable that the above method of this embodiment can be executed by a single device, such as a computer or server, etc., and can also be applied to a distributed scenario and completed by multiple devices cooperating with each other. In the case of a distributed scenario, one of the multiple devices can only execute one or more steps in the above method of this embodiment, and multiple devices interact to complete the above method. The processor can be implemented in the form of a general-purpose CPU, a microprocessor, an application-specific integrated circuit, or one or more integrated circuits, etc., for executing relevant programs to implement the above method of this embodiment. The memory can be implemented in the form of a read-only memory ROM, a random access memory RAM, a static storage device, and a dynamic storage device. The memory can store an operating system and other application programs. When the above method of this embodiment is implemented by software or firmware, the relevant program code is stored in the memory and called and executed by the processor.

[0100] This embodiment further provides a computer-readable storage medium storing a computer program, which implements the above method when executed by a processor.

[0101] Those skilled in the art will appreciate that the above-described embodiments of the present invention may be provided as methods, systems, or computer program products. Therefore, the present invention may take the form of a completely hardware embodiment, a completely software embodiment, or an embodiment combining software and hardware aspects. Moreover, the present invention may take the form of a computer program product implemented 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 the flowcharts and / or block diagrams of the methods, devices (systems), and computer program products according to the embodiments of the present invention. It should be understood that each process and / or box in the flowchart and / or block diagram, as well as the combination of processes and / or boxes in the flowchart and / or block diagram, may be implemented by computer program instructions. These computer program instructions may be provided to a processor of a general-purpose computer, a special-purpose computer, an embedded processor, or other programmable data processing device to produce a machine, so that the instructions executed by the processor of the computer or other programmable data processing device generate instructions for implementing the functions described in the processes. Figure 1 a process or multiple processes and / or boxes Figure 1 These computer program instructions can also be stored in a computer-readable memory that can guide a computer or other programmable data processing device to work in a specific way, so that the instructions stored in the computer-readable memory produce a product including the instruction device, which implements the function specified in the process. Figure 1 a process or multiple processes and / or boxes Figure 1 These computer program instructions can also be loaded onto a computer or other programmable data processing device, so that a series of operation steps are executed on the computer or other programmable device to produce a computer-implemented process, thereby providing instructions for implementing the process in the process. Figure 1 a process or multiple processes and / or boxes Figure 1 A step that specifies a function in one or more boxes.

[0102] The above description is merely a preferred embodiment of the present invention and does not constitute any form of limitation to the present invention. Although the present invention has been disclosed above with reference to the preferred embodiment, it is not intended to limit the present invention. Therefore, any simple modifications, equivalent variations, and modifications to the above embodiment that do not depart from the technical solution of the present invention and are based on the technical essence of the present invention shall fall within the scope of protection of the technical solution of the present invention.

Claims

1. A parallel prediction method for fluidized bed agglomeration and wear failure based on GATT-CNN-LSTM, characterized in that the steps include: Step S01, fault data collection: collecting state parameters of the gasifier and the combustion furnace at different monitoring positions under different fault severity conditions of the hot dual circulating fluidized bed system to form a fault data set, wherein the state parameters include temperature data and mass flow rate; Step S02, fault feature extraction: extract the change value of the state parameter of each monitoring location at different measurement time points and the degree of deviation from the mean value to obtain the fault feature according to the fault data set, and obtain a set of fault feature sequences at each measurement time point; Step S03, fault data classification: Calculate the corresponding fault severity index value based on the fault feature sequence at each measurement time point, determine whether it is in a faulty working condition and determine the fault severity and fault category based on the fault severity index value, and assign a fault severity label and a fault category label corresponding to each fault feature sequence based on the fault severity and fault category judgment results to obtain a labeled fault feature dataset; Step S04, fault prediction model training: using the labeled fault feature dataset to train a pre-built GATT-CNN-LSTM fault prediction classification model; Step S05, real-time fault parallel prediction: The state parameters of the gasifier and combustion furnace at different positions of the fluidized bed system under test are collected in real time during operation, and the corresponding change values ​​and deviation from the mean are extracted to obtain a real-time fault feature sequence, which is input into the trained GATT-CNN-LSTM fault prediction and classification model to obtain the fault degree prediction result and the fault type classification result output.

2. The method for parallel prediction of fluidized bed agglomeration and wear failure based on GATT-CNN-LSTM according to claim 1 is characterized in that: In step S01, the fault degree conditions include normal operating conditions, early fault conditions and severe fault conditions. The collected temperature data include the cross-sectional temperatures at multiple different heights from the combustion furnace air distribution plate in the gasifier, the temperature data that changes with time in the horizontal flue of the gasifier, and the temperature data that changes with time in the horizontal flue of the combustion furnace. The collected mass flow rate includes the mass flow rate that changes with time in the horizontal flue of the combustion furnace and the mass flow rate that changes with time in the horizontal flue of the gasifier. The early fault conditions include early heating surface wear faults in the gasifier and early agglomeration faults in the combustion furnace. The severe fault conditions include severe heating surface wear faults in the gasifier and severe agglomeration faults in the combustion furnace.

3. The parallel prediction method for fluidized bed agglomeration and wear failure based on GATT-CNN-LSTM according to claim 1 is characterized in that: In step S03, calculating the corresponding fault degree index value according to the fault feature sequence includes: The fault severity index is calculated 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 positions in the fault feature sequence. R The temperature change parameter includes the absolute value of the temperature change acceleration, the deviation from the mean value includes the temperature deviation from the mean value and the mass flow rate deviation from the mean value, the fault degree index R The calculation model includes agglomeration fault calculation items, wear fault calculation items and cross feedback calculation items. The agglomeration fault calculation item is calculated based on the temperature conversion parameter and the corresponding deviation from the mean value. The wear fault calculation item is calculated based on the mass flow rate change rate and the corresponding deviation from the mean value. The cross feedback calculation item is calculated based on the total temperature conversion parameter of each position and the total mass flow rate deviation from the mean value. The fault degree index R The calculation model is: in, R is the fault degree index value, 、 are the absolute values ​​of temperature change acceleration at two different locations in the gasifier away from the combustion furnace air distribution plate, 、 are the absolute values ​​of temperature change acceleration at the horizontal flue position of the gasifier and the combustion furnace, respectively. 、 are the mass flow rate change rates at the gasifier horizontal flue position and the combustion furnace horizontal flue position, respectively. ~ is the degree of temperature deviation from the mean at two different locations in the gasifier, the gasifier horizontal flue position, and the combustion furnace horizontal flue position. 、 R is the degree of deviation of the mass flow rate at the gasifier horizontal flue position and the combustion furnace horizontal flue position from the mean, TD is the degree of deviation of the total temperature at each location from the mean, R TC is the total temperature change rate at each location, R FC R is the total mass flow rate change rate in the gasifier horizontal flue and the combustion furnace horizontal flue, FD is the degree of deviation of the total mass flow rate at the gasifier horizontal flue position and the combustion furnace horizontal flue position from the mean, 、 are the global basic weight coefficients, γ The weights are used to reflect the mutual aggravating effects of different features in evaluating and quantifying caking and wear.

4. The method for parallel prediction of fluidized bed agglomeration and wear failure based on GATT-CNN-LSTM according to claim 3 is characterized in that: In step S03, judging whether the system is in a fault condition and judging the fault degree according to the fault degree index value includes: if the fault degree index value R is less than the preset early fault threshold, it is judged as a normal condition and the data tag is configured as a normal state; if the fault degree index value R is greater than the preset early fault threshold and less than the preset severe fault threshold, it is judged as an early fault condition and the corresponding data tag is configured as an early fault; if the fault degree index value R is greater than the preset severe fault threshold, it is judged as a severe fault condition and the corresponding data tag is configured as a severe fault.

5. The method for parallel prediction of fluidized bed agglomeration and wear failure based on GATT-CNN-LSTM according to claim 3, characterized in that: In step S03, when a fault condition is determined, the fault type is determined based on the contribution status of each parameter in the fault degree index value calculation process, including: Step S321. Determine the fault degree index value of the agglomeration fault calculation item and the wear fault calculation item R If the contribution ratio of the agglomeration fault calculation item is greater than the preset contribution ratio threshold, it is determined to be the agglomeration fault dominant mode, and the fault label is set to agglomeration fault. If the contribution ratio of the wear fault calculation item is greater than the preset contribution ratio threshold, it is determined to be the wear fault dominant mode, and the fault label is set to wear fault. If the contribution ratios of the agglomeration fault calculation item and the wear fault calculation item are balanced, then the process goes to step S322 for secondary judgment, where the temperature parameter includes the absolute value of the 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 cross feedback items in the calculation model of R If the proportion of the cross feedback item is greater than the preset double fault judgment threshold, it is judged as a double fault of agglomeration and wear and triggers an early warning; if the cross feedback item has a negative impact on the fault degree index value R If the proportion is less than the preset single fault judgment threshold, it is determined to be a single fault, and the fault type is determined based on the fault corresponding to the larger item in the contribution ratio of the agglomeration fault calculation item and the wear fault calculation item.

6. The method for parallel prediction of fluidized bed agglomeration and wear failure based on GATT-CNN-LSTM according to claim 5, characterized in that: When it is determined that the fault is a combination of agglomeration and wear, the weights in the cross-feedback items are dynamically adjusted according to the total temperature change rate of each position and the degree of deviation of the total mass flow rate from the mean. γ , the calculation expression is: in, Express Adjusted weight, R TC is the total temperature change rate at each location, R FD is the degree of deviation of the total mass flow rate at the gasifier horizontal flue position and the combustion furnace horizontal flue position from the mean, is the total temperature change rate The maximum value of is the degree of deviation of the total mass flow rate from the mean The maximum value of .

7. The method for parallel prediction of fluidized bed agglomeration and wear failure based on GATT-CNN-LSTM according to any one of claims 1 to 6, characterized in that: The GATT-CNN-LSTM fault prediction and classification model includes a feature extraction module, a fault classification branch, and a fault severity prediction branch. The feature extraction module includes a Gaussian convolution kernel temporal attention layer GATT, a CNN layer, and an LSTM layer connected in sequence. The Gaussian convolution model used in the Gaussian convolution kernel temporal attention layer GATT in the GATT-CNN-LSTM fault prediction and classification model is: in, Is an indicator of the degree of failure R The weight coefficient in the calculation model 、 , σ is the standard deviation of the Gaussian distribution used to control the degree of smoothing; The output of the Gaussian convolution kernel temporal attention layer GATT is expressed as: in, 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 of the nucleus, is the normalized Gaussian convolution kernel weight, and It is a feature The learnable parameters.

8. The method for parallel prediction of fluidized bed agglomeration and wear failure based on GATT-CNN-LSTM according to claim 7, characterized in that: The fault classification branch includes at least two CNN layers and one fully connected layer. The CNN layer at the input end receives the features output by the feature extraction module, and outputs the fault type classification result after passing 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 LSTM layer at the input end receives the features output by the feature extraction module, and outputs the fault degree prediction result after passing through each LSTM layer and the fully connected layer.

9. A computer device comprising a processor and a memory, wherein the memory is used to store a computer program, wherein: The processor is configured to execute the computer program to perform the method according to any one of claims 1 to 8.

10. A computer-readable storage medium storing a computer program, characterized in that: When the computer program is executed by a processor, the method according to any one of claims 1 to 8 is implemented.

Images