Method and device for diagnosing differential pressure anomaly of flue gas filter dust collector

Abstract

This invention discloses a method and apparatus for diagnosing pressure difference anomalies in flue gas filter dust collectors, solving the problem in existing technologies where it is difficult to identify the specific causes of pressure difference increases in flue gas filter dust collectors based on measured pressure difference. The method includes: collecting multi-source operating data during the operation of the flue gas filter dust collector; determining the filtration velocity or air-to-cloth ratio based on the flue gas flow rate and the effective filtration area of ​​the flue gas filter dust collector, and normalizing the measured pressure difference based on the filtration velocity or air-to-cloth ratio to obtain normalized filtration resistance; constructing a cleaning event window using the cleaning action signal as the event anchor point, and extracting cleaning response features from the cleaning event window; constructing a pressure difference anomaly time-series feature matrix based on the normalized filtration resistance, cleaning response features, and multi-source operating data; and inputting the pressure difference anomaly time-series feature matrix into a pre-trained deep learning time-series attribution model to obtain the anomaly type.

Description

Technical Field

[0001] This invention relates to the field of flue gas filtration and dust removal technology, and in particular to a method and device for diagnosing abnormal pressure differences in flue gas filtration and dust removal equipment. Background Technology

[0002] Flue gas filtration dust collectors are commonly used dust removal equipment in industrial flue gas treatment. They capture particulate matter in flue gas through filter media to reduce the concentration of particulate matter emissions. Flue gas filtration dust collectors can include bag filters, cartridge filters, and plate filters. Bag filters use filter bags as the filter media; cartridge filters use filter cartridges; and plate filters use filter plates. As the flue gas filtration dust collector continues to operate, particulate matter will deposit on the surface of the filter media, forming a dust layer. The resistance encountered by the flue gas as it passes through the filter media and dust layer gradually increases, leading to an increase in the measured pressure difference of the flue gas filtration dust collector.

[0003] The measured differential pressure of a flue gas filter dust collector is a crucial parameter reflecting its operating status. On one hand, the measured differential pressure reflects the degree of dust accumulation on the filter media surface, changes in filtration resistance, and the flue gas throughput capacity. On the other hand, an excessively high measured differential pressure can lead to increased induced draft fan load, higher system energy consumption, and decreased flue gas treatment capacity. In severe cases, it may even cause insufficient flue gas flow, unstable production system operation, or abnormal outlet particulate matter concentration. Therefore, during the operation of a flue gas filter dust collector, it is usually necessary to continuously monitor the increase in differential pressure to determine whether the flue gas filter dust collector is in a normal filtration state, requires cleaning, or is operating abnormally.

[0004] In existing flue gas filtration dust collector control systems, an increase in the differential pressure within the flue gas filtration dust collector is typically used as a condition for initiating dust removal. Specifically, when the measured differential pressure reaches or exceeds a preset differential pressure threshold, the control system activates the dust removal actuator to clean the filter media. For flue gas filtration dust collectors employing a reverse-flushing dust removal mechanism, the control system can activate the reverse-flushing dust removal mechanism when the measured differential pressure reaches a preset differential pressure threshold; among these, a pulse-jet dust removal mechanism can be considered a specific form of reverse-flushing dust removal mechanism, and the control system can trigger pulse-jet dust removal action when the measured differential pressure reaches a preset differential pressure threshold. For flue gas filtration dust collectors employing a rapping dust removal mechanism, the control system can activate the rapping dust removal action when the measured differential pressure reaches a preset differential pressure threshold. Therefore, an increase in the differential pressure of the flue gas filtration dust collector is not only an object of operational status monitoring but also an important control basis in existing dust removal control strategies.

[0005] However, while an increase in differential pressure in a flue gas filter dust collector can serve as a condition for initiating dust removal, it doesn't directly indicate the specific cause of the increase. For flue gas filter dust collectors, a measured increase in differential pressure essentially reflects increased resistance as flue gas passes through the filter media and its surface deposits, but the mechanisms leading to this increased resistance vary. Some increases in differential pressure are normal changes during the filtration process, while others correspond to abnormal changes in equipment operating status, dust removal execution status, flue gas conditions, or filter media conditions. Therefore, in actual operation, although on-site personnel can detect an increase in differential pressure through monitoring, it is usually difficult to further determine whether the increase is due to increased resistance during normal operation or an abnormal increase requiring intervention based solely on the measured differential pressure.

[0006] Specifically, during normal filtration, a dust layer gradually forms on the surface of the filter media. The resistance of flue gas passing through the filter media and dust layer increases with filtration time, manifested as a gradual increase in the measured pressure difference. When the dust removal actuator operates normally, part of the dust layer on the surface of the filter media is removed, and the measured pressure difference decreases accordingly, before entering the next filtration cycle. This pressure difference change, characterized by periodic increases and subsequent decreases after dust removal, usually corresponds to normal filter cake formation and should not be simply identified as a malfunction.

[0007] Unlike normal filter cake formation, when the dust load increases, the rate of dust accumulation on the filter media surface may accelerate, resulting in a significant increase in the slope of the measured differential pressure rise. When the cleaning cycle is too long or the cleaning triggering strategy is unreasonable, the particulate matter deposited on the filter media surface cannot be removed in time, which may lead to insufficient cleaning. When the back-blowing air pressure, back-blowing air volume, jet pressure, air tank pressure, pulse valve failure, rapping motor abnormality, or cleaning feedback signal abnormality occur, even if the control system has issued a cleaning action signal, the cleaning actuator may not achieve an effective cleaning effect, thus causing a cleaning system failure.

[0008] In addition, when the flue gas humidity increases or the flue gas outlet temperature approaches the water dew point or acid dew point, the surface of the filter media may become damp, sticky, caked, or have condensed deposits, leading to abnormal clogging of the filter media. When the ash hopper level rises, the current of the ash discharge equipment is abnormal, or the ash discharge equipment operates abnormally, the ash accumulation in the ash hopper may cause ash hopper blockage or abnormal ash level, and affect the flow state within the flue gas filter dust collector. When the flue gas flow rate, the frequency of the induced draft fan, or the current of the induced draft fan increases, the measured differential pressure may also increase accordingly. However, this increase does not necessarily indicate that the filter media is blocked, but may be due to abnormal airflow caused by an increase in filtration velocity or air-to-cloth ratio. When the cumulative running time of the filter media is long, the baseline of the residual differential pressure after multiple cleaning cycles continues to rise, or the cleaning effectiveness decreases over a long period of time, it may also correspond to aging or permanent blockage of the filter media.

[0009] Therefore, an increase in measured differential pressure is a common manifestation of various operating and abnormal states, potentially corresponding to different causes such as normal filter cake formation, increased dust load, insufficient dust removal, dust removal system malfunction, abnormal filter media adhesion and blockage, ash hopper blockage or abnormal material level, abnormal airflow, or filter media aging or permanent blockage. Existing control methods that use an increase in the differential pressure of the flue gas filter dust collector as the activation condition for the dust removal actuator, while triggering the dust removal action when the measured differential pressure reaches a preset threshold, are essentially still differential pressure threshold-based control methods. This method primarily addresses the question of "when to start dust removal" and cannot further determine the cause of the increased measured differential pressure. The aforementioned different causes may all manifest as an increase in measured differential pressure, but their mechanisms and subsequent handling methods are different. If dust removal control or operational judgment is based solely on an increase in measured differential pressure, it can easily lead to inaccurate identification of the cause of the increased differential pressure in the flue gas filter dust collector and improper subsequent handling. Summary of the Invention

[0010] The purpose of this invention is to provide a method and device for diagnosing abnormal pressure differences in flue gas filters, solving the problem in the prior art that it is difficult to identify the specific cause of the increase in pressure difference in flue gas filters by measuring the actual pressure difference.

[0011] Firstly, a method for diagnosing abnormal pressure differences in a flue gas filter dust collector is provided. The flue gas filter dust collector is a dust removal device with a cleaning actuator and uses a filter medium to capture particulate matter in flue gas. The method includes: collecting multi-source operating data during the operation of the flue gas filter dust collector; the multi-source operating data includes at least measured pressure difference, flue gas flow rate, cleaning action signal, cleaning execution status data, flue gas temperature, flue gas humidity, dew point related data, ash hopper status, ash unloading equipment status, and outlet particulate matter concentration; the dew point related data includes at least one of water dew point temperature, acid dew point temperature, flue gas humidity data used to calculate the water dew point temperature, and acidic gas concentration data used to calculate the acid dew point temperature; and diagnosing abnormalities based on the flue gas flow rate and the flue gas filter dust collector... The effective filtration area of ​​the filter determines the filtration velocity or air-to-cloth ratio, and the measured pressure difference is normalized according to the filtration velocity or air-to-cloth ratio to obtain the normalized filtration resistance. A cleaning event window is constructed using the cleaning action signal as the event anchor point, and cleaning response features are extracted from the cleaning event window. These cleaning response features include at least the pressure difference before cleaning, the pressure difference after cleaning, the pressure drop amplitude after cleaning, the pressure drop ratio after cleaning, the pressure difference re-rise slope, and the residual pressure difference after cleaning. A pressure difference anomaly time-series feature matrix is ​​constructed based on the normalized filtration resistance, the cleaning response features, and the multi-source operating data. This matrix includes features from multiple moments within a continuous time window, where the continuous time window is set according to a preset window length and preset steps. The data is a multivariate time series segment extracted from the multi-source operational data; the features include ash accumulation rate features, ash removal effectiveness features, residual resistance features, condensation risk features, ash removal execution consistency features, ash hopper blockage or material level anomaly features, air volume anomaly features, and filter media aging or permanent blockage features corresponding to candidate anomaly types; the pressure difference anomaly time series feature matrix is ​​input into a pre-trained deep learning time series attribution model to obtain the anomaly type corresponding to the pressure difference increase of the flue gas filter dust collector and the attribution probability of each candidate anomaly type, and the anomaly confidence is determined based on at least one of the following: maximum attribution probability, attribution probability interval between different candidate anomaly types, input data integrity, and model output distribution; wherein, the attribution probability... The rate is obtained by probabilistic processing of the classification output of the pre-trained deep learning time-series attribution model for each candidate anomaly type; the classification output, also known as the unnormalized diagnostic score, refers to at least one of the classification score, logistic value, or decision function value output by the pre-trained deep learning time-series attribution model for each candidate anomaly type before probabilistic processing; the pre-trained deep learning time-series attribution model is a deep learning model that completes parameter training based on training samples with anomaly type labels before diagnosis; the candidate anomaly types include at least two of the following: normal filter cake formation, increased dust load, insufficient dust removal, dust removal system failure, abnormal filter media adhesion and clogging, ash hopper clogging or abnormal material level, abnormal air volume, and filter media aging or permanent clogging.The filter media adhesion and clogging anomalies include at least one of condensation adhesion clogging, moisture absorption caking clogging, acidic condensate adhesion clogging, and oil mist or tar adhesion clogging; based on the anomaly type, attribution probability, and anomaly confidence level, output pressure difference anomaly diagnosis results, diagnostic basis information, and corresponding operation and maintenance handling suggestions for the anomaly type.

[0012] Further, the flue gas filtration dust collector is a flue gas filtration dust removal device that uses at least one filter medium selected from filter bags, filter cylinders, and filter plates; the dust removal actuator includes at least one selected from reverse-blowing dust removal mechanism and rapping dust removal mechanism; wherein, the reverse-blowing dust removal mechanism includes at least one selected from pulse-jet dust removal mechanism and reverse-blowing air dust removal mechanism; when the dust removal actuator is a pulse-jet dust removal mechanism, the dust removal execution status data includes at least one selected from jet pressure, air tank pressure, pulse valve status, pulse valve current, and jet duration; when the dust removal actuator is a reverse-blowing dust removal mechanism or a rapping dust removal mechanism, the dust removal execution status data includes at least one selected from reverse-blowing air pressure, reverse-blowing air volume, rapping motor status, rapping current, rapping duration, and dust removal feedback signal.

[0013] Further, the filtration velocity or air-to-cloth ratio is determined based on the flue gas flow rate and the effective filtration area of ​​the flue gas filter dust collector, and the measured pressure difference is normalized based on the filtration velocity or air-to-cloth ratio to obtain the normalized filtration resistance. This includes: determining the filtration velocity or air-to-cloth ratio based on the flue gas flow rate and the effective filtration area of ​​the flue gas filter dust collector; the effective filtration area is the area of ​​the filter media in the flue gas filter dust collector in the filtration state, or the effective filtration area is the total filter media area minus at least one of the following: offline compartment area, isolated filter unit area, and damaged filter unit area; the normalized filtration resistance is obtained using a first normalization method; the first normalization method includes determining the normalized filtration resistance as the ratio of the measured pressure difference to the filtration velocity or air-to-cloth ratio; or The normalized filter resistance is obtained using a second normalization method. This second normalization method includes: determining a reference pressure difference corresponding to the current filtration velocity or air-to-closing ratio based on the fitting relationship between the pressure difference and the filtration velocity or air-to-closing ratio under historical normal operating conditions; and determining the deviation or proportion of the measured pressure difference from the reference pressure difference as the normalized filter resistance. When the measured pressure difference increases, the flue gas flow rate increases, but the normalized filter resistance does not increase synchronously, the attribution probability or unnormalized diagnostic score for abnormal airflow is increased, and the attribution probability or unnormalized diagnostic score for abnormal filter media adhesion and blockage and filter media aging or permanent blockage is decreased. The "unsynchronized increase" refers to a change in the normalized filter resistance that is less than the change in the measured pressure difference, or less than the synchronous change threshold obtained statistically from historical normal operating conditions.

[0014] Further, a dust removal event window is constructed using the dust removal action signal as the event anchor point, and dust removal response features are extracted from the dust removal event window, including: using the moment the dust removal action is triggered as the event anchor point; acquiring pressure difference data within a preset time period before the event anchor point to determine the pressure difference before dust removal; acquiring pressure difference data within a preset time period after the event anchor point to determine the pressure difference after dust removal; determining the dust removal drop amplitude and dust removal drop ratio based on the pressure difference before and after dust removal; determining the pressure difference rise slope and residual pressure difference after dust removal based on the pressure difference change trend after the event anchor point; and then... The dust removal drop amplitude, dust removal drop ratio, pressure difference rise slope, and residual pressure difference after dust removal are used to construct the dust removal effectiveness characteristics, residual resistance characteristics, and dust removal execution consistency characteristics. The dust removal response characteristics are mapped to the corresponding time in the continuous time window according to the corresponding dust removal event timestamp, or the dust removal response characteristics are filled, interpolated, maintained, or statistically aggregated within the dust removal event window to form the event response characteristic channel of the pressure difference abnormal time series characteristic matrix. The preset duration is determined based on the dust removal action response time, dust removal cycle, sampling cycle, or historical normal dust removal response data.

[0015] Furthermore, the process for determining the condensation risk characteristics includes: when the water dew point temperature is obtained, determining the water dew point margin based on the difference between the flue gas outlet temperature and the water dew point temperature; when the acid dew point temperature is obtained, determining the acid dew point margin based on the difference between the flue gas outlet temperature and the acid dew point temperature; when the acid dew point temperature is not directly obtained, estimating the acid dew point temperature based on at least one of SO2 concentration, estimated SO3 value, HCl concentration, and flue gas humidity, and determining the acid dew point margin based on the difference between the flue gas outlet temperature and the estimated acid dew point temperature; and determining the condensation risk characteristics based on the obtained or estimated water dew point margin, acid dew point margin, and flue gas humidity. The condensation risk characteristics, which characterize an increased condensation risk, include at least one of the following: water dew point margin being lower than a preset water dew point margin threshold, acid dew point margin being lower than a preset acid dew point margin threshold, and flue gas humidity being higher than a preset humidity threshold. The determination process for the consistency characteristics of the dust removal execution includes: determining whether the dust removal execution status data is within the preset normal range of the corresponding dust removal actuator after the dust removal action signal is triggered, or whether a state change matching the dust removal action signal occurs; determining whether the measured pressure difference or normalized filter resistance decreases by a preset magnitude after the dust removal action signal is triggered; and determining whether the dust removal execution status data is based on... The consistency characteristics of the dust removal process are determined based on the pressure difference decrease. The process for determining the characteristics of ash hopper blockage or abnormal material level includes: determining the characteristics of ash hopper blockage or abnormal material level based on the ash hopper status and the status of the ash unloading equipment, wherein the ash hopper status includes at least the ash hopper material level and the ash unloading equipment status includes at least the ash unloading equipment current or the ash unloading equipment operating status; based on the condensation risk characteristics, the consistency characteristics of the dust removal process, and the characteristics of ash hopper blockage or abnormal material level, the attribution probability or unnormalized diagnostic score of each candidate abnormality type is corrected, wherein the correction includes: when the condensation risk characteristics indicate an increased condensation risk, and the dust removal is effective. When at least one of the following conditions is met: the characteristic of decreased cleaning effect, the residual resistance characteristic of increased residual resistance after cleaning, and the consistency characteristic of normal cleaning operation, the attribution probability or unnormalized diagnostic score of abnormal filter media adhesion and blockage is increased; when there is a cleaning action signal but the cleaning operation status data is abnormal, or the pressure difference decrease is lower than a preset threshold, the attribution probability or unnormalized diagnostic score of cleaning system failure is increased; when the ash hopper status indicates an increase in ash hopper level, or the ash unloading equipment status indicates abnormal operation of the ash unloading equipment, the attribution probability or unnormalized diagnostic score of ash hopper blockage or abnormal level is increased.

[0016] Further, the process for determining the ash accumulation rate characteristic includes: determining the ash accumulation rate characteristic based on the growth slope of the measured pressure difference or normalized filter resistance per unit time after ash removal; the process for determining the abnormal airflow characteristic includes: determining the abnormal airflow characteristic based on at least one of the measured pressure difference change amplitude, flue gas flow rate change amplitude, induced draft fan frequency change amplitude, induced draft fan current change amplitude, and normalized filter resistance change amplitude; the process for determining the filter media aging or permanent blockage characteristic includes: determining the filter media aging or permanent blockage characteristic based on at least one of the cumulative operating time of the filter media, baseline drift of the residual pressure difference after ash removal within multiple ash removal cycles, long-term upward trend of normalized filter resistance, and long-term downward trend of ash removal effectiveness characteristic; and correcting the attribution probability or unnormalized diagnostic score of each candidate abnormality type based on the ash accumulation rate characteristic, the abnormal airflow characteristic, and the filter media aging or permanent blockage characteristic. The corrections include: when the ash accumulation rate characteristic indicates an increase in the rate of increase of differential pressure or normalized filter resistance after cleaning, and the cleaning effectiveness characteristic is within the normal range, increasing the attribution probability or unnormalized diagnostic score for increased dust load; when the abnormal airflow characteristic indicates an increase in measured differential pressure, an increase in flue gas flow rate, and a failure of normalized filter resistance to increase synchronously, increasing the attribution probability or unnormalized diagnostic score for abnormal airflow, and decreasing the attribution probability or unnormalized diagnostic score for abnormal filter media adhesion and clogging and filter media aging or permanent clogging; when the cumulative operating time of the filter media exceeds the preset operating time, or the baseline of residual differential pressure after cleaning continues to rise within multiple cleaning cycles, or the cleaning effectiveness characteristic declines over a long period, increasing the attribution probability or unnormalized diagnostic score for filter media aging or permanent clogging; wherein, the preset operating time is determined based on the filter media design life, historical maintenance records, historical normal operating condition statistics, equipment operating experience values, or training sample quantiles.

[0017] Furthermore, the training process of the pre-trained deep learning time-series attribution model includes: acquiring historical operating samples or simulated operating samples of the flue gas filter dust collector under different operating conditions; labeling the historical operating samples or simulated operating samples with anomaly type tags, the anomaly type tags including at least two of the following: normal filter cake formation, increased dust load, insufficient dust removal, dust removal system failure, abnormal filter media adhesion and clogging, ash hopper clogging or abnormal material level, abnormal airflow, and filter media aging or permanent clogging; dividing the historical operating samples or simulated operating samples into multiple multivariate time-series training samples by sliding window segmentation according to a preset window length and preset step size; extracting the pressure difference anomaly time-series feature matrix from each multivariate time-series training sample; and using the pressure difference anomaly time-series feature matrix... The feature matrix is ​​used as input, and the anomaly type label is used as a supervision signal to train the deep learning model, resulting in the pre-trained deep learning time-series attribution model. The sliding window segmentation includes: sorting the historical or simulated running samples according to timestamps; forming a time window with a predetermined number of consecutive sampling points; moving the time window with a predetermined step size to generate multiple training samples; determining the anomaly type label at the end of the time window as the label of the training sample; fitting standardized parameters to the numerical features in the training set, and using the standardized parameters for the validation set, test set, and online inference data; the predetermined window length and predetermined step size are determined based on at least one of the sampling period, clearing period, anomaly duration, and model validation set performance.

[0018] Furthermore, the pre-trained deep learning temporal attribution model includes a temporal feature encoding module, an anomaly attribution classification module, and a probability output module. The temporal feature encoding module is used to encode the temporal feature matrix of the pressure difference anomaly to obtain encoded features. The anomaly attribution classification module is used to calculate the classification output for each candidate anomaly type based on the encoded features. The classification output is at least one of the following: a diagnostic score before probabilistic processing, a classification score, a logistic value, or a decision function value. The unnormalized diagnostic score has the same meaning as the classification output, referring to the score calculated by the anomaly attribution classification module for each candidate anomaly before probabilistic processing. The output type includes at least one of the following: diagnostic score, classification score, logistic value, or decision function value; the probability output module is used to perform probabilistic processing on the unnormalized diagnostic score to obtain the attribution probability of each candidate anomaly type, and determine the anomaly confidence based on at least one of the following: the maximum attribution probability, the difference between the highest and second-highest attribution probabilities, input data integrity, and model output entropy; the probabilistic processing includes Softmax normalization or probability calibration; the temporal feature encoding module includes at least one of the following: recurrent neural network, long short-term memory network, gated recurrent unit, temporal convolutional network, or Transformer encoder.

[0019] Further, after obtaining the anomaly type, the attribution probability of each candidate anomaly type, and the anomaly confidence, the method further includes: obtaining the rule diagnosis result based on the mechanism constraint rule, and fusing the rule diagnosis result with the anomaly type, attribution probability, and anomaly confidence output by the pre-trained deep learning temporal attribution model to obtain a hybrid diagnosis result; the rule diagnosis result includes the rule anomaly type, rule confidence, and rule diagnosis basis information; wherein, when the rule anomaly type is consistent with the anomaly type output by the pre-trained deep learning temporal attribution model, the confidence of the hybrid diagnosis result is increased; when the two are inconsistent and the anomaly confidence output by the pre-trained deep learning temporal attribution model is higher than a preset fusion threshold, the anomaly type output by the pre-trained deep learning temporal attribution model is preferred; when the two are inconsistent and the anomaly confidence output by the pre-trained deep learning temporal attribution model is not higher than the preset fusion threshold, the rule anomaly type is preferred; the preset fusion threshold is determined based on at least one of the validation set classification accuracy, macro average F1 value, false positive rate constraint, and false negative rate constraint, or is between 0.60 and 0.The values ​​are within the range of 95; the mechanism constraint rules include at least one of the following rules: when the condensation risk characteristic is lower than the preset condensation risk threshold, reduce the rule diagnosis score or attribution probability of abnormal filter media adhesion and blockage; when the dust removal execution status data is normal and the pressure difference after dust removal has effectively decreased, reduce the rule diagnosis score or attribution probability of dust removal system failure; when the measured pressure difference increases but the normalized filter resistance does not increase synchronously, reduce the rule diagnosis score or attribution probability of abnormal filter media adhesion and blockage and filter media aging or permanent blockage; when the ash hopper status does not indicate an increase in the ash hopper level and the ash unloading equipment does not show any abnormal operation, reduce the ash... The system outputs rule-based diagnostic scores or attribution probabilities for bucket blockage or material level anomalies. When the pre-trained deep learning temporal attribution model is missing, model parameter loading fails, standardized parameter loading fails, label mapping relationship loading fails, or the current running data length is insufficient for a preset time window, the system outputs a model unavailable status and calls a rule-based diagnostic model based on mechanistic constraints to generate a backup pressure difference anomaly diagnostic result. This backup pressure difference anomaly diagnostic result is stored in parallel with the pressure difference anomaly diagnostic result output when the pre-trained deep learning temporal attribution model is available, for temporary use by the operation and maintenance system until the pre-trained deep learning temporal attribution model becomes available again. The technical characteristics of this method, which primarily uses the pre-trained deep learning temporal attribution model as the diagnostic path, remain unchanged. The operation and maintenance recommendations are determined based on the anomaly type, including: when the anomaly type is normal filter cake formation, outputting a recommendation to maintain the existing dust removal strategy or optimize the dust removal cycle; when the anomaly type is increased dust load, outputting a recommendation to check production load, inlet dust concentration, and raw material changes; when the anomaly type is insufficient dust removal, outputting a recommendation to adjust the dust removal cycle, pulse jet duration, or differential pressure trigger threshold; when the anomaly type is a dust removal system malfunction, outputting a recommendation to check the dust removal actuator, pulse jet pipe, air tank pressure, backflushing mechanism, or rapping mechanism. The organization's recommendations are as follows: When the anomaly is filter media adhesion and clogging, the recommendations are to check for moisture, adhesion, caking, or condensation deposits on the filter media surface, increase flue gas temperature, improve insulation and heat tracing, and check for air leaks. When the anomaly is ash hopper clogging or abnormal material level, the recommendations are to check the ash hopper level gauge, ash discharge valve, ash conveying equipment, and clean the ash hopper. When the anomaly is abnormal airflow, the recommendations are to check the induced draft fan frequency, airflow setting, damper opening, and air-to-cloth ratio. When the anomaly is filter media aging or permanent clogging, the recommendations are to check the filter media operating hours, historical maintenance records, and replacement or offline cleaning plans.

[0020] Secondly, a flue gas filter dust collector pressure difference abnormality diagnosis device is provided, including a data acquisition interface, a processor, a memory, and an output interface. The data acquisition interface communicates with the multi-source sensors, DCS system, PLC system, or industrial gateway of the flue gas filter dust collector to acquire multi-source operating data during the operation of the flue gas filter dust collector. The memory stores a computer program, model parameters, standardized parameters, and label mapping relationships of a pre-trained deep learning time-series attribution model. The processor executes the computer program to call the model parameters, standardized parameters, and label mapping relationships of the pre-trained deep learning time-series attribution model and implement the flue gas filter dust collector pressure difference abnormality diagnosis method of the first aspect. The output interface outputs the pressure difference abnormality diagnosis results, diagnostic basis information, and operation and maintenance suggestions.

[0021] The aforementioned method and device for diagnosing pressure difference abnormalities in flue gas filter dust collectors collects multi-source operating data during the operation of the flue gas filter dust collector. It combines measured pressure difference, flue gas flow rate, dust removal action signals, dust removal execution status data, flue gas temperature, flue gas humidity, dew point related data, ash hopper status, ash unloading equipment status, and outlet particulate matter concentration data for pressure difference abnormality diagnosis. This allows the diagnosis process to move beyond solely relying on whether the measured pressure difference reaches a preset threshold. Instead, it combines various status information during the operation of the flue gas filter dust collector to identify the specific cause of the pressure difference increase. This helps solve the problem in existing technologies where it is difficult to identify the specific cause of the pressure difference increase in flue gas filter dust collectors using only the measured pressure difference.

[0022] The filtration velocity or air-to-cloth ratio is determined based on the flue gas flow rate and the effective filtration area of ​​the flue gas filter dust collector. The measured pressure difference is then normalized based on this filtration velocity or air-to-cloth ratio to obtain the normalized filtration resistance. This reduces the impact of changes in flue gas flow rate, filtration velocity, or air-to-cloth ratio on the judgment of the measured pressure difference. When the increase in measured pressure difference is mainly caused by an increase in flue gas flow rate, filtration velocity, or air-to-cloth ratio, the normalized filtration resistance can be used to distinguish between abnormal airflow and abnormalities such as filter media adhesion and blockage, filter media aging, or permanent blockage. This helps reduce the risk of misjudging abnormal airflow as filter media blockage.

[0023] A dust removal event window is constructed using the dust removal action signal as the event anchor point. Within this window, dust removal response features are extracted, including pre-duration pressure difference, post-duration pressure difference, dust removal drop amplitude, dust removal drop ratio, pressure difference re-rise slope, and residual pressure difference after dust removal. This establishes a correspondence between the dust removal action signal and the measured pressure difference change process after the dust removal action. These dust removal response features allow for determination of whether an effective dust removal response is formed after the dust removal action is triggered. This provides a basis for constructing dust removal effectiveness characteristics, residual resistance characteristics, and dust removal execution consistency characteristics, thus facilitating the differentiation between different causes such as normal filter cake formation, insufficient dust removal, and dust removal system malfunction.

[0024] A time-series feature matrix of pressure differential anomalies is constructed based on normalized filtration resistance, dust removal response characteristics, and multi-source operational data. This matrix includes features from multiple moments within a continuous time window, thus expressing the temporal changes in pressure differential during the rise of the flue gas filter dust collector. Compared to methods that rely solely on measured pressure differential at a single moment, the time-series feature matrix of pressure differential anomalies can simultaneously characterize features corresponding to candidate anomaly types, including ash accumulation rate, dust removal effectiveness, residual resistance, condensation risk, consistency of dust removal execution, ash hopper blockage or abnormal material level, abnormal airflow, and filter media aging or permanent blockage. This enhances the ability to differentiate the causes of pressure differential increases in the flue gas filter dust collector.

[0025] The time-series feature matrix of differential pressure anomalies is input into a pre-trained deep learning time-series attribution model to obtain the anomaly type corresponding to the increase in differential pressure in the flue gas filter dust collector and the attribution probability of each candidate anomaly type. The anomaly confidence level is determined based on at least one of the following: maximum attribution probability, attribution probability interval between different candidate anomaly types, input data integrity, and model output distribution. Therefore, this invention can not only output the anomaly type corresponding to the increase in differential pressure in the flue gas filter dust collector, but also the attribution probability and anomaly confidence level of each candidate anomaly type. This allows the differential pressure anomaly diagnosis results to have a probabilistic expression and confidence evaluation, facilitating comparison and judgment of candidate anomaly types such as normal filter cake formation, increased dust load, insufficient dust removal, dust removal system failure, abnormal filter media adhesion and clogging, ash hopper clogging or abnormal material level, abnormal airflow, and filter media aging or permanent clogging.

[0026] Based on the anomaly type, attribution probability, and anomaly confidence level, the system outputs the differential pressure anomaly diagnosis results, diagnostic basis information, and corresponding operation and maintenance recommendations. Therefore, after identifying the anomaly type corresponding to the increased differential pressure in the flue gas filter dust collector, it can further provide diagnostic basis information and operation and maintenance recommendations related to that anomaly type. This helps avoid problems such as inaccurate judgment of the cause of the increased differential pressure in the flue gas filter dust collector and improper subsequent handling, which can result from relying solely on measured differential pressure increases for dust removal control or operational judgment.

[0027] The present invention will now be further described in conjunction with the accompanying drawings and specific embodiments. Additional aspects and advantages of the invention will be set forth in part in the description which follows, and in part will be obvious from the description, or may be learned by practice. Attached Figure Description

[0028] The accompanying drawings, which form part of this specification, are used to aid in understanding the invention. The contents provided in the drawings and their related descriptions in this specification can be used to explain the invention, but do not constitute an undue limitation of the invention.

[0029] Figure 1 This is a schematic diagram of an embodiment of the flue gas filter dust collector pressure difference abnormality diagnosis device of the present invention.

[0030] Figure 2 This is a schematic diagram of an embodiment of the flue gas filter dust collector pressure difference abnormality diagnosis method of the present invention.

[0031] Figure 3 A schematic diagram of the time series characteristics of multi-source operational data corresponding to the abnormal types of normal filter cake formation.

[0032] Figure 4 A schematic diagram of the time series characteristics of multi-source operational data corresponding to the abnormal type of increased dust load.

[0033] Figure 5 A schematic diagram of the time series characteristics of multi-source operational data corresponding to the abnormal type of insufficient dust removal.

[0034] Figure 6 A schematic diagram of the timing characteristics of multi-source operational data corresponding to the abnormal types of malfunctions in the dust removal system.

[0035] Figure 7 This is a schematic diagram of the timing characteristics of multi-source operating data corresponding to abnormal adhesion and blockage of filter media.

[0036] Figure 8 This is a schematic diagram of the time sequence characteristics of multi-source operating data corresponding to ash hopper blockage or abnormal material level.

[0037] Figure 9 This is a schematic diagram of the time series characteristics of multi-source operational data corresponding to abnormal air volume.

[0038] Figure 10 A schematic diagram of the timing characteristics of multi-source operating data corresponding to the abnormal types of filter media aging or permanent blockage.

[0039] Figure 11 This is a schematic diagram of the online inference results display interface for a pre-trained deep learning temporal attribution model.

[0040] The components are labeled as follows: data acquisition interface 11, memory 12, processor 13, and output interface 14. Detailed Implementation

[0041] The present invention will now be clearly and completely described in conjunction with the accompanying drawings. Those skilled in the art will be able to implement the present invention based on these descriptions. Before describing the present invention in conjunction with the accompanying drawings, it should be particularly noted that: The technical solutions and features provided in the various sections, including the following description, can be combined with each other without conflict. Furthermore, where possible, these technical solutions, features, and related combinations can be given specific technical subject matter and protected by relevant patents.

[0042] The embodiments of the present invention described below are generally only some embodiments and not all embodiments. Based on these embodiments, all other embodiments obtained by those skilled in the art without creative effort should fall within the scope of patent protection.

[0043] Regarding terminology and units in this specification: The term "comprising" and any variations thereof in this specification, the corresponding claims, and related sections are intended to cover a non-exclusive inclusion. Furthermore, other relevant terms and units can be reasonably interpreted based on the information provided in this specification.

[0044] This embodiment provides a method and apparatus for diagnosing abnormal pressure differences in a flue gas filtration dust collector. The flue gas filtration dust collector is a dust removal device that has a cleaning actuator and uses a filter medium to capture particulate matter in flue gas. Flue gas filtration dust collectors can include bag filters, cartridge filters, and plate filters, among other types. Specifically, bag filters or baghouse filters use filter bags as the filter medium; cartridge filters use filter cartridges as the filter medium; and plate filters use filter plates as the filter medium. The cleaning actuator can include at least one of a reverse-air cleaning mechanism and a rapping cleaning mechanism; wherein, a pulse-jet cleaning mechanism can be a specific form of a reverse-air cleaning mechanism.

[0045] In one specific embodiment, the back-flushing cleaning mechanism refers to a cleaning actuator that cleans the filter medium by reverse airflow, compressed gas pulse, or back-flushing airflow. The back-flushing cleaning mechanism may include at least one of a pulse jet cleaning mechanism and a back-flushing air cleaning mechanism. Specifically, the pulse jet cleaning mechanism refers to a back-flushing cleaning mechanism that uses compressed gas pulses to instantly clean the filter medium, and it may include at least one of an air reservoir, a pulse valve, a jet pipe, a jet control unit, and a jet pressure detection component; the back-flushing air cleaning mechanism refers to a back-flushing cleaning mechanism that cleans the filter medium by reverse airflow or back-flushing airflow, and it may include at least one of a back-flushing fan, a back-flushing valve, a back-flushing duct, a back-flushing air pressure detection component, and a back-flushing air volume detection component.

[0046] In one specific embodiment, when the dust removal actuator is a pulse jet dust removal mechanism, the dust removal execution status data includes at least one of the following: jet pressure, air tank pressure, pulse valve status, pulse valve current, and jet duration; when the dust removal actuator is a reverse air dust removal mechanism, the dust removal execution status data includes at least one of the following: reverse air pressure, reverse air volume, reverse air valve status, reverse air duration, and dust removal feedback signal; when the dust removal actuator is a rapping dust removal mechanism, the dust removal execution status data includes at least one of the following: rapping motor status, rapping current, rapping duration, and dust removal feedback signal. Therefore, the dust removal execution status data can reflect the actual execution status after the dust removal action signal is triggered, depending on the type of dust removal actuator.

[0047] like Figure 1 As shown, the flue gas filter dust collector pressure difference abnormality diagnosis device provided in this embodiment includes a data acquisition interface 11, a processor 13, a memory 12, and an output interface 14. The data acquisition interface 11 is communicatively connected to a communication connection object, which may include at least one of a multi-source sensor, a DCS system, a PLC system, and an industrial gateway. The data acquisition interface 11 is used to acquire multi-source operating data during the operation of the flue gas filter dust collector. The multi-source operating data includes at least the measured pressure difference, flue gas flow rate, dust removal action signal, dust removal execution status data, flue gas temperature, flue gas humidity, dew point related data, ash hopper status, ash unloading equipment status, and outlet particulate matter concentration.

[0048] Multi-source sensors refer to field detection elements installed on flue gas filtration dust collectors and their related flue gas pipelines, dust removal actuators, ash hoppers, ash unloading equipment, or fan systems. These can include some or all of the following: differential pressure sensors, flow sensors, temperature sensors, humidity sensors, dew point sensors, particulate matter concentration sensors, pressure sensors, level sensors, current sensors, valve position status sensors, and equipment operation status sensors. A DCS system refers to a distributed control system that centrally collects and manages process data, interlock status, alarm status, and equipment operation data from the production system where the flue gas filtration dust collector is located. A PLC system refers to a programmable logic control system, typically used to collect field switch and analog signals and control the actions of dust removal actuators, ash unloading equipment, fans, or related actuators. An industrial gateway refers to a gateway device used for data conversion, protocol adaptation, and communication forwarding between field devices, sensors, DCS systems, PLC systems, and upper-level maintenance or diagnostic systems. It can support at least one of Modbus, OPC, OPC UA, Profinet, Profibus, EtherNet / IP, MQTT, or other industrial communication protocols.

[0049] Modbus is a commonly used industrial field communication protocol used to transmit register data, digital data, or analog data between sensors, instruments, PLC systems, industrial gateways, or host systems. OPC is a communication standard for industrial automation data access, used to exchange process data between control systems, monitoring systems, and data acquisition systems from different manufacturers. UA is a unified architecture communication specification developed based on OPC, supporting cross-platform data modeling, secure communication, and structured data exchange. It is suitable for data integration between DCS systems, PLC systems, industrial gateways, and upper-level diagnostic systems. Profinet is an automation communication protocol based on industrial Ethernet, commonly used for real-time data communication between PLC systems, remote I / O, drive devices, and field actuators. Profibus is a fieldbus communication protocol, commonly used for data communication between PLC systems and field instruments, remote I / O, frequency converters, valve positioners, and other devices. EtherNet / IP is a communication protocol based on industrial Ethernet, which can be used for data exchange between PLC systems, industrial controllers, frequency converters, remote I / O modules, and upper-level systems. MQTT is a message transmission protocol based on a publish / subscribe mechanism, which can be used to transmit flue gas filter dust collector operation data, alarm data, and diagnostic result data between industrial gateways, edge computing devices, cloud platforms, or operation and maintenance systems.

[0050] All of the above communication protocols can be used as communication methods to acquire multi-source operational data through data acquisition interfaces. Different flue gas filter dust collectors may use DCS systems, PLC systems, sensors, industrial gateways, or upper-level maintenance systems from different manufacturers, therefore the actual communication protocols may differ. This embodiment lists Modbus, OPC, OPC UA, Profinet, Profibus, EtherNet / IP, and MQTT only to illustrate that industrial gateways or data acquisition interfaces can adapt to various industrial communication environments. It does not limit the use of all the above communication protocols, nor does it exclude the use of other industrial communication protocols capable of transmitting flue gas filter dust collector operational data.

[0051] The "at least one" relationship is used between the aforementioned communication connection objects because the data access conditions at different flue gas filtration and dust collector sites vary. At some sites, multi-source operational data such as measured differential pressure, flue gas flow rate, flue gas temperature, flue gas humidity, dew point data, ash hopper status, ash discharge equipment status, and outlet particulate matter concentration can be directly accessed by multi-source sensors through the data acquisition interface. At other sites, the aforementioned multi-source operational data is centrally collected and stored by the DCS system, and the data acquisition interface can directly read from the DCS system. At still other sites, data such as cleaning action signals, cleaning execution status data, pulse valve status, vibrating motor status, and ash discharge equipment operating status may primarily be collected and controlled by the PLC system, and the data acquisition interface can read from the PLC system. Finally, at some sites, an industrial gateway may be used to perform protocol conversion and aggregation of data from multi-source sensors, the DCS system, or the PLC system, and the data acquisition interface can obtain multi-source operational data through the industrial gateway. Therefore, this embodiment does not limit the data acquisition interface to simultaneously connecting to multiple source sensors, DCS system, PLC system and industrial gateway. Instead, it allows the selection of one or more communication connection objects to obtain multi-source operating data during the operation of the flue gas filter dust collector, based on the existing control system architecture and data access conditions on site.

[0052] The memory 12 is used to store the computer program, model parameters of the pre-trained deep learning time-series attribution model, standardized parameters, and label mapping relationships. The standardized parameters are used to standardize the numerical features input to the pre-trained deep learning time-series attribution model, and the label mapping relationships are used to establish the correspondence between candidate anomaly types and the output categories of the pre-trained deep learning time-series attribution model.

[0053] Processor 13 executes computer programs and calls the model parameters, normalized parameters, and label mapping relationships of the pre-trained deep learning time-series attribution model stored in memory 12 to implement a method for diagnosing pressure difference anomalies in flue gas filtration dust collectors. Processor 13 can perform operations such as multi-source operating data preprocessing, normalized filter resistance calculation, dust removal event window construction, dust removal response feature extraction, pressure difference anomaly time-series feature matrix construction, pre-trained deep learning time-series attribution model inference, rule-based diagnostic result acquisition, mixed diagnostic result generation, and pressure difference anomaly diagnostic result output.

[0054] In one specific implementation, the model parameters of the pre-trained deep learning temporal attribution model refer to the neural network parameters obtained after parameter training based on training samples with anomaly type labels, as well as configuration parameters related to the neural network structure. The neural network parameters may include at least one of the weight parameters, bias parameters, and normalization layer parameters in the temporal feature encoding module and the anomaly attribution classification module. The configuration parameters related to the neural network structure may include at least one of the model type, input feature dimension, number of candidate anomaly types, hidden layer dimension, random deactivation ratio, preset window length, and feature column set used to construct the compression difference anomaly temporal feature matrix.

[0055] In a specific software implementation, the pre-trained deep learning temporal attribution model can employ a temporal convolutional network or a long short-term memory (LSTM) network. When the pre-trained deep learning temporal attribution model employs a temporal convolutional network, the temporal feature encoding module can include multiple one-dimensional convolutional layers, batch normalization layers, non-linear activation layers, random deactivation layers, and adaptive pooling layers. The anomaly attribution classification module can include fully connected layers, which are used to map the encoded features output by the temporal feature encoding module to the classification output corresponding to each candidate anomaly type. When the pre-trained deep learning temporal attribution model employs an LTM network, the temporal feature encoding module can include LTM network layers, and the anomaly attribution classification module can include random deactivation layers and fully connected layers. The fully connected layers are used to obtain the classification output corresponding to each candidate anomaly type based on the hidden states output by the LTM network layers. The aforementioned classification output is the classification score, logistic value, or unnormalized diagnostic score before probabilistic processing.

[0056] In one specific implementation, the model parameters can be stored in the memory 12 in the form of a model parameter file. The model parameter file may include the state parameters of the neural network, as well as the model type, input feature dimension, number of candidate anomaly types, hidden layer dimension, random deactivation ratio, preset window length, and feature column set required to construct the neural network. When the processor 13 calls the pre-trained deep learning time-series attribution model, it first reads the model parameter file, constructs the corresponding deep learning model structure according to the model type, input feature dimension, number of candidate anomaly types, hidden layer dimension, and random deactivation ratio, and then loads the state parameters of the neural network into the deep learning model structure, thereby obtaining a pre-trained deep learning time-series attribution model that can be used for inference.

[0057] In one specific implementation, the standardized parameters refer to parameters obtained after standardizing the numerical features in the training samples. The standardized parameters may include at least one of the mean, variance, or standard deviation parameters corresponding to each input feature. During training, the processor 13 performs sliding window segmentation on the historical or simulated running samples according to a preset window length and a preset step size to obtain multiple multivariate time series training samples; then, the stress difference abnormal time series feature matrix in the training set is expanded according to the feature dimensions, and the standardized parameters are fitted based on the expanded numerical features of the training set. Thus, the standardized parameters reflect the numerical distribution of the feature data in the training set.

[0058] In one specific software implementation, the standardized parameters can be generated by a feature normalizer and saved as a standardized parameter file. During online diagnostics, the processor 13 reads the standardized parameter file from the memory 12 and uses the same standardized parameters as in the training phase to standardize the pressure difference abnormal time-series feature matrix within the current continuous time window. Specifically, the processor 13 first reshapes the pressure difference abnormal time-series feature matrix corresponding to the current continuous time window into the input shape required by the pre-trained deep learning time-series attribution model, and then uses the standardized parameters to perform a standardization transformation on each numerical feature, so that the online inference data and the training data are on the same numerical scale. The standardized pressure difference abnormal time-series feature matrix is ​​then input into the pre-trained deep learning time-series attribution model to reduce the impact of differences in feature dimensions, numerical ranges, and acquisition units on the classification output.

[0059] In one specific implementation, the label mapping relationship refers to the correspondence between candidate anomaly types and the category numbers output by the pre-trained deep learning temporal attribution model. Since the anomaly attribution classification module of the pre-trained deep learning temporal attribution model outputs a classification output arranged according to the category numbers, the processor 13 needs to map the candidate anomaly types to integer category numbers used during training based on the label mapping relationship, and restore the category numbers output by the pre-trained deep learning temporal attribution model to the corresponding candidate anomaly types during inference.

[0060] In one specific implementation, the label mapping relationship may include a forward label mapping relationship and a reverse label mapping relationship. The forward label mapping relationship is used to map candidate anomaly types to integer category numbers. For example, candidate anomaly types such as normal filter cake formation, increased dust load, insufficient dust removal, dust removal system malfunction, abnormal filter media adhesion and clogging, ash hopper clogging or abnormal material level, abnormal airflow, and aging or permanent clogging of the filter media are mapped to different category numbers. The reverse label mapping relationship is used to restore the category numbers output by the pre-trained deep learning time-series attribution model to the corresponding candidate anomaly types. After obtaining the attribution probability corresponding to each candidate anomaly type, the processor 13 generates anomaly type names, attribution probabilities for each candidate anomaly type, and anomaly confidence levels based on the reverse label mapping relationship.

[0061] In one specific software implementation, the training process can first establish a label mapping relationship based on the anomaly type labels in historical or simulated running samples, and save this label mapping relationship to memory 12. Then, when loading the pre-trained deep learning temporal attribution model, processor 13 simultaneously reads the model parameters, normalized parameters, and label mapping relationship. If any of the model parameters, normalized parameters, or label mapping relationship is missing or fails to load, processor 13 can determine that the pre-trained deep learning temporal attribution model is unavailable and output a model unavailable status; if all three are successfully loaded, processor 13 constructs the pre-trained deep learning temporal attribution model based on the model parameters, processes the current pressure difference anomaly temporal feature matrix based on the normalized parameters, and interprets the class number output by the model based on the label mapping relationship.

[0062] In one specific implementation, model parameters, standardized parameters, and label mapping relationships together constitute the basic data required for online inference of the pre-trained deep learning time-series attribution model. Specifically, model parameters determine the structure and trained weights of the pre-trained deep learning time-series attribution model; standardized parameters ensure that the time-series feature matrix of pressure differential anomalies in the online input is consistent with the numerical scale during the training phase; and label mapping relationships ensure that the classification output, attribution probability, and maximum attribution probability of the pre-trained deep learning time-series attribution model can be correctly interpreted as candidate anomaly types such as normal filter cake formation, increased dust load, insufficient dust removal, dust removal system malfunction, abnormal filter media adhesion and clogging, ash hopper blockage or abnormal material level, abnormal airflow, or filter media aging or permanent blockage. By storing the above data in memory 12, processor 13 can stably call the pre-trained deep learning time-series attribution model during the diagnostic process to achieve inference and output of the anomaly type corresponding to the increase in pressure differential of the flue gas filter dust collector.

[0063] Output interface 14 is used to output differential pressure anomaly diagnosis results, diagnostic basis information, and operation and maintenance handling suggestions. Output interface 14 can communicate with the operation and maintenance system or display terminal, enabling the operation and maintenance system or display terminal to receive the information output by the output interface. Output interface 14 can also output at least one of the following: anomaly type, attribution probability of each candidate anomaly type, anomaly confidence, rule-based diagnosis results, mixed diagnosis results, model availability status, and training metrics.

[0064] Figure 1 The code file names shown are only for illustrating one software implementation method and do not constitute a limitation on the scope of protection of this invention. In a specific software implementation, the computer program can be implemented using a layered and modular software project structure. Specifically, a data preprocessing module can be set up to implement data reading, field validation, and time range filtering. For example, preprocessing.py and schemas.py can be used to load multi-source running data, validate required fields, process timestamps, and standardize data formats. A feature engineering module can be set up to implement normalized filter resistance calculation, dust removal event window construction, dust removal response feature extraction, and candidate anomaly type corresponding feature construction. For example, feature_engineering.py can be used to calculate filter velocity or air-to-cloth ratio, normalized filter resistance, dust removal fall-off amplitude, dust removal fall-off ratio, pressure difference re-rise slope, residual pressure difference after dust removal, and various diagnostic features.

[0065] The system can be configured to implement a rule-based diagnostic module, a pre-trained deep learning time-series attribution model module, an inference fusion module, and an output recommendation module. For example, `diagnosis.py` can generate rule-based diagnostic results based on ash accumulation rate characteristics, ash removal effectiveness characteristics, residual resistance characteristics, condensation risk characteristics, ash removal execution consistency characteristics, ash hopper blockage or material level anomaly characteristics, airflow anomaly characteristics, and filter media aging or permanent blockage characteristics; `deep_learning / models.py` can define the network structure of the pre-trained deep learning time-series attribution model; `deep_learning / inference.py` can load model parameters, standardized parameters, and label mapping relationships, and perform online inference of the pre-trained deep learning time-series attribution model and its fusion with rule-based diagnostic results; `recommendations.py` can generate operation and maintenance recommendations based on anomaly types.

[0066] In a specific software implementation, app.py can also be used as the main control program or interactive entry point at runtime. This program can call the data acquisition interface, various functional modules in the processor, model parameters in the memory, standardized parameters, and label mapping relationships. It can also output pressure difference anomaly diagnosis results, diagnostic basis information, and operation and maintenance suggestions through the output interface. Furthermore, visualization.py can be used to display the time-series characteristics of multi-source running data, pressure difference anomaly diagnosis results, training indicators, or classification evaluation results. The above-mentioned code file names and module divisions are only used to illustrate one engineering organization method of the computer program in this embodiment. In actual implementation, other file names, module division methods, programming languages, or deployment architectures can be used. Any implementation that can achieve the pressure difference anomaly diagnosis method for flue gas filter dust collectors described in this invention should be considered an implementation of this invention.

[0067] like Figure 2 As shown in the figure, the flue gas filter dust collector pressure difference abnormality diagnosis method provided in this embodiment includes the following steps.

[0068] S1: Collect multi-source operational data.

[0069] The processor 13 collects multi-source operating data during the operation of the flue gas filter dust collector via the data acquisition interface 11. The multi-source operating data includes at least the measured differential pressure, flue gas flow rate, dust removal action signal, dust removal execution status data, flue gas temperature, flue gas humidity, dew point related data, ash hopper status, ash unloading equipment status, and outlet particulate matter concentration. The dew point related data includes at least one of the following: water dew point temperature, acid dew point temperature, flue gas humidity data used to calculate the water dew point temperature, and acidic gas concentration data used to calculate the acid dew point temperature.

[0070] The measured differential pressure refers to the actual pressure difference between the inlet and outlet of the flue gas filter dust collector, obtained by differential pressure sensors, differential pressure transmitters, DCS systems, PLC systems, or industrial gateways during the operation of the flue gas filter dust collector. Alternatively, it can refer to the actual differential pressure value corresponding to a compartment, filter unit, or filter section of the flue gas filter dust collector. The measured differential pressure characterizes the resistance encountered by flue gas as it passes through the filter medium and its surface deposits. The measured differential pressure can be expressed in Pa, kPa, or other pressure units.

[0071] Flue gas flow rate refers to the volumetric flow rate or equivalent flow rate of flue gas passing through a flue gas filter and dust collector per unit time. Flue gas flow rate can be obtained from flow sensors, DCS systems, PLC systems, or industrial gateways, or it can be calculated based on fan operating parameters, damper opening, duct velocity, or flue gas cross-sectional area. Flue gas flow rate can be expressed in Nm³. 3 / h、m 3 / h or other flow rate units. Flue gas flow rate is used to determine the filtration velocity or air-to-cloth ratio, and to distinguish whether the increase in measured pressure difference is mainly caused by changes in flue gas flow rate.

[0072] A dust removal action signal is a signal used to indicate whether the dust removal actuator has been triggered or is performing a dust removal action. The dust removal action signal can be at least one of the following: a switching signal, a pulse signal, a control command signal, a valve action signal, a rapping action signal, or a backflushing action signal. For example, when the dust removal actuator is a pulse jet dust removal mechanism, the dust removal action signal can be a pulse jet control command or a pulse valve action signal; when the dust removal actuator is a backflushing dust removal mechanism, the dust removal action signal can be a backflushing valve opening signal or a backflushing blower start signal; when the dust removal actuator is a rapping dust removal mechanism, the dust removal action signal can be a rapping motor start signal or a rapping control command.

[0073] Dust cleaning execution status data refers to data used to characterize the actual operating status of the dust cleaning actuator after the dust cleaning action signal is triggered. When the dust cleaning actuator is a back-flushing dust cleaning mechanism, the dust cleaning execution status data may include at least one of the following: back-flushing air pressure, back-flushing air volume, back-flushing valve status, and dust cleaning feedback signal. When the dust cleaning actuator is a pulse jet dust cleaning mechanism, the dust cleaning execution status data may include at least one of the following: jet pressure, air tank pressure, pulse valve status, pulse valve current, and jet duration. When the dust cleaning actuator is a rapping dust cleaning mechanism, the dust cleaning execution status data may include at least one of the following: rapping motor status, rapping current, rapping duration, and dust cleaning feedback signal.

[0074] Flue gas temperature can include inlet and outlet flue gas temperatures. Flue gas humidity characterizes the moisture content in the flue gas. Dew point data is used to determine whether there is a risk of water vapor condensation or acidic gas condensation near the filter media. Ash hopper status can include at least one of the following: ash hopper level, ash hopper temperature, ash hopper rapping status, or ash hopper blockage alarm status. Ash discharge equipment status can include at least one of the following: ash discharge equipment current, ash discharge equipment operating status, ash discharge valve status, ash conveying equipment status, or ash discharge equipment alarm status. Outlet particulate matter concentration refers to the particulate matter concentration in the flue gas at the outlet of the flue gas filter dust collector, which can be used to help determine if the filter media is damaged, the seal is abnormal, or the dust removal efficiency is abnormal.

[0075] In one specific implementation, the multi-source operational data may further include compartment pressure differential, induced draft fan frequency, induced draft fan current, inlet dust concentration, production load, cumulative operating time of the filter media, maintenance markers, and anomaly type labels. The anomaly type labels are used for training, validating, or testing the pre-trained deep learning time-series attribution model and are not used as input features during online diagnosis. This avoids the pre-trained deep learning time-series attribution model directly reading anomaly type labels during the online diagnosis phase, thus ensuring that the online diagnosis results originate from the operational data itself.

[0076] After acquiring multi-source operational data, processor 13 preprocesses the data. Preprocessing may include sorting by timestamp, filtering by time range, handling missing values, handling outliers, converting numerical fields, converting units, changing the status of dust removal action signals, changing the status of dust removal execution data, and verifying field integrity. Through this preprocessing, processor 13 can transform multi-source operational data from different sources, with different units, different sampling periods, or different data formats into a unified data structure usable for subsequent calculations.

[0077] S2: Obtain the normalized filter resistance.

[0078] The processor 13 determines the filtration velocity or air-to-cloth ratio based on the flue gas flow rate and the effective filtration area of ​​the flue gas filter dust collector, and normalizes the measured pressure difference based on the filtration velocity or air-to-cloth ratio to obtain the normalized filtration resistance.

[0079] The effective filtration area refers to the area of ​​the filter media actually in use in a flue gas filtration dust collector. The effective filtration area can be the total area of ​​the filter media, or it can be the total area of ​​the filter media minus at least one of the following: the area of ​​offline compartments, the area of ​​isolated filter units, and the area of ​​damaged filter units. For baghouse dust collectors, the effective filtration area can be determined based on the number of filter bags in operation and the filtration area of ​​a single filter bag; for cartridge dust collectors, the effective filtration area can be determined based on the number of filter cartridges in operation and the filtration area of ​​a single filter cartridge; for plate dust collectors, the effective filtration area can be determined based on the number of filter plates in operation and the filtration area of ​​a single filter plate.

[0080] Filtration velocity or air-to-cloth ratio refers to the ratio between flue gas flow rate and the effective filtration area of ​​a flue gas filter, used to characterize the amount of flue gas passing through a unit area of ​​filter media. For bag filters, this parameter is usually called the air-to-cloth ratio; for cartridge filters, plate filters, or other flue gas filters, this parameter can be called filtration velocity. Processor 13 can determine the ratio of flue gas flow rate to effective filtration area as filtration velocity or air-to-cloth ratio. Processor 13 can also determine the ratio of measured pressure difference to filtration velocity or air-to-cloth ratio as normalized filtration resistance. Processor 13 can also determine the reference pressure difference corresponding to the current filtration velocity or air-to-cloth ratio based on the fitting relationship between measured pressure difference and filtration velocity or air-to-cloth ratio under historical normal operating conditions, and determine the deviation or deviation ratio between the measured pressure difference and the reference pressure difference as normalized filtration resistance.

[0081] The normalized filter resistance serves to mitigate the impact of changes in flue gas flow rate, filtration velocity, or air-to-cloth ratio on the determination of measured pressure difference. Since measured pressure difference is affected not only by the degree of ash accumulation on the filter media surface but also by flue gas flow rate and filtration velocity or air-to-cloth ratio, an increase in flue gas flow rate may lead to an increase in measured pressure difference. However, this increase in measured pressure difference does not necessarily indicate abnormal filter media adhesion and blockage, or filter media aging and permanent blockage. By obtaining the normalized filter resistance, processor 13 can distinguish between "increased measured pressure difference due to increased flue gas flow rate" and "increased measured pressure difference due to abnormal filter media resistance."

[0082] In one specific implementation, when the processor 13 determines that the measured differential pressure increases, the flue gas flow rate increases, but the normalized filter resistance does not increase synchronously, the processor 13 can increase the attribution probability or unnormalized diagnostic score of abnormal airflow, and decrease the attribution probability or unnormalized diagnostic score of abnormal filter media adhesion and blockage and abnormal filter media aging or permanent blockage. Therefore, the normalized filter resistance can serve as an important basis for distinguishing between abnormal airflow and abnormal filter media resistance.

[0083] S3: Create a dust removal event window.

[0084] The processor 13 constructs a dust removal event window using the dust removal action signal as the event anchor point, and extracts dust removal response features from the dust removal event window. The dust removal response features include at least the pressure difference before dust removal, the pressure difference after dust removal, the dust removal drop amplitude, the dust removal drop ratio, the pressure difference rise slope, and the residual pressure difference after dust removal.

[0085] The event anchor point refers to the moment when the dust removal action signal is triggered or the moment when the dust removal actuator begins to perform the dust removal action. The processor 13 uses the event anchor point as the center or starting point, and extracts the measured differential pressure data within a preset time period before and after the event anchor point to form a dust removal event window. The preset time period can be determined based on the dust removal action response time, dust removal cycle, sampling cycle, or historical normal dust removal response data.

[0086] In one specific implementation, the processor 13 acquires measured differential pressure data within a preset time period before the event anchor point to determine the differential pressure before cleaning; acquires measured differential pressure data within a preset time period after the event anchor point to determine the differential pressure after cleaning; determines the cleaning drop amplitude and cleaning drop ratio based on the differential pressure before and after cleaning; and determines the differential pressure rise slope and residual differential pressure after cleaning based on the measured differential pressure change trend after the event anchor point.

[0087] For example, processor 13 can take the average of the measured pressure differences at several sampling points before the dust removal action signal is triggered as the pre-duration pressure difference, and take the minimum measured pressure difference at several sampling points after the dust removal action signal is triggered as the post-duration pressure difference. The dust removal drop amplitude is the difference between the pre-duration and post-duration pressure differences. The dust removal drop ratio is the ratio between the dust removal drop amplitude and the pre-duration pressure difference. The pressure difference rise slope can be obtained by linearly fitting the measured pressure differences within a preset time range after the dust removal action signal is triggered. The residual pressure difference after dust removal can be determined by the post-duration pressure difference or the low-level measured pressure difference after the dust removal action signal is triggered.

[0088] The purpose of constructing the dust removal event window is to enable the processor 13 not only to know whether the control system has issued a dust removal action signal, but also to determine whether the measured differential pressure shows a decrease in response matching the dust removal action after the dust removal action signal is triggered. If a dust removal action signal exists, but the differential pressure decreases insufficiently after dust removal, the dust removal drop is small, or the dust removal drop ratio is low, it may indicate that the dust removal actuator has not formed an effective dust removal response. If the measured differential pressure decreases after dust removal, but the residual differential pressure gradually increases after dust removal, it may indicate that the residual resistance has increased. Therefore, the dust removal event window can provide a basis for distinguishing between normal filter cake formation, insufficient dust removal, dust removal system failure, abnormal filter media adhesion and clogging, and filter media aging or permanent clogging.

[0089] In one specific implementation, when the dust removal drop amplitude is greater than a preset dust removal drop amplitude threshold and the dust removal drop ratio is greater than a preset dust removal drop ratio threshold, the processor 13 can determine that the dust removal event has a valid dust removal response. If the dust removal drop amplitude is not greater than the preset dust removal drop amplitude threshold, or the dust removal drop ratio is not greater than the preset dust removal drop ratio threshold, the processor 13 can determine that the dust removal response of the dust removal event is insufficient.

[0090] In one specific implementation, after extracting the dust removal response features, the processor 13 maps the dust removal response features to the corresponding time points in a continuous time window according to the corresponding dust removal event timestamps, so that the dust removal response features maintain a correspondence with the time dimension in the differential pressure abnormal time sequence feature matrix. For example, when a dust removal action signal is triggered at the nth sampling point within the continuous time window, the processor 13 can map the pre-duration differential pressure, post-duration differential pressure, dust removal drop amplitude, dust removal drop ratio, differential pressure rise slope, and post-duration residual differential pressure corresponding to the dust removal event to the nth sampling point, or to multiple sampling points covered by the dust removal event window.

[0091] In one specific implementation, for multiple sampling points within a dust removal event window, the processor 13 can perform filling, interpolation, maintenance, or statistical aggregation on the dust removal response features to form an event response feature channel of the pressure difference abnormal time-series feature matrix. Filling refers to filling the corresponding dust removal response feature value of the dust removal event into the corresponding sampling point within the dust removal event window; interpolation refers to performing linear interpolation or other interpolation processing on the dust removal response feature value based on the time position between different sampling points within the dust removal event window; maintenance refers to maintaining the same dust removal response feature value within the dust removal event window or within a preset time range after the dust removal event; statistical aggregation refers to performing mean, maximum, minimum, quantile, cumulative, or recent value statistics on the dust removal response features corresponding to one or more dust removal events within a continuous time window.

[0092] In one specific embodiment, the event response feature channel may include at least one of the following: a dust removal drop amplitude channel, a dust removal drop ratio channel, a differential pressure rise slope channel, a residual differential pressure channel after dust removal, and an effective dust removal response marker channel. By incorporating the event response feature channel into the differential pressure anomaly time-series feature matrix, the processor 13 enables the pre-trained deep learning time-series attribution model to simultaneously receive normal operating features and dust removal event response features within a continuous time window, thereby allowing the differential pressure change process before and after the dust removal action signal to participate in anomaly type identification.

[0093] S4: Construct the time series feature matrix of pressure difference abnormality.

[0094] Processor 13 constructs a time-series feature matrix of pressure differential anomalies based on normalized filter resistance, dust removal response characteristics, and multi-source operating data. The time-series feature matrix of pressure differential anomalies includes features from multiple moments within a continuous time window. The continuous time window is a multivariate time series segment extracted from the multi-source operating data according to a preset window length and a preset step size.

[0095] Processor 13 sorts the multi-source runtime data according to timestamps, constructs a continuous time window with a preset number of consecutive sampling points, and moves the continuous time window according to a preset step size to obtain multiple multivariate time series segments. Each multivariate time series segment corresponds to a differential pressure anomaly time series feature matrix. The rows of the differential pressure anomaly time series feature matrix can correspond to different times within the continuous time window, and the columns can correspond to different feature channels.

[0096] In one specific embodiment, the feature channels in the differential pressure abnormal time-series feature matrix may include measured differential pressure, flue gas flow rate, filtration velocity or air-to-cloth ratio, normalized filtration resistance, ash removal action signal, ash removal execution status data, flue gas temperature, flue gas humidity, dew point related data, ash hopper status, ash unloading equipment status, outlet particulate matter concentration, and derived features constructed based on the above data. Derived features may include at least one of the following: measured differential pressure change slope, normalized filtration resistance change slope, measured differential pressure step change, water dew point margin, acid dew point margin, ash removal drop amplitude, ash removal drop ratio, differential pressure re-rise slope, and residual differential pressure after ash removal.

[0097] The role of the differential pressure anomaly time-series feature matrix is ​​to expand single-point data at a certain moment into multivariate time-series data within a continuous time window. The increase in differential pressure in a flue gas filter dust collector is usually not an isolated state at a single moment, but rather manifests as a trend in measured differential pressure over a period of time, changes in response before and after dust removal, changes in residual resistance, changes in dew point margin, changes in ash hopper level, changes in the status of ash unloading equipment, and changes in flue gas flow rate. By constructing the differential pressure anomaly time-series feature matrix, the processor 13 enables the pre-trained deep learning time-series attribution model to receive inputs with temporal order and multivariate correlations, thereby improving its ability to distinguish different candidate anomaly types.

[0098] In one specific implementation, the preset window length can be 60 sampling points, and the preset step size can be 5 sampling points. The preset window length and preset step size can be determined based on at least one of the sampling period, the dust clearing period, the anomaly duration, and the performance of the pre-trained deep learning time-series attribution model on the validation set. If the sampling period is short, the preset window length can be appropriately increased to cover the complete dust clearing period or the anomaly development process; if the sampling period is long, the preset window length can be appropriately decreased to avoid the continuous time window covering too long a period of time, thereby weakening the recent state representation ability.

[0099] S5: Determine the features corresponding to the candidate anomaly types.

[0100] Processor 13 determines the features corresponding to the candidate anomaly types based on the time-series feature matrix of pressure difference anomalies. Candidate anomaly types include at least two of the following: normal filter cake formation, increased dust load, insufficient dust removal, dust removal system malfunction, abnormal filter media adhesion and clogging, ash hopper clogging or abnormal material level, abnormal airflow, and filter media aging or permanent clogging. Abnormal filter media adhesion and clogging includes at least one of the following: condensation adhesion clogging, moisture absorption caking clogging, acidic condensate adhesion clogging, and oil mist or tar adhesion clogging.

[0101] The features corresponding to the candidate anomaly types include at least one of the following: ash accumulation rate feature, ash removal effectiveness feature, residual resistance feature, condensation risk feature, ash removal performance consistency feature, ash hopper blockage or material level anomaly feature, air volume anomaly feature, and filter media aging or permanent blockage feature.

[0102] The ash accumulation rate characteristic is used to characterize the rate at which particulate matter deposition on the filter media surface causes an increase in measured pressure difference or normalized filtration resistance. The processor 13 can determine the ash accumulation rate characteristic based on the slope of the increase in measured pressure difference or normalized filtration resistance per unit time after dust removal, or it can be determined in conjunction with changes in inlet dust concentration, production load, or dust load. When the ash accumulation rate characteristic increases, the attribution probability of the increased dust load or the unnormalized diagnostic score can be correspondingly improved.

[0103] The cleaning effectiveness characteristic is used to characterize whether the filter media resistance effectively decreases after the cleaning action. The processor 13 can determine the cleaning effectiveness characteristic based on the cleaning drop amplitude, cleaning drop ratio, residual pressure difference after cleaning, and pressure difference re-rise slope. The greater the cleaning drop amplitude, the higher the cleaning drop ratio, and the lower the residual pressure difference after cleaning, the higher the cleaning effectiveness characteristic; the smaller the cleaning drop amplitude, the lower the cleaning drop ratio, and the higher the residual pressure difference after cleaning, the lower the cleaning effectiveness characteristic.

[0104] Residual resistance characteristics are used to characterize the resistance level remaining on the filter media after cleaning. Processor 13 can determine the residual resistance characteristics based on the lower quantile of the measured differential pressure within a continuous time window, the lower quantile of the normalized filter resistance, the baseline change of the residual differential pressure after multiple cleaning events, and the upward trend of the residual differential pressure after cleaning. When the residual resistance characteristics increase, the attribution probability or unnormalized diagnostic score of candidate abnormality types such as abnormal filter media adhesion and clogging, filter media aging or permanent clogging, and insufficient cleaning may be affected accordingly.

[0105] The condensation risk characteristic is used to characterize the risk of water vapor condensation, acid gas condensation, or adhesion and deposition of flue gas near the filter medium. When the water dew point temperature is obtained, processor 13 determines the water dew point margin based on the difference between the flue gas outlet temperature and the water dew point temperature; when the acid dew point temperature is obtained, processor 13 determines the acid dew point margin based on the difference between the flue gas outlet temperature and the acid dew point temperature; when the acid dew point temperature is not directly obtained, processor 13 estimates the acid dew point temperature based on at least one of SO2 concentration, estimated SO3 value, HCl concentration, and flue gas humidity, and determines the acid dew point margin based on the difference between the flue gas outlet temperature and the estimated acid dew point temperature. Processor 13 determines the condensation risk characteristic based on the obtained or estimated water dew point margin, acid dew point margin, and flue gas humidity. The lower the water dew point margin, the lower the acid dew point margin, or the higher the flue gas humidity, the higher the condensation risk characteristic.

[0106] The consistency characteristic of dust removal execution is used to characterize whether there is consistency between the dust removal action signal, the dust removal execution status data, and the pressure difference decrease. Processor 13 determines whether the dust removal execution status data is within the preset normal range of the corresponding dust removal actuator after the dust removal action signal is triggered, or whether a state change matching the dust removal action signal occurs. Processor 13 further determines whether the measured pressure difference or normalized filter resistance decreases by a preset amount after the dust removal action signal is triggered. Processor 13 determines the consistency characteristic of dust removal execution based on the dust removal execution status data and the pressure difference decrease. If the dust removal execution status data is abnormal or the measured pressure difference decreases insufficiently after the dust removal action signal is triggered, the consistency characteristic of dust removal execution decreases, and the attribution probability or unnormalized diagnostic score of the dust removal system fault can be increased accordingly.

[0107] The ash hopper blockage or abnormal level characteristics are used to characterize the impact of ash accumulation in the ash hopper, abnormal ash level in the ash hopper, or abnormal ash discharge equipment on the differential pressure changes of the flue gas filter dust collector. Processor 13 determines the ash hopper blockage or abnormal level characteristics based on the ash hopper status and the ash discharge equipment status. The ash hopper status includes at least the ash level in the ash hopper, and the ash discharge equipment status includes at least the ash discharge equipment current or the ash discharge equipment operating status. When the ash hopper level rises, the ash discharge equipment current is abnormal, or the ash discharge equipment operating status is abnormal, the ash hopper blockage or abnormal level characteristics increase.

[0108] The airflow anomaly characteristic is used to characterize whether the increase in measured differential pressure is mainly caused by changes in flue gas flow rate, filtration velocity, or air-to-cloth ratio. The processor 13 determines the airflow anomaly characteristic based on at least one of the following: the magnitude of change in measured differential pressure, the magnitude of change in flue gas flow rate, the magnitude of change in induced draft fan frequency, the magnitude of change in induced draft fan current, and the magnitude of change in normalized filter resistance. When the measured differential pressure increases, the flue gas flow rate increases, the induced draft fan frequency increases, or the induced draft fan current increases, but the normalized filter resistance does not increase synchronously, the processor 13 can determine that the airflow anomaly characteristic has increased. The purpose of this characteristic is to avoid directly judging the increase in measured differential pressure caused by an increase in airflow as a filter media blockage anomaly.

[0109] Filter media aging or permanent clogging characteristics are used to characterize the irreversible increase in resistance that occurs after long-term use of the filter media. Processor 13 determines the filter media aging or permanent clogging characteristics based on at least one of the following: cumulative filter media operating time, baseline drift of residual pressure difference after cleaning over multiple cleaning cycles, long-term upward trend of normalized filtration resistance, and long-term downward trend of cleaning effectiveness characteristics. The longer the cumulative filter media operating time, the more pronounced the baseline drift of residual pressure difference after cleaning, the more pronounced the long-term upward trend of normalized filtration resistance, or the more pronounced the long-term downward trend of cleaning effectiveness characteristics, the higher the filter media aging or permanent clogging characteristics.

[0110] In one specific implementation, the processor 13 can correct the attribution probability or unnormalized diagnostic score of each candidate anomaly type based on ash accumulation rate characteristics, ash removal effectiveness characteristics, residual resistance characteristics, condensation risk characteristics, ash removal execution consistency characteristics, ash hopper blockage or material level anomaly characteristics, airflow anomaly characteristics, and filter media aging or permanent blockage characteristics. The correction can occur before or after probabilistic processing. When the correction occurs before probabilistic processing, the processor 13 can perform weighted increases, weighted decreases, threshold trigger adjustments, or rule constraint adjustments on the classification output or unnormalized diagnostic score of the pre-trained deep learning time-series attribution model, and input the corrected unnormalized diagnostic score into the probability output module for probabilistic processing. When the correction occurs after probabilistic processing, the processor 13 can perform weighted increases, weighted decreases, normalized and then redistributed, or threshold constraint processing on the attribution probabilities of each candidate anomaly type obtained from the probability output module.

[0111] In one specific implementation, the processor 13 corrects the attribution probability or unnormalized diagnostic score of each candidate anomaly type based on condensation risk characteristics, ash cleaning execution consistency characteristics, and ash hopper blockage or material level anomaly characteristics. Specifically, when at least one of the following conditions is met: condensation risk characteristics indicate increased condensation risk, ash cleaning effectiveness characteristics indicate decreased ash cleaning effect, residual resistance characteristics indicate increased residual resistance after ash cleaning, and ash cleaning execution consistency characteristics indicate normal ash cleaning execution, the processor 13 increases the attribution probability or unnormalized diagnostic score of filter media adhesion and blockage anomaly; when there is an ash cleaning action signal but the ash cleaning execution status data is abnormal, or the pressure difference decrease is lower than a preset threshold, the processor 13 increases the attribution probability or unnormalized diagnostic score of ash cleaning system failure; when the ash hopper status indicates an increase in ash hopper material level, or the ash unloading equipment status indicates abnormal operation of the ash unloading equipment, the processor 13 increases the attribution probability or unnormalized diagnostic score of ash hopper blockage or material level anomaly.

[0112] In one specific implementation, the processor 13 corrects the attribution probability or unnormalized diagnostic score of each candidate anomaly type based on ash accumulation rate characteristics, abnormal airflow characteristics, and filter media aging or permanent blockage characteristics. Specifically, when the ash accumulation rate characteristic indicates an increase in the rate of increase of differential pressure or normalized filter resistance after cleaning, and the cleaning effectiveness characteristic is within the normal range, the processor 13 increases the attribution probability or unnormalized diagnostic score of increased dust load; when the abnormal airflow characteristic indicates an increase in measured differential pressure, an increase in flue gas flow, and a failure of normalized filter resistance to increase synchronously, the processor 13 increases the attribution probability or unnormalized diagnostic score of abnormal airflow, and decreases the attribution probability or unnormalized diagnostic score of filter media adhesion and blockage anomalies and filter media aging or permanent blockage; when the cumulative operating time of the filter media exceeds the preset operating time, or the baseline of the residual differential pressure after cleaning continues to rise within multiple cleaning cycles, or the cleaning effectiveness characteristic declines for a long period, the processor 13 increases the attribution probability or unnormalized diagnostic score of filter media aging or permanent blockage.

[0113] In one specific implementation, after completing the above correction, the processor 13 can normalize the attribution probabilities of each candidate anomaly type so that the sum of the attribution probabilities of each candidate anomaly type satisfies a preset probability constraint; alternatively, the processor 13 can retain the corrected unnormalized diagnostic score, and the probability output module can perform Softmax normalization or probability calibration on the corrected unnormalized diagnostic score to obtain the corrected attribution probabilities of each candidate anomaly type. Through the above correction process, features with clear operational mechanism meanings can participate in the probability adjustment of candidate anomaly types, ensuring that the output results of the pre-trained deep learning time-series attribution model are consistent with the actual operational mechanism of the flue gas filter dust collector.

[0114] like Figures 3 to 10As shown, the time series characteristics of multi-source operational data corresponding to different pressure difference abnormality attribution types can be compared and observed through multiple time series channels. Figures 3 to 10 The horizontal direction is used to represent changes over time, and the vertical direction is used to represent changes in different multi-source operating data or derived features. Figures 3 to 10 The multi-source operating data time-series characteristics illustrated in the figure may include measured differential pressure, ash cleaning action signal, normalized filter resistance, flue gas outlet temperature, water dew point temperature, acid dew point temperature, ash hopper level, and outlet particulate matter concentration, etc.

[0115] Figures 3 to 10 The purpose is to illustrate that "increased measured differential pressure" can occur simultaneously with different combinations of multi-source operational data time-series characteristics under different candidate anomaly types. In other words, the processor 13 does not determine the anomaly type solely based on whether the measured differential pressure increases, but rather distinguishes candidate anomaly types based on the time-series characteristics of multi-source operational data, such as the change pattern of the measured differential pressure over time, the timing of the dust removal action signal, the differential pressure response after the dust removal action signal is triggered, the change in normalized filter resistance, the relationship between flue gas outlet temperature and dew point related data, changes in ash hopper level, and changes in outlet particulate matter concentration.

[0116] Specifically, normal filter cake formation typically manifests as a decrease in measured differential pressure after the cleaning action, followed by a gradual increase in the next filtration cycle as a dust layer forms on the filter media surface. Increased dust load is typically manifested as a stronger upward slope in the measured differential pressure or a greater upward trend in the normalized filtration resistance. Insufficient cleaning is typically manifested as a decreased drop in differential pressure after cleaning or an increase in residual differential pressure after cleaning. Cleaning system malfunctions are typically manifested as the presence of cleaning action signals, but abnormal cleaning execution status data or insufficient decrease in measured differential pressure after cleaning. Abnormal filter media adhesion and blockage are typically manifested as flue gas outlet temperature approaching the water dew point or acid dew point, increased condensation risk characteristics, accompanied by decreased cleaning effectiveness or increased residual resistance. Dust hopper blockage or abnormal material level is typically manifested as an increase in the dust hopper level, abnormal status of the ash discharge equipment, or a step change in the measured differential pressure. Abnormal airflow is typically manifested as an increase in measured differential pressure with increasing flue gas flow rate, induced draft fan frequency, or induced draft fan current, but without a corresponding increase in normalized filtration resistance. Filter media aging or permanent clogging is usually manifested by a long cumulative operating time of the filter media, a continuous increase in the baseline residual pressure difference after cleaning within multiple cleaning cycles, or a long-term decline in the cleaning effectiveness.

[0117] S6: Input the pre-trained deep learning temporal attribution model.

[0118] Processor 13 inputs the time-series feature matrix of the voltage difference anomaly into the pre-trained deep learning time-series attribution model. The pre-trained deep learning time-series attribution model is a deep learning model whose parameters are trained based on training samples with anomaly type labels before diagnosis.

[0119] The pre-trained deep learning temporal attribution model includes a temporal feature encoding module, an anomaly attribution classification module, and a probability output module. The temporal feature encoding module encodes the temporal feature matrix of differential pressure anomalies to obtain encoded features. The anomaly attribution classification module calculates the classification output for each candidate anomaly type based on the encoded features. The classification output is at least one of the following: diagnostic score, classification score, logistic value, or decision function value before probabilistic processing. The unnormalized diagnostic score has the same meaning as the classification output; it refers to at least one of the following: diagnostic score, classification score, logistic value, or decision function value output by the anomaly attribution classification module for each candidate anomaly type before probabilistic processing. The probability output module performs probabilistic processing on the unnormalized diagnostic score to obtain the attribution probability for each candidate anomaly type.

[0120] In one specific implementation, the temporal feature encoding module can employ a temporal convolutional network. The temporal convolutional network may include multiple one-dimensional convolutional layers, batch normalization layers, activation layers, random deactivation layers, and adaptive average pooling layers. After the processor 13 inputs the pressure difference anomaly temporal feature matrix into the temporal convolutional network, the network extracts temporal features along the time dimension, such as the pressure difference curve shape, dust removal response, residual resistance changes, temperature and dew point margin changes, ash hopper level changes, and airflow changes, and outputs encoded features. The anomaly attribution and classification module outputs the classification output corresponding to each candidate anomaly type based on the encoded features.

[0121] In another specific implementation, the temporal feature encoding module can employ a Long Short-Term Memory (LSTM) network. The LSM network receives the temporal feature matrix of the pressure difference anomaly and outputs the encoded features corresponding to the last hidden state. The temporal feature encoding module can also employ at least one of a recurrent neural network, a gated recurrent unit, or a Transformer encoder. All of these different network structures can be used to encode multivariate time series segments within a continuous time window, and their common function is to enable the pre-trained deep learning temporal attribution model to identify candidate anomaly types using time sequence and variable correlations.

[0122] S7: Obtain the attribution probability and the confidence level of the anomaly.

[0123] The pre-trained deep learning temporal attribution model outputs classification results for each candidate anomaly type. Processor 13 performs probabilistic processing on the classification output through a probability output module to obtain the attribution probability for each candidate anomaly type. Probabilistic processing may include Softmax normalization or probability calibration.

[0124] Processor 13 determines the anomaly confidence level based on at least one of the following: maximum attribution probability, attribution probability interval between different candidate anomaly types, input data integrity, and model output distribution. In one specific embodiment, processor 13 can identify the candidate anomaly type corresponding to the maximum attribution probability as the anomaly type corresponding to the increase in differential pressure of the flue gas filter dust collector, and use the maximum attribution probability as the anomaly confidence level. In another specific embodiment, processor 13 can modify the anomaly confidence level by combining the difference between the highest and second-highest attribution probabilities, input data integrity, or model output entropy.

[0125] The purpose of setting attribution probability and anomaly confidence is that processor 13 not only outputs a single anomaly type, but also the probabilistic results corresponding to each candidate anomaly type, and quantifies the reliability of the diagnostic results. Therefore, when the attribution probabilities of multiple candidate anomaly types are similar, the anomaly confidence can reflect the degree of distinction between candidate anomaly types; when the input data is incomplete or the model output distribution is scattered, the anomaly confidence can be reduced accordingly, thus avoiding the situation where only a single anomaly type is output and the diagnostic uncertainty cannot be reflected.

[0126] S8: Obtain the rule diagnosis result.

[0127] In one optional implementation, the processor 13 acquires the rule diagnosis results obtained based on the mechanism constraint rules. The rule diagnosis results include rule anomaly type, rule confidence level, and rule diagnosis basis information.

[0128] The rule-based diagnostic model can calculate the rule-based diagnostic score for each candidate anomaly type based on the characteristics of ash accumulation rate, ash removal effectiveness, residual resistance, condensation risk, consistency of ash removal execution, ash hopper blockage or abnormal material level, abnormal air volume, aging or permanent blockage of filter media, and trend characteristics within a continuous time window. The model then selects the candidate anomaly type with the highest rule-based diagnostic score as the rule anomaly type.

[0129] The mechanism constraint rules include at least one of the following rules: when the condensation risk characteristics are lower than the preset condensation risk threshold, the processor 13 reduces the rule diagnosis score or attribution probability of abnormal filter media adhesion and blockage; when the dust removal execution status data is normal and the differential pressure drops effectively after dust removal, the processor 13 reduces the rule diagnosis score or attribution probability of dust removal system failure; when the measured differential pressure increases but the normalized filter resistance does not increase synchronously, the processor 13 reduces the rule diagnosis score or attribution probability of abnormal filter media adhesion and blockage and filter media aging or permanent blockage; when the ash hopper status does not indicate an increase in ash hopper level and the ash unloading equipment does not show any abnormal operation, the processor 13 reduces the rule diagnosis score or attribution probability of ash hopper blockage or abnormal level.

[0130] The purpose of setting the rule-based diagnostic results is that the processor 13 can supplement the output results of the pre-trained deep learning time-series attribution model by utilizing the constraint relationships corresponding to the operating mechanism of the flue gas filter dust collector. The pre-trained deep learning time-series attribution model is used to learn the comprehensive patterns in multivariate time series segments, while the rule-based diagnostic results are used to reflect judgment conditions with clear mechanistic significance, such as normalized filter resistance, dust removal response, dew point margin, ash hopper status, and ash unloading equipment status. These two correspond to different information sources and judgment methods, providing a foundation for subsequent mixed diagnostic results.

[0131] When the pre-trained deep learning temporal attribution model is unavailable, model parameter loading fails, normalization parameter loading fails, label mapping relationship loading fails, or the current running data length is insufficient for the preset time window, the processor 13 can output a model unavailable status and call a rule-based diagnostic model based on mechanistic constraint rules to generate a backup pressure difference anomaly diagnostic result. Therefore, even when the pre-trained deep learning temporal attribution model is temporarily unavailable, the processor 13 can still output a backup pressure difference anomaly diagnostic result based on the rule-based diagnostic result.

[0132] S9: A mixed diagnostic result is obtained.

[0133] In one alternative implementation, the processor 13 fuses the rule-based diagnostic results with the anomaly type, attribution probability, and anomaly confidence output by the pre-trained deep learning temporal attribution model to obtain a hybrid diagnostic result.

[0134] When the rule-based anomaly type matches the anomaly type output by the pre-trained deep learning temporal attribution model, the processor 13 increases the confidence of the hybrid diagnostic result. This is because the rule-based diagnostic result and the output of the pre-trained deep learning temporal attribution model originate from mechanistic constraint rules and multivariate temporal pattern recognition, respectively. When both point to the same anomaly type, it indicates that the anomaly type simultaneously satisfies both mechanistic constraint conditions and temporal pattern conditions, thus increasing the confidence of the hybrid diagnostic result.

[0135] When the anomaly type of the rule is inconsistent with the anomaly type output by the pre-trained deep learning temporal attribution model, and the anomaly confidence of the pre-trained deep learning temporal attribution model is higher than the preset fusion threshold, the processor 13 prioritizes the anomaly type output by the pre-trained deep learning temporal attribution model. This is because, when multivariate temporal features are relatively complete, the model output distribution is relatively concentrated, and the maximum attribution probability is relatively high, the pre-trained deep learning temporal attribution model may better reflect the comprehensive change pattern within a continuous time window.

[0136] When the rule-based anomaly type is inconsistent with the anomaly type output by the pre-trained deep learning temporal attribution model, and the anomaly confidence of the pre-trained deep learning temporal attribution model is not higher than the preset fusion threshold, the processor 13 prioritizes the rule-based anomaly type. This is because when the output of the pre-trained deep learning temporal attribution model is not sufficiently concentrated or the anomaly confidence is insufficient, the rule-based diagnostic results based on mechanistic constraint rules can serve as a more robust diagnostic basis.

[0137] The preset fusion threshold is determined based on at least one of the following: validation set classification accuracy, macro-average F1 score, false positive rate constraint, and false negative rate constraint, or is a value in the range of 0.60 to 0.95. The backup pressure anomaly diagnosis results and the pressure anomaly diagnosis results output when the pre-trained deep learning temporal attribution model is available can be retained in parallel for the operation and maintenance system to use temporarily before the pre-trained deep learning temporal attribution model becomes available again, without changing the technical characteristic of this method that uses the pre-trained deep learning temporal attribution model as the main diagnostic path.

[0138] Through the above-described hybrid diagnostic process, the processor 13 can improve the diagnostic confidence when the pre-trained deep learning temporal attribution model and the rule-based diagnostic results are consistent, select the diagnostic source based on the abnormal confidence when they are inconsistent, and provide backup pressure difference anomaly diagnostic results when the pre-trained deep learning temporal attribution model is unavailable, thereby improving the stability and engineering applicability of the pressure difference anomaly diagnostic results.

[0139] S10: Output diagnostic results and recommendations.

[0140] Based on the anomaly type, attribution probability, and anomaly confidence, processor 13 outputs the differential pressure anomaly diagnosis results, diagnostic basis information, and operation and maintenance handling suggestions corresponding to the anomaly type through output interface 14.

[0141] The results of pressure difference anomaly diagnosis can include at least one of the following: anomaly type, attribution probability of each candidate anomaly type, anomaly confidence level, rule-based diagnosis results, mixed diagnosis results, and model availability status. Diagnostic basis information can include at least one of the following: cleaning effectiveness characteristics, residual resistance characteristics, condensation risk characteristics, cleaning execution consistency characteristics, ash hopper blockage or material level anomaly characteristics, airflow anomaly characteristics, filter media aging or permanent blockage characteristics, inlet dust concentration, production load, cleaning interval, cleaning fallback ratio, acid dew point margin, water dew point margin, ash hopper material level, ash discharge equipment current, flue gas flow rate, induced draft fan frequency, induced draft fan current, cumulative filter media runtime, outlet particulate matter concentration, and number of cleaning events.

[0142] Operation and maintenance recommendations are determined based on the type of anomaly. When the anomaly is normal filter cake formation, processor 13 outputs recommendations to maintain the existing dust removal strategy or optimize the dust removal cycle. When the anomaly is an increase in dust load, processor 13 outputs recommendations to check production load, inlet dust concentration, and raw material changes. When the anomaly is insufficient dust removal, processor 13 outputs recommendations to adjust the dust removal cycle, pulse-jet duration, or differential pressure trigger threshold. When the anomaly is a dust removal system malfunction, processor 13 outputs recommendations to check the dust removal actuator, pulse-jet pipe, air tank pressure, back-flushing mechanism, or rapping mechanism. When the anomaly is abnormal filter media adhesion and blockage, processor 13 outputs recommendations to check for moisture, adhesion, caking, or condensation deposits on the filter media surface, increase flue gas temperature, improve insulation and heat tracing, and check for air leaks. When the anomaly is ash hopper blockage or abnormal material level, processor 13 outputs recommendations to check the ash hopper level gauge, ash discharge valve, ash conveying equipment, and clean the ash hopper. When the anomaly type is airflow anomaly, processor 13 outputs suggestions to check the induced draft fan frequency, airflow setting, damper opening, and air-to-cloth ratio. When the anomaly type is filter media aging or permanent blockage, processor 13 outputs suggestions to check the filter media operating hours, historical maintenance records, and replacement or offline cleaning plans.

[0143] By outputting diagnostic information and operation and maintenance recommendations, processor 13 not only outputs the anomaly type corresponding to the increased differential pressure of the flue gas filter dust collector, but also outputs the corresponding evidence and handling direction. This reduces the risk of inaccurate judgment of the cause of the increased differential pressure in the flue gas filter dust collector and improper subsequent handling, which may result from relying solely on measured differential pressure increases for dust removal control or operational judgment.

[0144] Furthermore, this embodiment may also include a training process for a pre-trained deep learning time-series attribution model. The training process can be completed before online diagnosis of pressure difference anomalies in the flue gas filter dust collector, or it can be re-executed after obtaining new historical or simulated operating samples. The training process can be executed by the processor 13 or by a training device communicatively connected to the memory 12. After training, the obtained model parameters, standardized parameters, and label mapping relationships are stored in the memory 12 for the processor 13 to retrieve during online diagnosis.

[0145] In one specific embodiment, the processor 13 or training device acquires historical operating samples or simulated operating samples of the flue gas filter dust collector under different operating states. Historical operating samples refer to multi-source operating data collected during the actual operation of the flue gas filter dust collector; simulated operating samples refer to simulated operating data generated based on the operating mechanism of the flue gas filter dust collector, the range of on-site operating conditions, or known abnormal modes. Historical operating samples or simulated operating samples may include at least one of the following: measured differential pressure, flue gas flow rate, dust removal action signal, dust removal execution status data, flue gas temperature, flue gas humidity, dew point related data, ash hopper status, ash unloading equipment status, outlet particulate matter concentration, inlet dust concentration, production load, induced draft fan frequency, induced draft fan current, and cumulative operating time of the filter media.

[0146] In one specific implementation, the processor 13 or training device labels historical or simulated operating samples with anomaly type tags. These anomaly type tags characterize the actual anomaly types corresponding to the historical or simulated operating samples. The anomaly type tags include at least two of the following: normal filter cake formation, increased dust load, insufficient dust removal, dust removal system malfunction, abnormal filter media adhesion and clogging, ash hopper clogging or abnormal material level, abnormal airflow, and filter media aging or permanent clogging. These anomaly type tags are only used for training, validation, or testing of the pre-trained deep learning time-series attribution model and are not used as input features into the pre-trained deep learning time-series attribution model during online diagnosis.

[0147] In one specific implementation, the processor 13 or training device sorts historical or simulated running samples according to timestamps, and performs sliding window segmentation according to a preset window length and preset step size to obtain multiple multivariate time series training samples. Specifically, the processor 13 or training device constructs a time window with a consecutive preset number of sampling points, and moves the time window with a preset step size to generate multiple training samples. Each training sample corresponds to a multivariate time series segment within a consecutive time window. The processor 13 or training device can determine the anomaly type label at the end of the time window as the label of the training sample; alternatively, it can determine the anomaly type label that appears most frequently within the time window as the label of the training sample.

[0148] In one specific embodiment, the processor 13 or the training device extracts a differential pressure anomaly time-series feature matrix from each multivariate time-series training sample. The differential pressure anomaly time-series feature matrix may include measured differential pressure, flue gas flow rate, normalized filter resistance, ash removal action signal, ash removal execution status data, flue gas temperature, flue gas humidity, dew point related data, ash hopper status, ash discharge equipment status, outlet particulate matter concentration, and at least one of the following characteristics constructed from the above data: ash accumulation rate characteristic, ash removal effectiveness characteristic, residual resistance characteristic, condensation risk characteristic, ash removal execution consistency characteristic, ash hopper blockage or abnormal material level characteristic, abnormal airflow characteristic, and filter media aging or permanent blockage characteristic.

[0149] In one implementation, the processor 13 or training device divides multiple multivariate time series training samples into a training set, a validation set, and a test set. The processor 13 or training device fits standardized parameters to the numerical features in the training set and uses these standardized parameters for the validation set, the test set, and the online inference data. The standardized parameters may include at least one of the mean, variance, or standard deviation parameters of each input feature. By fitting standardized parameters only to the training set and using the same standardized parameters for the validation set, the test set, and the online inference data, the pre-trained deep learning time series attribution model can receive a stress-discretion anomaly time series feature matrix at a consistent numerical scale during training, evaluation, and online diagnostics.

[0150] In one specific implementation, the processor 13 or training device trains a deep learning model using a pressure difference anomaly temporal feature matrix as input and anomaly type labels as supervision signals to obtain a pre-trained deep learning temporal attribution model. The deep learning model may include at least one of a recurrent neural network, a long short-term memory network, a gated recurrent unit, a temporal convolutional network, or a Transformer encoder. During training, the processor 13 or training device may use a cross-entropy loss function or other classification loss function to calculate the error between the model output and the anomaly type labels, and update at least one of the weight parameters, bias parameters, and normalization layer parameters in the deep learning model through an optimizer. The processor 13 or training device may also determine whether the training has converged based on the validation set loss, validation set classification accuracy, or macro-average F1 score, or perform early stopping.

[0151] In one specific implementation, after training is completed, the processor 13 or training device stores the model parameters, normalized parameters, and label mapping relationships of the pre-trained deep learning time-series attribution model in the memory 12. The model parameters may include at least one of the following: neural network state parameters, model type, input feature dimension, number of candidate anomaly types, hidden layer dimension, random deactivation ratio, preset window length, and feature column set. The normalized parameters are used to normalize the time-series feature matrix of pressure differential anomalies during the online inference stage. The label mapping relationship is used to establish the correspondence between candidate anomaly types and category numbers, and is used to restore the category numbers output by the pre-trained deep learning time-series attribution model to anomaly types during the online diagnosis stage. Through the above training process, the processor 13 can call the pre-trained deep learning time-series attribution model with completed parameter training during the online diagnosis process to infer the anomaly type corresponding to the increase in pressure differential of the flue gas filter dust collector.

[0152] like Figure 11 As shown, the output interface 14 can display the diagnostic results of the pre-trained deep learning time-series attribution model output by the processor 13 in the form of a graphical interface. Figure 11 The left-hand area is used to display a summary of diagnostic results, which may include at least one of the following: anomaly type, anomaly confidence level, model type, input window length, and diagnostic basis information. The right-hand area is used to display the attribution probability corresponding to each candidate anomaly type. In embodiments without probabilistic processing, the unnormalized diagnostic score corresponding to each candidate anomaly type may also be displayed. Figure 11 In the example shown, the processor 13 performs online inference based on the pressure difference anomaly time series feature matrix corresponding to the most recent continuous time window. It finds that the attribution probability corresponding to "filter media aging or permanent blockage" is the highest. Therefore, "filter media aging or permanent blockage" is identified as an anomaly type, and the corresponding anomaly confidence level of 0.99 is output.

[0153] Specifically, Figure 11 The right-hand bar chart illustrates the attribution probability or diagnostic score output by the pre-trained deep learning time-series attribution model for each candidate anomaly type. Each candidate anomaly type can include normal filter cake formation, increased dust load, insufficient dust removal, dust removal system malfunction, abnormal filter media adhesion and clogging, hopper blockage or abnormal material level, abnormal airflow, and filter media aging or permanent blockage. The processor 13 can determine the candidate anomaly type with the highest attribution probability as the anomaly type, and determine the anomaly confidence level as the maximum attribution probability or the result corrected based on the maximum attribution probability, the attribution probability interval between different candidate anomaly types, input data integrity, and model output distribution. Thus, Figure 11 It can not only reflect the final anomaly type, but also the relative distinction between each candidate anomaly type.

[0154] Figure 11 The "Model Type" field indicates the specific network type of the pre-trained deep learning temporal attribution model being invoked, such as a Temporal Convolutional Network (TCNN). TCNNs encode temporal features of the differential pressure anomaly temporal feature matrix. They can extract multivariate change patterns within a continuous time window through one-dimensional convolutional layers and input the extracted encoded features into the anomaly attribution classification module to obtain the classification output, attribution probability, and anomaly confidence for each candidate anomaly type. The "Input Window Length" field indicates the number of sampling points contained in the continuous time window used for this online inference, such as 60 points. The "Diagnostic Basis Information" field describes the features or variables comprehensively considered by the pre-trained deep learning temporal attribution model during diagnosis, such as differential pressure, dust removal response, temperature and humidity, dew point margin, dust removal execution status data, ash hopper status, and dust load. Through the above display content, the operation and maintenance system or display terminal can intuitively understand the model input range, model output results, and diagnostic basis corresponding to this diagnosis.

[0155] It should be noted that, Figure 11 The `known_scenario` field displayed is an anomaly type label field, used to characterize the true anomaly type corresponding to the training sample, validation sample, or test sample. The `known_scenario` is only used for training, evaluation, or simulation data display of the pre-trained deep learning time-series attribution model and is not used as an input feature in the online inference stage. That is, when performing online diagnosis, the processor 13 does not input `known_scenario` into the pre-trained deep learning time-series attribution model. Instead, it outputs the anomaly type, the attribution probability of each candidate anomaly type, and the anomaly confidence level based solely on the measured pressure difference, flue gas flow rate, ash removal action signal, ash removal execution status data, flue gas temperature, flue gas humidity, dew point related data, ash hopper status, ash unloading equipment status, outlet particulate matter concentration, and the pressure difference anomaly time-series feature matrix constructed from these parameters. Figure 11 The display of known_scenario is only for illustrating the use of this field during the training and evaluation phases and does not constitute a limitation on the input features for online inference.

[0156] Furthermore, Figure 11 This is merely an illustration of one display method for output interface 14 and does not constitute a limitation on the scope of protection of this invention. In other embodiments, the anomaly type, attribution probability, anomaly confidence level, and diagnostic basis information output by processor 13 may also be displayed or output in the form of lists, tables, dashboards, alarm pages, report pages, or industrial monitoring screens.

[0157] The present invention has been described above. Those skilled in the art will be able to implement the present invention based on these descriptions. All other embodiments obtained by those skilled in the art based on the foregoing content of this specification without inventive effort should fall within the scope of patent protection.

Claims

1. A method for diagnosing abnormal pressure differences in a flue gas filter dust collector, wherein the flue gas filter dust collector is a dust removal device with a dust removal actuator and a filter medium for capturing particulate matter in flue gas, characterized in that: include: Multi-source operational data is collected during the operation of the flue gas filter dust collector. The multi-source operational data includes at least the measured differential pressure, flue gas flow rate, dust removal action signal, dust removal execution status data, flue gas temperature, flue gas humidity, dew point related data, ash hopper status, ash unloading equipment status, and outlet particulate matter concentration. The dew point related data includes at least one of the following: water dew point temperature, acid dew point temperature, flue gas humidity data used to calculate water dew point temperature, and acid gas concentration data used to calculate acid dew point temperature. The filtration velocity or air-to-cloth ratio is determined based on the flue gas flow rate and the effective filtration area of ​​the flue gas filter dust collector, and the measured pressure difference is normalized based on the filtration velocity or air-to-cloth ratio to obtain the normalized filtration resistance. A dust removal event window is constructed using the dust removal action signal as the event anchor point, and dust removal response features are extracted in the dust removal event window. The dust removal response features include at least the pressure difference before dust removal, the pressure difference after dust removal, the dust removal drop amplitude, the dust removal drop ratio, the pressure difference rise slope, and the residual pressure difference after dust removal. Based on the normalized filter resistance, the dust removal response characteristics, and the multi-source operating data, a pressure difference anomaly time series feature matrix is ​​constructed. The pressure difference anomaly time series feature matrix includes features at multiple moments within a continuous time window. The continuous time window is a multivariate time series segment extracted from the multi-source operating data according to a preset window length and a preset step size. The features include features corresponding to candidate anomaly types among features such as ash accumulation rate, dust removal effectiveness, residual resistance, condensation risk, dust removal execution consistency, ash hopper blockage or material level anomaly, airflow anomaly, and filter media aging or permanent blockage. The pressure difference anomaly time-series feature matrix is ​​input into a pre-trained deep learning time-series attribution model to obtain the anomaly type corresponding to the pressure difference increase in the flue gas filter dust collector and the attribution probability of each candidate anomaly type. The anomaly confidence is determined based on at least one of the following: maximum attribution probability, attribution probability interval between different candidate anomaly types, input data integrity, and model output distribution. The attribution probability is obtained by probabilistically processing the classification output of the pre-trained deep learning time-series attribution model for each candidate anomaly type. The classification output, also known as the unnormalized diagnostic score, refers to the score obtained by the pre-trained deep learning time-series attribution model before probabilistic processing. The classification score, logistic value, or decision function value output for each candidate anomaly type; the pre-trained deep learning time-series attribution model is a deep learning model that completes parameter training based on training samples with anomaly type labels before diagnosis; the candidate anomaly types include at least two of the following: normal filter cake formation, increased dust load, insufficient dust removal, dust removal system failure, abnormal filter media adhesion and clogging, ash hopper clogging or abnormal material level, abnormal air volume, and filter media aging or permanent clogging; the abnormal filter media adhesion and clogging includes at least one of the following: condensation adhesion and clogging, moisture absorption caking and clogging, acidic condensate adhesion and clogging, and oil mist or tar adhesion and clogging. Based on the anomaly type, attribution probability, and anomaly confidence level, output the pressure difference anomaly diagnosis result, diagnosis basis information, and operation and maintenance handling suggestions corresponding to the anomaly type.

2. The method according to claim 1, characterized in that: The flue gas filtration dust collector is a flue gas filtration dust removal device that uses at least one of the following filter media: filter bag, filter cylinder, and filter plate. The dust removal actuator includes at least one of a back-blowing dust removal mechanism and a rapping dust removal mechanism; wherein, the back-blowing dust removal mechanism includes at least one of a pulse jet dust removal mechanism and a back-blowing air dust removal mechanism; When the dust removal actuator is a pulse jet dust removal mechanism, the dust removal execution status data includes at least one of the following: jet pressure, air tank pressure, pulse valve status, pulse valve current, and jet duration. When the dust removal actuator is a back-blowing dust removal mechanism or a rapping dust removal mechanism, the dust removal execution status data includes at least one of the following: back-blowing air pressure, back-blowing air volume, rapping motor status, rapping current, rapping duration, and dust removal feedback signal.

3. The method according to claim 1, characterized in that: The filtration velocity or air-to-cloth ratio is determined based on the flue gas flow rate and the effective filtration area of ​​the flue gas filter dust collector. The measured pressure difference is then normalized based on the filtration velocity or air-to-cloth ratio to obtain the normalized filtration resistance, including: The filtration velocity or air-to-cloth ratio is determined based on the flue gas flow rate and the effective filtration area of ​​the flue gas filter dust collector; the effective filtration area is the area of ​​the filter media in the flue gas filter dust collector that is in the filtration state, or the effective filtration area is the total filter media area minus at least one of the following: offline compartment area, isolated filter unit area, and damaged filter unit area. The normalized filtration resistance is obtained by using a first normalization method; the first normalization method includes determining the normalized filtration resistance as the ratio of the measured pressure difference to the filtration velocity or air-to-cloth ratio. Alternatively, the normalized filter resistance can be obtained using a second normalization method; the second normalization method includes: determining the reference pressure difference corresponding to the current filter velocity or air-cloth ratio based on the fitting relationship between the pressure difference and the filter velocity or air-cloth ratio under historical normal operating conditions, and determining the deviation or deviation ratio between the measured pressure difference and the reference pressure difference as the normalized filter resistance. When the measured differential pressure increases, the flue gas flow rate increases, and the normalized filter resistance does not increase synchronously, the attribution probability or unnormalized diagnostic score for abnormal airflow is increased, and the attribution probability or unnormalized diagnostic score for abnormal filter media adhesion and blockage and filter media aging or permanent blockage is reduced; the "unsynchronized increase" means that the change in normalized filter resistance is less than the change in measured differential pressure, or less than the synchronous change threshold obtained from historical normal operating conditions.

4. The method according to claim 1, characterized in that: A dust removal event window is constructed using the dust removal action signal as the event anchor point, and dust removal response features are extracted from the dust removal event window, including: Use the moment that triggers the dust removal action as the event anchor point; Obtain the differential pressure data within a preset time period before the event anchor point to determine the differential pressure before dust removal; Obtain the differential pressure data within a preset time after the event anchor point to determine the differential pressure after dust removal; The dust removal fall-off amplitude and dust removal fall-off ratio are determined based on the pressure difference before and after dust removal. The slope of the pressure difference increase and the residual pressure difference after dust removal are determined based on the pressure difference change trend after the event anchor point. The dust removal drop amplitude, dust removal drop ratio, pressure difference rise slope, and residual pressure difference after dust removal are used to construct the dust removal effectiveness characteristics, residual resistance characteristics, and dust removal execution consistency characteristics. The dust removal response features are mapped to the corresponding timestamps of the corresponding dust removal events to the corresponding moments in the continuous time window, or the dust removal response features are filled, interpolated, preserved, or statistically aggregated within the dust removal event window to form the event response feature channel of the pressure difference abnormal time series feature matrix. The preset duration is determined based on the dust removal action response time, dust removal cycle, sampling cycle, or historical normal dust removal response data.

5. The method according to claim 4, characterized in that: The process for determining the condensation risk characteristics includes: when the water dew point temperature is obtained, determining the water dew point margin based on the difference between the flue gas outlet temperature and the water dew point temperature; when the acid dew point temperature is obtained, determining the acid dew point margin based on the difference between the flue gas outlet temperature and the acid dew point temperature; when the acid dew point temperature is not directly obtained, estimating the acid dew point temperature based on at least one of SO2 concentration, estimated SO3 value, HCl concentration, and flue gas humidity, and determining the acid dew point margin based on the difference between the flue gas outlet temperature and the estimated acid dew point temperature; determining the condensation risk characteristics based on the obtained or estimated water dew point margin, acid dew point margin, and flue gas humidity; wherein, the condensation risk characteristics characterizing an increase in condensation risk include at least one of: water dew point margin being lower than a preset water dew point margin threshold, acid dew point margin being lower than a preset acid dew point margin threshold, and flue gas humidity being higher than a preset humidity threshold. The process of determining the consistency characteristics of the dust removal execution includes: determining whether the dust removal execution status data is within the preset normal range of the corresponding dust removal actuator after the dust removal action signal is triggered, or whether a state change matching the dust removal action signal occurs; determining whether the measured differential pressure or normalized filter resistance decreases by a preset amount after the dust removal action signal is triggered; and determining the consistency characteristics of the dust removal execution based on the dust removal execution status data and the decrease in differential pressure. The process of determining the characteristics of ash hopper blockage or abnormal material level includes: determining the characteristics of ash hopper blockage or abnormal material level based on the state of the ash hopper and the state of the ash unloading equipment, wherein the state of the ash hopper includes at least the ash hopper material level and the state of the ash unloading equipment includes at least the ash unloading equipment current or the ash unloading equipment operating status. Based on the condensation risk characteristics, the dust removal execution consistency characteristics, and the ash hopper blockage or abnormal material level characteristics, the attribution probability or unnormalized diagnostic score of each candidate abnormality type is corrected. The correction includes: increasing the attribution probability or unnormalized diagnostic score of filter media adhesion and blockage abnormality when at least one of the following conditions is met: the condensation risk characteristic indicates an increased condensation risk, the dust removal effectiveness characteristic indicates a decreased dust removal effect, the residual resistance characteristic indicates an increased residual resistance after dust removal, and the dust removal execution consistency characteristic indicates normal dust removal execution; increasing the attribution probability or unnormalized diagnostic score of dust removal system failure when there is a dust removal action signal and the dust removal execution status data is abnormal, or the pressure difference decrease is lower than a preset threshold; and increasing the attribution probability or unnormalized diagnostic score of ash hopper blockage or abnormal material level when the ash hopper status indicates an increased ash hopper material level, or the ash unloading equipment status indicates abnormal operation of the ash unloading equipment.

6. The method according to claim 4, characterized in that: The process of determining the ash accumulation rate characteristic includes: determining the ash accumulation rate characteristic based on the growth slope of the measured pressure difference or normalized filter resistance per unit time after ash removal; The process of determining the abnormal air volume characteristics includes: determining the abnormal air volume characteristics based on at least one of the measured pressure difference change range, flue gas flow rate change range, induced draft fan frequency change range, induced draft fan current change range, and normalized filter resistance change range. The process of determining the aging or permanent clogging characteristics of the filter media includes: determining the aging or permanent clogging characteristics of the filter media based on at least one of the following: cumulative operating time of the filter media, baseline drift of residual pressure difference after cleaning in multiple cleaning cycles, long-term upward trend of normalized filtration resistance, and long-term downward trend of cleaning effectiveness characteristics. Based on the ash accumulation rate characteristics, the abnormal airflow characteristics, and the aging or permanent blockage characteristics of the filter media, the attribution probability or unnormalized diagnostic score of each candidate abnormality type is corrected. The correction includes: increasing the attribution probability or unnormalized diagnostic score of the increased dust load when the ash accumulation rate characteristic indicates an increase in the rate of increase in differential pressure or normalized filter resistance after dust removal, and the dust removal effectiveness characteristic is within the normal range; and increasing the attribution probability or unnormalized diagnostic score of the abnormal airflow when the abnormal airflow characteristic indicates an increase in measured differential pressure, an increase in flue gas flow rate, and a failure to synchronously increase the normalized filter resistance. The system normalizes diagnostic scores and reduces the attribution probability or unnormalized diagnostic score for abnormal filter media adhesion and clogging, as well as filter media aging or permanent clogging. When the cumulative operating time of the filter media exceeds a preset operating time, or the baseline of residual pressure difference after cleaning continues to rise within multiple cleaning cycles, or the cleaning effectiveness characteristics decline over a long period, the system increases the attribution probability or unnormalized diagnostic score for filter media aging or permanent clogging. The preset operating time is determined based on the filter media's design life, historical maintenance records, historical normal operating condition statistics, equipment operating experience values, or training sample quantiles.

7. The method according to claim 1, characterized in that: The training process of the pre-trained deep learning temporal attribution model includes: Obtain historical or simulated operating samples of the flue gas filter dust collector under different operating conditions; The historical or simulated operation samples are labeled with anomaly type tags, which include at least two of the following: normal filter cake formation, increased dust load, insufficient dust removal, dust removal system failure, abnormal filter media adhesion and blockage, ash hopper blockage or abnormal material level, abnormal air volume, and filter media aging or permanent blockage. According to the preset window length and preset step size, the historical running samples or simulated running samples are divided by a sliding window to obtain multiple multivariate time series training samples. Extract the differential pressure abnormal time series feature matrix from each multivariate time series training sample; Using the pressure difference anomaly time series feature matrix as input and the anomaly type label as supervision signal, the deep learning model is trained to obtain the pre-trained deep learning time series attribution model; The sliding window segmentation includes: sorting the historical running samples or simulation running samples according to timestamps; forming a time window with a preset number of consecutive sampling points; moving the time window with a preset step size to generate multiple training samples; determining the anomaly type label at the end of the time window as the label of the training sample; fitting standardized parameters to the numerical features in the training set, and using the standardized parameters for the validation set, test set, and online inference data. The preset window length and preset step size are determined based on at least one of the sampling period, dust removal period, anomaly duration, and model validation set performance.

8. The method according to claim 1, characterized in that: The pre-trained deep learning temporal attribution model includes a temporal feature encoding module, an anomaly attribution classification module, and a probability output module. The temporal feature encoding module is used to perform temporal feature encoding on the differential pressure abnormal temporal feature matrix to obtain encoded features; The anomaly attribution classification module is used to calculate the classification output of each candidate anomaly type according to the encoded features. The classification output is at least one of the diagnostic score, classification score, logical value or decision function value before probabilistic processing. The unnormalized diagnostic score has the same meaning as the classification output, referring to at least one of the diagnostic score, classification score, logical value, or decision function value output by the anomaly attribution classification module for each candidate anomaly type before probabilistic processing. The probability output module is used to perform probabilistic processing on the unnormalized diagnostic score to obtain the attribution probability of each candidate anomaly type, and to determine the anomaly confidence based on at least one of the following: the maximum attribution probability, the difference between the highest and second-highest attribution probabilities, the completeness of the input data, and the model output entropy. The probabilistic processing includes Softmax normalization or probability calibration; the temporal feature encoding module includes at least one of recurrent neural networks, long short-term memory networks, gated recurrent units, temporal convolutional networks, or Transformer encoders.

9. The method according to claim 1, characterized in that: After obtaining the anomaly type, the attribution probability of each candidate anomaly type, and the anomaly confidence level, the following steps are also included: Obtain rule-based diagnostic results from mechanistic constraint rules, and fuse these results with the anomaly type, attribution probability, and anomaly confidence level output by the pre-trained deep learning temporal attribution model to obtain a hybrid diagnostic result. The rule-based diagnostic result includes rule anomaly type, rule confidence level, and rule diagnostic basis information. Specifically, when the rule-based anomaly type matches the anomaly type output by the pre-trained deep learning temporal attribution model, the confidence level of the hybrid diagnostic result is increased; when they are inconsistent and the anomaly confidence level output by the pre-trained deep learning temporal attribution model is higher than a preset fusion threshold, the anomaly type output by the pre-trained deep learning temporal attribution model is preferred; when they are inconsistent and the anomaly confidence level output by the pre-trained deep learning temporal attribution model is not higher than the preset fusion threshold, the rule-based anomaly type is preferred; the preset fusion threshold is determined based on at least one of the validation set classification accuracy, macro-average F1 score, false positive rate constraint, and false negative rate constraint, or is a value in the range of 0.60 to 0.

95. The mechanism constraint rules include at least one of the following rules: when the condensation risk characteristics are lower than the preset condensation risk threshold, reduce the rule diagnosis score or attribution probability of abnormal filter media adhesion and blockage; when the dust removal execution status data is normal and the differential pressure drops effectively after dust removal, reduce the rule diagnosis score or attribution probability of dust removal system failure; when the measured differential pressure increases but the normalized filter resistance does not increase synchronously, reduce the rule diagnosis score or attribution probability of abnormal filter media adhesion and blockage and filter media aging or permanent blockage; when the ash hopper status does not indicate an increase in ash hopper level and the ash unloading equipment does not show any abnormal operation, reduce the rule diagnosis score or attribution probability of ash hopper blockage or abnormal level. When the pre-trained deep learning time-series attribution model is missing, model parameter loading fails, standardized parameter loading fails, label mapping relationship loading fails, or the current running data length is insufficient for the preset time window, the model is output as unavailable, and a rule-based diagnostic model based on mechanistic constraint rules is invoked to generate a backup pressure difference anomaly diagnostic result. The backup pressure difference anomaly diagnostic result is retained in parallel with the pressure difference anomaly diagnostic result output when the pre-trained deep learning time-series attribution model is available, for temporary use by the operation and maintenance system before the pre-trained deep learning time-series attribution model is restored to availability, without changing the technical feature of this method that uses the pre-trained deep learning time-series attribution model as the main diagnostic path. The operation and maintenance recommendations are determined based on the anomaly type, including: when the anomaly type is normal filter cake formation, outputting a recommendation to maintain the existing dust removal strategy or optimize the dust removal cycle; when the anomaly type is increased dust load, outputting a recommendation to check production load, inlet dust concentration, and raw material changes; when the anomaly type is insufficient dust removal, outputting a recommendation to adjust the dust removal cycle, pulse jet duration, or differential pressure trigger threshold; when the anomaly type is a dust removal system malfunction, outputting a recommendation to check the dust removal actuator, pulse jet pipe, air tank pressure, backflushing mechanism, or rapping mechanism; when the anomaly type is filter media adhesion... When a blockage occurs, the output should indicate whether there is moisture, adhesion, caking, or condensation deposits on the surface of the filter media, and provide suggestions to increase the flue gas temperature, improve insulation and heat tracing, and check for air leaks. When the abnormality is ash hopper blockage or material level abnormality, the output should indicate whether to check the ash hopper level gauge, ash discharge valve, ash conveying equipment, and clean the ash hopper. When the abnormality is airflow abnormality, the output should indicate whether to check the induced draft fan frequency, airflow setting, damper opening, and air-to-cloth ratio. When the abnormality is filter media aging or permanent blockage, the output should indicate whether to check the filter media operating hours, historical maintenance records, and replacement or offline cleaning plans.

10. A device for diagnosing abnormal pressure differences in a flue gas filter dust collector, characterized in that: Includes data acquisition interface, processor, memory, and output interface; The data acquisition interface communicates with the multi-source sensors of the flue gas filter dust collector, DCS system, PLC system or industrial gateway to acquire multi-source operating data during the operation of the flue gas filter dust collector. The memory stores computer programs, model parameters of a pre-trained deep learning time-series attribution model, standardized parameters, and label mapping relationships. The processor is used to execute the computer program to call the model parameters, normalization parameters and label mapping relationship of the pre-trained deep learning time series attribution model, and to implement the flue gas filter dust collector pressure difference abnormality diagnosis method based on the pre-trained deep learning time series attribution model as described in any one of claims 1 to 9. The output interface is used to output the abnormal voltage diagnosis results, diagnostic basis information, and operation and maintenance handling suggestions.

Images