Power plant boiler fault automatic alarm method and system based on frequency characteristics

Abstract

The application discloses a power plant boiler fault automatic alarm method and system based on frequency characteristics, and relates to the technical field of power plant equipment fault diagnosis and predictive maintenance. The method comprises the following steps: collecting boiler operation data in real time, and constructing a training sample set containing frequency domain labels; constructing a hybrid deep learning framework for performing collaborative training on the training sample set through a joint optimization strategy, wherein the hybrid deep learning framework is composed of a first model and a second model; calculating reconstruction errors between actual observation values and model prediction values and generating a residual sequence; analyzing the residual sequence by adopting an exponential weighted moving average control chart algorithm, dynamically generating an adaptive alarm threshold, and controlling the execution of the operation of triggering a fault alarm according to the situation that the residual sequence continuously exceeds the adaptive alarm threshold; and identifying key variables causing the reconstruction errors, combining fault source positioning, and performing visual display.

Description

Technical Field

[0001] This application relates to the technical field of fault diagnosis and predictive maintenance of thermal power plant equipment, and in particular to an automatic alarm method and system for power plant boiler faults based on frequency characteristics. Background Technology

[0002] Boilers are the core equipment of thermal power plants. Their key components (such as coal mills, fans, desuperheating water regulating valves, air preheaters, and superheater tube panels) are subjected to harsh conditions of high temperature, high pressure, and high wear for a long time, resulting in frequent failures.

[0003] Existing fault alarm technologies suffer from the following main drawbacks: 1. Limitations of time-domain analysis: Traditional methods primarily analyze signals in the time domain, making it difficult to capture subtle changes in the system's dynamic characteristics during fault development. Many mechanical faults (such as wear, imbalance, and jamming) manifest as changes in specific frequency components in the frequency domain, which time-domain analysis struggles to effectively identify. 2. Neglecting system evolution patterns: Boiler systems undergo slow performance evolution over long-term operation (e.g., equipment aging, ash accumulation, and wear). Traditional methods compare the current state with a fixed baseline, easily misjudging normal evolution as a fault, leading to a high false alarm rate. 3. Insufficient sensitivity to early faults: Existing methods are insensitive to early, subtle fault characteristics, often triggering alarms only when the fault has significantly worsened, failing to provide sufficient early warning time. 4. Weak fault tracing capabilities: Traditional alarms only indicate anomalies but cannot pinpoint specific fault sources or associated sensor variables, requiring maintenance personnel to spend considerable time troubleshooting.

[0004] Therefore, there is an urgent need for a solution that can overcome the above-mentioned shortcomings. Summary of the Invention

[0005] This application proposes an automatic alarm method and system for power plant boiler faults based on frequency characteristics, which is used to overcome the deficiencies of the prior art.

[0006] According to a first aspect of the embodiments of this application, an automatic fault alarm method for power plant boilers based on frequency characteristics is provided, comprising: Boiler operation data is collected in real time, and the operation data is cleaned, normalized and time-series aligned. Frequency domain analysis and fault features are extracted from the time-series signals in the operation data to construct a training sample set containing frequency domain labels. A hybrid deep learning framework is constructed to perform collaborative training on the training sample set through a joint optimization strategy. The hybrid deep learning framework consists of a first model and a second model. The first model is used to learn the frequency domain evolution law of the normal operation of the power plant boiler system. The second model is used to extract the typical change pattern of the power plant boiler system in the frequency domain before and after the fault occurs and to enhance the output of the first model, so as to improve the sensitivity of the hybrid deep learning framework to the fault to the target value. Real-time data is input into the trained hybrid deep learning framework to calculate the reconstruction error between the actual observations and the model predictions and to generate a residual sequence. The residual sequence is analyzed using an exponentially weighted moving average control chart algorithm to dynamically generate an adaptive alarm threshold. The execution of the operation to trigger a fault alarm is controlled based on the condition that the residual sequence continuously exceeds the adaptive alarm threshold. Based on the attention mechanism in the hybrid deep learning framework, key variables that cause reconstruction errors are identified, and the fault source is located by combining the device topology relationship. Alarm information, key variables, frequency domain feature changes and trend curves are visualized.

[0007] In some implementations, performing frequency domain analysis on the time-series signals in the runtime data includes: By employing fast Fourier transform, wavelet transform, or power spectral density estimation mechanisms, time-series signals are converted to frequency domain signals and analyzed. The amplitude changes, frequency drifts, or harmonic component increases of specific frequency components exhibited by key components during degradation or failure are extracted and analyzed.

[0008] In some embodiments, the method further includes: The time-series deep learning model trained based on historical normal operation data is used as the first model to establish the operating benchmark under multiple frequency bands and output the probability distribution of anomalies in each frequency band.

[0009] In some embodiments, the method further includes: The second model is built based on a convolutional neural network or attention mechanism network trained on the fault dataset. This model is used to learn typical variation patterns of frequency domain features from the fault data and to correct the output of the first model.

[0010] In some implementations, performing collaborative training on the training sample set using a joint optimization strategy includes: The first model and the second model are trained together using a total loss function, which includes the loss of the first model, the loss of the second model, and the consistency constraint loss between the two.

[0011] In some implementations, the operation of controlling the triggering of a fault alarm based on the determination that the residual sequence continuously exceeds the adaptive alarm threshold includes: If the reconstruction error value of the residual sequence at multiple consecutive sampling times is greater than the adaptive alarm threshold, and the duration of the continuous exceedance exceeds the preset fault tolerance time window, the operation to trigger a fault alarm will be executed; otherwise, the operation to trigger a fault alarm will not be executed.

[0012] In some implementations, identifying key variables that lead to reconstruction errors based on the attention mechanism in the hybrid deep learning framework includes: The attention mechanism in the hybrid deep learning framework outputs and identifies the contribution weight distribution of each input variable to the reconstruction error, and selects input variables whose weights exceed the target threshold or sensor variables that undergo abrupt changes before and after the fault as key variables causing the reconstruction error.

[0013] According to a second aspect of this application, an automatic alarm system for power plant boiler faults based on frequency characteristics is provided, comprising: The data acquisition and preprocessing module is used to acquire boiler operation data in real time, and to clean, normalize and align the operation data, and to perform frequency domain analysis and extract fault features from the time-series signals in the operation data, in order to construct a training sample set containing frequency domain labels. A hybrid deep learning modeling module is used to construct a hybrid deep learning framework for performing collaborative training on the training sample set through a joint optimization strategy. The hybrid deep learning framework consists of a first model and a second model. The first model is used to learn the frequency domain evolution law of the normal operation of the power plant boiler system, and the second model is used to extract the typical change pattern of the power plant boiler system in the frequency domain before and after the fault occurs and enhance the output of the first model, thereby improving the sensitivity of the hybrid deep learning framework to faults to the target value. The reconstruction error calculation module is used to input real-time data into the trained hybrid deep learning framework, and to calculate the reconstruction error between the actual observed values ​​and the model predicted values, as well as to generate a residual sequence. The dynamic threshold alarm module is used to analyze the residual sequence using an exponentially weighted moving average control chart algorithm, dynamically generate an adaptive alarm threshold, and control the execution of the operation to trigger a fault alarm based on the judgment that the residual sequence continuously exceeds the adaptive alarm threshold. The fault tracing and visualization module is used to identify key variables that cause reconstruction errors based on the attention mechanism in the hybrid deep learning framework, locate the fault source by combining the device topology relationship, and visualize alarm information, key variables, frequency domain feature changes and trend curves.

[0014] In some implementations, the boiler operation data collection targets include at least one key component among the following: coal mill, induced draft fan, forced draft fan, desuperheating water regulating valve, superheater, reheater, air preheater, and water-cooled wall.

[0015] In some implementations, the dynamic threshold alarm module is also used to calculate an exponentially weighted moving average statistic and dynamically adjust the adaptive alarm threshold based on the historical distribution of the exponentially weighted moving average statistic.

[0016] The beneficial effects of the automatic alarm method and system for power plant boiler faults based on frequency characteristics in this application include at least the following: This application's embodiments capture early, subtle fault characteristics through frequency domain analysis. Combined with fault feature enhancement from Model B, early warnings can be issued in the early stages of fault development, with an advance warning time of several hours to several days, improving the timeliness of warnings. This application's embodiments learn the normal evolution patterns of the system through Model A, distinguishing between normal aging and real faults; EWMA dynamic threshold filtering removes noise interference, significantly reducing the false alarm rate. This application's embodiments automatically identify key variables through an attention mechanism, quickly locating the fault source and reducing maintenance personnel's troubleshooting time by more than 50%, achieving accurate fault tracing. This application's embodiments, through a dual-model architecture, can adapt to the operating characteristics of different units, different operating conditions, and different coal qualities, exhibiting good generalization ability and strong adaptability. This application's embodiments provide visualized information such as frequency domain feature changes, residual trends, and key variable weights, providing intuitive basis for maintenance decisions and completing visualized decision support. Attached Figure Description

[0017] Figure 1 This is a flowchart illustrating an automatic alarm method for power plant boiler faults based on frequency characteristics, according to an embodiment of this application. Figure 2 This is a schematic diagram of the structure of the hybrid deep learning framework according to an embodiment of this application; Figure 3 This is a schematic diagram of the alarm visualization interface according to an embodiment of this application; Figure 4 This is a schematic diagram of the structure of an automatic alarm system for power plant boiler faults based on frequency characteristics, according to an embodiment of this application. Detailed Implementation

[0018] To make the objectives, technical solutions, and advantages of the embodiments of this disclosure clearer, the various embodiments of this disclosure will be described in detail below with reference to the accompanying drawings. However, those skilled in the art will understand that many technical details have been provided in the various embodiments of this disclosure to facilitate a better understanding of the disclosure. However, the technical solutions claimed in this disclosure can be implemented even without these technical details and with various variations and modifications based on the following embodiments. The division of the various embodiments below is for ease of description and should not constitute any limitation on the specific implementation of this disclosure. The various embodiments can be combined with and referenced by each other without contradiction.

[0019] Therefore, the following detailed description of the embodiments of the present application provided in the accompanying drawings is not intended to limit the scope of the claimed embodiments of the present application, but merely to illustrate selected embodiments of the present application. Other embodiments obtained by those skilled in the art based on the embodiments of the present application without inventive effort are all within the scope of protection of the embodiments of the present application.

[0020] This application discloses an automatic alarm method and system for power plant boiler faults based on frequency characteristics. The automatic alarm method for power plant boiler faults based on frequency characteristics is implemented based on the automatic alarm system for power plant boiler faults based on frequency characteristics. The purpose is to overcome the shortcomings of the prior art and to provide early warning of faults in key components of the boiler through timing signals, thereby improving the accuracy and timeliness of the alarm.

[0021] See attached document Figure 1 The diagram shows a flowchart illustrating an embodiment of the automatic alarm method for power plant boiler faults based on frequency characteristics. This automatic alarm method for power plant boiler faults based on frequency characteristics includes the following steps S1 to S5.

[0022] Step S1: Collect boiler operation data in real time, and clean, normalize and align the operation data. Perform frequency domain analysis and extract fault features from the time series signals in the operation data to construct a training sample set (i.e., multi-dimensional time series samples) containing frequency domain labels.

[0023] For example, the boiler operating data includes historical operating data and real-time operating data of the power plant boiler. Both the historical operating data and the real-time operating data include analog data and digital data.

[0024] For example, the analog data includes continuously changing parameters such as temperature, pressure, flow rate, current, vibration, differential pressure, and valve position; the digital data includes discrete state quantities such as equipment start / stop status, valve open / close status, and protection action signals.

[0025] The data collection covers key boiler components, including but not limited to: coal mill, induced draft fan, forced draft fan, desuperheating water regulating valve, superheater, reheater, air preheater, water-cooled wall, etc.

[0026] In some implementations, the objects for collecting the boiler operating data include, but are not limited to, at least one key component among: coal mill, induced draft fan, forced draft fan, desuperheating water regulating valve, superheater, reheater, air preheater, and water-cooled wall.

[0027] In some implementations, the cleaning, normalization, and time-series alignment of the operational data include: removing data from downtime periods, abnormal jump data caused by sensor malfunctions, and missing data caused by communication interruptions to complete the cleaning; normalization: standardizing the data of each measurement point according to the mean and standard deviation to eliminate the influence of dimensions to complete the normalization; and aligning data with different sampling frequencies to the same time axis to complete the time-series alignment.

[0028] In some implementations, the frequency domain analysis of the time-series signal in the operating data includes: using a Fast Fourier Transform (FFT), wavelet transform, or power spectral density estimation mechanism to convert the time-series signal to a frequency domain signal and perform frequency domain signal analysis, and extracting and analyzing the amplitude changes, frequency drifts, or harmonic component increases of specific frequency components exhibited by key components during degradation or failure.

[0029] For example, the training sample set is a training sample set containing frequency domain labels. Each sample in the training sample set includes at least: original time domain data, frequency domain features, fault type label, and fault severity label.

[0030] Step S2: Construct a hybrid deep learning framework for performing collaborative training on the training sample set through a joint optimization strategy; wherein the hybrid deep learning framework consists of a first model (model A) and a second model (model B). The first model is used to learn the frequency domain evolution law of the normal operation of the power plant boiler system, and the second model is used to extract the typical change pattern of the power plant boiler system in the frequency domain before and after the fault occurs and enhance the output of the first model, so as to improve the sensitivity of the hybrid deep learning framework to the fault to the target value.

[0031] In some implementations, the method further includes: using a time-series deep learning model trained based on historical normal operation data as the first model to establish an operating benchmark under multiple frequency bands and output the probability distribution of anomalies in each frequency band.

[0032] In some implementations, the method further includes: building a second model based on a convolutional neural network or attention mechanism network trained on the fault dataset, for learning typical variation patterns of frequency domain features from the fault data and correcting the output of the first model.

[0033] For example, Model A can be understood as a steady-state evolution model. Its function is to learn the performance evolution patterns of a power plant boiler system during long-term operation based on historical normal operating data, and to establish operating benchmarks across multiple frequency bands. Model A employs temporal deep learning models such as Long Short-Term Memory (LSTM), Temporal Convolutional Network (TCN), or Transformer. The input is historical multivariate time-series data, and the output is the predicted feature distribution for each frequency band. The core functions of Model A include at least: learning the frequency domain feature evolution patterns of a power plant boiler system under normal operating conditions, which at least include the impact of slow changes such as equipment aging, ash accumulation, and wear on frequency domain characteristics; and outputting the probability distribution of anomalies in each frequency band as a benchmark reference for fault diagnosis.

[0034] For example, Model B can be understood as a fault feature enhancement model. Model B's function is to extract typical frequency domain change patterns of the system before and after a fault, based on a fault dataset, to correct and enhance the output of Model A. Model B's model structure employs a convolutional neural network (CNN) or an attention mechanism network specifically designed to extract fault-related frequency domain feature patterns. The core functions of Model B include at least: learning typical frequency domain feature change patterns from fault data when a fault occurs; correcting the output of Model A to enhance the model's sensitivity to early faults; and reducing the interference of normal operating condition fluctuations on fault detection.

[0035] In some implementations, performing collaborative training on the training sample set through a joint optimization strategy includes: collaboratively training a first model and a second model through a total loss function, wherein the total loss function includes the loss of the first model, the loss of the second model, and the consistency constraint loss between the two.

[0036] For example, Model A and Model B are co-trained based on the following mathematical expression: Total loss function = α × Model A loss + β × Model B loss + γ × Consistency constraint loss; Model A is responsible for fitting the frequency domain evolution law under normal operating conditions; Model B is responsible for enhancing the ability to express fault characteristics; through joint optimization, the power plant boiler system can simultaneously possess the ability to accurately fit the normal evolution and the ability to respond quickly to fault characteristics.

[0037] Step S3: Input real-time data into the trained hybrid deep learning framework to calculate the reconstruction error between the actual observations and the model predictions, and to generate a residual sequence.

[0038] For example, the model prediction value is the reconstructed prediction value output by the hybrid model (i.e., the hybrid deep learning framework) after real-time running data is input into the trained hybrid model, including: the normal expected amplitude of each frequency band; the expected range of change of each frequency band; and the corrected prediction value after fault enhancement.

[0039] For example, the reconstruction error is determined based on both the frequency domain reconstruction error and the comprehensive reconstruction error.

[0040] Among them, frequency domain reconstruction error Calculated based on the following mathematical expression:

[0041] in, The function then represents taking the absolute value; Among them, the comprehensive reconstruction error is calculated based on the comprehensive error index formed by weighted fusion of multi-band errors.

[0042] For example, generating the residual sequence includes: arranging the reconstruction errors at consecutive time points in chronological order to form a residual sequence, which is represented, for example, as: ,in, Indicates duration, This represents the sequence length. The residual sequence reflects the degree of deviation between the current operating state of the power plant boiler system and the learned normal evolution pattern and fault characteristics; a continuous increase in residual indicates that the state of the power plant boiler system is gradually deviating from the normal baseline, and there may be a fault evolution trend.

[0043] Step S4: The residual sequence is analyzed using the Exponential Weighted Moving Average (EWMA) control chart algorithm to dynamically generate an adaptive alarm threshold. The execution of the operation to trigger a fault alarm is controlled based on the condition that the residual sequence continuously exceeds the adaptive alarm threshold.

[0044] In some implementations, the operation of controlling the triggering of a fault alarm based on the determination that the residual sequence continuously exceeds the adaptive alarm threshold includes: if the reconstruction error value of the residual sequence at multiple consecutive sampling times is greater than the adaptive alarm threshold, and the duration of the continuous exceedance exceeds a preset fault tolerance time window, the operation of controlling the triggering of a fault alarm is executed; otherwise, the operation of controlling the triggering of a fault alarm is not executed.

[0045] For example, the EWMA control chart algorithm is used to perform (real-time statistical) analysis on the residual sequence to dynamically generate adaptive alarm thresholds. It is determined based on the following mathematical expression:

[0046] in, For smoothing parameters , Indicates duration.

[0047] For example, the alarm threshold is dynamically adjusted based on the historical distribution of the EWMA statistic, specifically calculated based on the following mathematical expression:

[0048] in, For alarm thresholds, The mean of the EWMA statistic is... Standard deviation, This is the gain coefficient.

[0049] For example, a system fault is determined and an alarm is triggered when the following conditions are met: the residual continuously exceeds the dynamic threshold; the duration of the out-of-limit error exceeds the preset fault tolerance time window (such as 30 consecutive seconds or 5 consecutive sampling points). This mechanism effectively filters out transient noise interference and reduces the false alarm rate.

[0050] Step S5: Based on the attention mechanism in the hybrid deep learning framework, identify the key variables that cause reconstruction errors, locate the fault source by combining the device topology relationship, and visualize the alarm information, key variables, frequency domain feature changes and trend curves.

[0051] In some implementations, identifying key variables that lead to reconstruction errors based on the attention mechanism in the hybrid deep learning framework includes: The attention mechanism in the hybrid deep learning framework outputs and identifies the contribution weight distribution of each input variable to the reconstruction error, and selects input variables whose weights exceed the target threshold or sensor variables that undergo abrupt changes before and after the fault as key variables causing the reconstruction error.

[0052] In some implementations, identifying key variables leading to reconstruction errors based on the attention mechanism in the hybrid deep learning framework further includes: outputting and identifying the contribution weight distribution of each input variable to the reconstruction error through the attention mechanism in the hybrid deep learning framework, and selecting sensor combination variables that match the prior knowledge of the fault type as key variables leading to reconstruction errors.

[0053] In some implementations, locating the fault source by combining equipment topology relationships, that is, locating potential fault sources by combining key variable information with equipment topology relationships, includes: identifying the equipment or components associated with abnormal sensors; outputting possible fault types and confidence levels; and providing fault troubleshooting suggestions.

[0054] In some implementations, visualizing alarm information, key variables, frequency domain characteristic changes, and trend curves involves pushing the following information to the monitoring interface to achieve a visual representation of fault tracing: alarm information, including: fault time, fault type, and alarm level; key variables, including: the name of the sensor causing the anomaly, its current value, and its normal range; frequency domain characteristic changes, including: a comparison of frequency domain spectra before and after the fault; trend curves, including: residual sequences, dynamic thresholds, and trend graphs of periods exceeding limits; and troubleshooting suggestions, including: recommended checkpoints and handling measures.

[0055] This application's embodiments capture early, subtle fault characteristics through frequency domain analysis. Combined with fault feature enhancement from Model B, early warnings can be issued in the early stages of fault development, with an advance warning time of several hours to several days, improving the timeliness of warnings. This application's embodiments learn the normal evolution patterns of the system through Model A, distinguishing between normal aging and real faults; EWMA dynamic threshold filtering removes noise interference, significantly reducing the false alarm rate. This application's embodiments automatically identify key variables through an attention mechanism, quickly locating the fault source and reducing maintenance personnel's troubleshooting time by more than 50%, achieving accurate fault tracing. This application's embodiments, through a dual-model architecture, can adapt to the operating characteristics of different units, different operating conditions, and different coal qualities, exhibiting good generalization ability and strong adaptability. This application's embodiments provide visualized information such as frequency domain feature changes, residual trends, and key variable weights, providing intuitive basis for maintenance decisions and completing visualized decision support.

[0056] According to a second aspect of this application, an automatic alarm system for power plant boiler faults based on frequency characteristics is provided, comprising: The data acquisition and preprocessing module is used to acquire boiler operation data in real time, and to clean, normalize and align the operation data, and to perform frequency domain analysis and extract fault features from the time-series signals in the operation data, in order to construct a training sample set containing frequency domain labels. A hybrid deep learning modeling module is used to construct a hybrid deep learning framework for performing collaborative training on the training sample set through a joint optimization strategy. The hybrid deep learning framework consists of a first model and a second model. The first model is used to learn the frequency domain evolution law of the normal operation of the power plant boiler system, and the second model is used to extract the typical change pattern of the power plant boiler system in the frequency domain before and after the fault occurs and enhance the output of the first model, thereby improving the sensitivity of the hybrid deep learning framework to faults to the target value. The reconstruction error calculation module is used to input real-time data into the trained hybrid deep learning framework, and to calculate the reconstruction error between the actual observed values ​​and the model predicted values, as well as to generate a residual sequence. The dynamic threshold alarm module is used to analyze the residual sequence using an exponentially weighted moving average control chart algorithm, dynamically generate an adaptive alarm threshold, and control the execution of the operation to trigger a fault alarm based on the judgment that the residual sequence continuously exceeds the adaptive alarm threshold. The fault tracing and visualization module is used to identify key variables that cause reconstruction errors based on the attention mechanism in the hybrid deep learning framework, locate the fault source by combining the device topology relationship, and visualize alarm information, key variables, frequency domain feature changes and trend curves.

[0057] In some implementations, the boiler operation data collection targets include at least one key component among the following: coal mill, induced draft fan, forced draft fan, desuperheating water regulating valve, superheater, reheater, air preheater, and water-cooled wall.

[0058] In some implementations, the dynamic threshold alarm module is also used to calculate an exponentially weighted moving average statistic and dynamically adjust the adaptive alarm threshold based on the historical distribution of the exponentially weighted moving average statistic.

[0059] The data acquisition and preprocessing module of this application embodiment can acquire boiler operation data in real time, perform cleaning, normalization, and time-series alignment, with a focus on frequency domain analysis and label construction. The hybrid deep learning modeling module of this application embodiment consists of Model A (steady-state evolution model) and Model B (fault feature enhancement model), trained collaboratively through a joint optimization strategy. The reconstruction error calculation module of this application embodiment inputs real-time data into the model, calculates the reconstruction error between actual observed values ​​and predicted values, and forms a residual sequence. The dynamic threshold alarm module of this application embodiment uses the EWMA control chart algorithm to dynamically generate adaptive thresholds and determine whether to trigger an alarm. The fault tracing and visualization module of this application embodiment utilizes an attention mechanism to identify key variables, locate the fault source, and visualize the results.

[0060] To further illustrate the technical solutions of the embodiments of this application in specific operations, the embodiments of this application are specifically described based on the following embodiments 1 to 3. It is understood that the solutions of embodiments 1 to 3 do not affect the protection scope of this application, and are only used as illustrative examples of specific operations of the embodiments of this application.

[0061] Example 1 describes a coal mill roller wear fault alarm. The scheme described in this application is deployed on the coal mill of a 600MW unit boiler. It includes: collecting analog quantities such as coal mill current, vibration, inlet / outlet pressure difference, coal feed rate, and inlet / outlet air temperature at a 1-second sampling period, as well as switch quantities such as coal mill start / stop status, to complete data acquisition. The vibration signal is subjected to FFT transformation to extract the power spectral density characteristics in the 0-200Hz frequency band. Historical data shows that in the early stages of roller wear, the vibration signal exhibits an amplitude increase trend in the 50-80Hz frequency band, which is used for frequency domain analysis. The model training in Example 1 includes: Model A is trained using 3 months of normal operation data after a major overhaul to learn the frequency domain characteristic evolution law of the coal mill under different loads and coal feed rates; Model B is trained using historical fault data (roller wear case) to learn the frequency domain change pattern of the wear process. The alarm effect includes at least the following: when the coal mill is in operation for 8 months, Model B detects that the amplitude of the 50-80Hz frequency band is continuously rising, and Model A outputs that the probability of this frequency band deviating from the normal evolution range reaches 85%; the reconstruction error continuously exceeds the EWMA dynamic threshold, and a level 2 alarm is triggered after exceeding the limit for 30 consecutive seconds; the attention mechanism shows that "vibration signal (50-80Hz frequency band)" and "coal mill current" contribute the highest weight, and are located as "coal mill A grinding roller wear"; compared with the traditional vibration threshold alarm, the warning is issued about 72 hours earlier.

[0062] Example 2 is an alarm for a stuck desuperheating water regulating valve. This system, as described in this application, is deployed on the desuperheating water regulating valve of a 330MW unit boiler. The system includes: collecting valve position commands, valve position feedback, desuperheating water flow rate, upstream pressure, and downstream pressure at a sampling period of 0.5 seconds to complete data acquisition. Frequency domain analysis is performed on the valve position feedback signal. In the early stages of stuckness, an increase in the low-frequency oscillation component of the command-feedback deviation signal is observed, which is used to complete the frequency domain analysis. The model training in Example 2 includes: Model A learning the frequency domain response characteristics of the command-feedback during normal valve operation; and Model B learning the low-frequency oscillation characteristic pattern of the deviation signal during a stuck fault. The alarm effect includes at least the following: the system detects periodic oscillations in the valve position feedback deviation signal in the 0.1-0.5Hz frequency band, and Model B outputs an enhanced fault characteristic signal; the reconstruction error continues to exceed the threshold, triggering an alarm; the attention mechanism identifies "valve position feedback" and "desuperheating water flow" as key variables, positioning it as "superheater desuperheating water regulating valve jamming trend"; and maintenance personnel find slight jamming in the valve stem during inspection and handle it in a timely manner.

[0063] Example 3 is an air preheater blockage fault alarm. The scheme described in this application is deployed on the air preheater of a 1000MW ultra-supercritical unit boiler. This includes: collecting differential pressure on the flue gas side, primary air side, secondary air side, drive current, and load commands at a 5-second sampling period to complete data acquisition. In the initial stage of air preheater blockage, the differential pressure signal shows a trend of increasing in the low-frequency range (<0.01Hz), while the power frequency fluctuation amplitude of the current signal increases, used for frequency domain analysis. The model training in Example 3 includes: Model A: learning the evolution law of the differential pressure baseline and frequency domain characteristics of the air preheater under different loads; Model B: learning the characteristic patterns of low-frequency trend changes in differential pressure and enhanced current fluctuations during blockage faults. The alarm effect includes at least the following: the system detects a continuous deviation of the residual differential pressure on the flue gas side, and Model B identifies a typical blockage frequency domain characteristic pattern; the reconstruction error exceeds the threshold, triggering a level 2 alarm; the attention mechanism identifies "flue gas side differential pressure" and "drive current" as key variables, positioning it as "blockage trend on the flue gas side of air preheater A"; maintenance personnel increase the frequency of soot blowing, the differential pressure returns to normal, and the unit avoids load limiting.

[0064] This application first extracts fault features using frequency domain analysis and constructs a training sample set containing frequency domain labels. Next, Model A learns the frequency domain evolution patterns of normal system operation, while Model B enhances the sensitivity to fault features. Then, fault alarms are implemented based on reconstruction error and dynamic thresholds. Finally, an attention mechanism is used to trace the fault source.

[0065] The above description is merely a specific implementation of the embodiments of this application, but the protection scope of the embodiments of this application is not limited thereto. Any simple modifications, changes, and equivalent changes made by those skilled in the art to the above embodiments based on the technical essence of the embodiments of this application within the technical scope disclosed in the embodiments of this application shall still fall within the protection scope of the technical solution of the embodiments of this application.

Claims

1. An automatic fault alarm method for power plant boilers based on frequency characteristics, characterized in that, include: Boiler operation data is collected in real time, and the operation data is cleaned, normalized and time-series aligned. Frequency domain analysis and fault features are extracted from the time-series signals in the operation data to construct a training sample set containing frequency domain labels. A hybrid deep learning framework is constructed to perform collaborative training on the training sample set through a joint optimization strategy. The hybrid deep learning framework consists of a first model and a second model. The first model is used to learn the frequency domain evolution law of the normal operation of the power plant boiler system. The second model is used to extract the typical change pattern of the power plant boiler system in the frequency domain before and after the fault occurs and to enhance the output of the first model, so as to improve the sensitivity of the hybrid deep learning framework to the fault to the target value. Real-time data is input into the trained hybrid deep learning framework to calculate the reconstruction error between the actual observations and the model predictions and to generate a residual sequence. The residual sequence is analyzed using an exponentially weighted moving average control chart algorithm to dynamically generate an adaptive alarm threshold. The execution of the operation to trigger a fault alarm is controlled based on the condition that the residual sequence continuously exceeds the adaptive alarm threshold. Based on the attention mechanism in the hybrid deep learning framework, key variables that cause reconstruction errors are identified, and the fault source is located by combining the device topology relationship. Alarm information, key variables, frequency domain feature changes and trend curves are visualized.

2. The method according to claim 1, characterized in that, The frequency domain analysis of the time-series signals in the running data includes: By employing fast Fourier transform, wavelet transform, or power spectral density estimation mechanisms, time-series signals are converted to frequency domain signals and analyzed. The amplitude changes, frequency drifts, or harmonic component increases of specific frequency components exhibited by key components during degradation or failure are extracted and analyzed.

3. The method according to claim 1, characterized in that, The method further includes: The time-series deep learning model trained based on historical normal operation data is used as the first model to establish the operating benchmark under multiple frequency bands and output the probability distribution of anomalies in each frequency band.

4. The method according to claim 1, characterized in that, The method further includes: The second model is built based on a convolutional neural network or attention mechanism network trained on the fault dataset. This model is used to learn typical variation patterns of frequency domain features from the fault data and to correct the output of the first model.

5. The method according to claim 1, characterized in that, The step of performing collaborative training on the training sample set through a joint optimization strategy includes: The first model and the second model are trained together using a total loss function, which includes the loss of the first model, the loss of the second model, and the consistency constraint loss between the two.

6. The method according to claim 1, characterized in that, The operation of controlling the triggering of a fault alarm based on the condition that the residual sequence continuously exceeds the adaptive alarm threshold includes: If the reconstruction error value of the residual sequence at multiple consecutive sampling times is greater than the adaptive alarm threshold, and the duration of the continuous exceedance exceeds the preset fault tolerance time window, the operation to trigger a fault alarm will be executed; otherwise, the operation to trigger a fault alarm will not be executed.

7. The method according to claim 1, characterized in that, The attention mechanism in the hybrid deep learning framework identifies key variables that lead to reconstruction errors, including: The attention mechanism in the hybrid deep learning framework outputs and identifies the contribution weight distribution of each input variable to the reconstruction error, and selects input variables whose weights exceed the target threshold or sensor variables that undergo abrupt changes before and after the fault as key variables causing the reconstruction error.

8. An automatic alarm system for power plant boiler faults based on frequency characteristics, characterized in that, include: The data acquisition and preprocessing module is used to acquire boiler operation data in real time, and to clean, normalize and align the operation data, and to perform frequency domain analysis and extract fault features from the time-series signals in the operation data, in order to construct a training sample set containing frequency domain labels. A hybrid deep learning modeling module is used to construct a hybrid deep learning framework for performing collaborative training on the training sample set through a joint optimization strategy. The hybrid deep learning framework consists of a first model and a second model. The first model is used to learn the frequency domain evolution law of the normal operation of the power plant boiler system, and the second model is used to extract the typical change pattern of the power plant boiler system in the frequency domain before and after the fault occurs and enhance the output of the first model, thereby improving the sensitivity of the hybrid deep learning framework to faults to the target value. The reconstruction error calculation module is used to input real-time data into the trained hybrid deep learning framework, and to calculate the reconstruction error between the actual observed values ​​and the model predicted values, as well as to generate a residual sequence. The dynamic threshold alarm module is used to analyze the residual sequence using an exponentially weighted moving average control chart algorithm, dynamically generate an adaptive alarm threshold, and control the execution of the operation to trigger a fault alarm based on the judgment that the residual sequence continuously exceeds the adaptive alarm threshold. The fault tracing and visualization module is used to identify key variables that cause reconstruction errors based on the attention mechanism in the hybrid deep learning framework, locate the fault source by combining the device topology relationship, and visualize alarm information, key variables, frequency domain feature changes and trend curves.

9. The automatic alarm system for power plant boiler faults based on frequency characteristics according to claim 8, characterized in that, The boiler operation data collection targets include at least one key component from the following: coal mill, induced draft fan, forced draft fan, desuperheating water regulating valve, superheater, reheater, air preheater, and water-cooled wall.

10. The automatic alarm system for power plant boiler faults based on frequency characteristics according to claim 8, characterized in that, The dynamic threshold alarm module is also used to calculate the exponentially weighted moving average statistic and dynamically adjust the adaptive alarm threshold according to the historical distribution of the exponentially weighted moving average statistic.

Images