Thermal power generation coal-fired boiler fault risk early warning method based on period-trend network

Abstract

The invention discloses a thermal power generation coal-fired boiler fault risk early warning method based on a period-trend network, and the method comprises the steps: extracting a period signal of original data from a frequency domain, and dividing a period window; and designing a trend network, and predicting a future single period window working condition based on historical period window data. And extracting a characteristic normal form of periodic data from historical normal periodic data, calculating a normal uncertainty range of a future trend, and sending a fault alarm for a prediction result exceeding a reasonable range. The related method is verified in an overtemperature fault instance of the coal-fired boiler in the thermal power generation process, and effective early warning of the fault risk of the thermal power generation coal-fired boiler based on the period-trend network is achieved.

Description

Technical Field

[0001] The present invention designs an industrial process failure risk warning method, in particular a thermal power generation coal-fired boiler failure risk warning method based on a cycle-trend network Background Art

[0002] In modern industrial processes, the technology of identifying faults through real-time monitoring data is becoming increasingly mature. However, sometimes the occurrence of a fault has already brought irreparable losses. In fact, in some complex industrial processes, some parameters have already become abnormal before the fault occurs. Therefore, monitoring abnormal signals before the fault occurs, predicting fault trends, and anticipating the occurrence of faults are more practical in maintaining the smooth operation of industrial processes. Early research on fault warning relied on expert knowledge, and established physical models by analyzing the structure of the equipment to detect abnormal signals. Data-driven fault warning is lower cost and has better performance than expert knowledge-based methods. Data-driven fault warning methods can be roughly divided into three categories:

[0003] 1) Learn process features from historical data to classify process states. For example, support vector machines (SVMs) map operating data to a hypersphere to classify states, determine abnormal states, and warn of possible faults. Artificial neural networks (ANNs) are used to detect changes in photovoltaic systems and classify arc fault states. 2) Signal processing techniques are used to analyze the frequency domain information of process data to obtain features that are sensitive to abnormal signals. For example, feature vectors are constructed from the time and frequency domains of the data to reconstruct low-dimensional sensitive features for fault warning. Cost-sensitive gradient boosting decision trees (CS-GBDTs) use sensitive features to assist classifier training and improve fault warning accuracy. 3) Learn the serial correlation of historical data to predict future data and provide fault warnings. For example, CNN-LSTM captures the spatiotemporal features of process data to predict future operating conditions. D-former captures the long-term dependencies and degradation characteristics of the process from the original signal to achieve gear life prediction and fault warning.

[0004] However, in non-stationary industrial processes, the operating conditions are complex and difficult to identify. Therefore, the above methods still have the following limitations in non-stationary processes:

[0005] 1) State classification methods struggle to accurately identify process states within complex nonstationary phenomena. 2) Nonstationary characteristics mask abnormal signals, making them difficult to detect. 3) Time series prediction models are susceptible to the impact of prediction step size, and unstable data characteristics further reduce model accuracy. With increasing computing power, some research has begun to investigate operating condition prediction and risk warning under nonstationary conditions. For example, the causal-gated time series transformer (CGTST) utilizes causal gate structures to assist in time series modeling, improving causal discovery and time series prediction capabilities for nonstationary process data. The broad learning system (BLS) and relevance vector machine (RVM) combine statistical learning and machine learning techniques to interpret nonstationary characteristics and achieve short-term wind speed forecasting. However, all of these studies use time series models to predict future observations for fault warning. In complex nonstationary processes, faults are often hidden deeper than surface data changes. Furthermore, the mean and variance of nonstationary data vary over time, further complicating fault warning. Therefore, there is an urgent need for a method to deeply explore the characteristics of non-stationary data and analyze and model them to achieve early fault warning for non-stationary industrial processes.

[0006] To address the above issues, this chapter fully utilizes the trend and periodic characteristics that are prevalent in non-stationary processes and proposes a period-trend network to extract trend features from non-stationary sub-periods. The trend features of historical sub-periods are then transferred to the sub-period recursive network to achieve trend prediction and fault warning for future operating conditions. The proposed trend network has the following advantages:

[0007] (1) This scheme does not require clear state labels and can continuously train the model from continuously input online data, adjust control limits, improve warning accuracy, and reduce false alarms; (2) Trend characteristics indicate changes and trends in future working conditions. Trend characteristics within the sub-period are extracted through the trend network, and trend feature information interaction of different sub-periods is realized in the sub-period recursive network; (3) Single-period step-length prediction is performed backward based on the sub-period recursive network, avoiding the influence of step-length on prediction accuracy in traditional time series prediction models; (4) Trend constraints are designed as part of the loss function for accurate extraction of trend characteristics, while also enhancing the proposed interpretability. Summary of the Invention

[0008] 1. A method for early warning of coal-fired boiler failures in thermal power generation based on a cycle-trend network. The method is characterized by extracting periodic signals from raw data in the frequency domain to divide the data into periodic windows; inputting sub-periodic data into the cycle-trend network to obtain sub-periodic trend characteristics and predict future operating conditions; and extracting characteristic patterns of periodic data from historical normal periodic data to calculate the normal uncertainty range of future trends. A fault alarm is then issued for prediction results that exceed a reasonable range.

[0009] 2. The method for early warning of coal-fired boiler failure risk in thermal power generation based on a cycle-trend network according to claim 1, characterized in that:

[0010] For sampled data N is the observed variable and T is the sampling length. First, the periodic signal is extracted from the frequency domain by FFT. The frequency data obtained by FFT is recorded as In the frequency domain, the inverse of the frequency corresponding to the maximum amplitude is the main periodic signal of the data. The maximum amplitude is calculated by the mean of N frequency domain signals: in is the maximum main frequency value, u x (i)∈U x , i=1,…,N. X in The cycle length in yes The corresponding frequency. X in The set X is re-divided into multiple sub-periods according to the period length = {x1, x2, ..., x P},in P is the number of sub-periods, p = 1,…, P. Subsequently, the input data after period re-division will be used to extract sub-period trend features using Trendnet.

[0011] 3. The method for early warning of coal-fired boiler failure risk in thermal power generation based on a cycle-trend network according to claim 2, characterized in that:

[0012] A temporal attention convolutional network (TACN) is designed to extract the trend information and period information of the current sub-period data. TACN consists of two branches: an expanded causal convolutional network for trend information extraction and a short-term dynamic network for dynamic feature extraction. For the input sub-period x p , obtain trend information t through the trend extraction branch of TACN p =ReLU(Ψ(x p )),in It is the trend information produced by TACN, and m is the dimension of the hidden layer. Ψ(·) is the dilated causal convolution operation. The convolution result is activated by the ReLU(·) function to prevent the gradient explosion when the trend information is transmitted in the subsequent recursive network. Dynamic information is extracted by the self-attention mechanism of another branch. In the self-attention layer, the input sub-period data is first mapped into three key values ​​Q (query), K (key), and V (value) through the fully connected layer, where Q represents the information requirement of the current position for other positions, the other position information in K is used to match Q, and V is the information that needs to be extracted. The dynamic information is finally output in the form of attention weights. Dynamic information in It is dynamic information. Similar to the extraction of trend information, the self-attention weight is also activated by ReLU(·).

[0013] 4. The method for early warning of coal-fired boiler failure risk in thermal power generation based on a cycle-trend network according to claim 3, characterized in that:

[0014] The trend information is placed in the trend network to obtain the overall trend characteristics and describe the overall trend of the working condition. Each sub-cycle trend information obtained in claim 2 is regarded as a recursive step and learned in the trend recursive network. The trend characteristics after recursive learning can be expressed as τ p =B·t p +C·τ p-1 +ε, where is the trend feature after recursion and information sharing, B and C represent the recursively calculated transfer matrix, and ε is the residual. Similar to traditional recursive networks, the trend features of a sub-period are only related to the trend information at the current moment and the trend features at the previous moment. The dynamic information obtained by TACN also represents the dynamic information within the current sub-period. The impact of dynamic information is short-lived and rapid, so there is no need to share information through the recursive network. The sub-period trend features and dynamic information are used to predict the operating data of the next sub-period. A fully connected layer is used to implement this function: in At this time, the working condition data of the next sub-cycle is obtained by recursively learning the trend characteristics of historical trend information through the trend network p and the dynamic characteristics of the current sub-period d p Joint decision.

[0015] 5. The method for early warning of coal-fired boiler failure risk in thermal power generation based on a cycle-trend network according to claim 4, characterized in that:

[0016] A loss function is designed to constrain the reconstruction effect of the cycle step. The model constraints of the trend network include two loss functions: an error loss function used to improve the model's prediction accuracy and a trend slow-changing constraint loss function to improve the interpretability of trend features. The MSE loss function is selected for calculating the error loss function. The model is trained iteratively by comparing the difference between the predicted value and the true value of the working condition data of the next sub-cycle. The prediction error loss function of the trend network is as follows: This loss function strengthens the trend feature τ i-1 and dynamic information i-1 The ability to indicate future cycle-step operating conditions. Another trend feature slow-changing constraint function is used to train trend features to clarify their physical meaning. In long-term dependencies, trend features need to represent the overall trend of the process, so trend features are usually flat. Therefore, trend features need to change as little as possible in different sub-cycles. Loss constraints are implemented by comparing the trend feature distances between adjacent sub-cycles: Where distance(·) represents the distance calculation operation of the trend feature. In this chapter, Euclidean distance is selected to calculate the distance between features.

[0017] 6. The method for early warning of coal-fired boiler failure risk in thermal power generation based on a cycle-trend network according to claim 5, characterized in that:

[0018] Establish a fault warning model based on trend features. Through the prediction loss function constraint, the current cycle trend features have the ability to indicate future working conditions. Establish a fault warning model based on trend features. When the online data runs to the observation point x(a), the last cycle segment x(aq) to x(a) is obtained through the trained trend network to obtain the sub-cycle trend feature τ a Fault warnings are performed for the future cycle segment from x(a+1) to x(t+a). Support Vector Data Description (SVDD) is used to analyze trend characteristics for anomaly detection. Historical trend characteristics are mapped to a hypersphere that can contain the vast majority of normal data using SVDD. The current trend characteristic range of the future cycle step condition is determined to achieve single-cycle fault warnings. The trend characteristic statistics of the online sample are recorded as: where c(τ a ) is the trend characteristic τ a The statistic, Φ(·) is the hypersphere mapping function trained based on historical trend features, and β is the Lagrange multiplier used to optimize SVDD. The detailed derivation and solution of SVDD are given in Section 3.2.1. c(τ a ) exceeds the radius R of the hypersphere, which means that the process may fail in the next working cycle, and the maintenance personnel are warned to deploy maintenance in advance. On the contrary, if c(τ a) does not exceed the range of the hypersphere, the sampling true value of the future cycle step will be added to the historical data. After obtaining the sampling true value of the future cycle step, the trend network model and SVDD statistical threshold are updated. BRIEF DESCRIPTION OF THE DRAWINGS

[0019] Figure 1 The overall framework diagram of the cycle trend network of the present invention

[0020] Figure 2 .Performance results of early failure risk warning using different methods DETAILED DESCRIPTION

[0021] The detailed steps of the present invention are as follows:

[0022] Step 1: Period window acquisition, for sampling data N is the observed variable and T is the sampling length. First, the periodic signal is extracted from the frequency domain by FFT. The frequency data obtained by FFT is recorded as In the frequency domain, the inverse of the frequency corresponding to the maximum amplitude is the main periodic signal of the data. The maximum amplitude is calculated by the mean of N frequency domain signals: in is the maximum main frequency value, u x (i)∈U x , i=1,…,N. X in The cycle length in yes The corresponding frequency. X in The set X is re-divided into multiple sub-periods according to the period length = {x1, x2, ..., x P},in P is the number of sub-periods, p = 1,…, P. Subsequently, the input data after period re-division will be used to extract sub-period trend features using Trendnet.

[0023] Step 2: See Figure 1 The overall framework of the cycle trend network of the present invention is designed. The temporal attention convolutional network (TACN) is designed to extract the trend information and cycle information of the current sub-cycle data. TACN consists of two branches: the dilated causal convolutional network for trend information extraction and the short-term dynamic network for dynamic feature extraction. For the input sub-cycle x p , obtain trend information t through the trend extraction branch of TACN p =ReLU(Ψ(x p )),in It is the trend information produced by TACN, and m is the dimension of the hidden layer. Ψ(·) is the dilated causal convolution operation. The convolution result is activated by the ReLU(·) function to prevent the gradient explosion when the trend information is transmitted in the subsequent recursive network. Dynamic information is extracted by the self-attention mechanism of another branch. In the self-attention layer, the input sub-period data is first mapped into three key values ​​Q (query), K (key), and V (value) through the fully connected layer, where Q represents the information requirement of the current position for other positions, the other position information in K is used to match Q, and V is the information that needs to be extracted. The dynamic information is finally output in the form of attention weights. Dynamic information in It is dynamic information. Similar to the extraction of trend information, the self-attention weight is also activated by ReLU(·).

[0024] Step 3: Predict future trends based on trend features, place trend information into the trend network, obtain overall trend features, and describe the overall trend of the working condition. Treat each sub-cycle trend information obtained in claim 2 as a recursive step and learn in the trend recursive network. The trend feature after recursive learning can be expressed as τ p =B·t p +C·τ p-1 +ε, where is the trend feature after recursion and information sharing, B and C represent the recursively calculated transfer matrix, and ε is the residual. Similar to traditional recursive networks, the trend features of a sub-period are only related to the trend information at the current moment and the trend features at the previous moment. The dynamic information obtained by TACN also represents the dynamic information within the current sub-period. The impact of dynamic information is short-lived and rapid, so there is no need to share information through the recursive network. The sub-period trend features and dynamic information are used to predict the operating data of the next sub-period. A fully connected layer is used to implement this function: in At this time, the working condition data of the next sub-cycle is obtained by recursively learning the trend characteristics of historical trend information through the trend network p and the dynamic characteristics of the current sub-period d p Jointly determined. The reconstruction effect of the cycle step is constrained by designing a loss function. The model constraints of the trend network include two loss functions: the error loss function used to improve the model prediction accuracy and the trend slow change constraint loss function to improve the interpretability of trend features. The MSE loss function is selected for the calculation of the error loss function. The model is trained iteratively by comparing the difference between the predicted value and the true value of the working condition data of the next sub-cycle. The prediction error loss function of the trend network is as follows: This loss function strengthens the trend feature τ i-1 and dynamic information i-1The ability to indicate future cycle-step operating conditions. Another trend feature slow-changing constraint function is used to train trend features to clarify their physical meaning. In long-term dependencies, trend features need to represent the overall trend of the process, so trend features are usually flat. Therefore, trend features need to change as little as possible in different sub-cycles. Loss constraints are implemented by comparing the trend feature distances between adjacent sub-cycles: Where distance(·) represents the distance calculation operation of the trend feature. In this chapter, Euclidean distance is selected to calculate the distance between features.

[0025] Step 4: Risk warning is performed based on the predicted future trend. When the online data runs to the observation point x(a), the sub-period trend feature τ is obtained from the last period segment x(aq) to x(a) through the trained trend network. a Fault warnings are performed for the future cycle segment from x(a+1) to x(t+a). Support Vector Data Description (SVDD) is used to analyze trend characteristics for anomaly detection. Historical trend characteristics are mapped to a hypersphere that can contain the vast majority of normal data using SVDD. The current trend characteristic range of the future cycle step condition is determined to achieve single-cycle fault warnings. The trend characteristic statistics of the online sample are recorded as: where c(τ a ) is the trend characteristic τ a The statistic, Φ(·) is the hypersphere mapping function trained based on historical trend features, and β is the Lagrange multiplier used to optimize SVDD. The detailed derivation and solution of SVDD are given in Section 3.2.1. c(τ a ) exceeds the radius R of the hypersphere, which means that the process may fail in the next working cycle, and the maintenance personnel are warned to deploy maintenance in advance. On the contrary, if c(τ a ) does not exceed the range of the hypersphere, the sampling true value of the future cycle step will be added to the historical data. After obtaining the sampling true value of the future cycle step, the trend network model and SVDD statistical threshold are updated.

[0026] Step 5: Verify the method in a real industrial environment and compare it with the LSTM-AE and Timesnet models. The visualization results are shown in Figure 2 A real-world boiler combustion process in a thermal power plant was used to validate the proposed method. The boiler combustion process used in this case study recorded a total of 34 process variables, including three unit groups: working fluid temperature and heat absorption, boiler load water-coal ratio, and boiler combustion (see Table 1):

[0027] Table 1 Coal-fired boiler combustion variables

[0028]

[0029] Experiment 1: In this case, the fault type in the next sub-cycle is a water-cooled wall overtemperature fault. In addition, in order to further verify the effectiveness of the proposed method, two time series prediction models, Timenet and LSTM-AE, are used as comparison groups for fault warning. Similar to the proposed Trendnet, Timenet and LSTM predict the future cycle step length, and use SVDD to judge the process characteristics and calculate statistics to achieve fault warning. In this case, the training set consisting of 6000 sampling points is pre-processed periodically to determine that the amplitude corresponding to this working condition is the largest when the cycle length is 108. Subsequently, another two data sets consisting of 2000 sample points are used as validation sets and test sets to verify the effectiveness of the model. In order to demonstrate the effectiveness of each method, a future cycle step is predicted and a fault warning is issued in this case. For the future prediction results of the model, this experiment uses two evaluation indicators:

[0030] (1) Mean absolute error

[0031] (2) Root mean square error

[0032] The soft sensing experimental results of TE simulation are shown in Table 3.

[0033] Table 3 Evaluation indicators of the results of each method in future trend prediction

[0034]

[0035] Figure 2 The results show the effectiveness of different methods in fault early warning. The envelope represents the 95% confidence interval of the statistical indicator under normal operating conditions. When the statistic exceeds this range, a fault has occurred. The fault occurred at time 52 in a future cycle. As shown in the figure, TimeNet experienced a callback after briefly exceeding the envelope, while LSTM-AE and TrendNet were more sensitive to the fault, with TrendNet providing a more pronounced and robust warning after the fault occurred. Combined with the future operating condition prediction results shown in Table 3, the proposed method demonstrates excellent performance in fault early warning and prediction.

Claims

1. A method for early warning of coal-fired boiler failure risk in thermal power generation based on a cycle-trend network, characterized by: By extracting the periodic signal of the original data from the frequency domain, the periodic window is divided; the sub-periodic data is input into the period-trend network to obtain the sub-periodic trend characteristics and predict the trend of future working conditions; at the same time, the characteristic paradigm of the periodic data is extracted from the historical normal periodic data, the normal uncertainty range of the future trend is calculated, and a fault alarm is issued for the prediction results that exceed the reasonable range.

2. The method for early warning of coal-fired boiler failure risk in thermal power generation based on a cycle-trend network according to claim 1, characterized in that: For sampled data N is the observed variable and T is the sampling length. First, the periodic signal is extracted from the frequency domain by FFT. The frequency data obtained by FFT is recorded as In the frequency domain, the inverse of the frequency corresponding to the maximum amplitude is the main periodic signal of the data. The maximum amplitude is calculated by the mean of N frequency domain signals: in is the maximum main frequency value, u x (i)∈U x , i=1,…,N. X in The cycle length in yes The corresponding frequency. X in The set X is re-divided into multiple sub-periods according to the period length = {x1, x2, ..., x P },in P is the number of sub-periods, p = 1,…, P. Subsequently, the input data after period re-division will be used to extract sub-period trend features using Trendnet.

3. The method for early warning of coal-fired boiler failure risk in thermal power generation based on a cycle-trend network according to claim 2, characterized in that: A temporal attention convolutional network (TACN) is designed to extract the trend information and period information of the current sub-period data. TACN consists of two branches: an expanded causal convolutional network for trend information extraction and a short-term dynamic network for dynamic feature extraction. For the input sub-period x p , obtain trend information t through the trend extraction branch of TACN p =ReLU(Ψ(x p )),in It is the trend information produced by TACN, and m is the dimension of the hidden layer. Ψ(·) is the dilated causal convolution operation. The convolution result is activated by the ReLU(·) function to prevent the gradient explosion when the trend information is transmitted in the subsequent recursive network. Dynamic information is extracted by the self-attention mechanism of another branch. In the self-attention layer, the input sub-period data is first mapped into three key values ​​Q (query), K (key), and V (value) through the fully connected layer, where Q represents the information requirement of the current position for other positions, the other position information in K is used to match Q, and V is the information that needs to be extracted. The dynamic information is finally output in the form of attention weights. Dynamic information in It is dynamic information. Similar to the extraction of trend information, the self-attention weight is also activated by ReLU(·).

4. The method for early warning of coal-fired boiler failure risk in thermal power generation based on a cycle-trend network according to claim 3, characterized in that: Place trend information into the trend network to obtain overall trend characteristics, Describe the overall changing trend of the working conditions. Each sub-period trend information obtained in claim 2 is regarded as a recursive step and learned in the trend recursive network. The trend feature after recursive learning can be expressed as τ p =B·t p +C·τ p-1 +ε, where is the trend feature after recursion and information sharing, B and C represent the recursively calculated transfer matrix and ε is the residual. Similar to traditional recurrent networks, the trend characteristics of a sub-period are only related to the trend information at the current moment and the trend characteristics at the previous moment. Similarly, the dynamic information obtained by TACN represents the dynamic information within the current sub-period. The impact of dynamic information is short-lived and rapid, so there is no need to share information through the recurrent network. The sub-period trend characteristics and dynamic information are used to predict the operating data of the next sub-period. A fully connected layer is used to implement this function: in At this time, the working condition data of the next sub-cycle is obtained by recursively learning the trend characteristics of historical trend information through the trend network p and the dynamic characteristics of the current sub-period d p Joint decision.

5. The method for early warning of coal-fired boiler failure risk in thermal power generation based on a cycle-trend network according to claim 4, characterized in that: A loss function is designed to constrain the reconstruction effect of the cycle step. The model constraints of the trend network include two loss functions: an error loss function used to improve the model's prediction accuracy and a trend slow-changing constraint loss function to improve the interpretability of trend features. The MSE loss function is selected for calculating the error loss function. The model is trained iteratively by comparing the difference between the predicted value and the true value of the working condition data of the next sub-cycle. The prediction error loss function of the trend network is as follows: This loss function strengthens the trend feature τ i-1 and dynamic information i-1 The ability to indicate future cycle-step operating conditions. Another trend feature slow-changing constraint function is used to train trend features to clarify their physical meaning. In long-term dependencies, trend features need to represent the overall trend of the process, so trend features are usually flat. Therefore, trend features need to change as little as possible in different sub-cycles. Loss constraints are implemented by comparing the trend feature distances between adjacent sub-cycles: Where distance(·) represents the distance calculation operation of the trend feature. In this chapter, Euclidean distance is selected to calculate the distance between features.

6. The method for early warning of coal-fired boiler failure risk in thermal power generation based on a cycle-trend network according to claim 5, characterized in that: Establish a fault warning model based on trend features. Through the prediction loss function constraint, the current cycle trend features have the ability to indicate future working conditions. Establish a fault warning model based on trend features. When the online data runs to the observation point x(a), the last cycle segment x(aq) to x(a) is obtained through the trained trend network to obtain the sub-cycle trend feature τ a Fault warnings are performed for the future cycle segment from x(a+1) to x(t+a). Support Vector Data Descriptors (SVDDs) are used to analyze trend characteristics for anomaly detection. Historical trend characteristics are mapped to a hypersphere that encompasses the vast majority of normal data using SVDDs. The range of current trend characteristics pointing to future cycle-step conditions is then determined to achieve fault warnings for single-cycle steps. The trend characteristic statistics of online samples are recorded as: where c(τ a ) is the trend characteristic τ a The statistic, Φ(·) is the hypersphere mapping function trained based on historical trend features, and β is the Lagrange multiplier used to optimize SVDD. The detailed derivation and solution of SVDD are given in Section 3.2.

1. c(τ a ) exceeds the radius R of the hypersphere, which means that the process may fail in the next working cycle, and the maintenance personnel are warned to deploy maintenance in advance. On the contrary, if c(τ a ) does not exceed the range of the hypersphere, the sampling true value of the future cycle step will be added to the historical data. After obtaining the sampling true value of the future cycle step, the trend network model and SVDD statistical threshold are updated.

Images