A rapid tracing method for the surge of induced draft fans in power plants

Abstract

The present invention relates to a method for quickly tracing the surge of a power plant induced draft fan, including: conducting a surge simulation test on a unit model with a divided surge source area to obtain n samples of multiple fault categories containing r relevant parameter data, adding fault labels to all samples to construct a surge fault library; selecting K key parameters beneficial to surge tracing from r relevant parameters, and dividing the samples into a training set and a validation set for each parameter combination; converting the time-domain data of the key parameters in each sample into a two-dimensional feature map of the spectrum; using the two-dimensional feature map as the input and the probability distribution of the fault categories as the output to conduct convolutional neural network training and validation; for the newly occurring surge of the unit, obtaining the two-dimensional feature map of the measured point data corresponding to the positions of K key parameters, inputting it into the network model to obtain the probability distribution of the fault categories, and the fault category with the highest probability is the surge tracing result. It can quickly lock the fault range and reduce the losses caused by surges, which is beneficial to the intelligent operation of the power plant.

Description

Technical Field

[0001] The present invention relates to the technical field of mechanical fault diagnosis, and in particular to a method for quickly tracing the surge of a power plant induced draft fan. Background Art

[0002] The power plant induced draft fan maintains the normal flow of flue gas by sucking in flue gas and stabilizes the negative pressure in the furnace. It is an important auxiliary equipment to ensure the safe and economic operation of the unit. During the operation of the induced draft fan, due to the start and stop of the boiler or the sudden increase in the local resistance of the flue duct, there may be a phenomenon of unstable and repeated periodic changes in flow rate and head in an instant, that is, surge. While the surge causes a huge impact inside the fan, it will also couple the fan and the large-capacity pipeline into a periodic elastic aerodynamic system, and in severe cases, resonance may occur, causing major damage to the entire unit equipment and building.

[0003] Surge is a common fault in the operation of induced draft fans. In surge cases, the surge caused by the increase in local resistance of equipment such as air preheater fouling in the tail flue accounts for the vast majority. At present, the location of the cause of the surge fault generally requires a comprehensive inspection of the induced draft fan and the air and flue gas system during unit shutdown, finding the equipment with serious scaling or ash accumulation, and eliminating the fault through thorough cleaning. This requires a large amount of manpower and material resources, will cause certain economic losses, and the efficiency of fault tracing is not high, and sometimes repeated inspections are needed. Research shows that the surges caused by the increase in resistance in different regions of the unit have different characteristics due to the different elastic aerodynamic systems formed. This characteristic related to the spatial range can be realized by arranging multiple high-frequency pressure measuring points at different positions in the tail flue to capture information and reflected through spectrum analysis.

[0004] Most of the existing surge fault identification methods aim to diagnose the surge fault of the fan by collecting and using the sound signals of on-site equipment and using a convolutional neural network to diagnose the probability of fault occurrence based on the characteristics of the surge sound signals. Such methods do not require sensors to be installed on the fan and can usually achieve a high diagnostic accuracy through model optimization. However, such methods cannot truly trace the source of the surge and cannot quickly eliminate and determine the abnormal resistance events that cause the surge. Summary of the Invention

[0005] Aiming at the deficiencies of the existing technology, the present invention provides a method for quickly tracing the surge of a power plant induced draft fan, aiming to quickly trace the fault by classifying the surges of the induced draft fan, improve the fault diagnosis efficiency, and reduce the surge loss.

[0006] The technical solution adopted by the present invention is as follows:

[0007] A method for quickly tracing the surge of a power plant induced draft fan, comprising:

[0008] Build a surge fault library: Construct a simulation model of unit operation, divide the surge source area, conduct surge simulation tests, obtain n samples of S fault categories containing r relevant parameter data, add fault labels to all samples to build a surge fault library, where the fault categories correspond one-to-one to the surges caused by abnormal resistances in each surge source area;

[0009] Determine key parameters: Select K key parameters that are conducive to surge traceability from r relevant parameters to form C r K kinds of parameter combinations. For each parameter combination, divide the samples into a training set and a validation set;

[0010] Data preprocessing: Preprocess the data in the surge fault library, convert the time-domain data of the K key parameters in each sample into two-dimensional feature maps of the frequency spectrum respectively, and obtain a two-dimensional feature map set of the K key parameters;

[0011] Surge traceability model training and verification: Use the two-dimensional feature map set as the input and the probability distribution of S fault categories as the output to conduct convolutional neural network model training and verification;

[0012] For the newly occurred surge of the unit, obtain the measuring point data. The measuring point positions correspond to the positions of the K key parameters respectively. Preprocess the measuring point data to obtain the two-dimensional feature maps of the measuring point data, input them into the trained convolutional neural network model, output the probability distribution of S fault categories, and the fault category with the highest probability obtained is the surge traceability result.

[0013] The further technical solution is:

[0014] The preprocessing of the data in the surge fault library, and the conversion of the time-domain data of the K key parameters in each sample into two-dimensional feature maps of the frequency spectrum respectively, includes:

[0015] Obtain the frequency spectrum diagrams \(\{(f j ,A j )\}\ i of the amplitude-frequency sequences of the time-domain data of each key parameter in the sample through fast Fourier transform, where f is the frequency, A is the amplitude, j is the serial number of the amplitude-frequency sequence, \(j = 1, 2, 3, \cdots, m\), i is the sample serial number, \(i = 1, 2, 3, \cdots, n\), and k is the serial number of the key parameter, \(k = 1, 2, 3, \cdots, K\);

[0016] Normalize the spectral amplitudes of all key parameters to obtain the spectral amplitude sequence x of each key parameter. x is a one-dimensional sequence containing m·n elements:

[0017] Arrange the spectral amplitude sequence x into a two-dimensional feature map T in matrix form, and the corresponding relationship expression:

[0018] T[i, j] = x[(i - 1)·b + j]

[0019] The above formula means that the value T[i, j] of the i-th row and j-th column of the two-dimensional feature map T corresponds to the (i·j)-th element of the spectral amplitude sequence x, where i = 1, 2, 3, ···, a (a is the number of rows of the two-dimensional feature map), and j = 1, 2, 3, ···, b (b is the number of columns of the two-dimensional feature map).

[0020] The convolutional neural network model includes a feature extraction layer, a transition layer, and a classification layer. The feature extraction layer includes two extraction units connected in sequence. Each extraction unit includes a convolutional layer, an activation layer, and a pooling layer. The transition layer is a flattening layer and two fully connected layers. An activation layer and a Dropout layer are added after the first fully connected layer. The classification layer uses the Softmax function as the activation function. The Softmax function converts the input value of each node into a value within [0, 1] and outputs it in the form of a probability distribution. The expression of the Softmax function is as follows:

[0021]

[0022] In the above formula, z i represents the output value of the i-th node in the transition layer, and S is the number of fault categories.

[0023] The loss function J(w) of the convolutional neural network model is:

[0024]

[0025] In the above formula, w represents the model parameters, S is the number of fault categories, p(x) is the true fault probability distribution of the sample x, that is, the fault label after one-hot encoding, and q(x) is the fault probability distribution output by the model.

[0026] The r relevant parameters include the inlet and outlet pressures of the key equipment of the unit.

[0027] The K key parameters include the pressures at the SCR outlet and the induced draft fan outlet.

[0028] The beneficial effects of the present invention are as follows:

[0029] The present invention proposes a classification basis for induced draft fan surge. By establishing a surge fault library through simulation and constructing a surge traceability model, rapid surge traceability is realized based on the real-time data of high-frequency pressure measurement points, simplifying the complex process of determining the cause of surge by checking equipment one by one.

[0030] The present invention proposes to extract the frequency-domain characteristics of surge using Fourier transform, and use the frequency-domain characteristics as the main classification basis. By constructing a two-dimensional feature map, preliminary extraction of various surge characteristics can be achieved. Utilizing the frequency-domain feature expression ability of Fourier transform and the feature extraction ability of convolutional neural network to trace the source of surge can quickly lock the fault range, reduce the losses caused by surge, and facilitate the intelligent operation of power plants.

[0031] The present invention proposes to trace the source of surge by using a convolutional neural network to fuse the frequency-domain characteristic information of multiple measurement points (parameters), breaking through the limitations of manual analysis characteristics and manual feature extraction, and realizing end-to-end fault diagnosis. It can quickly determine the cause of surge faults at a low cost, ensure the timeliness of surge elimination, give full play to the advantages of data-driven methods, and has certain engineering value.

[0032] Other features and advantages of the present invention will be described in the following specification, and partly will be obvious from the specification, or will be understood by implementing the present invention. Brief Description of the Drawings

[0033] Figure 1 is the flowchart of the method of the embodiment of the present invention.

[0034] Figure 2 is the original signal, signal spectrogram and two-dimensional feature map of a key parameter in the embodiment of the present invention.

[0035] Figure 3 is the visualization image after the surge source tracing model of the embodiment of the present invention extracts the characteristics of the surge samples in the training set.

[0036] Figure 4 is the tracing and classification result of the surge source tracing model of the embodiment of the present invention for newly occurring surge. Detailed Embodiments

[0037] The following describes the detailed embodiments of the present invention with reference to the drawings.

[0038] See Figure 1 , a method for quickly tracing the source of surge of a power plant induced draft fan in the present application, includes:

[0039] Construct a surge fault library: Build a unit simulation model, divide the surge source area, conduct a surge simulation test, obtain n samples of S fault categories containing r relevant parameter data, and add fault labels to all samples to construct a surge fault library, where the fault categories correspond one-to-one to the surges caused by abnormal resistance in each surge source area;

[0040] Determine key parameters: Select K key parameters that are beneficial to surge source tracing from r relevant parameters to form C r KA variety of parameter combinations are used. For each parameter combination, the samples are divided into a training set and a validation set.

[0041] Data preprocessing: Preprocess the data in the surge fault database. Convert the time-domain data of K key parameters in each sample into two-dimensional feature maps of the frequency spectrum respectively, and obtain a two-dimensional feature map set of K key parameters.

[0042] Surge source tracing model training and validation: Using the two-dimensional feature map set as the input and the probability distribution of S fault categories as the output, perform the training and validation of the convolutional neural network model.

[0043] For the newly occurring surge of the unit, obtain the measured point data. The positions of the measured points correspond to the positions of K key parameters respectively. Preprocess the measured point data to obtain the two-dimensional feature maps of each measured point data, and input them into the trained convolutional neural network model. Output the probability distribution of S fault categories. The fault category with the highest probability obtained is the surge source tracing result.

[0044] The rapid surge source tracing method for the induced draft fan of the power station in this application constructs a surge fault database according to the characteristics of the induced draft fan surge, and makes full use of the Fourier transform frequency domain feature expression and the end-to-end learning ability of the CNN, and can realize the recording of key surge information, the fusion of measured point information, the automatic extraction of features, and the rapid surge source tracing.

[0045] In recent years, deep learning methods represented by CNN have been widely studied in the field of fault diagnosis. This method has stronger feature extraction ability, higher efficiency and higher accuracy compared with expert systems and traditional machine learning methods. This application applies deep learning to the problem of surge source tracing. For real-time surges, the same feature map construction method is used to process multi-measured point data, and the trained model is used to complete the surge source tracing. It can quickly determine the cause of the surge fault at a low cost, ensure the timeliness of surge elimination, give full play to the advantages of the data-driven method, and has certain engineering value.

[0046] The following further illustrates the rapid surge source tracing method for the induced draft fan of the power station in this application with specific embodiments, including the following steps:

[0047] (1) Construct a surge fault database:

[0048] Use the advanced process simulation software Apros to build a unit model, including the boiler tail flue gas system model, verify the steady-state accuracy and dynamic accuracy of the model, and ensure the full-condition simulation accuracy of the model.

[0049] Define the surge categories. Surges caused by abnormal resistances in different regions of the boiler's tail flue have different characteristics. The tail flue is divided into several surge source areas around key equipment, such as the SCR area, the air preheater area, etc. Surges caused by abnormal resistance in each source area are recorded as different categories, with a total of S categories.

[0050] Based on the simulation model, set different magnitudes of abnormal resistances in each surge source area to obtain n surge samples. For each sample, obtain the surge label and record the inlet and outlet pressures of each key equipment, a total of r parameters P1 to Pr, to construct a surge fault library. The maximum data output frequency allowed by Apros is 1 MHz. Combine with the selection of high-frequency pressure sensors to read Apros data at the same frequency.

[0051] (2) Determine the key parameters:

[0052] Select K key parameters that are conducive to surge traceability from the r parameters. The number of combinations of the K key parameters is types. Based on each parameter combination, divide the samples in the surge fault library into a training set and a validation set.

[0053] (3) Surge data preprocessing:

[0054] Export the surge data from the surge fault library for processing. Through fast Fourier transform (FFT), spectral amplitude normalization, and rearrangement, convert the time-domain data of the K key parameters P1 to PK in each fault sample into two-dimensional feature maps T1 to T K , specifically including:

[0055] For the n samples, use FFT to obtain the spectrogram of the time-domain data sequence of each parameter Pk in the sample {(f j , A j )} i , k, where: f is the frequency, A is the amplitude, j is the serial number of the amplitude-frequency sequence, j = 1, 2, 3, ···, m, i is the sample serial number, i = 1, 2, 3, ···, n; k is the serial number of the key parameter, k = 1, 2, 3, ···, K.

[0056] Record the data set of the spectral amplitude of parameter Pk as X k = [x k,1 , … x k,n T , and normalize the spectral amplitudes of the K key parameters. The normalization method of the spectral amplitude is as follows:

[0057]

[0058] x k,i = {A j}i,k , x k,max and x k,min are respectively the row vectors obtained by taking the maximum and minimum values in each column; k

[0059] Arrange the normalized spectral amplitude sequence x into a two-dimensional feature map T in matrix form. The value T[i, j] at the i-th row and j-th column of the feature map corresponds to the i·j-th element of x. The specific correspondence can be expressed by the following formula:

[0060] T[i, j] = x[(i - 1)·b + j]

[0061] In the above formula, i = 1, 2, 3, ···, a, where a is the number of rows of the feature map; j = 1, 2, 3, ···, b, where b is the number of columns of the feature map

[0062] (4) Surge source tracing model training and verification:

[0063] Construct a convolutional neural network model, using the two-dimensional feature maps T1 to T of multiple measurement points K as multi-input channel data, and using the probability distribution of S types of faults as the output. The category with the highest probability is the surge source tracing result. Use the training set to train the model and use the validation set to verify the model accuracy. The specific steps are as follows:

[0064] Determine the neural network composition: The constructed convolutional neural network includes a feature extraction layer, a transition layer, and a classification layer. The feature extraction layer is 2 groups of consecutive convolutional layers, activation layers, and pooling layers. The transition layer is a flattening layer and 2 fully connected layers. An activation layer and a Dropout layer are added after the first fully connected layer to prevent overfitting. The classification layer uses the Softmax as the activation function. The Softmax function converts the input values of each node into values within [0, 1] and outputs them in the form of a probability distribution. The Softmax function expression is as follows:

[0065]

[0066] In the above formula, z i represents the output value of the i-th node in the transition layer, and S is the number of fault categories.

[0067] Clarify the model loss. Denote the fault labels of S types of surges as 1 to S, and use one-hot encoding to convert the fault labels into binary form, as shown below:

[0068]

[0069] Use cross-entropy as the loss function J(w):

[0070]

[0071] ​In the above formula, w represents the model parameters, S is the number of fault categories, p(x) is the true fault probability distribution of the sample x, that is, the fault label after one-hot encoding, and q(x) is the fault probability distribution output by the model.

[0072] Training and validation: Set model hyperparameters such as the learning rate, number of iterations, number of channels in each layer, size of the convolutional kernel, and moving step size. Those skilled in the art can understand that the number of channels in the input layer of the model is equal to the number of key parameters K, and there are S nodes in the output layer. Use the feature maps transformed from the training set to train the model, calculate the model loss through forward propagation, and update the network parameters using the Adam algorithm. After training, use the feature maps transformed from the test set to verify the model accuracy.

[0073] After data preprocessing and model training, verify the surge source tracing accuracy. Set the accuracy threshold for model verification as Th, and select the smallest K while ensuring the accuracy.

[0074] (5) Installation of measuring points and surge source tracing:

[0075] Corresponding to the K key parameters, install high-frequency pressure measuring points at the corresponding positions in the boiler tail flue for experiments. According to the experimental results, the K high-frequency measuring points at least include the SCR outlet pressure and the induced draft fan outlet pressure. For newly occurring surges, obtain the measuring point data, apply the same data preprocessing method, and input the obtained multi-measuring-point two-dimensional feature map into the trained model to achieve rapid surge source tracing.

[0076] The following uses a specific example to illustrate the effectiveness of the prediction method of this embodiment.

[0077] Taking a certain coal-fired power generation unit as the research object, a simulation model is built, and a surge simulation experiment is carried out on the simulation software platform to obtain samples of various surges under N working conditions, and the effectiveness of this method is verified based on the simulation data.

[0078] The specific steps are as follows:

[0079] 1) Construct a surge fault library:

[0080] A simulation model of this unit is built. There are 6 key equipment in the tail flue of this unit. The tail flue is divided into 6 regions, and resistance anomalies are set respectively to form surge samples. The surge labels are also divided into 6 categories according to the regions where the surge causes are located. Organize the surge samples to construct a surge fault library. The sampling frequency of the samples is 100Hz, and the sampling time is 120s.

[0081] Divide the samples into a training set, a validation set, and a test set, with a ratio of (N - 2):1:1, and regard the test set as real-time surge data.

[0082] 2) Determine the key parameters:

[0083] After several rounds of experiments, it is determined that two key parameters, namely the SCR outlet and the induced draft fan outlet, are used as data sources, and the accuracy rate is above 99%. Therefore, K = 2 is taken.

[0084] 3) Surge data preprocessing:

[0085] Derive the surge data from the surge fault library for processing. Through FFT, spectral amplitude normalization, and rearrangement, the time-domain data of two key parameters P1 and P2 in each sample are converted into a spectral two-dimensional feature map. Referring to the effective frequency range in the spectrogram, both the number of rows a and the number of columns b of the two-dimensional feature map are set to 48, and the corresponding frequency range is 0 - 19.19 Hz. Figure 2 As shown in (a), (b), and (c), they are respectively the original signal, spectrogram, and the generated two-dimensional feature map of the key parameter P2 in a certain sample.

[0086] 4) Surge traceability model training and verification:

[0087] Construct a convolutional neural network model. The detailed structure of the constructed model is shown in Table 1. Using the two-dimensional feature maps T1 and T2 of two key parameters as multi-input channel data, the model is trained with surge samples. Set the learning rate to 0.001, the number of training rounds to 64, and adopt batch training with a batch size of 16.

[0088] Table 1 CNN model structure

[0089]

[0090] Figure 3 It is the t-SNE visualization result of the output value of the transition layer after the model extracts features from the input data. It can be seen that various types of faults can be well distinguished in the 3D space.

[0091] 5) Measuring point installation and surge traceability:

[0092] Assume that high-frequency pressure measuring points have been installed at the SCR outlet and the induced draft fan outlet of the unit. Regard the test set as a newly occurring surge. Use the same data preprocessing method for the new surge to obtain a two-dimensional feature map, and input the two-dimensional feature maps of the two measuring points into the trained model to achieve fault diagnosis. Figure 4 It is the confusion matrix of the traceability results of 18 surge samples in the test set. It can be seen that the coincidence accuracy rate between the predicted label (fault distribution) and the true label (fault distribution) reaches 100%, which proves the effectiveness of the method in this embodiment.

[0093] Those of ordinary skill in the art can understand that the above are only preferred embodiments of the present invention and are not intended to limit the present invention. Although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art can still modify the technical solutions described in the foregoing embodiments or perform equivalent replacements for some of the technical features. Any modifications, equivalent replacements, improvements, etc. made within the spirit and principle of the present invention shall be included within the protection scope of the present invention.

Claims

1. A method for quickly tracing the surge of a power station induced draft fan, characterized in that, Including: Construct a surge fault library: Build a unit operation simulation model, divide the surge source area, conduct a surge simulation test, obtain n samples of S fault categories containing r relevant parameter data, add fault labels to all samples to construct a surge fault library, and the fault categories correspond one-to-one to the surges caused by abnormal resistance in each surge source area; Determine key parameters: Select K key parameters that are beneficial to surge traceability from r relevant parameters to form parameter combinations. For each parameter combination, divide the samples into a training set and a validation set; Data preprocessing: Preprocess the data in the surge fault library, convert the time-domain data of K key parameters in each sample into a two-dimensional feature map of the spectrum, and obtain a two-dimensional feature map set of K key parameters; Surge source tracing model training and verification: Use the two-dimensional feature map set as the input and the probability distribution of S fault categories as the output to conduct convolutional neural network model training and verification; For the newly occurred surge of the unit, obtain the measured point data. The measurement point positions correspond to the positions where K key parameters are obtained respectively. Preprocess the measured point data to obtain the two-dimensional feature map of each measured point data, input it into the trained convolutional neural network model, output the probability distribution of S fault categories, and the fault category with the highest probability obtained is the surge source tracing result.

2. The method for quickly tracing the surge of a power station induced draft fan according to claim 1, characterized in that, The preprocessing of the data in the surge fault library, which converts the time-domain data of K key parameters in each sample into a two-dimensional feature map of the spectrum, includes: Obtain the spectrogram of the amplitude-frequency sequence of the time-domain data of each key parameter in the sample through the fast Fourier transform \(\{(f j ,A j )\} i,k , where \(f\) is the frequency, \(A\) is the amplitude, \(j\) is the serial number of the amplitude-frequency sequence, \(j = 1, 2, 3, \cdots, m\), \(i\) is the sample serial number, \(i = 1, 2, 3, \cdots, n\), and \(k\) is the serial number of the key parameter, \(k = 1, 2, 3, \cdots, K\); Normalize the spectrum amplitudes of all key parameters to obtain the spectrum amplitude sequence x of each key parameter. x is a one-dimensional sequence containing m·n elements: Arrange the spectrum amplitude sequence x into a two-dimensional feature map T in matrix form, and the corresponding relationship expression: T[i,j] = x[(i - 1)·b + j] The above formula means that the value T[i,j] of the i-th row and j-th column of the two-dimensional feature map T corresponds to the (i·j)-th element of the spectrum amplitude sequence x, where i = 1, 2, 3, ···, a, a is the number of rows of the two-dimensional feature map, and j = 1, 2, 3, ···, b, b is the number of columns of the two-dimensional feature map.

3. The method for quickly tracing the surge of a power station induced draft fan according to claim 1, characterized in that, The convolutional neural network model includes a feature extraction layer, a transition layer, and a classification layer. The feature extraction layer includes two extraction units connected in front and back. Each extraction unit includes a convolutional layer, an activation layer, and a pooling layer. The transition layer is a flattening layer and two fully connected layers. An activation layer and a Dropout layer are added after the first fully connected layer. The classification layer uses the Softmax as the activation function. The Softmax function converts the input value of each node into a value within [0,1] and outputs it in the form of a probability distribution. The Softmax function expression is as follows: In the above formula, z i represents the output value of the i-th node in the transition layer, and S is the number of fault categories.

4. The method for quickly tracing the surge of a power station induced draft fan according to claim 1, characterized in that, The loss function J(w) of the convolutional neural network model is: In the above formula, w represents the model parameters, S is the number of fault categories, p(x) is the true fault probability distribution of the sample x, that is, the fault label after one-hot encoding, and q(x) is the fault probability distribution output by the model.

5. The method for quickly tracing the surge of a power station induced draft fan according to claim 1, characterized in that, The r relevant parameters include the inlet and outlet pressures of key equipment of the unit.

6. The method for quickly tracing the surge of a power station induced draft fan according to claim 1, characterized in that, The K key parameters include the pressures at the SCR outlet and the induced draft fan outlet.

Images