Waste heat boiler economizer fault diagnosis method and system based on BP neural network

Abstract

The invention provides a waste heat boiler economizer fault diagnosis method and system based on a BP neural network, and relates to the technical field of thermal power plant equipment fault diagnosis. Aiming at the problem that in the prior art, accurate identification cannot be carried out at the early stage of fault occurrence of the waste heat boiler economizer, characteristic parameters strongly related to leakage of the economizer are screened, an optimized BP neural network diagnosis model is constructed, accurate identification of normal state, low-pressure, medium-pressure and high-pressure leakage faults of the economizer is achieved, and the fault identification accuracy is improved. Especially, tiny features at the initial stage of the fault can be effectively recognized, the problems that in the prior art, recognition at the initial stage of the fault is not timely, and accuracy is low are solved, and reliable guarantee is provided for safe operation of a heat-engine plant waste heat boiler.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of waste heat boiler equipment fault diagnosis technology in thermal power plants, and particularly relates to a waste heat boiler economizer fault diagnosis method and system based on a BP neural network. BACKGROUND

[0002] In the operation of a waste heat boiler in a thermal power plant, the economizer as a key heat exchange equipment, its leakage fault can cause the boiler to shut down, economic losses, and even endanger personnel safety. In the prior art, the economizer fault diagnosis mainly relies on traditional sensor monitoring or a general neural network model, but there are significant defects: Early fault identification lag: the traditional technology can only identify when the fault characteristics are obvious, at which time irreversible damage such as pipe corrosion aggravation has already occurred; and the characteristic parameter fluctuation is small at the early stage of the fault, and the traditional method is prone to misjudgment as normal fluctuation.

[0003] Insufficient pertinence: the existing neural network model is generally designed, and is not optimized for the fault characteristics of the economizer - such as the characteristic parameter selection relying on experience (without verifying the correlation with leakage), and the network structure not being adapted to the fitting demand of small fluctuation, resulting in low early diagnosis accuracy.

[0004] Therefore, there is an urgent need for a diagnosis technology that can accurately identify the early stage of the economizer fault, so as to realize early warning and early treatment. SUMMARY

[0005] To solve the above problems, the present application aims to provide a waste heat boiler economizer fault diagnosis method and system based on a BP neural network, which realizes early diagnosis of the economizer fault by screening sensitive characteristic parameters, optimizing the BP neural network structure and training method.

[0006] To achieve the above purpose, the technical scheme of the present application is as follows: A waste heat boiler economizer fault diagnosis method based on a BP neural network, comprising the following steps: S1, characteristic parameter acquisition and screening: selecting characteristic parameters related to economizer leakage, the characteristic parameters including turbine load, low-pressure main steam flow, low-pressure economizer recirculation pump inlet pressure, medium-pressure economizer inlet pressure, medium-pressure economizer outlet pressure, high-pressure feedwater flow, and high-pressure bypass desuperheating water pressure; S2, sample data processing: under steady-state conditions, sample data of the characteristic parameters under normal state and low-pressure, medium-pressure, and high-pressure economizer leakage state are collected, and the sample data is normalized to the [-1, 1] interval; S3, BP neural network model construction: a three-layer BP neural network is constructed, the number of input layer neurons is 7, the number of output layer neurons is 2, and the number of hidden layer neurons is 10; S4, model training: the neural network is trained by using the trainlm training function and the tansig activation function until the network error reaches the preset threshold; S5, fault diagnosis: input the real-time collected characteristic parameters into the trained neural network, and determine the economizer fault type according to the output signal.

[0007] Further, the selection of the characteristic parameters in step S1 is based on sensitivity analysis of 100 initial variable fault data, and parameters with high correlation coefficient related to the initial stage of the economizer leakage degree fault are selected.

[0008] Further, the sample data in step S2 include fault initial stage samples, and the fault initial stage samples are expanded by a data enhancement algorithm, and the proportion of the expanded samples is not less than 70%.

[0009] Further, the number of hidden layer neurons in step S3 is determined by comparison experiment: test the network error and training efficiency when the number of hidden layer neurons is 3-13, and select 10 neurons with network error 0.724 and training times 39 as the optimal structure.

[0010] Further, the training process in step S4 adopts improved normalization processing: the original data are first centered by mean, and then normalized to the interval [-1, 1] to enhance the recognition ability of small fluctuations in the initial stage of fault.

[0011] In order to achieve the above purpose, the application also provides a waste heat boiler economizer fault diagnosis system based on BP neural network, comprising the following modules: Data acquisition module: used for real-time acquisition of 7 characteristic parameters including turbine load, low-pressure main steam flow, low-pressure economizer recirculation pump inlet pressure, medium-pressure economizer inlet pressure, medium-pressure economizer outlet pressure, high-pressure feedwater flow and high-pressure bypass desuperheating water pressure; Data preprocessing module: used for normalizing the collected characteristic parameters to the interval [-1, 1]; BP neural network diagnosis module: the BP neural network constructed by the above method receives the preprocessed characteristic parameters and outputs the fault diagnosis result; Result output module: converts the neural network output signal into fault type and displays.

[0012] Beneficial effects: the application selects the characteristic parameters strongly related to the economizer leakage, constructs an optimized BP neural network diagnosis model, realizes accurate identification of the normal state of the economizer and low-pressure, medium-pressure and high-pressure leakage faults, especially can effectively identify the small characteristics in the initial stage of fault, solves the problems of untimely identification and low accuracy in the initial stage of fault in the traditional technology, and provides reliable guarantee for safe operation of the waste heat boiler of the thermal power plant. BRIEF DESCRIPTION OF DRAWINGS

[0013] The accompanying drawings, which form a part of the specification, are included to provide a further understanding of the application and are incorporated herein in conjunction with the description of the application. The embodiments of the application, together with its description, serve to explain the application. In the drawings: Figure 1 Flow chart of the BP neural network-based waste heat boiler economizer fault diagnosis method according to the embodiments of the application; Figure 2 Structure diagram of the three-layer BP neural network in the BP neural network-based waste heat boiler economizer fault diagnosis method according to the embodiments of the application; Figure 3 Changes in the steam turbine load when the low-pressure economizer fails in the BP neural network-based waste heat boiler economizer fault diagnosis method according to the embodiments of the application; Figure 4 Changes in the low-pressure main steam flow when the low-pressure economizer fails in the BP neural network-based waste heat boiler economizer fault diagnosis method according to the embodiments of the application; Figure 5 Changes in the low-pressure economizer recirculation pump inlet pressure when the low-pressure economizer fails in the BP neural network-based waste heat boiler economizer fault diagnosis method according to the embodiments of the application; Figure 6 Neural network training in the BP neural network-based waste heat boiler economizer fault diagnosis method according to the embodiments of the application; Figure 7 Network training result one in the BP neural network-based waste heat boiler economizer fault diagnosis method according to the embodiments of the application; Figure 8 Network training result two in the BP neural network-based waste heat boiler economizer fault diagnosis method according to the embodiments of the application; Figure 9 Structure schematic diagram of the BP neural network-based waste heat boiler economizer fault diagnosis system according to the embodiments of the application. DETAILED DESCRIPTION

[0014] It should be noted that the embodiments in the application and the features in the embodiments can be combined with each other without conflict.

[0015] The application will be described in detail below with reference to the accompanying drawings and in conjunction with the embodiments.

[0016] Embodiment 1 Reference Figures 1-8 A BP neural network-based waste heat boiler economizer fault diagnosis method, comprising the following steps: S1. Feature parameter acquisition and screening: Select feature parameters related to economizer leakage. The feature parameters include turbine load, low-pressure main steam flow rate, low-pressure economizer recirculation pump inlet pressure, medium-pressure economizer inlet pressure, medium-pressure economizer outlet pressure, high-pressure feedwater flow rate, and high-pressure bypass desuperheating water pressure. S2. Sample data processing: Under steady-state conditions (800MW), sample data of the characteristic parameters are collected under normal conditions and under low-pressure, medium-pressure, and high-pressure economizer leakage conditions, and the sample data are normalized to the [-1,1] interval. S3. Construction of BP Neural Network Model: Construct a three-layer BP neural network with 7 neurons in the input layer (corresponding to 7 feature parameters), 2 neurons in the output layer (the output signals (0,0), (0,1), (1,0), and (1,1) correspond to the normal state and low-pressure, medium-pressure, and high-pressure leakage faults, respectively), and 10 neurons in the hidden layer. S4. Model Training: Train the neural network using the trainlm training function and tansig activation function until the network error reaches a preset threshold. S5. Fault Diagnosis: Input the real-time collected feature parameters into the trained neural network, and determine the economizer fault type based on the output signal.

[0017] This embodiment constructs an optimized BP neural network diagnostic model by screening feature parameters that are strongly correlated with economizer leakage, thereby achieving accurate identification of economizer normal state and low-pressure, medium-pressure, and high-pressure leakage faults. In particular, it can effectively identify minute features in the early stage of faults, solving the problems of untimely and low accuracy in the early stage of fault identification of traditional technologies, and providing reliable protection for the safe operation of waste heat boilers in thermal power plants.

[0018] In this specific implementation, a three-layer BP neural network containing a single hidden layer is selected, and its structure is as follows: Figure 2 As shown: Indicates the input layer's first... The input of each neuron node, .

[0019] Indicates the hidden layer number 1 The nth neuron node is connected to the input layer. The weights between each neuron node.

[0020] Indicates the hidden layer number 1 The threshold of each neuron node, .

[0021] This represents the transfer function of the hidden layer.

[0022] denotes the weight between the output layer first neuron node and the hidden layer first neuron node, .

[0023] denotes the threshold of the output layer kth .

[0024] denotes the transfer function of the output layer. denotes the output of the output layer first neuron node.

[0025] Each fault is shown in Table 1, in order to simplify the network structure, the output signal of U1 is (0, 0), the output signal of U2 is (0, 1), the output signal of U3 is (1, 0), and the output signal of U4 is (1, 1). Then the output layer has two neurons.

[0026] Table 1 Neural network output

[0027] In order to find the characteristic parameters most closely related to the low-pressure, medium-pressure and high-pressure economizer leakage, 100 kinds of variable change data are collected every 1s in the normal state, low-pressure economizer fault state, medium-pressure economizer fault state and high-pressure economizer fault state, as shown in Table 1, Table 2 and Table 3. Figure 3 、 4 、

[0028] Finally, the final characteristic parameters are selected after comprehensive consideration of each parameter, as shown in Table 2.

[0029] Table 2 Selection of economizer leakage characteristic parameters

[0030] For the 7 characteristic parameters in Table 2, 855 groups of sample data (sampling time 1s) of different fault data are collected at 800MW steady state. A sample point is extracted every 1s from the extracted data, and a total of 855 groups of data are extracted for neural network training. In order to reduce the adjustment range of network weights in the training process and eliminate the influence of the size of the original data value on the network learning process, the data need to be normalized to [-1, 1]. In the input layer structure, there are a total of 7 neurons.

[0031] In a specific implementation, the selection of the characteristic parameters in step S1 is based on: sensitivity analysis is performed on the fault data of 100 initial variables, and parameters with high correlation coefficient related to the initial stage of economizer leakage degree fault are selected.

[0032] In a specific implementation, the parameters are screened through 'feature importance analysis': sensitivity calculation (using Pearson correlation coefficient) is performed on the initial failure data of 100 variables (leakage of 0.5L / min), and parameters with a correlation greater than 0.8 with the degree of failure are selected (redundant parameters with a correlation less than 0.5, such as furnace temperature, are excluded). For example: the correlation coefficient between the fluctuation of the steam turbine load in the initial failure period and the leakage is 0.82, and the correlation coefficient between the low-pressure main steam flow is 0.85, so the feature parameters are included; and the correlation coefficient between the flue gas temperature is 0.3, which is excluded. Sensitivity verification of parameters in the initial failure period: "7 parameters are tested for initial failure response: when the leakage is 0.5L / min, the low-pressure economizer recirculation pump inlet pressure drops by 0.02MPa (which can be accurately captured), and the medium-pressure economizer outlet pressure drops by 0.015MPa (although the fluctuation is small, it can be identified after being amplified by the neural network), verifying its sensitivity to the initial failure." In a specific implementation, the sample data in step S2 includes initial failure samples, and the initial failure samples are expanded through a data augmentation algorithm, and the proportion of the expanded samples is not less than 70%.

[0033] This embodiment expands the samples through data augmentation to compensate for the potential defect of "possible insufficient initial samples", ensuring that the network has enough learning samples for the tiny features in the initial failure period and avoiding insufficient recognition ability of the model for the initial failure.

[0034] In a specific implementation, the number of hidden layer neurons in step S3 is determined through comparative experiments: the network error and training efficiency when the number of hidden layer neurons is 3-13 are tested, and 10 neurons with a network error of 0.724 and a training frequency of 39 times are selected as the optimal structure.

[0035] This embodiment clearly defines the determination method of the number of hidden layer neurons, and the advantage lies in proving the rationality of the network structure design rather than blind selection: through the above test of errors and training times of 3-13 neurons, 10 neurons are finally selected (error 0.724, training times 39, and comprehensive optimal). This process is converted into the technical feature of "determination through comparative experiments", avoiding "setting the number of neurons based on experience", and strengthening the rationality of the network structure design.

[0036] In a specific implementation, the training process in step S4 adopts improved normalization processing: the original data is first centered by mean (subtracting the mean value of the parameters in the normal state), and then normalized to the interval [-1, 1] to enhance the recognition ability of the initial failure period.

[0037] The embodiment can adopt an improved method of "mean centering first and then normalizing", highlighting the ability to capture the initial fault of the small features, and solving the technical details of the "small fluctuation recognition" which is not clear.

[0038] In order to test the effectiveness of the waste heat boiler economizer fault diagnosis system of the BP neural network, the BP neural network for boiler economizer leakage fault is designed and trained, and the training error is shown in Table 3.

[0039] Table 3 Network training error

[0040] From the table, the network error of the hidden layer neuron number of 3, 9, 10 and 12 is smaller. The network error and training times of the hidden layer of 12 are compared with the hidden layer of 10, which can be excluded. The training times of the hidden layer of 3 are larger, which can be excluded. Although the training times of the hidden layer of 9 are less than the hidden layer of 10, the network error is nearly 0.2. Therefore, the BP neural network with 10 hidden layer neurons is selected.

[0041] Determination of activation function In the BP neural network, the commonly used transfer function is S-shaped logarithmic function logsig, S-shaped tangent function tansig and pure linear function purelin. The logsig function maps the value of (-∞, +∞) to (0, 1); the tansig function maps the value of (-∞, +∞) to (-1, +1); and the purelin function maps the value of (-∞, +∞) to (-∞, +∞).

[0042] In this neural network, because the normalization (normalized to [-1, +1]) is adopted, the hidden layer adopts the tansig function, and the output layer also adopts the tansig function.

[0043] Determination of other functions In order to achieve the function approximation effect, and meet the above requirements of the BP neural network, improve the network training precision and performance, the network training function uses trainlm, the learning function takes the default value learngdm, the performance function takes the default value mse, and the threshold range function minmax of the input vector elements.

[0044] After 23 times of training, the network error reaches the set minimum value, and the training results are shown in Figure 6 , 7, 8.

[0045] Finally, the real-time data is used to test the above diagnosis system, and the test results are shown in Table 4.

[0046] Table 4 Real-time test results

[0047] According to the results of real-time data test, the first group of data (0, 0) and the second group of data (0, 1) respectively represent the normal state and the low-pressure economizer leakage fault, which is correct. For the third group of data (0.9951, 0.0087), it can be accurate to (1, 0) due to the small error of the network, which represents the medium-pressure economizer leakage fault. For the fourth group of output data (0.9933, 0.9764), it can be accurate to (1, 1), which represents the high-pressure economizer fault. The output results are compared with Table 4, which shows that the designed BP neural network is reasonable and can be put into practical engineering application.

[0048] Embodiment 2 In order to achieve the above-mentioned purpose, referring to Figure 9 : The embodiment also provides a waste heat boiler economizer fault diagnosis system based on a BP neural network, comprising the following modules: A data acquisition module is used to acquire 7 feature parameters in real time, including the steam turbine load, the low-pressure main steam flow, the low-pressure economizer recirculation pump inlet pressure, the medium-pressure economizer inlet pressure, the medium-pressure economizer outlet pressure, the high-pressure feedwater flow and the high-pressure bypass desuperheating water pressure; A data preprocessing module is used to normalize the acquired feature parameters to the interval [-1, 1]; A BP neural network diagnosis module receives the preprocessed feature parameters and outputs a fault diagnosis result by using the BP neural network constructed by the above-mentioned method; A result output module converts the neural network output signal into a fault type (normal state U1, low-pressure economizer leakage U2, medium-pressure economizer leakage U3, and high-pressure economizer leakage U4) and displays it.

[0049] The waste heat boiler economizer fault diagnosis system based on the BP neural network of the embodiment has the same advantages as the waste heat boiler economizer fault diagnosis method based on the BP neural network of the above-mentioned embodiment relative to the prior art, and will not be described here.

[0050] The above only describes the preferred embodiments of the present application and should not be used to limit the present application. Any modification, equivalent replacement, improvement, etc. made within the spirit and principles of the present application should be included in the protection scope of the present application.

Claims

1. A waste heat boiler economizer fault diagnosis method based on a BP neural network, characterized in that, The method comprises the following steps: S1, characteristic parameter acquisition and screening: selecting characteristic parameters related to the coal economizer leakage, the characteristic parameters comprising the steam turbine load, the low-pressure main steam flow, the low-pressure coal economizer recirculation pump inlet pressure, the medium-pressure coal economizer inlet pressure, the medium-pressure coal economizer outlet pressure, the high-pressure feedwater flow, and the high-pressure bypass desuperheating water pressure; S2, sample data processing: under a steady state condition, sample data of the characteristic parameters under normal state and low-pressure, medium-pressure, and high-pressure coal economizer leakage state are collected, and the sample data is normalized to the [-1, 1] interval; S3, BP neural network model construction: a three-layer BP neural network is constructed, the number of input layer neurons is 7, the number of output layer neurons is 2, and the number of hidden layer neurons is 10; S4, model training: the neural network is trained by using the trainlm training function and the tansig activation function until the network error reaches a preset threshold; S5, fault diagnosis: the real-time collected characteristic parameters are input into the trained neural network, and the output signal is used to determine the coal economizer fault type.

2. The BP neural network-based waste heat boiler economizer fault diagnosis method according to claim 1, characterized in that, The screening of the characteristic parameters in the step S1 is based on the sensitivity analysis of 100 initial variable fault data, and the parameters with high correlation coefficient in the initial stage of the coal economizer leakage degree fault are selected. 3.The waste heat boiler economizer fault diagnosis method based on BP neural network according to claim 1, characterized in that, The sample data in the step S2 comprises initial stage sample data, and the initial stage sample data is expanded by using a data enhancement algorithm, and the proportion of the expanded sample data is not less than 70%.

4. The BP neural network-based waste heat boiler economizer fault diagnosis method according to claim 1, characterized in that, The number of hidden layer neurons in the step S3 is determined by comparison experiments: the network error and training efficiency when the number of hidden layer neurons is 3-13 are tested, 10 neurons with the network error of 0.724 and the training number of 39 times are selected as the optimal structure.

5. The BP neural network-based waste heat boiler economizer fault diagnosis method according to claim 1, characterized in that, In the training process in the step S4, improved normalization processing is used: the original data is first centered by mean, and then normalized to the [-1, 1] interval, so as to enhance the recognition ability of the initial stage of the fault.

6. A waste heat boiler economizer fault diagnosis system based on a BP neural network, characterized in that, The method comprises the following modules: A data acquisition module: used for real-time acquisition of seven characteristic parameters, including the steam turbine load, the low-pressure main steam flow, the low-pressure coal economizer recirculation pump inlet pressure, the medium-pressure coal economizer inlet pressure, the medium-pressure coal economizer outlet pressure, the high-pressure feedwater flow, and the high-pressure bypass desuperheating water pressure; A data preprocessing module: used for normalizing the collected characteristic parameters to the [-1, 1] interval; A BP neural network diagnosis module: the BP neural network constructed by using the method in any one of claims 1-5, receiving the preprocessed characteristic parameters, and outputting the fault diagnosis result; A result output module: converting the neural network output signal into the fault type and displaying.

Images