Multi-parameter fusion nuclear power emergency generator fault diagnosis method

Abstract

The invention aims to provide a multi-parameter fusion nuclear power emergency generator fault diagnosis method. Various electrical faults of the emergency diesel generator are interactively simulated through a test device and a protection device, change rules of fault transient characteristics in time domain and frequency domain dimensions are described based on time-frequency transformation and image fusion, internal relations between the fault transient characteristics and fault types are mined, initial input data are preprocessed to form a fault scene data set, and the fault scene data set is analyzed. The CNN-LSTM network is used for fault classification, and accurate diagnosis of electrical faults is realized. The method is suitable for the field of nuclear power station emergency diesel generator fault diagnosis and is high in universality.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of fault diagnosis, in particular to a method for diagnosing faults of a nuclear power emergency diesel generator by using electrical characteristics of the nuclear power emergency diesel generator and fusing multiple parameters. BACKGROUND

[0002] The nuclear power emergency diesel generator is a backup power supply of a safety system device, and when the normal power supply and the backup power supply of a nuclear island fail, the nuclear power emergency diesel generator is quickly started to supply power to nuclear safety devices, to ensure that the reactor heat removal system works normally, and thus the reactor core is normally shut down, to avoid the occurrence of a nuclear safety accident. The nuclear power emergency diesel generator is an important safety power supply device, and is the last line of defense of a nuclear power plant. Due to the complex structure and electrical characteristics, the harsh working environment and the high requirement for rapid start performance, the nuclear power emergency diesel generator has a high fault frequency.

[0003] There are many electrical state parameters of the nuclear power emergency diesel generator, and the traditional fault diagnosis method has limited means, and the fault mechanism and fault performance of some faults are not clear. The traditional fault diagnosis method realizes fault judgment and distinction according to a threshold value, but needs to deeply mine and analyze the fault background and principle, and the diagnosis effect depends on the accurate analysis of the principle and the accurate selection of the threshold value.

[0004] The traditional fault diagnosis method mainly depends on single index analysis, and lacks data fusion and analysis and comprehensive processing of multi-dimensional data, resulting in insufficient diagnosis accuracy and frequent misjudgment and omission. Fault information processing is difficult, and the lack of fault cases is also a prominent problem, which leads to low efficiency of the traditional fault diagnosis method in actual application, and the method is difficult to meet the requirement of rapid fault positioning. The limitation of intelligent diagnosis technology is that due to the lack of sufficient sample data, the diagnosis effect is often unsatisfactory.

[0005] In actual engineering, a small amount of fault data has labels, and a large amount of fault data is untagged. The fault data quality is low, and presents obvious imbalance, which is difficult to meet the training requirement of a deep neural network.

[0006] At present, there is still a lack of a general method for diagnosing faults based on electrical characteristics. SUMMARY

[0007] The purpose of the present application is to provide a nuclear power emergency generator fault diagnosis method based on multiple parameter fusion.

[0008] The present application discloses a nuclear power emergency generator fault diagnosis method based on multiple parameter fusion, which comprises the following steps:

[0009] Step 1, through the interaction between the relay protection tester and the digital comprehensive protection device, various types of electrical faults of the emergency diesel generator are simulated to obtain transient voltage and current data of each type of electrical fault;

[0010] Step 2, through time-frequency transformation and image fusion, the change rule of the fault transient characteristics in the time domain and frequency domain dimensions is described, and the initial input data is preprocessed into the reconstructed time-frequency graph data set corresponding to the fault scene required by the neural network;

[0011] Step 3, the reconstructed time-frequency graph data set is divided into a training sample set and a test sample set of fault types, the former is sent into the CNN-LSTM network for training to optimize the network hyperparameters, and the latter is used to test the trained network to check whether the classification accuracy meets the standard;

[0012] Step 4, after the network training meets the standard, the network model is deployed on the embedded device, and when the fault occurs again, intelligent identification of the electrical fault can be realized.

[0013] Further, in step 2, the wavelet basis function is respectively convolved with the three-phase voltage and current signals to obtain the distribution of different frequency components of the signals in the time scale, and the wavelet function wavelet time-frequency graph is obtained:

[0014]

[0015] denotes the scale, denotes the translation, denotes the wavelet function of the two variables scale and translation , denotes the wavelet basis function, denotes the input three-phase voltage or current signal, denotes time.

[0016] Further, in step 2, the specific transformation formula of the reconstructed time-frequency graph is:

[0017]

[0018]

[0019]

[0020]

[0021]

[0022]

[0023] M1 to M6 are wavelet time-frequency maps after three-phase voltage and current conversion; 、 、 respectively represent three-phase voltage, 、 、 respectively represent three-phase current, represent the wavelet function of two variables as scale and translation amount ; to are fusion time-frequency maps after Alpha fusion processing, represent the transparency coefficient, which is manually specified according to application scenarios and requirements; is the data point set of input data to ; represent normalization processing results, respectively corresponding to ; and respectively represent the minimum value and the maximum value in the data point set of each input data; represent dimensionless data after normalization processing, and P is the reconstructed time-frequency map obtained by fusion superposition.

[0024] Further, in step 3, the CNN-LSTM model is as shown in Figure 3 , including 1 input layer, 3 local feature learning blocks, 1 expansion layer, 1 LSTM layer, 1 fully connected layer, 1 softmax layer and 1 output layer, wherein the local feature learning block includes 1 convolutional layer, 1 batch normalization layer, 1 activation layer and 1 maximum pooling layer.

[0025] The beneficial effects achieved by the present application are:

[0026] Compared with the traditional method, the fault diagnosis method does not depend on the setting of the protection setting value, only collects three-phase voltage and current waveform data before and after the fault point, and performs time-frequency change and fusion processing on the data; in the process of time-frequency conversion, the change law of the fault transient characteristics in the time domain and the frequency domain is described, and the two-dimensional time-frequency spectrum presents more fault transient characteristics than the one-dimensional time domain waveform; in the process of fusion reconstruction, the transient voltage and transient current characteristics are mapped to each other, the internal relationship between the fault transient characteristics and the fault type is excavated, and finally the CNN-LSTM network is relied on to realize accurate classification of electrical faults. BRIEF DESCRIPTION OF DRAWINGS

[0027] Figure 1 is a multi-parameter fusion nuclear power emergency generator fault diagnosis method flow chart;

[0028] Figure 2 A schematic diagram for continuous wavelet transform processing of fault voltage and current signals;

[0029] Figure 3 A schematic diagram of a CNN-LSTM model structure;

[0030] Figure 4 A schematic diagram of an LSTM layer structure. DETAILED DESCRIPTION

[0031] The advantages and features of the present application will become more apparent with the description. However, these embodiments are only exemplary and do not constitute any limitation on the scope of the present application. Those skilled in the art should understand that the details and forms of the technical solutions of the present application can be modified or replaced without departing from the spirit and scope of the present application, and such modifications and replacements all fall within the protection scope of the present application.

[0032] As shown in Figure 1 , the specific technical solution of the present application provides a multi-parameter fusion nuclear power emergency generator fault diagnosis method, comprising the following steps:

[0033] Step 1: Through the interaction between the relay protection tester and the digital comprehensive protection device, the relay protection tester inputs multiple groups of voltage and current signals to the digital comprehensive protection device, jointly simulates various electrical faults of the emergency diesel generator, collects the time domain waveforms of each 5 cycles before and after the fault point in the digital comprehensive protection device, and obtains the transient voltage and current data of each type of electrical fault.

[0034] Step 2: Through time-frequency transformation and image fusion, the change rule of the fault transient characteristics in the time domain and frequency domain dimensions is described, the internal relationship between the fault transient characteristics and the fault type is mined, and the initial input data is preprocessed into the fault scene data set required by the neural network;

[0035] Through wavelet transform, two variables are scale and translation , the scale controls the stretching of the wavelet function , and the translation controls the translation of the wavelet function. The scale is inversely proportional to the frequency, and the translation is proportional to the time . After the wavelet basis function is convolved with the three-phase voltage and current signals , the distribution of different frequency components of the signals in the time scale can be obtained, i.e., the wavelet time-frequency diagram of the wavelet function.

[0036]

[0037] representing scale, representing translation, representing wavelet function with two variables as scale and translation , representing wavelet basis function, representing input signal, representing time.

[0038] The one-dimensional time-domain signal obtained is transformed into a two-dimensional time-frequency diagram by using the above wavelet transform. The wavelet transform has the characteristic of being localized in time. By scaling and translating the wavelet function, the one-dimensional time-domain signal can be decomposed at different scales and positions, thereby obtaining a two-dimensional representation of the one-dimensional time-domain signal and generating a two-dimensional array composed of wavelet coefficients, wherein each wavelet coefficient represents the wavelet component of the two-dimensional array at that position and specific scale. These wavelet coefficients can be visualized by grayscale or color mapping. The size of the wavelet coefficient is mapped to the change of color or brightness, and the two-dimensional time-frequency diagram is visualized. The fault voltage and current signals are processed by continuous wavelet transform, as shown in Figure 2 .

[0039] The specific transformation formula is as follows:

[0040]

[0041] (1)

[0042]

[0043] (2)

[0044] (3)

[0045] (4)

[0046] Equation (1) represents time-frequency transform, , , representing three-phase voltage, , , representing three-phase current, representing wavelet function with two variables as scale and translation , M1 to M6 are the wavelet time-frequency diagrams after transformation of three-phase voltage and current, and the surrounding redundant display information of the time-frequency diagram is removed and compressed to 224x224x3 size; equation (2) represents Alpha fusion processing, To is the fusion time-frequency map after Alpha fusion processing, represents the transparency coefficient, which is manually specified according to the application scene and requirements; formula (3) represents normalization, is the input data To is the data point set, represents the normalization processing result, which corresponds to respectively; and respectively represent the minimum value and the maximum value in the data point set of each input data; respectively represent the dimensionless data after normalization processing, formula (4) represents fusion reconstruction, and P is the reconstructed time-frequency map obtained by fusion superposition, wherein the gradual degree of color reflects the complexity of the distribution law of the fault transient characteristic.

[0047] Step 3, the reconstructed time-frequency map obtained in step 2 is divided into a training sample set and a test sample set of fault types, the former is sent into the CNN-LSTM network for training, and the hyperparameters of the network are optimized, and the latter is used to test the trained network to see if the classification accuracy meets the standard;

[0048] The constructed CNN-LSTM model is as shown in Figure 3 , which includes 1 input layer, 3 designed local feature learning blocks (block), 1 expansion layer, 1 LSTM layer, 1 fully connected layer, 1 softmax layer and 1 output layer, wherein the block includes 1 convolutional layer, 1 batch normalization layer, 1 activation layer and 1 maximum pooling layer.

[0049] The network is constructed to learn fault features from the reconstructed time-frequency spectrum. When the reconstructed time-frequency spectrum in the form of a matrix is input into the network, 3 blocks can learn local features related to fault transients. The features output by the block are reshaped into a time sequence form and expanded to be input into the LSTM layer, and then the time dependence relationship is learned from these local features. The learning of local features and time dependence is as shown in the figure, the features output by the LSTM layer contain fault transient local information and global input information correlation, the fully connected layer is used to generalize these features to the output space, the softmax is used to predict and convert the learned features containing fault local features and global front-back dependence into probabilities, and the output layer outputs the classification result.

[0050] Step 4, after the network training meets the standard, the network model is deployed to an embedded device, and when a fault occurs again, intelligent identification of electrical faults can be realized.

[0051] The above merely describes the preferred embodiments of the present application, and is not intended to limit the scope of protection of the present application; any technical solution formed by equivalent transformation or equivalent replacement falls within the scope of protection of the present application; and the part not described in detail in the present application belongs to the common knowledge of the skilled in the art.

Claims

1. A multi-parameter fusion nuclear power emergency generator fault diagnosis method, characterized in that, The multi-parameter fusion nuclear power emergency generator fault diagnosis method comprises the following steps: Step 1, interact with the digital integrated protection device through the relay protection tester, jointly simulate various electrical faults of the emergency diesel generator, and obtain transient voltage and current data of various types of electrical faults; Step 2, through time-frequency transformation and image fusion, the change rule of the fault transient characteristics in the time domain and frequency domain dimensions is described, and the initial input data is preprocessed into the reconstructed time-frequency graph data set corresponding to the fault scene required by the neural network; Step 3, the reconstructed time-frequency graph data set is divided into a fault type training sample set and a test sample set, the former is sent into the CNN-LSTM network for training, the hyperparameters of the network are optimized, and the latter is used to test the trained network to see if the classification accuracy meets the standard; Step 4, after the network training meets the standard, the network model is deployed on the embedded device, and when the fault occurs again, the intelligent identification of the electrical fault can be realized.

2. The multi-parameter fusion nuclear power emergency generator fault diagnosis method according to claim 1, characterized in that, In step 2, the wavelet base function respectively with three-phase voltage, current signals After convolution, the distribution of different frequency components of the signal in the time scale is obtained, and the wavelet function wavelet time-frequency diagram is obtained: ； denotes a scale, denotes a translation, denotes a wavelet function of two variables being a scale and a translation denotes a wavelet basis function, denotes an input three-phase voltage or current signal, denotes time.​ 3. The multi-parameter fusion nuclear power emergency generator fault diagnosis method according to claim 2, characterized in that, In step 2, the specific transformation formula of the reconstructed time-frequency graph is: ， ； ， ； ； ； M1 to M6 are the wavelet time-frequency maps after three-phase voltage and current conversion; 、 、 respectively represent three-phase voltage, 、 、 respectively represent three-phase current, represents the wavelet function of two variables as scale and translation ; to are the fusion time-frequency maps after Alpha fusion processing, represents the transparency coefficient, which is manually specified according to the application scene and requirements; is the data point set of input data to , represents normalization processing results, respectively corresponding to ; and respectively represent the minimum value and the maximum value in the data point set of each input data; respectively represent non-dimensional data after normalization processing, and P is the reconstructed time-frequency map obtained by fusion superposition.

4. The multi-parameter fusion nuclear power emergency generator fault diagnosis method according to claim 1, characterized in that, In step 3, the CNN-LSTM model is shown in Figure 3, which includes 1 input layer, 3 local feature learning blocks, 1 expansion layer, 1 LSTM layer, 1 full connection layer, 1 softmax layer and 1 output layer, wherein the local feature learning block includes 1 convolution layer, 1 batch normalization layer, 1 activation layer and 1 maximum pooling layer.

Images