Power plant equipment fault diagnosis system based on information fusion

Abstract

The invention discloses a power plant equipment fault diagnosis system based on information fusion, and relates to the technical field of fault diagnosis, the power plant equipment fault diagnosis system comprises a fault classification unit, a fault reflection unit and a data acquisition unit, the data acquisition unit uses SCADA time sequence data to establish a bearing temperature prediction model based on light GBM, and in the operation process of a temperature monitoring instrument, the bearing temperature prediction model based on the light GBM is established; healthy historical data, abnormal historical data and real-time monitoring state data of the running state of the inner ring bearing are continuously collected, and the sampled data are classified, recorded and stored according to the forms of time, rotating speed, active power, resonant frequency and bandwidth. The data acquisition unit and the data processing unit are used for cleaning, feature extraction and feature integration of obtained sampling records, so that memory consumption is reduced, the fault prediction unit constructs a fault decision model prediction result, the abnormal residual processing unit effectively monitors the running state of the bearing, and the fault prediction result is more accurate. And finally, the fault diagnosis unit and the fault classification unit accurately diagnose the fault problem of the inner ring bearing of the temperature detection instrument.

Description

Technical Field

[0001] The present invention relates to the technical field of fault diagnosis, and in particular to a power plant equipment fault diagnosis system based on information fusion. Background Art

[0002] In various industrial sites, there are a large number of key equipment, whose normal and reliable operation is crucial to the operation and production economic benefits of the entire factory. If these equipment fail, it will cause huge economic losses to the entire industrial operation. If key equipment fails, the production line will stop running, the factory output will decrease, and huge economic losses will also be caused. In the patent with application number 202211248303.6, "a data fusion-based industrial equipment fault diagnosis system and method, belonging to the technical field of industrial equipment fault diagnosis, collects industrial equipment vibration signals, and collects vibration data of industrial equipment over time; converts the vibration signal into a time-frequency image through short-time Fourier transform, generates a spatial frequency image, calculates the energy of the central component of the spatial frequency in the spatial frequency image, and calculates the energy modal value; constructs a deep neural network model, inputs multiple energy modal values ​​into the neural network model, and obtains the error between the output of the neural network and the expected target through forward propagation; uses a synchronous transformation algorithm to analyze the output value of the neural network, and performs fault diagnosis on industrial equipment."

[0003] The above-mentioned documents can effectively identify and judge the working status and fault information of industrial equipment, and conduct status monitoring and health management of industrial equipment. However, at present, the transformers mainly used for power transmission in power plants are the core of the traction power supply system. The transformer capacity is extremely large and the reliability of the required work is extremely strong. The probability of accidents is small, but once an accident occurs, it will cause power equipment failure and cause fire, which will seriously affect the normal power supply safety of the power plant. The main transformer fault diagnosis is mainly achieved through temperature sensor monitoring instruments. Therefore, taking the inner ring bearing fault in the main transformer temperature monitoring instrument as the research object can effectively prevent the occurrence of safety accidents. In addition, if fault diagnosis is performed only when the power equipment is damaged, the safety risk is high. Therefore, it is also important to predict future trends based on past monitoring status. Summary of the invention

[0004] The purpose of the present invention is to provide a power plant equipment fault diagnosis system based on information fusion to solve the problems raised in the above background technology.

[0005] To achieve the above-mentioned object, the present invention provides the following technical solutions: A power plant equipment fault diagnosis system based on information fusion, comprising a fault classification unit and a fault reflection unit;

[0006] A data acquisition unit, which uses SCADA time series data to establish a bearing temperature prediction model based on lightGBM. During the operation of the temperature monitoring instrument, the data acquisition unit continuously collects the healthy history data, abnormal history data and real-time monitoring status data of the inner ring bearing operation status, and records and stores the sampled data in the form of time, speed, active power, resonance frequency and bandwidth;

[0007] A data processing unit, which cleans data mixed with abnormal noise, efficiently extracts hidden features from the original perception data acquired from the sensor, normalizes the extracted feature data with high correlation with the bearing, and screens and integrates the feature data for subsequent data provision;

[0008] A fault prediction unit, wherein the fault prediction unit calculates abnormal state residual features for the collected abnormal historical feature data, calculates healthy state residual features for the collected healthy historical feature data, and calculates monitoring state residual features for the collected real-time monitoring state feature data;

[0009] Abnormal residual processing unit, which uses the SPC method to screen the abnormal residual of bearing temperature within a controllable range according to the residual characteristics of abnormal state, and uses the AC-GAN algorithm to reconstruct the temperature residual of the residual sequence, so as to extract the residual characteristics of normal and abnormal bearing temperature, and then establishes a bearing state monitoring model based on NGBoost to effectively monitor the bearing operation state;

[0010] A fault diagnosis unit, wherein the fault diagnosis unit utilizes a fast spectral kurtosis method to adaptively extract the resonance frequency and bandwidth of the inner ring bearing of multiple groups of temperature monitoring instruments at different times, thereby obtaining complex Morlet wavelet coefficients, and then realizes signal processing of the temperature monitoring instrument and fault diagnosis based on the bandpass filtering and signal envelope characteristics of the complex Morlet wavelet, thereby improving the fault diagnosis accuracy and adaptability of the monitoring instrument.

[0011] Preferably, the data processing unit includes a data cleaning module, a feature extraction module and a feature integration module. The data cleaning module cleans the data collected by the data acquisition unit due to power failure and communication failure during equipment operation, and divides the interval according to the speed of 900-1000r / min as the standard, and removes the data exceeding the speed range and the data with active power less than and equal to 0. The feature extraction module uses a bidirectional long short-term memory group BG-LSTM combined with an LSTM moving forward along the sensed data and another LSTM moving backward along the sensed data to fully extract the characteristics of the bearing under complex operating conditions, and normalizes the data with different dimensional units. The feature integration module uses the gradient unilateral sampling technology through the LightGBM algorithm to filter most of the small gradient data, and only uses the data information with high gradient when calculating information. The independent feature merging technology is used to realize the bundling of mutually exclusive features, reduce the redundancy of sample feature data, and thus reduce memory consumption;

[0012] The specific formula of the bidirectional long short-term memory group BG-LSTM is:

[0013]

[0014] In the formula, and Represent the output of the forward layer and the backward layer respectively, w i represents the weight between two LSTM elements, x t represents the data sample, σ represents the specific conditions, b is the deviation vector, and Represent the bias vectors of the forward and backward layer outputs respectively.

[0015] Preferably, the fault prediction unit constructs a fault decision model with respect to the abnormal state residual features, the healthy state residual features and the monitoring state residual features to further predict the results.

[0016] Preferably, the temperature residual reconstruction in the abnormal residual processing unit includes the following steps:

[0017] A1. Input the characteristic data of the bearing one month before the failure into the LightGBM bearing temperature prediction model to calculate the bearing temperature abnormality residual within one month;

[0018] A2. Use SPC control chart to eliminate the normal temperature residuals of bearing temperature abnormalities within the control range;

[0019] A3. Define the residual data outside the control range as training sample data, and use the AC-GAN algorithm to generate residual data with a distribution similar to the real sample data to reconstruct the normal temperature residual eliminated in step A2.

[0020] Preferably, the fault diagnosis unit includes a parameter determination module and a signal processing module, the parameter determination module specifically analyzes the vibration signal collected by the sensor through a fast spectral kurtosis method, and determines the resonance frequency and bandwidth of the sensor signal based on the maximum spectral kurtosis principle, and determines the center frequency and scale of the complex Morlet wavelet as wavelet coefficients based on these two parameters, the signal processing module performs wavelet transform on the signal based on the determined wavelet coefficients, then performs structural analysis and envelope demodulation on the signal to obtain an envelope signal, and performs FFT transform on the envelope signal, thereby finding the fault frequency;

[0021] The complex Morlet wavelet is specifically:

[0022]

[0023] In the formula, t and f 0 denotes time and center frequency respectively, σ and i denote the control wavelet window width and imaginary unit respectively, and ψ(t) denotes the wavelet coefficient;

[0024] The envelope signal is specifically:

[0025]

[0026] Where X(t) represents the enveloped signal, Re and Im represent the real and imaginary parts of the analytical signal respectively, and S(t) represents the analytical signal.

[0027] Preferably, the fault diagnosis unit also includes a feature learning module and a fault analysis module. The feature learning module inputs the bearing fault features at different times into the feature learning model, and outputs the hidden state at the next time, and uses the generated hidden state as the input of the attention mechanism. The attention mechanism is used to change the weight of the feature learning model to strengthen the weight of the key features in the model. The fault analysis module calculates the characteristic value according to the amplitude of the fault frequency obtained in the signal processing module, and judges the state of the temperature monitoring instrument according to the threshold of the characteristic value to achieve fault diagnosis.

[0028] Preferably, the K final features obtained by the fault classification unit through the feature learning module are used for bearing fault classification, and global average pooling combined with the SoftMax layer is used to complete accurate bearing fault classification, including the following steps:

[0029] B1. Use the concatenation operation to fuse K features and calculate the feature vector;

[0030] B2, the feature vector is input into the SoftMax layer to derive the probability distribution of different fault categories;

[0031] B3. If the device has multiple states, the feature vector is first fed back to a fully connected layer to obtain the probability distribution parameters of each state.

[0032] Preferably, the fault reflection unit extracts characteristic data after transmitting the collected sample data to the data processing unit. If a device fault occurs, signal diagnosis of the characteristic data is required and the data is transferred to the fault diagnosis unit for processing. The diagnosis results, original data and predicted results are uploaded to the host, which stores the uploaded data and displays it in a visual form.

[0033] Compared with the prior art, the present invention has the following beneficial effects:

[0034] 1. The present invention establishes a bearing temperature prediction model based on lightGBM through a data acquisition unit, which has high prediction accuracy. During the operation of the temperature monitoring instrument, the healthy history data, abnormal history data and real-time monitoring status data of the inner ring bearing operation status are continuously collected to obtain sampling records. The collected data is cleaned, feature extracted and integrated through the data processing unit for subsequent data provision. Feature integration reduces the redundancy of sample feature data, thereby reducing memory consumption. The fault prediction unit uses the residual features of the abnormal state, the residual features of the healthy state and the residual features of the monitoring state to form a fault decision model, thereby predicting the result.

[0035] 2. The present invention solves the related problems of artificially setting thresholds by screening the abnormal temperature residuals of the bearing within a controllable range and reconstructing the temperature residuals of the residual sequence through the abnormal residual processing unit, improves the accuracy of the labeling of abnormal and normal data of the bearing, effectively monitors the operating status of the bearing, and not only facilitates the positioning of faults at the inner ring bearing of the temperature monitoring instrument, but also predicts future operating conditions from the acquired historical data to achieve effective prevention. The fault diagnosis unit realizes the signal processing of the temperature monitoring instrument according to the bandpass filtering and signal envelope characteristics of the complex Morlet wavelet. The algorithm can adaptively extract the resonant frequency of the signal, effectively filter out the interference of external noise, and improve the accuracy of bearing fault diagnosis. The fault classification unit is used to accurately classify the faults of the inner ring bearing of the temperature detection instrument. The fault reflection unit uses the host to upload the diagnosis results, original data and prediction results and displays them in a visual form, which is more intuitive and clear. BRIEF DESCRIPTION OF THE DRAWINGS

[0036] Figure 1 A schematic diagram of the overall system flow provided by an embodiment of the present invention;

[0037] Figure 2An internal module block diagram of a data processing unit provided by an embodiment of the present invention;

[0038] Figure 3 An internal module block diagram of a fault diagnosis unit provided by an embodiment of the present invention;

[0039] Figure 4 A flow chart of temperature residual reconstruction provided by an embodiment of the present invention;

[0040] Figure 5 A flowchart of fault classification provided by an embodiment of the present invention.

[0041] In the figure: 1. Data acquisition unit; 2. Data processing unit; 201. Data cleaning module; 202. Feature extraction module; 203. Feature integration module; 3. Fault prediction unit; 4. Abnormal residual processing unit; 5. Fault diagnosis unit; 501. Parameter determination module; 502. Signal processing module; 503. Feature learning module; 504. Fault analysis module; 6. Fault classification unit; 7. Fault reflection unit. DETAILED DESCRIPTION

[0042] The following will be combined with the drawings in the embodiments of the present invention to clearly and completely describe the technical solutions in the embodiments of the present invention. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without creative work are within the scope of protection of the present invention.

[0043] See also Figure 1-5 ,The present invention provides a technical solution: a power plant equipment fault diagnosis system based on information fusion, comprising a fault classification unit 6 and a fault reflection unit 7;

[0044] Data acquisition unit 1 uses SCADA time series data to establish a bearing temperature prediction model based on lightGBM. During the operation of the temperature monitoring instrument, the healthy history data, abnormal history data and real-time monitoring status data of the inner ring bearing operation status are continuously collected, and the sampled data is classified, recorded and stored in the form of time, speed, active power, resonance frequency and bandwidth;

[0045] Data processing unit 2: Data processing unit 2 cleans the data mixed with abnormal noise, efficiently extracts the hidden features in the original perception data obtained from the sensor, normalizes the extracted feature data with high correlation with the bearing, and screens and integrates the feature data for the provision of subsequent data;

[0046] The fault prediction unit 3 calculates the abnormal state residual feature for the collected abnormal historical feature data, calculates the healthy state residual feature for the collected healthy historical feature data, and calculates the monitoring state residual feature for the collected real-time monitoring state feature data;

[0047] Abnormal residual processing unit 4 uses the SPC method to screen the abnormal residual of the bearing temperature within a controllable range according to the residual characteristics of the abnormal state, and uses the AC-GAN algorithm to reconstruct the temperature residual of the residual sequence, thereby extracting the residual characteristics of normal and abnormal bearing temperature, and then establishing a bearing state monitoring model based on NGBoost to effectively monitor the operating state of the bearing. It is not only convenient to locate the fault at the inner ring bearing of the temperature monitoring instrument, but also predicts the future operating status from the acquired historical data to achieve effective prevention;

[0048] The fault diagnosis unit 5 uses a fast spectral kurtosis method to adaptively extract the resonant frequency and bandwidth of the inner ring bearing of multiple groups of temperature monitoring instruments at different times, thereby obtaining complex Morlet wavelet coefficients, and then implements fault diagnosis based on the bandpass filtering and signal envelope characteristics of the complex Morlet wavelet, thereby improving the fault diagnosis accuracy and adaptability of the monitoring instrument.

[0049] The data processing unit 2 includes a data cleaning module 201, a feature extraction module 202 and a feature integration module 203. The data cleaning module 201 cleans the data collected by the data acquisition unit due to power failure and communication failure during equipment operation. The data is divided into intervals based on the speed of 900-1000r / min as the standard, and the data exceeding the speed range and the data with active power less than or equal to 0 are removed. The feature extraction module 202 uses a bidirectional long short-term memory group BG-LSTM combined with an LSTM moving forward along the sensed data and another LSTM moving backward along the sensed data to fully extract the characteristics of the bearing under complex operating conditions, and normalize the data with different dimensional units. The feature integration module 203 uses the gradient unilateral sampling technology through the LightGBM algorithm to filter most of the small gradient data, and only uses the data information with high gradient when calculating information, and uses the independent feature merging technology to achieve the bundling of mutually exclusive features;

[0050] The specific formula of the bidirectional long short-term memory group BG-LSTM is:

[0051]

[0052] In the formula, and Represent the output of the forward layer and the backward layer respectively, w i represents the weight between two LSTM elements, xt represents the data sample, σ represents the specific conditions, b is the deviation vector, and Represent the bias vectors of the forward and backward layer outputs respectively.

[0053] The fault prediction unit 3 combines the residual features of the abnormal state, the residual features of the healthy state and the residual features of the monitoring state to form a fault decision model and further predict the results;

[0054] The temperature residual reconstruction in the abnormal residual processing unit 4 includes the following steps:

[0055] A1. Input the characteristic data of the bearing one month before the failure into the LightGBM bearing temperature prediction model to calculate the bearing temperature abnormality residual within one month;

[0056] A2. Use SPC control chart to eliminate the normal temperature residuals of bearing temperature abnormalities within the control range;

[0057] A3, define the residual data outside the control range as training sample data, use the AC-GAN algorithm to generate residual data with a distribution similar to the real sample data, and use it to reconstruct the normal temperature residual eliminated in step A2;

[0058] The fault diagnosis unit 5 includes a parameter determination module 501 and a signal processing module 502. The parameter determination module 501 specifically analyzes the vibration signal collected by the sensor through the fast spectral kurtosis method, and determines the resonance frequency and bandwidth of the sensor signal based on the maximum spectral kurtosis principle, and determines the center frequency and scale of the complex Morlet wavelet as the wavelet coefficients according to these two parameters. The signal processing module 502 performs wavelet transform on the signal based on the determined wavelet coefficients, then performs structural analysis and envelope demodulation on the signal, and performs FFT transform on the envelope signal, so as to find the fault frequency;

[0059] The complex Morlet wavelet is specifically:

[0060]

[0061] In the formula, t and f 0 denote time and center frequency respectively, σ and i denote the control wavelet window width and imaginary unit respectively, ψ(t) denotes the wavelet coefficient;

[0062] The envelope signal is specifically:

[0063]

[0064] Where X(t) represents the enveloped signal, Re and Im represent the real and imaginary parts of the analytical signal respectively, and S(t) represents the analytical signal;

[0065] The fault diagnosis unit 5 also includes a feature learning module 503 and a fault analysis module 504. The feature learning module 503 inputs the bearing fault features at different times into the feature learning model, and outputs the hidden state at the next time, and uses the generated hidden state as the input of the attention mechanism. The attention mechanism is used to change the weight of the feature learning model to strengthen the weight of the key features in the model. The fault analysis module 504 calculates the characteristic value according to the amplitude of the fault frequency obtained in the signal processing module 502, and judges the state of the temperature monitoring instrument according to the threshold of the characteristic value to realize fault diagnosis;

[0066] The K final features obtained by the fault classification unit 6 through the feature learning module 503 are used for bearing fault classification, and global average pooling combined with the SoftMax layer is used to complete accurate bearing fault classification, including the following steps:

[0067] B1. Use the concatenation operation to fuse K features and calculate the feature vector;

[0068] B2, the feature vector is input into the SoftMax layer to derive the probability distribution of different fault categories;

[0069] B3. If the device has multiple states, the feature vector is first fed back to a fully connected layer to obtain the probability distribution parameters of each state;

[0070] The fault reflection unit 7 extracts characteristic data after transmitting the collected sample data to the data processing unit 2. If a device fault occurs, signal diagnosis of the characteristic data is required and the data is transferred to the fault diagnosis unit 5 for processing. The diagnosis results, original data and prediction results are uploaded to the host, which uses the host to store the uploaded data and display it in a visual form.

[0071] The bearing temperature prediction model based on lightGBM is as follows:

[0072]

[0073] In the formula, y i is the actual measured value of the i-th bearing temperature, is the predicted value of the i-th bearing temperature, is the mean of the true measured values ​​of the main bearing temperature, RMSE is the root mean square error, r 2 is the determination coefficient.

[0074] Working principle: When the present invention is used, the SCADA time series data is used to establish a bearing temperature prediction model based on lightGBM through the data acquisition unit 1. During the operation of the temperature monitoring instrument, the health history data, abnormal history data and real-time monitoring status data of the inner ring bearing operation status are continuously collected, and the sampled data are classified and recorded and stored in the form of time, rotation speed, active power, resonance frequency and bandwidth. The data collected by the data acquisition unit 1 due to power failure and communication failure during equipment operation is cleaned by the data processing unit 2. A bidirectional long short-term memory group BG-LSTM is used to combine an LSTM that moves forward along the sensed data and another LSTM that moves backward along the sensed data to fully extract the characteristics of the bearing under complex operating conditions. The gradient unilateral sampling technology is used by the LightGBM algorithm to filter most of the small gradient data. In order to reduce the redundancy of sample feature data when calculating information, only data information with high gradient is used. The independent feature merging technology is used to achieve the bundling of mutually exclusive features, thereby reducing memory consumption.

[0075] The fault prediction unit 3 calculates the abnormal state residual features of the collected abnormal historical feature data, calculates the healthy state residual features of the collected healthy historical feature data, and calculates the monitoring state residual features of the collected real-time monitoring state feature data. These three categories are combined to form a fault decision model to further predict the results. The abnormal residual processing unit 4 uses the SPC method to screen the abnormal bearing temperature residual within a controllable range based on the abnormal state residual features, and uses the AC-GAN algorithm to reconstruct the temperature residual of the residual sequence. The feature data one month before the bearing failure is input into the LightGBM bearing temperature prediction model to calculate the bearing temperature abnormality within one month. Residuals: The SPC control chart is used to eliminate the normal temperature residuals within the control range of the abnormal bearing temperature residuals. The residual data outside the control range is defined as the training sample data. The AC-GAN algorithm is used to generate residual data with a distribution similar to the real sample data to reconstruct the eliminated normal temperature residuals. This solves the problems related to artificially set thresholds and improves the accuracy of labeling abnormal and normal data for bearings, thereby extracting the residual features of normal and abnormal bearing temperatures. Then, a bearing status monitoring model based on NGBoost is established to effectively monitor the operating status of the bearing. This not only facilitates the location of faults at the inner ring bearing of the temperature monitoring instrument, but also predicts future operating conditions from the acquired historical data to achieve effective prevention.

[0076] The fault diagnosis unit 5 uses the fast spectral kurtosis method to adaptively extract the resonance frequency and bandwidth of the inner ring bearing of multiple groups of temperature monitoring instruments at different times, thereby obtaining the complex Morlet wavelet coefficients. Then, according to the bandpass filtering and signal envelope characteristics of the complex Morlet wavelet, the temperature monitoring instrument signal processing is realized, thereby improving the fault diagnosis accuracy and adaptability of the monitoring instrument. The K final features obtained by the fault classification unit 6 through the feature learning module 503 will be used for bearing fault classification, and the GAP combined with the SoftMax layer will be used to complete the accurate bearing fault classification. The fault reflection unit 7 extracts the feature data after transmitting the collected sample data to the data processing unit 2. If the equipment fails, the feature data needs to be diagnosed by signal, and is transferred to the fault diagnosis unit 5 for processing. The diagnosis results, original data and prediction results are uploaded to the host, and the uploaded data is stored in the host and displayed in a visual form.

[0077] Although embodiments of the present invention have been shown and described, it will be appreciated by those skilled in the art that various changes, modifications, substitutions and variations may be made to the embodiments without departing from the principles and spirit of the present invention, and that the scope of the present invention is defined by the appended claims and their equivalents.

Claims

1. A power plant equipment fault diagnosis system based on information fusion, comprising a fault classification unit (6) and a fault reflection unit (7), characterized in that: A data acquisition unit (1), wherein the data acquisition unit (1) uses SCADA time series data to establish a bearing temperature prediction model based on lightGBM, and continuously collects healthy history data, abnormal history data and real-time monitoring status data of the inner ring bearing operation status during the operation of the temperature monitoring instrument, and records and stores the sampled data in the form of time, rotation speed, active power, resonance frequency and bandwidth; A data processing unit (2), wherein the data processing unit (2) cleans data mixed with abnormal noise, efficiently extracts hidden features from original sensing data acquired from the sensor, normalizes the extracted feature data with high correlation with the bearing, and screens and integrates the feature data for providing subsequent data; A fault prediction unit (3), wherein the fault prediction unit (3) calculates abnormal state residual features for the collected abnormal historical feature data, calculates healthy state residual features for the collected healthy historical feature data, and calculates monitoring state residual features for the collected real-time monitoring state feature data; An abnormal residual processing unit (4), wherein the abnormal residual processing unit (4) uses an SPC method to screen the abnormal residual of the bearing temperature within a controllable range according to the abnormal state residual characteristics, and uses an AC-GAN algorithm to reconstruct the temperature residual of the residual sequence, thereby extracting the residual characteristics of normal and abnormal bearing temperatures, and then establishing a bearing state monitoring model based on NGBoost to effectively monitor the bearing operation state; A fault diagnosis unit (5) is provided, wherein the fault diagnosis unit (5) uses a fast spectral kurtosis method to adaptively extract the resonance frequency and bandwidth of the inner ring bearing of multiple groups of temperature monitoring instruments at different times, thereby obtaining complex Morlet wavelet coefficients, and then realizes fault diagnosis based on the bandpass filtering and signal envelope characteristics of the complex Morlet wavelet.

2. The power plant equipment fault diagnosis system based on information fusion according to claim 1 is characterized in that: The data processing unit (2) comprises a data cleaning module (201), a feature extraction module (202) and a feature integration module (203). The data cleaning module (201) cleans the data collected by the data collection unit (1) from abnormal noise data caused by power failure and communication failure during equipment operation, divides the interval into intervals based on a rotation speed of 900 to 1000 r / min, and removes data exceeding the rotation speed range and data with active power less than or equal to 0. The feature extraction module (202) uses a bidirectional long short-term memory group BG-LSTM combined with an LSTM that moves forward along the sensed data and another LSTM that moves backward along the sensed data to fully extract the characteristics of the bearing under complex operating conditions, and normalizes data with different dimensional units. The feature integration module (203) uses the gradient unilateral sampling technology through the LightGBM algorithm to filter most of the small gradient data, only uses data information with high gradient when calculating information, and uses independent feature merging technology to achieve bundling of mutually exclusive features.

3. The power plant equipment fault diagnosis system based on information fusion according to claim 1 is characterized in that: The fault prediction unit (3) constructs a fault decision model based on the abnormal state residual features, the healthy state residual features and the monitoring state residual features, and further predicts the results.

4. The power plant equipment fault diagnosis system based on information fusion according to claim 1 is characterized in that: The temperature residual reconstruction in the abnormal residual processing unit (4) comprises the following steps: A1. Input the characteristic data of the bearing one month before the failure into the LightGBM bearing temperature prediction model to calculate the bearing temperature abnormality residual within one month; A2. Use SPC control chart to eliminate the normal temperature residuals of bearing temperature abnormalities within the control range; A3. Define the residual data outside the control range as training sample data, and use the AC-GAN algorithm to generate residual data with a distribution similar to the real sample data to reconstruct the normal temperature residual eliminated in step A2.

5. The power plant equipment fault diagnosis system based on information fusion according to claim 1 is characterized in that: The fault diagnosis unit (5) comprises a parameter determination module (501) and a signal processing module (502). The parameter determination module (501) specifically analyzes the vibration signal collected by the sensor through a fast spectral kurtosis method, and determines the resonance frequency and bandwidth of the sensor signal based on the maximum spectral kurtosis principle, and determines the center frequency and scale of the complex Morlet wavelet as wavelet coefficients based on the two parameters. The signal processing module (502) performs wavelet transform on the signal based on the determined wavelet coefficients, then performs structural analysis and envelope demodulation on the signal, and performs FFT transform on the envelope signal, thereby finding the fault frequency.

6. The power plant equipment fault diagnosis system based on information fusion according to claim 5 is characterized in that: The fault diagnosis unit (5) further comprises a feature learning module (503) and a fault analysis module (504). The feature learning module (503) inputs the bearing fault features at different moments into the feature learning model, and outputs the hidden state at the next moment, and uses the generated hidden state as the input of the attention mechanism. The attention mechanism is used to change the weight of the feature learning model to strengthen the weight of the key features in the model. The fault analysis module (504) calculates the characteristic value according to the amplitude of the fault frequency obtained in the signal processing module (502), and judges the state of the temperature monitoring instrument according to the threshold of the characteristic value to achieve fault diagnosis.

7. The power plant equipment fault diagnosis system based on information fusion according to claim 6 is characterized in that: The K final features obtained by the fault classification unit (6) through the feature learning module (503) are used for bearing fault classification, and global average pooling combined with a SoftMax layer is used to complete accurate bearing fault classification, including the following steps: B1. Use the concatenation operation to fuse K features and calculate the feature vector; B2, the feature vector is input into the SoftMax layer to derive the probability distribution of different fault categories; B3. If the device has multiple states, the feature vector is first fed back to a fully connected layer to obtain the probability distribution parameters of each state.

8. The power plant equipment fault diagnosis system based on information fusion according to claim 1 is characterized in that: The fault reflection unit (7) extracts characteristic data after transmitting the collected sample data to the data processing unit (2). If a fault occurs in the equipment, signal diagnosis of the characteristic data is required and the data is transferred to the fault diagnosis unit (5) for processing. The diagnosis result, original data and prediction result are uploaded to the host, and the uploaded data is stored in the host and displayed in a visual form.

Images