A power station fan status early warning method, system and application thereof

Abstract

A power station fan status early warning method, system and application thereof, comprising the following steps: collecting normal operating signals of the fan, collecting time series of process parameters of the fan, extracting and constructing wavelet energy spectrum from the collected vibration data time series under normal operating conditions of the fan using a wavelet packet decomposition method; extracting and constructing frequency band energy spectrum from the collected noise data time series under normal operating conditions of the fan using a 1 / 3 octave decomposition method; obtaining an early warning model; constructing Euclidean distance and cosine distance similarity indicators between model observation values ​​and reconstructed values; using a Bayesian maximum a posteriori probability estimation method to construct a fan status early warning indicator based on multi-source information fusion, setting a status early warning threshold, and judging whether a fault occurs according to the fault early warning threshold; the present invention realizes online real-time early warning of the fan status, issues an early warning before a fault occurs, buys time for on-site personnel, and reduces safety and economic losses.

Description

Technical Field

[0001] The present invention relates to the technical field of status monitoring of fans of thermal power generation units, and in particular to a method and system for early warning status of fans of power stations, and applications thereof. Background Art

[0002] Under the current background of "dual carbon" goals, with the continuous deepening of the reform of the power market system, traditional thermal power companies are facing increasingly fierce industry competition, and higher requirements are also put forward for the operation flexibility and peak-shaving capacity of the units. It is increasingly important to maintain the long-term stable operation of the units in terms of safety, economy and environmental protection. How to reasonably arrange the maintenance cycle of the units and improve the utilization rate of various main and auxiliary equipment has become a key link. However, in practice, the status monitoring of power station auxiliary equipment mainly includes the limit alarm of DCS and the inspection and maintenance of on-site personnel. Such methods are less dynamic and real-time, and belong to post-maintenance. With the increase of operating years and the increase in the frequency of unit load changes, the degree of wear and aging increases, which seriously affects the economy and safety of the units.

[0003] At present, the existing equipment monitoring technology is prone to false alarms and failure to identify faults due to the lack of fault data and the single detection method. Since the amount of data transmitted during the monitoring process is usually large, the host computer has a heavy processing burden, resulting in poor real-time warning. In addition, during the deep peak-shaving process, a series of safety problems will arise in the equipment (fan stall, surge, and large vibration caused by small flow), which further amplifies the shortcomings of the existing technology. Therefore, it is necessary to develop a highly real-time quantitative evaluation technology for the auxiliary machine status to complete the transformation from "post-maintenance" to "pre-prediction". Summary of the invention

[0004] In order to overcome the above-mentioned defects of the prior art and meet the basic requirements of safe and reliable operation of power station fans, the present invention provides a power station fan status warning method and system to improve the monitoring efficiency of the fan status. First, the online detection of multi-dimensional state parameters such as current, pressure, flow, temperature, vibration, and noise is realized by using traditional and new detection technologies. A feature extraction method for vibration and noise signals is proposed at the front end, and a fault state assessment and warning algorithm based on a stacked sparse denoising autoencoder network is proposed at the terminal. An important auxiliary machine state monitoring and warning system based on an edge computing architecture is developed, and the functional verification of the auxiliary machine state assessment and warning system is completed.

[0005] To achieve the above object, the present invention provides the following technical solutions:

[0006] The present invention provides a power station fan status early warning method, which specifically comprises the following steps:

[0007] Step 1: Collect vibration and noise signals of normal operation of the fan;

[0008] Step 2: Collecting the time series of process parameters such as fan flow, current, and power; the fan process parameter time series includes fan operation data and the fan operation status corresponding to the fan operation data, and the fan status includes a normal operation status and a fault status;

[0009] Step 3: Using a wavelet packet decomposition method, extract and construct a wavelet energy spectrum from the vibration data time series collected under normal operating conditions of the fan;

[0010] Step 4: Use the 1 / 3 octave decomposition method to extract and construct the frequency band energy spectrum from the noise data time series collected under the normal operation state of the fan;

[0011] Step 5: Taking the process parameters, vibration, and noise energy spectrum data as input, respectively, a stacked sparse denoising autoencoder network is constructed to perform model training to obtain a process parameter, fan vibration, and noise warning model;

[0012] Step 6: Obtain the process parameter, vibration, and noise energy spectrum data Xobs within the time period to be evaluated, input the process parameter, fan vibration, and noise warning models respectively, and obtain the model reconstruction value Xrec;

[0013] Step 7: Construct the model based on the observed values ​​X obs and the reconstruction value X rec Similarity indicators of Euclidean distance and cosine distance between them;

[0014] Step 8: Use the Bayesian maximum a posteriori probability estimation method to fuse the similarity indicators of the above single warning models, construct a wind turbine status warning indicator based on multi-source information fusion, set the status warning threshold, and determine whether a fault occurs based on the fault warning threshold.

[0015] The present invention discloses a method for applying the above-mentioned power station fan status early warning method to a power station fan status early warning system.

[0016] The invention discloses a method for monitoring the status of a fan of a thermal power unit, which is characterized by comprising the above-mentioned method for early warning the status of a fan of a power station.

[0017] Beneficial effects:

[0018] 1. The present invention adopts advanced sound pressure sensors and acceleration sensors to achieve high-precision noise and vibration signal collection, making up for the shortcomings of insufficient detection means of current power plant equipment.

[0019] 2. The present invention adopts a vibration signal feature extraction method based on wavelet packet decomposition, which can not only dynamically adjust the time resolution by frequency, but also obtain high-frequency signals while extracting low-frequency signals.

[0020] 3. The present invention selects generator power, motor current, motor winding temperature, drive end bearing temperature, free end bearing temperature, hydraulic oil pressure, hydraulic oil temperature, lubricating oil pressure, lubricating oil temperature, fan blade opening, inlet air temperature, outlet air temperature and outlet air pressure as fan status monitoring parameters. In the actual operation of the thermal power unit, it is very necessary and effective to select the above monitoring parameters to monitor the fan.

[0021] 4. The present invention adopts a stacked sparse denoising autoencoder network algorithm, which increases the robustness of the model compared to existing machine learning or deep learning models.

[0022] 5. The multi-source information fusion method adopted in the present invention improves the reliability of the overall evaluation of the system.

[0023] 6. The fan status warning system designed by the present invention can realize online real-time warning of the fan status, send out a warning signal before a fault occurs, buy processing time for on-site personnel, and reduce safety and economic losses caused by the fault. BRIEF DESCRIPTION OF THE DRAWINGS

[0024] Figure 1 The original signal of fan vibration and its frequency band energy;

[0025] Where a) original signal of horizontal vibration of fan;

[0026] b) Horizontal vibration signal feature extraction results;

[0027] c) Original signal of vertical vibration of fan;

[0028] d) Vertical vibration signal feature extraction results;

[0029] Figure 2 Original signal of fan noise and 1 / 3 octave feature extraction results;

[0030] Where: a) is the original signal of fan noise

[0031] b) is: Noise signal feature extraction result

[0032] Figure 3 Training flow chart for stacked sparse denoising autoencoder network model;

[0033] Figure 4 Training error for stacked sparse denoising autoencoder network model

[0034] Figure 5 This is the wind turbine online status warning flow chart;

[0035] Figure 6 This is a structural diagram of the fan status warning system. DETAILED DESCRIPTION

[0036] A power station fan status early warning method specifically comprises the following steps:

[0037] Step 1: Collect vibration and noise signals of normal operation of the fan;

[0038] Step 2: Collecting the time series of process parameters such as fan flow, current, and power; the fan process parameter time series includes fan operation data and the fan operation status corresponding to the fan operation data, and the fan status includes a normal operation status and a fault status;

[0039] Step 3: Using a wavelet packet decomposition method, extract and construct a wavelet energy spectrum from the vibration data time series collected under normal operating conditions of the fan;

[0040] Step 4: Use the 1 / 3 octave decomposition method to extract and construct the frequency band energy spectrum from the noise data time series collected under the normal operation state of the fan;

[0041] Step 5: Taking the process parameters, vibration, and noise energy spectrum data as input, respectively, a stacked sparse denoising autoencoder network is constructed to perform model training to obtain a process parameter, fan vibration, and noise warning model;

[0042] Step 6: Obtain the process parameter, vibration, and noise energy spectrum data Xobs within the time period to be evaluated, input the process parameter, fan vibration, and noise warning models respectively, and obtain the model decoding output, which is the reconstructed value Xrec;

[0043] Step 7: Construct the model based on the observed values ​​X obs and the reconstruction value X rec Similarity indicators of Euclidean distance and cosine distance between them;

[0044] Step 8: Use the Bayesian maximum a posteriori probability estimation method to fuse the similarity indicators of the above single warning models, construct a wind turbine status warning indicator based on multi-source information fusion, set the status warning threshold, and determine whether a fault occurs based on the fault warning threshold.

[0045] Furthermore, the step 1 specifically includes the following steps:

[0046] Step 1.1, install an acoustic sensor with a chassis fixed bracket at a distance of 50 cm from the fan motor and the fan body, and install a vibration acceleration sensor in the X / Y direction of the fan bearing housing with strong glue;

[0047] Step 1.2: collect two noise signals of the fan at a sampling frequency of 25.6 kHz, and collect two vibration signals at a sampling frequency of 51.2 kHz.

[0048] Further, step 2 includes the following steps:

[0049] Step 2.1, collecting generator power, motor current, motor winding temperature, drive end bearing temperature, free end bearing temperature, hydraulic oil pressure, hydraulic oil temperature, lubricating oil pressure, lubricating oil temperature, fan blade opening, inlet air temperature, outlet air temperature and outlet air pressure data through a distributed control system or safety instrument system connected to the fan;

[0050] Step 2.2, extract 20,000 sets of data from the historical data, calculate the absolute value of the Pearson correlation coefficient between each parameter, and retain only one of the parameters with an absolute value of the correlation coefficient between each pair greater than 0.6 as the modeling variable.

[0051] Further, step 3 includes the following steps:

[0052] Step 3.1, perform spectrum analysis on the collected vibration signal to determine the appropriate frequency band for wavelet packet decomposition;

[0053] Step 3.2, determine the number of layers of wavelet packet decomposition. When performing wavelet packet decomposition, the number of decomposition layers has a significant correlation with the analysis accuracy of time-frequency domain signals;

[0054] Step 3.3, decompose and reconstruct the signal using wavelet packets, and calculate the energy of each sub-band signal to obtain the reconstructed signal energy spectrum;

[0055] Step 3.4, divide the reconstructed signal energy spectrum into adjacent and equal frequency bands, and calculate the normalized signal energy ratio of each frequency band.

[0056] In step 3.3, the wavelet packet function expression is:

[0057]

[0058] Where: l is the oscillation coefficient, k is the position index, h k and g k They are the high-pass and low-pass orthogonal filter coefficients derived from the scaling function respectively; Daubechies wavelet is used to perform third-order wavelet packet decomposition on the vibration signal.

[0059] The calculation formula of frequency band energy is:

[0060]

[0061] Where: E j is the energy of the jth node of the subband signal; S j (t) is the original signal, x j is the discrete point amplitude of the original signal, m / s2; N is the number of sampling points.

[0062] In step 3.4, the third-order four low-frequency components and four high-frequency components obtained by three-layer wavelet packet decomposition are combined into a feature vector T with 8 columns.

[0063]

[0064] Where T is the normalized feature vector of a single sample; E j is the jth reconstructed signal of the third layer. The feature vectors are normalized, and each vector is a percentage of energy, which is convenient for removing the dimension.

[0065] Further, step 4 includes the following steps:

[0066] Step 4.1, performing fast Fourier transform on the collected noise signal;

[0067] Step 4.2, add the Hanning window according to the power equal restoration coefficient to obtain the preliminary spectrum of the noise;

[0068] Step 4.3, extract the nominal value of the center frequency and the 1 / 3 octave range, and use the vibration level method to convert the graph into an unweighted 1 / 3 octave sound pressure level spectrum diagram;

[0069] Step 4.4, after C-weighted sound level correction, the C-weighted 1 / 3 octave band sound pressure level spectrum of the noise is obtained.

[0070] In step 4.2, the time domain expression of the normalized Hanning window length is:

[0071]

[0072] The frequency domain expression is:

[0073]

[0074] In order to prevent the windowed power spectrum or amplitude spectrum from being affected by the window function, the restoration function must be derived according to certain principles. The restoration coefficient calculation formula is:

[0075] K t =A / A'A is the amplitude of the harmonic signal; A' is the amplitude at the peak frequency after windowing.

[0076] The window function can be directly analyzed for spectrum, A' makes the amplitude spectrum of the window function at zero frequency, and A is equal to 1. According to this method, the amplitude recovery coefficient of the Hanning window is calculated to be 2, so the power equal recovery coefficient is 1.633.

[0077] In step 4.3, the definition of sound pressure level is as follows:

[0078]

[0079] Where: L p is the sound pressure level, dB; p0 = 2 × 10 －5 Pa is the reference sound pressure, and the sound pressure level of the reference sound pressure itself is 0dB; p e is the effective sound pressure.

[0080] In step 4.3, the frequency extraction range of the noise signal is 20Hz-20kHz, and each frequency band of the octave has a corresponding upper limit frequency and a lower limit frequency ratio set to 2 1 / 3 ,Right now:

[0081]

[0082] where f u is the upper frequency limit, f d is the lower frequency limit.

[0083] In step 4.4, the calculation formula of C weight is:

[0084]

[0085] Further, step 5 includes the following steps:

[0086] Step 5.1, normalize the historical data to prevent the attributes with larger magnitude from dominating due to the difference in data magnitude and alleviate the slowdown of iterative convergence caused by the difference in magnitude;

[0087] Step 5.2, divide the normalized data into training set data and test set data;

[0088] Step 5.3: Add noise to the input data x to get a new input Generate a stacked denoising autoencoder;

[0089] Step 5.4, determine the number of hidden layers, the number of neurons in each layer, and the activation function of the initial SSDAE network;

[0090] Step 5.5, adding constraints to the loss function of the stacked denoising autoencoder to generate a deep sparse denoising autoencoder, i.e., a stacked sparse denoising autoencoder network;

[0091] Step 5.6, using an optimizer to optimize the parameters of the stacked sparse denoising autoencoder network;

[0092] Step 5.7, save the optimal training model (i.e., early warning model) and network parameters.

[0093] Step 5.8, repeat steps 5.1-5.7, and train and save the early warning model using process parameters, fan vibration, and noise as inputs respectively.

[0094] In the step 5.1, the normalization process adopts the MinMax normalization process method.

[0095] The step 5.2 is specifically as follows: the normalized historical data is divided into multiple groups of data of the same length to form a matrix as all the data of the training set and the test set, and 3 / 4 of the groups of data in the matrix are used as the training set and the remaining 1 / 4 of the groups of data are used as the test set using the holdout method, and the training set and the test set data are randomly arranged.

[0096] In step 5.3, the method of adding noise to the data is:

[0097] Step 5.3.1, generate uniformly distributed noise noise with the same dimension as x;

[0098] Step 5.3.2, take the Hadamard product of x and noise, the specific formula is:

[0099]

[0100] Optionally, the noise may be increased by using Salt-and-Pepper Noise, Gaussian Noise, Masking Noise, or White Noise.

[0101] In step 5.4, the output layer uses the Sigmoid activation function, and the remaining layers use the Leaky Relu activation function.

[0102] In step 5.5, the optimization goal of the stacked sparse denoising autoencoder is:

[0103]

[0104]

[0105] In the formula, is the activation value of the hidden layer neuron; β is the weight of the sparsity penalty factor; x is the preprocessed running data, is the running data after adding noise; are respectively ρ and The relative entropy between two Bernoulli random variables with the same mean is hour, KL increases with ρ and The difference between them increases monotonically, and when When it is close to 0 or 1, the relative entropy becomes infinite, so minimizing the penalty factor can make Approaching ρ.

[0106] In step 5.6, the optimizer adopts the Adam optimization algorithm.

[0107] Further, step 7 includes the following steps:

[0108] Step 7.1, calculate the observed values ​​X respectively obs and the reconstruction value X rec Euclidean distance and cosine distance between them;

[0109] Step 7.2, calculate the model similarity based on Euclidean distance and cosine distance.

[0110] In step 7.1, the calculation formula of the Euclidean distance is:

[0111]

[0112] The calculation formula for cosine distance is:

[0113]

[0114] Here, dot(·) represents the inner product operation, and ||·|| represents the 2-norm of the vector.

[0115] In step 7.2, the calculation formula of model similarity is:

[0116]

[0117] In the formula, sim(X, Y) represents the similarity between vectors X and Y. The cosine distance value constructed in the present invention is between 0 and 1, and the closer it is to 0, the lower the morphological similarity between the reconstructed value and the input value.

[0118] In step 8, the Bayesian maximum a posteriori probability estimation calculation formula is:

[0119]

[0120] Among them, x f Indicates the fan status indicator, z i represents the similarity index of the i-th information source, σ i Represents the standard deviation of the ith information source.

[0121] The warning threshold is 0.8. f When the temperature falls below this threshold, the system issues a warning of abnormal fan status.

[0122] Compared with the prior art, the present invention has the following advantages:

[0123] 1. The present invention adopts advanced sound pressure sensors and acceleration sensors to achieve high-precision noise and vibration signal collection, making up for the shortcomings of insufficient detection means of current power plant equipment.

[0124] Second, the present invention adopts a vibration signal feature extraction method based on wavelet packet decomposition, which can not only dynamically adjust the time resolution by frequency, but also obtain high-frequency signals while extracting low-frequency signals. It effectively makes up for the shortcomings of traditional signal analysis methods, such as short-time Fourier transform has only fixed resolution in different frequency domains; empirical mode decomposition lacks theoretical proof; wavelet transform only extracts low-frequency information, while high-frequency detail information is easily lost during the analysis process. The 1 / 3 octave noise signal feature extraction method adopted successfully avoids spectrum leakage and energy leakage caused by signal truncation by performing windowed fast Fourier transform on the noise signal, and uses C weighting algorithm to correct the frequency band sound pressure level to restore the auditory characteristics of the human ear to sound. At the same time, by deploying the acquisition front end at the edge of the device for real-time calculation, on the one hand, the data transmission volume is reduced and the network pressure is relieved; on the other hand, moving the calculation process forward can effectively alleviate the calculation pressure of the host computer.

[0125] 3. The present invention selects generator power, motor current, motor winding temperature, drive end bearing temperature, free end bearing temperature, hydraulic oil pressure, hydraulic oil temperature, lubricating oil pressure, lubricating oil temperature, fan blade opening, inlet wind temperature, outlet wind temperature and outlet wind pressure as fan status monitoring parameters. Since these monitoring parameters are field operation parameters that are more prone to abnormalities in the fan monitoring system, the operation of the fan can be well judged by monitoring these parameters. For example, the failure of the fan bearing or the deterioration of the lubricating oil quality will cause the fan bearing temperature and oil temperature to rise; the wear of the blades, the misalignment of the rotor or the rotational stall will cause the fan to vibrate abnormally. Therefore, it is very necessary and effective to select the above monitoring parameters to monitor the fan.

[0126] 4. The present invention adopts a stacked sparse denoising autoencoder network algorithm. Compared with existing machine learning or deep learning models, the use of a deep sparse denoising autoencoder can effectively improve the generalization ability of the model; since a part of the original training set data is destroyed, the present invention can alleviate the generation gap between the training set and the test set to a certain extent, so it is closer to the test data to a certain extent, thereby increasing the robustness of the model.

[0127] 5. The multi-source information fusion method adopted by the present invention makes full use of multiple information sources, fuses the output similarity of multiple models, and improves the reliability of the overall evaluation of the system. The above shows and describes the basic principles, main features and advantages of the present invention. Those skilled in the art should understand that the present invention is not limited to the above-mentioned embodiments. The above-mentioned embodiments and the specification only describe the principles of the present invention. The present invention will have various changes and improvements without departing from the spirit and scope of the present invention. These changes and improvements all fall within the scope of the present invention to be protected. The scope of protection claimed by the present invention is defined by the attached claims and their equivalents.

Claims

1. A power station fan status early warning method, specifically comprising the following steps: Step 1: Collect vibration and noise signals of normal operation of the fan; Step 2: Collecting a time series of fan process parameters; the time series of fan process parameters includes fan operation data and a fan operation state corresponding to the fan operation data, wherein the fan operation state includes a normal operation state and a fault state; Step 3: Using a wavelet packet decomposition method, extract and construct a wavelet energy spectrum from the vibration data time series collected under normal operating conditions of the fan; Step 4: Use the 1 / 3 octave decomposition method to extract and construct the frequency band energy spectrum from the noise data time series collected under the normal operation state of the fan; Step 4.1, performing fast Fourier transform on the collected noise signal; Step 4.2, add the Hanning window according to the power equal restoration coefficient to obtain the preliminary spectrum of the noise; the time domain expression of the Hanning window after normalization is: The frequency domain expression is: In order to prevent the windowed power spectrum or amplitude spectrum from being affected by the window function, the restoration function must be derived. The restoration coefficient calculation formula is: K t =A / A'A is the amplitude of the harmonic signal; A' is the amplitude at the peak frequency after windowing; Directly perform spectrum analysis on the window function, A' makes the amplitude spectrum of the window function at zero frequency, and A is equal to 1 at this time; according to this method, the amplitude recovery coefficient of the Hanning window is calculated to be 2, so the power equal recovery coefficient is 1.633; Step 4.3, extract the nominal value of the center frequency and the 1 / 3 octave range, and use the vibration level method to convert the graph into an unweighted 1 / 3 octave sound pressure level spectrum diagram; The sound pressure level is defined as follows: Where: L p is the sound pressure level, dB; p0 = 2 × 10 －5 Pa is the reference sound pressure, and the sound pressure level of the reference sound pressure itself is 0dB; p e is the effective sound pressure; Step 4.4, after C-weighted sound level correction, the C-weighted 1 / 3 octave band sound pressure level spectrum of the noise is obtained. The calculation formula of C-weighting is: Step 5: respectively taking the process parameters, vibration, and noise energy spectrum data as input, and constructing a stacked sparse denoising autoencoder network for model training to obtain early warning models of process parameters, fan vibration, and noise; Step 6: Obtain process parameters, vibration, and noise energy spectrum data X during the time period to be evaluated obs , respectively input the early warning model of process parameters, fan vibration and noise, and obtain the model reconstruction value X rec ; Step 7: Construct the model based on the observed values ​​X obs and the reconstruction value X rec Similarity indicators of Euclidean distance and cosine distance between them; Step 8: Use the Bayesian maximum a posteriori probability estimation method to fuse the similarity indicators of the single warning models of the above process parameters, fan vibration, and noise, construct a fan status warning indicator based on multi-source information fusion, set the status warning threshold, and determine whether a fault occurs based on the fault warning threshold.

2. A power station fan status early warning method according to claim 1, characterized in that: The step 3 specifically includes the following steps: Step 3.1, perform spectrum analysis on the collected vibration signal to determine the appropriate frequency band for wavelet packet decomposition; Step 3.2, determine the number of layers of wavelet packet decomposition. When performing wavelet packet decomposition, the number of decomposition layers has a significant correlation with the analysis accuracy of time-frequency domain signals; Step 3.3, decompose and reconstruct the signal using wavelet packets, and calculate the energy of each sub-band signal to obtain the reconstructed signal energy spectrum; Step 3.4, divide the reconstructed signal energy spectrum into adjacent and equal frequency bands, and calculate the normalized signal energy ratio of each frequency band.

3. A power station fan status early warning method according to claim 1, characterized in that: The step 5 specifically comprises the following steps: Step 5.1, normalize the historical data to prevent the attributes with larger magnitude from dominating due to the difference in data magnitude and alleviate the slowdown of iterative convergence caused by the difference in magnitude; Step 5.2, divide the normalized data into training set data and test set data; Step 5.3: Add noise to the input data x to get a new input Generate a stacked denoising autoencoder; Step 5.4, determine the number of hidden layers, the number of neurons in each layer, and the activation function of the initial SSDAE network; Step 5.5, adding constraints to the loss function of the stacked denoising autoencoder to generate a deep sparse denoising autoencoder, i.e., a stacked sparse denoising autoencoder network; Step 5.6, using an optimizer to optimize the parameters of the stacked sparse denoising autoencoder network; Step 5.7, save the optimal training model, that is, the early warning model and network parameters; Step 5.8, repeat steps 5.1-5.7, and train and save the early warning model using process parameters, fan vibration, and noise as inputs respectively.

4. A power station fan status early warning method according to claim 3, characterized in that: In step 5.1, the normalization process adopts the MinMax normalization process method; the specific formula is: Among them, d i is the original historical data, Max and Min are the maximum and minimum values ​​in the historical data respectively, and x is d i The normalized result of The step 5.2 is specifically as follows: the normalized historical data is divided into multiple groups of data of the same length to form a matrix as all the data of the training set and the test set, and 3 / 4 of the groups of data in the matrix are used as the training set by the holdout method, and the remaining 1 / 4 of the groups of data are used as the test set, and the training set and the test set data are randomly arranged; In step 5.3, the method of adding noise to the data is: Step 5.3.1, generate uniformly distributed noise noise with the same dimension as x; Step 5.3.2, take the Hadamard product of x and noise, the specific formula is: In step 5.4, the output layer uses the Sigmoid activation function, and the remaining layers use the Leaky Relu activation function; In step 5.5, the optimization goal of the stacked sparse denoising autoencoder is: In the formula, is the activation value of the hidden layer neuron; β is the weight of the sparsity penalty factor; x is the preprocessed running data, is the running data after adding noise; are respectively ρ and The relative entropy between two Bernoulli random variables with the same mean is hour, KL increases with ρ and The difference between them increases monotonically, and when When it is close to 0 or 1, the relative entropy becomes infinite, so minimize the penalty factor to make Approaching ρ; In step 5.6, the optimizer adopts the Adam optimization algorithm.

5. A power station fan status early warning method according to claim 1, characterized in that: The step 7 specifically includes the following steps: Step 7.1, calculate the observed values ​​X respectively obs and the reconstruction value X rec Euclidean distance and cosine distance between them; Step 7.2, calculate the model similarity based on Euclidean distance and cosine distance.

6. A power station fan status early warning method according to claim 1, characterized in that: In step 8, the Bayesian maximum a posteriori probability estimation calculation formula is: in, x f Indicates the fan status indicator, z i represents the similarity index of the i-th information source, σ i represents the standard deviation of the i-th information source; the warning threshold is 0.8, when the wind turbine status index x f When the temperature falls below this threshold, the system issues a warning of abnormal fan status.

7. A method for monitoring the status of a fan of a thermal power unit, characterized by: It includes the power station fan status early warning method as described in any one of claims 1-5.

Images