An intelligent diagnosis system and method for state of a thermal power plant

Abstract

The application discloses a thermal power plant state intelligent diagnosis system and method, and relates to the technical field of acoustic diagnosis; the method comprises the following steps: collecting original acoustic signals of a boiler in real time, and pre-acquiring micro-leakage acoustic samples, coking rupture acoustic samples and boiler background noise samples; through frequency domain power spectrum density analysis, frequency domain features of the micro-leakage acoustic samples, the coking rupture acoustic samples and the boiler background noise samples are extracted; based on the frequency domain features, a fault feature frequency band is determined, and a band-pass filter and a notch filter are set; through the band-pass filter and the notch filter, noise reduction processing is performed on the original acoustic signals, and the noise-reduced acoustic signals are obtained; the application can timely issue an early warning, avoid problems such as boiler heating surface tube burning, heat efficiency reduction and unplanned shutdown caused by micro-leakage and coking rupture fault missing detection of the boiler, and guarantee safe and efficient operation of a generator set.

Description

Technical Field

[0001] This invention relates to the field of acoustic diagnostic technology, and more specifically, to an intelligent diagnostic system and method for the condition of thermal power generation equipment. Background Technology

[0002] In thermal power generation systems, boiler micro-leakage and early coking are two typical potential faults. If they are not detected and addressed in a timely manner, they can easily lead to more serious equipment damage. To achieve early warning of these two types of faults, the industry generally adopts acoustic sensing monitoring technology. This involves placing microphones or acoustic emission sensors in the boiler body and surrounding key areas to collect various acoustic signals generated during boiler operation, thereby indirectly reflecting the internal operating conditions of the boiler and providing data support for fault diagnosis.

[0003] The core principle of existing boiler acoustic monitoring technology is as follows: First, acoustic sensors are used to convert acoustic signals such as mechanical vibration, fluid flow, and fault-related phenomena during boiler operation into electrical signals. Then, a signal preprocessing module is used to reduce noise in the original electrical signals. This process typically employs a bandpass filtering algorithm to retain only acoustic signal components within a preset frequency range, or uses conventional wavelet analysis to decompose the signal and extract energy characteristic parameters. Next, the preprocessed signal characteristics are compared with a pre-set fault judgment threshold. If the signal characteristic parameters exceed the threshold range, it is determined that the boiler may have a fault, and the corresponding alarm mechanism is triggered to prompt maintenance personnel to conduct an inspection.

[0004] However, in the actual operating environment of thermal power generating units, large auxiliary equipment such as primary air fans and coal mills are often installed around the boiler. These auxiliary equipment generate strong-amplitude periodic mechanical noise during operation. Simultaneously, boiler operation is accompanied by ambient wind noise and broadband noise generated by pipeline vibration, forming a complex multi-source strong background noise field. In contrast, the jet noise generated by boiler micro-leakage faults and the acoustic signals generated by early coking layer ruptures have inherently low energy amplitudes and exhibit significant non-stationary characteristics, resulting in substantial overlap in frequency distribution between these target fault signals and background noise. Existing bandpass filtering algorithms can only filter out noise components outside a preset frequency range and cannot effectively separate background noise that overlaps with the target fault signal frequency. Simple wavelet analysis methods can only extract and characterize the overall signal energy, making it difficult to distinguish between the target fault contribution and the background noise contribution. Ultimately, the target fault signal is completely masked by strong background noise and cannot be effectively identified by existing monitoring systems.

[0005] The aforementioned signal obstruction issues directly reduce the detection rate of micro-leakage and early coking faults in existing boiler acoustic monitoring technologies, making it prone to missed fault detection. If a boiler micro-leakage fault is not detected in time, the leak point will gradually expand over time, eventually causing the boiler heating surface tubes to burn out and rupture, leading to a boiler steam-water system leakage accident. If early coking faults are not addressed for a long time, the coking layer will continue to thicken on the boiler heating surface, not only hindering heat transfer inside the boiler and causing a significant decrease in boiler thermal efficiency and increased coal consumption for power generation, but also, in severe cases, causing boiler furnace flameout due to coking layer detachment, forcing the generator unit to shut down unplanned for coking removal or equipment replacement. This not only causes huge direct economic losses, but also seriously affects the continuous and stable operation of thermal power generating units, threatening the reliability of power supply in the power system.

[0006] In view of this, the present invention proposes an intelligent diagnostic system and method for the condition of thermal power generation equipment to solve the above problems. Summary of the Invention

[0007] To overcome the aforementioned deficiencies of the prior art and achieve the above objectives, the present invention provides the following technical solution: a method for intelligent diagnosis of the status of thermal power generation equipment, comprising:

[0008] The original acoustic signals of the boiler are acquired in real time, and micro-leakage acoustic samples, coking and cracking acoustic samples and boiler background noise samples are obtained in advance.

[0009] Frequency domain power spectral density analysis was used to extract the frequency domain features of micro-leakage acoustic samples, coking and cracking acoustic samples, and boiler background noise samples; based on the frequency domain features, the fault characteristic frequency bands were determined, and bandpass filters and notch filters were set.

[0010] The original acoustic signal is denoised by using a bandpass filter and a notch filter to obtain the denoised acoustic signal.

[0011] Based on the preset fault judgment threshold, the denoised acoustic signal is marked as a suspected fault signal, and the suspected fault signal is marked as a candidate fault signal.

[0012] The feature vector of the candidate fault signal is obtained by using the corresponding time-frequency transformation processing method;

[0013] Based on the pre-built fault identification model, the feature vectors are comprehensively judged to obtain the fault determination result.

[0014] Furthermore, methods for obtaining the feature vectors of candidate fault signals include:

[0015] The feature vectors of candidate fault signals include instantaneous energy and spectral kurtosis;

[0016] The instantaneous energy is calculated as follows: if it is calculated based on the time-frequency spectrum matrix of the candidate fault signal, the element value in the column vector set of the fault feature frequency band is the power spectral density at the corresponding time-frequency point, thus obtaining the instantaneous energy at the corresponding time point; if it is calculated based on the wavelet coefficient matrix of the candidate fault signal, the wavelet coefficients are first squared, and then the squared coefficients in the column vector set of the corresponding scale of the fault feature frequency band at the corresponding time point are summed to obtain the instantaneous energy at the corresponding time point.

[0017] The spectral kurtosis is calculated as follows: the time-frequency spectrum matrix or wavelet coefficient matrix of the candidate fault signal is divided into several local time windows according to the time dimension; first, the second central moment of the corresponding element of the fault characteristic frequency band in each local time window is calculated; then, the fourth central moment of the corresponding element of the fault characteristic frequency band in each local time window is calculated; spectral kurtosis = fourth central moment / (square of second central moment) - 3.

[0018] Furthermore, the methods for obtaining the time-frequency spectrum matrix and wavelet coefficient matrix of the candidate fault signal include:

[0019] When the time-domain duration of the candidate fault signal is greater than the preset duration threshold, and the width of the fluctuation range of the frequency components of the candidate fault signal within the fault characteristic frequency band is less than the product of the width of the fault characteristic frequency band and the preset fluctuation ratio coefficient, a short-time Fourier transform is used; otherwise, a continuous wavelet transform is used.

[0020] If a short-time Fourier transform is used, the candidate fault signal is divided into several frame signals according to the set window length and frame shift. A Fourier transform is performed on each frame signal to obtain the complex spectrum of each frame signal in the frequency domain. The power spectral density of each frame signal is obtained by taking the modulus and squaring the complex spectrum. The power spectral densities of all frame signals are arranged in the order of the frame signals and the order of frequency from low to high to form a two-dimensional matrix, which is the time-frequency spectrum matrix. The rows of the time-frequency spectrum matrix correspond to the time dimension, and the columns correspond to the frequency dimension.

[0021] If continuous wavelet transform is used, the convolution integral of the candidate fault signal with the preset wavelet basis function at multiple set scales is calculated to obtain the wavelet coefficient sequence at the corresponding scale. The wavelet coefficient sequences of all scales are arranged in order from the smallest scale to the largest scale and in the sampling time order of the candidate fault signal to form a two-dimensional matrix, which is the wavelet coefficient matrix of the candidate fault signal at different scales. The rows of the wavelet coefficient matrix correspond to the time dimension and the columns correspond to the scale dimension.

[0022] Furthermore, the method for marking suspected fault signals as candidate fault signals includes:

[0023] The suspected fault signal is preprocessed to remove the DC component and residual low-frequency interference, resulting in the preprocessed suspected fault signal.

[0024] For the pre-processed suspected fault signals, calculate the cross-correlation function between the suspected fault signals output by any two acoustic sensors, obtain the cross-correlation function curve through cross-correlation operation, and extract the peak position and peak amplitude of the suspected fault signal from the cross-correlation function curve.

[0025] If the time difference corresponding to the peak position of the suspected fault signal falls within the preset theoretical propagation time difference range, and the peak amplitude of the suspected fault signal is greater than the preset coherence threshold, then the corresponding suspected fault signal will be marked as a candidate fault signal; otherwise, the corresponding suspected fault signal will be remarked as a normal signal.

[0026] Furthermore, methods for marking suspected fault signals in the denoised acoustic signal include:

[0027] Fault determination thresholds include the micro-leakage characteristic energy determination threshold and the coking rupture pulse amplitude determination threshold;

[0028] Calculate the mean power spectral density of the noise-reduced acoustic signal within the fault characteristic frequency band;

[0029] If the mean power spectral density is greater than the threshold for determining microleakage characteristic energy, the corresponding acoustic signal is marked as a suspected microleakage signal; otherwise, the corresponding acoustic signal is marked as a normal signal.

[0030] The instantaneous amplitude of the denoised acoustic signal is obtained by traversing the denoised acoustic signal through a sliding window. The number of pulses with an instantaneous amplitude greater than the threshold for judging the amplitude of coking and cracking pulses is counted. If the number of pulses is greater than the preset counting threshold, the corresponding acoustic signal is marked as a suspected coking and cracking signal; otherwise, the corresponding acoustic signal is marked as a normal signal.

[0031] Acoustic signals marked as suspected micro-leakage signals and suspected coking and cracking signals in the noise-reduced acoustic signals were selected as suspected fault signals.

[0032] Furthermore, methods for determining the fault characteristic frequency band include:

[0033] The maximum power spectral density of the micro-leak acoustic sample is denoted as the first peak power, the maximum power spectral density of the coking and cracking acoustic sample is denoted as the second peak power, and the maximum power spectral density of the boiler background noise sample is denoted as the third peak power.

[0034] Based on the power spectral density value and the maximum power spectral density value of each type of data, the energy proportion concentration area of ​​each type of data is determined, and the first frequency range, the second frequency range and the third frequency range are obtained.

[0035] Based on the average energy proportion of each type of data in the first frequency range, the second frequency range, and the third frequency range, the energy concentration range of the fault signal is calculated.

[0036] Configure a digital bandpass filter as a filtering component, set the low cutoff frequency of the digital bandpass filter as the lower limit of the fault signal energy concentration range, and set the high cutoff frequency as the upper limit of the fault signal energy concentration range; calculate the highest frequency value of the boiler background noise sample in the third frequency range, and the low cutoff frequency of the digital bandpass filter is greater than the highest frequency value in the third frequency range.

[0037] The passband frequency band corresponding to the digital bandpass filter is the fault characteristic frequency band.

[0038] Furthermore, the methods for obtaining the first frequency range, the second frequency range, and the third frequency range include:

[0039] In the micro-leak acoustic sample, the continuous frequency range formed by the frequency points whose power spectral density value is greater than or equal to the first peak power preset first proportion is the first frequency range, and the energy proportion corresponding to the first frequency range is greater than the preset fault energy proportion threshold.

[0040] In the acoustic sample of coking and cracking, the continuous frequency range formed by the frequency points whose power spectral density value is greater than or equal to the second peak power preset second ratio is the second frequency range, and the energy ratio corresponding to the second frequency range is greater than the fault energy ratio threshold.

[0041] In the boiler background noise sample, the continuous frequency range formed by the frequency points whose power spectral density values ​​are greater than or equal to the third peak power preset third proportion is the third frequency range, and the energy proportion corresponding to the third frequency range is greater than the preset background energy proportion threshold.

[0042] The upper limit of the third frequency range is less than the lower limit of the first frequency range and the lower limit of the second frequency range, and there is a continuous frequency segment that partially overlaps with the first frequency range and the second frequency range. The frequency points in the overlapping continuous frequency segment belong to both the first frequency range and the second frequency range.

[0043] Furthermore, methods for obtaining the energy concentration range of the fault signal include:

[0044] The overlapping area between the first frequency range and the second frequency range is taken as the initial interval. The average energy ratio of micro-leakage acoustic samples and coking cracking acoustic samples in the initial interval, as well as the average energy ratio of boiler background noise samples in the initial interval are calculated.

[0045] If the average energy proportion of both types of fault samples in the initial interval is greater than the preset fault energy proportion threshold, and the average energy proportion of boiler background noise samples is less than the preset background energy proportion threshold, then the initial interval is determined as the fault signal energy concentration interval; otherwise, the upper and lower limits of the initial interval are adjusted, and the average energy proportion of the three types of data in the adjusted interval is recalculated until the conditions of the average energy proportion of both types of fault samples and the average energy proportion of background noise are met, and the fault signal energy concentration interval is finally determined.

[0046] Furthermore, methods for obtaining fault determination results include:

[0047] The mean value of the instantaneous energy and the spectral kurtosis value are used to form a feature vector;

[0048] The feature vector is input into the fault identification model, and the output includes probability values ​​for three types of results: normal, micro-leakage fault, and coking cracking fault.

[0049] If the probability value of a microleak fault is greater than the preset microleak confidence threshold, then a microleak fault is determined to have occurred.

[0050] If the probability value of coking failure is greater than the preset coking confidence threshold, then coking failure is determined to have occurred.

[0051] If the probability values ​​of both micro-leakage fault and coking cracking fault are less than the corresponding confidence threshold, then it is determined that no fault has occurred.

[0052] A smart diagnostic system for the condition of thermal power generation equipment, comprising:

[0053] The data acquisition module is used to acquire the original acoustic signals of the boiler in real time, and to pre-acquire micro-leakage acoustic samples, coking and cracking acoustic samples, and boiler background noise samples.

[0054] The feature setting module is used to extract the frequency domain features of micro-leakage acoustic samples, coking and cracking acoustic samples, and boiler background noise samples through frequency domain power spectral density analysis; determine the fault characteristic frequency band based on the frequency domain features, and set the bandpass filter and notch filter.

[0055] The noise reduction processing module is used to perform noise reduction processing on the original acoustic signal through a bandpass filter and a notch filter to obtain a noise-reduced acoustic signal;

[0056] The signal marking module is used to mark suspected fault signals on the noise-reduced acoustic signals according to a preset fault judgment threshold, and to mark the suspected fault signals as candidate fault signals.

[0057] The feature vector module is used to obtain the feature vector of the candidate fault signal by employing the corresponding time-frequency transformation processing method.

[0058] The fault determination module, based on a pre-built fault identification model, performs comprehensive discrimination on feature vectors to obtain fault determination results.

[0059] Compared with the prior art, the technical effects and advantages of the intelligent diagnostic system and method for thermal power generation equipment of the present invention are as follows:

[0060] This invention acquires the raw acoustic signals of a boiler in real time and pre-obtains acoustic samples of micro-leakage, coking and cracking, and boiler background noise. Frequency domain features of these three types of samples are extracted through power spectral density analysis to determine the fault characteristic frequency band. Corresponding bandpass and notch filters are configured, and these filters are used to perform noise reduction processing on the raw acoustic signals to obtain denoised acoustic signals. Suspected fault signals are marked on the denoised acoustic signals according to a preset fault judgment threshold. Then, sensor array correlation analysis is used to screen the suspected fault signals and mark candidate fault signals. Based on the time-domain duration and frequency component variation characteristics of the candidate fault signals, a selection is made. Short-time Fourier transform or continuous wavelet transform is used to obtain the time-frequency spectrum matrix or wavelet coefficient matrix. Instantaneous energy, spectral kurtosis and other characteristic parameters are calculated and feature vectors are constructed. Finally, the feature vectors are input into a pre-constructed one-dimensional convolutional neural network fault recognition model. The model outputs probability values ​​for three types of results: normal, micro-leakage fault, and coking and cracking fault. The fault determination result is obtained by comparing the probability value with the preset signal threshold. At the same time, the timestamp of the fault determination and the fault characteristic frequency band are recorded. The fault location is determined by combining the time difference of sensor correlation analysis. After accumulating the analysis results of multiple frames to reduce false alarms, alarm information is generated and a warning is issued. The acoustic feature data related to the fault is also saved for maintenance personnel to further diagnose.

[0061] This invention solves the problem in existing boiler acoustic monitoring technologies where the inability to effectively separate strong background noise overlapping with fault signal frequencies leads to the masking of micro-leakage and early coking fault signals, resulting in a high rate of missed detection. It can specifically suppress strong periodic mechanical noise and low-frequency noise generated by auxiliary equipment such as primary air fans and coal mills, effectively preserving signals within the fault characteristic frequency band and improving the signal-to-noise ratio. Through a two-stage signal screening process from suspected to candidate faults, it reduces the impact of interference signals. Combined with time-frequency transformation, it accurately extracts fault features, and then, through a trained fault identification model, comprehensively judges and accurately identifies micro-leakage and coking cracking faults. Accumulating multi-frame analysis results reduces the false alarm rate, and timely early warnings can prevent problems such as boiler heating surface tube burnout, decreased thermal efficiency, and unplanned shutdowns caused by missed detections, ensuring the safe and efficient operation of the generator unit. Attached Figure Description

[0062] Figure 1 This is a schematic diagram of an intelligent diagnostic system for the status of thermal power generation equipment according to an embodiment of the present invention;

[0063] Figure 2This is a flowchart of an intelligent diagnostic method for the status of thermal power generation equipment according to an embodiment of the present invention;

[0064] Figure 3 This is a flowchart illustrating the method for obtaining the time-frequency spectrum matrix and wavelet coefficient matrix of candidate fault signals according to an embodiment of the present invention.

[0065] Figure 4 This is a flowchart of a method for obtaining fault determination results according to an embodiment of the present invention. Detailed Implementation

[0066] The technical solutions of the embodiments of the present invention will be described in detail, clearly, and completely below with reference to the accompanying drawings. It should be particularly noted that the specific embodiments described below are only for better illustrating and explaining the technical solutions of the present invention, and are intended to enable those skilled in the art to better understand and implement the present invention, and should not be construed as limiting the scope of protection of the present invention. Without departing from the spirit and substance of the present invention, those skilled in the art can modify, adjust, or make equivalent substitutions based on the content disclosed in the present invention, and these should all be considered within the scope of protection of the present invention.

[0067] Example 1:

[0068] Please see Figure 1 As shown in the figure, this embodiment discloses an intelligent diagnostic system for the status of thermal power generation equipment, including a data acquisition module, a feature setting module, a noise reduction processing module, a signal marking module, a feature vector module, and a fault determination module. Each module is connected by wires and / or wirelessly to realize data transmission.

[0069] The data acquisition module is used to acquire the original acoustic signals of the boiler in real time, and to pre-acquire micro-leakage acoustic samples, coking and cracking acoustic samples, and boiler background noise samples.

[0070] The raw acoustic signal is acoustic data collected in real time by multiple microphones or acoustic emission sensors distributed around the boiler body. This includes mechanical vibration noise, fluid flow noise, and possible fault sound signals generated during boiler operation. The raw acoustic signal is obtained by continuously recording the sound signals during boiler operation using high-sampling-rate data acquisition equipment.

[0071] Boiler background noise samples are environmental acoustic data samples of the boiler and its auxiliary equipment under normal and fault-free operation. Examples include periodic mechanical noise generated by large auxiliary equipment such as primary air fans and coal mills, as well as broadband noise generated by wind noise and pipe vibration within the furnace. Environmental sounds are recorded for extended periods under normal boiler operating conditions, or sound signals are individually collected from specific noise sources such as fans to obtain sample data representing typical background noise.

[0072] Micro-leak acoustic samples are sample data of jet sound signals generated when a tiny leak occurs in the boiler's heating surface pipes. Micro-leak acoustic samples can be obtained through experiments or historical data. For example, during maintenance, a controllable micro-leak point can be artificially set up, and the resulting acoustic signal can be recorded using proximity sensors; or, acoustic signal fragments recorded by the monitoring system at the time of a past actual boiler tube micro-leak accident can be extracted. This data reflects the typical acoustic characteristics of micro-leak faults, such as high-frequency continuous noise components.

[0073] Acoustic samples of coke rupture refer to acoustic signal samples generated when the early coke layer on the boiler's heating surface detaches or ruptures. These samples can be obtained through simulation experiments, such as monitoring small-scale events of coke material falling during boiler operation and recording the corresponding sound signals; or by using acoustic emission sensors to capture transient acoustic signals when the actual coke layer cracks and detaches within the furnace. This data reflects the acoustic characteristics of coke layer fracture, such as short-duration, sudden impact sounds and pulse ringing.

[0074] The feature setting module is used to extract the frequency domain features of micro-leakage acoustic samples, coking and cracking acoustic samples and boiler background noise samples through frequency domain power spectral density analysis; determine the fault characteristic frequency band based on the frequency domain features, and set bandpass filter and notch filter.

[0075] Frequency domain features of micro-leakage acoustic samples, coke cracking acoustic samples, and boiler background noise samples were extracted using frequency domain power spectral density analysis. Specific methods included: using frequency domain power spectral density analysis to obtain the center frequency, bandwidth, and power spectral peak of the micro-leakage acoustic samples in the high-frequency band, determining the specific characteristics of the micro-leakage acoustic samples exhibiting continuous broadband high-frequency components, such as a center frequency concentrated in the 2kHz–10kHz range, a bandwidth not less than 5kHz, and a power spectral peak stably maintained at more than three times the peak value of the background noise power spectral spectrum; and using frequency domain power spectral density analysis to obtain the dominant frequency component and energy decay rate of the coke cracking acoustic samples, determining the specific characteristics of the coke cracking acoustic samples exhibiting transient pulse signals. For example, the rise time is no more than 5ms, the fall time is no more than 8ms, the pulse amplitude is more than 5 times the maximum fluctuation amplitude of the background noise, and the main frequency components are also concentrated in the 2kHz to 10kHz range. Frequency domain power spectral density analysis is used to obtain the characteristic frequencies and energy distribution of the boiler background noise samples, and to determine the specific characteristics of the boiler background noise samples containing strong periodic components and broadband random noise. For example, the strong periodic components correspond to the primary air fan rotation fundamental frequency of 50Hz and harmonic 100Hz, the coal mill rotation fundamental frequency of 30Hz and harmonic 60Hz, and the broadband random noise is mainly distributed in the 200Hz to 1kHz range and the power spectral amplitude is less than 1 / 2 of the fault signal in the 2kHz to 10kHz range.

[0076] Based on the frequency domain characteristics of micro-leakage acoustic samples, coking cracking acoustic samples, and boiler background noise samples, the fault characteristic frequency bands and preprocessing parameters are determined. The specific process is as follows:

[0077] Frequency domain power spectral density analysis was performed on the micro-leakage acoustic samples, coking cracking acoustic samples, and boiler background noise samples. The specific method included: using Fast Fourier Transform to transform the micro-leakage acoustic samples, coking cracking acoustic samples, and boiler background noise samples from the time domain to the frequency domain, and calculating the power spectral density values ​​of each data type at different frequency points; first, the maximum power spectral density value of each data type was determined, with the maximum power spectral density value of the micro-leakage acoustic samples recorded as the first peak power, the maximum power spectral density value of the coking cracking acoustic samples recorded as the second peak power, and the maximum power spectral density value of the boiler background noise samples recorded as the third peak power; then, all frequency points in each data type whose power spectral density value was greater than or equal to a preset proportion of the corresponding peak power were statistically analyzed. This preset proportion was based on... Based on the requirement of significant fault characteristics, it is ensured that the selected frequency points can reflect the core energy distribution of each type of data. The continuous frequency intervals formed by all frequency points are respectively determined as the frequency range corresponding to the power spectrum peak of each type of data. The total power spectral density value of each type of data in the frequency range corresponding to its respective power spectrum peak is calculated, which is the sum of the power spectral density values ​​of all frequency points in the corresponding frequency interval. Then, the total power spectral density value of each type of data in the entire frequency range is calculated, which is the sum of the power spectral density values ​​of all frequency points. The ratio of the total power spectral density value in the frequency range corresponding to the power spectrum peak to the total power spectral density value in the entire frequency range is defined as the energy proportion. The energy proportion is used to quantify the proportion of energy in a certain frequency interval in the overall energy of the data.

[0078] In the micro-leaking acoustic sample, the continuous frequency range formed by frequency points with power spectral density values ​​greater than or equal to the first peak power at a preset first proportion is the first frequency range, and the energy proportion corresponding to the first frequency range is greater than the preset fault energy proportion threshold. The fault energy proportion threshold is set based on the fault signal identifiability requirements to ensure that the first frequency range can concentrate the core energy of the micro-leaking acoustic sample. The preset first proportion is set by determining the frequency coverage range corresponding to the power spectrum attenuation characteristics of the micro-leaking acoustic sample by statistically analyzing the frequency coverage range when the power spectrum of the micro-leaking acoustic sample drops from the first peak power to different proportions. Combined with the integrity requirements of the core frequency points for subsequent fault feature extraction, the proportion value that allows the selected frequency points to cover the main energy distribution area of ​​the micro-leaking acoustic sample while excluding low-energy interference frequency points on the edge is selected as the preset first proportion. In the acoustic sample of coking fracture, the continuous frequency range formed by the frequency points whose power spectral density value is greater than or equal to the second peak power preset second ratio is the second frequency range, and the energy ratio corresponding to the second frequency range is greater than the fault energy ratio threshold. The preset second ratio is set by determining it based on the pulse signal power spectrum characteristics of the acoustic sample of coking fracture. The coking fracture signal is a transient pulse signal, and its power spectrum peak has a strong concentration. By analyzing the energy distribution around the power spectrum peak of the acoustic sample of coking fracture, the ratio value that can accurately retain the main frequency component of the pulse signal while eliminating irrelevant clutter frequency points is selected as the preset second ratio to ensure that the second frequency range can reflect the core frequency domain characteristics of the coking fracture fault. In the boiler background noise sample, the continuous frequency range formed by frequency points with power spectral density values ​​greater than or equal to the third peak power at a preset third proportion is defined as the third frequency range. The energy proportion corresponding to the third frequency range is greater than a preset background energy proportion threshold, which is set based on the requirements for identifying the core frequency band of background noise, ensuring that the third frequency range concentrates the core energy of the boiler background noise sample. The preset third proportion is determined based on the degree of energy dispersion in the power spectrum of the boiler background noise sample. Since the energy distribution of boiler background noise is relatively dispersed, the cumulative energy proportion of its power spectrum at different proportions is statistically analyzed. For example, a proportion value that concentrates more than 80% of the main energy of the background noise at the selected frequency points and clearly distinguishes it from the fault sample frequency range is selected as the preset third proportion, providing a clear boundary for the subsequent division of the fault frequency band and the background frequency band. The upper limit of the third frequency range is less than the lower limit of the first frequency range and the lower limit of the second frequency range, and there are partially overlapping continuous frequency segments between the first and second frequency ranges. The frequency points within these overlapping continuous frequency segments belong to both the first and second frequency ranges.

[0079] The overlapping area between the first and second frequency ranges is taken as the initial interval. The average energy proportion of micro-leakage acoustic samples and coking cracking acoustic samples, as well as the average energy proportion of boiler background noise samples, are calculated within the initial interval. If the average energy proportion of both types of fault samples within the initial interval is greater than the preset fault energy proportion threshold, and the average energy proportion of boiler background noise samples is less than the preset background energy proportion threshold, then the initial interval is determined as the fault signal energy concentration interval. If not, the upper and lower limits of the initial interval are adjusted, with each adjustment being a preset frequency step size. The average energy proportion of the three types of data within the adjusted interval is recalculated until the conditions of the average energy proportion of both types of fault samples and the average energy proportion of background noise are met. Finally, the fault signal energy concentration interval is determined.

[0080] A digital bandpass filter is configured as the core filtering component. The low cutoff frequency of the digital bandpass filter is set as the lower limit of the fault signal energy concentration range, and the high cutoff frequency is set as the upper limit of the fault signal energy concentration range. The highest frequency value of the boiler background noise sample in the third frequency range is calculated. The low cutoff frequency of the digital bandpass filter is greater than the highest frequency value in the third frequency range to ensure that the high-frequency segment with prominent fault signal energy is retained after passing through the digital bandpass filter, while filtering out the low-frequency mechanical noise in the boiler background noise sample. At this time, the passband frequency band corresponding to the digital bandpass filter is the fault characteristic frequency band.

[0081] For the strong periodic components in the boiler background noise sample, frequency domain power spectral density analysis is first performed on the boiler background noise sample. The frequency domain analysis method is consistent with the one used to determine the energy concentration range of the fault signal. Frequency points with power spectral density values ​​significantly higher than the surrounding frequencies and exhibiting periodic repetition characteristics are identified. These frequency points are determined as the characteristic frequencies of the strong periodic components. Among them, the characteristic frequency corresponding to the periodic noise generated by the operation of the primary fan is recorded as the first periodic characteristic frequency, and the characteristic frequency corresponding to the periodic noise generated by the operation of the coal mill is recorded as the second periodic characteristic frequency.

[0082] Based on the identified characteristic frequencies, digital notch filters are configured as follows: Two sets of digital notch filters are configured. The center frequency of the first set of digital notch filters is set as the first period characteristic frequency, and its bandwidth is determined by calculating the half-width at half-maximum (WHM) of the power spectrum peak corresponding to the first period characteristic frequency. That is, it is determined by the frequency range corresponding to when the power spectral density value drops to half of the peak value. For example, the bandwidth value is set to 1.2-1.5 times the WHM to ensure that the frequency fluctuation range of the primary fan period noise is completely covered. The center frequency of the second set of digital notch filters is set as the second period characteristic frequency, and its bandwidth is also determined by calculating the half-width at half-maximum (WHM) of the power spectrum peak corresponding to the second period characteristic frequency. For example, the bandwidth value is set to 1.2-1.5 times the WHM to ensure that the frequency fluctuation range of the coal mill period noise is completely covered.

[0083] Two sets of digital notch filters are used to suppress the periodic noise interference generated by the primary air fan and the coal mill, respectively: the first set of digital notch filters attenuates the frequency components centered on the first periodic characteristic frequency in the boiler background noise sample, and the second set of digital notch filters attenuates the frequency components centered on the second periodic characteristic frequency, and finally the configured filter parameters are obtained.

[0084] Methods for setting fault determination thresholds include:

[0085] The fault determination thresholds include the micro-leakage characteristic energy determination threshold and the coking cracking pulse amplitude determination threshold. The mean power spectral density of the micro-leakage acoustic sample's frequency domain characteristics within the fault characteristic frequency band is calculated. Statistical analysis is used to obtain the characteristic difference coefficient between the mean power spectral density of the micro-leakage acoustic sample within the fault characteristic frequency band and the mean background noise. Combined with a preset false alarm rate under fault-free conditions, the micro-leakage characteristic energy determination threshold is determined. This threshold must ensure that the probability of the power spectral density value of the boiler background noise sample triggering the threshold within the fault characteristic frequency band does not exceed the false alarm rate under fault-free conditions, while also ensuring that the power spectral density values ​​of the micro-leakage acoustic sample within the corresponding frequency band can effectively trigger the micro-leakage characteristic energy determination threshold. For the coking cracking pulse amplitude determination threshold, statistical analysis is performed on the pulse amplitude in the time domain characteristics of the coking cracking acoustic sample. All effective pulse signals that can characterize the coking cracking fault are screened out, and transient interference pulses mixed in with the sample are excluded. To meet the preset fault detection rate requirements, such as a false detection rate of less than 0.05%, a threshold for judging the amplitude of coking cracking pulses is determined. This threshold must cover the amplitude of the vast majority of valid pulse signals in the acoustic sample of coking cracking. At the same time, it must be ensured that the amplitude of any occasional pulse interference signals in the boiler background noise sample is lower than the threshold for judging the amplitude of coking cracking pulses to avoid false alarms triggered by interference signals. The method for determining the false alarm rate under fault-free operating conditions is as follows: based on historical monitoring data of boilers in thermal power generating units operating without faults in the industry, the frequency of false triggering of the acoustic signal threshold within the historical fault-free period is statistically analyzed. Combined with the limit value of the false alarm rate of the boiler acoustic monitoring system in the industry standard, the false alarm rate under fault-free operating conditions is determined. The false alarm rate under fault-free operating conditions must meet the requirement that the number of times the background noise triggers the threshold does not exceed the preset number within the preset duration of continuous fault-free boiler operation.

[0086] The noise reduction module is used to perform noise reduction processing on the original acoustic signal through a bandpass filter and a notch filter to obtain the noise-reduced acoustic signal.

[0087] Continuous monitoring of acoustic signals during boiler operation. Multiple acoustic sensors synchronously acquire data based on the same time base signal. The arrangement of the acoustic sensors needs to cover different areas of the boiler body to ensure comprehensive signal acquisition, and the sampling rate of each sensor is kept consistent during the acquisition process. The raw acoustic signal is divided into frames according to a preset time window. The duration of the time window is determined based on the time domain characteristics of the fault signal to ensure that the time domain waveform of the fault signal is completely included. At the same time, a reasonable inter-frame overlap rate is set to avoid missing transient fault signals. After division, a frame of acoustic signal to be processed with a timestamp is obtained. The frame of acoustic signal to be processed contains mechanical vibration sound, fluid flow noise and possible fault sound signals generated during boiler operation.

[0088] Methods for performing noise reduction processing on the original acoustic signal using bandpass filters and notch filters to obtain the noise-reduced acoustic signal include:

[0089] Two pre-configured digital notch filters are used to filter the acoustic signal frame to be processed. The first set of digital notch filters attenuates the frequency components centered at the first periodic characteristic frequency in the acoustic signal frame, reducing the periodic noise generated by the primary fan operation. The second set of digital notch filters attenuates the frequency components centered at the second periodic characteristic frequency in the acoustic signal frame, reducing the periodic noise generated by the coal mill operation. After this processing, the strong periodic mechanical noise components in the acoustic signal frame to be processed are significantly suppressed.

[0090] The configured digital bandpass filter is invoked to filter the signal frame after periodic noise suppression. During the filtering process, the digital bandpass filter retains the components of the signal that are within the energy concentration range of the fault signal, and filters out all signal components that are below the low cutoff frequency or above the high cutoff frequency, thereby filtering out the low-frequency mechanical noise and other non-frequency noise in the boiler background noise sample.

[0091] After the above two steps of processing, the strong periodic noise, low-frequency mechanical noise and out-of-band noise in the acoustic signal frame to be processed are effectively suppressed, and only the signal components in the fault characteristic frequency band are retained. The section with prominent fault signal energy is retained, the signal-to-noise ratio is improved, and the foundation is laid for subsequent fault feature extraction.

[0092] The signal marking module is used to mark suspected fault signals on the noise-reduced acoustic signals according to a preset fault judgment threshold, and to mark the suspected fault signals as candidate fault signals.

[0093] The method for marking suspected fault signals in the denoised acoustic signal based on a preset fault determination threshold includes:

[0094] The mean power spectral density of the denoised acoustic signal within the fault characteristic frequency band is calculated. The mean power spectral density is compared with the micro-leakage characteristic energy judgment threshold. If the mean power spectral density is greater than the micro-leakage characteristic energy judgment threshold, the corresponding acoustic signal is marked as a suspected micro-leakage signal. If the mean power spectral density is less than or equal to the micro-leakage characteristic energy judgment threshold, the corresponding acoustic signal is marked as a normal signal.

[0095] Temporal pulse detection is performed on the denoised acoustic signal of the current frame. The instantaneous amplitude is obtained by traversing the signal through a sliding window. The number of pulses with instantaneous amplitude greater than the threshold for judging the amplitude of coking and cracking pulses is counted. If the number of pulses is greater than the preset counting threshold, the corresponding acoustic signal is marked as a suspected coking and cracking signal; if it is less than or equal to the counting threshold, the corresponding acoustic signal is marked as a normal signal.

[0096] Acoustic signals marked as suspected micro-leakage signals and suspected coking and cracking signals in the noise-reduced acoustic signals are selected as suspected fault signals, while acoustic signals marked as normal signals are not analyzed for the time being. This reduces the amount of data for in-depth analysis and improves the response speed of the overall fault identification process.

[0097] Methods for marking suspected fault signals as candidate fault signals include:

[0098] The suspected fault signals identified through rapid initial screening are preprocessed to remove DC components and residual low-frequency interference, resulting in preprocessed suspected fault signals. For these preprocessed signals, the cross-correlation function between the suspected fault signals output by any two acoustic sensors is calculated. The cross-correlation function curve is obtained through cross-correlation operations, and the peak position and peak amplitude are extracted from the curve. The peak position corresponds to the time difference between the two acoustic sensors receiving the same signal, and the peak amplitude reflects the coherence of the suspected fault signals output by the two acoustic sensors. Based on the preset installation positions of each acoustic sensor on the boiler body, the propagation distance difference from any possible fault point inside the boiler to each acoustic sensor is calculated. The theoretical propagation time difference range is obtained by calculating the ratio of the distance difference to the speed of sound propagating inside the boiler. The time difference corresponding to the peak position of the cross-correlation function curve is compared with the theoretical propagation time difference range, and the peak amplitude is compared with a preset coherence threshold. If the time difference corresponding to the peak position falls within the theoretical propagation time difference range and the peak amplitude is greater than the preset coherence threshold, the suspected fault signal is determined to originate from the same sound source inside the boiler, and the corresponding suspected fault signal is marked as a candidate fault signal. If the time difference corresponding to the peak position exceeds the theoretical propagation time difference range, or the peak amplitude is less than the preset coherence threshold, the suspected fault signal is determined to be environmental noise interference, and the corresponding suspected fault signal is remarked as a normal signal.

[0099] The feature vector module is used to obtain the feature vector of the candidate fault signal by using the corresponding time-frequency transformation processing method.

[0100] Methods for extracting feature vectors of candidate fault signals using appropriate time-frequency transformation processing include:

[0101] Please see Figure 3 As shown, the time-frequency spectrum matrix and wavelet coefficient matrix of the candidate fault signal are obtained through time-frequency transformation processing. The specific method includes: when the time-domain duration of the candidate fault signal is greater than a preset duration threshold, and the width of the fluctuation range of the frequency components of the candidate fault signal in the fault characteristic frequency band is less than the product of the width of the fault characteristic frequency band and the preset fluctuation ratio coefficient, a short-time Fourier transform is used; otherwise, a continuous wavelet transform is used; the fluctuation ratio coefficient is a constant less than 1; the duration threshold is set to an integer multiple of the period corresponding to the lowest frequency of the fault characteristic frequency band.

[0102] If a short-time Fourier transform is used, the window function is first selected based on the edge attenuation requirements and main lobe energy concentration requirements of the candidate fault signal: when the edge attenuation of the candidate fault signal meets the preset attenuation threshold, the Hanning window is selected. The edge attenuation rate of the Hanning window can make the length of the attenuation interval from the peak value to zero of the signal edge amplitude less than the preset attenuation interval length; the attenuation threshold is the ratio of the maximum amplitude of the candidate fault signal to the minimum amplitude of the edge; when the main lobe width of the candidate fault signal needs to meet the preset main lobe proportion, the Hamming window is selected. The main lobe energy proportion of the Hamming window can make the proportion of the total energy in the main lobe to the total energy of the entire window function greater than the preset energy proportion threshold; the main lobe proportion is the ratio of the main lobe width to the effective range of the entire frequency axis. The window length of the window function is the ratio of the sampling frequency to the preset frequency resolution, where the sampling frequency is the fixed sampling rate when the acoustic sensor acquires the candidate fault signal, and the sampling frequency satisfies the Nyquist sampling theorem; the inter-frame overlap rate is determined based on the characteristic continuity requirements of adjacent frame signals, and the inter-frame overlap rate = (window length - frame shift) / window length, where the frame shift is the starting time difference between two adjacent frame signals, and the frame shift is less than half of the window length, so as to ensure that the overlapping area of ​​adjacent frame signals in the time dimension can cover the minimum domain feature length of the candidate fault signal and avoid feature loss. The candidate fault signal is divided into several frame signals according to the window length and frame shift of the determined window function. Fourier transform is performed on each frame signal to obtain the complex spectrum of each frame signal in the frequency domain. The power spectral density of each frame signal is obtained by taking the modulus and squaring the complex spectrum. The power spectral densities of all frame signals are arranged in the order of the frame signals and the order of frequency from low to high to form a two-dimensional matrix, which is the time-frequency spectrum matrix in the time-frequency domain. The rows of the time-frequency spectrum matrix correspond to the time dimension, and each row vector corresponds to a time point of a frame signal. The columns correspond to the frequency dimension, and each column vector corresponds to a frequency point.

[0103] If continuous wavelet transform is used, the wavelet basis function is first selected based on the oscillation frequency distribution and amplitude variation characteristics of the candidate fault signal: when the oscillation frequency interval of the candidate fault signal, that is, the frequency difference corresponding to two adjacent oscillation peaks, is less than the preset interval threshold, the db wavelet is selected. The tight support characteristic of the db wavelet allows the length of the non-zero interval of the wavelet function in the time domain to be less than the minimum oscillation period of the candidate fault signal, which can accurately capture the dense oscillation characteristics; when the amplitude change rate of adjacent sampling points of the candidate fault signal, that is, the ratio of the amplitude difference between two adjacent sampling points to the sampling interval, is less than the preset change rate threshold, the sym wavelet is selected. The symmetry of the sym wavelet allows the amplitude change trend of the wavelet-transformed signal to be consistent with the amplitude change trend of the original candidate fault signal, satisfying the signal smoothness requirement. The scale range corresponding to the fault characteristic frequency band is calculated through a scale-frequency mapping relationship, where scale is inversely proportional to frequency. The mapping coefficient is the center frequency of the wavelet basis function. The minimum scale is the ratio of the mapping coefficient to the highest frequency of the fault characteristic frequency band, and the maximum scale is the ratio of the mapping coefficient to the lowest frequency of the fault characteristic frequency band. The multi-scale decomposition process covers all integer scales from the minimum to the maximum scale, ensuring that the frequency corresponding to each scale can cover all frequency points within the fault characteristic frequency band. Wavelet transform is performed on the candidate fault signal at each integer scale, i.e., the convolution integral of the candidate fault signal with the wavelet basis function at the corresponding scale is calculated to obtain the wavelet coefficient sequence at the corresponding scale. The wavelet coefficient sequences of all scales are arranged in order from the minimum to the maximum scale and in the sampling time order of the candidate fault signal to form a two-dimensional matrix, which is the wavelet coefficient matrix of the signal at different scales. The rows of the wavelet coefficient matrix correspond to the time dimension, and each row vector corresponds to a sampling time point of the candidate fault signal. The columns correspond to the scale dimension, and each column vector corresponds to a scale.

[0104] The characteristic parameters of candidate fault signals include instantaneous energy and spectral kurtosis, and the specific calculation methods include:

[0105] Instantaneous energy is the energy intensity of a candidate fault signal within the fault characteristic frequency band at a specific time point. It is calculated based on a time-frequency spectrum matrix or wavelet coefficient matrix. The specific method includes: selecting the set of column vectors corresponding to the fault characteristic frequency band from the time-frequency spectrum matrix or wavelet coefficient matrix; first, determining the scale range corresponding to the fault characteristic frequency band through a scale-frequency mapping relationship; then, selecting the set of column vectors within the scale range; for each row in the matrix, i.e., each time point, summing all elements in the corresponding fault characteristic frequency band column vector set. If calculated based on the time-frequency spectrum matrix, the element value in the fault characteristic frequency band column vector set is the power spectral density at the corresponding time-frequency point, and the summation result is the instantaneous energy at the corresponding time point. If calculated based on the wavelet coefficient matrix, the wavelet coefficients need to be squared first to represent the energy, and then the squared coefficients in the scale column vector set corresponding to the fault characteristic frequency band at the corresponding time point are summed, and the summation result is the instantaneous energy at the corresponding time point.

[0106] Spectral kurtosis is a high-order statistic that quantifies the degree to which the impulse components in a candidate fault signal deviate from Gaussian noise. Its calculation is based on the second and fourth moments of the candidate fault signal in the time-frequency domain. The specific method includes: dividing the time-frequency spectrum matrix or wavelet coefficient matrix into several local time windows along the time dimension, with the window length determined according to the time-domain characteristics of the candidate fault signal to ensure that each window contains complete local signal features; first calculating the second central moment of the elements corresponding to the fault feature frequency bands within each local time window (i.e., first calculating the element mean, then summing the squared differences between each element and the mean and taking the average); then calculating the fourth central moment of the elements corresponding to the fault feature frequency bands within each local time window (i.e., first calculating the element mean, then summing the fourth power differences between each element and the mean and taking the average); based on spectral kurtosis = four... The spectral kurtosis value for each local time window is calculated as (first central moment / (square of second central moment)) - 3. The theoretical spectral kurtosis value for Gaussian noise is 0. This formula reduces the spectral kurtosis corresponding to Gaussian noise to 0 by subtracting 3, which facilitates the differentiation of impulse components. By subtracting 3 from the spectral kurtosis formula, the calculation result for Gaussian distributed signals is reduced to 0. Signals containing impulse components will have a calculation result greater than 0 because the fourth central moment / (square of second central moment) is greater than 3. This forms a discrimination criterion with 0 as a clear boundary. A spectral kurtosis value greater than 0 indicates that the signal contains impulse components, while a spectral kurtosis value equal to 0 indicates that the signal conforms to a Gaussian distribution and has no impulse components. The setting of subtracting 3 can intuitively and accurately quantify the degree to which the signal deviates from the Gaussian distribution, which is convenient for subsequent differentiation of impulse components in candidate fault signals from background Gaussian noise. Replacing the "3" in the spectral kurtosis formula with other values ​​would disrupt the distinction between Gaussian distributed signals and signals containing impulse components. For example, replacing it with 2 would change the spectral kurtosis value of the Gaussian distributed signal to 1, requiring the spectral kurtosis value of the signal containing impulse components to be greater than 1 to be considered pulse-containing. This not only alters the physical meaning of the quantification of spectral kurtosis's deviation from the Gaussian distribution but also leads to incompatibility with existing fault feature discrimination logic, increasing the complexity of subsequent feature parameter comparisons and even causing misjudgments or omissions of impulse components. Therefore, it must be fixed at 3. If the spectral kurtosis value is greater than 0, it indicates that the candidate fault signal within the corresponding local time window contains impulse components. The larger the spectral kurtosis value, the larger the ratio of the fourth central moment to the second central moment, and the more prominent the peak value of the transient impulse component in the signal, and the more significant the deviation from Gaussian noise. If the spectral kurtosis value is less than or equal to 0, it indicates that the candidate fault signal within the corresponding local time window is mainly Gaussian noise and does not contain obvious impulse components.

[0107] The mean value of the instantaneous energy and the spectral kurtosis value are used to form an eigenvector.

[0108] The fault determination module, based on a pre-built fault identification model, performs comprehensive discrimination on feature vectors to obtain fault determination results.

[0109] Please see Figure 4As shown, the methods for obtaining fault determination results by comprehensively judging feature vectors based on pre-built fault identification models include:

[0110] Methods for constructing fault identification models include:

[0111] A sample set was constructed by collecting acoustic samples of micro-leakage, coking and cracking, and boiler background noise. Specifically, the micro-leakage acoustic samples were noise-reduced samples of acoustic signals collected when the boiler experienced a micro-leakage fault; the coking and cracking acoustic samples were noise-reduced samples of acoustic signals collected when the boiler experienced a coking and cracking fault; and the boiler background noise samples were noise-reduced samples of acoustic signals collected during normal boiler operation without fault components. Labels were assigned to the three types of samples: micro-leakage acoustic samples were labeled as micro-leakage fault, coking and cracking acoustic samples were labeled as coking and cracking fault, and boiler background noise samples were labeled as normal.

[0112] For each sample in the sample set, feature parameters are extracted to construct a feature vector: the mean of the instantaneous energy of the sample is calculated, which is the arithmetic mean of the instantaneous energy of the sample at all time points in the full time range; the spectral kurtosis value of the sample is calculated, which is the arithmetic mean of the spectral kurtosis values ​​of the sample at all local time windows in the full time range; the mean of the instantaneous energy and the spectral kurtosis value are arranged in a fixed order to form the feature vector corresponding to each sample.

[0113] The fault identification model employs a one-dimensional convolutional neural network architecture, comprising an input layer, convolutional layer, pooling layer, fully connected layer, and output layer connected sequentially. The input layer receives feature vectors, with the input dimension matching the length of the feature vectors. The convolutional layer performs convolution operations on the input features using a predetermined number of one-dimensional convolutional kernels, extracting locally correlated features from the feature vectors. The kernel size is set based on the dimension of the feature vectors and the local correlation of the fault features. The pooling layer uses max pooling to downsample the feature map output by the convolutional layer, retaining key features while reducing the feature dimension. The pooling window size and stride are set based on the size of the feature map output by the convolutional layer. The fully connected layer maps the features output by the pooling layer into high-dimensional feature vectors, enhancing the fault identification model's ability to fit complex features through a non-linear activation function. The output layer uses a softmax activation function, outputting probability values ​​for three categories: normal, minor leak faults, and coking / cracking faults, with the sum of these probabilities equal to 1.

[0114] The sample set is divided into a training set and a validation set. The training set is used for learning the parameters of the fault identification model, and the validation set is used to monitor the generalization ability of the fault identification model during training. A one-dimensional convolutional neural network is trained with feature vectors as input and sample labels as the target output. The cross-entropy loss function is used to calculate the error between the output probability value of the fault identification model and the sample label, and the label is represented in one-hot encoding form. A gradient descent optimizer is used to iteratively update the parameters of each layer of the fault identification model through backpropagation algorithm to minimize the loss function value. A maximum number of training iterations and an early stopping mechanism are set. Training is stopped when the loss of the validation set does not decrease for a preset number of consecutive iterations to avoid overfitting of the fault identification model. After training, the fault identification model has the ability to learn the patterns of fault signals of micro-leakage faults and coking cracking faults, and can output the probability values ​​of three types of results based on the input feature vector.

[0115] The performance of the trained fault identification model is evaluated based on the validation set, and the recognition accuracy of the fault identification model for three types of samples is calculated. If the recognition accuracy does not reach the preset accuracy threshold, the architectural parameters of the one-dimensional convolutional neural network are adjusted, including the number of convolutional kernels, the size of the convolutional kernels, the size of the pooling window, and the number of neurons in the fully connected layer, and the training process is re-executed. If the recognition accuracy reaches the preset accuracy threshold, it is determined as the final fault identification model.

[0116] The feature vector is input into the fault identification model, which adopts a pre-trained one-dimensional convolutional neural network architecture. The fault identification model has been trained with micro-leakage acoustic samples, coking and cracking acoustic samples, and boiler background noise samples. It has the ability to learn the patterns of fault signals of micro-leakage faults and coking and cracking faults. The fault identification model outputs probability values ​​of three categories of results: normal, micro-leakage fault, and coking and cracking fault. If the probability value of micro-leakage fault is greater than the preset micro-leakage confidence threshold, a micro-leakage fault is determined to have occurred. If the probability value of coking and cracking fault is greater than the preset coking confidence threshold, a coking and cracking fault is determined to have occurred. If the probability values ​​of both micro-leakage fault and coking and cracking fault are less than the corresponding confidence thresholds, no fault is determined to have occurred.

[0117] For cases where a fault is detected, the system records the timestamp of the fault determination and the fault characteristic frequency band. Combined with the time difference from the sensor array correlation analysis, the approximate location of the fault sound source inside the boiler is deduced, thus determining the fault location. After accumulating multiple frames of analysis results to reduce false alarms, the system generates an alarm message. The alarm message includes the fault type determination, as well as additional information such as time and sensor location. The system can also save the acoustic characteristic data related to the fault event for further diagnosis by maintenance personnel. When the target fault signal is successfully identified, the system issues a warning notification, such as an audible and visual alarm or a control signal alarm, reminding maintenance personnel to promptly inspect the boiler equipment.

[0118] This embodiment achieves effective early warning of boiler micro-leakage and early coking faults in the presence of strong noise through an innovative signal processing scheme, significantly reducing the failure rate and ensuring the safe and efficient operation of the generator set.

[0119] Example 2:

[0120] Please see Figure 2 As shown, this embodiment provides a method for intelligent diagnosis of the status of thermal power generation equipment, including:

[0121] The original acoustic signals of the boiler are acquired in real time, and micro-leakage acoustic samples, coking and cracking acoustic samples and boiler background noise samples are obtained in advance.

[0122] Frequency domain power spectral density analysis was used to extract the frequency domain features of micro-leakage acoustic samples, coking and cracking acoustic samples, and boiler background noise samples; based on the frequency domain features, the fault characteristic frequency bands were determined, and bandpass filters and notch filters were set.

[0123] The original acoustic signal is denoised by using a bandpass filter and a notch filter to obtain the denoised acoustic signal.

[0124] Based on the preset fault judgment threshold, the denoised acoustic signal is marked as a suspected fault signal, and the suspected fault signal is marked as a candidate fault signal.

[0125] The feature vector of the candidate fault signal is obtained by using the corresponding time-frequency transformation processing method;

[0126] Based on the pre-built fault identification model, the feature vectors are comprehensively judged to obtain the fault determination result.

[0127] The above description is merely a specific embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any variations or substitutions that can be easily conceived by those skilled in the art within the technical scope disclosed in the present invention should be included within the scope of protection of the present invention. Therefore, the scope of protection of the present invention should be determined by the scope of the claims.

[0128] In conclusion, the above description is only a preferred embodiment of the present invention and is not intended to limit the present invention. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the protection scope of the present invention.

Claims

1. A method for intelligent diagnosis of the condition of thermal power generation equipment, characterized in that, include: The original acoustic signals of the boiler are acquired in real time, and micro-leakage acoustic samples, coking and cracking acoustic samples and boiler background noise samples are obtained in advance. Frequency domain power spectral density analysis was used to extract frequency domain features from micro-leakage acoustic samples, coking cracking acoustic samples, and boiler background noise samples. Based on these frequency domain features, fault characteristic frequency bands were determined, and bandpass and notch filters were set. Specifically, the maximum power spectral density of the micro-leakage acoustic samples was recorded as the first peak power, the maximum power spectral density of the coking cracking acoustic samples as the second peak power, and the maximum power spectral density of the boiler background noise samples as the third peak power. Based on the power spectral density value and the maximum power spectral density of each data type, the energy concentration region of each data type was determined, resulting in a first frequency range, a second frequency range, and a third frequency range. Based on the average energy proportion of each data type in the first, second, and third frequency ranges, the energy concentration interval of the fault signal was calculated. Configure a digital bandpass filter as a filtering component, and set the low cutoff frequency of the digital bandpass filter as the lower limit of the fault signal energy concentration range and the high cutoff frequency as the upper limit of the fault signal energy concentration range. Calculate the highest frequency value of the boiler background noise sample in the third frequency range. The low cutoff frequency of the digital bandpass filter is greater than the highest frequency value in the third frequency range. The original acoustic signal is denoised by using a bandpass filter and a notch filter to obtain the denoised acoustic signal. Based on the preset fault judgment threshold, the denoised acoustic signal is marked as a suspected fault signal, and the suspected fault signal is marked as a candidate fault signal. The feature vector of the candidate fault signal is obtained by using the corresponding time-frequency transformation processing method; Based on the pre-built fault identification model, the feature vectors are comprehensively judged to obtain the fault determination result.

2. The intelligent diagnostic method for the condition of thermal power generation equipment according to claim 1, characterized in that, Methods for obtaining the feature vectors of candidate fault signals include: The feature vectors of candidate fault signals include instantaneous energy and spectral kurtosis; The instantaneous energy is calculated as follows: if it is calculated based on the time-frequency spectrum matrix of the candidate fault signal, the element value in the column vector set of the fault feature frequency band is the power spectral density at the corresponding time-frequency point, thus obtaining the instantaneous energy at the corresponding time point; if it is calculated based on the wavelet coefficient matrix of the candidate fault signal, the wavelet coefficients are first squared, and then the squared coefficients in the column vector set of the corresponding scale of the fault feature frequency band at the corresponding time point are summed to obtain the instantaneous energy at the corresponding time point. The spectral kurtosis is calculated as follows: the time-frequency spectrum matrix or wavelet coefficient matrix of the candidate fault signal is divided into several local time windows according to the time dimension; first, the second central moment of the corresponding element of the fault characteristic frequency band in each local time window is calculated; then, the fourth central moment of the corresponding element of the fault characteristic frequency band in each local time window is calculated; spectral kurtosis = fourth central moment / (square of second central moment) - 3.

3. The intelligent diagnostic method for the condition of thermal power generation equipment according to claim 2, characterized in that, Methods for obtaining the time-frequency spectrum matrix and wavelet coefficient matrix of candidate fault signals include: When the time-domain duration of the candidate fault signal is greater than the preset duration threshold, and the width of the fluctuation range of the frequency components of the candidate fault signal within the fault characteristic frequency band is less than the product of the width of the fault characteristic frequency band and the preset fluctuation ratio coefficient, a short-time Fourier transform is used; otherwise, a continuous wavelet transform is used. If a short-time Fourier transform is used, the candidate fault signal is divided into several frame signals according to the set window length and frame shift. A Fourier transform is performed on each frame signal to obtain the complex spectrum of each frame signal in the frequency domain. The power spectral density of each frame signal is obtained by taking the modulus and squaring the complex spectrum. The power spectral densities of all frame signals are arranged in the order of the frame signals and the order of frequency from low to high to form a two-dimensional matrix, which is the time-frequency spectrum matrix. The rows of the time-frequency spectrum matrix correspond to the time dimension, and the columns correspond to the frequency dimension. If continuous wavelet transform is used, the convolution integral of the candidate fault signal with the preset wavelet basis function at multiple set scales is calculated to obtain the wavelet coefficient sequence at the corresponding scale. The wavelet coefficient sequences of all scales are arranged in order from the smallest scale to the largest scale and in the sampling time order of the candidate fault signal to form a two-dimensional matrix, which is the wavelet coefficient matrix of the candidate fault signal at different scales. The rows of the wavelet coefficient matrix correspond to the time dimension and the columns correspond to the scale dimension.

4. The intelligent diagnostic method for the condition of thermal power generation equipment according to claim 1, characterized in that, Methods for marking suspected fault signals as candidate fault signals include: The suspected fault signal is preprocessed to remove the DC component and residual low-frequency interference, resulting in the preprocessed suspected fault signal. For the pre-processed suspected fault signals, calculate the cross-correlation function between the suspected fault signals output by any two acoustic sensors, obtain the cross-correlation function curve through cross-correlation operation, and extract the peak position and peak amplitude of the suspected fault signal from the cross-correlation function curve. If the time difference corresponding to the peak position of the suspected fault signal falls within the preset theoretical propagation time difference range, and the peak amplitude of the suspected fault signal is greater than the preset coherence threshold, then the corresponding suspected fault signal will be marked as a candidate fault signal; otherwise, the corresponding suspected fault signal will be remarked as a normal signal.

5. The intelligent diagnostic method for the condition of thermal power generation equipment according to claim 1, characterized in that, Methods for marking suspected fault signals in denoised acoustic signals include: Fault determination thresholds include the micro-leakage characteristic energy determination threshold and the coking rupture pulse amplitude determination threshold; Calculate the mean power spectral density of the noise-reduced acoustic signal within the fault characteristic frequency band; If the mean power spectral density is greater than the threshold for determining microleakage characteristic energy, the corresponding acoustic signal is marked as a suspected microleakage signal; otherwise, the corresponding acoustic signal is marked as a normal signal. The instantaneous amplitude of the denoised acoustic signal is obtained by traversing the denoised acoustic signal through a sliding window. The number of pulses with an instantaneous amplitude greater than the threshold for judging the amplitude of coking and cracking pulses is counted. If the number of pulses is greater than the preset counting threshold, the corresponding acoustic signal is marked as a suspected coking and cracking signal; otherwise, the corresponding acoustic signal is marked as a normal signal. Acoustic signals marked as suspected micro-leakage signals and suspected coking and cracking signals in the noise-reduced acoustic signals were selected as suspected fault signals.

6. The intelligent diagnostic method for the condition of thermal power generation equipment according to claim 1, characterized in that, The passband frequency band corresponding to the digital bandpass filter is the fault characteristic frequency band.

7. The intelligent diagnostic method for the condition of thermal power generation equipment according to claim 6, characterized in that, Methods for obtaining the first frequency range, the second frequency range, and the third frequency range include: In the micro-leak acoustic sample, the continuous frequency range formed by the frequency points whose power spectral density value is greater than or equal to the first peak power preset first proportion is the first frequency range, and the energy proportion corresponding to the first frequency range is greater than the preset fault energy proportion threshold. In the acoustic sample of coking and cracking, the continuous frequency range formed by the frequency points whose power spectral density value is greater than or equal to the second peak power preset second ratio is the second frequency range, and the energy ratio corresponding to the second frequency range is greater than the fault energy ratio threshold. In the boiler background noise sample, the continuous frequency range formed by the frequency points whose power spectral density values ​​are greater than or equal to the third peak power preset third proportion is the third frequency range, and the energy proportion corresponding to the third frequency range is greater than the preset background energy proportion threshold. The upper limit of the third frequency range is less than the lower limit of the first frequency range and the lower limit of the second frequency range, and there is a continuous frequency segment that partially overlaps with the first frequency range and the second frequency range. The frequency points in the overlapping continuous frequency segment belong to both the first frequency range and the second frequency range.

8. The intelligent diagnostic method for the condition of thermal power generation equipment according to claim 6, characterized in that, Methods for obtaining the energy concentration range of fault signals include: The overlapping area between the first frequency range and the second frequency range is taken as the initial interval. The average energy ratio of micro-leakage acoustic samples and coking cracking acoustic samples in the initial interval, as well as the average energy ratio of boiler background noise samples in the initial interval are calculated. If the average energy proportion of both types of fault samples in the initial interval is greater than the preset fault energy proportion threshold, and the average energy proportion of boiler background noise samples is less than the preset background energy proportion threshold, then the initial interval is determined as the fault signal energy concentration interval; otherwise, the upper and lower limits of the initial interval are adjusted, and the average energy proportion of the three types of data in the adjusted interval is recalculated until the conditions of the average energy proportion of both types of fault samples and the average energy proportion of background noise are met, and the fault signal energy concentration interval is finally determined.

9. The intelligent diagnostic method for the condition of thermal power generation equipment according to claim 1, characterized in that, Methods for obtaining fault determination results include: The mean value of the instantaneous energy and the spectral kurtosis value are used to form a feature vector; The feature vector is input into the fault identification model, and the output includes probability values ​​for three types of results: normal, micro-leakage fault, and coking cracking fault. If the probability value of a microleak fault is greater than the preset microleak confidence threshold, then a microleak fault is determined to have occurred. If the probability value of coking failure is greater than the preset coking confidence threshold, then coking failure is determined to have occurred. If the probability values ​​of both micro-leakage fault and coking cracking fault are less than the corresponding confidence threshold, then it is determined that no fault has occurred.

10. A state-of-the-art diagnostic system for thermal power generation equipment, used to implement the state-of-the-art diagnostic method for thermal power generation equipment as described in any one of claims 1-9, characterized in that, include: The data acquisition module is used to acquire the original acoustic signals of the boiler in real time, and to pre-acquire micro-leakage acoustic samples, coking and cracking acoustic samples, and boiler background noise samples. The feature setting module is used to extract the frequency domain features of micro-leakage acoustic samples, coking and cracking acoustic samples, and boiler background noise samples through frequency domain power spectral density analysis; determine the fault characteristic frequency band based on the frequency domain features, and set the bandpass filter and notch filter. The noise reduction processing module is used to perform noise reduction processing on the original acoustic signal through a bandpass filter and a notch filter to obtain a noise-reduced acoustic signal; The signal marking module is used to mark suspected fault signals on the noise-reduced acoustic signals according to a preset fault judgment threshold, and to mark the suspected fault signals as candidate fault signals. The feature vector module is used to obtain the feature vector of the candidate fault signal by employing the corresponding time-frequency transformation processing method. The fault determination module, based on a pre-built fault identification model, performs comprehensive discrimination on feature vectors to obtain fault determination results.

Images