Intelligent auscultation diagnosis method and device for steam turbine

Abstract

The present application relates to the field of steam turbine equipment state monitoring, and particularly relates to a steam turbine intelligent auscultation diagnosis method and device, comprising obtaining double-path effective sound signals; after completing standardization processing of characteristic parameters, a standardized characteristic set directly used for fault matching is constructed; a characteristic matching benchmark library suitable for the current measured steam turbine unit is generated; through an association mechanism, fault root measuring points and influence measuring points caused by fault sound propagation are distinguished, and defect positioning is completed. The present application replaces artificial subjective judgment with quantitative analysis, combines a fault sound database and a machine learning algorithm, can accurately identify different types of faults, reduces the risk of missed judgment and misjudgment, displays sound wave graphs and diagnosis results on site, provides intuitive basis for operation and maintenance personnel, and improves diagnosis efficiency.

Description

Technical Field

[0001] This invention relates to the field of steam turbine equipment condition monitoring, and specifically to an intelligent auscultation diagnostic method and device for steam turbines. Background Technology

[0002] Steam turbines are core power equipment in energy generation, heavy equipment manufacturing, and other fields. The stability of their operation directly determines the safety and efficiency of the entire production system. Traditional steam turbine condition inspections rely on listening rods. Maintenance personnel listen closely to detect abnormal noises from key components such as bearings, rotors, and steam seals by placing their ears close to the ground, thereby diagnosing equipment malfunctions.

[0003] Traditional listening rod inspection methods have significant limitations: First, they rely heavily on personnel experience, and different maintenance personnel have different auditory sensitivity and judgment standards, which can easily lead to missed or misjudged cases; Second, manual inspection cannot achieve 24-hour uninterrupted monitoring and is difficult to capture intermittent fault signals; Third, it is impossible to quantify, store, and analyze sound data, which is not conducive to tracking equipment fault trends and early warning.

[0004] With the rapid development of the Industrial Internet and intelligent sensing technology, there is an urgent need for a device that can replace the traditional listening rod and realize the digital acquisition and intelligent analysis of steam turbine sound signals, so as to improve the level of intelligence in equipment status monitoring. Summary of the Invention

[0005] This invention addresses the technical problems existing in the prior art by providing a method and apparatus for intelligent auscultation diagnosis of steam turbines.

[0006] The technical solution of the present invention to solve the above technical problems is as follows: A steam turbine intelligent auscultation diagnosis method, including the following steps: S100, based on the original sound signal synchronously collected by the contact sound sensor and the environmental noise sensor, and the timing alignment reference signal generated by the embedded microprocessor, to obtain dual effective sound signals, one of which is the original sound signal of the steam turbine operation collected by the contact sound sensor, and the other is the ambient background noise signal collected by the environmental noise sensor. S200, based on time-aligned dual-channel effective sound signals, eliminates environmental noise interference through an adaptive filtering algorithm, and purifies to obtain effective operating signals containing only the mechanical vibrations inside the steam turbine. Then, it performs full-dimensional analysis in the time and frequency domains, extracts feature parameters through a fast Fourier transform algorithm, and constructs a standardized feature set that can be directly used for fault matching after completing the standardization of the feature parameters. S300: Retrieve the preset fault sound database, compare the standardized feature set with the standard normal feature set of the same model unit extracted from the database item by item, introduce the adaptation strategy, generate the corresponding working condition adaptation coefficient based on the comparison structure and then make corrections, and generate the feature matching benchmark library that is adapted to the current steam turbine unit under test. S400 intelligently compares the standardized feature set to be diagnosed with the feature matching benchmark library, completes the matching and identification of fault features through the convolutional neural network algorithm, and outputs the fault diagnosis results including fault location, fault type and cosine similarity value. Then, through the association mechanism, it distinguishes the fault root cause measurement point from the influence measurement point caused by the propagation of fault sound, and completes the defect location.

[0007] In a preferred embodiment, S100 sends a timing alignment reference signal to the contact sound sensor and the ambient noise sensor at the same time the embedded microprocessor starts collecting data. The original sound signals collected all contain the timing alignment reference signal, which marks the dual-channel acquisition signals with a unified time reference, thus solving the problem of misalignment of the acquisition start time caused by the difference in hardware circuit and response speed between the two sensors. The embedded microprocessor extracts the start time of the timing alignment reference signal from the original sound signal collected by the contact sound sensor and the start time of the timing alignment reference signal from the original sound signal collected by the environmental noise sensor, respectively, and calculates the time difference between the two start times. Furthermore, this application provides the above-mentioned means to transform the abstract time misalignment into a calculable specific value. The embedded microprocessor performs time-domain shifting on the original audio signal with a time lag based on the time difference, so that the start times of the timing alignment reference signals of the two original audio signals completely coincide, thereby completing the timing alignment of the two signals and obtaining two valid audio signals.

[0008] In a preferred embodiment, S200 sets the ambient background noise signal collected by the environmental noise sensor in the dual-channel effective sound signal as the reference input signal of the adaptive filtering algorithm, and sets the original turbine operation sound signal collected by the contact sound sensor in the dual-channel effective sound signal as the main input signal of the adaptive filtering algorithm. A transverse filter is constructed based on the reference input signal. The initial weight coefficients of the filter are set to equal values, and the filter order is set to 128. The reference input signal is input into the transverse filter to generate an analog noise output signal. The difference between the main input signal and the analog noise output signal is calculated to obtain the error signal. The error signal refers to the preliminarily purified turbine operating sound signal. The difference calculation preliminarily removes the environmental noise component from the main input signal, completing the core calculation step of noise elimination, and obtaining a preliminarily purified signal that retains only the turbine operating characteristics. Based on the minimum mean square error criterion, with the goal of minimizing the error signal, the weight coefficients of the transverse filter are iteratively updated. After each iteration, the analog noise output signal and the error signal are recalculated until the mean square value of the error signal stabilizes. Then the iteration stops, and the purified effective operating signal of the steam turbine is obtained. During the iteration process, the filter weight coefficients are continuously optimized so that the analog noise output signal can infinitely approximate the real environmental noise component in the main input signal.

[0009] In a preferred embodiment, S200 further includes: Based on an embedded microprocessor, the effective operating signals of the steam turbine are subjected to time-domain feature extraction to obtain the time-domain feature parameters of the corresponding effective operating signal time-domain waveform, including time-domain peak value, time-domain mean, time-domain variance, kurtosis, and margin index. These indexes can capture the instantaneous fluctuation and impact characteristics of the steam turbine operating signals from the time-domain dimension. These features can directly reflect the sudden anomalies of the steam turbine mechanical components, such as dynamic and static friction, component loosening, bearing cracks, etc. Then, the effective operating signal of the steam turbine in the time domain is converted into a spectrum signal in the frequency domain by using the Fast Fourier Transform algorithm. Specifically, the effective operating signal of the steam turbine is windowed, and the Hanning window is selected as the window function to avoid spectrum leakage. Then, the windowed effective operating signal is subjected to Fast Fourier Transform calculation to convert the time domain signal into a frequency domain signal and obtain the frequency domain spectrum.

[0010] In a preferred embodiment, S200 further includes: extracting frequency domain feature parameters from the frequency domain spectrum, including spectral peak value, spectral centroid, spectral variance, main frequency amplitude, sideband amplitude, and harmonic amplitude. Unlike the above feature parameters, the frequency domain feature parameters extracted in this application correspond to different types and locations of faults, such as the characteristic frequency corresponding to bearing wear, the broadband characteristic corresponding to steam seal leakage, and the main frequency harmonic characteristic corresponding to rotor imbalance. Therefore, they complement the time domain features. All extracted time-domain and frequency-domain feature parameters are merged into an initial feature parameter group. For each feature parameter in the initial feature parameter group, feature parameter normalization is performed to obtain standardized feature values. All normalized feature values ​​are arranged in a fixed order of time-domain and frequency-domain features to construct the corresponding standardized feature set.

[0011] In a preferred embodiment, step S300 retrieves a preset fault sound database, extracts a standard normal feature set under normal operating conditions of the unit that matches the model of the turbine being tested, and a standard fault feature set corresponding to the model unit, such as typical faults such as bearing wear, dynamic and static rubbing, and steam seal leakage. It should be noted that the preset fault sound database is obtained by filtering from a general database, and the benchmark feature data that matches the current model unit is selected from the general database to eliminate feature deviations caused by differences in structure and parameters between different models of turbines. Because of the inherent differences in operating characteristics between the current on-site units and the standard units of the same model in the database, the operating condition differences abstracted in this application are transformed into calculable correction coefficients. For example, the obtained standardized feature set is compared item by item with the standard normal feature set of the same model extracted from the database, the ratio of each feature parameter is calculated, and the weighted average of all ratios is taken to generate the operating condition adaptation coefficient corresponding to the unit. The product of all standard normal feature values ​​in the marked normal feature set and the corresponding operating condition mismatch coefficient is used as the adapted fault feature value. After item-by-item correction, it is classified and archived according to fault type, fault location and severity level. Each type of fault corresponds to a unique adapted fault feature set. At the same time, the measurement point location information of the current unit is supplemented to generate a feature matching benchmark library that is only applicable to the current steam turbine unit under test.

[0012] In a preferred embodiment, S400 further includes: All adapted fault feature sets in the feature matching benchmark library are used as training benchmark samples for the convolutional neural network algorithm. After initializing the input layer parameters of the convolutional neural network algorithm, the number of neurons in the input layer is made to be completely consistent with the number of feature parameters in the standardized feature set. The standardized feature set under the working condition to be diagnosed is then input into the input layer of the initialized convolutional neural network algorithm as the input data to be identified. The convolutional layer performs feature depth extraction on the standardized feature set to be identified. In some other specific embodiments, this application uses three convolutional layers, each with 64 convolutional kernels, to perform sliding convolution calculation on the input standardized feature set, extract the correlation features between feature parameters, and generate a feature map. These correlation features can accurately reflect the unique feature patterns of different faults, which is different from the surface comparison of single feature parameters. Therefore, it can improve the algorithm's ability to identify weak fault features and composite fault features, and generate a feature map that can be used for accurate matching. The feature map output by the convolutional layer is then reduced in dimensionality by a pooling layer. Max pooling is used to retain the feature values ​​in the feature map except for redundant features. The dimensionality-reduced features output by the pooling layer are then matched one by one with each adapted fault feature set in the feature matching benchmark library using cosine similarity. This generates a cosine similarity value between the feature set to be identified and each fault feature set. The fault feature set corresponding to the highest cosine similarity value is taken as the diagnostic result. The fault location, fault type, and cosine similarity value of all measurement points in the fault feature set corresponding to the highest cosine similarity value are integrated to form the fault diagnosis result.

[0013] In a preferred embodiment, after the S400 uploads the diagnostic results of each measuring point through the communication module, it retrieves the pre-stored turbine measuring point deployment location distribution map and spatial association weight matrix. Therefore, this application summarizes the diagnostic data of all measuring points of the entire unit, and at the same time retrieves the unit's spatial location information and measuring point association weight data. The test points that indicate faults in the diagnostic results of each test point are selected and marked as fault-related test points. The fault type, cosine similarity value and test point location information of each fault-related test point are extracted. Based on the spatial association weight matrix, the propagation influence value of each fault-related test point on all other fault-related test points is calculated. For each fault-related measurement point, the difference between the corresponding feature value and the sum of the propagation influence values ​​of all other fault-related measurement points is calculated and used as the inherent fault feature value of the measurement point. Finally, the influence of the fault sound propagation of other measurement points on the current measurement point is removed, and the inherent fault feature value generated by the current measurement point itself is extracted. This value directly reflects whether the measurement point itself has a real fault. The higher the value, the more obvious the fault characteristics of the measurement point itself are. The inherent fault characteristic values ​​of all fault-related measurement points are sorted in descending order. The measurement point with the highest inherent fault characteristic value is identified as the root cause measurement point of the fault, and the remaining fault-related measurement points are identified as measurement points that affect the fault propagation. At the same time, a fault root cause location report is generated based on the spatial correlation weight matrix, and the defect location is completed through the correlation mechanism.

[0014] In a preferred embodiment, the turbine measuring point deployment location distribution map and spatial correlation weight matrix pre-stored in S400 include: By retrieving the factory structural drawings of the turbine under test, a base map of the main structure is constructed using proportional scaling and drawing software. The material properties and acoustic parameters of each component are entered into the base map, including the carbon steel material of the bearing housing, the alloy steel structure of the cylinder, the stainless steel material of the connecting pipes, as well as the longitudinal wave velocity, transverse wave velocity, and solid-borne sound attenuation coefficient of the corresponding materials. All deployed measuring points are marked in the base map, including bearing housing measuring points, steam seal measuring points, speed control valve measuring points, and coupling measuring points. Each measuring point is marked with a unique and non-repeatable measuring point number, and all measuring point attribute information is entered, namely the specific location of the measuring point installation, the turbine component corresponding to the measuring point, the base material type of the measuring point, the plane coordinates of the measuring point, the straight-line distance between adjacent measuring points, and the relative position of the measuring point to the core rotating components such as the turbine rotor, forming a measuring point deployment location distribution map. Retrieve the distribution map of the measurement point deployment locations, count the total number of actual deployed measurement points, and construct a two-dimensional matrix consistent with the total number of measurement points. The rows and columns of the matrix correspond to the unique measurement point numbers in the distribution map. The rows represent the sound source measurement points, and the columns represent the affected measurement points, that is, the measurement points that receive the sound of the fault propagation. The sound source measurement points are the measurement points where the fault occurs. The weight coefficients of the diagonal elements of the two-dimensional matrix are fixed at 1, which means that the influence weight of the fault characteristics of the measuring point on itself is 100%, with no propagation attenuation. The initial values ​​of all off-diagonal elements of the two-dimensional matrix are uniformly set to 0, thus forming an initial spatial correlation weight matrix framework that perfectly matches the number of measuring points. After calculating the sound propagation attenuation coefficient for each off-diagonal element of the two-dimensional matrix, the propagation efficiency between each measurement point is obtained, and the weight coefficients are filled into the spatial correlation weight matrix to complete the assignment.

[0015] The present invention also provides an intelligent auscultation diagnostic device for steam turbines, the device comprising: Contact-type sound sensor; used to be attached to the surface of measuring points including turbine bearing housings, steam seals, speed control valves, and couplings to collect raw sound signals generated by mechanical vibrations inside the turbine. An ambient noise sensor; used to synchronously collect background noise signals in industrial settings surrounding steam turbines with a contact-type sound sensor, and to acquire raw sound signals. Embedded microprocessor; used as the core display carrier to display the time-domain waveform and frequency-domain spectrum of the turbine's operating sound signal in real time, and simultaneously present the diagnostic results of the measurement point location, operating status, fault location, and fault type. Wireless communication chip; used as the communication function carrier to support Bluetooth, Wi-Fi, and industrial IoT communication protocols, realize the uploading of raw sound data, analysis results, and measurement point location information, as well as network communication and centralized data transmission management. Magnetic fixing end; used to use strong magnetic materials to be adsorbed onto the surface of metal measuring points such as turbine bearing seats and cylinder blocks to achieve equipment installation and fixation, adapting to the stable deployment requirements of high-speed operation and strong vibration of the unit. Adhesive fixing end; used for installing and fixing equipment with high temperature and high viscosity industrial adhesive tape, suitable for non-magnetic material measuring points of steam turbines, forming a double fixing structure with magnetic fixing end, covering the installation and deployment of all types of measuring points of steam turbines; Rechargeable lithium battery; used as a built-in power supply unit to provide operating power; DC power interface; used for connecting an external industrial DC power supply to provide uninterrupted power supply, forming a dual power supply system with a rechargeable lithium battery.

[0016] The beneficial effects of this invention are: by replacing subjective human judgment with quantitative analysis, and by combining a fault sound database with machine learning algorithms, different types of faults can be accurately identified, reducing the risk of missed or misjudged cases; the on-site display of acoustic waveforms and diagnostic results provides intuitive information for maintenance personnel and improves diagnostic efficiency.

[0017] Adopting a dual fixing design of magnetic attraction and adhesive, it can be deployed at multiple key measuring points of the steam turbine at the same time, supporting 24-hour uninterrupted monitoring, effectively capturing intermittent fault signals, and eliminating the pain points of numerous, wide-ranging, and low-frequency manual inspections.

[0018] Ensuring personnel safety and reducing maintenance intensity: No personnel need to have close contact with the high-temperature and rotating parts of the steam turbine. Data can be viewed remotely through local displays or upper-level systems, improving the safety of inspections; automated data collection and analysis reduce the workload of manual recording and judgment, thus reducing maintenance intensity. Attached Figure Description

[0019] Figure 1 This is a flowchart of the present invention; Figure 2 This is a schematic diagram of the modular structure of the present invention. Detailed Implementation

[0020] The technical solutions of the embodiments of this application will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of this application, and not all embodiments. Based on the embodiments of this application, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of this application.

[0021] As attached Figure 1-2 As shown, this embodiment provides: an intelligent auscultation diagnostic method for steam turbines, including the following steps: S100: Based on the original sound signals synchronously collected by the contact sound sensor and the environmental noise sensor, and the timing alignment reference signal generated by the embedded microprocessor, dual effective sound signals are obtained. One channel is the original sound signal of the turbine operation collected by the contact sound sensor, and the other channel is the ambient background noise signal collected by the environmental noise sensor. At the same time that the embedded microprocessor starts collecting data from the contact sound sensor and the ambient noise sensor, it sends a timing alignment reference signal to both sensors. The original sound signals collected all contain the timing alignment reference signal, which marks the dual-channel acquisition signals with a unified time reference, thus solving the problem of misaligned acquisition start times caused by differences in hardware circuits and response speeds between the two sensors. The embedded microprocessor extracts the start time of the timing alignment reference signal from the original sound signal collected by the contact sound sensor and the start time of the timing alignment reference signal from the original sound signal collected by the environmental noise sensor, respectively, and calculates the time difference between the two start times. Furthermore, this application provides the above-mentioned means to transform the abstract time misalignment into a calculable specific value. The embedded microprocessor performs time-domain shifting on the original audio signal with a time lag based on the time difference, so that the start times of the timing alignment reference signals of the two original audio signals completely coincide, thereby completing the timing alignment of the two signals and obtaining two valid audio signals.

[0022] In some other specific implementations, the amplitude, signal-to-noise ratio, and time-domain continuity of the two signals can be calculated separately. If any parameter of any signal exceeds the preset verification threshold range, the acquired signal is determined to be invalid, and the dual-channel signal acquisition is restarted.

[0023] S200, based on time-aligned dual-channel effective sound signals, eliminates environmental noise interference through an adaptive filtering algorithm, and purifies to obtain effective operating signals containing only the mechanical vibrations inside the steam turbine. Then, it performs full-dimensional analysis in the time and frequency domains, extracts feature parameters through a fast Fourier transform algorithm, and constructs a standardized feature set that can be directly used for fault matching after completing the standardization of the feature parameters. The ambient background noise signal collected by the environmental noise sensor in the dual-channel effective sound signal is set as the reference input signal of the adaptive filtering algorithm, and the original turbine operation sound signal collected by the contact sound sensor in the dual-channel effective sound signal is set as the main input signal of the adaptive filtering algorithm. A transverse filter is constructed based on the reference input signal. The initial weight coefficients of the filter are set to equal values, and the filter order is set to 128. The reference input signal is input into the transverse filter to generate an analog noise output signal. The difference between the main input signal and the analog noise output signal is calculated to obtain the error signal. The error signal refers to the preliminarily purified turbine operating sound signal. The difference calculation preliminarily removes the environmental noise component from the main input signal, completing the core calculation step of noise elimination, and obtaining a preliminarily purified signal that retains only the turbine operating characteristics. Based on the minimum mean square error criterion, with the goal of minimizing the error signal, the weight coefficients of the transverse filter are iteratively updated. After each iteration, the analog noise output signal and the error signal are recalculated until the mean square value of the error signal stabilizes. Then the iteration stops, and the purified effective operating signal of the steam turbine is obtained. During the iteration process, the filter weight coefficients are continuously optimized so that the analog noise output signal infinitely approximates the real environmental noise component in the main input signal. This is quite common in existing technologies, so it will not be described in detail. Furthermore, after the iteration stops, the embedded microprocessor determines the final output error signal as the effective operating signal of the turbine after the adaptive filtering algorithm. At the same time, it performs a second signal-to-noise ratio check on the effective signal to ensure that the signal-to-noise ratio is not lower than 40dB. If it is lower than this value, the filtering process is re-executed.

[0024] Based on an embedded microprocessor, the effective operating signals of the steam turbine are subjected to time-domain feature extraction to obtain the time-domain feature parameters of the corresponding effective operating signal time-domain waveform, including time-domain peak value, time-domain mean, time-domain variance, kurtosis, and margin index. These indexes can capture the instantaneous fluctuation and impact characteristics of the steam turbine operating signals from the time-domain dimension. These features can directly reflect the sudden anomalies of the steam turbine mechanical components, such as dynamic and static friction, component loosening, bearing cracks, etc. Then, the effective operating signal of the steam turbine in the time domain is converted into a spectrum signal in the frequency domain by using the Fast Fourier Transform algorithm. Specifically, the effective operating signal of the steam turbine is windowed, and the Hanning window is selected as the window function to avoid spectrum leakage. Then, the windowed effective operating signal is subjected to Fast Fourier Transform calculation to convert the time domain signal into a frequency domain signal and obtain the frequency domain spectrum.

[0025] Then, frequency domain feature parameters are extracted from the frequency domain spectrum, including spectral peak, spectral centroid, spectral variance, main frequency amplitude, sideband amplitude, and harmonic amplitude. Unlike the feature parameters mentioned above, the frequency domain feature parameters extracted in this application correspond to different types and locations of faults, such as the characteristic frequency corresponding to bearing wear, the broadband characteristic corresponding to steam seal leakage, and the main frequency harmonic characteristic corresponding to rotor imbalance. Therefore, they complement the time domain features. All extracted time-domain and frequency-domain feature parameters are merged into an initial feature parameter group. For each feature parameter in the initial feature parameter group, feature parameter normalization is performed to obtain standardized feature values. All normalized feature values ​​are arranged in a fixed order of time-domain and frequency-domain features to construct the corresponding standardized feature set.

[0026] S300: Retrieve the preset fault sound database, compare the standardized feature set with the standard normal feature set of the same model unit extracted from the database item by item, introduce the adaptation strategy, generate the corresponding working condition adaptation coefficient based on the comparison structure and then make corrections, and generate the feature matching benchmark library that is adapted to the current steam turbine unit under test. By retrieving the preset fault sound database, the standard normal feature set under normal operating conditions of the unit that matches the current turbine model under test is extracted from the preset fault database, as well as the standard fault feature set corresponding to the model unit, such as typical faults such as bearing wear, dynamic and static friction, and steam seal leakage. It should be noted that the preset fault sound database is obtained by filtering from the general database. The benchmark feature data that matches the current unit model is selected from the general database to eliminate feature deviations caused by differences in structure and parameters of different turbine models. Because of the inherent differences in operating characteristics between the current on-site units and the standard units of the same model in the database, the operating condition differences abstracted in this application are transformed into calculable correction coefficients. For example, the obtained standardized feature set is compared item by item with the standard normal feature set of the same model extracted from the database, the ratio of each feature parameter is calculated, and the weighted average of all ratios is taken to generate the operating condition adaptation coefficient corresponding to the unit. The product of all standard normal feature values ​​in the marked normal feature set and the corresponding operating condition mismatch coefficient is used as the adapted fault feature value. After item-by-item correction, it is classified and archived according to fault type, fault location and severity level. Each type of fault corresponds to a unique adapted fault feature set. At the same time, the measurement point location information of the current unit is supplemented to generate a feature matching benchmark library that is only applicable to the current steam turbine unit under test.

[0027] S400 intelligently compares the standardized feature set to be diagnosed with the feature matching benchmark library, completes the matching and identification of fault features through the convolutional neural network algorithm, and outputs the fault diagnosis results including fault location, fault type and cosine similarity value. Then, through the association mechanism, it distinguishes the fault root cause measurement point from the influence measurement point caused by the propagation of fault sound, and completes the defect location.

[0028] All adapted fault feature sets in the feature matching benchmark library are used as training benchmark samples for the convolutional neural network algorithm. After initializing the input layer parameters of the convolutional neural network algorithm, the number of neurons in the input layer is made to be completely consistent with the number of feature parameters in the standardized feature set. The standardized feature set under the working condition to be diagnosed is then input into the input layer of the initialized convolutional neural network algorithm as the input data to be identified. The convolutional layer performs feature depth extraction on the standardized feature set to be identified. In some other specific embodiments, this application uses three convolutional layers, each with 64 convolutional kernels, to perform sliding convolution calculation on the input standardized feature set, extract the correlation features between feature parameters, and generate a feature map. These correlation features can accurately reflect the unique feature patterns of different faults, which is different from the surface comparison of single feature parameters. Therefore, it can improve the algorithm's ability to identify weak fault features and composite fault features, and generate a feature map that can be used for accurate matching. The feature map output by the convolutional layer is then reduced in dimensionality by a pooling layer. Max pooling is used to retain the feature values ​​in the feature map except for redundant features. The dimensionality-reduced features output by the pooling layer are then matched one by one with each adapted fault feature set in the feature matching benchmark library using cosine similarity. This generates a cosine similarity value between the feature set to be identified and each fault feature set. The fault feature set corresponding to the highest cosine similarity value is taken as the diagnostic result. The fault location, fault type, and cosine similarity value of all measurement points in the fault feature set corresponding to the highest cosine similarity value are integrated to form the fault diagnosis result.

[0029] After uploading the diagnostic results of each measuring point through the communication module, the pre-stored turbine measuring point deployment location distribution map and spatial association weight matrix are retrieved. Therefore, this application summarizes the diagnostic data of all measuring points of the entire unit, and at the same time retrieves the spatial location information of the unit and the measuring point association weight data. The test points that indicate faults in the diagnostic results of each test point are selected and marked as fault-related test points. The fault type, cosine similarity value and test point location information of each fault-related test point are extracted. Based on the spatial association weight matrix, the propagation influence value of each fault-related test point on all other fault-related test points is calculated. It should be noted that the propagation impact value refers to the product of the cosine similarity value of the measuring point and the spatial correlation weight coefficient between the two measuring points, thereby quantifying the degree of sound propagation impact of each fault-related measuring point on other measuring points, and transforming the abstract sound propagation attenuation law into a calculable propagation impact value. For each fault-related measurement point, the difference between the corresponding feature value and the sum of the propagation influence values ​​of all other fault-related measurement points is calculated and used as the inherent fault feature value of the measurement point. Finally, the influence of the fault sound propagation of other measurement points on the current measurement point is removed, and the inherent fault feature value generated by the current measurement point itself is extracted. This value directly reflects whether the measurement point itself has a real fault. The higher the value, the more obvious the fault characteristics of the measurement point itself are. The inherent fault characteristic values ​​of all fault-related measurement points are sorted in descending order. The measurement point with the highest inherent fault characteristic value is identified as the root cause measurement point of the fault, and the remaining fault-related measurement points are identified as measurement points that affect the fault propagation. At the same time, a fault root cause location report is generated based on the spatial correlation weight matrix, and the defect location is completed through the correlation mechanism.

[0030] The pre-stored turbine measuring point deployment location distribution map and spatial correlation weight matrix include: By retrieving the factory structural drawings of the turbine under test, a base map of the main structure is constructed using proportional scaling and drawing software. The material properties and acoustic parameters of each component are entered into the base map, including the carbon steel material of the bearing housing, the alloy steel structure of the cylinder, the stainless steel material of the connecting pipes, as well as the longitudinal wave velocity, transverse wave velocity, and solid-borne sound attenuation coefficient of the corresponding materials. All deployed measuring points are marked in the base map, including bearing housing measuring points, steam seal measuring points, speed control valve measuring points, and coupling measuring points. Each measuring point is marked with a unique and non-repeatable measuring point number, and all measuring point attribute information is entered, namely the specific location of the measuring point installation, the turbine component corresponding to the measuring point, the base material type of the measuring point, the plane coordinates of the measuring point, the straight-line distance between adjacent measuring points, and the relative position of the measuring point to the core rotating components such as the turbine rotor, forming a measuring point deployment location distribution map. Retrieve the distribution map of the measurement point deployment locations, count the total number of actual deployed measurement points, and construct a two-dimensional matrix consistent with the total number of measurement points. The rows and columns of the matrix correspond to the unique measurement point numbers in the distribution map. The rows represent the sound source measurement points, and the columns represent the affected measurement points, that is, the measurement points that receive the sound of the fault propagation. The sound source measurement points are the measurement points where the fault occurs. The weight coefficients of the diagonal elements of the two-dimensional matrix are fixed at 1, which means that the influence weight of the fault characteristics of the measuring point on itself is 100%, with no propagation attenuation. The initial values ​​of all off-diagonal elements of the two-dimensional matrix are uniformly set to 0, thus forming an initial spatial correlation weight matrix framework that perfectly matches the number of measuring points. After calculating the sound propagation attenuation coefficient for each off-diagonal element of the two-dimensional matrix, the propagation efficiency between each measurement point is obtained, and the weight coefficients are filled into the spatial correlation weight matrix to complete the assignment.

[0031] This application further describes the construction and assignment of the matrix framework, including: For each off-diagonal element in the matrix, corresponding to a set of independent sound source measurement points A and affected measurement points B, perform the following calculation operations: From the distribution map of the measurement point deployment locations, extract the straight-line propagation distance L between the sound source measurement point A and the affected measurement point B, the type of metal medium on the propagation path, the segmented propagation distance corresponding to each medium, and the solid-borne sound transmission attenuation coefficient α of each medium; Calculate the total attenuation of sound propagation between two measuring points, i.e., total attenuation = attenuation coefficient of each segment of the medium along the propagation path × sum of the propagation distances of the corresponding segments + natural attenuation due to distance, where natural attenuation due to distance is inversely proportional to the square of the propagation distance between the two measuring points; The basic sound propagation efficiency between two measuring points is calculated based on the total attenuation. The basic propagation efficiency = 10^(-total attenuation / 20), thus strictly limiting the value range of the basic propagation efficiency to between 0 and 1. It represents the proportion of the remaining effective signal after the fault sound signal from the sound source measuring point is propagated through the solid to the affected measuring point.

[0032] Furthermore, in some other specific embodiments, the following modification steps are also included: For the same set of sound source measuring points A and affected measuring points B, perform the following correction operations: Based on the distribution map of the measurement point locations, the mechanical connection relationship between the corresponding components of two measurement points is determined and divided into four levels: rigid direct connection (e.g., adjacent steam seal measurement points on the same cylinder block, or adjacent measurement points on the same bearing housing); rigid indirect connection (e.g., a bearing housing and another bearing housing measurement point are rigidly connected through the rotor); flexible connection (e.g., a speed control valve measurement point and a cylinder block measurement point are flexibly connected through a pipeline); and no direct mechanical connection. Corresponding mechanical connection correlation coefficients are set for different connection types: a coefficient of 1.0 for rigid direct connection, 0.8 for rigid indirect connection, 0.3 for flexible connection, and 0.05 for no direct mechanical connection. These coefficients are determined based on solid-state sound transmission test data from industrial rotating machinery. Finally, the basic propagation efficiency is corrected using the mechanical connection correlation coefficients.

[0033] The modified propagation efficiency is determined as the spatial correlation weight coefficient corresponding to the sound source measurement point A and the affected measurement point B. The value range of the weight coefficient is strictly controlled between 0 and 1, which perfectly matches the numerical range of the initial matrix frame. The calculated weight coefficient is filled into the spatial correlation weight matrix at the intersection of the row corresponding to the sound source measurement point A and the column corresponding to the affected measurement point B, and the coefficient is assigned at that position. The operation is repeated to complete the calculation and filling of all off-diagonal elements in the matrix, and finally a complete spatial correlation weight matrix without blank values ​​is formed.

[0034] The present invention also provides an intelligent auscultation diagnostic device for steam turbines, the device comprising: Contact-type sound sensor; used to be attached to the surface of measuring points including turbine bearing housings, steam seals, speed control valves, and couplings to collect raw sound signals generated by mechanical vibrations inside the turbine. An ambient noise sensor; used to synchronously collect background noise signals in industrial settings surrounding steam turbines with a contact-type sound sensor, and to acquire raw sound signals. Embedded microprocessor; used as the core display carrier to display the time-domain waveform and frequency-domain spectrum of the turbine's operating sound signal in real time, and simultaneously present the diagnostic results of the measurement point location, operating status, fault location, and fault type. Wireless communication chip; used as the communication function carrier to support Bluetooth, Wi-Fi, and industrial IoT communication protocols, realize the uploading of raw sound data, analysis results, and measurement point location information, as well as network communication and centralized data transmission management. Magnetic fixing end; used to use strong magnetic materials to be adsorbed onto the surface of metal measuring points such as turbine bearing seats and cylinder blocks to achieve equipment installation and fixation, adapting to the stable deployment requirements of high-speed operation and strong vibration of the unit. Adhesive fixing end; used for installing and fixing equipment with high temperature and high viscosity industrial adhesive tape, suitable for non-magnetic material measuring points of steam turbines, forming a double fixing structure with magnetic fixing end, covering the installation and deployment of all types of measuring points of steam turbines; Rechargeable lithium battery; used as a built-in power supply unit to provide operating power; DC power interface; used for connecting an external industrial DC power supply to provide uninterrupted power supply, forming a dual power supply system with a rechargeable lithium battery.

Claims

1. A method for intelligent auscultation diagnosis of steam turbines, characterized in that, Includes the following steps: S100: Based on the original sound signals synchronously acquired by the contact sound sensor and the environmental noise sensor, and the timing alignment reference signal generated by the embedded microprocessor, dual effective sound signals are obtained; S200, based on time-aligned dual-channel effective sound signals, eliminates environmental noise interference through an adaptive filtering algorithm, and purifies to obtain effective operating signals containing only the mechanical vibrations inside the steam turbine. Then, it performs full-dimensional analysis in the time and frequency domains, extracts feature parameters through a fast Fourier transform algorithm, and constructs a standardized feature set that can be directly used for fault matching after completing the standardization of the feature parameters. S300: Retrieve the preset fault sound database, compare the standardized feature set with the standard normal feature set of the same model unit extracted from the database item by item, introduce the adaptation strategy, generate the corresponding working condition adaptation coefficient based on the comparison structure and then make corrections, and generate the feature matching benchmark library that is adapted to the current steam turbine unit under test. S400 intelligently compares the standardized feature set to be diagnosed with the feature matching benchmark library, completes the matching and identification of fault features through the convolutional neural network algorithm, and outputs the fault diagnosis results including fault location, fault type and cosine similarity value. Then, through the association mechanism, it distinguishes the fault root cause measurement point from the influence measurement point caused by the propagation of fault sound, and completes the defect location.

2. The intelligent auscultation diagnostic method for steam turbines according to claim 1, characterized in that, The S100 sends a timing alignment reference signal to the contact sound sensor and the environmental noise sensor at the same time that the contact sound sensor and the environmental noise sensor start collecting data through the embedded microprocessor. All the collected raw sound signals contain the timing alignment reference signal. The embedded microprocessor extracts the start time of the timing alignment reference signal from the original sound signal collected by the contact sound sensor and the start time of the timing alignment reference signal from the original sound signal collected by the environmental noise sensor, and calculates the time difference between the two start times. The embedded microprocessor performs time-domain shifting on the original audio signal with a time lag based on the time difference, so that the start times of the timing alignment reference signals of the two original audio signals completely coincide, thereby completing the timing alignment of the two signals and obtaining two valid audio signals.

3. The intelligent auscultation diagnostic method for steam turbines according to claim 1, characterized in that, S200 sets the ambient background noise signal collected by the environmental noise sensor in the dual-channel effective sound signal as the reference input signal of the adaptive filtering algorithm, and sets the original turbine operation sound signal collected by the contact sound sensor in the dual-channel effective sound signal as the main input signal of the adaptive filtering algorithm. A transverse filter is constructed based on the reference input signal. The initial weight coefficients of the filter are set to equal values. The reference input signal is input into the transverse filter to generate an analog noise output signal. The difference between the main input signal and the analog noise output signal is calculated to obtain the error signal. Based on the minimum mean square error criterion, with the goal of minimizing the error signal, the weight coefficients of the transverse filter are iteratively updated. After each iteration, the analog noise output signal and the error signal are recalculated until the mean square value of the error signal stabilizes. Then the iteration stops, and the purified effective operating signal of the steam turbine is obtained.

4. The intelligent auscultation diagnostic method for steam turbines according to claim 1, characterized in that, The S200 further includes: Based on the embedded microprocessor, the time-domain features of the effective operating signal of the steam turbine are extracted to obtain the time-domain feature parameters of the corresponding effective operating signal time-domain waveform, including time-domain peak value, time-domain mean, time-domain variance, kurtosis, and margin index. Then, the effective operating signal of the steam turbine in the time domain is converted into a spectrum signal in the frequency domain by using the Fast Fourier Transform algorithm. Specifically, the effective operating signal of the steam turbine is windowed, with the Hanning window as the window function. Then, the windowed effective operating signal is subjected to Fast Fourier Transform calculation to convert the time domain signal into a frequency domain signal, thus obtaining the frequency domain spectrum.

5. The intelligent auscultation diagnostic method for steam turbines according to claim 1, characterized in that, S200 further includes: extracting frequency domain feature parameters from the frequency domain spectrum, including frequency domain dimensions such as frequency peak, frequency centroid, frequency variance, main frequency amplitude, sideband amplitude, and harmonic amplitude; All extracted time-domain and frequency-domain feature parameters are merged into an initial feature parameter group. For each feature parameter in the initial feature parameter group, feature parameter normalization is performed to obtain standardized feature values. All normalized feature values ​​are arranged in a fixed order of time-domain and frequency-domain features to construct the corresponding standardized feature set.

6. The intelligent auscultation diagnostic method for steam turbines according to claim 1, characterized in that, The S300 retrieves a preset fault sound database and extracts the standard normal feature set under normal operating conditions of the unit that matches the model of the steam turbine being tested, as well as the standard fault feature set corresponding to the model unit. The obtained standardized feature set is compared item by item with the standard normal feature set of the same model unit extracted from the database. The ratio of each feature parameter is calculated, and the weighted average of all ratios is taken to generate the operating condition adaptation coefficient corresponding to the unit. The product of all standard normal feature values ​​in the marked normal feature set and the corresponding operating condition mismatch coefficient is used as the adapted fault feature value. After item-by-item correction, it is classified and archived according to fault type, fault location and severity level. Each type of fault corresponds to a unique adapted fault feature set. At the same time, the measurement point location information of the current unit is supplemented to generate a feature matching benchmark library that is only applicable to the current steam turbine unit under test.

7. The intelligent auscultation diagnostic method for steam turbines according to claim 1, characterized in that, The S400 also includes: All adapted fault feature sets in the feature matching benchmark library are used as training benchmark samples for the convolutional neural network algorithm. After initializing the input layer parameters of the convolutional neural network algorithm, the number of neurons in the input layer is made to be completely consistent with the number of feature parameters in the standardized feature set. The standardized feature set under the working condition to be diagnosed is then input into the input layer of the initialized convolutional neural network algorithm as the input data to be identified. The convolutional layer performs feature depth extraction on the standardized feature set to be identified, performs sliding convolution calculation on the input standardized feature set, extracts the correlation features between feature parameters, and generates a feature map. The feature map output by the convolutional layer is then reduced in dimensionality by a pooling layer. Max pooling is used to retain the feature values ​​in the feature map except for redundant features. The dimensionality-reduced features output by the pooling layer are then matched one by one with each adapted fault feature set in the feature matching benchmark library using cosine similarity. This generates a cosine similarity value between the feature set to be identified and each fault feature set. The fault feature set corresponding to the highest cosine similarity value is taken as the diagnostic result. The fault location, fault type, and cosine similarity value of all measurement points in the fault feature set corresponding to the highest cosine similarity value are integrated to form the fault diagnosis result.

8. The intelligent auscultation diagnostic method for steam turbines according to claim 7, characterized in that, After the S400 uploads the diagnostic results of each measuring point through the communication module, it retrieves the pre-stored turbine measuring point deployment location distribution map and spatial correlation weight matrix. The test points that indicate faults in the diagnostic results of each test point are selected and marked as fault-related test points. The fault type, cosine similarity value and test point location information of each fault-related test point are extracted. Based on the spatial association weight matrix, the propagation influence value of each fault-related test point on all other fault-related test points is calculated. For each fault-related measurement point, calculate the difference between the corresponding characteristic value and the sum of the propagation influence values ​​of all other fault-related measurement points, and use it as the inherent fault characteristic value of the measurement point; The inherent fault characteristic values ​​of all fault-related measurement points are sorted in descending order. The measurement point with the highest inherent fault characteristic value is identified as the root cause measurement point of the fault, and the remaining fault-related measurement points are identified as measurement points that affect the fault propagation. At the same time, a fault root cause location report is generated based on the spatial correlation weight matrix, and the defect location is completed through the correlation mechanism.

9. The intelligent auscultation diagnostic method for steam turbines according to claim 8, characterized in that, The turbine measuring point deployment location distribution map and spatial correlation weight matrix pre-stored in the S400 include: By retrieving the factory structural drawings of the steam turbine under test, a base map of the main structure is constructed in the drawing software using proportional scaling. The material properties and acoustic parameters of each component are entered into the base map of the main structure. All deployed measurement points are marked in the base map of the main structure, and the attribute information of all measurement points is entered to form a distribution map of the deployment location of measurement points. Retrieve the distribution map of the measurement point deployment locations, count the total number of actual deployed measurement points, and construct a two-dimensional matrix consistent with the total number of measurement points. The rows and columns of the matrix correspond to the unique measurement point numbers in the distribution map, where the rows represent the sound source measurement points and the columns represent the affected measurement points. The weight coefficients of the diagonal elements of the two-dimensional matrix are fixed at 1, and the initial values ​​of all the off-diagonal elements of the two-dimensional matrix are uniformly set to 0, thus forming an initial spatial correlation weight matrix framework that perfectly matches the number of measurement points. After calculating the sound propagation attenuation coefficient for each off-diagonal element of the two-dimensional matrix, the propagation efficiency between each measurement point is obtained, and the weight coefficients are filled into the spatial correlation weight matrix to complete the assignment.

10. A steam turbine intelligent auscultation diagnostic device, applied to the steam turbine intelligent auscultation diagnostic method according to any one of claims 1-9, characterized in that, The device includes: Contact-type sound sensor; used to be attached to the surface of measuring points including turbine bearing housings, steam seals, speed control valves, and couplings to collect raw sound signals generated by mechanical vibrations inside the turbine. An ambient noise sensor; used to synchronously collect background noise signals in industrial settings surrounding steam turbines with a contact-type sound sensor, and to acquire raw sound signals. Embedded microprocessor; used as the core display carrier to display the time-domain waveform and frequency-domain spectrum of the turbine's operating sound signal in real time, and simultaneously present the diagnostic results of the measurement point location, operating status, fault location, and fault type. Wireless communication chip; used as the communication function carrier to support Bluetooth, Wi-Fi, and industrial IoT communication protocols, realize the uploading of raw sound data, analysis results, and measurement point location information, as well as network communication and centralized data transmission management. Magnetic fixing end; used to attach the device to the surface of the measuring point using a strong magnetic material, thus achieving device installation and fixation; Adhesive fixing end; used for installing and fixing equipment with high temperature and high viscosity industrial adhesive tape, suitable for non-magnetic material measuring points of steam turbines, forming a double fixing structure with magnetic fixing end, covering the installation and deployment of all types of measuring points of steam turbines; Rechargeable lithium battery; used as a built-in power supply unit to provide operating power; DC power interface; used for connecting an external industrial DC power supply to provide uninterrupted power supply, forming a dual power supply system with a rechargeable lithium battery.

Images