Double-water internal cooling phase modifier fault diagnosis method and system fusing multi-dimensional information

Abstract

The invention discloses a multi-dimensional information fused dual-water internal cooling phase modifier fault diagnosis method and system, and the method comprises the steps: synchronously collecting voiceprint data, vibration data and output current data of a dual-water internal cooling phase modifier, and obtaining an original data set with time synchronization; performing hierarchical preprocessing on the original data set to generate preprocessed standardized data; fault sensitive features are extracted from the voiceprint data, the vibration data and the output current data, and a voiceprint feature vector, a vibration feature vector and an output current feature vector are constructed; performing dynamic weight distribution and fusion based on each feature vector to generate a fusion feature vector; and inputting the fusion feature vector into a CNN-LSTM model, extracting local spatial features through a convolutional neural network, capturing a time sequence evolution rule through a long and short-term memory network, and outputting a fault type and a corresponding confidence coefficient. The method has the advantages of high fault diagnosis precision and the like.

Description

Technical Field

[0001] This invention mainly relates to the field of synchronous condenser operation and maintenance technology, specifically to a fault diagnosis method and system for dual water-cooled synchronous condensers that integrates multi-dimensional information. Background Technology

[0002] As the power system transitions towards a higher proportion of renewable energy and greater power electronics, the demand for reactive power regulation in the power grid continues to increase. Synchronous condensers, as key reactive power compensation devices, directly determine the voltage stability and power supply reliability of the power grid. Synchronous condensers operate under high load and multiple interference conditions for extended periods, making them prone to three typical faults: loose screws, shaft asymmetry, and output current asymmetry. Failure to identify these faults in a timely manner can lead to a chain reaction of problems such as equipment shutdowns and grid voltage fluctuations. Therefore, efficient and accurate fault diagnosis technology has become a core requirement for ensuring the safe operation of synchronous condensers.

[0003] Existing synchronous condenser fault diagnosis technologies exhibit significant limitations when dealing with complex operating conditions: traditional diagnostic methods often rely on single-dimensional data, failing to fully integrate acoustic signature information directly related to the fault—such as friction noise from loose screws and bearing noise caused by shaft asymmetry, which are key indicators of early-stage latent faults. The single data dimension leads to a high rate of missed diagnoses for these faults. In recent years, domestic and international scholars have made partial optimizations to address these issues. Some studies have introduced acoustic signature feature extraction techniques such as MFCC and LPCC, or adopted deep learning models such as CNN and LSTM to improve diagnostic intelligence. However, significant shortcomings remain: schemes relying solely on acoustic signature or vibration data fail to achieve multi-source information synergy; while some studies have attempted multi-feature fusion, they mostly employ fixed weight allocation strategies, unable to dynamically adjust feature contribution based on fault type. Current research on synchronous condenser fault diagnosis focuses on three main areas: multi-source data synergy and fusion, the integration of intelligent algorithms with physical mechanisms, and the standardization of diagnostic procedures.

[0004] Existing solutions still fail to effectively address the entire chain of challenges, including "multi-source data interference suppression, dynamic feature weight allocation, accurate fault location, and model condition adaptation," and are insufficient to meet the actual needs of efficient operation and maintenance of synchronous condensers in complex power grid environments. Summary of the Invention

[0005] To address the technical problems existing in the prior art, this invention provides a method and system for accurate fault diagnosis of dual underwater cold-tuned cameras that integrates multi-dimensional information.

[0006] To solve the above-mentioned technical problems, the technical solution proposed by this invention is as follows: A fault diagnosis method for dual underwater cold-tuning cameras that integrates multi-dimensional information includes the following steps: The acoustic data, vibration data and output current data of the dual underwater cooling cameras are collected simultaneously to obtain a time-synchronized raw data set. The original dataset is subjected to hierarchical preprocessing, including targeted denoising and correction of acoustic print data, vibration data and output current data, and unified data dimensions to generate preprocessed standardized data. Based on standardized data, fault-sensitive features are extracted from acoustic data, vibration data, and output current data respectively, and acoustic feature vectors, vibration feature vectors, and output current feature vectors are constructed. Based on the fault type, dynamic weight allocation and fusion are performed on the acoustic signature feature vector, vibration feature vector and output current feature vector to generate a fused feature vector; the weights are dynamically adjusted according to the fault sensitivity. The fused feature vectors are input into the CNN-LSTM model, local spatial features are extracted through the convolutional neural network, and the temporal evolution pattern is captured through the long short-term memory network to output the fault type and corresponding confidence level.

[0007] Preferably, the hierarchical preprocessing of the original dataset specifically includes: The voiceprint data is subjected to wavelet threshold denoising, pre-emphasis processing, and frame-segmentation windowing processing. Wavelet denoising uses a db4 wavelet basis for three-level decomposition and soft thresholding. Pre-emphasis uses a 100Hz high-pass filter. Frame-segmentation windowing uses a 20ms frame length, a 10ms frame shift, and a Hanning window. The vibration data were subjected to low-pass filtering, power frequency notch filtering, and baseline correction. The low-pass filtering used a 4th-order Butterworth filter with a cutoff frequency of 500Hz. The power frequency notch filtering used an adaptive notch filter with a center frequency of 50Hz and a bandwidth of 2Hz. Baseline correction was achieved by subtracting the signal mean. The output current data is subjected to Kalman filtering, moving average processing, and phase correction; the moving average uses a 5-point moving average filter; the phase correction is based on the synchronously acquired phase information to align the phases of the three-phase currents.

[0008] Preferably, the specific process for extracting fault-sensitive features from acoustic signature data, vibration data, and output current data is as follows: Mel frequency cepstral coefficients (MFCC), linear prediction cepstral coefficients (LPCC), short-time energy, and zero-crossing rate are extracted from the voiceprint data to construct a 27-dimensional voiceprint feature vector; where the first 13 dimensions of MFCC are used, the first 12 dimensions of LPCC are used, and the short-time energy and zero-crossing rate are each 1 dimension. Frequency domain features and time domain features are extracted from vibration data; frequency domain features include the amplitude and peak frequency of the 50-200Hz fault frequency band; time domain features include peak-to-peak value and root mean square value, and a 4-dimensional vibration feature vector is constructed. The current imbalance, harmonic distortion rate (THD), statistical features, and phase deviation features are extracted from the output current data to construct a 6-dimensional output current feature vector. The statistical features include the maximum value of the mean and the maximum value of the standard deviation of the three-phase current, and the phase deviation features include the phase difference between phases AB and BC.

[0009] Preferably, the specific process for generating the fused feature vector is as follows: Initial weights are preset based on fault types; the fault types include loose screws, shaft asymmetry, and output current asymmetry. Calculate the fault sensitivity of each feature vector; The initial weights are dynamically adjusted based on fault sensitivity. The modified weights are used to sum the voiceprint feature vector, vibration feature vector, and output current feature vector to generate a fused feature vector.

[0010] Preferably, fault sensitivity The calculation formula is:

[0011] in For the current eigenvalue, These are characteristic values ​​under normal operating conditions.

[0012] Preferably, the specific process of dynamically adjusting the initial weights based on fault sensitivity is as follows:

[0013]

[0014]

[0015] in The corrected voiceprint weight; The corrected vibration weights; The corrected output current weight; where ; For voiceprint weighting, For vibration weights, Weighting for output current; For voiceprint fault sensitivity, For vibration fault sensitivity, For output current fault sensitivity.

[0016] Preferably, the CNN-LSTM model includes: The CNN layer contains two convolutional layers and two max pooling layers. The convolutional kernel size is 3×3, and the number of convolutional kernels is 37 and 74 respectively. The activation function is ReLU, and the max pooling kernel size is 2×2. The LSTM layer contains one hidden layer with 64 neurons, and the activation function is tanh. The output layer uses the Softmax activation function to output four types of results: normal, loose screw, asymmetrical shaft, and asymmetrical output current, along with their corresponding confidence levels.

[0017] The present invention also discloses a computer program product, comprising a computer program that, when executed by a processor, performs the steps of the method described above.

[0018] The present invention further discloses a computer-readable storage medium having a computer program stored thereon, the computer program executing the steps of the method described above when run by a processor.

[0019] The present invention also discloses a computer system including a memory and a processor interconnected thereon, wherein the memory stores a computer program that, when run by the processor, performs the steps of the method described above.

[0020] This invention further discloses a fault diagnosis system for a dual underwater cold-tuning camera that integrates multi-dimensional information, comprising: The data acquisition module synchronously acquires acoustic data, vibration data, and output current data from the dual underwater cooling cameras to obtain a time-synchronized raw data set. The preprocessing module is used to perform hierarchical preprocessing on the original dataset, including targeted denoising and correction of acoustic data, vibration data and output current data, and unifying data dimensions to generate preprocessed standardized data. The feature extraction module is used to extract fault-sensitive features from acoustic print data, vibration data, and output current data based on standardized data, and to construct acoustic print feature vectors, vibration feature vectors, and output current feature vectors. The feature fusion module is used to dynamically assign weights and fuse acoustic feature vectors, vibration feature vectors, and output current feature vectors based on fault type to generate a fused feature vector; the weights are dynamically adjusted according to fault sensitivity. The intelligent diagnostic module is used to input the fused feature vector into the CNN-LSTM model, extract local spatial features through the convolutional neural network, capture the temporal evolution pattern through the long short-term memory network, and output the fault type and corresponding confidence level.

[0021] Preferably, the hierarchical preprocessing of the original dataset in the preprocessing module specifically includes: The voiceprint data is subjected to wavelet threshold denoising, pre-emphasis processing, and frame-segmentation windowing processing. Wavelet denoising uses a db4 wavelet basis for three-level decomposition and soft thresholding. Pre-emphasis uses a 100Hz high-pass filter. Frame-segmentation windowing uses a 20ms frame length, a 10ms frame shift, and a Hanning window. The vibration data were subjected to low-pass filtering, power frequency notch filtering, and baseline correction. The low-pass filtering used a 4th-order Butterworth filter with a cutoff frequency of 500Hz. The power frequency notch filtering used an adaptive notch filter with a center frequency of 50Hz and a bandwidth of 2Hz. Baseline correction was achieved by subtracting the signal mean. The output current data is subjected to Kalman filtering, moving average processing, and phase correction; the moving average uses a 5-point moving average filter; the phase correction is based on the synchronously acquired phase information to align the phases of the three-phase currents.

[0022] Preferably, the specific process of extracting fault-sensitive features from acoustic signature data, vibration data, and output current data in the feature extraction module is as follows: Mel frequency cepstral coefficients (MFCC), linear prediction cepstral coefficients (LPCC), short-time energy, and zero-crossing rate are extracted from the voiceprint data to construct a 27-dimensional voiceprint feature vector; where the first 13 dimensions of MFCC are used, the first 12 dimensions of LPCC are used, and the short-time energy and zero-crossing rate are each 1 dimension. Frequency domain features and time domain features are extracted from vibration data; frequency domain features include the amplitude and peak frequency of the 50-200Hz fault frequency band; time domain features include peak-to-peak value and root mean square value, and a 4-dimensional vibration feature vector is constructed. The current imbalance, harmonic distortion rate (THD), statistical features, and phase deviation features are extracted from the output current data to construct a 6-dimensional output current feature vector. The statistical features include the maximum value of the mean and the maximum value of the standard deviation of the three-phase current, and the phase deviation features include the phase difference between phases AB and BC.

[0023] Preferably, the specific process of generating the fused feature vector in the feature fusion module is as follows: Initial weights are preset based on fault types; the fault types include loose screws, shaft asymmetry, and output current asymmetry. Calculate the fault sensitivity of each feature vector; The initial weights are dynamically adjusted based on fault sensitivity. The modified weights are used to sum the voiceprint feature vector, vibration feature vector, and output current feature vector to generate a fused feature vector.

[0024] Preferably, fault sensitivity The calculation formula is:

[0025] in For the current eigenvalue, These are characteristic values ​​under normal operating conditions.

[0026] Preferably, the specific process of dynamically adjusting the initial weights based on fault sensitivity is as follows:

[0027]

[0028]

[0029] in The corrected voiceprint weight; The corrected vibration weights; The corrected output current weight; where ; For voiceprint weighting, For vibration weights, Weighting for output current; For voiceprint fault sensitivity, For vibration fault sensitivity, For output current fault sensitivity.

[0030] Preferably, the CNN-LSTM model in the intelligent diagnostic module includes: The CNN layer contains two convolutional layers and two max pooling layers. The convolutional kernel size is 3×3, and the number of convolutional kernels is 37 and 74 respectively. The activation function is ReLU, and the max pooling kernel size is 2×2. The LSTM layer contains one hidden layer with 64 neurons, and the activation function is tanh. The output layer uses the Softmax activation function to output four types of results: normal, loose screw, asymmetrical shaft, and asymmetrical output current, along with their corresponding confidence levels.

[0031] Preferably, in the data acquisition module, the voiceprint data acquisition adopts an electret microphone linear array with a signal-to-noise ratio of not less than 60dB, a sampling frequency of 20-20000Hz, a sampling interval of 100ms, a single sampling duration of 200ms, and a storage format of WAV. Vibration data acquisition uses a piezoelectric vibration sensor with a sensitivity of 100mV / g, a measurement range of 0-50g, a sampling frequency of 1000-2000Hz, a sampling interval of 100ms, and synchronously records the sensor number. The output current data acquisition uses a Hall current sensor with an accuracy class of 0.5, a measurement range of 0-1.2 times the rated output current of the phase shifter, a sampling frequency of 100Hz, a sampling interval of 100ms, and outputs the instantaneous values ​​of the three-phase current and phase information. The acquisition time of acoustic print, vibration and output current data is synchronized, with a timestamp error of no more than 1ms.

[0032] Compared with the prior art, the advantages of the present invention are as follows: This invention innovatively integrates multi-dimensional data acquisition and intelligent analysis technologies of acoustic signature, vibration, and output current to construct an intelligent fault diagnosis system for dual underwater cooling cameras. This effectively overcomes the technical bottlenecks of traditional diagnostic methods, such as single data dimensions, rigid feature fusion, and insufficient diagnostic accuracy. Through a standardized multi-dimensional data synchronous acquisition scheme, it achieves comprehensive capture of key fault information. Combined with a hierarchical preprocessing and attention mechanism dynamic fusion (DFF) strategy, it significantly improves fault feature representation capabilities and greatly reduces the missed detection problems of traditional single-data diagnostic schemes.

[0033] This invention, in standardized multidimensional data synchronous acquisition, establishes a unified acquisition parameter and time synchronization mechanism for three types of data—acoustic fingerprint, vibration, and output current—for dual underwater cold-tuning cameras. Acoustic fingerprint acquisition utilizes an electret linear array with a signal-to-noise ratio of no less than 60dB and a sampling frequency of 20-20000Hz; vibration acquisition employs a piezoelectric sensor with a sensitivity of 100mV / g and a sampling frequency of 1000-2000Hz; and output current acquisition uses a Hall effect sensor with an accuracy class of 0.5 and a sampling frequency of 100Hz. The sampling interval for all three types of data is set to 100ms, with a timestamp error not exceeding 1ms. This scheme solves the feature distortion problems caused by data asynchrony and chaotic acquisition parameters in traditional diagnostics, laying a high-quality data foundation for subsequent multi-source fusion.

[0034] In the multi-dimensional data hierarchical preprocessing stage, this invention customizes processing strategies for different types of data interference characteristics: For acoustic fingerprint data, a combination of db4 wavelet 3-level decomposition, soft thresholding noise reduction, and 100Hz high-pass pre-emphasis is used; for vibration data, a 4th-order Butterworth low-pass filter, 50Hz adaptive notch filtering, and baseline correction are employed; and for output current data, Kalman filtering, 5-point moving average, and phase correction are used. After targeted processing, 16-bit quantization and min-max normalization unify the dimensions of the three types of data. This strategy can accurately filter out interference from computer room ventilation, power grid frequency, and sensor zero drift, while preserving fault characteristics to the maximum extent, such as high-frequency friction signals in acoustic fingerprints and low-frequency abnormal signals in vibration, effectively improving the accuracy of feature extraction.

[0035] In the feature extraction stage, this invention constructs a comprehensive feature dimension system covering both mechanical and electrical faults, extracting sensitive features from acoustic signature, vibration, and output current data respectively: Acoustic signature features extract Mel-frequency cepstral coefficients, linearly predicted cepstral coefficients, and time-domain short-time energy and zero-crossing rate, forming a comprehensive acoustic signature feature vector, significantly enhancing the ability to characterize the frequency characteristics and intensity changes of acoustic signatures; vibration features extract the amplitude and peak frequency of the fault frequency band, combined with peak-to-peak value and root mean square value; current features calculate current imbalance and harmonic distortion rate, combined with three-phase current statistical features and phase deviation features, improving the sensitivity of current asymmetry fault identification and enabling early detection of fault signs. Specifically, the acoustic signature feature vector includes 13-dimensional Mel-frequency cepstral coefficients, 12-dimensional linearly predicted cepstral coefficients, 1-dimensional short-time energy, and 1-dimensional zero-crossing rate, totaling 27 dimensions; the vibration feature vector includes 2-dimensional frequency domain features, 2-dimensional time domain features, and sensor position encoding, totaling 4 dimensions; the output current feature vector includes 1-dimensional current imbalance, 1-dimensional harmonic distortion rate, 2-dimensional statistical features, and 2-dimensional phase deviation features, totaling 6 dimensions. This system breaks through the limitations of traditional single-dimensional feature information and can achieve comprehensive characterization of early latent faults such as slight loosening of screws and obvious faults such as current asymmetry.

[0036] This invention improves the fault discriminativeness of fused features through dynamic weight allocation in fault adaptive attention mechanism feature fusion. First, initial weights are preset based on fault type; for example, when diagnosing a loose screw, the weights of acoustic and vibration features are emphasized, while when diagnosing current asymmetry, the weights of output current features are emphasized. Then, the fault sensitivity of each feature vector is calculated, derived from the percentage difference between the current feature value and the feature value under normal operating conditions. The initial weights are then dynamically adjusted based on the fault sensitivity to ensure that the sum of the weights for each dimension is 1 after adjustment. Finally, the acoustic, vibration, and output current feature vectors are weighted and summed according to the adjusted weights to obtain the fused feature vector. This method overcomes the shortcomings of traditional fixed-weight fusion, which cannot adapt to different fault types, reduces feature information entropy, and significantly improves feature discriminativeness.

[0037] This invention utilizes a CNN-LSTM intelligent diagnostic model, combining the advantages of extracting spatial local features and temporal series features to achieve high accuracy and real-time performance in fault diagnosis. The CNN layer comprises two convolutional layers and two max-pooling layers, with 3×3 kernels (37 and 74 respectively) and ReLU activation. The max-pooling layers, with 2×2 kernels, extract local fault patterns from fused features, such as acoustic peaks and abrupt changes in vibration amplitude. The LSTM layer consists of one hidden layer with 64 neurons, using tanh activation to capture the temporal evolution of fault features, such as changes in features from minor to severe faults. The output layer employs the Softmax activation function, outputting four categories of results: normal, loose screw, asymmetrical axis, and asymmetrical output current, along with their corresponding confidence levels. Under complex operating conditions, this model achieves an overall diagnostic accuracy of no less than 98%, with a single-sample diagnosis time not exceeding 120ms, meeting the requirements for real-time diagnosis.

[0038] In summary, this invention achieves accurate and rapid identification of faults in dual-water-cooled synchronous condensers through a full-process innovation involving multi-dimensional data acquisition, hierarchical preprocessing, dynamic feature fusion, and intelligent diagnosis. This effectively reduces equipment downtime, minimizes grid voltage fluctuations, extends the operation and maintenance cycle of the synchronous condensers, provides key technical support for the safe operation of reactive power compensation equipment in grids with a high proportion of renewable energy access, and significantly improves the voltage stability and power supply reliability of the power system. Attached Figure Description

[0039] Figure 1 The flowchart is shown in an embodiment of the fault diagnosis method for dual water-cooled cameras that integrates multi-dimensional information according to the present invention.

[0040] Figure 2 This is a structural diagram of the CNN-LSTM intelligent diagnostic model of the present invention in an embodiment.

[0041] Figure 3 This is a structural diagram of the LSTM layer in an embodiment of the present invention. Detailed Implementation

[0042] The present invention will be further described below with reference to the accompanying drawings and specific embodiments.

[0043] like Figure 1 As shown in the figure, the fault diagnosis method for dual underwater cold-tuning cameras that integrates multi-dimensional information provided in this embodiment of the invention includes the following specific steps: S1. Synchronously acquire acoustic data, vibration data and output current data of the dual underwater cooling cameras to obtain a time-synchronized raw data set; Multidimensional data synchronous acquisition is a fundamental step in fault diagnosis. Acquisition equipment needs to be scientifically deployed in key parts of the synchronous condenser to ensure the synchronization and integrity of the three core data types: acoustic signature, vibration, and output current, so as to provide high-quality raw data support for subsequent diagnostic work.

[0044] In terms of voiceprint data acquisition, a linear array composed of electret microphones is used, with a signal-to-noise ratio of ≥60dB. This high signal-to-noise ratio characteristic can effectively reduce the interference of background noise on the voiceprint signal and ensure that the acquired voiceprint information can truly reflect the operating status of the camera condenser. The installation location should be chosen in areas with acoustic radiation intensity ≥50dB, such as the bearing end cover and densely connected screw areas of the phase condenser housing (the weakest acoustic signal is about 50-65dB; the background noise at the phase condenser operation site (ventilation, power grid interference) is usually stable at 30-45dB; the difference between the fault acoustic signal intensity and the background noise intensity should be at least 5dB (50dB-45dB=5dB). Combined with the "signal superposition enhancement and noise cancellation" capability of the linear array (which can improve the signal-to-noise ratio by 55-60dB), the "total signal-to-noise ratio ≥60dB" can be easily met). These areas are the main radiation sources of fault acoustic signals, maximizing the capture of fault-related acoustic information. The acquisition parameters should be set to a sampling frequency of 20-20000Hz (for loose screw friction noise, mainly concentrated in...). The frequency range of 500Hz to 10kHz is used for bearing noise caused by axial asymmetry (mainly concentrated in the range of 1kHz to 20kHz). This frequency range covers the frequency band of acoustic signature signals generated by common faults in synchronous condensers. The sampling interval is 100ms (the fastest fault acoustic signature pulse is 50ms. A single effective signal lasts for tens to hundreds of milliseconds: slight loosening of screws results in "intermittent short pulses" of friction, each lasting about 50 to 150ms; bearing noise caused by axial asymmetry is "continuous, but the frequency / intensity changes slowly within one second," without "millisecond-level sudden changes"). The single sampling duration is 200ms, which ensures both real-time data acquisition and avoids data redundancy. The storage format is WAV, which is a lossless audio format that can completely preserve the original characteristics of the acoustic signature signal.

[0045] Vibration data acquisition utilizes piezoelectric vibration sensors with a sensitivity of 100mV / g and a measurement range of 0-50g. These sensors can accurately detect minute vibration changes during the operation of the synchronous condenser and promptly capture vibration anomalies caused by faults. During installation, the sensors are magnetically attached to the surfaces of the end cover screws and base fixing screws of the synchronous condenser. These locations are closely connected to the core moving parts of the synchronous condenser, and the vibration signals directly reflect the operating status of these components. The acquisition parameters are set to a sampling frequency of 1000-2000Hz (vibration frequency band for loose screws (800-1500Hz); vibration frequency band for axial asymmetry (1200-1800Hz)). This frequency effectively acquires vibration signals related to mechanical faults in the synchronous condenser. The sampling interval is 100ms, consistent with the acoustic fingerprint data sampling interval, ensuring time synchronization of multi-dimensional data. Sensor numbers are recorded synchronously.

[0046] Output current data acquisition utilizes a Hall effect current sensor with an accuracy class of 0.5. The measurement range is 0-1.2 times the rated output current of the synchronous condenser. Its high precision ensures accurate current data, and its wide measurement range adapts to current variations under different load conditions of the synchronous condenser. The installation method is series connection to the three-phase output circuit. A , B , C (Phase), after setting the sampling frequency to 100Hz and the sampling interval to 100ms after the input circuit phase reference, output the instantaneous value of the three-phase current ( i A , i B , i C ) and phase information ( φ A , φ B , φ C ).

[0047] The final output of the entire acquisition process is a set of three types of raw data that are synchronized in time.

[0048] S2. Perform hierarchical preprocessing on the original dataset, including targeted denoising and correction of acoustic data, vibration data and output current data, and unify data dimensions to generate preprocessed standardized data; The raw data inevitably contains various interference signals, such as noise from the computer room ventilation, electromagnetic interference from the power grid, and sensor zero drift. These interferences can mask fault characteristics and affect the accuracy of subsequent diagnostic results. Therefore, targeted layered preprocessing of acoustic signature, vibration, and output current data is necessary to eliminate interference and unify data dimensions, laying a solid foundation for feature extraction.

[0049] The preprocessing of the voiceprint data first employs wavelet threshold denoising (separating the high-frequency components where noise is concentrated from the high- and low-frequency components where fault features are concentrated, laying the foundation for subsequent "targeted denoising"), and then uses the db4 wavelet basis to perform a three-level decomposition of the voiceprint signal (splitting the original voiceprint signal from "high frequency to low frequency" into three high-frequency detail components). d 1. d 2. d 3) and one low-frequency approximation component ( a 3) The audio signal is decomposed into four independent components: "pure noise > 5kHz", "screw fault characteristics 2~5kHz", "shaft fault characteristics 0.5~2kHz", and "useless background noise < 0.5kHz". This 3-layer decomposition can break down the audio signal into coefficients of different frequency bands, making it easier to process noise in different frequency bands.

[0050] Soft thresholding is applied to high-frequency coefficients. λ The calculation formula is: (1) in The standard deviation of noise. N The signal length; When the absolute value of a certain high-frequency coefficient is < λ When a signal is identified as "pure noise figure" (such as a random small electromagnetic interference signal), it is directly set to 0 according to the soft threshold rule to completely filter out this type of noise. When the absolute value of a certain high-frequency coefficient is greater than λ When: If it is determined to be an "effective coefficient containing fault characteristics" (such as a high-frequency signal generated by loose screws), the "coefficient value -" is calculated according to the soft threshold rule. λ The smooth correction retains the key information reflecting the fault state; This threshold calculation method can be adaptively adjusted according to the signal noise level, effectively filtering out environmental noise such as computer room ventilation and power grid electromagnetic interference, while preserving fault characteristics in the voiceprint signal to the greatest extent.

[0051] Then, pre-emphasis processing is performed. The high-frequency fault features of the voiceprint are enhanced by a 100Hz high-pass filter (100Hz is the optimal threshold value for "filtering out useless low-frequency noise" and "fully preserving high-frequency fault features". Voiceprint signals below 100Hz are mostly "low-frequency background noise without fault significance" (such as the rotation sound of computer room ventilation fans, the humming sound of the 50Hz power frequency and its harmonics of the power grid)). For example, the high-frequency components generated by screw friction. These high-frequency components are often important early manifestations of faults and are more easily captured by subsequent feature extraction stages after enhancement.

[0052] Finally, frame segmentation and windowing are performed (the purpose of "windowing" is to solve the "spectral leakage" problem of the signal after framing). The processed signal is framed with a frame length of 20ms (the minimum period of the fault feature is about 10~15ms) and a frame shift of 10ms (to ensure 10ms overlap). The 20ms frame length ensures that each frame contains sufficient fault feature information, and the 10ms frame shift avoids loss of feature information. A Hanning window is superimposed on each frame, and the window function is... Specifically: (2) N The frame length.

[0053] The Hanning window can effectively reduce abrupt changes in signal at frame edges, avoid spectral leakage, and ensure the accuracy of subsequent spectral analysis.

[0054] The first step in vibration data preprocessing is low-pass filtering, using a Butterworth low-pass filter (cutoff frequency 500Hz, 4th order (20dB / decimal frequency per order, 80dB / decimal frequency attenuation rate designed for the 4th order). Excessive attenuation at higher orders can cause frame loss in mid-to-low frequency signals, while slower attenuation at lower orders preserves high-frequency signals. The 4th order design ensures effective filtering, and the 500Hz cutoff frequency filters out high-frequency vibration interference generated during the operation of the synchronous condenser, retaining mid-to-low frequency vibration signals related to mechanical faults.

[0055] The second step involves power frequency notch filtering, using an adaptive notch filter (center frequency 50Hz, bandwidth 2Hz) to eliminate the coupling interference of the power grid frequency to the vibration signal. Power grid frequency interference is quite common in vibration signals, and if it is not eliminated, it will affect the extraction of fault characteristics. The adaptive notch filter can automatically adjust the filtering parameters according to changes in the interference signal to ensure the stability of the filtering effect.

[0056] The third step is to perform baseline correction and calculate the vibration signal. mean : (3) Subtract the signal μ This eliminates the influence of sensor zero drift. Sensor zero drift causes a fixed deviation in the vibration signal, affecting the judgment of the actual vibration amplitude. Baseline correction allows the vibration signal to accurately reflect the vibration state of the synchronous condenser.

[0057] The output current data preprocessing first employs Kalman filtering to establish the current signal state equation and observation equation; the state equation is as follows: (4) The observation equation is: (5) inA The state transition matrix ( A =1), H For the observation matrix ( H =1), w k , v k These are process noise and observation noise, respectively.

[0058] Kalman filtering can effectively smooth current fluctuations and eliminate random noise in current signals by predicting and updating the state, making the current signal closer to the true value.

[0059] Next, a moving average processing is performed, using a 5-point moving average filter to further eliminate instantaneous interference, such as sudden current changes caused by instantaneous fluctuations in the power grid, making the current signal more stable.

[0060] The formula for 5-point moving average filtering is: (6) in y k For the filtered output signal at k The value at each moment; z k The original input signal at the 1st k The sampled values ​​at each time point; z k-2 , z k-1 These are the current times. k The sampled values ​​from the second and first time points; z k-2 , z k-1 These are the current times. k The sampled values ​​at the first and second subsequent time intervals; Finally, phase correction is performed. Based on the synchronously acquired phase information, the three-phase currents are phase-aligned. The accuracy of the three-phase current phase directly affects the calculation of characteristics such as current imbalance. Phase correction ensures the accuracy of current characteristic calculations, providing a reliable basis for subsequent fault diagnosis.

[0061] To ensure the compatibility of multidimensional data in subsequent fusion and diagnostic stages, the preprocessed acoustic signature, vibration, and output current data need to be standardized. The data is converted into 16-bit digital signals (16-bit data avoids both the "insufficient precision of 8-bit" and the "excessive storage pressure caused by 32-bit data"). 16-bit precision meets the data accuracy requirements of fault diagnosis while controlling the data size.

[0062] Use min-max normalization: (7) x The data is after normalization; x This is the original data; The minimum value in the original data; The maximum value in the original data; Mapping data to the [0,1] interval eliminates the impact of differences in units and numerical ranges between different dimensions of data, ensuring the consistency of multidimensional data dimensions and laying the foundation for feature fusion.

[0063] S3. Based on standardized data, fault-sensitive features are extracted from acoustic fingerprint data, vibration data, and output current data respectively, and acoustic fingerprint feature vector, vibration feature vector, and output current feature vector are constructed. Feature extraction is the core step in fault diagnosis. It is necessary to extract sensitive features that can effectively characterize the fault state from the preprocessed acoustic signature, vibration, and output current data, and construct a multi-dimensional feature vector to provide key information support for subsequent fault identification.

[0064] Voiceprint feature extraction mainly includes Mel frequency cepstral coefficients (MFCC) (which can filter out high-frequency redundancy unrelated to faults and focus on "low-frequency voiceprint changes" caused by phase shifter faults (such as bearing wear and component loosening); the final 13-dimensional coefficients can also be reduced in dimensionality by DCT to retain the core feature of "spectral envelope" (the offset of the spectral envelope directly reflects the fault state) and avoid high-dimensional interference of the original signal), linear predictive cepstral coefficients (LPCC) (which are highly sensitive to "minor fluctuations" in voiceprints - early faults in phase shifters (such as slight loosening of screws and minor wear of bearing balls) will cause subtle changes in the channel propagation characteristics of voiceprints. These changes cannot be captured by the macroscopic spectrum of the original signal or MFCC, but can be accurately identified by LPCC coefficients to achieve "early fault warning") and time domain statistical features. In the MFCC extraction process, the voiceprint signal is passed through a Mel filter bank. The frequency response of the Mel filter bank conforms to the characteristics of human hearing and can better capture the auditory features in the voiceprint signal. The energy of each filter bank is calculated, and after logarithmic transformation, discrete cosine transform (DCT) is performed on the energy. The first 13 dimensions (the 13-dimensional feature vector has a moderate dimension, which can meet the requirements of "feature accuracy" for camera fault diagnosis, and will not cause excessive computation of subsequent algorithms (such as machine learning models) due to excessive dimensionality, thus ensuring real-time performance) DCT coefficients are taken as MFCC features. These 13-dimensional features can effectively characterize the spectral envelope information of the voiceprint signal and are important features for voiceprint recognition and fault diagnosis. LPCC extraction uses a 12th-order model (in the acoustic signature spectrum during phase shifter operation, there are approximately 6-8 key resonance peaks (mainly distributed in the 200-5000Hz frequency band, covering low-frequency and mid-to-high-frequency components related to faults); the order of the linear prediction model needs to satisfy "order ≈ 2 × number of resonance peaks" to fully fit the "frequency, amplitude, and bandwidth" characteristics of the resonance peaks. If the order is too low (e.g., 8th order), only 4-5 resonance peaks can be fitted, resulting in the loss of fault information from high resonance peaks; if the order is slightly higher than "2 × number of resonance peaks" (e.g., 12th order), it can fully cover 6-8 key resonance peaks while avoiding "model underfitting" due to insufficient order, ensuring... The voiceprint generation process is precisely simulated. Linear Predictive Analysis (LPC) uses a linear prediction model to predict future sampled values ​​based on past sampled values. The calculated prediction coefficients reflect the vocal tract characteristics of the voiceprint signal. Cepstral transform is applied to the prediction coefficients, and the first 12 dimensions (which accurately preserve the core information sensitive to subtle changes in faults, meeting the requirement of "capturing changes in voiceprint characteristics caused by faults") are used to control the feature dimensions and adapt to the input requirements of subsequent fault diagnosis algorithms. The LPCC feature is highly sensitive to subtle changes in the voiceprint signal and can effectively capture changes in voiceprint characteristics caused by faults. Temporal statistical features are used to calculate the voiceprint signal for each frame. short-term energy and zero crossing rate : (8) (9) Short-time energy reflects changes in the intensity of the voiceprint signal; abnormal fluctuations in voiceprint energy often accompany fault occurrences. Zero-crossing rate represents the number of times the voiceprint signal crosses the zero level per unit time, reflecting the frequency characteristics of the voiceprint signal; frequency changes in the voiceprint caused by a fault can be reflected through the zero-crossing rate. Based on these characteristics, a 27-dimensional voiceprint feature vector is constructed. (13-dimensional MFCC + 12-dimensional LPCC + 1-dimensional short-time energy + 1-dimensional zero-crossing rate) comprehensively characterizes fault information in the voiceprint signal.

[0065] Vibration feature extraction is conducted in both the frequency and time domains. In frequency domain feature extraction, a Fast Fourier Transform (FFT) is performed on the vibration signal to convert the time-domain signal into a frequency-domain signal. The amplitude of the fault frequency band (50-200Hz) is extracted (the vibration frequency of loose screws is mostly between 50-150Hz; the axis rotation frequency of an asymmetrical phase shifter is typically 50-100Hz; other vibration frequencies are also mostly concentrated in the 50-200Hz range). A f and peak frequency f p This frequency band is the main frequency band causing abnormal vibrations due to common mechanical faults in synchronous condensers (such as loose screws or shaft asymmetry). A f Amplitude reflects the intensity of fault vibration, peak frequency f p It can indicate characteristic frequencies related to faults, providing a basis for fault type judgment.

[0066] Time-domain characteristic calculation of peak-to-peak value of vibration signal and root mean square value : (10) (11) Peak-to-peak value is the maximum value of the vibration signal. and minimum value The difference between the peak and peak values ​​directly reflects the maximum amplitude of the vibration; the severity of the fault is often related to the peak-to-peak value. The root mean square value is the vibration signal... The effective value of the vibration signal comprehensively reflects the energy level of the vibration signal and is an important indicator for assessing vibration intensity. A 4-dimensional vibration feature vector is constructed by integrating frequency domain and time domain characteristics. (2-dimensional frequency domain features + 2-dimensional time domain features), which includes fault feature information.

[0067] Output current feature extraction focuses on current imbalance, harmonic distortion (THD), and statistical characteristics, combined with phase deviation information. Current imbalance... During calculation, the positive sequence current is calculated based on the three-phase current. and negative sequence current The formula is: (12) During normal operation, the three-phase current of the synchronous condenser is basically balanced, and the negative sequence current is relatively small. When an output current imbalance fault or other related faults occur, the negative sequence current will increase, and the current imbalance will increase accordingly. This indicator is the key basis for judging whether the current is normal and the severity of the fault.

[0068] When calculating the harmonic distortion rate (THD), a Fast Fourier Transform (FFT) is performed on the single-phase current to calculate: (13) This is the effective value of the fundamental current. The effective value of each harmonic current; Power grid harmonics can adversely affect the operation of synchronous condensers, and faults may also lead to an increase in harmonic content. THD can reflect the degree of harmonic pollution of current signals and provide supplementary information for fault diagnosis.

[0069] Statistical characteristics are used to calculate the mean of three-phase currents. ) and standard deviation Take the maximum value of the mean. Maximum value of standard deviation The mean reflects the average level of the current, while the standard deviation reflects the degree of fluctuation in the current. The mean and fluctuation of the current may change when a fault occurs. and It can comprehensively reflect the overall statistical characteristics of three-phase current. The phase deviation characteristics are extracted by analyzing the phase relationship of the three-phase current, which is two-dimensional (the phase difference between AB is the first dimension, the phase difference between BC is the second dimension, and AC can be derived from the first two, i.e., AB+BC+AC=0). The phase deviation information can help determine the specific cause of current asymmetry fault.

[0070] Based on the above characteristics, a 6-dimensional output current feature vector is constructed. (1-dimensional current imbalance + 1-dimensional THD + 1-dimensional) +1 dimension (+2D phase deviation characteristics) comprehensively reflect fault information in the current signal.

[0071] S4. Dynamically assign and fuse acoustic signature feature vector, vibration feature vector and output current feature vector based on fault type to generate fused feature vector; where the weights are dynamically adjusted according to fault sensitivity. Features of different dimensions have varying importance in fault diagnosis, and their importance changes with the type and severity of the fault. Traditional fixed-weight feature fusion methods struggle to adapt to this dynamic change, resulting in poor fusion performance. Therefore, an attention mechanism is employed for feature fusion. Through a dynamic weight allocation strategy, important features are given higher weights during the fusion process, thereby enhancing the fault representation capability of the fused features.

[0072] In the feature weight initialization phase, initial weights are preset based on the characteristics of common fault types. For screw loosening faults, the fault characteristics are mainly reflected in sound and vibration, with the output current being less affected; therefore, a sound weight is set. =0.45, vibration weight =0.4, output current weighting =0.15. Axial asymmetry faults will cause significant vibration abnormalities and generate specific acoustic signature signals. The output current change is relatively small, therefore the acoustic signature weight is set to 0.15. =0.4, vibration weight =0.45, output current weighting =0.15. The core characteristics of output current asymmetry faults are concentrated in the current signal, while the acoustic and vibration signals are less affected; therefore, the acoustic weight is set to 0.15. =0.1, vibration weight =0.2, output current weighting =0.7. The initial weight setting provides a reasonable starting point for subsequent dynamic adjustments, ensuring that key characteristics can be initially highlighted under different fault types.

[0073] During the dynamic weight adjustment phase, the initial weights are corrected by calculating the fault sensitivity of each feature vector. Fault sensitivity The calculation formula is: (14) in f For the current eigenvalue, f 0 represents the characteristic value under normal operating conditions. Fault sensitivity. S The larger the value, the more significant the difference between the feature vector and the normal operating condition, and the richer the fault information it contains. Therefore, it should be given a higher weight during the fusion process.

[0074] The formula for adjusting the weight based on fault sensitivity is: (15) Similarly, the corrected vibration weights can be calculated. and output current weight And guarantee: (16) This dynamic adjustment method can optimize the weight allocation in real time according to the actual fault situation, so that the fusion features always highlight the information most sensitive to the current fault, thereby improving the adaptability and accuracy of the fusion effect.

[0075] In the feature vector construction stage, the voiceprint feature vector is fused. Vibration characteristic vector Output current characteristic vector The 37-dimensional fused feature vector is obtained by weighted summation based on the corrected weights. (17) This fusion feature vector integrates key fault information from three dimensions: acoustic signature, vibration, and output current. The weights of each dimension are dynamically adjusted based on fault sensitivity, enabling a more comprehensive and accurate characterization of the synchronous condenser's fault status and providing high-quality input features for subsequent intelligent diagnosis.

[0076] S5. Input the fused feature vector into the CNN-LSTM model, extract local spatial features through the convolutional neural network, and capture the temporal evolution pattern through the long short-term memory network to output the fault type and corresponding confidence level.

[0077] CNN (Convolutional Neural Network) excels at extracting local spatial features from data, while LSTM (Long Short-Term Memory Network) can effectively capture the temporal evolution of data. Combining the two to build a CNN-LSTM intelligent diagnostic model can fully leverage the advantages of both networks to achieve accurate identification and localization of camera fault types.

[0078] In the model architecture, the CNN layer is responsible for extracting local fault patterns from the fused features. This layer contains two convolutional layers (first capturing basic local fault patterns from the fused features; then further extracting finer local fault details from the basic features) and two max-pooling layers (corresponding to the two convolutional layers, performing pooling to compress the data). The convolutional kernel size is 3×3, with 37 and 74 kernels respectively, and the ReLU activation function is used. The 3×3 convolutional kernel (for vocalization peaks, there needs to be a peak and two adjacent points; for vibration amplitude, two adjacent points of the abrupt change point) can effectively capture local features. The increasing number of convolutional kernels of 32 and 64 allows for progressively refined feature extraction. The ReLU activation function has non-linear transformation capabilities, which can enhance the model's expressive power and avoid the gradient vanishing problem. The max-pooling layer uses a 2×2 kernel size (this 2×2 size ensures that peak information related to the fault is not missed when "taking the maximum" while avoiding loss of local fault distribution details due to an excessively large range by covering a local range of two consecutive feature points). By downsampling the output feature map of the convolutional layer, the amount of data is reduced, computational complexity is lowered, key features are preserved, and the generalization ability of the model is improved. The CNN layer can effectively extract local fault patterns from the fused features, such as peak values ​​of the acoustic signature at specific frequencies and sudden changes in vibration amplitude, providing local feature support for fault type judgment.

[0079] The LSTM layer is primarily used to capture the temporal evolution trend of fault features. It contains one hidden layer with 64 neurons and uses the tanh activation function. The LSTM network effectively handles long-sequence data and avoids long-term dependencies through its gating mechanism (input gate, forget gate, output gate). The 64 neurons ensure the model's ability to capture temporal features. The tanh activation function maps the output value to the [-1, 1] interval, making it suitable for handling fluctuations in temporal data. The LSTM layer can capture the changing patterns of fault features over time, such as the progression from slight to severe loosening of screws, providing temporal feature support for fault type identification and fault development trend analysis.

[0080] LSTM layer structure as follows Figure 3 As shown, the output layer uses the Softmax activation function to output four types of results (normal, loose screw, asymmetrical shaft, and asymmetrical output current) and their corresponding confidence levels. P The Softmax function can convert the model output into a probability distribution form, and the probability value corresponding to each category is the confidence level, which makes it easy to intuitively judge the fault type and the reliability of the diagnosis results.

[0081] Figure 2This diagram illustrates the CNN-LSTM model structure for camera condenser fault diagnosis, clearly showing the complete core process from fused feature input to fault type output. The model input is a multi-dimensional feature vector fused through an attention mechanism. This vector integrates three core fault features from the camera condenser's operation: acoustic signature, vibration, and current. This ensures the input information comprehensively covers possible fault-related signals, providing a sufficient data foundation for subsequent diagnosis. The core CNN module contains two progressive structures: "convolutional layer + pooling layer." The first set, "convolutional layer 1 + pooling layer 1," performs preliminary processing on the fused features, focusing on extracting local correlation patterns, such as capturing the spatial correlation between acoustic signature peaks and vibration amplitude abrupt changes, and the matching relationship between abnormal current features and specific vibration frequency bands. The second set, "convolutional layer 2 + pooling layer 2," further compresses the data volume while strengthening key fault features and filtering redundant information, achieving precise focusing on local fault patterns. Subsequently, the LSTM module (containing one hidden layer) receives local features output from the CNN module. Leveraging its temporal memory capability, it analyzes the evolution of fault features over time. For example, it captures the continuous changing trends of features such as acoustic energy, vibration amplitude, and current imbalance as the fault progresses from a minor to a severe state, compensating for the limitations of CNN in processing temporal features. Finally, the Softmax output layer performs classification operations on the features processed by the LSTM module, outputting specific fault types (including normal, axial asymmetry, loose screws, and asymmetrical output current), providing a clear basis for maintenance personnel to judge the operating status of the synchronous condenser and formulate maintenance strategies, thus fully realizing the intelligent diagnostic closed loop for synchronous condenser faults.

[0082] This model combines the advantages of spatial and temporal feature extraction to achieve high accuracy and real-time diagnosis of synchronous condenser faults.

[0083] Through the above complete fault diagnosis process, the accurate identification and location of common faults in dual water-cooled synchronous condensers can be achieved, providing strong technical support for the maintenance and safe and stable operation of synchronous condensers, effectively reducing downtime due to faults, and improving the reliability and stability of the power system.

[0084] This invention innovatively integrates multi-dimensional data acquisition and intelligent analysis technologies of acoustic signature, vibration, and output current to construct an intelligent fault diagnosis system for dual underwater cooling cameras. This effectively overcomes the technical bottlenecks of traditional diagnostic methods, such as single data dimensions, rigid feature fusion, and insufficient diagnostic accuracy. Through a standardized multi-dimensional data synchronous acquisition scheme, it achieves comprehensive capture of key fault information. Combined with a hierarchical preprocessing and attention mechanism dynamic fusion (DFF) strategy, it significantly improves fault feature representation capabilities and greatly reduces the missed detection problems of traditional single-data diagnostic schemes.

[0085] In standardized multidimensional data synchronous acquisition, a unified acquisition parameter and time synchronization mechanism was established for three types of data—acoustic fingerprint, vibration, and output current—for dual underwater cooling cameras. Acoustic fingerprint acquisition utilizes an electret linear array with a signal-to-noise ratio of no less than 60dB and a sampling frequency of 20-20000Hz; vibration acquisition employs a piezoelectric sensor with a sensitivity of 100mV / g and a sampling frequency of 1000-2000Hz; and output current acquisition uses a Hall effect sensor with an accuracy class of 0.5 and a sampling frequency of 100Hz. The sampling interval for all three types of data is set to 100ms, with a timestamp error not exceeding 1ms. This scheme solves the feature distortion problems caused by data asynchrony and chaotic acquisition parameters in traditional diagnostics, laying a high-quality data foundation for subsequent multi-source fusion.

[0086] In the multi-dimensional data hierarchical preprocessing stage, customized processing strategies were implemented for different types of data interference characteristics: acoustic fingerprint data underwent a combination of 3-level decomposition using db4 wavelet, soft thresholding denoising, and 100Hz high-pass pre-emphasis; vibration data employed a 4th-order Butterworth low-pass filter, 50Hz adaptive notch filtering, and baseline correction; and output current data utilized Kalman filtering, 5-point moving average, and phase correction. After targeted processing, 16-bit quantization and min-max normalization were used to unify the dimensions of the three data types. This strategy effectively filters out interference from computer room ventilation, power grid frequency, and sensor zero drift, while preserving fault characteristics to the maximum extent, such as high-frequency friction signals in acoustic fingerprints and low-to-mid-frequency abnormal signals in vibration, effectively improving the accuracy of feature extraction.

[0087] In the feature extraction stage, a full feature dimension system covering mechanical and electrical faults is constructed, extracting sensitive features from acoustic signature, vibration, and output current data respectively: acoustic signature features extract Mel frequency cepstral coefficients, linear prediction cepstral coefficients, and time-domain short-time energy and zero-crossing rate to form a comprehensive acoustic signature feature vector, which greatly enhances the ability to characterize the frequency characteristics and intensity changes of acoustic signatures; vibration features extract the amplitude and peak frequency of the fault frequency band, and combine peak-to-peak value, root mean square value, and sensor position coding to achieve accurate fault location; current features calculate current imbalance and harmonic distortion rate, and combine with three-phase current statistical features and phase deviation features to improve the sensitivity of current asymmetry fault identification and detect fault signs in advance. Specifically, the acoustic signature feature vector includes 13-dimensional Mel frequency cepstral coefficients, 12-dimensional linear prediction cepstral coefficients, 1-dimensional short-time energy, and 1-dimensional zero-crossing rate, totaling 27 dimensions; the vibration feature vector includes 2-dimensional frequency domain features, 2-dimensional time domain features, and sensor position encoding, totaling 4 dimensions; the output current feature vector includes 1-dimensional current imbalance, 1-dimensional harmonic distortion rate, 2-dimensional statistical features, and 2-dimensional phase deviation features, totaling 6 dimensions. This system overcomes the limitations of traditional single-dimensional feature information, enabling comprehensive characterization of early latent faults such as slight screw loosening and overt faults such as current asymmetry.

[0088] In the fault adaptive attention mechanism feature fusion, dynamic weight allocation improves the fault discriminativeness of the fused features. First, initial weights are preset based on the fault type; for example, when diagnosing a loose screw, the weights of acoustic and vibration features are emphasized, while when diagnosing current asymmetry, the weights of output current features are emphasized. Then, the fault sensitivity of each feature vector is calculated, derived from the percentage difference between the current feature value and the feature value under normal operating conditions. The initial weights are then dynamically adjusted based on the fault sensitivity to ensure that the sum of the weights for each dimension is 1 after adjustment. Finally, the acoustic, vibration, and output current feature vectors are weighted and summed according to the adjusted weights to obtain the fused feature vector. This method overcomes the shortcomings of traditional fixed-weight fusion, which cannot adapt to different fault types, reduces feature information entropy, and significantly improves feature discriminativeness.

[0089] In the CNN-LSTM intelligent diagnostic model, the hybrid CNN-LSTM model combines the advantages of extracting spatial local features and temporal series features to achieve high accuracy and real-time performance in fault diagnosis. The CNN layer in the model contains two convolutional layers and two max-pooling layers, with 3×3 kernels (37 and 74 respectively) and ReLU activation function. The max-pooling layer has a 2×2 kernel and is responsible for extracting local fault patterns from fused features, such as acoustic peaks and sudden changes in vibration amplitude. The LSTM layer is a single hidden layer with 64 neurons and tanh activation function, responsible for capturing the temporal evolution of fault features, such as changes in features from minor to severe faults. The output layer uses the Softmax activation function and outputs four categories of results: normal, loose screw, asymmetrical axis, and asymmetrical output current, along with their corresponding confidence levels. Under complex working conditions, the model achieves an overall diagnostic accuracy of no less than 98%, with a single-sample diagnosis time of no more than 120ms, meeting the requirements for real-time diagnosis.

[0090] In summary, this invention achieves accurate and rapid identification and location of faults in dual-water-cooled synchronous condensers through a full-process innovation involving multi-dimensional data acquisition, hierarchical preprocessing, dynamic feature fusion, and intelligent diagnosis. This effectively reduces equipment downtime, minimizes grid voltage fluctuations, extends the operation and maintenance cycle of the synchronous condensers, provides key technical support for the safe operation of reactive power compensation equipment in grids with a high proportion of renewable energy access, and significantly improves the voltage stability and power supply reliability of the power system.

[0091] This invention also provides a fault diagnosis system for a dual underwater cold-tuning camera that integrates multi-dimensional information, comprising: The data acquisition module synchronously acquires acoustic data, vibration data, and output current data from the dual underwater cooling cameras to obtain a time-synchronized raw data set. The preprocessing module is used to perform hierarchical preprocessing on the original dataset, including targeted denoising and correction of acoustic data, vibration data and output current data, and unifying data dimensions to generate preprocessed standardized data. The feature extraction module is used to extract fault-sensitive features from acoustic print data, vibration data, and output current data based on standardized data, and to construct acoustic print feature vectors, vibration feature vectors, and output current feature vectors. The feature fusion module is used to dynamically assign weights and fuse acoustic feature vectors, vibration feature vectors, and output current feature vectors based on fault type to generate a fused feature vector; the weights are dynamically adjusted according to fault sensitivity. The intelligent diagnostic module is used to input the fused feature vector into the CNN-LSTM model, extract local spatial features through the convolutional neural network, capture the temporal evolution pattern through the long short-term memory network, and output the fault type and corresponding confidence level.

[0092] Specifically, in the data acquisition module, the voiceprint data acquisition adopts an electret microphone linear array with a signal-to-noise ratio of not less than 60dB, a sampling frequency of 20-20000Hz, a sampling interval of 100ms, a single sampling duration of 200ms, and a storage format of WAV. Vibration data acquisition uses a piezoelectric vibration sensor with a sensitivity of 100mV / g, a measurement range of 0-50g, a sampling frequency of 1000-2000Hz, a sampling interval of 100ms, and synchronously records the sensor number and corresponding installation position. The output current data acquisition uses a Hall current sensor with an accuracy class of 0.5, a measurement range of 0-1.2 times the rated output current of the phase shifter, a sampling frequency of 100Hz, a sampling interval of 100ms, and outputs the instantaneous values ​​of the three-phase current and phase information. The acquisition time of acoustic print, vibration and output current data is synchronized, with a timestamp error of no more than 1ms.

[0093] Specifically, the hierarchical preprocessing of the original data set in the preprocessing module includes: The voiceprint data is subjected to wavelet threshold denoising, pre-emphasis processing, and frame-segmentation windowing processing. Wavelet denoising uses a db4 wavelet basis for three-level decomposition and soft thresholding. Pre-emphasis uses a 100Hz high-pass filter. Frame-segmentation windowing uses a 20ms frame length, a 10ms frame shift, and a Hanning window. The vibration data were subjected to low-pass filtering, power frequency notch filtering, and baseline correction. The low-pass filtering used a 4th-order Butterworth filter with a cutoff frequency of 500Hz. The power frequency notch filtering used an adaptive notch filter with a center frequency of 50Hz and a bandwidth of 2Hz. Baseline correction was achieved by subtracting the signal mean. The output current data is subjected to Kalman filtering, moving average processing, and phase correction; the moving average uses a 5-point moving average filter; the phase correction is based on the synchronously acquired phase information to align the phases of the three-phase currents.

[0094] Specifically, the process of extracting fault-sensitive features from acoustic signature data, vibration data, and output current data in the feature extraction module is as follows: Mel frequency cepstral coefficients (MFCC), linear prediction cepstral coefficients (LPCC), short-time energy, and zero-crossing rate are extracted from the voiceprint data to construct a 27-dimensional voiceprint feature vector; where the first 13 dimensions of MFCC are used, the first 12 dimensions of LPCC are used, and the short-time energy and zero-crossing rate are each 1 dimension. Frequency domain features and time domain features are extracted from vibration data; frequency domain features include the amplitude and peak frequency of the 50-200Hz fault frequency band; time domain features include peak-to-peak value and root mean square value, and a 4-dimensional vibration feature vector is constructed. The current imbalance, harmonic distortion rate (THD), statistical features, and phase deviation features are extracted from the output current data to construct a 6-dimensional output current feature vector. The statistical features include the maximum value of the mean and the maximum value of the standard deviation of the three-phase current, and the phase deviation features include the phase difference between phases AB and BC.

[0095] Specifically, the process of generating the fused feature vector in the feature fusion module is as follows: Initial weights are preset based on fault types; the fault types include loose screws, shaft asymmetry, and output current asymmetry. Calculate the fault sensitivity of each eigenvector; fault sensitivity The calculation formula is:

[0096] in For the current eigenvalue, These are characteristic values ​​under normal operating conditions.

[0097] The initial weights are dynamically adjusted based on fault sensitivity; the specific adjustment formula is as follows:

[0098]

[0099]

[0100] in The corrected voiceprint weight; The corrected vibration weights; The corrected output current weight; where ; For voiceprint weighting, For vibration weights, Weighting for output current; For voiceprint fault sensitivity, For vibration fault sensitivity, For output current fault sensitivity.

[0101] The modified weights are used to sum the voiceprint feature vector, vibration feature vector, and output current feature vector to generate a fused feature vector.

[0102] Specifically, the CNN-LSTM model in the intelligent diagnostic module includes: The CNN layer contains two convolutional layers and two max pooling layers. The convolutional kernel size is 3×3, and the number of convolutional kernels is 37 and 74 respectively. The activation function is ReLU, and the max pooling kernel size is 2×2. The LSTM layer contains one hidden layer with 64 neurons, and the activation function is tanh. The output layer uses the Softmax activation function to output four types of results: normal, loose screw, asymmetrical shaft, and asymmetrical output current, along with their corresponding confidence levels.

[0103] The fault diagnosis system for dual underwater cold-tuning cameras of this invention integrates multi-dimensional information. First, it synchronously acquires the sound, vibration, and current signals of the camera during operation to provide raw data for fault analysis. Then, it performs multi-dimensional data preprocessing, differentially denoising and correcting the three types of signals: sound signals are denoised and feature extracted using wavelet denoising, soft thresholding, and frame-by-frame windowing; vibration signals are filtered to eliminate interference and restore true vibration using low-pass filtering, power frequency notch filtering, and baseline correction; and current signals are stabilized and aligned with electrical phase using Kalman filtering, moving average, and phase correction. Subsequently, features strongly correlated with the fault (such as acoustic signature spectrum features, vibration time / frequency domain features, and current electrical features) are extracted from the preprocessed signals. Then, through attention mechanism feature fusion, feature weights (DFF) are dynamically assigned based on the fault's sensitivity to the three types of signals, integrating them into a unified fused feature vector to achieve multi-dimensional information complementarity. Finally, through a CNN-LSTM intelligent diagnosis stage, CNN extracts local fault patterns from the fused features, and LSTM captures the temporal evolution of the fault, ultimately outputting the fault type and completing the intelligent diagnosis.

[0104] This invention further provides a computer program product, including a computer program that, when run by a processor, performs the steps of the method described above.

[0105] This invention also provides a computer-readable storage medium having a computer program stored thereon, the computer program executing the steps of the method described above when run by a processor.

[0106] This invention further provides a computer system including a memory and a processor interconnected, wherein the memory stores a computer program, and the computer program executes the steps of the method described above when run by the processor.

[0107] The products, media, and systems of the present invention, corresponding to the methods described above, also possess the advantages described above.

[0108] The present invention can implement all or part of the processes in the methods of the above embodiments, or it can be implemented by hardware related to computer program instructions. The computer program can be stored in a computer-readable storage medium. When the computer program is executed by a processor, it can implement the steps of the above method embodiments. The computer program includes computer program code, which can be in the form of source code, object code, executable file, or some intermediate form. The computer-readable storage medium includes: any entity or device capable of carrying computer program code, recording media, USB flash drive, portable hard drive, magnetic disk, optical disk, computer memory, read-only memory (ROM), random access memory (RAM), electrical carrier signals, telecommunication signals, and software distribution media, etc. The memory is used to store computer programs and / or modules. The processor implements various functions by running or executing the computer programs and / or modules stored in the memory, and by calling data stored in the memory. The memory may include high-speed random access memory, as well as non-volatile memory, such as hard disks, RAM, plug-in hard disks, smart media cards (SMC), secure digital (SD) cards, flash cards, at least one disk storage device, flash memory device, or other volatile solid-state storage devices.

[0109] The above are merely preferred embodiments of the present invention. The scope of protection of the present invention is not limited to the above embodiments. All technical solutions falling within the scope of the present invention's concept are within the scope of protection of the present invention. It should be noted that for those skilled in the art, any improvements and modifications made without departing from the principles of the present invention should be considered within the scope of protection of the present invention.

Claims

1. A method for diagnosing faults of a dual water inner cooling phase modulation camera by fusing multi-dimensional information, characterized by, The method comprises the steps of: synchronously collecting soundprint data, vibration data and output current data of the dual-water internal cooling phase modulation camera to obtain a time-synchronized original data set; performing hierarchical preprocessing on the original data set, including respectively performing targeted denoising and correction on the soundprint data, vibration data and output current data, and unifying the data dimensions to generate standardized data after preprocessing; extracting fault-sensitive features from the soundprint data, vibration data and output current data based on the standardized data to construct a soundprint feature vector, a vibration feature vector and an output current feature vector; performing dynamic weight allocation and fusion on the soundprint feature vector, the vibration feature vector and the output current feature vector based on the fault type to generate a fusion feature vector; wherein the weight is dynamically adjusted according to the fault sensitivity; inputting the fusion feature vector into a CNN-LSTM model, extracting local spatial features through a convolutional neural network, and capturing the time evolution law through a long short-term memory network to output the fault type and the corresponding confidence.

2. The dual water inside cooling camera fault diagnosis method of fusing multi-dimensional information according to claim 1, characterized in that, The hierarchical preprocessing of the original data set specifically comprises: wavelet threshold denoising, pre-emphasis processing and frame windowing processing are performed on the soundprint data; wherein the wavelet denoising adopts db4 wavelet basis for three-layer decomposition and soft threshold processing; the pre-emphasis adopts a 100Hz high-pass filter; the frame windowing adopts a 20ms frame length, a 10ms frame shift and a Hanning window; low-pass filtering, power frequency notch filtering and baseline correction are performed on the vibration data; wherein the low-pass filtering adopts a 4th order Butterworth filter with a cutoff frequency of 500Hz; the power frequency notch filtering adopts an adaptive notch filter with a center frequency of 50Hz and a bandwidth of 2Hz; the baseline correction is achieved by subtracting the signal mean value; Kalman filtering, moving average processing and phase correction are performed on the output current data; wherein the moving average adopts a 5-point moving average filter; the phase correction aligns the phases of the three-phase currents based on the phase information collected synchronously.

3. The dual water inside cooling phase modulation camera failure diagnosis method of fusing multi-dimensional information according to claim 1 or 2, characterized in that, The specific process of extracting fault-sensitive features from the soundprint data, vibration data and output current data is as follows: Mel-frequency cepstral coefficients (MFCC), linear prediction cepstral coefficients (LPCC), short-time energy and zero-crossing rate are extracted from the soundprint data to construct a 27-dimensional soundprint feature vector; wherein the first 13 coefficients of MFCC are taken, the first 12 coefficients of LPCC are taken, and the short-time energy and zero-crossing rate are each 1-dimensional; frequency domain features and time domain features are extracted from the vibration data; the frequency domain features include the amplitude and peak frequency in the 50-200Hz fault frequency band, and the time domain features include the peak-to-peak value and root mean square value to construct a 4-dimensional vibration feature vector; current unbalance degree, total harmonic distortion (THD), statistical features and phase deviation features are extracted from the output current data to construct a 6-dimensional output current feature vector; wherein the statistical features include the maximum value of the three-phase current mean and the maximum value of the standard deviation, and the phase deviation features include the phase difference between the AB phases and the phase difference between the BC phases.

4. The dual water inside cooling phase modulation camera failure diagnosis method of fusing multi-dimensional information according to claim 1 or 2, characterized in that, The specific process of generating the fusion feature vector is as follows: preset initial weights based on the fault type; wherein the fault types include screw loosening, shaft asymmetry and output current asymmetry; calculate the fault sensitivity of each feature vector; dynamically correct the initial weights according to the fault sensitivity; The voiceprint feature vector, the vibration feature vector and the output current feature vector are weighted and summed using the corrected weight to generate a fusion feature vector.

5. The dual water inside cooling camera fault diagnosis method of fusing multi-dimensional information according to claim 4, characterized in that, Fault sensitivity The formula for calculating the fault sensitivity is: wherein is the current characteristic value, is the characteristic value under normal operating conditions.

6. The dual water inside cooling camera fault diagnosis method of fusing multi-dimensional information according to claim 5, characterized in that, The specific process of dynamically correcting the initial weight according to the fault sensitivity is as follows: wherein is the revised voiceprint weight; is the revised vibration weight; is the revised output current weight; wherein ; is a voiceprint weight, is a vibration weight, is an output current weight; is a voiceprint fault sensitivity, is a vibration fault sensitivity, is an output current fault sensitivity.

7. The fusion multi-dimensional information-based diagnosis method of a twin-water inner-cooled phase modulation camera according to claim 1 or 2, characterized in that, The CNN-LSTM model comprises: The CNN layer comprises two convolution layers and two maximum pooling layers, the convolution kernel size is 3*3, the number of convolution kernels is 37 and 74 respectively, the activation function adopts ReLU, and the maximum pooling layer has a pooling kernel size of 2*2; The LSTM layer comprises one hidden layer and has 64 neurons, and the activation function adopts tanh; The output layer adopts a Softmax activation function and outputs four categories of results: normal, loose screw, asymmetric shaft and asymmetric output current, and corresponding confidence.

8. A computer program product comprising a computer program, characterized in that, The computer program, when executed by a processor, performs the steps of the method of any one of claims 1-7.

9. A computer readable storage medium having stored thereon a computer program, characterized in that, The computer program, when executed by a processor, performs the steps of the method of any one of claims 1-7.

10. A computer system comprising a memory and a processor connected to each other, said memory having stored thereon a computer program, characterized in that, The computer program, when executed by a processor, performs the steps of the method of any one of claims 1-7.

11. A dual water inner cooling phase modulation camera fault diagnosis system fusing multi-dimensional information, characterized in that, It comprises: A data acquisition module synchronously acquires voiceprint data, vibration data and output current data of a double-water internal cooling camera to obtain a time-synchronized original data set; A preprocessing module is configured to perform hierarchical preprocessing on the original data set, including denoising and correction of the voiceprint data, the vibration data and the output current data respectively, and unifying the data dimensions to generate standardized data after preprocessing; A feature extraction module is configured to extract fault sensitive features from the voiceprint data, the vibration data and the output current data based on the standardized data to construct a voiceprint feature vector, a vibration feature vector and an output current feature vector; A feature fusion module is configured to dynamically allocate and fuse the voiceprint feature vector, the vibration feature vector and the output current feature vector based on the fault type to generate a fusion feature vector; The weight is dynamically adjusted according to the fault sensitivity; An intelligent diagnosis module is configured to input the fusion feature vector into a CNN-LSTM model, extract local spatial features through a convolutional neural network, capture the time evolution law through a long short-term memory network, and output the fault type and the corresponding confidence. 12.The dual water inside cooling camera fault diagnosis system of fusing multi-dimensional information according to claim 11, characterized in that, The hierarchical preprocessing of the original data set in the preprocessing module comprises: Wavelet threshold denoising, pre-emphasis processing and frame windowing processing are performed on the voiceprint data; the wavelet denoising adopts db4 wavelet basis for three-layer decomposition and soft threshold processing; the pre-emphasis adopts a 100Hz high-pass filter; the frame windowing adopts a 20ms frame length, a 10ms frame shift and a Hanning window; Low-pass filtering, power frequency notch filtering and baseline correction are performed on the vibration data; the low-pass filtering adopts a 4th order Butterworth filter with a cutoff frequency of 500Hz; the power frequency notch filtering adopts an adaptive notch filter with a center frequency of 50Hz and a bandwidth of 2Hz; the baseline correction is achieved by subtracting the signal mean value; The output current data is subjected to Kalman filtering, moving average processing, and phase correction; the moving average uses a 5-point moving average filter; the phase correction is based on the synchronously acquired phase information to align the phases of the three-phase currents.

13. The dual water inside cooling phase modulation camera failure diagnostic system fusing multi-dimensional information according to claim 11 or 12, characterized by, The specific process of extracting fault-sensitive features from acoustic signature data, vibration data, and output current data in the feature extraction module is as follows: Mel frequency cepstral coefficients (MFCC), linear prediction cepstral coefficients (LPCC), short-time energy, and zero-crossing rate are extracted from the voiceprint data to construct a 27-dimensional voiceprint feature vector. in MFCC uses the first 13 dimensions of coefficients, LPCC uses the first 12 dimensions of coefficients, and short-time energy and zero-crossing rate each have 1 dimension. Extract frequency domain and time domain features from vibration data; The frequency domain features include the amplitude and peak frequency of the 50-200Hz fault frequency band; the time domain features include the peak-to-peak value and root mean square value, constructing a 4-dimensional vibration feature vector. Extract current imbalance, harmonic distortion rate (THD), statistical characteristics, and phase deviation characteristics from the output current data to construct a 6-dimensional output current feature vector; The statistical characteristics include the maximum value of the mean and the maximum value of the standard deviation of the three-phase current, and the phase deviation characteristics include the phase difference between phases AB and phase difference between phases BC.

14. The dual water inside cooling phase modulation camera failure diagnostic system fusing multi-dimensional information according to claim 11 or 12, characterized by, The specific process of generating the fused feature vector in the feature fusion module is as follows: Initial weights are preset based on fault types; the fault types include loose screws, shaft asymmetry, and output current asymmetry. Calculate the fault sensitivity of each feature vector; The initial weights are dynamically adjusted based on fault sensitivity. The modified weights are used to sum the voiceprint feature vector, vibration feature vector, and output current feature vector to generate a fused feature vector.

15. The dual water inside cooling phase modulation camera failure diagnostic system fusing multi-dimensional information of claim 14, wherein, Fault sensitivity The formula for calculating the fault sensitivity is: wherein is the current characteristic value, is the characteristic value under normal operating conditions.

16. The dual water inside cooling camera fault diagnosis system fusing multi-dimensional information according to claim 15, characterized in that, The specific process of dynamically adjusting the initial weights based on fault sensitivity is as follows: wherein is the revised voiceprint weight; is the revised vibration weight; is the revised output current weight; wherein ; is a voiceprint weight, is a vibration weight, is an output current weight; is a voiceprint fault sensitivity, is a vibration fault sensitivity, is an output current fault sensitivity.

17. The dual water inside cooling phase modulation camera failure diagnostic system fusing multi-dimensional information according to claim 11 or 12, characterized by, The CNN-LSTM model in the intelligent diagnostic module includes: The CNN layer contains two convolutional layers and two max pooling layers. The convolutional kernel size is 3×3, and the number of convolutional kernels is 37 and 74 respectively. The activation function is ReLU, and the max pooling kernel size is 2×2. The LSTM layer contains one hidden layer with 64 neurons, and the activation function is tanh. The output layer uses the Softmax activation function to output four types of results: normal, loose screw, asymmetrical shaft, and asymmetrical output current, along with their corresponding confidence levels.

18. The dual water inside cooling phase modulation camera failure diagnostic system fusing multi-dimensional information according to claim 11 or 12, characterized by, In the data acquisition module, the voiceprint data acquisition uses an electret microphone linear array with a signal-to-noise ratio of not less than 60dB, a sampling frequency of 20-20000Hz, a sampling interval of 100ms, a single sampling duration of 200ms, and a storage format of WAV. Vibration data acquisition uses a piezoelectric vibration sensor with a sensitivity of 100mV / g, a measurement range of 0-50g, a sampling frequency of 1000-2000Hz, a sampling interval of 100ms, and synchronously records the sensor number and corresponding installation position. The output current data acquisition uses a Hall current sensor with an accuracy class of 0.5, a measurement range of 0-1.2 times the rated output current of the phase shifter, a sampling frequency of 100Hz, a sampling interval of 100ms, and outputs the instantaneous values ​​of the three-phase current and phase information. The collection time of the voiceprint, vibration and output current data is synchronized, and the timestamp error is not more than 1 ms.

Images