Rotor acoustic anomaly detection method, system and equipment based on auto-encoder and wavelet packet energy entropy, and medium

Abstract

The invention discloses a rotor acoustic anomaly detection method, system and device based on an auto-encoder and wavelet packet energy entropy and a medium, and belongs to the technical field of hydroelectric generating sets, and the method comprises the steps: collecting an original acoustic signal, carrying out the noise reduction of the original acoustic signal, and carrying out the multi-channel data fusion to obtain an integrated acoustic signal; inputting the integrated acoustic signal into a wavelet packet for transformation processing to obtain wavelet packet coefficients under different scales; calculating energy values of different frequency band signals based on wavelet packet coefficients to form an energy entropy feature vector; constructing an implicit feature model of the hydroelectric generating set rotor in a normal state according to the energy entropy feature vector; and carrying out feature reconstruction on the energy entropy feature vector, calculating a reconstruction error, carrying out dynamic statistical analysis, and carrying out real-time monitoring and abnormity judgment on the operation state of the hydroelectric generating set rotor. According to the method, the detection precision and the response speed are improved in actual hydroelectric generating set rotor acoustic anomaly detection, and automatic identification and dynamic threshold adaptive adjustment of irregular perturbation are realized.

Description

Technical Field

[0001] This invention relates to the field of hydropower unit technology, specifically to a rotor acoustic anomaly detection method, system, equipment, and medium based on self-encoder and wavelet packet energy entropy. Background Technology

[0002] With the continuous development of intelligent sensing technology and deep learning models, equipment condition monitoring based on acoustic signals has been gradually applied to the operation and maintenance of large-scale generating units. As a key piece of equipment in the national energy system, the rotor operating status of hydropower units is directly related to power generation efficiency and system safety. Acoustic signals, as a non-contact and highly continuous sensing method, have the advantages of high sensitivity and fast response speed. In recent years, they have been gradually applied to the detection of anomalies during rotor operation. However, in actual operation, the acoustic signals generated by the rotor of hydropower units are often affected by a variety of non-fault factors such as environmental noise, water flow disturbance, and electromagnetic interference, making it difficult to distinguish abnormal features from normal fluctuations. The time-varying nature of the signals and interference from multiple operating conditions also pose a severe challenge to traditional anomaly detection methods.

[0003] Currently, common acoustic anomaly detection methods mainly include feature analysis methods based on empirical thresholds, statistical modeling methods, and shallow machine learning methods. Existing methods usually rely on predefined frequency domain or time domain features for classification or judgment. The core problem is that the feature extraction process is not responsive enough to the overall changes in the signal. In high-noise and multi-perturbation environments, feature aliasing and decreased detection accuracy are prone to occur. In addition, traditional methods often rely on expert experience to set fixed thresholds or rules, which are difficult to effectively identify small anomalies in complex backgrounds, and are prone to false alarms or missed alarms, thus limiting their application in real-world scenarios.

[0004] In recent years, deep learning methods have been gradually applied to fault identification. Among them, autoencoder models are used to construct equipment feature models under normal conditions due to their advantages in unsupervised learning tasks, and then perform anomaly detection through reconstruction errors. However, existing autoencoder-based methods mostly use the original acoustic signal or its traditional features as input, lacking fine-grained modeling of frequency locality and distribution structure, and cannot fully reflect the energy transfer characteristics of the rotor under abnormal conditions.

[0005] To address the problems of existing technologies, there is an urgent need to propose a new method that integrates time-frequency structure enhancement feature extraction, adaptive unsupervised modeling, and dynamic anomaly determination mechanism to improve the accuracy and stability of acoustic anomaly detection in hydropower unit rotors. Summary of the Invention

[0006] In view of the above-mentioned problems, the present invention is proposed.

[0007] Therefore, one objective of this invention is to propose a rotor acoustic anomaly detection method based on an autoencoder and wavelet packet energy entropy. In actual hydropower unit rotor acoustic anomaly detection, this invention not only improves detection accuracy and response speed, but also realizes automatic identification of irregular perturbations and dynamic threshold adaptive adjustment.

[0008] To solve the above-mentioned technical problems, the present invention provides the following technical solution: a rotor acoustic anomaly detection method based on self-encoder and wavelet packet energy entropy, comprising,

[0009] The process involves collecting raw acoustic signals generated during the operation of the hydropower unit rotor, denoising the raw acoustic signals, performing multi-channel data fusion on the denoised raw acoustic signals to merge the different raw acoustic signals into an integrated acoustic signal, inputting the integrated acoustic signal into wavelet packet transform processing to obtain wavelet packet coefficients at different scales, calculating energy values ​​for signals in different frequency bands based on the wavelet packet coefficients to form energy entropy feature vectors, constructing a latent feature model of the hydropower unit rotor under normal conditions based on the energy entropy feature vectors, reconstructing the energy entropy feature vectors, calculating the reconstruction error during the reconstruction process, performing dynamic statistical analysis on the reconstruction error, and conducting real-time monitoring and anomaly detection of the hydropower unit rotor's operating status.

[0010] As a preferred embodiment of the rotor acoustic anomaly detection method based on autoencoder and wavelet packet energy entropy described in this invention, the original acoustic signal includes: arranging acoustic sensors around the rotor of the hydropower unit, collecting the original acoustic signal generated during the operation of the hydropower unit rotor, and constructing it into an original acoustic signal set.

[0011] Each original acoustic signal in the original acoustic signal set is converted to a uniform format to obtain a standardized acoustic signal set;

[0012] The standardized acoustic signal set is denoised to obtain the denoised acoustic signal set.

[0013] As a preferred embodiment of the rotor acoustic anomaly detection method based on autoencoder and wavelet packet energy entropy described in this invention, the integrated acoustic signal includes classifying and organizing the denoised acoustic signal set to construct a denoised acoustic signal subset.

[0014] An adaptive weighting factor is assigned to a subset of acoustic signals, and the acoustic signals are weighted and summed according to the assigned adaptive weighting factor to form an integrated acoustic signal value consistent with the original time axis.

[0015] The integrated acoustic signal values ​​are arranged sequentially in time to construct an integrated acoustic signal sequence set.

[0016] As a preferred embodiment of the rotor acoustic anomaly detection method based on autoencoder and wavelet packet energy entropy described in this invention, wherein obtaining wavelet packet coefficients at different scales includes inputting a constructed integrated acoustic signal sequence set into wavelet packet transform processing;

[0017] The integrated acoustic signal is decomposed into sub-band signals under the conditions of preset wavelet basis function and decomposition level;

[0018] Extract the corresponding wavelet packet coefficients from the sub-band signal and calculate the sub-band energy value;

[0019] The energy values ​​of all sub-bands are normalized to construct a set of wavelet packet coefficients.

[0020] As a preferred embodiment of the rotor acoustic anomaly detection method based on self-encoder and wavelet packet energy entropy described in this invention, wherein: the formation of energy entropy feature vector includes inputting a wavelet packet coefficient set into an energy calculation module for processing, the wavelet packet coefficient set containing all decomposition levels and their corresponding sub-frequency band numbers of the wavelet packet coefficient sequence, wherein each wavelet packet coefficient sequence is composed of wavelet packet coefficients under the corresponding time sampling point index;

[0021] Calculate the energy value corresponding to each sub-band for each wavelet packet coefficient sequence, normalize the energy values ​​of all sub-bands, and construct the energy proportion of each sub-band.

[0022] The wavelet packet energy entropy characteristics are calculated based on the energy proportion of each sub-band, representing the entropy characteristics of the energy distribution of the integrated acoustic signal under wavelet packet decomposition.

[0023] The energy proportions of each sub-band are combined according to their index order in the wavelet packet decomposition structure to form an energy entropy feature vector.

[0024] As a preferred embodiment of the rotor acoustic anomaly detection method based on autoencoder and wavelet packet energy entropy described in this invention, the implicit feature model includes: inputting the energy entropy feature vector into the input layer of the gated autoencoder model, and constructing a frequency band gated activation factor based on the energy proportion of each sub-band;

[0025] The frequency band gate activation factor is applied to each frequency band channel in the energy entropy feature vector. A weighted input vector is generated by weighting the elements one by one. The weighted input vector is then input into the encoder structure. After weight matrix mapping and nonlinear activation processing, the hidden feature vector is output.

[0026] The latent feature vector is input into the decoder to obtain the initial reconstructed vector. At the same time, a residual compensation path is introduced to generate the residual correction amount. Combined with the frequency band gated activation factor, differential weighted compensation is performed to obtain the reconstructed energy entropy feature vector.

[0027] As a preferred embodiment of the rotor acoustic anomaly detection method based on autoencoder and wavelet packet energy entropy described in this invention, the feature reconstruction includes: inputting the energy entropy feature vector to be detected into the trained gated autoencoder model, and using the trained gated autoencoder model to encode and decode the energy entropy feature vector to obtain the corresponding hidden feature representation and reconstructed energy entropy feature vector respectively.

[0028] Calculate the reconstruction error vector between the original energy entropy eigenvector and the reconstructed energy entropy eigenvector, and calculate the overall reconstruction error value based on the reconstruction error vector;

[0029] The overall reconstruction error values ​​corresponding to all historical samples are constructed into a historical reconstruction error sample set. Dynamic statistical analysis is performed on the historical reconstruction error sample set to determine the dynamic reconstruction error judgment threshold used in the current detection task.

[0030] The overall reconstruction error value corresponding to the current detection sample is compared with the dynamic reconstruction error judgment threshold, and the current operating state of the hydropower unit rotor is determined accordingly.

[0031] The overall reconstruction error value of the current detection sample is added to the historical reconstruction error sample set, and the historical reconstruction error sample set is updated based on the sliding time window mechanism or the cumulative sample update mechanism, dynamically adjusting the historical reconstruction error mean and the historical reconstruction error standard deviation.

[0032] Another objective of this invention is to provide a rotor acoustic anomaly detection system based on an autoencoder and wavelet packet energy entropy.

[0033] To solve the above-mentioned technical problems, the present invention provides the following technical solution: a rotor acoustic anomaly detection system based on autoencoder and wavelet packet energy entropy, comprising: a data preprocessing module, a data fusion module, a wavelet packet transform module, an energy calculation module, a hidden feature module, and a feature reconstruction module;

[0034] The data preprocessing module collects the raw acoustic signals generated during the operation of the hydropower unit rotor and performs noise reduction processing on the raw acoustic signals.

[0035] The data fusion module performs multi-channel data fusion on the original acoustic signal after noise reduction, and merges the different original acoustic signals into an integrated acoustic signal;

[0036] The wavelet packet transform module takes the integrated acoustic signal as input to the wavelet packet transform process to obtain wavelet packet coefficients at different scales.

[0037] The energy calculation module calculates energy values ​​for signals in different frequency bands based on wavelet packet coefficients, forming an energy entropy feature vector.

[0038] The hidden feature module constructs a hidden feature model of the hydropower unit rotor under normal conditions based on the energy entropy feature vector;

[0039] The feature reconstruction module reconstructs the energy entropy feature vector, calculates the reconstruction error during the reconstruction process, performs dynamic statistical analysis on the reconstruction error, and monitors and judges the rotor operation status of the hydropower unit in real time.

[0040] The present invention provides a computer device, including a memory and a processor, wherein the memory stores a computer program, characterized in that the processor executes the computer program to implement the steps of the rotor acoustic anomaly detection method based on autoencoder and wavelet packet energy entropy.

[0041] The present invention provides a computer-readable storage medium having a computer program stored thereon, characterized in that, when the computer program is executed by a processor, it implements the steps of the rotor acoustic anomaly detection method based on autoencoder and wavelet packet energy entropy.

[0042] The beneficial effects of this invention are as follows: This invention performs wavelet packet multi-scale decomposition on the integrated acoustic signal sequence, extracts multi-layer and multi-frequency band wavelet packet coefficients, constructs a frequency band energy proportion vector, and calculates the energy entropy feature vector based on this, comprehensively characterizing the time-frequency structure and energy uncertainty of the acoustic signal during rotor operation. The feature extraction method has the advantages of high resolution accuracy and sensitivity to local energy disturbances, effectively improving the ability to capture minor faults under complex background noise conditions, and enhancing the sensitivity and anti-interference ability of the fault early warning system.

[0043] The gated autoencoder model proposed in this invention takes the energy entropy feature vector as input, fuses frequency band activation factors, dynamically adjusts the representation weights of the input channels, and performs enhanced reconstruction of abnormal frequency bands through residual compensation branches. A sparse regularization term is added to the encoder to extract the most discriminative hidden features, and a compensation constraint mechanism for frequency abnormal segments is introduced in the decoder to prevent redundant information from interfering with the model expression. This significantly improves the modeling ability for local frequency anomalies, especially in small sample and perturbation anomaly scenarios, where it exhibits higher recognition accuracy and adaptability.

[0044] This invention does not rely on static threshold settings. Instead, it dynamically statistically models the reconstruction error under historical normal conditions, updates the mean and standard deviation in real time, generates a dynamic reconstruction error judgment threshold, and classifies abnormal states based on the comparison between the current detection error and the threshold. This effectively addresses the feature distribution drift caused by changes in operating status over time, improves the stability and false alarm control capability of the model in long-term monitoring, and the dynamic mechanism is more real-time and self-adaptive, making it widely adaptable to the multi-condition and multi-stage operating environment of hydropower unit rotors. Attached Figure Description

[0045] To more clearly illustrate the technical solutions of the embodiments of the present invention, the accompanying drawings used in the description of the embodiments will be briefly introduced below. Obviously, the accompanying drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0046] Figure 1 The above is a flowchart of the overall process for a rotor acoustic anomaly detection method based on an autoencoder and wavelet packet energy entropy, provided as an embodiment of the present invention. Detailed Implementation

[0047] To make the above-mentioned objects, features, and advantages of the present invention more apparent and understandable, specific embodiments of the present invention will be described in detail below with reference to the accompanying drawings. Obviously, the described embodiments are only a part of the embodiments of the present invention, and not all of them. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort should fall within the protection scope of the present invention.

[0048] Example 1, referring to Figure 1 This is one embodiment of the present invention, which provides a rotor acoustic anomaly detection method based on an autoencoder and wavelet packet energy entropy, including:

[0049] S100: Collects the original acoustic signal generated during the operation of the hydropower unit rotor, and performs noise reduction processing on the original acoustic signal.

[0050] S200 performs multi-channel data fusion on the original acoustic signal after noise reduction processing, and merges the different original acoustic signals into an integrated acoustic signal.

[0051] S300: Input the integrated acoustic signal into wavelet packet transform processing to obtain wavelet packet coefficients at different scales.

[0052] S400: Calculate the energy value of signals in different frequency bands based on wavelet packet coefficients to form an energy entropy feature vector.

[0053] S500. Construct a hidden feature model of the rotor of the hydropower unit under normal conditions based on the energy entropy feature vector.

[0054] S600. Perform feature reconstruction on the energy entropy eigenvector and calculate the reconstruction error during the reconstruction process.

[0055] S700 performs dynamic statistical analysis on reconfiguration errors and performs real-time monitoring and anomaly detection on the rotor operating status of hydropower units.

[0056] It should be noted that the acoustic signals generated by the rotor of a hydropower unit are often affected by a variety of non-fault factors such as environmental noise, water flow disturbance, and electromagnetic interference, making it difficult to distinguish abnormal features from normal fluctuations. The feature extraction process is not responsive enough to the overall changes in the signal. In high-noise and multi-disturbance environments, feature aliasing and decreased detection accuracy are likely to occur. It is difficult to effectively identify small anomalies in complex backgrounds, and false alarms or missed alarms are very likely to occur, which limits its application in practical scenarios. Existing methods based on autoencoders mostly lack fine-grained modeling of frequency locality and distribution structure, and cannot fully reflect the energy transfer characteristics of the rotor under abnormal conditions.

[0057] Therefore, to address the aforementioned problems, the steps S100-S700 are used to integrate time-frequency structure enhancement feature extraction, adaptive unsupervised modeling, and dynamic anomaly determination mechanism to improve the accuracy and stability of acoustic anomaly detection in hydropower unit rotors.

[0058] Example 2, refer to Figure 1 This is one embodiment of the present invention, which provides a rotor acoustic anomaly detection method based on an autoencoder and wavelet packet energy entropy, including:

[0059] In this embodiment of the invention, step S100 involves collecting the original acoustic signal generated during the operation of the hydropower unit rotor, and then performing noise reduction processing on the original acoustic signal. This includes the following steps S101-S103:

[0060] S101. Multiple acoustic sensors are arranged around the rotor of the hydropower unit to collect the raw acoustic signals generated during the operation of the rotor, and the acoustic signals collected by each acoustic sensor are uniformly constructed into a raw acoustic signal set D. raw :

[0061]

[0062] Among them, s i (t) represents the raw acoustic signal acquired at the i-th time point, t i At the signal sampling time, a i l represents the sound pressure amplitude at that moment. i The sensor channel identifier for the acoustic signal, where N is the total number of sampling points for the original acoustic signal;

[0063] In an embodiment of the present invention, S102, for each original acoustic signal s in the original acoustic signal set... i (t) Perform a unified format conversion to obtain a standardized acoustic signal set. The unified format conversion includes the following steps A1-A3:

[0064] A1. Sampling frequency standardization: Adjust the sampling frequency of all acoustic sensors to the target frequency, and use linear interpolation or downsampling techniques to process raw signals with different sampling rates to ensure consistent data length.

[0065] A2. Unit normalization: Normalizes the amplitude of each channel signal to a uniform range. Maximum and minimum value normalization or Z-score normalization can be used to eliminate the influence of sensor sensitivity differences.

[0066] A3. Time axis alignment: Based on hardware synchronization timestamps or signal cross-correlation analysis, time offset compensation is performed on signals of each channel to ensure that all sensor data are aligned under the same time base.

[0067] In an optional implementation, the unified format conversion in step S102 can also be performed by segmented resampling, dividing the original acoustic signal into segments according to a fixed time window, and adjusting the sampling frequency of each segment to a unified target value using linear interpolation; amplitude scaling is performed, with the maximum amplitude of a single sensor channel as a reference, and the scaling ratio of other channel signals is calculated to linearly compress all signal amplitudes to the [-1,1] interval; dynamic time warping and alignment is performed, selecting the main sensor signal as a reference, and performing dynamic time warping matching on segments of other channel signals to eliminate local time delay differences and maintain overall waveform synchronization.

[0068] In another optional implementation, the unified format conversion in step S102 can also be achieved through master-slave channel frequency synchronization, specifying a master sensor channel as a reference, and matching the master channel frequency to other channels through downsampling or upsampling, using nearest neighbor interpolation to simplify calculations; sliding window amplitude normalization, calculating the average signal energy within the sliding window as a unit, and adjusting the amplitude of each channel proportionally to a unified energy level; and global timestamp alignment, performing overall translation compensation on all channel signals based on the initial trigger timestamp recorded by the acquisition system to ensure consistent start times.

[0069] In an embodiment of the present invention, S103, the standardized acoustic signal set is subjected to noise reduction processing to obtain a noise-reduced acoustic signal set. The noise reduction processing includes the following steps B1-B3:

[0070] B1. Perform preliminary noise reduction processing on the standardized acoustic signal set;

[0071] B2. Use a bandpass filter to remove high-frequency and low-frequency noise that exceeds the operating frequency band of the device;

[0072] B3. After removal, the ambient background interference is reduced by the moving average method to obtain the noise-reduced acoustic signal set D. denoise .

[0073] In an optional implementation, the noise reduction process in step S103 can also be performed by wavelet decomposition of the standardized acoustic signal, selecting a suitable wavelet basis to decompose to the third layer, using a soft thresholding method to remove noise components from the high-frequency coefficients of each layer, retaining the main characteristic frequency bands, superimposing median filtering to eliminate pulse interference, and outputting the final noise-reduced signal.

[0074] In another optional implementation, the noise reduction process in step S103 can also be performed by standardizing the acoustic signal and performing EMD decomposition on each signal to obtain several IMF components; removing the first two high-frequency IMF components and the last residual trend term; and performing adaptive filtering optimization on the remaining IMF components to reconstruct the noise-reduced signal set.

[0075] In this embodiment of the invention, step S200 involves multi-channel data fusion of the original acoustic signal after noise reduction processing, fusing different original acoustic signals into an integrated acoustic signal, including the following steps S201-S203:

[0076] S201. Perform multi-channel data fusion on the original acoustic signal after noise reduction, including the following steps C1-C2:

[0077] C1, the acoustic signal set D after noise reduction processing denoise According to the acoustic sensor channel label i Organize and categorize;

[0078] C2. Construct the noise-reduced acoustic signal subsets corresponding to each acoustic sensor channel. Each subset of acoustic signals consists of all acoustic signals under the same sensor channel number;

[0079] C3. All acoustic signals include the sampling time t. i The sound pressure amplitude a′ after noise reduction i,l and sensor channel identifier l i ;

[0080] In an optional implementation, the multi-channel data fusion method in step S201 can also be based on time-domain weighted multi-channel fusion. The acoustic signals of each channel after noise reduction are time-aligned and calibrated to ensure that the timestamps of all sensor data are synchronized. The weighting coefficient is set according to the distance between the sensor and the rotor, with higher weight for near-end sensors and lower weight for far-end sensors. The sound pressure amplitude of each channel at the same time is superimposed according to the weight to generate the fused integrated acoustic signal.

[0081] In another optional implementation, the multi-channel data fusion method in step S201 can also be channel fusion based on frequency band selection. Fast Fourier transform frequency band energy analysis is performed on the noise-reduced signal of each sensor channel, and the channel data with the highest proportion of rotor characteristic frequency band energy is selected. The sound pressure amplitude of the selected channel is directly used as the fusion output, and the remaining channels are only used for redundancy verification.

[0082] S202 represents each subset of acoustic signals after noise reduction. Assign adaptive weighting factor ω l The adaptive weighting factor of each acoustic sensor channel is used to control its contribution in the integrated acoustic signal fusion process and ensure that the sum of all weighting factors is equal to 1.

[0083] S203, at each time sampling point t i Above, the noise-reduced sound pressure amplitude a′ corresponding to each acoustic sensor channel is used. i,l and the corresponding adaptive weighting factor ω l The acoustic signals from multiple acoustic sensor channels are weighted and summed to form an integrated acoustic signal value x that is consistent with the original time axis. fused (t i The sequences are arranged chronologically to construct an integrated acoustic signal sequence set D. fused .

[0084] In summary, this step improves the stability of the signal under disturbance conditions. Through adaptive weighting or frequency band selection, this step can suppress local noise and highlight the characteristics of the rotor core frequency band, solving the problem of insufficient characteristic response under multiple disturbances, thereby reducing false alarms. For example, in the frequency band selection fusion of Example 2, using only high-energy channel data can avoid the error introduced by low-quality channels.

[0085] In an embodiment of the present invention, step S300 involves inputting the integrated acoustic signal to wavelet packet transform processing to obtain wavelet packet coefficients at different scales, including the following steps S301-S303:

[0086] S301, Construct the integrated acoustic signal sequence set D fused The input is fed into the wavelet packet transform module. Based on the typical acoustic frequency characteristics of the hydropower unit rotor during operation, the corresponding wavelet basis function ψ(t) and the maximum decomposition level J are selected. In each decomposition level j of the maximum decomposition level J, the integrated acoustic signal is decomposed into 2j j Each sub-band signal is composed of a frequency band and a time sampling point t. i The corresponding sound pressure amplitude composition, the sub-band signals under different decomposition levels have independent frequency coverage ranges, which are used to distinguish the characteristics of different frequency components in the operation of the hydropower unit rotor;

[0087] S302. Extract the corresponding wavelet packet coefficients for each sub-band signal. All wavelet packet coefficients are organized in order according to their decomposition level j, sub-band number k, and time sampling point index n.

[0088] S303. During the wavelet packet decomposition and wavelet packet coefficient extraction process, maintain the one-to-one correspondence between the wavelet packet coefficients and their source sub-frequency band signals and original sampling time points, so that all wavelet packet decomposition results are mapped in the time-frequency space, and construct the wavelet packet coefficient set C. The wavelet packet coefficient set consists of the wavelet packet coefficients corresponding to the sub-frequency bands under all decomposition levels.

[0089] In summary, this step enhances frequency locality modeling. Through multi-scale decomposition, this step can cover the rotor operating frequency band in a fine-grained manner, solving the problem of insufficient modeling of distributed structures in existing technologies and improving the sensitivity of identifying minute anomalies (such as early bearing wear).

[0090] In an embodiment of the present invention, step S400, which calculates energy values ​​for signals in different frequency bands based on wavelet packet coefficients to form an energy entropy feature vector, includes the following steps S401-S404:

[0091] S401. Input the wavelet packet coefficient set C into the energy calculation module. The wavelet packet coefficient set C contains the wavelet packet coefficient sequence C of all decomposition levels j and their corresponding sub-frequency band numbers k. j,k Each wavelet packet coefficient sequence consists of the wavelet packet coefficients c at the corresponding time sampling point index n. j,k,n constitute;

[0092] S402, For each wavelet packet coefficient sequence C j,k Calculate the energy value E corresponding to the sub-band j,k :

[0093]

[0094] Among them, E j,k M represents the energy value corresponding to the k-th sub-band of the j-th layer. j,k This indicates the number of wavelet packet coefficients in the corresponding sub-band;

[0095] In an optional implementation, the method for calculating the sub-band energy value in step S402 can also be based on absolute value summation energy calculation, while preserving the wavelet packet coefficients c. j,k,n The input structure and hierarchical division remain unchanged. The absolute value of the wavelet packet coefficient sequence of each sub-band is taken and then linearly summed. The summation result is used as the energy characterization value of that sub-band.

[0096] In another optional implementation, the method for calculating the sub-band energy value in step S402 can also be based on the square mean energy calculation. The wavelet packet coefficient grouping method of the original scheme is used. The wavelet packet coefficients of each sub-band are first squared, and the arithmetic mean of the squared coefficients is calculated as the energy value corresponding to the sub-band.

[0097] S403. Normalize the energy values ​​of all sub-bands to construct the energy proportion P of each sub-band. j,k Based on the energy proportion P of each sub-band j,k The wavelet packet energy entropy feature H is calculated. H represents the entropy characteristic of the energy distribution of the integrated acoustic signal under wavelet packet decomposition, reflecting the uncertainty and distribution complexity of the hydroelectric generator rotor acoustic signal in its frequency structure.

[0098]

[0099] S404, the energy percentage P of each sub-band j,k The energy entropy feature vector is formed by combining the features according to their index order in the wavelet packet decomposition structure. The energy entropy eigenvector is used to fully describe the energy distribution of the integrated acoustic signal under the wavelet packet decomposition structure.

[0100] In summary, this step provides fine-grained time-frequency structural features. Energy entropy can sensitively capture energy transfer (such as entropy changes caused by anomalies), solving the problem of insufficient response of feature extraction to overall changes. It avoids feature aliasing in high-noise environments and improves detection accuracy. In particular, in the square mean energy calculation in Example 2, it can amplify the contribution of anomalous frequency bands.

[0101] In this embodiment of the invention, step S500 involves constructing a hidden feature model of the hydropower unit rotor under normal conditions based on the energy entropy feature vector, including the following steps S501-S505:

[0102] S501, the energy entropy feature vector F entropy The energy entropy feature vector is input to the input layer of the gated autoencoder model to reflect the energy distribution of the integrated acoustic signal at different decomposition scales and frequency bands.

[0103] S502, Based on the energy percentage P of each sub-band j,k Constructing band-gated activation factor γ j,k The frequency band gating activation factor is calculated by multiplying the energy percentage by the adjustment coefficient w. g With bias term b g The activation function transformation result of the sum is used to dynamically adjust the activation intensity of each input channel and give higher energy frequency bands greater representation weight.

[0104] S503. Apply the frequency band gate activation factor to each frequency band channel in the energy entropy feature vector, generate a weighted input vector by element-wise weighting, and input the weighted input vector into the encoder structure. After weight matrix mapping and nonlinear activation processing, output the hidden feature vector h. The hidden feature vector represents the deep acoustic feature expression of the hydropower unit rotor under normal operating conditions.

[0105] S504. Input the hidden feature vector h into the decoder to obtain the preliminary reconstructed vector. Simultaneously, a residual compensation path is introduced to generate a residual correction amount ΔF, which is then combined with a frequency band gated activation factor for differentiated weighted compensation to obtain the reconstructed energy entropy feature vector:

[0106]

[0107] Wherein, γ is the frequency band activation factor, used to amplify the residual compensation space of abnormal frequency bands;

[0108] S505. Construct a weighted loss function during the training phase. The weighted loss function consists of a reconstruction error term, a latent feature sparsity regularization term, and a frequency band compensation penalty term:

[0109]

[0110] in, This represents the original energy entropy feature vector corresponding to the i-th training sample. This represents the reconstructed energy entropy feature vector output by the i-th training sample after the encoding-decoding process of the gated autoencoder model. Let represent the squared Euclidean distance between the original and reconstructed energy entropy feature vectors of the i-th sample, used to measure the model's ability to reconstruct normal acoustic features, and ||h||₁ represent the latent feature vector h. Norm, ΔF j,k This represents the residual compensation amount of the wavelet packet in the j-th layer and k-th frequency band, (1-γ) j,k )·|ΔF j,k | represents the penalty mechanism that imposes strong constraints on the residual compensation term in the low activation frequency band, λ1 is the regularization adjustment coefficient of the implicit feature sparse constraint, and λ2 is the adjustment coefficient of the penalty intensity of the residual compensation term.

[0111] In summary, this step achieves adaptive fine-grained modeling, and frequency band gating and residual compensation enhance the model's sensitivity to abnormal frequency bands. It solves the problem of existing autoencoders lacking distributed structure modeling, improves the ability to capture small energy transfers, and reduces false negatives. For example, sparse regularization terms during the training phase prevent overfitting and ensure model generalization.

[0112] In an embodiment of the present invention, step S600 involves reconstructing the energy entropy feature vector and calculating the reconstruction error during the reconstruction process, including the following steps S601-S602:

[0113] S601, The energy entropy feature vector to be detected The input is fed into the trained gated autoencoder model, which is then used to process the energy entropy feature vector. Encoding and decoding operations are performed to obtain the corresponding latent feature representations h. test With reconstructing the energy entropy eigenvector

[0114] S602. Calculate the original energy entropy eigenvector. With reconstructing the energy entropy eigenvector The reconstruction error vector ε between the two points is used to calculate the overall reconstruction error value E. rec :

[0115]

[0116] in, This represents the energy percentage of the test sample in the k-th sub-band of the j-th layer. This indicates the corresponding reconstructed value.

[0117] In summary, a direct amplification mechanism for abnormal signals is provided, and the reconstruction error can reflect minor anomalies (such as energy distribution distortion caused by initial faults) with high sensitivity, avoiding the problem of existing methods being insensitive to overall changes, and providing reliable input for real-time monitoring.

[0118] In an embodiment of the present invention, dynamic statistical analysis of the reconfiguration error is performed in S700, and real-time monitoring and anomaly judgment are performed on the operating status of the hydropower unit rotor, including the following steps S701-S704:

[0119] S701. Collect the overall reconstruction error value output by the gated self-encoder model in the historical stage, and construct the overall reconstruction error value corresponding to all historical samples into a historical reconstruction error sample set. Each reconstruction error value in the historical reconstruction error sample set represents the degree of overall deviation generated by the reconstruction of the energy entropy feature vector under the normal operating state of the hydropower unit rotor.

[0120] S702. Perform dynamic statistical analysis on the historical reconstruction error sample set, calculate the mean and standard deviation of the sample set respectively, express the mean as the historical reconstruction error mean, and express the standard deviation as the historical reconstruction error standard deviation. Based on the historical reconstruction error mean and historical reconstruction error standard deviation, determine the dynamic reconstruction error judgment threshold used in the current detection task. The dynamic reconstruction error judgment threshold represents the maximum acceptable normal deviation value and is adaptively generated from historical data.

[0121] S703. Compare the overall reconstruction error value corresponding to the current detection sample with the dynamic reconstruction error judgment threshold, and classify the current operating state of the hydropower unit rotor accordingly.

[0122] Optionally, the current operating state includes:

[0123] If the current overall reconstruction error value is less than or equal to the dynamic reconstruction error judgment threshold, the current operating status is determined to be normal.

[0124] If the current overall reconstruction error value is greater than the dynamic reconstruction error judgment threshold, and less than or equal to the sum of the dynamic reconstruction error judgment threshold and the abnormal buffer threshold, the current running state is determined to be an abnormal warning state.

[0125] If the current overall reconstruction error value is greater than the sum of the dynamic reconstruction error judgment threshold and the abnormal buffer threshold, the current running state is determined to be an abnormal state.

[0126] The anomaly buffer threshold is used to set the transition range between minor and severe anomalies.

[0127] S704. Add the overall reconstruction error value of the current detection sample to the historical reconstruction error sample set, and update the historical reconstruction error sample set based on the sliding time window mechanism or the cumulative sample update mechanism. Dynamically adjust the mean and standard deviation of the historical reconstruction error so that the dynamic reconstruction error judgment threshold reflects the state evolution trend of the hydropower unit rotor in different operating stages in real time, and maintains the timeliness and adaptability of the anomaly judgment standard.

[0128] In summary, this step uniquely introduces a timeliness and adaptability mechanism. Through dynamic thresholds and state grading, it can adapt to different operating stages (such as start-up and shutdown transitions), solve the false alarm problem caused by static thresholds, and ensure the long-term stability of the system. For example, the sliding window update mechanism allows the threshold to evolve with the data, reducing the impact of environmental drift.

[0129] Example 3 is an embodiment of the present invention, which provides a rotor acoustic anomaly detection method based on self-encoder and wavelet packet energy entropy. In order to verify the beneficial effects of the present invention, scientific demonstration is carried out through experiments.

[0130] During routine operation and inspection, a large hydropower station utilizes an acoustic anomaly detection system deployed beneath the rotor of Unit 5 turbine generator to monitor equipment status in real time. The system consists of six equidistantly positioned high-sensitivity MEMS acoustic sensors with a signal sampling frequency of 24kHz and a sampling window of one segment every 10 seconds. The monitoring data is transmitted to the detection platform in real time via an edge server.

[0131] During the detection process, the system detected energy disturbances in multiple frequency bands in sensor channel "CH-03". The system automatically input the acquired raw acoustic signal into the wavelet packet decomposition module, performed 5-level db6 wavelet basis decomposition, extracted the energy proportion features of 32 sub-frequency bands, and constructed an energy entropy feature vector. Subsequently, this vector was fed into a frequency band-gated autoencoder model that had been trained under normal operating conditions for reconstruction analysis.

[0132] The system detected an overall reconstruction error of 0.0417 in the 872nd sample, significantly higher than the current adaptively generated dynamic anomaly threshold of 0.0302. Further examination of the residual heatmap revealed that energy anomalies were concentrated in the 12th and 14th layer sub-bands (frequency range 2.7kHz–3.5kHz), with gating activation factor values ​​all above 0.92, and the fluctuation range of the anomalous energy residuals exceeding twice the mean.

[0133] The platform immediately generated an anomaly report with the number "AR-20241022-005", which recorded the following:

[0134] Anomaly occurred at 10:43:50 on October 22, 2024.

[0135] Abnormal location: Unit 5, CH-03 passage;

[0136] Abnormal frequency bands: Sub-band 12 (2.7–3.0kHz), Sub-band 14 (3.3–3.5kHz);

[0137] Anomalies: Reconstruction error 0.0417, maximum residual amplitude 0.0134;

[0138] Anomaly type: Suspected low-frequency periodic vibration anomaly, initially judged to be initial loosening of rotor bearing.

[0139] The system pushed the anomaly report to the dispatch and monitoring control room in real time via the northbound interface. After receiving the alarm, the on-duty engineer, Wu, immediately conducted a physical re-inspection of the unit and detected a slight misalignment in the second shaft section, which corresponded perfectly to the system report. The engineer tightened the bearings and adjusted the lubrication. Subsequently, the reconstruction error in the data re-collected by the system recovered to 0.0194, which was below the threshold. The anomaly was eliminated, and the monitoring status returned to green.

[0140] To compare the results, the experimental records are shown in Table 1. The team used the traditional short-time Fourier transform method to backtest the acoustic signal during this period. They found that although there was a slight energy perturbation in the amplitude diagram (maximum sound pressure amplitude was 0.17 Pa, background mean was 0.13 Pa), it was not marked as an anomaly because the overall change did not exceed the empirically set threshold of 0.20 Pa. If the original method were used, this anomaly would be missed, potentially leading to instability in subsequent equipment operation.

[0141] Table 1. Experimental records of the present invention and the traditional STFT threshold method.

[0142]

[0143]

[0144] Subsequently, three similar low-amplitude acoustic anomalies were detected in late October, namely at 15:10 on October 23, 08:32 on October 25, and 21:57 on October 27. The system successfully issued early warnings for all of them, which were later confirmed by manual inspection to correspond to: unbalanced cooling system fan, slight misalignment of coupler, and abnormal gear meshing, respectively.

[0145] Analysis of operational data over two consecutive weeks shows that the method of this invention achieves an anomaly identification accuracy of 96.2% under high noise and high operating load conditions, with an average response delay controlled within 3.1 seconds. Compared with traditional methods, the false alarm rate is reduced by 71.4%, effectively avoiding multiple potential unit damage risks and significantly improving the safe operation and maintenance level of hydropower units.

[0146] Example 1 verifies the accurate modeling and dynamic adaptation capabilities of the present invention in complex acoustic scenarios, providing a practical and feasible technical path for realizing intelligent monitoring of hydropower equipment throughout its entire life cycle.

[0147] This invention performs wavelet packet multi-scale decomposition on the integrated acoustic signal sequence, extracts multi-layer and multi-frequency band wavelet packet coefficients, constructs a frequency band energy proportion vector, and calculates the energy entropy feature vector based on this. This comprehensively characterizes the time-frequency structure and energy uncertainty of the acoustic signal during rotor operation. The feature extraction method has the advantages of high resolution accuracy and sensitivity to local energy disturbances, effectively improving the ability to capture minor faults under complex background noise conditions and enhancing the sensitivity and anti-interference ability of the fault early warning system.

[0148] The gated autoencoder model proposed in this invention takes the energy entropy feature vector as input, fuses frequency band activation factors, dynamically adjusts the representation weights of the input channels, and performs enhanced reconstruction of abnormal frequency bands through residual compensation branches. A sparse regularization term is added to the encoder to extract the most discriminative hidden features, and a compensation constraint mechanism for frequency abnormal segments is introduced in the decoder to prevent redundant information from interfering with the model expression. This significantly improves the modeling ability for local frequency anomalies, especially in small sample and perturbation anomaly scenarios, where it exhibits higher recognition accuracy and adaptability.

[0149] This invention does not rely on static threshold settings. Instead, it dynamically statistically models the reconstruction error under historical normal conditions, updates the mean and standard deviation in real time, generates a dynamic reconstruction error judgment threshold, and classifies abnormal states based on the comparison between the current detection error and the threshold. This effectively addresses the feature distribution drift caused by changes in operating status over time, improves the stability and false alarm control capability of the model in long-term monitoring, and the dynamic mechanism is more real-time and self-adaptive, making it widely adaptable to the multi-condition and multi-stage operating environment of hydropower unit rotors.

[0150] Example 4 is an embodiment of the present invention. The above is an illustrative scheme of the rotor acoustic anomaly detection method based on self-encoder and wavelet packet energy entropy. It should be noted that the technical solution of the rotor acoustic anomaly detection system based on self-encoder and wavelet packet energy entropy is of the same concept as the above-described rotor acoustic anomaly detection method based on self-encoder and wavelet packet energy entropy. Details not described in detail in this embodiment can be found in the description of the above-described rotor acoustic anomaly detection method based on self-encoder and wavelet packet energy entropy.

[0151] This embodiment provides a rotor acoustic anomaly detection system based on autoencoder and wavelet packet energy entropy, including: a data preprocessing module, a data fusion module, a wavelet packet transform module, an energy calculation module, a hidden feature module, and a feature reconstruction module;

[0152] The data preprocessing module collects the raw acoustic signals generated during the operation of the hydropower unit rotor and performs noise reduction processing on the raw acoustic signals.

[0153] The data fusion module performs multi-channel data fusion on the original acoustic signal after noise reduction, and merges the different original acoustic signals into an integrated acoustic signal;

[0154] The wavelet packet transform module takes the integrated acoustic signal as input to the wavelet packet transform process to obtain wavelet packet coefficients at different scales.

[0155] The energy calculation module calculates energy values ​​for signals in different frequency bands based on wavelet packet coefficients, forming an energy entropy feature vector.

[0156] The hidden feature module constructs a hidden feature model of the hydropower unit rotor under normal conditions based on the energy entropy feature vector;

[0157] The feature reconstruction module reconstructs the energy entropy feature vector, calculates the reconstruction error during the reconstruction process, performs dynamic statistical analysis on the reconstruction error, and monitors and judges the rotor operation status of the hydropower unit in real time.

[0158] This embodiment also provides an electronic device applicable to the rotor acoustic anomaly detection method based on self-encoder and wavelet packet energy entropy, comprising: a memory and a processor; the memory is used to store computer-executable instructions, and the processor is used to execute the computer-executable instructions to implement the rotor acoustic anomaly detection method based on self-encoder and wavelet packet energy entropy as proposed in the above embodiment.

[0159] This embodiment also provides a storage medium storing a computer program that, when executed by a processor, implements the rotor acoustic anomaly detection method based on self-encoder and wavelet packet energy entropy as proposed in the above embodiments.

[0160] The storage medium proposed in this embodiment and the rotor acoustic anomaly detection method based on self-encoder and wavelet packet energy entropy proposed in the above embodiments belong to the same inventive concept. Technical details not described in detail in this embodiment can be found in the above embodiments, and this embodiment has the same beneficial effects as the above embodiments.

[0161] Based on the above description of the implementation methods, those skilled in the art can clearly understand that the present invention can be implemented using software and necessary general-purpose hardware, and of course, it can also be implemented using hardware, but in many cases the former is a better implementation method. Based on this understanding, the technical solution of the present invention, or the part that contributes to the prior art, can be embodied in the form of a software product. This computer software product can be stored in a computer-readable storage medium, such as a computer floppy disk, read-only memory (ROM), random access memory (RAM), flash memory, hard disk, or optical disk, etc., including several instructions to cause a computer device (which may be a personal computer, server, or network device, etc.) to execute the methods of the various embodiments of the present invention.

[0162] It should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention and not to limit it. Although the present invention has been described in detail with reference to preferred embodiments, those skilled in the art should understand that modifications or equivalent substitutions can be made to the technical solutions of the present invention without departing from the spirit and scope of the technical solutions of the present invention, and all such modifications or substitutions should be covered within the scope of the claims of the present invention.

Claims

1. A rotor acoustic anomaly detection method based on self-encoder and wavelet packet energy entropy, characterized in that: include, The original acoustic signals generated during the operation of the hydropower unit rotor are collected, and the original acoustic signals are processed to reduce noise to obtain the original acoustic signal set; Multi-channel data fusion is performed on the original acoustic signal after noise reduction to fuse the different original acoustic signals into an integrated acoustic signal; The integrated acoustic signal is input into wavelet packet transform processing to obtain wavelet packet coefficients at different scales; Energy values ​​are calculated for signals in different frequency bands based on wavelet packet coefficients to form an energy entropy feature vector. Construct a hidden feature model of the hydropower unit rotor under normal conditions based on the energy entropy feature vector; The energy entropy eigenvector is reconstructed, and the reconstruction error during the reconstruction process is calculated. Dynamic statistical analysis is performed on the reconfiguration error, and the operating status of the hydropower unit rotor is monitored in real time and anomalies are identified.

2. The rotor acoustic anomaly detection method based on autoencoder and wavelet packet energy entropy as described in claim 1, characterized in that: The original acoustic signals include acoustic sensors arranged around the rotor of the hydropower unit to collect the original acoustic signals generated during the operation of the hydropower unit rotor and construct them into an original acoustic signal set. Each original acoustic signal in the original acoustic signal set is converted to a uniform format to obtain a standardized acoustic signal set; The standardized acoustic signal set is denoised to obtain the denoised acoustic signal set.

3. The rotor acoustic anomaly detection method based on autoencoder and wavelet packet energy entropy as described in claim 2, characterized in that: The integrated acoustic signal includes classifying and organizing the noise-reduced acoustic signal set to construct a noise-reduced acoustic signal subset; An adaptive weighting factor is assigned to a subset of acoustic signals, and the acoustic signals are weighted and summed according to the assigned adaptive weighting factor to form an integrated acoustic signal value consistent with the original time axis. The integrated acoustic signal values ​​are arranged sequentially in time to construct an integrated acoustic signal sequence set.

4. The rotor acoustic anomaly detection method based on autoencoder and wavelet packet energy entropy as described in claim 3, characterized in that: The process of obtaining wavelet packet coefficients at different scales includes inputting the constructed integrated acoustic signal sequence set into wavelet packet transform processing. The integrated acoustic signal is decomposed into sub-band signals under the conditions of preset wavelet basis function and decomposition level; Extract the corresponding wavelet packet coefficients from the sub-band signal and calculate the sub-band energy value; The energy values ​​of all sub-bands are normalized to construct a set of wavelet packet coefficients.

5. The rotor acoustic anomaly detection method based on autoencoder and wavelet packet energy entropy as described in claim 4, characterized in that: The formation of the energy entropy feature vector includes inputting the wavelet packet coefficient set into the energy calculation module for processing. The wavelet packet coefficient set contains wavelet packet coefficient sequences under all decomposition levels and their corresponding sub-frequency band numbers, wherein each wavelet packet coefficient sequence is composed of wavelet packet coefficients under the corresponding time sampling point index. Calculate the energy value corresponding to each sub-band for each wavelet packet coefficient sequence, normalize the energy values ​​of all sub-bands, and construct the energy proportion of each sub-band. The wavelet packet energy entropy characteristics are calculated based on the energy proportion of each sub-band, representing the entropy characteristics of the energy distribution of the integrated acoustic signal under wavelet packet decomposition. The energy proportions of each sub-band are combined according to their index order in the wavelet packet decomposition structure to form an energy entropy feature vector.

6. The rotor acoustic anomaly detection method based on autoencoder and wavelet packet energy entropy as described in claim 5, characterized in that: The implicit feature model includes inputting the energy entropy feature vector into the input layer of the gated autoencoder model and constructing a frequency band gated activation factor based on the energy proportion of each sub-band. The frequency band gate activation factor is applied to each frequency band channel in the energy entropy feature vector. A weighted input vector is generated by weighting the elements one by one. The weighted input vector is then input into the encoder structure. After weight matrix mapping and nonlinear activation processing, the hidden feature vector is output. The latent feature vector is input into the decoder to obtain the initial reconstructed vector. At the same time, a residual compensation path is introduced to generate the residual correction amount. Combined with the frequency band gated activation factor, differential weighted compensation is performed to obtain the reconstructed energy entropy feature vector.

7. The rotor acoustic anomaly detection method based on autoencoder and wavelet packet energy entropy as described in claim 6, characterized in that: The feature reconstruction includes inputting the energy entropy feature vector to be detected into the trained gated autoencoder model, and using the trained gated autoencoder model to encode and decode the energy entropy feature vector to obtain the corresponding latent feature representation and reconstruct the energy entropy feature vector, respectively. Calculate the reconstruction error vector between the original energy entropy eigenvector and the reconstructed energy entropy eigenvector, and calculate the overall reconstruction error value based on the reconstruction error vector; The overall reconstruction error values ​​corresponding to all historical samples are constructed into a historical reconstruction error sample set. Dynamic statistical analysis is performed on the historical reconstruction error sample set to determine the dynamic reconstruction error judgment threshold used in the current detection task. The overall reconstruction error value corresponding to the current detection sample is compared with the dynamic reconstruction error judgment threshold, and the current operating state of the hydropower unit rotor is determined accordingly. The overall reconstruction error value of the current detection sample is added to the historical reconstruction error sample set, and the historical reconstruction error sample set is updated based on the sliding time window mechanism or the cumulative sample update mechanism, dynamically adjusting the historical reconstruction error mean and the historical reconstruction error standard deviation.

8. A rotor acoustic anomaly detection system based on autoencoder and wavelet packet energy entropy, employing the rotor acoustic anomaly detection method based on autoencoder and wavelet packet energy entropy as described in any one of claims 1 to 7, characterized in that, include: Data preprocessing module, data fusion module, wavelet packet transform module, energy calculation module, hidden feature module, feature reconstruction module; The data preprocessing module collects the raw acoustic signals generated during the operation of the hydropower unit rotor and performs noise reduction processing on the raw acoustic signals. The data fusion module performs multi-channel data fusion on the original acoustic signal after noise reduction, and merges the different original acoustic signals into an integrated acoustic signal; The wavelet packet transform module takes the integrated acoustic signal as input to the wavelet packet transform process to obtain wavelet packet coefficients at different scales. The energy calculation module calculates energy values ​​for signals in different frequency bands based on wavelet packet coefficients, forming an energy entropy feature vector. The hidden feature module constructs a hidden feature model of the hydropower unit rotor under normal conditions based on the energy entropy feature vector; The feature reconstruction module reconstructs the energy entropy feature vector, calculates the reconstruction error during the reconstruction process, performs dynamic statistical analysis on the reconstruction error, and monitors and judges the rotor operation status of the hydropower unit in real time.

9. A computer device comprising a memory and a processor, wherein the memory stores a computer program, characterized in that, When the processor executes the computer program, it implements the steps of the rotor acoustic anomaly detection method based on autoencoder and wavelet packet energy entropy as described in any one of claims 1 to 7.

10. A computer-readable storage medium having a computer program stored thereon, characterized in that, When the computer program is executed by the processor, it implements the steps of the rotor acoustic anomaly detection method based on self-encoder and wavelet packet energy entropy as described in any one of claims 1 to 7.

Images