Rotary electromechanical equipment acoustic and vibration monitoring method, system, medium and equipment

Abstract

The invention discloses a rotary electromechanical equipment acoustics and vibration monitoring method and system, a medium and equipment, and the method comprises the steps: constructing a signal collection system, collecting sound and vibration signals in real time, carrying out the DC offset removal preprocessing, and removing the low-frequency interference; then, obtaining a time domain signal after noise reduction through adaptive noise reduction; then, multi-band energy characteristics of the noise reduction signals are extracted through wavelet packet decomposition; then, the order spectrum features are fused, and a multi-modal feature vector is constructed; and finally, constructing a multi-modal collaborative enhancement network model, and carrying out working condition anomaly classification by taking the multi-modal feature vector as input.

Description

Technical Field

[0001] This invention relates to the field of signal processing technology, and more specifically to methods, systems, media, and devices for acoustic and vibration monitoring of rotating electromechanical equipment. Background Technology

[0002] Rotating electromechanical equipment, as core equipment in energy systems and industrial production, directly impacts energy supply security and industrial production continuity through its operational status. During operation, the acoustic and vibration signals generated by rotating electromechanical equipment carry a wealth of equipment condition information. These signals originate from various physical processes such as rotor dynamics, bearing friction, gear meshing, and hydrodynamic interactions. Traditional monitoring methods primarily rely on single vibration sensors, which, while capable of capturing some mechanical fault characteristics, have significant limitations when dealing with the unique and complex operating conditions of rotating electromechanical equipment.

[0003] Current acoustic and vibration signal analysis faces the following key technical challenges: 1. Distortion in spectrum analysis caused by rotational speed fluctuations. Rotating electromechanical equipment inevitably experiences rotational speed fluctuations (typically within ±0.5%) during actual operation. These fluctuations lead to severe "spectral ambiguity" in traditional Fourier transform-based spectrum analysis. Key fault characteristic frequencies (such as bearing passage frequency, blade passage frequency, and gear meshing frequency) diffuse in the spectrum, significantly reducing amplitude measurement accuracy and severely affecting the identification of early faults. For example, minute amplitude changes caused by early bearing wear are often masked by spectral diffusion, leading to a higher false negative rate in fault detection. 2. Lack of physical meaning in feature extraction methods. While currently widely used time-domain statistical features (such as root mean square, kurtosis, and peak factor) and frequency-domain features (such as power spectral density) are computationally simple, they have weak correlation with the specific physical mechanisms of the equipment. These features cannot effectively distinguish different fault types and lack sensitivity to early faults. Especially in the analysis of broadband acoustic signals and complex vibration signals unique to rotating electromechanical equipment, traditional features often fail to capture key information with diagnostic value. 3. Insufficient multimodal signal fusion mechanisms. Acoustic and vibration signals, as two important modes reflecting the state of equipment, have obvious complementary characteristics: vibration signals are sensitive to mechanical structural faults, while acoustic signals can better capture features such as aerodynamic noise and electromagnetic noise. However, most existing methods use simple feature splicing or decision-level fusion, failing to fully explore the deep correlation between the two signals and unable to achieve true synergistic enhancement. 4. Signal processing challenges in non-stationary noise environments. The noise in the operating environment of rotating electromechanical equipment has significant non-stationary characteristics, including hydrodynamic noise, electromagnetic interference, and background noise. Traditional noise reduction methods such as spectral subtraction and Wiener filtering are based on the assumption of stationary noise, which can easily lead to signal distortion in practical applications, especially weakening the transient impact components in fault features. Although wavelet denoising methods can handle non-stationary signals, the threshold selection depends on experience and has poor adaptability under different operating conditions. 5. Limitations of adaptability of existing deep learning models. Although deep learning technology has made some progress in the field of fault diagnosis, existing models still have significant shortcomings in the monitoring of rotating electromechanical equipment: on the one hand, the models lack the embedding of the unique physical laws of rotating equipment, resulting in poor generalization ability when training data is insufficient; on the other hand, most models adopt fixed fusion strategies and cannot dynamically adjust the weight allocation of acoustic and vibration modes according to signal characteristics, making it difficult to adapt to the monitoring needs of different equipment types. 6. Cross-equipment type universality issues. Different types of rotating electromechanical equipment (such as hydroelectric generators, wind turbines, industrial motors, etc.) have their own unique operating characteristics and fault modes. Existing monitoring methods are often designed for specific equipment and lack universality. This limitation makes technology promotion difficult and increases system development and maintenance costs.

[0004] In summary, existing acoustic and vibration monitoring technologies for rotating machinery and equipment suffer from core problems such as distorted spectral analysis, lack of characteristic physical meaning, insufficient multimodal fusion, and poor cross-equipment adaptability. These issues result in an abnormal condition identification accuracy that fails to meet practical engineering requirements. There is an urgent need for an innovative method that can deeply integrate physical mechanisms, adapt to speed fluctuations, achieve intelligent multimodal fusion, and possess good versatility, in order to improve the reliability and intelligence level of condition monitoring for rotating machinery and equipment. Summary of the Invention

[0005] In view of this, the present invention provides a method, system, medium and device for acoustic and vibration monitoring of rotating machinery and equipment, so as to at least solve the problem that the existing acoustic and vibration monitoring technology for rotating machinery and equipment is difficult to efficiently suppress noise and accurately extract features.

[0006] To achieve the above objectives, the present invention adopts the following technical solution: A method for acoustic and vibration monitoring of rotating mechanical equipment, comprising the following steps: S1. Real-time acquisition of raw acoustic and vibration signals during the operation of rotating electromechanical equipment. ; ,in Acoustic signal, It is a vibration signal. The signal length; S2. A dual-time-scale noise estimation method is employed, which involves analyzing the original acoustic and vibration signals. The corresponding number Frame, First Amplitude spectrum at frequency points Calculate the short-time noise amplitude based on the statistical distribution along the time axis. and long-term noise amplitude ,based on and To estimate the noise amplitude And then according to To reconstruct the original acoustic and vibration signals, and obtain the denoised acoustic and vibration signals. , ,in For signal length, the denoised acoustic and vibration signals Including the noise-reduced acoustic signal and vibration signals ; S3. Based on wavelet packet decomposition, energy features are extracted from the acoustic and vibration signals of the denoised rotating electromechanical equipment, and then order spectrum features are fused to construct a multimodal feature vector; S4. The multimodal collaborative augmentation network MMSAN, trained with multimodal feature vectors, is used for anomaly detection. The multimodal collaborative augmentation network MMSAN is trained using denoised acoustic and vibration signals.

[0007] Preferably, after obtaining the original acoustic and vibration signals in S2, preprocessing is performed, specifically including: Obtained through time-frequency transformation The corresponding time-frequency representation matrix The time-frequency transformation methods specifically include: Short Time Fourier Transform (STFT), Fast Fourier Transform (FFT), wavelet transform, or wavelet packet transform. In the time-frequency representation matrix Based on this, the amplitude spectrum is calculated. : .

[0008] Preferably, in S2, the original acoustic and vibration signals are analyzed. Corresponding amplitude spectrum Calculate the short-time noise amplitude based on the statistical distribution along the time axis. and long-term noise amplitude The specific content includes: A dual-timescale noise estimation method is employed, and the amplitude spectrum is analyzed. Calculate the short-time noise amplitude based on the statistical distribution along the time axis. and long-term noise amplitude : ; ; In the formula, and They represent the first quantiles and the quantiles; and These are the quantile thresholds for short-time and long-time noise estimation, respectively; Reflects the local background level of noise. Capture the overall trend and possible sudden changes in noise.

[0009] Preferably, in S2 based on and To estimate the noise amplitude The specific content includes: Introducing dynamic weights To balance the contributions of short-time and long-time estimates: ; ; In the formula, As the initial weights, For time frame indexing, , For frequency index, , , The number of points in the time-frequency transformation. The imaginary unit, and for The upper and lower bounds of the range satisfy 0 ≤ < ≤1; Then the basic noise amplitude for: ; Based on the fundamental noise amplitude Calculate the signal-to-noise ratio Used to quantize the relative strength of signal and noise: ; Based on signal-to-noise amplitude ratio Calculate the minimum amplitude threshold : ; In the formula, The scaling factor controls the range of the minimum amplitude threshold to ensure that key signal features are not weakened. Introducing a dynamic noise scaling factor To dynamically adjust the noise amplitude: ; in, This is the scaling factor, which controls the sensitivity of the adjustment factor. These are the initial weights; Based on dynamic noise scaling factor Calculate and estimate noise amplitude : .

[0010] Preferably, in S2 according to To reconstruct the original acoustic and vibration signals, and obtain the denoised acoustic and vibration signals. The specific content includes: in accordance with and Calculate the noise reduction spectrum amplitude : ; Preserve the original phase and recover the denoised complex spectrum. : ; The inverse transform is used to reconstruct the time-domain signal from the denoised spectrum. : ; In the formula, Overlapping windowing functions; denoised acoustic and vibration signals Including the noise-reduced acoustic signal and vibration signals .

[0011] Preferably, the specific content of S3 includes: fusion of order analysis and wavelet packet energy proportion features: according to , and speed signal Calculate the cumulative rotation angle: ; in: Indicates cumulative angular displacement; This indicates the sampling frequency, i.e., the number of sampling points per second; , representing the sampling period, is used for time-angle conversion of angle increments; in a uniform angle grid, k represents the index of the current sampling point, corresponding to its position in the angle grid, and is used to identify the order of the signal sequence; To represent the cumulative angular displacement at the k-th sampling point, reflecting the cumulative angle of the device's rotation, used for subsequent angular domain signal generation; Grid spacing is , The number of sampling points per revolution, the output Used for spline interpolation to generate angular domain signals: ; Where M represents the length of the angular domain signal, which is determined by the original signal duration and the average rotational speed; For acoustic angular domain signals; The signal is a vibration angular domain signal; interpolation generates the angular domain signal to ensure that the signal is uniformly distributed in the angular domain, which facilitates subsequent order spectrum calculation and is suitable for the analysis of non-stationary rotation signals. Applying Discrete Fourier Transform to acoustic or vibration angular domain signals to perform order spectrum calculations: ; ; in: Indicates order, These represent the acoustic and vibrational order spectra, respectively. It represents the imaginary unit, used for phase representation in complex exponentiation; For noise reduction signals and Perform wavelet packet decomposition, using the db3 wavelet, to the Lth level; calculate the energy of each sub-band: ; ; in, Indicates the acoustic signal at the th Layer, First The energy of the frequency band; calculated using the sum of squared moduli of the wavelet coefficients; Indicates the decomposition hierarchy, from 1 to Used to control the depth of decomposition. Total number of floors; This indicates the frequency band index of the current level, divided into sub-bands; Indicates the acoustic signal at the th Layer, First The wavelet coefficients of the frequency band are the output of the wavelet packet transform, capturing local time-frequency features; Indicates the vibration signal at the 1st Layer, First The energy of the frequency band is calculated using the sum of squared moduli of the wavelet coefficients; Indicates the vibration signal at the 1st Layer, First Wavelet coefficients of the frequency band; The number of sampling points in this frequency band is represented as follows: ; ; in Indicates the acoustic signal at the th Layer, First The energy percentage of a frequency band is a normalized result, ranging from [0,1]. This represents the total energy of the acoustic signal across all levels and frequency bands, used to normalize the denominator; Indicates the acoustic signal at the th Layer, First The energy percentage of the frequency band. This value is a normalized result, ranging from [0,1]. This represents the total energy of the acoustic signal across all levels and frequency bands, used to normalize the denominator; The fusion of order spectrum and wavelet packet features is as follows: ; in, This represents the fused multimodal feature vector; It represents the acoustic order characteristics, including the critical frequency BPF and the amplitude and phase of its harmonics; Indicates the vibration order characteristics; This indicates the percentage of acoustic wavelet packet energy. This represents the proportion of vibration wavelet packet energy.

[0012] Preferably, the Convolutional Neural Network (CNN) model specifically includes: convolutional layers, pooling layers, fully connected layers, and an output layer, and is suitable for pattern recognition of non-stationary signals; The Multimodal Collaborative Augmentation Network (MMSAN) employs a dual-branch CNN to process acoustic and vibration features separately, ensuring modal specificity. For the acoustic branch, acoustic feature vectors are calculated, and then a convolution operation is performed: ; in: Indicates the branch of acoustics in the The output feature map of the layer contains the extracted local features, which are used for propagation in subsequent layers; This indicates the index of the current convolutional layer, identifying the depth position of the network; This represents the convolution operation function, used to perform convolution calculations; This represents the acoustic feature map output from the previous layer; For convolution kernel weights, For bias; BatchNorm represents batch normalization to stabilize and accelerate neural network training; For the vibration branch, calculate the vibration eigenvector: ; in, Indicates the vibration branch at the 1st The output feature map of the layer contains the extracted local features, which are used for subsequent layer propagation; This represents the vibration characteristic map output from the previous layer; For convolution kernel weights, For bias; By dynamically fusing acoustic and vibrational features through SAM, key modal information is amplified. Fusion characteristics: ; Collaborative weights are calculated using global average pooling and learnable weights to dynamically amplify key modal information: ; ; in, The collaborative weights of acoustic modes, ranging from [0,1], are used to quantify the importance of acoustic features; The cooperative weights of the vibration modes are defined in the range [0,1], and satisfy the following conditions: + =1; The activation function maps the input to [0,1] and is used to generate probabilistic weights; The learnable weight matrix representing the acoustic branch; A learnable weight matrix representing the vibration branches, used for vibration modes; This represents the global average pooling function, used to compress spatial dimensions, calculate the average value of each channel of the feature map, and output a one-dimensional vector; Final fusion features: ; Final fusion features Input the classifier to identify anomalies.

[0013] Preferably, it also includes: S5. Generating alarm signals based on detection results and generating decision suggestions, combined with a real-time feedback mechanism, to achieve closed-loop risk management.

[0014] An acoustic and vibration monitoring system for rotating mechanical equipment, based on the aforementioned acoustic and vibration monitoring method for any rotating mechanical equipment, includes: The signal acquisition module is used to acquire raw acoustic and vibration signals from rotating electromechanical equipment in real time during operation. ; ,in Acoustic signal, It is a vibration signal. The signal length; The noise reduction module is used to employ a dual-timescale noise estimation method by analyzing the original acoustic and vibration signals. The corresponding number Frame, First Amplitude spectrum at frequency points Calculate the short-time noise amplitude based on the statistical distribution along the time axis. and long-term noise amplitude ,based on and To estimate the noise amplitude And then according to To reconstruct the original acoustic and vibration signals, and obtain the denoised acoustic and vibration signals. , ,in For signal length, the denoised acoustic and vibration signals Including the noise-reduced acoustic signal and vibration signals ; The feature extraction module is used to extract energy features from the acoustic and vibration signals of the rotating electromechanical equipment after noise reduction based on wavelet packet decomposition. The detection module is used to input energy features into the trained multimodal collaborative augmentation network MMSAN for anomaly detection. The multimodal collaborative augmentation network MMSAN is trained using denoised acoustic and vibration signals.

[0015] Preferably, the noise reduction module analyzes the original acoustic and vibration signals. Corresponding amplitude spectrum Calculate the short-time noise amplitude based on the statistical distribution along the time axis. and long-term noise amplitude The specific content includes: A dual-timescale noise estimation method is employed, and the amplitude spectrum is analyzed. Calculate the short-time noise amplitude based on the statistical distribution along the time axis. and long-term noise amplitude : ; ; In the formula, and They represent the first quantiles and the quantiles; and These are the quantile thresholds for short-time and long-time noise estimation, respectively; Reflects the local background level of noise. Capture the overall trend and possible sudden changes in noise.

[0016] Preferably, the noise reduction module is based on and To estimate the noise amplitude The specific content includes: Introducing dynamic weights To balance the contributions of short-time and long-time estimates: ; ; In the formula, As the initial weights, For time frame indexing, , For frequency index, , , The number of points in the time-frequency transformation. The imaginary unit, and for The upper and lower bounds of the range satisfy 0 ≤ < ≤1; Then the basic noise amplitude for: ; Based on the fundamental noise amplitude Calculate the signal-to-noise ratio Used to quantize the relative strength of signal and noise: ; based on Calculate the minimum amplitude threshold : ; In the formula, The scaling factor controls the range of the minimum amplitude threshold to ensure that key signal features are not weakened. Introducing a dynamic noise scaling factor To dynamically adjust the noise amplitude: ; in, This is the scaling factor, which controls the sensitivity of the adjustment factor. These are the initial weights; Based on dynamic noise scaling factor Calculate and estimate noise amplitude : .

[0017] A computer-readable storage medium having a computer program stored thereon, which, when executed by a processor, implements the acoustic and vibration monitoring method for rotating electromechanical equipment as described above.

[0018] An electronic device includes a processor and a memory for storing one or more programs; when the processor executes the one or more programs, it implements the acoustic and vibration monitoring method for rotating electromechanical equipment as described above.

[0019] As can be seen from the above technical solution, compared with the prior art, the present invention discloses a method, system, medium, and device for acoustic and vibration monitoring of rotating electromechanical equipment. By integrating dual-timescale noise estimation, adaptive adjustment of signal-to-noise ratio (SNMR), feature calculation of wavelet packet basis, and anomaly detection using a CNN model, it achieves efficient suppression of broadband weak-amplitude noise and identification of abnormal operating conditions. This enables effective suppression of non-stationary noise and accurate detection of abnormal operating conditions in the acoustic and vibration signals of rotating electromechanical equipment, specifically including the following beneficial effects: 1. Dual-time-scale noise estimation and adaptive adjustment: In the time-frequency analysis stage, by extracting short-time and long-time noise amplitudes and dynamically fusing the basic noise amplitude, calculating SNMR and optimizing the noise reduction threshold, the dynamic distribution of noise is accurately captured, overcoming the limitations of static models and improving the suppression effect of broadband weak amplitude noise. 2. Feature fusion based on order analysis and wavelet packet energy proportion: The time-domain signal is converted into an angular-domain signal by equal-angle resampling technology, mitigating the impact of rotational speed fluctuations on the spectrum analysis. Combined with the multi-resolution analysis capabilities provided by wavelet packet energy proportion features, the constructed fused feature vector can effectively capture full-band fault information from low-frequency structural vibrations to high-frequency acoustic emissions; 3. Innovative Architectural Effects of Multimodal Cooperative Enhancement Network: The MMSAN network achieves adaptive weight allocation and dynamic information fusion of acoustic and vibration modes through the Cooperative Enhancement Module (SAM). Moreover, the dual-branch structure fully preserves the sensitivity of acoustic signals to hydrodynamic faults and the specificity of vibration signals to mechanical structural faults. Attached Figure Description

[0020] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the 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.

[0021] Figure 1 This is a schematic flowchart of the acoustic and vibration monitoring method for rotating electromechanical equipment provided by the present invention; Figure 2 This is a schematic diagram of the hardware system structure provided in an embodiment of the present invention; Figure 3 The present invention provides a time-frequency diagram of abnormal operating conditions of hydropower units; Figure 4 The noise reduction results for abnormal operating condition sounds provided in the embodiments of the present invention; Figure 5 The sub-band energy proportion characteristics provided in the embodiments of the present invention; wherein Figure 5 (a) ~ Figure 5 (h) are the curves showing the change of energy proportion of the first to eighth sub-bands over time, respectively; Figure 6 A schematic diagram illustrating the verification results of the model provided in this embodiment of the invention. Detailed Implementation

[0022] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0023] This invention provides a method for acoustic and vibration monitoring of rotating mechanical equipment, comprising the following steps: S1. Real-time acquisition of raw acoustic and vibration signals during the operation of rotating electromechanical equipment. ; ,in Acoustic signal, It is a vibration signal. The signal length; S2. A dual-time-scale noise estimation method is employed, which involves analyzing the original acoustic and vibration signals. The corresponding number Frame, First Amplitude spectrum at frequency points Calculate the short-time noise amplitude based on the statistical distribution along the time axis. and long-term noise amplitude ,based on and To estimate the noise amplitude And then according to To reconstruct the original acoustic and vibration signals, and obtain the denoised acoustic and vibration signals. , ,in For signal length, the denoised acoustic and vibration signals Including the noise-reduced acoustic signal and vibration signals ; S3. Based on wavelet packet decomposition, energy features are extracted from the acoustic and vibration signals of the denoised rotating electromechanical equipment, and then order spectrum features are fused to construct a multimodal feature vector; S4. The multimodal collaborative augmentation network MMSAN, trained with multimodal feature vectors, is used for anomaly detection. The multimodal collaborative augmentation network MMSAN is trained using denoised acoustic and vibration signals.

[0024] To further implement the above technical solution, the raw acoustic and vibration signals acquired in S2 are preprocessed, specifically including: Obtained through time-frequency transformation The corresponding time-frequency representation matrix The time-frequency transformation methods specifically include: Short Time Fourier Transform (STFT), Fast Fourier Transform (FFT), wavelet transform, or wavelet packet transform. In the time-frequency representation matrix Based on this, the amplitude spectrum is calculated. : .

[0025] It should be noted that: Input signal The acoustic and vibration signals of the rotating electromechanical equipment were collected, among which Let be the signal length. In the initial stage, the signal is preprocessed to remove the DC offset, effectively eliminating DC component interference, resulting in a preprocessed time-domain signal. Removing DC offset can be achieved using various techniques, including but not limited to traditional mean subtraction, high-pass filtering to suppress low-frequency DC components, and adaptive filtering techniques.

[0026] To further implement the above technical solution, S2 analyzes the original acoustic and vibration signals. Corresponding amplitude spectrum Calculate the short-time noise amplitude based on the statistical distribution along the time axis. and long-term noise amplitude The specific content includes: A dual-timescale noise estimation method is employed, and the amplitude spectrum is analyzed. Calculate the short-time noise amplitude based on the statistical distribution along the time axis. and long-term noise amplitude : ; ; In the formula, and They represent the first quantiles and the quantiles; and These are the quantile thresholds for short-time and long-time noise estimation, respectively; Reflects the local background level of noise. Capture the overall trend and possible sudden changes in noise.

[0027] To further implement the above technical solution, S2 is based on and To estimate the noise amplitude The specific content includes: Introducing dynamic weights To balance the contributions of short-time and long-time estimates: ; ; In the formula, As the initial weights, For time frame indexing, , The frequency index is the frequency index of the time-frequency representation matrix. , , For the number of points in the time-frequency transformation, The imaginary unit, and for The upper and lower bounds of the range satisfy 0 ≤ < ≤1; Then the basic noise amplitude for: ; Based on the fundamental noise amplitude Calculate the signal-to-noise ratio Used to quantize the relative strength of signal and noise: ; based on Calculate the minimum amplitude threshold : ; In the formula, The scaling factor controls the range of the minimum amplitude threshold to ensure that key signal features are not weakened. Introducing a dynamic noise scaling factor To dynamically adjust the noise amplitude: ; in, This is the scaling factor, which controls the sensitivity of the adjustment factor. As the initial weights, according to the formula above calculate; Based on dynamic noise scaling factor Calculate and estimate noise amplitude : .

[0028] It should be noted that: The noise in the acoustic and vibration signals of rotating electromechanical equipment exhibits both short-term bursts and long-term stationarity, making it difficult to accurately capture its dynamic distribution using single-scale noise estimation. To address this issue, this invention employs dual-time-scale noise estimation. and The noise estimation thresholds for short-term and long-term noise are respectively. They can be calculated using statistical methods or other methods, including but not limited to multiples of the mean plus the standard deviation or K-means clustering, to optimize noise estimation performance.

[0029] Reflects the local background level of noise. Capture the overall trend and potential sudden changes in noise. To integrate short-term and long-term noise characteristics, dynamic weights are introduced. To balance the contributions of the two estimates. Weights It reflects the average strength of the signal relative to long-term noise and is used to adjust the contribution ratio of short-time and long-time estimates.

[0030] To effectively denoise, this invention calculates the signal-to-noise ratio (SNMR) based on the baseline noise amplitude, which is used to quantify the relative strength of the signal and noise. To avoid over-suppression of the signal during denoising, a minimum amplitude threshold is further calculated based on SNMR.

[0031] To further implement the above technical solution, S2 is based on... To reconstruct the original acoustic and vibration signals, and obtain the denoised acoustic and vibration signals. The specific content includes: in accordance with and Calculate the noise reduction spectrum amplitude : ; Preserve the original phase and recover the denoised complex spectrum. : ; The inverse transform is used to reconstruct the time-domain signal from the denoised spectrum. : ; In the formula, Overlapping windowing functions; denoised acoustic and vibration signals Including the noise-reduced acoustic signal and vibration signals .

[0032] To further implement the above technical solution, S3 specifically includes: fusion of order analysis and wavelet packet energy proportion features. Rotating electromechanical equipment (such as generator sets) has its rotational speed controlled within a target range (e.g., 100–120 RPM), but actual speeds exhibit minute fluctuations (0.1–1%) due to load variations or hydraulic turbulence. These fluctuations cause spectral blurring in conventional FFT: energy dispersion at critical frequencies (e.g., BPF) reduces fault detection accuracy. Order analysis, by standardizing the rotational period, provides... Radius maps the signal to the angular domain, ensuring that the characteristic frequency is fixed at a specific order, enabling comparability across operating conditions: consistent characteristics at different speeds, facilitating trend analysis; fault source separation: distinguishing between high-order (e.g., gear meshing ~100th order), medium-order (e.g., bearing failure ~5–10th order), and low-order (e.g., blade passage ~3–6th order); nonlinear detection: identifying non-integer orders (e.g., 1.5th order) caused by cracks. Wavelet packet features supplement broadband non-stationary characteristics (e.g., transient impacts), complementing the order spectrum to form a comprehensive feature representation.

[0033] First, the acoustic signal is resampled at an equal angle, and then the noise reduction signal is input. (Acoustics) and (Vibration), and rotational speed signal (Unit: RPM, obtained from the unit's speed sensor via the Fast API interface). Calculate the cumulative rotation angle: ; in: Indicates cumulative angular displacement; This indicates the sampling frequency, measured in Hertz (Hz), which is the number of sampling points per second. , representing the sampling period, is used for time-angle conversion of angle increments.

[0034] In a uniform angle grid , This represents the index of the current sampling point, corresponding to its position in the angle grid, for example, the index of the sampling point. The accumulated values ​​of each angle grid are used to identify the order of the signal sequence; To indicate the first The cumulative angular displacement at each sampling point reflects the cumulative angle of the device's rotation and is used for subsequent angular domain signal generation.

[0035] Grid spacing is , The number of sampling points per revolution, the output Used for spline interpolation to generate angular domain signals: ; Where M represents the length of the angular domain signal, which is determined by the original signal duration and the average rotational speed; For acoustic angular domain signals; This is for the vibration angular domain signal. This formula ensures the signal is uniformly distributed in the angular domain, facilitating subsequent order spectrum calculations and making it suitable for the analysis of non-stationary rotational signals. Applying Discrete Fourier Transform to acoustic or vibration angular domain signals to perform order spectrum calculations: ; ; in: Indicates order, These represent the acoustic and vibrational order spectra, respectively, and M represents the angular domain signal length. The imaginary unit is used for phase representation in complex exponentiation. This formula outputs the order spectrum, which is used for subsequent feature fusion.

[0036] For noise reduction signals and Perform wavelet packet decomposition, using the db3 wavelet, to the Lth level; calculate the energy of each sub-band: ; ; in, Indicates the acoustic signal at the th Layer, First The energy of the frequency band is calculated using the sum of squared moduli of the wavelet coefficients. : Indicates the decomposition level, from 1 to This parameter controls the decomposition depth. This represents the total number of floors. : Indicates the frequency band index of the current level, which is a sub-frequency band. Indicates the acoustic signal at the th Layer, First Wavelet coefficients for the frequency band. These coefficients are the output of the wavelet packet transform, capturing local time-frequency features. Furthermore, Indicates the vibration signal at the 1st Layer, First The energy of the frequency band is calculated using the sum of squared moduli of the wavelet coefficients. Indicates the vibration signal at the 1st Layer, First Wavelet coefficients of the frequency band. The number of sampling points in this frequency band is represented as follows: ; ; in, Indicates the acoustic signal at the th Layer, First The energy percentage of the frequency band. This value is a normalized result, ranging from [0,1]. Indicates the first Layer, First Acoustic energy of the frequency band. Calculated using the previous formula. This represents the total energy of the acoustic signal across all levels and frequency bands, used to normalize the denominator. Indicates the acoustic signal at the th Layer, First The energy percentage of the frequency band. This value is a normalized result, ranging from [0,1]. Indicates the first Layer, First Acoustic energy of the frequency band. Calculated using the previous formula. This represents the total energy of the acoustic signal across all levels and frequency bands, used to normalize the denominator.

[0037] The fusion of order spectrum and wavelet packet features is as follows: ; in, This represents the fused multimodal feature vector; It represents the acoustic order characteristics, including the critical frequency BPF and the amplitude and phase of its harmonics; Indicates the vibration order characteristics; This indicates the percentage of acoustic wavelet packet energy. This represents the proportion of vibration wavelet packet energy.

[0038] To further implement the above technical solutions, the Convolutional Neural Network (CNN) model specifically includes: convolutional layers, pooling layers, fully connected layers, and an output layer, which is suitable for pattern recognition of non-stationary signals. MMSAN employs a dual-branch CNN to process acoustic and vibration features separately, ensuring modality specificity. For the acoustic branch, acoustic feature vectors are further calculated, followed by convolution operations. ; in: Indicates the branch of acoustics in the The output feature map of a layer is usually a multidimensional tensor containing extracted local features, which is used for propagation in subsequent layers. This indicates the index of the current convolutional layer, identifying the depth position of the network. This represents the convolution operation function. This function performs convolution calculations. This represents the acoustic feature map output from the previous layer. For convolution kernel weights, This is the bias. Batch Normalization represents batch normalization to stabilize and accelerate neural network training.

[0039] For the vibration branch, the vibration eigenvector is further calculated: ; in, Indicates the vibration branch at the 1st The output feature map of a layer is usually a multidimensional tensor containing extracted local features, which is used for propagation in subsequent layers. This indicates the index of the current convolutional layer, identifying the depth position of the network. This represents the convolution operation function. This function performs convolution calculations. This represents the vibration characteristic map output from the previous layer. For convolution kernel weights, This is the bias. Batch Normalization represents batch normalization to stabilize and accelerate neural network training.

[0040] Furthermore, by dynamically fusing acoustic and vibrational features through SAM, key modal information is amplified. This is achieved by first further fusing the features: ; This process concatenates the final output feature maps (Lth layer) of the two branches along a specified dimension to form joint features. The initial fusion of acoustic and vibration information was completed. Further, collaborative weights were calculated using global average pooling (GAP) and learnable weights to dynamically amplify key modal information.

[0041] ; ; in, This represents the collaborative weight of the acoustic modes. The value ranges from [0,1] and quantifies the importance of the acoustic features. This represents the cooperative weight of the vibration modes. The value ranges from [0,1] and satisfies... + =1. This is the activation function. It maps the input to [0,1] and is used to generate probabilistic weights. This represents the learnable weight matrix for the acoustic branch. This is the learnable weight matrix representing the vibration branch. This parameter is similar to W_s and is used for vibration modes. This represents the global average pooling function. This function compresses the spatial dimension, calculates the average value of each channel of the feature map, and outputs a one-dimensional vector.

[0042] Final fusion features: ; Final fusion features Input the classifier to identify anomalies.

[0043] To further implement the above technical solution, it also includes: S5. Generating alarm signals based on the detection results and generating decision suggestions, combined with a real-time feedback mechanism, to achieve closed-loop risk management.

[0044] An acoustic and vibration monitoring system for rotating mechanical equipment, based on the aforementioned acoustic and vibration monitoring method for any rotating mechanical equipment, includes: The signal acquisition module is used to acquire raw acoustic and vibration signals from rotating electromechanical equipment in real time during operation. ; ,in Acoustic signal, It is a vibration signal. The signal length; The noise reduction module is used to employ a dual-timescale noise estimation method by analyzing the original acoustic and vibration signals. The corresponding number Frame, First Amplitude spectrum at frequency points Calculate the short-time noise amplitude based on the statistical distribution along the time axis. and long-term noise amplitude ,based on and To estimate the noise amplitude And then according to To reconstruct the original acoustic and vibration signals, and obtain the denoised acoustic and vibration signals. , ,in For signal length, the denoised acoustic and vibration signals Including the noise-reduced acoustic signal and vibration signals ; The feature extraction module is used to extract energy features from the acoustic and vibration signals of the rotating electromechanical equipment after noise reduction based on wavelet packet decomposition. The detection module is used to input energy features into the trained multimodal collaborative augmentation network MMSAN for anomaly detection. The multimodal collaborative augmentation network MMSAN is trained using denoised acoustic and vibration signals.

[0045] It should be noted that: The signal acquisition module includes signal acquisition devices installed at different locations on the rotating electromechanical equipment and at the main transformer. The raw signals are amplified and filtered by conditioning circuits to adapt to wideband noise environments and improve signal quality. It also includes a signal acquisition card, which is used to digitally acquire the filtered signals and transmit them to the host computer via Ethernet.

[0046] The front-end data acquisition device consists of audio sensors, vibration sensors, and synchronous data acquisition cards, which are placed at various key locations on the rotating electromechanical equipment. The number and location of the sensors can be flexibly set according to the on-site acquisition requirements. A noise reduction module, feature extraction module, and detection module run in the host computer, serving as the data storage and processing unit. Installed in the central control room, these modules primarily perform subsequent signal filtering, feature calculation, and abnormal operating condition detection.

[0047] To further implement the above technical solution, the noise reduction module analyzes the original acoustic and vibration signals. Corresponding amplitude spectrum Calculate the short-time noise amplitude based on the statistical distribution along the time axis. and long-term noise amplitude The specific content includes: A dual-timescale noise estimation method is employed, and the amplitude spectrum is analyzed. Calculate the short-time noise amplitude based on the statistical distribution along the time axis. and long-term noise amplitude : ; ; In the formula, and They represent the first quantiles and the quantiles; and These are the quantile thresholds for short-time and long-time noise estimation, respectively; Reflects the local background level of noise. Capture the overall trend and possible sudden changes in noise.

[0048] To further implement the above technical solution, the noise reduction module is based on... and To estimate the noise amplitude The specific content includes: Introducing dynamic weights To balance the contributions of short-time and long-time estimates: ; ; In the formula, As the initial weights, For time frame indexing, , For frequency index, , , The number of points in the time-frequency transformation. The imaginary unit, and for The upper and lower bounds of the range satisfy 0 ≤ < ≤1; Then the basic noise amplitude for: ; Based on the fundamental noise amplitude Calculate the signal-to-noise ratio Used to quantize the relative strength of signal and noise: ; based on Calculate the minimum amplitude threshold : ; In the formula, The scaling factor controls the range of the minimum amplitude threshold to ensure that key signal features are not weakened. Introducing a dynamic noise scaling factor To dynamically adjust the noise amplitude: ; in, This is the scaling factor, which controls the sensitivity of the adjustment factor. These are the initial weights; Based on dynamic noise scaling factor Calculate and estimate noise amplitude : .

[0049] A computer-readable storage medium having a computer program stored thereon, which, when executed by a processor, implements the acoustic and vibration monitoring method for rotating electromechanical equipment as described above.

[0050] An electronic device includes a processor and a memory for storing one or more programs; when the processor executes the one or more programs, it implements the acoustic and vibration monitoring method for rotating electromechanical equipment as described above.

[0051] The invention will be further illustrated below through specific experiments: This invention applies to rotating electromechanical equipment, which mainly includes all rotating machinery that realizes the mutual conversion of electrical energy and mechanical energy. Its core consists of two main categories: generators (such as hydropower, thermal power, and wind power generators) and electric motors (such as motors that drive water pumps, fans, and machine tools). In this embodiment, a hydropower unit is used as an example to further illustrate the method.

[0052] like Figure 2 As shown, the multi-channel distributed high-resolution sound and vibration signal acquisition scheme for hydroelectric generating units mainly includes audio sensors, vibration sensors, signal acquisition cards, and a host computer. Analog signals are acquired by the sensors, then processed by conditioning circuits and the acquisition card to output digital signals, which are then transmitted to the server via Ethernet for processing and analysis.

[0053] In this example, the main parameter is configured as follows: window length The settings are: 256 sampling points, overlap length: 128 sampling points, FFT point count: 512, and short-time noise amplitude. and long-term noise amplitude The quantiles are taken as 25 percentile and 75 percentile, respectively. Set the lower bound to 0.5 for weight calculation. Set the weight calculation to 0.9 as the upper limit. Set 0.005 as the base value for the minimum amplitude threshold. Set it to 1.5 as the dynamic noise scaling factor parameter.

[0054] Figure 3 This diagram presents a spectral analysis of the collected abnormal operating sound from the hydroelectric generator. The horizontal axis represents time (in seconds), ranging from 0 to 10 seconds; the vertical axis represents frequency (in Hz), ranging from 0 to 16000 Hz. The amplitude (in dB) gradually transitions from a higher amplitude (approximately -10 dB) to a lower amplitude (approximately -40 dB) to display the intensity distribution of the sound signal. This abnormal sound was caused by a foreign object entering the hydroelectric generator, resulting in abnormal frequency bands and sound characteristics. The abnormal frequency bands are shown in the marked area of ​​the figure, forming a stark contrast with the background noise.

[0055] The acoustic and vibration signals were sampled at a rate of 32,000 Hz, with a total duration of approximately 10 seconds (320,000 sampling points). After normalization and removal of DC offset, a time-frequency representation matrix was generated, and the amplitude spectrum was further calculated. The resulting amplitude spectrum size was (257, 2501), corresponding to 257 frequency components and 2501 time frames, reflecting the time-frequency characteristics of the approximately 10-second signal. The amplitude spectrum statistics are shown in Table 1.

[0056]

[0057] By analyzing the 10-second amplitude spectrum, short-time and long-time noise amplitudes were calculated using a multi-scale method. Short-time and long-time noise matrices were generated based on the 25th and 75th quantiles, respectively, with a matrix size of (257, 1). Then, based on the formula... ,calculate The value is 0.1374, and its output is shown in Tables 2 and 3.

[0058]

[0059]

[0060] Using the fundamental noise matrix and amplitude spectrum The SNMR, minimum amplitude threshold, dynamic noise scaling factor, and final noise amplitude are calculated and their outputs are shown in Tables 4 and 5.

[0061]

[0062]

[0063] Using the final noise amplitude, the final amplitude spectrum matrix is ​​calculated, the complex spectrum matrix is ​​reconstructed, and the denoised signal is reconstructed. The output is shown in Table 6.

[0064]

[0065] Figure 4 The audio spectrum analysis after applying this noise reduction method is shown. The horizontal axis represents time (in seconds), ranging from 0 to 10 seconds; the vertical axis represents frequency (in Hz), ranging from 0 to 16000 Hz. Amplitude range and Figure 3 Consistent. Through with Figure 3 As can be seen from the comparison, the method of the present invention significantly suppresses the background noise frequency band (manifested as an increase in large areas of low amplitude in the figure), while effectively preserving the abnormal characteristic frequency band caused by foreign objects inside the hydropower unit (manifested as a more prominent high amplitude area). This indicates that the noise reduction method can effectively maintain the relative intensity of key abnormal features while removing interference noise.

[0066] In the feature extraction stage, this example employs a multi-level feature fusion strategy to achieve a comprehensive characterization of the operating status of the hydropower unit. For example... Figure 5 As shown, the denoised and reconstructed signal is first decomposed using a three-level decomposition based on the db3 wavelet basis, and the wavelet packet energy proportion characteristics of each sub-band are extracted. The analysis results show that the energy of the eight sub-bands exhibits obvious periodic peak fluctuations during the 0-10 second monitoring period, accurately reflecting the transient energy concentration phenomenon caused by the impact of foreign objects inside the turbine.

[0067] Building upon this foundation, this example innovatively integrates wavelet packet energy proportion features with order analysis. Addressing the speed fluctuation issue (measured fluctuation range ±0.7%) in hydropower unit operation, an equal-angle resampling technique is employed to convert the time-domain signal into an angular-domain signal, effectively eliminating the impact of speed fluctuations on feature extraction. Through angular-domain Fourier transform, the amplitude and phase information of key fault features such as blade throughput frequency and bearing throughput frequency are accurately extracted, with a phase consistency coefficient reaching 0.89, indicating excellent feature stability. The final fused feature vector F contains 42 dimensions, specifically: 8 dimensions of acoustic signal order features (including amplitude and phase from fundamental frequency to fourth harmonic); 8 dimensions of vibration signal order features; 8 dimensions of acoustic signal wavelet packet energy proportion features; 8 dimensions of vibration signal wavelet packet energy proportion features; and 10 dimensions of cross-modal correlation features (including order spectrum correlation coefficient and energy distribution similarity). This feature fusion scheme fully leverages the advantages of each feature: the wavelet packet energy proportion features provide multi-resolution analysis capabilities across the 0-16kHz frequency band, effectively capturing transient impact components; the order features ensure feature stability under fluctuating rotational speed conditions, with the coefficient of variation controlled within 5%; and the cross-modal correlation features deeply explore the coupling relationship between acoustic and vibration signals, laying the foundation for collaborative analysis of multi-source information.

[0068] Based on the aforementioned 42-dimensional fused feature vector, this example constructs an innovative Multimodal Collaborative Augmentation Network (MMSAN) architecture to achieve accurate identification and classification of abnormal operating conditions in hydropower units. The network employs a dual-branch symmetrical structure, processing acoustic and vibration feature flows separately, and maintaining modal specificity through parameter-independent convolutional channels. The specific network configuration is as follows: the input layer receives a 42-dimensional feature vector, which is reconstructed and then input into the acoustic and vibration branches respectively; each branch contains three convolutional layers with kernel numbers of 16, 32, and 64 respectively, all using 3×3 convolutional kernels with a stride of 1, combined with ReLU activation and batch normalization; the pooling layer uses 2×2 max pooling with a stride of 2.

[0069] The core innovation of the network lies in the design of the Collaborative Enhancement Module (SAM), which dynamically evaluates the contribution of each modality through an attention mechanism. In this example, the weight assignments calculated by the SAM module are 0.6 and 0.37, respectively, indicating that acoustic features have a higher importance in the current fault detection task. The fused features are then input into a fully connected classifier after global average pooling, and finally the classification probabilities of normal and abnormal states are output through the Softmax function.

[0070] Performance test results show that the MMSAN model achieved excellent performance on 215 independent test samples: overall accuracy reached 98.14%, precision 96.23%, recall 96.23%, and F1 score 96.23%. Of particular note is the excellent balance between high accuracy and recall, with precision and recall being completely consistent, demonstrating the model's ability to balance normal and abnormal operating conditions. In terms of error analysis, the false positive rate was controlled at 1.23%, and the false negative rate at 3.77%, indicating that the model maintains good sensitivity to real anomalies while avoiding misjudgments. Regarding real-time performance, the average inference time is only 86ms, fully meeting the real-time requirements of online monitoring of hydropower units.

[0071] The dataset of 2150 samples shows that 1620 samples were under normal operating conditions (75.3%), and 530 samples were under abnormal operating conditions (24.7%). This distribution accurately reflects the dominance of normal conditions in the actual operation of hydropower units. In the test set of 215 samples, there were 162 normal samples and 53 abnormal samples. The confusion matrix of the model on this test set shows: 160 true normal samples, 2 false abnormal samples, 2 false normal samples, and 51 true abnormal samples. This result verifies the reliability and stability of the MMSAN model in practical engineering applications, providing strong technical support for intelligent operation and maintenance of hydropower units.

[0072] The above embodiments are only used to illustrate the technical solutions of this application, and are not intended to limit them. Although this application has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some of the technical features. Such modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the spirit and scope of the technical solutions of the embodiments of this application, and should all be included within the protection scope of this application.

Claims

1. A method of acoustic and vibration monitoring of a rotating electromechanical equipment, characterized in that, Comprising the following steps: S1. Collecting raw acoustic and vibration signals of a rotating electromechanical equipment in real time ; wherein is the acoustic signal, is the vibration signal, is the signal length; S2. Adopting the double-time-scale noise estimation method, by analyzing the original acoustic and vibration signals The corresponding first frame, the first frequency point of the amplitude spectrum Along the time axis, the statistical distribution is calculated respectively short-time noise amplitude And long-time noise amplitude , based on And Estimate the estimated noise amplitude , and then according to Reconstruct the original acoustic and vibration signals to get the noise-reduced acoustic and vibration signals , , wherein The signal length is The noise-reduced acoustic and vibration signals Include noise-reduced acoustic signal And vibration signal S3. After energy feature extraction of the denoised acoustic and vibration signals of the rotating electromechanical equipment based on wavelet packet decomposition, the order spectrum features are fused to construct a multi-modal feature vector; S4. The multi-modal collaborative enhancement network MMSAN trained by the multi-modal feature vector is used for working condition anomaly detection, wherein the multi-modal collaborative enhancement network MMSAN is trained by the denoised acoustic and vibration signals.

2. The rotating electro-mechanical device acoustic and vibration monitoring method of claim 1, wherein, After the original acoustic and vibration signals are obtained in S2, preprocessing is performed, and the specific content includes: obtained by time-frequency transformation a corresponding time-frequency representation matrix wherein the method by time-frequency transformation specifically comprises a short-time Fourier transform (STFT), a fast Fourier transform (FFT), a wavelet transform or a wavelet packet transform. On the basis of the time-frequency representation matrix the amplitude spectrum is calculated 。 3. The rotating electro-mechanical device acoustic and vibration monitoring method of claim 1, wherein, S2 in which the original acoustic and vibration signals are analyzed the corresponding amplitude spectrum the statistical distribution along the time axis, respectively calculating the short-time noise amplitude and the long-time noise amplitude The specific content includes: Using a double-time-scale noise estimation, the short-time noise amplitude and the long-time noise amplitude are calculated by analyzing the amplitude spectrum along the time axis ； ； wherein and denote the 1st fractile and the 99th fractile, respectively; and are the fractile thresholds for the short-time and long-time noise estimates, respectively; reflects the local background level of the noise, captures the overall trend and possible abrupt changes of the noise.

4. The rotating electro-mechanical device acoustic and vibration monitoring method of claim 3, wherein, S2 based on and to estimate the estimated noise amplitude The details include: Introducing dynamic weights to balance the contribution of both short and long estimates: ； ； wherein is an initial weight, is a time frame index, , is a frequency index, , , is a number of points of the time-frequency transform, is an imaginary unit, and is upper and lower bounds of a range, satisfying 0≤ <1 <1 the base noise amplitude is: ； Based on the base noise amplitude Computing the signal noise amplitude ratio for quantifying the relative strength of signal and noise: ； Based on signal to noise amplitude ratio Computing a minimum amplitude threshold : ； wherein is a scaling factor that controls the range of the minimum amplitude threshold, ensuring that the signal's key features are not attenuated; Introducing dynamic noise scaling factor to dynamically adjust the noise amplitude: ； wherein, is a scaling factor, controlling the sensitivity of the adjustment factor, is an initial weight; Dynamic noise scaling factor Computing an estimated noise amplitude : 。 5. The rotating electro-mechanical device acoustic and vibration monitoring method of claim 4, wherein, S2 in accordance with reconstruct the original acoustic and vibration signals to obtain the noise-reduced acoustic and vibration signals The specific content includes: In accordance with and computing the noise reduction spectral magnitudes : ； Preserving original phase, restoring complex spectrum after noise reduction : ； reconstructing the noise-reduced spectrum into a time-domain signal using an inverse transform : ； wherein is an overlap window function; denoised acoustic and vibration signals comprising denoised acoustic signals and vibration signals .

6. The rotating electromechanical device acoustic and vibration monitoring method of claim 1, wherein, The specific content of S3 includes order analysis and wavelet packet energy proportion feature fusion; According to , and the rotational speed signal the cumulative rotation angle is calculated: ； wherein: represents the cumulative angular displacement; represents the sampling frequency, i.e. the number of samples per second; , represents the sampling period, for time-angle conversion of angular increments; in a uniform angular grid, , represents the index of the current sampling point, corresponding to the position in the angular grid, for identifying the sequence order of the signal; is the cumulative angular displacement at the k-th sampling point, reflecting the cumulative angle of the device rotation, for subsequent angular domain signal generation; The grid spacing is , The number of samples per revolution is For spline interpolation to generate angular domain signal: ； Wherein, M represents the angle domain signal length, determined by the original signal time length and average rotating speed; is an acoustic angle domain signal; is a vibration angle domain signal; the interpolation generated angle domain signal is used to ensure the uniform distribution of the signal in the angle domain, facilitate the subsequent order spectrum calculation, and is suitable for the analysis of non-stationary rotating signals; Discrete Fourier transform is applied to the acoustic or vibration angular domain signal to complete order spectrum calculation: ； ； where: represents an order, respectively represent the acoustic and vibration order spectrum; represents the imaginary unit, used for phase representation in complex exponential operations; To a noise reduction signal And Wavelet packet decomposition is performed, db3 wavelet is selected, and decomposition is performed to the Lth layer; and energy of each sub-band is calculated: ； ； wherein, represents the energy of the acoustic signal in the jth layer, the ith frequency band; calculated by the sum of the modulus square of the wavelet coefficients; represents the decomposition level, from 1 to , for controlling the decomposition depth, is the total number of layers; represents the frequency band index of the current level, distinguishing the sub-frequency band; represents the wavelet coefficients of the acoustic signal in the jth layer, the ith frequency band; is the output of the wavelet packet transform, capturing the local time-frequency features; represents the energy of the vibration signal in the jth layer, the ith frequency band; calculated by the sum of the modulus square of the wavelet coefficients; represents the wavelet coefficients of the vibration signal in the jth layer, the ith frequency band; represents the number of sampling points in the frequency band, then the energy proportion is represented as:​​​​ ； ； wherein represents the energy proportion of the acoustic signal in the i-th layer, j-th frequency band. The value is a normalized result, ranging from [0, 1]; represents the total energy of the acoustic signal in all layers and frequency bands, used for the normalization denominator; represents the energy proportion of the acoustic signal in the i-th layer, j-th frequency band. The value is a normalized result, ranging from [0, 1]; represents the total energy of the acoustic signal in all layers and frequency bands, used for the normalization denominator; represents the energy proportion of the acoustic signal in the i-th layer, j-th frequency band. The value is a normalized result, ranging from [0, 1]; represents the total energy of the acoustic signal in all layers and frequency bands, used for the normalization denominator; represents the energy proportion of the acoustic signal in the i-th layer, j-th frequency band. The value is a normalized result, ranging from [0, 1]; represents the total energy of the acoustic signal in all layers and frequency bands, used for the normalization denominator; The order spectrum and wavelet packet features are fused as: ； wherein, represents a fused multimodal feature vector; represents acoustic order features including amplitudes and phases of the key frequencies BPF and their harmonics; represents vibration order features; represents acoustic wavelet packet energy proportion; represents vibration wavelet packet energy proportion.

7. The rotating electro-mechanical device acoustic and vibration monitoring method of claim 6, wherein, The convolutional neural network CNN model specifically includes a convolutional layer, a pooling layer, a fully connected layer, and an output layer, and is suitable for pattern recognition of non-stationary signals; The multi-modal collaborative enhancement network MMSAN adopts a double-branch CNN to process acoustic and vibration features respectively, ensuring modality specificity; For the acoustic branch, the acoustic feature vector is calculated, and then convolution operation is performed: ； wherein: represents the output feature map of the acoustic branch at the layer, containing the extracted local features for subsequent hierarchical transmission; represents the current convolutional layer index, identifying the depth position of the network; represents the convolution operation function, used to perform convolution calculation; represents the acoustic feature map output by the previous layer; is the convolution kernel weight, is the bias; BatchNorm represents batch normalization, to stabilize and accelerate the neural network training; For the vibration branch, the vibration feature vector is calculated: ； wherein, represents the output feature map of the vibration branch in the i-th layer, containing the extracted local features, for subsequent layer transmission; represents the output feature map of the vibration branch in the i-th layer, containing the extracted local features, for subsequent layer transmission; represents the vibration feature map output by the previous layer; is a convolution kernel weight, is a bias; The SAM dynamically fuses acoustic and vibration features to ensure that key modality information is amplified: The fused features are: ； The collaborative weight is calculated through global average pooling and a learnable weight, which is used to dynamically amplify key modality information: ； ； wherein, represents the collaborative weight of the acoustic modal, ranging from [0, 1], for quantifying the importance of the acoustic feature; represents the collaborative weight of the vibration modal, ranging from [0, 1], satisfying + = 1; is an activation function mapping the input to [0, 1] for generating the probabilistic weight; represents the learnable weight matrix of the acoustic branch; is a learnable weight matrix representing the vibration branch for the vibration modal; represents a global average pooling function for compressing the spatial dimension, calculating the average value of each channel of the feature map, and outputting a one-dimensional vector; The final fused features are: ； final fused features input classifier, which determines abnormalities.

8. The rotating electromechanical device acoustic and vibration monitoring method of claim 1, wherein, Also including: S5. Based on the detection result, an alarm signal is generated, and a decision suggestion is generated, and a real-time feedback mechanism is combined to realize closed-loop risk management.

9. A rotating electromechanical equipment acoustic and vibration monitoring system based on the rotating electromechanical equipment acoustic and vibration monitoring method according to any one of claims 1-8, characterized by, Including: A signal acquisition module is configured to acquire original acoustic and vibration signals of the rotating electromechanical equipment in real time. ; wherein is an acoustic signal, is a vibration signal, is a signal length. The noise reduction module is used to employ a dual-timescale noise estimation method by analyzing the original acoustic and vibration signals. The corresponding number Frame, First Amplitude spectrum at frequency points Calculate the short-time noise amplitude based on the statistical distribution along the time axis. and long-term noise amplitude ,based on and To estimate the noise amplitude And then according to To reconstruct the original acoustic and vibration signals, and obtain the denoised acoustic and vibration signals. , ,in For signal length, the denoised acoustic and vibration signals Including the noise-reduced acoustic signal and vibration signals ; A feature extraction module for energy feature extraction of the denoised acoustic and vibration signals of the rotating electromechanical equipment based on wavelet packet decomposition; A detection module for inputting the energy features into the trained multi-modal collaborative enhancement network MMSAN for working condition anomaly detection, wherein the multi-modal collaborative enhancement network MMSAN is trained by the denoised acoustic and vibration signals.

10. A rotating electro-mechanical device acoustic and vibration monitoring system according to claim 9, characterised in that, In the noise reduction module, the original acoustic and vibration signals are analyzed The corresponding amplitude spectrum The statistical distribution along the time axis, respectively calculating the short-time noise amplitude And long-time noise amplitude The specific content includes: A dual-timescale noise estimation method is employed, and the amplitude spectrum is analyzed. Calculate the short-time noise amplitude based on the statistical distribution along the time axis. and long-term noise amplitude : ； ； wherein and denote the first decile and the first decile, respectively; and are decile thresholds for short-time and long-time noise estimates, respectively; reflects the local background level of noise, captures the overall trend and possible abrupt changes in noise.

11. A rotating electro-mechanical device acoustic and vibration monitoring system according to claim 10, wherein, In the noise reduction module based on and To estimate the noise amplitude The specific content includes: Introducing dynamic weights to balance the contribution of both short and long estimates: ； ； wherein is an initial weight, is a time frame index, , is a frequency index, , , is a number of points of the time-frequency transform, is an imaginary unit, and is upper and lower bounds of a range, satisfying 0≤ <1 ≤1; the base noise amplitude is: ； Based on the base noise amplitude Computing the signal noise amplitude ratio For quantifying the relative strength of signal and noise: ； Based on signal to noise amplitude ratio Computing a minimum amplitude threshold : ； wherein is a scaling factor that controls the range of the minimum amplitude threshold, ensuring that the signal's key features are not attenuated; Introducing dynamic noise scaling factor to dynamically adjust the noise amplitude: ； wherein, is a scaling factor, controlling the sensitivity of the adjustment factor, is an initial weight; Dynamic noise scaling factor Computing an estimated noise amplitude : 。 12. A computer readable storage medium having stored thereon a computer program, characterized in that, The computer program is executed by the processor to realize the rotating electromechanical equipment acoustic and vibration monitoring method of any one of claims 1-8.

13. An electronic device comprising: The processor and the memory, the memory is used to store one or more programs; characterized in that, when the one or more programs are executed by the processor, the rotating electromechanical equipment acoustic and vibration monitoring method of any one of claims 1-8 is realized.

Images