A dynamic gradient-guided cross-domain collaborative anti-noise acoustic diagnosis method

Abstract

The application provides a dynamic gradient guide-based cross-domain cooperative anti-noise acoustic diagnosis method, and belongs to the field of mechanical fault diagnosis. The method comprises collecting acoustic data and performing pretreatment; constructing a first counter-attack network mechanism for filtering time-varying non-static noise; constructing a second counter-attack network mechanism based on feature attention mask and local fine-grained self-adaptation; placing the first counter-attack network mechanism and the second counter-attack network mechanism in the same end-to-end cross-domain anti-noise diagnosis framework, cooperatively optimizing the cross-domain anti-noise diagnosis framework under the guidance of the feature attention mask based on cosine similarity, and outputting fault prediction and identification results by using the optimized cross-domain anti-noise diagnosis framework. In the cross-domain scene with extremely low signal-to-noise ratio, a new domain migration paradigm capable of intelligently identifying and ignoring non-static noise and accurately aligning core fault features is constructed, thereby completely avoiding negative migration and realizing stable anti-noise diagnosis.

Description

Technical Field

[0001] This invention belongs to the field of mechanical fault diagnosis, and in particular relates to a cross-domain collaborative noise immunity diagnostic method based on dynamic gradient guidance. Background Technology

[0002] Rotating machinery, as a core component of modern industrial systems such as aircraft engines, high-speed trains, wind turbine generators, and various production line drive units, directly impacts production safety, economic efficiency, and even the smooth operation of national infrastructure through its operational stability and reliability. Therefore, developing efficient and precise fault diagnosis technologies for rotating machinery, enabling effective early warning and identification of minor faults in critical components such as bearings and gears, has irreplaceable strategic value for ensuring the safety of major equipment, implementing predictive maintenance strategies, and reducing unplanned downtime losses.

[0003] Acoustic signal-based diagnostic methods, with their unique non-contact measurement advantages, have shown broad application prospects among numerous fault diagnosis technologies. Compared to typical contact measurement methods such as vibration diagnosis, the acquisition of radiated sound waves in the air domain by placing microphone arrays or single microphones near the equipment requires no physical contact with the equipment. This "non-invasive" characteristic gives them unique advantages such as flexible deployment, uninterrupted equipment operation, and the ability to achieve long-distance and multi-target synchronous monitoring, making them particularly suitable for health status assessment of large, complex, or harsh mechanical systems. However, acoustic diagnostic methods also face a severe technical challenge in the air propagation path that gives them their unique advantages—the problem of strong interference from complex and variable non-static background noise. In real industrial environments, the fault sound signals generated by the target equipment are often extremely weak and submerged in a strong background noise field coupled by multiple noise sources such as other nearby equipment, ambient wind fields, and personnel activities. Unlike the relatively steady-state and predictable noise in laboratory environments, background noise in industrial sites has significant non-static and time-varying characteristics, with its spectral components, energy intensity, and statistical distribution fluctuating drastically and irregularly over time. The presence of this non-static background noise can, on the one hand, directly "pollute" or "overwhelm" the characteristic frequency components that indicate early faults (such as harmonics of bearing inner and outer ring fault frequencies); on the other hand, its strong time-varying characteristics can seriously interfere with the effectiveness of traditional signal processing methods such as traditional fixed thresholds or static filtering, resulting in extremely low input signal-to-noise ratios for diagnostic models, thereby triggering a large number of false alarms and missed alarms, which greatly weakens the reliability and practicality of acoustic diagnostic systems.

[0004] Based on the foregoing analysis, the core technical problem currently facing the field of noise immunity diagnostics lies in the fact that when faced with non-static background noise with extremely significant randomness and time-varying characteristics, a huge distribution gap with no physical laws emerges between the noisy target domain and the clean source domain. Under this extreme distribution shift, existing domain migration methods, unable to effectively remove noise components, are prone to "negative migration" when performing blind global alignment, thus causing the diagnostic model to fail.

[0005] Therefore, in order to fully realize the enormous potential of non-contact measurement in acoustic diagnostics and promote its large-scale application in real industrial scenarios, researching and developing noise-resistant diagnostics that can effectively suppress non-static background noise interference and accurately extract fault features from signals with extremely low signal-to-noise ratios, even in the absence of prior information about the target noise environment, has become a key bottleneck that urgently needs to be overcome in the current field of acoustic diagnostics. Exploring advanced noise-resistant diagnostic technologies is not only the core of improving the performance of acoustic diagnostic methods, but also a necessary technical prerequisite for ensuring industrial intelligence and achieving the ambitious goal of predictive maintenance of equipment, possessing significant theoretical research significance and engineering application value. Summary of the Invention

[0006] To address the aforementioned shortcomings in existing technologies, this invention provides a cross-domain collaborative noise immunity diagnostic method based on dynamic gradient guidance. The macroscopic problem it aims to solve is: how to construct a novel domain migration paradigm in cross-domain scenarios with extremely low signal-to-noise ratios that can intelligently identify and ignore non-static noise and accurately align core fault features, thereby completely avoiding negative migration and achieving robust noise immunity diagnostics.

[0007] To achieve the above objectives, the technical solution adopted by this invention is as follows: a cross-domain collaborative noise immunity diagnostic method based on dynamic gradient guidance, comprising the following steps: S1. For the radiated sound field generated by the operation of mechanical equipment in real industrial sites, acoustic data is collected from all directions using a spatial array, and the acoustic data is preprocessed to obtain a two-dimensional time-frequency characteristic spectrum. S2. Construct a first-level adversarial network mechanism for filtering out time-varying non-static noise, wherein the first-level adversarial network mechanism is used to perform front-end noise reduction and reconstruction on the two-dimensional time-frequency feature map; S3. Construct a second adversarial network mechanism based on feature attention mask and local fine-grained adaptation. The second adversarial network mechanism is used to perform cross-domain feature alignment based on the front-end noise-resistant reconstruction results under the action of dynamic gradient reversal mechanism. S4. The first and second adversarial network mechanisms are placed in the same end-to-end cross-domain noise immunity diagnosis framework. The cross-domain noise immunity diagnosis framework is collaboratively optimized under the guidance of feature attention mask based on cosine similarity. The optimized cross-domain noise immunity diagnosis framework is used to output fault prediction and identification results, so as to realize cross-domain collaborative noise immunity diagnosis under real industrial non-static background noise interference.

[0008] Further, S1 includes the following steps: Construct a multi-channel spherical microphone array; For the radiated sound field generated by the operation of mechanical equipment in real industrial sites, a multi-channel spherical microphone array is used to collect acoustic data from all directions. During the acquisition process, the physical sound pressure of the multi-microphone channels is synchronously converted from analog to digital at a high-frequency sampling rate, and the original one-dimensional time-domain acoustic signal is saved as a lossless WAV audio format. Based on WAV lossless audio signals, continuous time-domain acoustic signals are sliced ​​into independent acoustic data analysis samples by performing equal-length slicing. The acoustic data analysis sample signals are subjected to frame segmentation and smoothing truncation processing. By applying the power spectrum of each frame of acoustic signal after framing and smoothing, a one-dimensional waveform is converted into a two-dimensional Mel-time spectrum feature map by using a Mel-frequency standard filter bank. The amplitude extrema of the two-dimensional Mel-time spectral feature map are normalized to output a standardized two-dimensional time-frequency feature map, thus completing the preprocessing.

[0009] Furthermore, the first layer of adversarial network mechanism includes a backbone generation network with a symmetric encoding-decoding architecture and a convolutional discriminant network connected to the backbone generation network, wherein the convolutional discriminant network constitutes the first layer of adversarial constraints; The backbone generation network is used to reconstruct the two-dimensional time-frequency feature map using a three-layer downsampling-three-layer upsampling topology to obtain the reconstructed spectrum, wherein the reconstructed spectrum is a target domain noise-reduced spectrum sample. The convolutional network is used to perform feature compression processing on the reconstructed spectrum and the clean spectrum samples of real industrial machinery in a noise-free environment using a three-layer fully convolutional network. Based on the compressed features, a global average pooling layer is used to flatten the multi-dimensional spatial information and output a scalar probability value that represents the authenticity / purity of the input sample, thus completing the front-end noise-resistant reconstruction. The title probability value is used to optimize the backbone generation network.

[0010] Furthermore, the second layer of adversarial network mechanism includes: The feature mask generation module is used to reconstruct the target domain denoised spectrum samples in the frequency domain by calculating the first-level adversarial network mechanism. With the corresponding clean spectrum samples of the source domain Based on the cosine similarity of spatial locations, weighted clean spectrum samples are generated. and weighted noise reduction spectrum samples ; A multi-scale feature extraction module is used for extracting features based on weighted clean spectral samples. and weighted noise reduction spectrum samples The parallel convolutional branches output tensor features, and the tensor features of each convolutional branch are stacked and spliced ​​to form a fused high-dimensional feature. The fused high-dimensional feature is then compressed and nonlinearly combined to output a common basic feature matrix, which includes weighted clean spectral features and weighted denoised spectral features. The local fine-grained domain discrimination module is used to output a spatial domain discrimination matrix based on the common basic feature matrix and the introduced local cross-domain adversarial architecture. The spatial domain discrimination matrix is ​​then fed back into the optimization of the common basic matrix using gradient inversion. The local cross-domain adversarial architecture is equipped with a gradient inversion layer at the front end. Each position in the spatial domain discrimination matrix represents the discrimination probability result for each local difference in the common basic feature matrix. The independent fault classification module is used to output a probability prediction matrix based on the weighted clean spectral features output by the multi-scale feature extraction module, and based on the probability prediction matrix, select the maximum extreme value term in the probability vector and output the fault type identification result of the mechanical equipment corresponding to the term.

[0011] Furthermore, the generation of weighted clean spectrum samples and weighted noise reduction spectrum samples ,include: In the frequency domain, the target domain denoised spectrum samples generated by the first-level adversarial network mechanism are calculated. With the corresponding clean spectrum samples of the source domain Cosine similarity in spatial location; Based on the calculated cosine similarity, a feature attention mask matrix is ​​generated through a nonlinear activation function mapping. ; Using feature attention mask matrix The clean spectral samples from the source domain and the denoised spectral samples from the target domain entering the second layer of the adversarial network mechanism are subjected to an element-wise Hadamard product weighting operation to generate weighted clean samples. and weighted noise reduction spectrum samples ,in, This represents the Hadama product.

[0012] Furthermore, the output spatial domain discrimination matrix includes: The common basic feature matrix is ​​continuously passed through multiple concatenated convolutional layers with a kernel dimension of 4×4 to compress the spatial receptive field; After spatial receptive field compression, channel-scale compression is performed by single-channel convolution with a 1×1 kernel to output a multi-dimensional spatial domain discrimination matrix. In this matrix, a single element represents the domain prediction attribute of the local time-frequency sensing region on the original two-dimensional time spectrum.

[0013] Furthermore, S4 includes: The first and second adversarial network mechanisms are placed within the same end-to-end cross-domain noise reduction diagnostic framework. Set the feature attention mask based on cosine similarity as the physical hub for backpropagation; A joint loss function is constructed, which is used to align the diagnosis of the second layer of adversarial network mechanism with cross-domain features in the form of dynamic gradients to penetrate the physical hub, reversely reshape and dominate the update trajectory of the first layer of adversarial network mechanism's regression reconstruction target, so as to perform collaborative optimization of the cross-domain noise resistance diagnosis framework. The optimized cross-domain noise immunity diagnostic framework is used to output fault prediction and identification results, so as to realize cross-domain collaborative noise immunity diagnosis under real industrial non-static background noise interference.

[0014] Furthermore, the expression for the joint loss function is as follows:

[0015]

[0016]

[0017]

[0018]

[0019]

[0020]

[0021]

[0022]

[0023]

[0024]

[0025]

[0026] in, Denotes the joint loss function. This represents the supervised mean squared error regression loss in the first layer of adversarial network mechanism. This represents the least-squares generative adversarial loss in the first-level adversarial network mechanism. This represents the local fine-grained adversarial loss of the second layer of adversarial network mechanism. This indicates the source domain purification classification diagnostic loss of the second layer of adversarial network mechanism. This represents the parameters of the backbone generation network in the first layer of adversarial network mechanism. This represents the backbone generation network in the first layer of adversarial network mechanism. This represents a noisy spectrum sample. This represents the convolutional discriminant network in the first layer of adversarial network mechanism. This represents the z-th target domain noise reduction spectrum sample. This represents the z-th clean spectrum sample from the source domain. This represents the weighted denoised spectrum sample. This represents the target domain denoised spectrum sample. This represents a weighted clean spectrum sample. Represents the feature attention mask matrix. This represents a clean spectrum sample in the source domain. Represents the Hadama product. Represents a classifier. This indicates a multi-scale parallel feature extraction module. This indicates the number of data samples in the training set. z Indicates the first z A dataset sample, This represents the expectation of the noisy spectral sample. x Represents a random variable. This represents the expectation of a clean spectral sample in the source domain. This represents the output of the backbone network. Indicates the weighted clean spectral characteristics Expectations and These represent the number of frequency bands and the number of frames in the original two-dimensional time spectrum, respectively. This represents the output of the local fine-grained domain discrimination module. This represents a local fine-grained domain discrimination module. Indicates the weighted noise reduction spectral characteristics Expectations This represents the true fault category in the clean sample. This indicates the predicted fault category output by the independent fault classification module.

[0027] Compared to existing acoustic diagnostics and conventional domain adaptive techniques for mechanical equipment, the cross-domain collaborative noise immunity diagnostic method proposed in this invention, through its innovative network architecture and underlying derivative optimization mechanism design, brings the following significant technical effects: 1. Construct a dual-adversarial integrated architecture to avoid the "negative migration" effect caused by adversarial adaptation: Faced with highly random and time-varying nonstatic background noise in industrial environments, traditional domain adaptive forced full-band alignment often leads to severe model failure. This invention innovatively constructs a dual-domain adversarial system architecture. The first layer of adversarial network mechanism dynamically tracks and suppresses nonstatic random time-varying fluctuations without losing high-frequency dimensional information, completing initial noise-resistant reconstruction. Subsequently, a second layer of adversarial network mechanism is tightly integrated, generating a feature attention mask based on cosine similarity to adaptively amplify structured fault bands and block residual noise bands. Combined with fine-grained domain adversarial mechanisms at the local block level, this powerful synergy enables the model to accurately mine deep domain-invariant features only in core fault regions, fundamentally eliminating the interference of irrelevant time-varying noise on cross-domain alignment and completely avoiding the "negative transfer" trap caused by blind alignment.

[0028] 2. Pioneering a dynamic gradient guidance mechanism to achieve deep synergy between "noise resistance" and "diagnosis": Traditional serial acoustic diagnostic or preprocessing methods typically aim to minimize the global signal mean square error, which can easily "falsely kill" weak high-frequency fault features during the noise reduction process. This invention innovatively uses a "feature attention mask" as the physical hub for forward and backward propagation, establishing a dynamic gradient channel where the backend (second-layer adversarial network mechanism) diagnostic task provides targeted feedback to the frontend noise reduction module. This mechanism prevents the noise reduction network from blindly pursuing global waveform smoothness; instead, it is "indirectly guided" by the backend classification loss error, forcing the frontend model to prioritize preserving the core diagnostic frequency band for optimization updates. This achieves deep coupling between noise-resistant feature extraction and fault diagnosis, ensuring that "noise-resistant reconstruction serves to improve the final diagnostic accuracy."

[0029] 3. High practicality of project deployment: The ingenuity of this architecture lies in the fact that the classification decision boundary, trained solely on pure, noise-free laboratory data, can be directly and seamlessly applied to heavily noisy industrial environments. This significantly reduces the data labeling cost of collecting labeled noise samples on-site and promotes the generalization of condition monitoring to extreme operating conditions. Attached Figure Description

[0030] Figure 1 This is a flowchart of the method of the present invention.

[0031] Figure 2 This is a schematic diagram of the visualization analysis of cross-domain alignment ablation.

[0032] Figure 3 This is a diagram illustrating the visualization and comparison of clustering results. Detailed Implementation

[0033] The specific embodiments of the present invention are described below to enable those skilled in the art to understand the present invention. However, it should be understood that the present invention is not limited to the scope of the specific embodiments. For those skilled in the art, various changes are obvious as long as they are within the spirit and scope of the present invention as defined and determined by the appended claims. All inventions utilizing the concept of the present invention are protected.

[0034] Example In order to address the macro-level problem, this invention, in its specific implementation scheme, focuses on solving the following two progressively key technical challenges: I. Construct a cross-domain adversarial mechanism that can effectively separate and suppress non-static noise while extracting domain-invariant diagnostic features. This mechanism is implemented through a dual adversarial network mechanism.

[0035] The first layer of adversarial training is task-oriented, and its core lies in using adversarial game theory to assist the mean squared error (MSE) loss function to reconstruct a predicted spectrum from the input noise spectrum that is highly similar in distribution to the corresponding clean spectrum. The key to this stage is to establish a generative adversarial noise reduction front-end that can effectively suppress or filter non-static noise components, providing a "cleaned" data foundation for subsequent accurate diagnosis.

[0036] The second layer of adversarial processing focuses on the transfer of diagnostic knowledge. Its core is to construct a domain-adaptive mechanism to accurately extract the "domain-invariant features" (weighted clean spectrum samples) shared by the denoised spectrum after the first layer of adversarial processing and the original clean spectrum, which characterize the device's health status. and weighted noise reduction spectrum samples The input multi-scale feature extraction module, through its collaboration with the local fine-grained domain discriminant network to construct a second adversarial network mechanism, learns the ability to extract "domain-invariant features." These domain-invariant features are the inherent, invariant features shared by the clean and denoised spectrum samples. The key at this stage is that adversarial training forces the feature extraction network to ignore unavoidable residual noise or artifacts in the denoised spectrum, focusing only on the structured fault information common to the clean spectrum. This ensures that diagnostic knowledge learned in the clean environment can be effectively transferred to the noisy environment, fundamentally avoiding negative transfer caused by random, time-varying noise.

[0037] II. How to establish a collaborative optimization bridge between the above-mentioned dual-domain adversarial mechanisms to avoid falling into the old dilemma of separating "noise reduction" and "diagnosis" tasks.

[0038] Simply stacking two adversarial network mechanisms degenerates into a separate "denoising first, then diagnosis" architecture, failing to achieve true end-to-end optimization. Therefore, the key challenge this invention addresses is establishing a dynamic gradient propagation mechanism to connect the regression task in the first adversarial layer with the classification task in the second adversarial layer. Specifically, an innovative information flow path needs to be designed so that the performance of the backend (second adversarial layer) diagnostic classification task can, in the form of gradient signals, dynamically and inversely "guide" the parameter updates of the frontend (first adversarial layer) denoising module. By establishing this mutually responsive and collaborative optimization relationship between the frontend and backend, it ensures that the frontend denoising behavior is not only about reducing global noise but also about generating feature representations most beneficial for accurate backend diagnosis. This achieves deep coupling and integrated collaboration between the two major goals of denoising and diagnosis, resulting in comprehensive system-level performance gains.

[0039] like Figure 1 As shown, this invention provides a cross-domain collaborative noise immunity diagnostic method based on dynamic gradient guidance, the implementation method of which is as follows: S1. For the radiated sound field generated by the operation of mechanical equipment in a real industrial site, acoustic data is collected from all directions using a spatial array, and the acoustic data is preprocessed to obtain a two-dimensional time-frequency characteristic spectrum. The implementation method is as follows: Construct a multi-channel spherical microphone array; For the radiated sound field generated by the operation of mechanical equipment in real industrial sites, a multi-channel spherical microphone array is used to collect acoustic data from all directions. During the acquisition process, the physical sound pressure of the multi-microphone channels is synchronously converted from analog to digital at a high-frequency sampling rate, and the original one-dimensional time-domain acoustic signal is saved as a lossless WAV audio format. Based on WAV lossless audio signals, continuous time-domain acoustic signals are sliced ​​into independent acoustic data analysis samples by performing equal-length slicing. The acoustic data analysis sample signals are subjected to frame segmentation and smoothing truncation processing. By applying the power spectrum of each frame of acoustic signal after framing and smoothing, a one-dimensional waveform is converted into a two-dimensional Mel-time spectrum feature map by using a Mel-frequency standard filter bank. The amplitude extrema of the two-dimensional Mel-time spectral feature map are normalized to output a standardized two-dimensional time-frequency feature map, thus completing the preprocessing.

[0040] In this embodiment, a data processing pipeline is constructed from high-fidelity sound field capture to standardized two-dimensional feature input, specifically including two sub-stages: acoustic data acquisition and preprocessing.

[0041] 1. Acoustic data acquisition based on multi-channel spherical array To address the radiated sound field generated by the operation of machinery in real industrial settings, this invention employs spatial array technology for omnidirectional data capture. Specifically, a four-channel spherical microphone array is constructed, with four identical measurement-grade microphones rigidly fixed to the surface of a custom-designed spherical array support in a spatially symmetrical geometric configuration. This array topology is designed to closely match the physical characteristics of mechanical fault acoustic radiation, which typically manifests as an approximately point source spreading outwards as a spherical wave. This effectively avoids spatial aliasing effects in single-direction measurements, achieving complete and seamless capture of the target's three-dimensional sound field physical information. During the signal acquisition phase, a multi-channel data acquisition front-end performs rigorous synchronous analog-to-digital conversion (A / D acquisition) on the physical sound pressure signals of the four microphone channels at a high-frequency sampling rate of 16kHz. The acquired raw one-dimensional time-domain acoustic signals are then stored in WAV lossless audio format as the data source for subsequent processing.

[0042] 2. Joint time-frequency domain transformation and data sample standardization preprocessing To transform the original one-dimensional audio sequence into an input format suitable for high-order feature mapping in deep adversarial neural networks, the following preprocessing steps are performed on the acquired WAV audio signal: 1) Sample Slicing and Frame Windowing: First, the continuous time-domain acoustic signal is sliced ​​into equal-length slices, with each 1-second time series strictly divided as an independent acoustic data analysis sample. Then, the Short-Time Fourier Transform (STFT) concept is introduced, employing a Hamming window with a window length of 64ms and setting a 50% frame overlap rate to frame and smoothly truncate each sample signal. This effectively suppresses spectral leakage while preserving the time-domain evolution characteristics.

[0043] 2) Mel-Spectrogram Feature Mapping: A Mel-band filter bank containing 40 triangular filters is applied to the power spectrum of each frame of the signal. By simulating the nonlinear perception characteristics of the human auditory system, the frequency axis is compressed logarithmically, and finally the one-dimensional waveform is converted into a two-dimensional Mel-spectrum feature map with a dimension of 40×32 (number of frequency bands × number of time frames).

[0044] 3) Amplitude Extremum Normalization: To accelerate the convergence speed of the subsequent deep adversarial network mechanism and eliminate the absolute energy differences caused by different acquisition conditions, an extremum normalization strategy is applied to the extracted Mel two-dimensional time spectrum. Through a linear mapping operator, the feature amplitudes in the spectrum matrix are uniformly scaled and limited to the standard closed interval [-1, 1]. This completes the feature engineering extraction of the original acoustic signal, outputting a standardized two-dimensional acoustic sample set, which can be directly used as the input carrier for the source domain (clean) and target domain (noisy) in the downstream dual-domain adversarial model.

[0045] S2. Construct a first-level adversarial network mechanism for filtering out time-varying non-static noise. The first-level adversarial network mechanism is used to perform front-end noise reduction reconstruction on the two-dimensional time-frequency feature map. The first-level adversarial network mechanism includes a backbone generation network with a symmetric encoding-decoding architecture and a convolutional discriminant network connected to the backbone generation network. The convolutional discriminant network constitutes the first-level adversarial constraint. The backbone generation network is used to reconstruct the two-dimensional time-frequency feature map using a three-layer downsampling-three-layer upsampling topology to obtain the reconstructed spectrum, which includes denoised spectrum samples. In this embodiment, the reconstructed spectrum is the denoised spectrum sample of the target domain, while the clean spectrum sample only serves as a supervisory tool for the real target. In other words, the denoised spectrum sample, as the output of the backbone generation network, needs to infinitely approximate the clean spectrum sample.

[0046] The convolutional discriminant network is used to perform feature compression processing on reconstructed spectrum and clean spectrum samples of real industrial machinery in a noise-free environment using a three-layer fully convolutional network. Based on the compressed features, a global average pooling layer is used to flatten the multi-dimensional spatial information and output a scalar probability value representing the authenticity / purity of the input sample, thus completing the front-end noise-resistant reconstruction. The title probability value is used to optimize the backbone generation network.

[0047] In this embodiment, the output scalar probability value is used to reflect the adversarial process. Specifically, the goal of the convolutional discriminant network is to clearly distinguish between real samples and reconstructed samples, that is, to output a high probability value (e.g., 0.9) for real samples and a low probability value (e.g., 0.1) for reconstructed samples; while the role of the backbone generator network is to reconstruct realistic samples so that the convolutional discriminant network cannot distinguish between real and fake (fake represents reconstructed samples), thus forming an adversarial relationship between the backbone generator network (generator) and the convolutional discriminant network (discriminator), with the aim of prompting the backbone generator network to reconstruct an extremely realistic reconstructed spectrum.

[0048] In this embodiment, the first layer of adversarial network (ADN) mechanism, serving as the front-end preprocessing hub of the entire noise reduction diagnostic system, aims to dynamically track and remove highly random and time-varying non-static noise components from the spectrum through a deep generative adversarial architecture. This first layer of ADN focuses on providing a highly pure "cleaned" spectrum for the subsequent second layer of ADN (cross-domain feature alignment), thereby cutting off the possibility of non-static noise causing a "negative migration" effect at its source. This mechanism is composed of a backbone generative network (generator) with a symmetrical encoder-decoder architecture and a convolutional discriminant network working together.

[0049] 1. Backbone Reconstruction Network (Generator) Based on Self-Attention Mechanism The backbone generation network accepts a 40×32 two-dimensional time-frequency feature map with severe non-static noise output by S1 as input. It adopts a deep topology structure of "three-layer downsampling (encoding) - three-layer upsampling (decoding)" to achieve accurate reconstruction from noisy spectrum to high-dimensional latent space and then to clean spectrum.

[0050] 1) Down-sampling stage: The network sequentially performs three cascaded down-sampling operations, progressively compressing the original noisy 40×32 input into a 5×4 high-order semantic abstraction feature map. Each down-sampling module contains three core functional units: (a) Patch Partition: A dimension-based adaptive block partitioning strategy is employed. The input time-frequency feature map is divided into equal-area segments based on the least common multiple (or common divisor) of its time and frequency dimensions. For example, for a 40×32 input in the first layer, it is divided into several 10×8 local image patches using a 4×4 grid, with its greatest common divisor of 4 as the benchmark. Similarly, if the third layer input is 10×8, it is divided into 5×4 blocks using a 2×2 grid. This operation structurally restructures the original (frequency × time) two-dimensional matrix into a three-dimensional tensor (number of blocks × frequency × time), establishing local semantic fragments.

[0051] (b) Multi-head Self-Attention and Information Fusion: Each image patch is assigned a specific attention head attribute, and linear projection and global context awareness are performed through a multi-head self-attention mechanism. Subsequently, stacking and merging operations are performed along the channel dimensions of each head to losslessly restore the 3D tensor back to a 2D matrix format (frequency × time). This process aims to perform deep linear encoding of time-frequency features and global interaction without incurring any spatial dimensional decay.

[0052] (c) Patch Merging: After completing the attention interaction, feature pixels are extracted and fused at fixed intervals (span step size of 2) along the time and frequency dimensions, so as to strictly achieve a 1 / 2 reduction (i.e. downsampling) of the time and frequency dimensions of the current layer input in terms of physical size.

[0053] 2) Up-sampling stage: Receives 5×4 high-dimensional abstract features and restores them to the original 40×32 acoustic time-frequency scale through three cascaded up-sampling modules. Each up-sampling module also contains three core functional units: (a) Patch Expanding: First, the channel dimension of the input feature map is significantly expanded to 4 times the original through convolution or linear mapping; then, a spatial rearrangement mechanism is executed to distribute the newly added channel dimension evenly to the physical dimensions of time and frequency, thereby achieving a physical expansion of spatial resolution by 2 times.

[0054] (b) Skip Connection: To prevent the loss of high-frequency fault details during upsampling reconstruction, long-distance skip connection links are introduced. The downsampled output features of the corresponding layer at the encoding end (e.g., the downsampled features of the third layer correspond to the upsampled stage of the first layer, and so on) are concatenated and fused with the currently patched and expanded features along the channel dimension, which greatly enhances the time-frequency fidelity of the reconstructed spectrum.

[0055] (c) Feature Refinement: The concatenated features are sequentially processed through "patch partitioning" and "multi-head self-attention mechanism" consistent with the encoding logic, completing the feature smoothing and refinement reconstruction from a two-dimensional matrix to three-dimensional blocks and then fused back into a two-dimensional matrix. After three iterations, the final output is a 40×32 reconstructed spectrum after "noise filtering".

[0056] 2. Convolutional Discriminant Networks and Adversarial Collaborative Optimization To ensure that the backbone generation network not only performs traditional smoothing and denoising, but also strictly approximates the spectrum distribution of the real physical health state, a discriminant network is introduced to form the first layer of adversarial constraints.

[0057] Convolutional Discriminant Structure: The convolutional discriminant network simultaneously receives reconstructed spectral samples from the backbone generator network and clean spectral samples from a real device in a noise-free environment. A three-layer fully convolutional network architecture is employed, with each layer having a receptive field (convolutional kernel) of 4×4 and a backward receptive stride of 2×2. A Leaky ReLU nonlinear activation function is used to progressively compress the input 40×32 spectral features into a 5×4 microscopic receptive field. Finally, a global average pooling layer flattens the multidimensional spatial information, outputting a scalar probability value representing the "real / clean" degree of the input sample.

[0058] Optimization Objective: This invention's first layer of adversarial processing does not employ conventional cross-entropy adversarial loss, but instead uses Least Squares GAN Loss (LSGAN). Within the least squares squared loss metric framework, the discriminator strives to accurately assess the absolute error between the generated denoised spectral samples and the real clean spectral samples in a high-dimensional distribution; and continuously guides the backbone generation network to update its weights through the gradient of this numerical error. This least squares-based local adversarial penalty forces the backbone generation network to precisely focus on and filter out "time-varying non-static noise" that deviates from the statistical distribution of the clean spectrum, achieving dynamic tracking and effective suppression of unpredictable noise statistical characteristics. This lays a high-quality data foundation for the subsequent construction of a second layer of noise-resistant alignment network with a cross-domain collaborative guidance mechanism.

[0059] S3. Construct a second adversarial network mechanism based on feature attention masks and local fine-grained adaptation. This second adversarial network mechanism is used to perform cross-domain feature alignment based on the front-end noise-resistant reconstruction results under the action of a dynamic gradient inversion mechanism. The second adversarial network mechanism includes: The feature mask generation module is used to reconstruct the target domain denoised spectrum samples in the frequency domain by calculating the first-level adversarial network mechanism. With the corresponding clean spectrum samples of the source domain Based on the cosine similarity of spatial locations, weighted clean spectrum samples are generated. and weighted noise reduction spectrum samples Specifically, it is as follows: In the frequency domain, the target domain denoised spectrum samples generated by the first-level adversarial network mechanism are calculated. With the corresponding clean spectrum samples of the source domain Cosine similarity in spatial location; based on the calculated cosine similarity, a feature attention mask matrix is ​​generated through a nonlinear activation function mapping. ; Utilizing the feature attention mask matrix The clean spectral samples from the source domain and the denoised spectral samples from the target domain entering the second layer of the adversarial network mechanism are subjected to an element-wise Hadamard product weighting operation to generate weighted clean samples. and weighted noise reduction spectrum samples ,in, This represents the Hadama product.

[0060] A multi-scale feature extraction module is used for extracting features based on weighted clean spectral samples. and weighted noise reduction spectrum samples The parallel convolutional branches output tensor features, and the tensor features of each convolutional branch are stacked and spliced ​​to form a fused high-dimensional feature. The fused high-dimensional feature is then compressed and nonlinearly combined to output a common basic feature matrix, which includes weighted clean spectral features and weighted denoised spectral features. In this embodiment, the multi-scale feature extraction module has two inputs: weighted clean spectral samples. With weighted noise reduction spectrum samples The common fundamental feature matrix actually contains samples derived from weighted clean spectrum samples. Weighted denoising (reconstruction) samples The system consists of two parts. In the local fine-grained domain discriminant module, both outputs from the multi-scale feature extraction module—weighted clean spectral features and weighted denoised spectral features—are simultaneously received. In the independent fault classification module, only the weighted clean sample features are received. The ultimate goal is to make these two output features identical through adversarial learning between the multi-scale feature extraction module and the local fine-grained domain discriminant network; this is called "domain-invariant features." The local fine-grained domain discriminant module outputs a spatial domain discriminant matrix based on a common fundamental feature matrix, utilizing an introduced local cross-domain adversarial architecture. The local cross-domain adversarial architecture is fronted by a gradient inversion layer, and the method is as follows: The common basic feature matrix is ​​continuously passed through multiple cascaded convolutional layers with a kernel dimension of 4×4 to compress the spatial receptive field. After spatial receptive field compression, channel scale compression is performed through single-channel convolution with a kernel of 1×1 to output a multi-dimensional spatial domain discrimination matrix. In this matrix, a single element represents the domain prediction attribute of the local time-frequency sensing region on the original two-dimensional time spectrum.

[0061] In this embodiment, in the first level of adversarial communication, the convolutional discriminant network outputs a scalar probability value, representing a global judgment result on the entire reconstructed spectrum and the clean spectrum. In the second level of adversarial communication, the output is a probability matrix, i.e., a spatial domain discriminant matrix. Each position in this matrix represents the discrimination probability result for each local region in the common basic features. The clean spectrum samples and the reconstructed spectrum samples are respectively processed by the common basic feature matrix to obtain their respective feature matrices. These two feature matrices are fed into the local fine-grained domain discriminant module (which is also a discriminant network) to obtain not a unique global discrimination result, but a result matrix. The output probability value at each position in the matrix represents the discrimination of whether the corresponding local region originates from the clean spectrum or the reconstructed spectrum. Compared with the global discrimination in the first level of adversarial communication, this local discrimination narrows the discrimination area and target, which is beneficial for the reconstruction of local spectral features. Similarly, the summation of the output probabilities of each local discrimination by the air domain discrimination matrix ideally guides the common fundamental feature matrix to extract domain-invariant features (common features) from the clean spectrum and the reconstructed spectrum. This results in each position of the spatial domain discrimination matrix outputting a probability of 0.5, meaning that no local region of the common features can distinguish between a clean spectrum and a reconstructed spectrum, thus demonstrating the effectiveness of the common fundamental feature matrix in extracting domain-invariant features. The output spatial domain discrimination matrix is ​​fed back into the optimization of the common fundamental feature matrix. This feedback is achieved by a gradient inversion layer, creating an adversarial relationship between the common fundamental feature matrix and the local fine-grained domain discrimination module (discriminator network).

[0062] The independent fault classification module is used to output a probability prediction matrix based on the weighted clean spectral features output by the multi-scale feature extraction module, and based on the probability prediction matrix, select the maximum extreme value term in the probability vector and output the fault type identification result of the mechanical equipment corresponding to the term.

[0063] In this embodiment, in the second layer of adversarial network mechanism, the cross-domain alignment result is reflected in the multi-scale feature extraction module, i.e., the common basic feature matrix; the criterion for whether alignment is achieved is reflected in the output probability matrix of the spatial domain discriminant matrix. If each element in the output of the spatial domain discriminant matrix is ​​0.5, it means that the local fine-grained domain discriminant module cannot distinguish whether the common basic feature matrix originates from clean spectrum samples or reconstructed spectrum samples. This indicates that the multi-scale feature extraction module can extract the common features between clean spectrum samples and reconstructed spectrum samples, achieving the effect of cross-domain feature alignment. Therefore, the final presentation of cross-domain feature alignment is reflected in the output result of the multi-scale feature extraction module.

[0064] In this embodiment, the weighted clean spectral features output by the multi-scale feature extraction module are input into the independent fault classification module. The common basic feature matrix actually contains features derived from the weighted clean spectral samples. Weighted denoising (reconstruction) samples The system consists of two parts, with the weighted clean spectral features entering the independent fault classification module. Simultaneously, both the weighted clean spectral features and the weighted denoised spectral features are entered into the local fine-grained domain discrimination module for feature sample source determination. The fault classification module and the local fine-grained domain discrimination module execute concurrently.

[0065] In this embodiment, the second adversarial network mechanism, as the core backend of the noise immunity diagnostic framework model, aims to build a deep domain adaptation channel based on the first adversarial layer (noise reduction). Its fundamental purpose is to accurately extract the "deep domain-invariant features" that characterize the health status of the device, shared by the real clean source domain spectrum and the reconstructed target domain (noise reduction) spectrum. Through local adversarial training, the network is forced to actively ignore residual unpredictable noise and artifacts in the noise reduction spectrum, focusing only on structured common fault information, thereby ensuring that the diagnostic boundary learned in a noise-free environment can be losslessly transferred to the complex noise domain. Structurally, this stage consists of four collaborative parts: a feature mask generation module, a multi-scale feature extraction network, a local fine-grained domain discrimination network, and an independent fault classification network, and contains two key innovations: Innovation Point 1: Based on cosine similarity, this invention generates and adaptively weights a feature attention mask. It breaks through the limitations of traditional domain adaptive full-band blind alignment and innovatively constructs a feature mask generation module.

[0066] Masking calculation: The target domain denoised spectrum sample generated by the first layer of adversarial network is used for masking calculation. With the corresponding clean spectrum samples of the source domain A synchronous input feature mask generation module is used. In the frequency domain, the cosine similarity between the corresponding spatial locations is calculated, and then mapped using a non-linear activation function to generate a feature attention mask matrix. .

[0067] Adaptive weighting: using the generated mask The source and target domain spectral samples entering the backend are weighted element-wise by the Hadamard product to generate weighted clean spectral samples. With weighted noise reduction spectrum samples .

[0068] Mechanism of Action: This mask physically acts as a "dynamic filter." Regions with high similarity represent key fault bands related to the operating status information of rotating machinery; these are amplified and preserved with a mask value close to 1. Regions with low similarity represent useless residual non-static noise bands; these are masked with a mask value close to 0. This significantly reduces the blind zone for subsequent extraction of domain-invariant features and eliminates interference from irrelevant noise on feature alignment.

[0069] To fully capture high-frequency features and long-term structured contours caused by localized minor impacts, the multi-scale feature extraction module employs a parallel multi-branch mapping architecture. Three parallel convolutional branches are constructed, with kernel sizes of 7×7, 5×5, and 3×3, and a stride of 1×1 for each. After feature mapping, each branch is connected to a max pooling layer with a stride of 2 for spatial downsampling, transforming the output dimension of each single-level branch into a 20×16×16 tensor feature. The features extracted from the three independent receptive fields are stacked and concatenated along the channel dimension to form a fused high-dimensional feature matrix of size 20×16×48. Finally, a feature aggregation network with a mean convolution kernel of 5×5 is used to compress and nonlinearly combine redundant channel dimension information, outputting a common basic feature matrix of size 20×16×16. This common basic matrix includes weighted clean spectral features and weighted denoised spectral features.

[0070] Innovation Point Two: Asymmetric Cross-Domain Adversarial Flow Based on Local Time-Frequency Sensing Level After obtaining the common basic features, the second layer of adversarial network executes an asymmetric task flow allocation mechanism: the target domain weighted features are only input to the local fine-grained domain discrimination module, while the source domain weighted features are split into two paths and input to the local fine-grained domain discrimination module and the independent fault classification module, respectively.

[0071] Local Fine-Grained Domain Discriminant Module: Conventional domain adaptive networks often use global pooling to collapse features at the end of the network when performing cross-domain discrimination, ultimately outputting only a single scalar probability value for a complete input (approaching 1 indicates the source domain, approaching 0 indicates the target domain). This easily leads to coarse alignment and negative transfer in time-frequency details. This invention innovatively introduces a local cross-domain adversarial architecture that sinks to the "time-frequency block region". Specifically, the extracted 20×16×16 feature matrix is ​​continuously passed through three cascaded convolutional layers with a kernel dimension of 4×4 for spatial receptive field compression, and then through a single-channel convolution with a kernel of 1×1 for channel scale compression. The final output is not a single scalar, but a spatial domain discriminant matrix of size 5×4×1. The single element in this spatial domain discriminant matrix (e.g., a point in the matrix) This directly maps and represents the domain prediction attributes of a specific local time-frequency sensing region on the original two-dimensional time-frequency spectrum. A gradient reversal layer (GRL) is mounted at its front end. This architecture forcibly guides the objective function to achieve a fine-grained distribution game of reverse gradients at the local region level. This local adversarial nature forces the feature extraction network to strive to mine deep physical domain-invariant features within each tiny frequency band and time window. In this embodiment, a method of compressing global time-frequency information into local region labels is used through a concatenated convolutional network, so that each matrix element corresponds to a block region of the physical time-frequency spectrum. Based on this, a gradient reversal layer (GRL) is applied to achieve fine-grained intra-domain feature game and purification at the local block level.

[0072] The source-domain driven independent fault classifier module independently inputs the weighted clean spectral features output from the multi-scale feature extraction module into the fault classification branch. This independent fault classification module consists of two concatenated convolutional layers with identical physical parameters (4×4 kernel, 2×2 stride), followed by a global average pooling (GAP) layer to globalize the spatial information, and finally connected to a fully connected layer and a softmax activation function classifier. The output of this module is fixed as a 12-dimensional probability prediction matrix, which physically represents the predicted confidence probability of the fault mode of the current rotating machinery under 12 different typical operating conditions. By selecting the maximum extremum term (Argmax) in the probability vector, the final accurate fault type identification result of the current equipment under the operating condition is output. Since the classification boundary is entirely dominated by noise-free weighted clean samples (source domain), it ensures that the disease and health scale learned by the classifier has the highest purity and theoretical upper limit.

[0073] S4. The first and second adversarial network mechanisms are placed within the same end-to-end cross-domain noise immunity diagnostic framework. The framework is then collaboratively optimized under the guidance of feature attention masks based on cosine similarity. The optimized framework outputs fault prediction and identification results to achieve cross-domain collaborative noise immunity diagnostics under real-world industrial non-static background noise interference. The implementation method is as follows: The first and second adversarial network mechanisms are placed within the same end-to-end cross-domain noise reduction diagnostic framework. Set the feature attention mask based on cosine similarity as the physical hub for backpropagation; A joint loss function is constructed, which is used to align the diagnosis of the second layer of adversarial network mechanism with cross-domain features in the form of dynamic gradients to penetrate the physical hub, reversely reshape and dominate the update trajectory of the first layer of adversarial network mechanism's regression reconstruction target, so as to perform collaborative optimization of the cross-domain noise resistance diagnosis framework. The optimized cross-domain noise immunity diagnostic framework is used to output fault prediction and identification results, so as to realize cross-domain collaborative noise immunity diagnosis under real industrial non-static background noise interference.

[0074] In this embodiment, the dynamic gradient guidance and global collaborative optimization mechanism based on the feature mask hub is the core computational flow for achieving "deep collaboration between noise reduction and noise resistance". This invention innovatively places the dual adversarial mechanism within the same end-to-end synchronous optimization framework and sets the "feature attention mask" based on cosine similarity as the physical hub for backpropagation. By constructing a joint loss function covering four sub-tasks, the backend diagnostic and alignment requirements can penetrate the mask hub in the form of "dynamic gradients," reshaping and dominating the backbone generator network (generator) of the first-layer adversarial network mechanism in reverse. The update trajectory of the regression reconstruction goal.

[0075] In this embodiment, the overall synchronous optimization of the system is based on the comprehensive synergy between the first layer of adversarial processing (denoising regression) and the second layer of adversarial processing (cross-domain features and classification). A complete global objective loss function is defined. It comprises the following four core components: front-end supervised mean squared error regression loss. Suppressing large fluctuations in non-static background noise; front-end least-squares generative adversarial loss. Constraints to make denoised samples approximate clean samples on a high-dimensional statistical distribution; backend local fine-grained domain adversarial loss. : Driven by the mining of deep domain invariant features; back-end source domain purification, classification, and diagnostic loss. Perform fault classification for weighted clean samples.

[0076] The global objective function is defined as follows:

[0077]

[0078]

[0079]

[0080]

[0081] in, Denotes the joint loss function. This represents the supervised mean squared error regression loss in the first layer of adversarial network mechanism. This represents the least-squares generative adversarial loss in the first-level adversarial network mechanism. This represents the local fine-grained adversarial loss of the second layer of adversarial network mechanism. This indicates the source domain purification classification diagnostic loss of the second layer of adversarial network mechanism. This represents the number of data samples in the training set (the number of clean samples and reconstructed and denoised samples in the training set is the same, both being N). z Indicates the first z A dataset sample, This represents the expectation of the noisy spectral sample. x Represents a random variable. This represents the expectation of a clean spectral sample in the source domain (source domain sample). This represents the output of the backbone network. Indicates the weighted clean spectral characteristics Expectations and These represent the number of frequency bands and the number of frames in the original two-dimensional time spectrum, respectively. This represents the output of the local fine-grained domain discrimination module. This represents a local fine-grained domain discrimination module. Indicates the weighted noise reduction spectral characteristics Expectations This represents the true fault category in the clean sample. This indicates the predicted fault category output by the independent fault classification module.

[0082] In this embodiment, the mechanism of dynamic propagation and guidance of the back gradient, with the mask as the hub, is as follows: During the backpropagation of the synchronous network, the front-end denoising and reconstruction network - the first layer of adversarial network (generator) ) parameter update amount The gradient is determined by the sum of all partial derivatives penetrating to that node. According to the differential chain rule, the total gradient is... It consists of four components:

[0083] The specific chain rule and physical meaning of the forward propagation of various losses are as follows: Sub-items 1 and 2: Front-end baseline regression and direct propagation of adversarial gradients: These two losses directly affect the target domain denoising spectrum output of the first layer of adversarial processing. (in , Indicates the noisy spectrum. The updated parameters of the backbone generation network constitute the driving force for the front-end to filter out time-varying noise.

[0084] .

[0085] Sub-items three and four present two unconventional "dynamically guided" paths, completely reshaping the front-end noise reduction goals: Path 1 (Sub-item 3): Direct Constraint Guidance for Domain Adversarial Targets Physical meaning: Local fine-grained resistance to loss at the back end Gradient flowing through the multi-scale feature extraction module After that, it will directly encounter weighted noise-reduced spectrum samples. .because The error gradient is directly decoupled and propagated to the weighted noise reduction spectrum sample. This then flows into the backbone generator network of the first layer of adversarial network mechanism. This guidance forces the front-end noise reduction model to prioritize stripping out those frequency band features that are "most easily detected by the local domain discriminator as having residual non-static noise characteristics of the target domain" while denoising based on MSE.

[0086] Due to the local fine-grained resistance loss at the back end Simultaneously receive weighted source domain (clean) spectrum samples. Weighted target domain (denoising) spectrum samples Due to the mask It is also a weighted denoised spectrum sample. The function, whose gradient propagation exhibits two-path coupling:

[0087] After expanding the mask hub, we get:

[0088] Path Two (Sub-item Four): Indirect Guidance via Mask Hub for Classification Targets This invention innovatively proposes a collaborative mechanism. Although the asymmetric classifier only receives and evaluates source-domain weighted spectral samples... That is, source domain purification classification and diagnostic loss The forward computation involves no noisy samples, but its error gradient depends entirely on... This hub "remotely" penalizes the front-end backbone generation network. .

[0089] Its mathematical mechanism lies in: feature attention mask It is the first layer of adversarial network mechanism for denoising output weighted denoised spectrum samples. Clean spectral samples from the source domain A function for calculating cosine similarity in the frequency domain. Therefore, according to the chain rule:

[0090] Expanding the computational path of the feature extraction and classification network, we can obtain the final propagation chain:

[0091] Physical meaning: If the backbone generates a network Blindly over-smoothing leads to a decrease in the cosine similarity of key diagnostic frequency bands, affecting the classifier. The resulting huge partial derivative will pass through instantly This bridge backfires on the backbone generation network. This forcibly constrains the first layer of adversarial mechanisms at the front end to prioritize preserving diagnostic classification features, with noise reduction and reconstruction serving only the final diagnosis.

[0092] Based on the aforementioned dynamic transmission mechanism, in each main training loop, the adaptive optimizer, based on the constructed partial derivative system, synchronously optimizes the first layer of adversarial network. With the second layer of adversarial network The parameters are used for collaborative iteration. Since the second-level classification and domain alignment penalty always exists in the form of a dynamic correction term of the front-end regression loss, the system, through a combination of "soft" and "hard" constraints, finally trains a cross-domain collaborative noise resistance model that can be dynamically adjusted based on gradient guidance in order to "accurately mine local domain invariant features and serve the final diagnostic task".

[0093] In summary, targeting the radiated sound field generated by the operation of mechanical equipment in real industrial settings, this invention employs spatial array technology. Four identical measurement-grade microphones are rigidly fixed to a custom-designed spherical array support in a spatially symmetrical geometric configuration for omnidirectional data capture. During the signal acquisition phase, a multi-channel data acquisition front-end performs rigorous synchronous analog-to-digital conversion (A / D acquisition) on the physical sound pressure signals of the four microphone channels at a high-frequency sampling rate of 16kHz. The acquired raw one-dimensional time-domain acoustic signals are then stored in WAV lossless audio format. Subsequently, the continuous time-domain acoustic signals are sliced ​​into equal-length segments, strictly divided into 1-second time series as independent acoustic data analysis samples. The Short-Time Fourier Transform (STFT) concept is introduced, employing a 64ms Hamming window with a 50% frame overlap rate for frame segmentation and smooth truncation of each sample signal. Finally, a Mel-scale filter bank containing 40 triangular filters is applied to the power spectrum of each frame of the signal. By simulating the nonlinear perception characteristics of the human auditory system, the frequency axis is compressed logarithmically, ultimately transforming the one-dimensional waveform into a two-dimensional Mel-time spectral feature map with a dimension of 40×32 (number of frequency bands × number of time frames). Through an extremum normalization strategy, a linear mapping operator is used to uniformly scale and constrain the feature amplitudes in the spectral matrix within the standard closed interval [-1, 1]. This completes the feature engineering extraction of the original acoustic signal, outputting standardized two-dimensional acoustic sample training and testing sets for real industrial background noise scenarios and noise-free semi-anechoic chamber scenarios. The training set contains 72,000 clean and 72,000 noisy samples, while the testing set contains 14,400 clean and 14,400 noisy samples.

[0094] During the training phase, the first adversarial network mechanism takes two-dimensional acoustic samples containing non-static background noise interference collected from real industrial scenarios as input and clean two-dimensional acoustic samples as labels. Supervised regression training is conducted during the adversarial learning process between the backbone generator network (generator) and the convolutional discriminant network. The reconstructed spectrum output from the first adversarial network mechanism, i.e., the denoised spectrum sample and the clean spectrum sample, simultaneously serves as input to the second adversarial network mechanism, performing an asymmetric cross-domain transfer task. Specifically, the denoised spectrum sample and the clean spectrum sample pass through a multi-scale feature extraction module and a local fine-grained domain discriminant module, undergoing cross-domain feature alignment under the gradient inversion mechanism. Simultaneously, the clean spectrum sample independently flows through the fault classification network, undergoing supervised multi-class classification training under the supervision of its corresponding real fault category labels. The cross-domain noise-resistant diagnostic framework based on the dual-domain adversarial mechanism is collaboratively optimized under the guidance of a feature attention mask based on cosine similarity until convergence is achieved, meaning that the total loss (including mean squared error loss, adversarial loss, and cross-entropy loss) and the single cross-entropy loss evaluation system no longer show a significant decrease in loss values.

[0095] During the testing phase, 14,400 test set sample data with non-static noise components were used as input to the cross-domain collaborative noise reduction framework. The reconstructed (denoised) spectrum output by the backbone generator of the first-level adversarial network mechanism was used as input to the second-level adversarial network mechanism. Then, the fault prediction and identification results were obtained through the multi-scale feature extraction module and the fault classification module, realizing cross-domain collaborative noise reduction diagnosis under real industrial non-static background noise interference.

[0096] To verify the effectiveness of the method of this invention, structural ablation experiments and comparative experiments were designed and conducted under the same training parameter configuration. The Adam optimizer was used with a fixed learning rate of 0.0002, and the batch size was set to 200 (treating 200 data samples as one batch). The early stopping strategy was observed within 10 iterations to determine the final number of optimization iterations. The purpose of the structural ablation experiments was to verify the effectiveness of each functional module proposed in this invention, demonstrating the advantages and value of the proposed solution. Three different structural variants related to this invention were selected for verification analysis: a base network without a dual adversarial mechanism (i.e., retaining only the multi-scale feature extraction network and the fault classification network), a back-end noise-resistant architecture without a front-end adversarial mechanism (i.e., the first layer of adversarial network mechanism), and a front-end noise-resistant architecture without a second layer of adversarial network mechanism. Through the above ablation comparison, the effectiveness of the gradient-guided dual adversarial cross-domain collaborative noise-resistant diagnostic method proposed in this invention can be effectively verified. The effectiveness of the present invention and three variant schemes in aligning the data feature spaces of the source and target domains was verified using mean square error and mean absolute error. The final noise reduction diagnostic quality was measured using accuracy, and the results are shown in Table 1.

[0097] Table 1

[0098] Analysis of the test set validation results in Table 1 reveals that after processing by the proposed dual-adversarial cross-domain mechanism, the source domain (clean samples) and target domain (noisy samples) data exhibit the smallest Euclidean and Manhattan distances in the feature space. This indicates that the source and target domain features processed by this invention can be effectively aligned, demonstrating that the diagnostic knowledge learned in the source domain feature space can be effectively transferred to the target domain for cross-domain application. The final diagnostic results also confirm this conclusion, showing that the proposed method achieves a fault accuracy of 98.99% in non-static background noise scenarios, exhibiting the best noise-resistant diagnostic performance. In contrast, other ablation schemes, lacking a dual-adversarial mechanism in their basic models, rely solely on multi-scale feature extraction networks, making it difficult to align the source and target domain data in the feature space. The Euclidean and Manhattan distances between feature points differ significantly, exhibiting substantial differences in feature distribution. This ultimately leads to obstacles in the transfer of source domain diagnostic knowledge to the target domain, resulting in a collapse in diagnostic performance due to the inability to cope with cross-domain feature perturbations introduced by random and time-varying noise characteristics. For architectures with only a single noise mitigation mechanism, although the model is endowed with a certain feature alignment capability, reducing the Euclidean-Manhattan distance between source and target domain features in the feature space and promoting the transfer of source domain diagnostic knowledge to the target domain, thus enabling the model to possess a certain noise mitigation diagnostic capability under non-static background noise interference, the lack of a dual adversarial mechanism hinders the synergistic effect of the front-end and back-end, weakens the anti-interference capability against noise components, and inevitably leads to a "negative transfer" phenomenon, resulting in a lower noise mitigation diagnostic accuracy than the proposed scheme. To further verify the performance differences of different model architecture configurations in cross-domain feature alignment from the perspective of time-frequency feature visualization, sample data was randomly sampled from the test two-dimensional acoustic samples, and a visual analysis and comparison of the source and target domain features under each configuration architecture was conducted. The cross-domain feature alignment effect of each scheme was explored from a microscopic level, such as... Figure 2 As shown. Observation Figure 2 It can be observed that the features extracted by the proposed scheme in this invention have extremely strong similarity between the source and target domains. That is, the extracted features are unaffected by external environmental variables, and regardless of the presence or absence of time-varying noise, they output domain-invariant features with consistent structure in the feature space. This allows diagnostic knowledge to be smoothly transferred between the source and target domains, achieving excellent noise-resistant diagnostic performance. In contrast, other comparative architectures exhibit varying degrees of difference between their source and target domain features, indicating different degrees of "underfitting" (no transfer) and "overfitting" (negative transfer) problems, inevitably affecting diagnostic performance during cross-domain noise-resistant diagnosis. (See Table 1 and...) Figure 2By combining macroscopic statistical results with microscopic sampling analysis, it can be seen that the present invention, through the effective synergy of a two-stage adversarial mechanism, can extract domain-invariant features in the source and target domains. This enables the diagnostic knowledge learned from clean acoustic samples to ignore the interference of random and time-varying noise components and successfully migrate across domains to non-static background noise environments, ultimately achieving the purpose of noise-resistant diagnosis and verifying the effectiveness of the present invention.

[0099] Based on this, to further verify the contribution and value of this invention to the noise immunity diagnostic technology of rotating machinery, it was compared with existing mature cross-domain transfer techniques under the same evaluation index. The comparative experimental results are shown in Table 2. The comparison of experimental data in Table 2 shows that, in terms of feature space metrics, this invention achieves global minima in both MSE and MAE distance metrics. This strongly confirms that this method can drive the source and target domains to achieve high fit and overlap in the feature space, and the extracted features possess perfect "domain-invariant properties." It is precisely thanks to this high-quality feature alignment that the accurate diagnostic boundary (diagnostic knowledge) learned in the clean source domain can be transferred to the complex non-static noise target domain with low loss and high robustness, thus achieving optimal noise immunity diagnostic accuracy in macroscopic testing. Meanwhile, lateral correlation analysis shows that the three domain transfer comparison methods also basically follow the transfer rule of "the closer the feature space distance, the higher the cross-domain diagnostic accuracy" in overall performance. Although there are a few nonlinear special cases caused by statistical bias in high-dimensional space during the measurement process (for example, the MAE value of the LMMD method is about 50% higher than that of CORAL, but its MSE is a whole order of magnitude lower than that of CORAL. In the end, LMMD achieves higher diagnostic accuracy than CORAL by virtue of its MSE convergence advantage, which is more sensitive to outliers), this indirectly confirms that when faced with extremely complex non-static background noise, single or traditional general-purpose feature alignment methods often cannot take into account the comprehensive consistency of feature distribution.

[0100] Table 2

[0101] Finally, as Figure 3As shown, the output features are visualized in two dimensions using the t-SNE (t-distributed random neighborhood embedding) manifold dimensionality reduction algorithm. The samples processed by this invention exhibit excellent clustering with highly compact intra-class structures and clear inter-class boundaries. In contrast, other comparative methods reveal varying degrees of inter-class overlap, confusion, and even severe class drift misjudgment at the edges of the distributed clusters. The fundamental mechanism lies in the fact that traditional domain adaptation methods, when faced with high-intensity non-static random noise, easily fall into the trap of blindly aligning by excessively extracting useless background features, thus triggering a severe "negative transfer" effect. The mutual corroboration between the numerical distance evaluation and the visualized clustering results fully highlights the core technological superiority of this invention, which utilizes a mask hub collaborative dual-adversarial mechanism to implement dynamic gradient-guided cross-domain transfer of diagnostic knowledge, thereby accurately avoiding negative transfer and achieving cross-domain collaborative noise reduction.

Claims

1. A cross-domain collaborative noise immunity diagnostic method based on dynamic gradient guidance, characterized in that, Includes the following steps: S1. For the radiated sound field generated by the operation of mechanical equipment in real industrial sites, acoustic data is collected from all directions using a spatial array, and the acoustic data is preprocessed to obtain a two-dimensional time-frequency characteristic spectrum. S2. Construct a first-level adversarial network mechanism for filtering out time-varying non-static noise, wherein the first-level adversarial network mechanism is used to perform front-end noise reduction and reconstruction on the two-dimensional time-frequency feature map; S3. Construct a second adversarial network mechanism based on feature attention mask and local fine-grained adaptation. The second adversarial network mechanism is used to perform cross-domain feature alignment based on the front-end noise-resistant reconstruction results under the action of dynamic gradient reversal mechanism. S4. The first and second adversarial network mechanisms are placed in the same end-to-end cross-domain noise immunity diagnosis framework. The cross-domain noise immunity diagnosis framework is collaboratively optimized under the guidance of feature attention mask based on cosine similarity. The optimized cross-domain noise immunity diagnosis framework is used to output fault prediction and identification results, so as to realize cross-domain collaborative noise immunity diagnosis under real industrial non-static background noise interference.

2. The cross-domain collaborative noise immunity diagnostic method based on dynamic gradient guidance according to claim 1, characterized in that, S1 includes the following steps: Construct a multi-channel spherical microphone array; For the radiated sound field generated by the operation of mechanical equipment in real industrial sites, a multi-channel spherical microphone array is used to collect acoustic data from all directions. During the acquisition process, the physical sound pressure of the multi-microphone channels is synchronously converted from analog to digital at a high-frequency sampling rate, and the original one-dimensional time-domain acoustic signal is saved as a lossless WAV audio format. Based on WAV lossless audio signals, continuous time-domain acoustic signals are sliced ​​into independent acoustic data analysis samples by performing equal-length slicing. The acoustic data analysis sample signals are subjected to frame segmentation and smoothing truncation processing. By applying the power spectrum of each frame of acoustic signal after framing and smoothing, a one-dimensional waveform is converted into a two-dimensional Mel-time spectrum feature map by using a Mel-frequency standard filter bank. The amplitude extrema of the two-dimensional Mel-time spectral feature map are normalized to output a standardized two-dimensional time-frequency feature map, thus completing the preprocessing.

3. The cross-domain collaborative noise immunity diagnostic method based on dynamic gradient guidance according to claim 1, characterized in that, The first layer of adversarial network mechanism includes a backbone generation network with a symmetric encoding-decoding architecture and a convolutional discriminant network connected to the backbone generation network, wherein the convolutional discriminant network constitutes the first layer of adversarial constraint; The backbone generation network is used to reconstruct the two-dimensional time-frequency feature map using a three-layer downsampling-three-layer upsampling topology to obtain the reconstructed spectrum, wherein the reconstructed spectrum is a target domain noise-reduced spectrum sample. The convolutional discriminant network is used to perform feature compression processing on the reconstructed spectrum and the clean spectrum samples of real industrial machinery in a noise-free environment using a three-layer fully convolutional network. Based on the compressed features, a global average pooling layer is used to flatten the multi-dimensional spatial information and output a scalar probability value that represents the authenticity / purity of the input sample, thus completing the front-end noise-resistant reconstruction. The title probability value is used to optimize the backbone generation network.

4. The cross-domain collaborative noise immunity diagnostic method based on dynamic gradient guidance according to claim 1, characterized in that, The second layer of adversarial network mechanism includes: The feature mask generation module is used to reconstruct the target domain denoised spectrum samples in the frequency domain by calculating the first-level adversarial network mechanism. With the corresponding clean spectrum samples of the source domain Based on the cosine similarity of spatial locations, weighted clean spectrum samples are generated. and weighted noise reduction spectrum samples ; A multi-scale feature extraction module is used for extracting features based on weighted clean spectral samples. and weighted noise reduction spectrum samples The parallel convolutional branches output tensor features, and the tensor features of each convolutional branch are stacked and spliced ​​to form a fused high-dimensional feature. The fused high-dimensional feature is then compressed and nonlinearly combined to output a common basic feature matrix, which includes weighted clean spectral features and weighted denoised spectral features. The local fine-grained domain discrimination module is used to output a spatial domain discrimination matrix based on the common basic feature matrix and the introduced local cross-domain adversarial architecture. The spatial domain discrimination matrix is ​​then fed back into the optimization of the common basic matrix using gradient inversion. The local cross-domain adversarial architecture is equipped with a gradient inversion layer at the front end. Each position in the spatial domain discrimination matrix represents the discrimination probability result for each local difference in the common basic feature matrix. The independent fault classification module is used to output a probability prediction matrix based on the weighted clean spectral features output by the multi-scale feature extraction module, and based on the probability prediction matrix, select the maximum extreme value term in the probability vector and output the fault type identification result of the mechanical equipment corresponding to the term.

5. The cross-domain collaborative noise immunity diagnostic method based on dynamic gradient guidance according to claim 4, characterized in that, The weighted clean spectrum sample and weighted noise reduction spectrum samples ,include: In the frequency domain, the target domain denoised spectrum samples generated by the first-level adversarial network mechanism are calculated. With the corresponding clean spectrum samples of the source domain Cosine similarity in spatial location; Based on the calculated cosine similarity, a feature attention mask matrix is ​​generated through a nonlinear activation function mapping. ; Using feature attention mask matrix The clean spectral samples from the source domain and the denoised spectral samples from the target domain entering the second layer of the adversarial network mechanism are subjected to an element-wise Hadamard product weighting operation to generate weighted clean spectral samples. and weighted noise reduction spectrum samples ,in, This represents the Hadama product.

6. The cross-domain collaborative noise immunity diagnostic method based on dynamic gradient guidance according to claim 4, characterized in that, The output spatial domain discrimination matrix includes: The common basic feature matrix is ​​continuously passed through multiple concatenated convolutional layers with a kernel dimension of 4×4 to compress the spatial receptive field; After spatial receptive field compression, channel-scale compression is performed by single-channel convolution with a 1×1 kernel to output a multi-dimensional spatial domain discrimination matrix. In this matrix, a single element represents the domain prediction attribute of the local time-frequency sensing region on the original two-dimensional time spectrum.

7. The cross-domain collaborative noise immunity diagnostic method based on dynamic gradient guidance according to claim 1, characterized in that, S4 includes: The first and second adversarial network mechanisms are placed within the same end-to-end cross-domain noise reduction diagnostic framework. Set the feature attention mask based on cosine similarity as the physical hub for backpropagation; A joint loss function is constructed, which is used to align the diagnosis of the second layer of adversarial network mechanism with cross-domain features in the form of dynamic gradients to penetrate the physical hub, reversely reshape and dominate the update trajectory of the first layer of adversarial network mechanism's regression reconstruction target, so as to perform collaborative optimization of the cross-domain noise resistance diagnosis framework. The optimized cross-domain noise immunity diagnostic framework is used to output fault prediction and identification results, so as to realize cross-domain collaborative noise immunity diagnosis under real industrial non-static background noise interference.

8. The cross-domain collaborative noise immunity diagnostic method based on dynamic gradient guidance according to claim 7, characterized in that, The expression for the joint loss function is as follows: in, Denotes the joint loss function. This represents the supervised mean squared error regression loss in the first layer of adversarial network mechanism. This represents the least-squares generative adversarial loss in the first-level adversarial network mechanism. This represents the local fine-grained adversarial loss of the second layer of adversarial network mechanism. This indicates the source domain purification classification diagnostic loss of the second layer of adversarial network mechanism. This represents the parameters of the backbone generation network in the first layer of adversarial network mechanism. This represents the backbone generation network in the first layer of adversarial network mechanism. This represents a noisy spectrum sample. This represents the convolutional discriminant network in the first layer of adversarial network mechanism. This represents the z-th target domain noise reduction spectrum sample. This represents the z-th clean spectrum sample from the source domain. This represents the weighted denoised spectrum sample. This represents the target domain denoised spectrum sample. This represents a weighted clean spectrum sample. Represents the feature attention mask matrix. This represents a clean spectrum sample in the source domain. Represents the Hadama product. Represents a classifier. This indicates a multi-scale parallel feature extraction module. This indicates the number of data samples in the training set. z Indicates the first z A dataset sample, This represents the expectation of the noisy spectral sample. x Represents a random variable. This represents the expectation of a clean spectral sample in the source domain. This represents the output of the backbone network. Indicates the weighted clean spectral characteristics Expectations and These represent the number of frequency bands and the number of frames in the original two-dimensional time spectrum, respectively. This represents the output of the local fine-grained domain discrimination module. This represents a local fine-grained domain discrimination module. Indicates the weighted noise reduction spectral characteristics Expectations This represents the true fault category in the clean sample. This indicates the predicted fault category output by the independent fault classification module.

Images