Fault detection method of auditory features and pg-cae in noisy environment

Abstract

The application discloses a method for hearing feature and PG-CAE fault detection in a noise environment, belongs to the technical field of bearing state monitoring and intelligent fault diagnosis, and aims to solve the problems of insufficient fault sample and poor fault detection stability in an industrial noise environment. The method extracts the hearing cortex response feature of a bearing sound signal, performs signed nonlinear logarithmic mapping, constructs a grouping decoupling convolution autoencoder, and only uses normal working condition samples for unsupervised training. Based on the structural similarity of input features and reconstructed features, the reconstruction error of each channel is calculated, the maximum value of the maximum value after normal working condition statistical benchmark standardization is taken as an abnormal score, and the 99th percentile of the normal working condition sample abnormal score is taken as a classification threshold, so that the normal state and fault state of the bearing are determined. The application is used for bearing operation state monitoring and fault detection in a noise environment.

Description

Technical Field

[0001] This invention relates to the field of industrial equipment condition monitoring and fault diagnosis technology, and in particular to a method for equipment fault detection based on acoustic signals under conditions of strong background noise interference and scarce fault samples. Specifically, this invention mainly discloses a fault detection method based on auditory characteristics and PG-CAE in a noisy environment, applicable to health status monitoring of rotating machinery such as bearings and gearboxes under complex operating conditions. Background Technology

[0002] As a core component in high-end equipment manufacturing, bearings' operating condition directly determines the safety and stability of the entire mechanical system. While vibration signal-based analysis methods are mature in equipment condition monitoring, their sensors require contact installation, which is challenging to deploy on-site due to limitations imposed by spatial structures, high temperatures, and high pressures. In contrast, acoustic signals offer advantages such as non-contact measurement, rich information, and flexible sensor deployment, providing a more convenient solution for equipment fault diagnosis.

[0003] However, the application of acoustic signal-based fault diagnosis technology in real-world industrial scenarios still faces two major challenges. First, strong background noise interference (such as machine noise, human voices, and environmental reverberation) is prevalent in industrial environments. These complex, non-stationary noises easily mask subtle acoustic features of faults, leading to a significant decrease in the robustness of traditional diagnostic models. Second, equipment operates normally for the vast majority of the time, resulting in extremely scarce and costly fault samples. This makes it difficult to effectively train traditional supervised learning models, limiting their generalization performance.

[0004] To suppress noise interference, some researchers have employed hardware methods such as acoustic metamaterials or lock-in amplifiers to enhance signals. However, these methods are typically limited to specific frequency bands and are susceptible to environmental factors. In terms of software algorithms, traditional signal processing methods such as Fourier decomposition and empirical wavelet transform still lack sufficient adaptability and robustness when facing multi-source, time-varying composite noise in real factory workshops. In recent years, some studies have attempted to use deep neural networks to learn features end-to-end from noisy signals. However, these methods heavily rely on large amounts of labeled fault data, making it difficult to address the problem of scarce fault samples, and they are prone to overfitting to noise components in high-noise environments.

[0005] Meanwhile, the human auditory system exhibits remarkable resistance to interference in noisy environments. Modern biological research indicates that the complete auditory processing involves two stages: the peripheral auditory system and the auditory cortex. The peripheral auditory system performs fine frequency analysis through the cochlear filter bank, while the auditory cortex deeply decouples the spectral envelope and temporal evolution through multi-scale time-frequency modulation, thereby accurately locating the target sound source in noise. How to integrate this efficient noise-resistant processing mechanism of the auditory center with industrial fault detection, and address the problem of scarce fault samples, is a bottleneck that urgently needs to be overcome in the current technological field.

[0006] In summary, existing acoustic signal-based fault detection technologies generally suffer from weak feature representation capabilities, poor model robustness, and strong dependence on fault labels in complex industrial scenarios with strong noise interference and scarce samples. Therefore, from a biomimetic perspective, providing a method that can achieve highly reliable fault detection using only normal samples in low signal-to-noise ratio environments is a technical challenge that urgently needs to be solved by those skilled in the art. Summary of the Invention

[0007] In view of the above problems, the present invention provides a fault detection method based on auditory features and PG-CAE in noisy environments. It aims to mimic the noise resistance mechanism of the human auditory center, extract robust auditory features, and combine them with an unsupervised autoencoder network to solve the problem of decreased detection performance caused by the scarcity of fault samples under strong background noise interference.

[0008] To achieve the above objectives, the present invention adopts the following technical solution:

[0009] A fault detection method based on auditory characteristics and PG-CAE in a noisy environment, characterized by comprising the following steps:

[0010] Step S1: Extract auditory cortex response features from the audio segment to obtain the three-dimensional feature tensor. , dimension ,in The number of channels for auditory feature slices. Number of frequency channels For time frames;

[0011] Step S2: Process the feature tensor obtained in step S1 A signed nonlinear logarithmic mapping is performed to obtain the enhanced feature tensor. ;

[0012] Step S3: Construct a grouped decoupled convolutional autoencoder PG-CAE, which consists of an encoder and a decoder; the encoder contains three cascaded grouped convolutional layers, each followed by a batch normalization layer and a ReLU activation function; the decoder is a mirror-symmetric structure of the encoder, containing three cascaded grouped transposed convolutional layers.

[0013] Step S4: Using bearing acoustic signal samples containing only normal operating conditions, process them according to steps S1 to S2 to obtain the feature tensor of the normal samples. Input the PG-CAE model constructed in step S3 for unsupervised training; the training uses the mean absolute error loss function and the optimizer is AdamW;

[0014] Step S5: Input the normal operating condition benchmark samples into the trained grouped decoupled convolutional autoencoder PG-CAE to obtain the corresponding reconstructed features; for each auditory feature channel, calculate the structural similarity index (SSIM) between the input features and the reconstructed features, and use it as... The reconstruction error is used as the corresponding channel's reconstruction error; based on the reconstruction errors of each channel in multiple normal operating condition benchmark samples, the mean reconstruction error corresponding to each auditory feature channel is determined. and reconstruction error standard deviation Establish a statistical benchmark for reconfiguration error under normal operating conditions.

[0015] Step S6: Input the sample to be detected into the trained grouped decoupled convolutional autoencoder PG-CAE, standardize the reconstruction error of each auditory feature channel based on the normal working condition reconstruction error statistical benchmark, and determine the abnormal score of the sample to be detected based on the standardized reconstruction error of each channel.

[0016] Step S7: Select a threshold benchmark sample set containing only normal operating condition samples. Using the same anomaly score calculation method as the sample to be detected, obtain the anomaly score of each normal operating condition sample in the threshold benchmark sample set, and determine the 99th percentile of the anomaly score of each normal operating condition sample as the anomaly judgment threshold. The abnormal score of the sample to be detected. With the aforementioned anomaly detection threshold When comparing, When the bearing is in a faulty state, it is determined that the bearing is in a faulty state; when When the bearing is in normal condition, the system determines that the bearing is in normal condition and outputs the bearing fault detection result.

[0017] Preferably, the extraction process of auditory cortex response features in step S1 includes the following sub-steps:

[0018] Step S1-1: Input the audio signal into the cochlear filter bank, and obtain the auditory spectrogram through inner hair cell conversion and lateral inhibition network processing;

[0019] Step S1-2: Input the auditory spectrogram into the auditory cortex model and fix the spectral modulation scale. cycles / oct and cycles / oct, respectively, are combined with the time modulation rate , , , A total of 8 parameter combinations were formed, and 8 two-dimensional spectrum-time feature slices were extracted.

[0020] Step S1-3: Stack the eight two-dimensional feature slices according to the channel dimension to form a three-dimensional feature tensor. Its dimensions are The number of channels in the auditory feature slice Number of frequency channels Time frame count .

[0021] Preferably, the encoder in step S3 has a single sample input tensor dimension of 8×128×313; the encoder includes three cascaded grouped convolutional layers, each with a kernel size of 3×3, a stride of 2, padding of 1, and a group number of G=8, with output channels of 64, 128, and 256 respectively; after three layers of grouped convolution, the output bottleneck layer feature dimension is 256×16×40, where the 256 bottleneck layer channels are grouped into 8 groups, each group corresponding to 32 feature channels; the decoder includes three cascaded grouped transposed convolutional layers, each with a kernel size of 3×3, a stride of 2, padding of 1, output padding of 1 in the frequency dimension and 0 in the time frame dimension, a group number of 8, and output channels of 128, 64, and 8 respectively, and the final output reconstructed tensor dimension is restored to 8×128×313.

[0022] Preferably, the training process in step S4 uses only normal operating condition samples; random batch sampling is used during training, with a batch size of... The optimizer uses AdamW, whose momentum parameters... , The learning rate is fixed at ; No attenuation is applied; weight attenuation coefficient Number of iteration rounds .

[0023] Preferably, during the SSIM calculation in step S5, the local window size is... ,constant , Calculate the SSIM value independently for each channel to obtain the reconstruction error value for each channel. , .

[0024] The fault detection method based on auditory characteristics and PG-CAE in a noisy environment provided by this invention has the following advantages compared with the prior art:

[0025] 1. This invention extracts multi-scale auditory features by constructing an auditory central model. These features mimic the selective sensitivity of the human auditory cortex to spectral envelope and temporal modulation, effectively highlighting weak fault impact components under strong background noise interference, and significantly improving the noise resistance robustness of feature expression.

[0026] 2. This invention designs a grouped decoupled convolutional autoencoder (PG-CAE). By utilizing the physical isolation characteristics of grouped convolution, an independent parallel feature extraction channel with the same number of input channels is established inside the network. This effectively prevents background noise from penetrating across channels and compensating for information in different sub-bands, allowing the model to focus on learning the essential manifold of normal working conditions and avoiding overfitting of the model to random noise during training.

[0027] 3. This invention adopts an unsupervised learning framework, which only requires acoustic signal samples under normal working conditions to complete model training. It quantifies the degree of anomaly by calculating the structural similarity error between test samples and reconstructed samples, thus solving the practical problems of scarce fault samples and difficulty in obtaining labeled data in industrial sites and lowering the threshold for engineering applications.

[0028] 4. This invention introduces a structural similarity index (SSIM) based on each channel and a maximum value standardization strategy as anomaly scores. Compared with traditional detection methods based on global pixel-level errors, it is more sensitive to subtle local structural distortions and can more accurately capture weak reconstruction failures caused by faults. Thus, it can maintain a detection accuracy of over 90% even in low signal-to-noise ratio (e.g., -5dB) environments, exhibiting extremely high detection sensitivity and reliability. Attached Figure Description

[0029] To more clearly illustrate the specific embodiments of the present invention or the technical solutions in the prior art, the accompanying drawings used in the description of the specific embodiments or the prior art will be briefly introduced below. In all the drawings, similar elements or parts are generally identified by similar reference numerals. In the drawings, the elements or parts are not necessarily drawn to scale.

[0030] Figure 1 This is a flowchart illustrating the overall implementation of the fault detection method of the present invention.

[0031] Figure 2 The first channel visualization comparison chart shows the original and reconstructed features of four states (bearing normal, inner ring pitting, outer ring pitting, cage failure) at a signal-to-noise ratio of 5dB.

[0032] Figure 3 Visualized error plots of the first channel for four state samples at a signal-to-noise ratio of 5dB.

[0033] Figure 4 A visualization of the t-SNE feature distribution of the test set samples in the bottleneck layer at a signal-to-noise ratio of 5dB.

[0034] Figure 5 A visualization of the t-SNE feature distribution of the test set samples in the bottleneck layer at a signal-to-noise ratio of 0dB.

[0035] Figure 6 A visualization of the t-SNE feature distribution of the test set samples in the bottleneck layer at a signal-to-noise ratio of -5dB.

[0036] Figure 7 The bearing fault detection results are based on auditory features and grouped decoupled autoencoders under three signal-to-noise ratios.

[0037] Figure 8 The bearing fault diagnosis results are shown for different models under three signal-to-noise ratios. Detailed Implementation

[0038] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some, not all, of the embodiments of the present invention. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0039] It should be understood that, when used in this specification and the appended claims, the terms "comprising" and "including" indicate the presence of the described features, integrals, steps, operations, elements and / or components, but do not exclude the presence or addition of one or more other features, integrals, steps, operations, elements, components and / or collections thereof.

[0040] It should also be understood that the terminology used in this specification is for the purpose of describing particular embodiments only and is not intended to limit the invention. As used in this specification and the appended claims, the singular forms “a,” “an,” and “the” are intended to include the plural forms unless the context clearly indicates otherwise.

[0041] It should also be further understood that the term "and / or" as used in this specification and the appended claims refers to any combination of one or more of the associated listed items and all possible combinations, and includes such combinations.

[0042] like Figure 1As shown, a specific embodiment of the present invention provides a fault detection method for auditory characteristics and PG-CAE in a noisy environment, including the following steps:

[0043] Step S1: Auditory cortex response feature extraction. The acquired bearing sound signal audio segment is first input into a cochlear filter bank, and then processed by inner hair cell transformation and lateral inhibition network to obtain an auditory spectrogram. Subsequently, the auditory spectrogram is input into the auditory cortex model, and the spectral modulation scale is fixed. cycles / oct and cycles / oct, respectively, are combined with the time modulation rate , , , A total of 8 parameter combinations were formed, resulting in 8 two-dimensional spectral-temporal feature slices. Finally, these 8 two-dimensional feature slices were stacked according to the channel dimension to form a three-dimensional feature tensor. , dimension The number of channels in the auditory feature slice Number of frequency channels Time frame count .

[0044] Step S2: Nonlinear logarithmic mapping enhancement, applying the feature tensor obtained in step S1... A signed nonlinear logarithmic mapping is performed to obtain the enhanced feature tensor. This is used as input for subsequent models.

[0045] Step S3: Construct the Grouped Decoupled Convolutional Autoencoder (PG-CAE). The PG-CAE consists of an encoder and a decoder. The encoder's single-sample input tensor has a dimension of 8×128×313 and contains three cascaded grouped convolutional layers. Each layer has a kernel size of 3×3, a stride of 2, padding of 1, and a group number of [missing information]. The number of output channels is 64, 128, and 256 respectively; each convolutional layer is followed by a batch normalization layer and a ReLU activation function. After three layers of grouped convolutions, the feature dimension of the output bottleneck layer is 256×16×40, where the 256 bottleneck layer channels are grouped into 8 groups, with each group corresponding to 32 feature channels. The decoder is a mirror-symmetric structure of the encoder, containing three cascaded grouped transposed convolutional layers. Each layer has a kernel size of 3×3, a stride of 2, and padding of 1. The output padding is 1 in the frequency dimension and 0 in the time frame dimension. The number of groups is 8, and the number of output channels is 128, 64, and 8 respectively. Finally, the reconstructed tensor dimension is restored to 8×128×313.

[0046] Step S4: Unsupervised training. Using bearing acoustic signal samples containing only normal operating conditions, process them according to steps S1-S2 to obtain the feature tensor of the normal samples. The PG-CAE model constructed in step S3 is then used for unsupervised training. Random batch sampling is used during training, with a batch size of... The optimizer uses AdamW with momentum parameters. , The learning rate is fixed at 1 And no attenuation is applied, the weight attenuation coefficient Number of training iterations Unsupervised training uses the mean absolute error as the loss function.

[0047] Step S5: Input the normal operating condition benchmark samples into the trained grouped decoupled convolutional autoencoder PG-CAE to obtain the corresponding reconstructed features; for each auditory feature channel, calculate the structural similarity index SSIM between the input features and the reconstructed features, and the local window size. ,constant , ; and with The reconstruction error is used as the corresponding channel's reconstruction error; based on the reconstruction errors of each channel in multiple normal operating condition benchmark samples, the mean reconstruction error corresponding to each auditory feature channel is determined. and reconstruction error standard deviation Establish a statistical benchmark for reconfiguration error under normal operating conditions.

[0048] Step S6: Input the sample to be detected into the trained grouped decoupled convolutional autoencoder PG-CAE, standardize the reconstruction error of each auditory feature channel based on the normal working condition reconstruction error statistical benchmark, and determine the abnormal score of the sample to be detected based on the standardized reconstruction error of each channel.

[0049] Step S7: Select a threshold benchmark sample set containing only normal operating condition samples. Using the same anomaly score calculation method as the sample to be detected, obtain the anomaly score of each normal operating condition sample in the threshold benchmark sample set, and determine the 99th percentile of the anomaly score of each normal operating condition sample as the anomaly judgment threshold. The abnormal score of the sample to be detected. With the aforementioned anomaly detection threshold When comparing, When the bearing is in a faulty state, it is determined that the bearing is in a faulty state; when When the bearing is in normal condition, the system determines that the bearing is in normal condition and outputs the bearing fault detection result.

[0050] To visually verify the reconstruction effect of PG-CAE, this invention extracted the first channel visualization comparison images of the original features and reconstructed features of four state samples at a 5dB signal-to-noise ratio, as shown below. Figure 2 As shown, the reconstruction error map is further calculated and visualized, as follows. Figure 3 As shown, to explore the separability of latent space features in PG-CAE, this invention uses t-SNE to perform nonlinear dimensionality reduction visualization of bottleneck layer features in the test set at signal-to-noise ratios of 5dB, 0dB, and -5dB, as shown. Figure 4 , Figure 5 , Figure 6 As shown.

[0051] Figure 7 The bar chart shows the detection accuracy of the method of the present invention at three signal-to-noise ratios of 5dB, 0dB and -5dB. The model accuracy reaches 100.00% at a signal-to-noise ratio of 5dB, and 97.50% at 0dB. Even when the signal-to-noise ratio drops to -5dB, the accuracy remains at 90.83%. The bar chart clearly shows the gradual decay trend of each indicator as the signal-to-noise ratio decreases.

[0052] To further illustrate the advantages of the feature extraction method of the present invention, Figure 8 The bar chart shows the comparative accuracy of four methods, including the method of this invention, at three signal-to-noise ratios. The present invention achieves 100.00% accuracy at 5dB, significantly outperforming Isolation Forest (94.17%), Single-class SVM (98.33%), and Deep SVDD (83.33%). At 0dB, the present invention achieves 97.50%, while the comparative methods achieve 95.83%, 95.00%, and 85.83%, respectively. At -5dB, the present invention maintains 90.83%, while Isolation Forest drops to 84.17%, Single-class SVM plummets to 76.67%, and Deep SVDD remains at 83.33%. The bar chart clearly demonstrates that the present invention significantly outperforms the comparative methods at all signal-to-noise ratios, fully validating the effectiveness of the grouped decoupling architecture in accurately capturing the essential manifold.

[0053] Those skilled in the art will recognize that the units of the various examples described in connection with the embodiments disclosed herein can be implemented in electronic hardware, computer software, or a combination of both. To clearly illustrate the interchangeability of hardware and software, the components of the various examples have been generally described in terms of functionality in the foregoing description. Whether these functions are implemented in hardware or software depends on the specific application and design constraints of the technical solution. Those skilled in the art can use different methods to implement the described functions for each specific application, but such implementations should not be considered beyond the scope of the invention.

[0054] In the embodiments provided in this application, it should be understood that the division of units is only a logical functional division. In actual implementation, there may be other division methods, such as multiple units can be combined into one unit, one unit can be split into multiple units, or some features can be ignored.

[0055] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention, and not to limit them. Although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some or all of the technical features therein. Such modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the scope of the technical solutions of the embodiments of the present invention, and they should all be covered within the scope of the claims and specification of the present invention.

Claims

1. A fault detection method based on auditory characteristics and PG-CAE in a noisy environment, characterized in that, Includes the following steps: Step S1: Extract auditory cortex response features from the audio segment to obtain a three-dimensional feature tensor. , dimension ,in The number of channels for auditory feature slices. Number of frequency channels For time frames; Step S2: Process the feature tensor obtained in step S1 A signed nonlinear logarithmic mapping is performed to obtain the enhanced feature tensor. ; Step S3: Construct a grouped decoupled convolutional autoencoder PG-CAE, which consists of an encoder and a decoder; the encoder contains three cascaded grouped convolutional layers, each followed by a batch normalization layer and a ReLU activation function; the decoder is a mirror-symmetric structure of the encoder, containing three cascaded grouped transposed convolutional layers. Step S4: Using bearing acoustic signal samples containing only normal operating conditions, process them according to steps S1 to S2 to obtain the feature tensor of the normal samples. Input the PG-CAE model constructed in step S3 for unsupervised training; the training uses the mean absolute error loss function and the optimizer is AdamW; Step S5: Input the normal operating condition benchmark samples into the trained grouped decoupled convolutional autoencoder PG-CAE to obtain the corresponding reconstructed features; for each auditory feature channel, calculate the structural similarity index (SSIM) between the input features and the reconstructed features, and use it as... The reconstruction error is used as the corresponding channel's reconstruction error; based on the reconstruction errors of each channel in multiple normal operating condition benchmark samples, the mean reconstruction error corresponding to each auditory feature channel is determined. and reconstruction error standard deviation Establish a statistical benchmark for reconfiguration error under normal operating conditions. Step S6: Input the sample to be detected into the trained grouped decoupled convolutional autoencoder PG-CAE, standardize the reconstruction error of each auditory feature channel based on the normal working condition reconstruction error statistical benchmark, and determine the abnormal score of the sample to be detected based on the standardized reconstruction error of each channel. Step S7: Select a threshold benchmark sample set containing only normal operating condition samples. Using the same anomaly score calculation method as the sample to be detected, obtain the anomaly score of each normal operating condition sample in the threshold benchmark sample set, and determine the 99th percentile of the anomaly score of each normal operating condition sample as the anomaly judgment threshold. The abnormal score of the sample to be detected. With the aforementioned anomaly detection threshold When comparing, When the bearing is in a faulty state, it is determined that the bearing is in a faulty state; when When the bearing is in normal condition, the system determines that the bearing is in normal condition and outputs the bearing fault detection result.

2. The fault detection method based on auditory characteristics and PG-CAE in a noisy environment according to claim 1, characterized in that, The extraction process of auditory cortex response features in step S1 includes the following sub-steps: Step S1-1: Input the audio signal into the cochlear filter bank, and obtain the auditory spectrogram through inner hair cell conversion and lateral inhibition network processing; Step S1-2: Input the auditory spectrogram into the auditory cortex model and fix the spectral modulation scale. cycles / oct and cycles / oct, respectively, are combined with the time modulation rate , , , A total of 8 parameter combinations were formed, and 8 two-dimensional spectrum-time feature slices were extracted. Step S1-3: Stack the eight two-dimensional feature slices according to the channel dimension to form a three-dimensional feature tensor. Its dimensions are The number of channels in the auditory feature slice Number of frequency channels Time frame count .

3. The fault detection method based on auditory characteristics and PG-CAE in a noisy environment according to claim 1, characterized in that, In step S3, the encoder's single-sample input tensor dimension is 8×128×313. The encoder includes three cascaded grouped convolutional layers, each with a kernel size of 3×3, a stride of 2, padding of 1, and a group number of G=8. The number of output channels is 64, 128, and 256 respectively. After the three-layer grouped convolution, the output bottleneck layer feature dimension is 256×16×40, where the 256 bottleneck layer channels are grouped into 8 groups, with each group corresponding to 32 feature channels. The decoder includes three cascaded grouped transposed convolutional layers, each with a kernel size of 3×3, a stride of 2, padding of 1, and output padding of 1 in the frequency dimension and 0 in the time frame dimension. The number of groups is 8, and the number of output channels is 128, 64, and 8 respectively. Finally, the output reconstructed tensor dimension is restored to 8×128×313.

4. The fault detection method based on auditory characteristics and PG-CAE in a noisy environment according to claim 1, characterized in that, The training process described in step S4 uses only normal operating condition samples; random batch sampling is used during training, with a batch size of... The optimizer uses AdamW, whose momentum parameters... , The learning rate is fixed at ; No attenuation is applied; weight attenuation coefficient Number of iteration rounds .

5. The fault detection method based on auditory characteristics and PG-CAE in a noisy environment according to claim 1, characterized in that, During the SSIM calculation described in step S5, the local window size... ,constant , Calculate the SSIM value independently for each channel to obtain the reconstruction error value for each channel. , .

Images