Non-contact bearing fault detection method and system for industrial rotating equipment

Abstract

This invention discloses a non-contact bearing fault detection method and system for industrial rotating equipment. The method involves collecting the operating sound of the rotating equipment to construct a one-dimensional time-series vector and a two-dimensional spectrogram. Global time-frequency distribution characteristics are extracted from the two-dimensional spectrogram, and acoustic features are reconstructed. The obtained two-dimensional features are then passed through a cascaded network to capture features and perform weighted operations using a coordinate attention mechanism to obtain a high-order two-dimensional feature map. A one-dimensional auxiliary branch is used to extract the one-dimensional time-series vector features and map them to channel weights. Residual modulation is then applied to obtain the final two-dimensional feature map. Statistical parameters of the one-dimensional time-series vector are extracted to perceive the energy distribution of the original acoustic signal. A classifier outputs the fault detection result. This invention solves the problem of sound signal differences caused by different acquisition environments and the problem of fault feature confusion caused by background noise interference, significantly improving the accuracy and reliability of non-contact bearing fault detection for industrial rotating equipment.

Description

Technical Field

[0001] This invention belongs to the field of industrial bearing fault diagnosis technology, specifically relating to a non-contact bearing fault detection method and system for industrial rotating equipment. Background Technology

[0002] Rotating equipment is widely used in modern industrial production systems. As a key component supporting rotating parts, the operating condition of bearings determines the safety and stability of the entire machine. Because bearings operate under high loads for extended periods, they are prone to fatigue peeling, wear, and other damage. If such faults are not detected in time through monitoring, they may lead to catastrophic equipment damage and huge economic losses.

[0003] Traditional bearing monitoring primarily relies on contact vibration monitoring technology. However, in certain industrial scenarios, contact monitoring faces significant technical limitations. First, when rotating equipment operates in environments with high temperatures, high pressures, or strong electromagnetic radiation, vibration sensors struggle to maintain stable operation over extended periods. Second, during equipment operation, physical space and safety boundaries often prevent the convenient placement or installation of contact sensors. In contrast, acoustic monitoring technology offers advantages such as non-contact data acquisition and minimal disruption to equipment operation, making it an effective approach to bearing diagnosis in these scenarios.

[0004] Despite the advantages of acoustic monitoring, existing acoustic detection methods still face several bottlenecks in practical applications. Firstly, there's the issue of sound signal variations due to different acquisition environments. Limited by on-site installation conditions, the distance between the sensor and the bearing under test is difficult to fix, and the characteristics of different working spaces (open fields versus enclosed indoor spaces) differ, leading to varying attenuation and distortion of sound waves during propagation. Existing models are mostly trained based on signals from specific fixed environments, lacking the ability to adapt to fluctuations in acquisition conditions, thus affecting the diagnostic accuracy of the models under different acquisition environments. Secondly, there's the problem of feature confusion caused by background noise interference. Industrial scenarios are accompanied by a large amount of background noise (such as metallic clanging sounds and airflow pulsations). This background noise mixes with the sound of bearing operation, affecting the identification of bearing sound characteristics and consequently impacting the accuracy of bearing fault detection.

[0005] Therefore, there is an urgent need to develop a non-contact bearing fault detection method that can effectively overcome background noise interference, adapt to changes in the acquisition environment, and accurately distinguish fault characteristics, so as to improve the accuracy and reliability of bearing detection in industrial rotating equipment. Summary of the Invention

[0006] To address the shortcomings of existing technologies and improve the accuracy and reliability of bearing fault detection in industrial rotating equipment, this invention adopts the following technical solution:

[0007] A non-contact bearing fault detection method for industrial rotating equipment includes the following steps:

[0008] Raw acoustic signals from industrial rotating equipment during operation were collected, and one-dimensional time-series vectors and two-dimensional spectrograms were constructed respectively.

[0009] Based on the two-dimensional spectrogram, the environmental frequency characteristics are perceived, and the two-dimensional spectrogram is calibrated through the environmental frequency characteristics to adaptively correct the acoustic signal changes caused by environmental changes, and output a two-dimensional feature map.

[0010] The two-dimensional feature map is processed through a cascaded network to extract bearing fault features in a hierarchical manner from shallow to deep. Within each unit, asymmetric convolution is used to capture time-frequency morphological features, and a coordinate attention mechanism is introduced to generate a saliency weight mask. The saliency feature regions are weighted to obtain a higher-order two-dimensional feature map. The step-by-step processing of the cascaded network realizes the hierarchical stripping and purification of complex acoustic features, effectively enhancing the ability to distinguish fault features in industrial noise environments.

[0011] A one-dimensional auxiliary branch is constructed to extract the temporal features of the one-dimensional time-series vector and map them as channel weight coefficients. The channel weight coefficients are used as soft guides to weight the higher-order two-dimensional feature map, so as to perform channel-level gain adjustment on the two-dimensional spatial features and obtain the final two-dimensional feature map. Thus, the adaptive modulation of the two-dimensional higher-order features by the one-dimensional time-domain auxiliary information effectively improves the model's ability to capture fault features in complex industrial backgrounds.

[0012] The statistical prior parameters of the one-dimensional time-series vector are extracted, and adaptive energy modulation is performed on the final two-dimensional feature map to perceive the energy distribution of the original acoustic signal. Then, the bearing fault detection results are classified and output according to the modulated final two-dimensional feature map, thereby effectively combining deep semantic features with the original statistical prior and further improving the fault diagnosis accuracy of the model under different energy level samples.

[0013] Furthermore, the original acoustic signal is discretized and sampled, and the resulting original discrete sound sequence is used as the one-dimensional time-series vector. The one-dimensional time-series vector is then subjected to a short-time Fourier transform to map it from the pure time domain space to the time-frequency domain space, thereby obtaining a two-dimensional spectrum that characterizes the dynamic changes of energy with time and frequency.

[0014] Furthermore, the environmental frequency feature perception is to perform a global average pooling operation along the frequency dimension of the two-dimensional spectrogram, and extract the frequency distribution envelope that reflects the current input frequency distribution characteristics by performing an arithmetic average on each row of the feature matrix.

[0015] The frequency distribution envelope is used to generate a calibration weight mask through a mapping network. The calibration weight mask is then multiplied element-wise with the two-dimensional spectrogram to obtain the reweighted features. In this process, by dynamically reshaping the spectral energy distribution, the model can adaptively correct frequency response distortion caused by different acquisition environments.

[0016] The reweighted features are reconstructed using time-frequency dual-axis coordinate features to correct feature changes caused by spatial reverberation. The reconstructed time-dimensional features are added to the frequency-dimensional features, spatial information is associated, and then added to the reweighted features to obtain the superimposed reconstructed features.

[0017] The superimposed reconstructed features are subjected to strength normalization to eliminate the absolute amplitude differences between different sampled samples, resulting in a two-dimensional feature map.

[0018] In existing technologies, the input end generally performs feature extraction operations directly on the spectrogram, with the default acquisition environment being absolutely fixed. It does not take into account the changes in features caused by different environments and lacks adaptive capability. Therefore, this invention inserts this step before performing deep extraction of fault features to solve the adaptation problem caused by different acquisition environments. It realizes feature calibration of the original sound signal at the acoustic channel level, effectively eliminates the problem of inconsistent feature distribution caused by different acquisition environments, and ensures that the model has adaptive capability in different acquisition environments.

[0019] Furthermore, a cascaded network consisting of multiple spatially cascaded processing units is constructed. As the number of levels increases, the spatial resolution (i.e., width) of the feature map also increases. and height The channel dimensions representing feature richness decrease in a stepwise manner, and the channel dimensions decrease stepwise. This design aims to transform the original spectral distribution into a high-order two-dimensional feature map containing high-order semantic information by condensing spatial information layer by layer.

[0020] Furthermore, the spatial cascaded processing unit includes a spatial feature processing block, a spatial downsampling fusion layer, and a coordinate attention enhancement unit;

[0021] The spatial feature processing block captures time-frequency morphological features and performs dual receptive field refinement on the time-frequency morphological features to obtain refined spatial features.

[0022] The spatial downsampling fusion layer performs spatial dimension feature aggregation on the refined spatial features.

[0023] The coordinate attention enhancement unit performs global average pooling operations along the time and frequency dimensions simultaneously on the aggregated features. After mapping and activation operations, a saliency weight mask is generated. The saliency weight mask is multiplied with the principal element of the aggregated features to weight the saliency regions associated with faults, resulting in a high-order two-dimensional feature map.

[0024] Most existing sound diagnostic models use ordinary symmetrical square convolutions (such as the common 3×3 convolution) for feature extraction. This method extracts all information indiscriminately. However, this invention, in addition to feature extraction, also refines the features. Within each spatially cascaded processing unit, asymmetric convolutions (1×7 and 7×1) are used to capture features in the time and frequency domains respectively. These two convolutions can better extract the corresponding features, rather than mixing time and frequency domain features together, resulting in more detailed feature extraction. Then, dual receptive field convolutions are used to refine details and identify large-area background noise. This step not only allows the model to learn fault features but also to understand what the background noise is like, which can more accurately identify faults. Existing models often only focus on learning fault features. Finally, this invention combines a coordinate attention mechanism, which can more accurately locate significant fault areas, but existing models often lack similar modules.

[0025] Furthermore, the spatial feature processing module includes an asymmetric feature extraction unit and a frequency refinement unit;

[0026] The asymmetric feature extraction unit captures temporal morphological features using parallel horizontal asymmetric convolution kernels and captures frequency morphological features using vertical asymmetric convolution kernels. The temporal and frequency morphological features are then concatenated along the channel dimension. After batch normalization and activation operations, the time-frequency morphological features are obtained, thereby achieving time-frequency decoupling of complex acoustic signals.

[0027] The frequency refinement unit divides the time-frequency morphological features into two parts along the channel dimension using channel segmentation, and inputs them into parallel standard convolution and dilated convolution respectively. This process uses dual receptive field modeling, with standard convolution capturing voiceprint details and dilated convolution extracting environmental background distribution. The outputs of standard convolution and dilated convolution are concatenated, and after batch normalization and activation operations, spatial refined features are obtained.

[0028] Furthermore, the one-dimensional auxiliary branch utilizes a multi-level cascaded temporal downsampling convolutional layer to perform temporal feature mining on the one-dimensional temporal vector;

[0029] Then, global average pooling is performed on the deep temporal features to compress and aggregate the information of the features in the time axis dimension, and channel weight coefficients matching the number of channels are generated through activation operations, which respectively represent the significance gain of the corresponding channel in the time domain dimension.

[0030] In order to enable the weights of the one-dimensional vector form to act on the two-dimensional spatial features, the present invention performs a spatial dimension expansion operation on the channel weight coefficients to align them with the higher-order two-dimensional feature map in terms of spatial size, thereby obtaining the temporal modulation weights.

[0031] The final two-dimensional feature map is obtained by performing residual modulation fusion using the temporal modulation weights and the higher-order two-dimensional feature map.

[0032] Existing acoustic detection models often use only two-dimensional spectrograms as a single input, which easily leads to the loss of micro-temporal features in the sound signal. Therefore, this invention uses fine-grained time information in one-dimensional time-series waveforms as soft guidance, enabling the model to have a more sensitive feature perception ability before the final classification decision.

[0033] Furthermore, the multi-level cascaded temporal downsampling convolutional layer includes three cascaded one-dimensional convolutional layers, which perform stepwise length compression and channel mapping on the one-dimensional temporal vector. Each one-dimensional temporal downsampling convolutional layer is followed by a one-dimensional batch normalization layer to stabilize the feature distribution.

[0034] The first-level convolutional layer performs preliminary feature extraction; the output features of the second-level convolutional layer are filtered using a gated activation mechanism to enhance significant impact beats in the time-domain signal and suppress background noise; the third-level one-dimensional convolutional layer performs high-order mapping on the filtered features to obtain deep temporal features.

[0035] Furthermore, global statistics are obtained from the one-dimensional time-series vector to generate statistical prior parameters, which reflect the basic energy state of the original sound signal at the time of acquisition.

[0036] The statistical prior parameters are processed by fully connected layers and nonlinear activation layers to learn scaling factors and offset factors respectively. Adaptive energy modulation is then performed on the final two-dimensional feature map to obtain energy modulation features. In this way, the energy distribution characteristics of the original signal are injected into the deep features as supplementary information through affine transformation, thereby achieving enhanced perception of sample features.

[0037] Global average pooling is performed on the energy modulation features, and the extracted global saliency feature vector is input into the classifier to obtain the probability distribution of each fault category, so as to select the final bearing fault detection result.

[0038] This invention utilizes the final gain of the original physical statistics information on the deep features to enhance the model's perception of the energy distribution of the input signal before classification decision, thereby outputting the final bearing fault detection result.

[0039] A non-contact bearing fault detection system for industrial rotating equipment includes a feature extraction unit, a global perception calibration unit, a cascaded network unit, a one-dimensional auxiliary branch unit, and an adaptive energy modulation and fault detection unit, which sequentially execute the non-contact bearing fault detection method for industrial rotating equipment.

[0040] The advantages and beneficial effects of this invention are as follows:

[0041] This invention possesses adaptive acquisition environment capabilities. By sensing the frequency distribution characteristics of the current input features through a global perception calibration unit and generating dynamic weights, it achieves adaptive correction for differences in different acoustic transmission channels. This effectively mitigates signal feature distribution shifts caused by variations in non-contact acquisition distance and spatial characteristics, ensuring the model's adaptive detection capability under different operating environments. Furthermore, this invention enhances fault feature discrimination capabilities under background noise. By utilizing asymmetric convolution to capture time-frequency morphological features and combining it with a coordinate attention mechanism to lock salient feature regions, it effectively weakens interference from background noise such as metal impacts and airflow pulsations, enabling the model to maintain accuracy even in industrial noise environments. This invention effectively isolates significant fault features, resolving the false alarm problem caused by feature confusion. It achieves deep coupling of multi-dimensional features and energy distribution perception. By capturing the microscopic temporal pulse patterns in the original waveform through a one-dimensional auxiliary branch, residual modulation is performed on the two-dimensional main features. This mechanism, combining temporal information with two-dimensional features, takes into account both the energy evolution and microscopic phase details of the fault signal. Furthermore, by incorporating adaptive energy modulation and introducing statistical priors of the one-dimensional temporal vector, the model's ability to perceive the energy distribution of the original acoustic signal is enhanced. This multi-dimensional feature enhancement method compensates for the loss of scale information in conventional processing, effectively improving the accuracy and reliability of bearing detection. Attached Figure Description

[0042] Figure 1 This is a flowchart of the method in an embodiment of the present invention.

[0043] Figure 2 This is a flowchart of constructing one-dimensional and two-dimensional inputs in an embodiment of the present invention.

[0044] Figure 3 This is a schematic diagram of the network structure of the global perception calibration unit in an embodiment of the present invention.

[0045] Figure 4 This is an overall architecture diagram of the cascaded network in an embodiment of the present invention.

[0046] Figure 5This is a schematic diagram of the internal structure of the spatially cascaded processing unit in an embodiment of the present invention.

[0047] Figure 6 This is a schematic diagram of a one-dimensional auxiliary branch structure in an embodiment of the present invention.

[0048] Figure 7 This is a flowchart of adaptive energy modulation and fault detection in an embodiment of the present invention. Detailed Implementation

[0049] The specific embodiments of the present invention will be described in detail below with reference to the accompanying drawings. It should be understood that the specific embodiments described herein are for illustration and explanation only and are not intended to limit the present invention.

[0050] like Figure 1 As shown, this invention proposes a non-contact bearing fault detection method for industrial rotating equipment, specifically including the following steps:

[0051] Step S1: Acquire the raw acoustic signals during the operation of the industrial rotating equipment. Based on the raw acoustic signals, simultaneously construct a one-dimensional time-series vector and a two-dimensional spectrogram, which serve as the one-dimensional and two-dimensional inputs to the subsequent fault detection model, respectively. Figure 2 As shown, the specific process is as follows:

[0052] First, raw acoustic signals from the operation of the industrial rotating equipment are acquired using a non-contact acoustic sensor. In this embodiment, the raw acoustic signals are discretized to obtain a digitized raw discrete sound sequence, which is then used as a one-dimensional time-series vector, denoted as... One-dimensional time vectors are directly used as one-dimensional inputs to the model to ensure that the information and characteristics of the signal in the time domain are fully preserved.

[0053] Simultaneously, to construct the two-dimensional input of the model, this embodiment copies the one-dimensional time-series vector to obtain a one-dimensional time-series backup vector. Subsequently, a Short-Time Fourier Transform (STFT) is performed on the one-dimensional time-series backup vector, mapping it from the pure time domain to the time-frequency domain. The calculation process of the STFT is expressed as follows:

[0054]

[0055] in, It is a one-dimensional time-series backup vector. For window functions, Indicates time and frequency The signal components. In this embodiment, when performing the short-time Fourier transform, the window function... A Hamming window was used. To ensure both frequency resolution and positioning accuracy in the time dimension, the window size was set to 2048, and the sliding step size between adjacent windows was set to 512. The above transformation yielded a two-dimensional spectrum representing the dynamic changes of energy with time and frequency, denoted as... And use it as a two-dimensional input to the model.

[0056] Thus, through step S1, the synchronous construction of one-dimensional and two-dimensional inputs based on the same original acoustic signal source was achieved, providing the model with one-dimensional and two-dimensional inputs respectively.

[0057] Step S2: Input the two-dimensional spectrogram into the global sensing calibration unit, extract the global time-frequency distribution characteristics of the current input, reconstruct the acoustic features, and adaptively correct for changes in the sound signal caused by changes in the acquisition environment, thereby obtaining a two-dimensional feature map, such as... Figure 3 As shown, the specific process is as follows:

[0058] First, the environmental frequency characteristic sensing process is performed, and the two-dimensional spectrum is generated. Input the cell and perform a global average pooling operation along the frequency dimension. Specifically, this operation extracts the frequency distribution envelope reflecting the frequency distribution characteristics of the current input by taking the arithmetic mean of each row of the feature matrix. The specific mathematical calculation process is expressed as follows:

[0059]

[0060] in, Represents a two-dimensional spectrum The width of the time dimension and These represent the row index and column index of the matrix, respectively.

[0061] After completing frequency feature perception, the calibration weight mask generation and feature reweighting processes are performed. The extracted frequency distribution envelope is then input into a mapping network consisting of 1×1 convolutional layers and nonlinear activation functions. In the middle, the Sigmoid activation function is used. Generate a calibration weight mask with values ​​between 0 and 1. Then, the weight mask will be labeled. Two-dimensional spectrum of the input Perform element-wise multiplication to obtain the reweighted features. The mathematical formula corresponding to the above process is:

[0062]

[0063]

[0064] In this process, by dynamically reshaping the spectral energy distribution, the model can adaptively correct frequency response distortions caused by different acquisition environments.

[0065] Subsequently, the reweighted features The coordinate features of the time-frequency dual axes are reconstructed. To correct feature variations caused by spatial reverberation, this process utilizes a strip pooling operator to simultaneously extract time-dimensional and frequency-dimensional features, the calculation methods of which are expressed as follows:

[0066]

[0067]

[0068] in, For reweighted features Frequency dimension height, Its time dimension width, This represents the global average pooling operation performed along the time dimension. This represents a global average pooling operation performed along the frequency dimension.

[0069] Next, the sum of the bidirectional pooling features is fed into a 3×3 convolutional layer. Spatial information correlation modeling is performed, and the output results are linked with reweighted features via residual connections. The features are added together to obtain the superimposed reconstructed features.

[0070] Finally, perform environment alignment and normalization. Input the overlay reconstructed features into the instance normalization layer and perform instance normalization. As a specific implementation method, instance normalization operation The specific mathematical calculation process is expressed as follows:

[0071]

[0072] in, The input is the feature to be processed in this operation. and They represent Mean and variance within the corresponding channel A preset constant is used to prevent the denominator from being zero. This operation effectively eliminates the absolute amplitude differences between different sampled samples. Therefore, the overall feature flow structure of step S2 can be expressed as follows:

[0073]

[0074] in, This represents the output two-dimensional feature map. After the above processing steps, the model outputs a two-dimensional feature map. This provides environmentally robust input for subsequent cascaded feature extraction.

[0075] Thus, step S2 completes the feature calibration of the original sound signal at the acoustic channel level. This effectively eliminates the problem of inconsistent feature distribution caused by different acquisition environments, ensuring that the model has adaptive capabilities under different acquisition environments.

[0076] Step S3, calibrate the two-dimensional feature map The input is a cascaded network to achieve hierarchical extraction of bearing fault features from shallow to deep. Within each unit, asymmetric convolution is used to capture time-frequency morphological features, and a coordinate attention mechanism is introduced to generate saliency weight masks. These masks are then used to weight salient feature regions, resulting in a high-order two-dimensional feature map, such as... Figure 4 As shown.

[0077] The cascaded network consists of three levels of spatial cascaded processing units (denoted as the first-level unit, the second-level unit, and the third-level unit, respectively) connected in series. In particular, the first-level unit, the second-level unit, and the third-level unit all adopt the same internal topology. Each level of spatial cascaded processing unit contains, in sequence, a spatial feature processing block, a spatial downsampling fusion layer, and a coordinate attention enhancement unit.

[0078] During feature transfer, as the number of processing levels increases... The increase in the spatial resolution (i.e., width) of the feature map and height The spectral density decreases in a stepwise manner, while the channel dimension, representing feature richness, increases progressively. This design aims to transform the original spectral distribution into a high-order two-dimensional feature map containing high-order semantic information by progressively condensing spatial information. .

[0079] Specifically, the overall characteristic flow structure of a cascaded network can be expressed by the following set of formulas:

[0080]

[0081]

[0082]

[0083] in, Indicates the first Feature maps output by cascaded spatial processing units Indicates input up to the number Feature maps of cascaded processing units in a spatial hierarchy. Specifically, This represents the two-dimensional feature map generated in step S2 received by the first stage of the cascaded network. , Represents the final stage output of the cascaded network and serves as a high-order two-dimensional feature map. .also, , , They represent the first The spatial feature processing block, spatial downsampling fusion layer, and coordinate attention enhancement unit are located within the cascaded spatial processing unit.

[0084] Through this design, features undergo morphological construction, spatial compression, and significantly enhanced processing within each unit level, ultimately resulting in an output... It has strong fault identification capabilities.

[0085] Features undergo processing within each level of the spatial cascaded processing unit. To further clarify its specific internal implementation logic, this embodiment focuses on the first... The internal structure of the cascaded processing unit is explained, such as... Figure 5 As shown, the specific process is as follows:

[0086] First, execute the spatial feature processing block. Time-frequency morphological feature capture. Input feature map. Upon entering this processing block, the signal first passes through an asymmetric feature extraction unit. This unit utilizes parallel 1×7 horizontal asymmetric convolution kernels to capture temporal morphological features and 7×1 vertical asymmetric convolution kernels to capture frequency-domain morphological features, thereby achieving time-frequency decoupling of complex acoustic signals. Its mathematical extraction process is expressed as follows:

[0087]

[0088] in, and These represent asymmetric convolution operations of 1×7 horizontal and 7×1 vertical, respectively. This indicates a batch normalization operation. This indicates that the concatenation is performed along the channel dimension, resulting in... It represents time-frequency morphological characteristics.

[0089] Secondly, the time-frequency morphological characteristics are refined using the indexed frequency refinement unit. Perform dual receptive field refinement. This involves refining time-frequency morphological features. Input frequency refining unit, using channel splitting operation to... Divided into two parts along the channel dimension (denoted as respectively) and The data are then fed into parallel standard convolutions and dilated convolutions, respectively. This process uses dual receptive fields for modeling, with standard convolutions capturing voiceprint details and dilated convolutions extracting environmental background distributions. The specific refinement process can be expressed as follows:

[0090]

[0091] in, This indicates a standard convolution with a 3×3 kernel. This represents a 3×3 dilated convolution with a dilation rate of 2, and the final output is... This is represented as a spatial refinement feature.

[0092] Next, a spatial downsampling fusion layer is used. Perform spatial dimension feature aggregation. Refine the spatial features. Input spatial downsampling fusion layer. It achieves halving of the spatial size and doubling of the channel dimension through a convolution operation with a stride of 2, aiming to expand the receptive field and filter out spatially redundant information. Its flow structure is expressed as:

[0093]

[0094] in, This represents a 2×2 convolution operation with a stride of 2. This indicates a batch normalization operation. This represents the activation function and its output. This is the feature map after downsampling.

[0095] Finally, through coordinate attention enhancement units Perform a significant weighting. Input coordinate attention enhancement unit, synchronously execute global average pooling operation along the time dimension. Global average pooling operation performed along the frequency dimension After mapping and activation, a saliency weight mask is generated, which is used to weight the saliency regions associated with faults. The specific mathematical calculation process is expressed as follows:

[0096]

[0097] in, This is the output feature map of the spatial cascaded processing unit at this level. This represents a 1×1 convolution operation. This represents the Sigmoid activation function. This indicates element-wise multiplication.

[0098] Thus, the first The internal structure of the cascaded processing unit has been explained. Through the step-by-step processing of the cascaded network in step S3, the model achieves hierarchical stripping and purification of complex acoustic features, effectively enhancing its ability to distinguish fault characteristics in industrial noise environments.

[0099] Step S4: Extract one-dimensional time-series vector features using a one-dimensional auxiliary branch and map them to channel weight coefficients. Then, weight the higher-order two-dimensional feature map using residual modulation to obtain the final two-dimensional feature map, as shown below. Figure 6 As shown, the specific process is as follows:

[0100] First, the one-dimensional time series vector obtained in step S1 The input is fed into a one-dimensional auxiliary branch. This branch utilizes multiple cascaded temporal downsampling convolutional layers... Temporal feature mining is performed. The multi-level cascaded temporal downsampling convolutional layer consists of three cascaded one-dimensional convolutional layers, which are used to perform stepwise length compression and channel mapping on the one-dimensional temporal vector. Each one-dimensional convolutional layer is followed by a one-dimensional batch normalization layer to stabilize the feature distribution.

[0101] First, preliminary feature extraction is performed by the first-level convolutional layer. Then, a gated activation mechanism is applied to the output features of the intermediate (second-level) one-dimensional convolutional layer to non-linearly filter significant temporal features. The specific mathematical expression is as follows:

[0102]

[0103] in, This represents the intermediate features output by the intermediate-level one-dimensional convolutional layer. This represents the Sigmoid activation function. This represents the features after gating. This mechanism, through adaptive gain adjustment, can enhance significant impulse beats in the time-domain signal and suppress background noise. Next, the filtered features... The third-level one-dimensional convolutional layer performs higher-order mapping and outputs deep temporal features.

[0104] Subsequently, the deep temporal features are processed by global average pooling. The information of features along the time axis is compressed and aggregated, and channel weight coefficients matching the number of channels are generated by the Sigmoid activation function. Its expression is as follows:

[0105]

[0106] in, This refers to the deep temporal features output by the third-level one-dimensional convolutional layer. The generated channel weight coefficients... Each value represents the significance gain of the corresponding channel in the time domain dimension.

[0107] Next, in order to enable the weights of the one-dimensional vector form to act on the features in two-dimensional space, this embodiment adjusts the channel weight coefficients. Perform a spatial dimension expansion operation to make it spatially comparable to the higher-order two-dimensional feature map. Alignment, thereby obtaining the timing modulation weights. Its mathematical expression process is as follows:

[0108]

[0109] Finally, the generated temporal modulation weights are used With higher-order two-dimensional feature maps Perform residual modulation fusion to obtain the final two-dimensional feature map. Its expression is:

[0110]

[0111] In this process, the temporal features extracted by the one-dimensional auxiliary branch serve as soft guidance for channel-level gain adjustment of the two-dimensional spatial features. Thus, through step S4, adaptive modulation of two-dimensional high-order features using one-dimensional temporal auxiliary information is achieved, effectively improving the model's ability to capture fault features in complex industrial contexts.

[0112] Step S5 involves extracting the statistical prior parameters of the one-dimensional time-series vector, performing adaptive energy modulation on the final two-dimensional feature map to perceive the energy distribution of the original acoustic signal, and finally outputting the bearing fault detection result through a classifier. Figure 7 As shown, the specific execution process is as follows:

[0113] First, the one-dimensional time series vector obtained from step S1 Calculate global statistics, including the time-domain mean. With time domain standard deviation The parameters are combined and used as statistical prior parameters. These parameters reflect the fundamental energy state of the original sound signal at the time of acquisition.

[0114] Subsequently, the statistical prior parameters are input into the mapping unit. This mapping unit learns the scaling factor through processing by fully connected layers and nonlinear activation layers, respectively. With offset factor The mapping process is expressed as follows:

[0115]

[0116] in, To statistically analyze prior parameters, This represents the mapping function consisting of fully connected layers and nonlinear activation layers. The generated scaling factor is then used. With offset factor The final two-dimensional feature map produced in step S4 Perform adaptive energy modulation to obtain energy modulation characteristics. The specific mathematical expression process is as follows:

[0117]

[0118] in, This represents channel-by-channel multiplication. In this process, affine transformation is used to inject the energy distribution characteristics of the original signal as supplementary information into the deep features, thereby enhancing the perception of sample features.

[0119] Next, the obtained energy modulation characteristics were analyzed. Perform global average pooling to extract the globally salient feature vector. This vector is then input into a classifier containing fully connected layers and a Softmax layer. The classifier uses the Softmax function to calculate the probability distribution of each fault category; its calculation expression is as follows:

[0120]

[0121] in, The first output of the fully connected layer in the classifier The original scores for each category, The total number of categories, This indicates that the model predicts the sample belongs to the first... The model outputs the probability distribution for each fault category, and then selects the category with the highest probability value as the final bearing fault detection result.

[0122] Furthermore, during model training, the cross-entropy loss function is used to calculate the difference between the predicted results and the true labels, and the network parameters are updated accordingly. The expression for the cross-entropy loss function is:

[0123]

[0124] in, This represents the true fault category label corresponding to the sample. Thus, through step S5, the final gain of the deep features using the original physical statistics information is achieved, enhancing the model's perception of the input signal energy distribution before classification decision, thereby outputting the final bearing fault detection result. This step effectively combines deep semantic features with the original statistical prior, further improving the model's fault diagnosis accuracy under samples of different energy levels.

[0125] In summary, the non-contact bearing fault detection method for industrial rotating equipment provided by this invention effectively solves the problems of sound signal distortion and energy scale loss caused by variable sensor distance and spatial characteristic differences in non-contact monitoring through pre-processing global perception calibration, significantly improving the model's adaptability to fluctuations in the acquisition environment. Utilizing a cascaded asymmetric feature extraction architecture and coordinate attention enhancement mechanism, it achieves accurate capture of transient morphological features of bearing faults, effectively overcoming the feature confusion problem caused by strong industrial background noise interference. Furthermore, by using the temporal features extracted by the one-dimensional auxiliary branch to perform residual modulation on the two-dimensional path features, it achieves information complementarity between the time and frequency domains, further enhancing the model's perception depth of weak fault features under complex industrial conditions. This scheme significantly improves the accuracy, reliability, and environmental robustness of bearing fault detection in industrial rotating equipment while maintaining efficient feature extraction.

[0126] This invention also provides a non-contact bearing fault detection system for industrial rotating equipment, including a feature extraction unit, a global perception calibration unit, a cascaded network unit, a one-dimensional auxiliary branch unit, and an adaptive energy modulation and fault detection unit, which sequentially execute the non-contact bearing fault detection method for industrial rotating equipment. The specific implementation method is the same as the above method, and will not be repeated here.

[0127] The above embodiments are only used to illustrate the technical solutions of the present invention, and are not intended to limit it. 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.

Claims

1. A non-contact bearing fault detection method for industrial rotating equipment, characterized in that: Raw acoustic signals from industrial rotating equipment during operation were collected, and one-dimensional time-series vectors and two-dimensional spectrograms were constructed respectively. Based on the two-dimensional spectrogram, the ambient frequency characteristics are perceived, and the two-dimensional spectrogram is calibrated through the ambient frequency characteristics to adaptively correct changes in acoustic signals caused by environmental changes, and output a two-dimensional feature map. The bearing fault features of the two-dimensional feature map are extracted in a hierarchical manner from shallow to deep. At each level, time-frequency morphological features are captured, and a coordinate attention mechanism is introduced to generate a saliency weight mask. The saliency feature regions are weighted to obtain a higher-order two-dimensional feature map. A one-dimensional auxiliary branch is constructed to extract the temporal features of the one-dimensional time-series vector and map them as channel weight coefficients. The channel weight coefficients are used as soft guides to weight the higher-order two-dimensional feature map, so as to adjust the channel-level gain of the two-dimensional spatial features and obtain the final two-dimensional feature map. The statistical prior parameters of the one-dimensional time-series vector are extracted, and adaptive energy modulation is performed on the final two-dimensional feature map to perceive the energy distribution of the original acoustic signal. Then, the bearing fault detection result is output according to the modulation of the final two-dimensional feature map.

2. The non-contact bearing fault detection method for industrial rotating equipment according to claim 1, characterized in that: The original acoustic signal is discretized and sampled, and the resulting original discrete sound sequence is used as the one-dimensional time-series vector. The one-dimensional time-series vector is then subjected to a short-time Fourier transform to map it from the pure time domain space to the time-frequency domain space, resulting in a two-dimensional spectrum.

3. The non-contact bearing fault detection method for industrial rotating equipment according to claim 1, characterized in that: The environmental frequency feature perception is achieved by performing a global average pooling operation along the frequency dimension of the two-dimensional spectrum. By performing an arithmetic average on each row of the feature matrix, a frequency distribution envelope reflecting the current input frequency distribution characteristics is extracted. The frequency distribution envelope is used to generate a calibration weight mask through a mapping network, and then the calibration weight mask is multiplied element-wise with the two-dimensional spectrogram to obtain the reweighted features. The reweighted features are reconstructed using time-frequency dual-axis coordinate features. The reconstructed time dimension features are added to the frequency dimension features, and then spatial information is associated. Finally, the reconstructed features are added to the reweighted features to obtain the superimposed reconstructed features. The superimposed and reconstructed features are subjected to strength normalization to obtain a two-dimensional feature map.

4. The non-contact bearing fault detection method for industrial rotating equipment according to claim 1, characterized in that: A cascaded network consisting of multiple spatially cascaded processing units is constructed. As the number of levels increases, the spatial resolution of the feature map decreases stepwise, while the channel dimension representing feature richness increases stepwise.

5. The non-contact bearing fault detection method for industrial rotating equipment according to claim 4, characterized in that: The spatial cascaded processing unit includes a spatial feature processing block, a spatial downsampling fusion layer, and a coordinate attention enhancement unit. The spatial feature processing block captures time-frequency morphological features and performs dual receptive field refinement on the time-frequency morphological features to obtain refined spatial features. The spatial downsampling fusion layer performs spatial dimension feature aggregation on the refined spatial features. The coordinate attention enhancement unit performs global average pooling operations along the time and frequency dimensions simultaneously on the aggregated features. After mapping and activation operations, a saliency weight mask is generated. The saliency weight mask is multiplied with the principal element of the aggregated features to weight the saliency regions associated with faults, resulting in a high-order two-dimensional feature map.

6. The non-contact bearing fault detection method for industrial rotating equipment according to claim 5, characterized in that: The spatial feature processing module includes an asymmetric feature extraction unit and a frequency refinement unit. The asymmetric feature extraction unit captures temporal morphological features using parallel horizontal asymmetric convolution kernels and captures frequency morphological features using vertical asymmetric convolution kernels. The temporal and frequency morphological features are then concatenated along the channel dimension and subjected to batch normalization and activation operations to obtain the time-frequency morphological features. The frequency refinement unit uses channel segmentation to divide the time-frequency morphological features into two parts along the channel dimension, and inputs them into parallel standard convolution and dilated convolution respectively. The standard convolution captures the voiceprint details, and the dilated convolution extracts the environmental background distribution. The outputs of the standard convolution and dilated convolution are concatenated, and after batch normalization and activation operations, the spatial refined features are obtained.

7. The non-contact bearing fault detection method for industrial rotating equipment according to claim 1, characterized in that: The one-dimensional auxiliary branch utilizes multi-level cascaded temporal downsampling convolutional layers to mine temporal features from the one-dimensional temporal vector; Then, global average pooling is performed on the deep temporal features to compress and aggregate the information of the features in the time axis dimension, and channel weight coefficients matching the number of channels are generated through activation operations, which respectively represent the significance gain of the corresponding channel in the time domain dimension. A spatial dimension expansion operation is performed on the channel weight coefficients to align them with the higher-order two-dimensional feature map in terms of spatial size, thereby obtaining the temporal modulation weights; The final two-dimensional feature map is obtained by performing residual modulation fusion using the temporal modulation weights and the higher-order two-dimensional feature map.

8. The non-contact bearing fault detection method for industrial rotating equipment according to claim 7, characterized in that: The multi-level cascaded temporal downsampling convolutional layer includes three cascaded one-dimensional convolutional layers, which perform progressive length compression and channel mapping on the one-dimensional temporal vector. Each one-dimensional temporal downsampling convolutional layer is followed by a one-dimensional batch normalization layer. The first-level convolutional layer performs preliminary feature extraction; The output features of the second-level convolutional layer are filtered using a gated activation mechanism to enhance significant impulse beats in the time-domain signal and suppress background noise. The third-level one-dimensional convolutional layer performs high-order mapping on the filtered features to obtain deep temporal features.

9. The non-contact bearing fault detection method for industrial rotating equipment according to claim 1, characterized in that: The global statistics in the one-dimensional time-series vector are obtained and used to generate statistical prior parameters to reflect the basic energy state of the original sound signal at the time of acquisition. The statistical prior parameters are processed by fully connected layers and nonlinear activation layers to learn scaling factors and offset factors respectively. Adaptive energy modulation is then performed on the final two-dimensional feature map to obtain energy modulation features. Global average pooling is performed on the energy modulation features, and the extracted global saliency feature vector is input into the classifier to obtain the probability distribution of each fault category, so as to select the final bearing fault detection result.

10. A non-contact bearing fault detection system for industrial rotating equipment, comprising a feature extraction unit, a global perception calibration unit, a cascaded network unit, a one-dimensional auxiliary branch unit, and an adaptive energy modulation and fault detection unit, characterized in that: The non-contact bearing fault detection method for industrial rotating equipment as described in any one of claims 1 to 9 shall be executed sequentially.

Images