Rotary machinery fault diagnosis method and device based on physical information enhancement and medium

Abstract

The invention provides a rotary machine fault diagnosis method and device based on physical information enhancement, and a medium. The method comprises the following steps: collecting an original vibration signal of a rotary machine; generating a fault characteristic frequency based on the original vibration signal of the rotating machine; based on the fault characteristic frequency, converting the original vibration signal of the rotating machine into a physically enhanced characteristic signal; and generating a fault probability from the physically enhanced feature signal by using a probability deep neural network model, and generating intelligent data so as to visually explain physical information. Based on the physical information enhancement technology, the accuracy and reliability of fault diagnosis are improved.

Description

Technical Field

[0001] The present disclosure relates to a method for rotating machinery fault diagnosis based on physical information enhancement, and in particular to a method, device and medium for rotating machinery fault diagnosis based on physical information enhancement. Background Art

[0002] The statements in this section merely provide background information related to the present disclosure and do not necessarily constitute prior art.

[0003] As industrial equipment continues to increase in complexity and automation, the operating status of rotating machinery, a core component of industrial machinery, directly impacts the safety, reliability, and operating costs of the entire system. Rotating machinery is widely used in various industrial systems, such as power generation, petrochemicals, aerospace, and manufacturing. In these applications, rotating machinery often requires long periods of high-intensity operation, making its core components susceptible to fatigue and wear, which can lead to failure. Consequently, failures can lead to significant economic losses if timely detection and early warning are not provided. Therefore, developing an efficient and reliable fault diagnosis method that can monitor and analyze the operating status of rotating machinery in real time is crucial to ensuring the safe operation of mechanical equipment.

[0004] In the field of vibration signal analysis, time-frequency analysis technology has been widely used in practice as an effective method for processing non-stationary signals. Common time-frequency analysis methods include short-time Fourier transform, wavelet transform, and empirical mode decomposition. Among them, CWT, due to its multi-scale analysis capabilities, can flexibly analyze the local characteristics of signals at different time scales and is widely used in rotating machinery fault diagnosis based on physical information enhancement. CWT can simultaneously provide time and frequency information of the signal, thereby enabling accurate analysis of non-stationary signals, which is crucial for complex machinery fault diagnosis.

[0005] With the advancement of industrial modernization, rotating machinery (such as engines, turbines, and centrifuges) is increasingly used in industrial production. These rotating machines often require high-intensity continuous operation, which can lead to mechanical fatigue, structural failures, and safety hazards, resulting in expensive maintenance costs. Therefore, improving the reliability and operational safety of rotating machinery and reducing operating and maintenance costs have become key issues that need to be addressed in modern industry. To this end, it is particularly important to develop accurate and reliable monitoring and assessment frameworks to predict and manage the health status of mechanical components.

[0006] Currently, the application of artificial intelligence (AI) technology in industrial machinery condition monitoring, particularly in machinery fault diagnosis, is attracting significant attention. The application of AI technology is primarily focused on prognostics and health management (PHM) processes, aiming to optimize maintenance strategies for machinery systems. The effectiveness and reliability of PHM systems directly impact the stability and safety of fault diagnosis methods. Therefore, ensuring the accuracy and robustness of fault diagnosis methods is a key requirement for the effective operation of PHM systems. Effective fault diagnosis methods not only improve operator safety but also significantly reduce operational and maintenance costs.

[0007] With the rise of deep learning (DL) as a key branch of artificial intelligence, it has been widely used in rotating machinery fault diagnosis. Deep learning models can significantly reduce the risks of relying on manual feature selection in traditional machine learning methods by automatically learning data features. In particular, deep learning technology, supported by deep neural network models such as convolutional neural networks (CNNs), recurrent neural networks (RNNs), and attention networks (ANs), has demonstrated strong capabilities in the application of mechanical fault diagnosis. Currently, deep learning models can effectively adapt to changing operating conditions, process small sample data, and address imbalanced datasets, demonstrating great potential in mechanical fault diagnosis.

[0008] While AI technology has made significant progress in mechanical fault diagnosis, several limitations remain that hinder the widespread adoption of large-scale AI models. First, the dynamic data of mechanical systems differs significantly from data types like natural language. This makes it difficult to directly transfer many AI models successfully applied to fields like natural language processing to mechanical fault diagnosis.

[0009] To address the above challenges, the present disclosure urgently needs a new fault diagnosis method for rotating machinery, which can significantly improve the accuracy and robustness of fault diagnosis. Summary of the Invention

[0010] In view of the above-mentioned deficiencies in the prior art, the purpose of the present disclosure is to provide a rotating machinery fault diagnosis method based on physical information enhancement, so as to improve the accuracy, real-time performance and robustness of rotating machinery fault diagnosis.

[0011] To achieve the above objectives, the present disclosure provides a rotating machinery fault diagnosis method based on physical information enhancement, comprising the following steps:

[0012] Collect the original vibration signal of rotating machinery;

[0013] generating a fault characteristic frequency based on the original vibration signal of the rotating machinery;

[0014] Based on the fault characteristic frequency, converting the original vibration signal of the rotating machinery into a physically enhanced characteristic signal;

[0015] Use probabilistic deep neural network models to generate fault probabilities from physically enhanced feature signals and generate intelligent data to visually interpret physical information.

[0016] To achieve the above objectives, the present disclosure further provides a rotating machinery fault diagnosis device based on physical information enhancement, comprising a memory, a processor, and a computing program, wherein the memory is connected to the processor:

[0017] a memory configured to store the computing program;

[0018] a processor configured to run the computing program;

[0019] When the computing program is executed by the processor, the rotating machinery fault diagnosis method based on physical information enhancement is executed.

[0020] In order to achieve the above-mentioned objectives, the present disclosure also provides a non-transitory computer-readable storage medium, which is configured to store computer-readable instructions, wherein when the computer-readable instructions are executed by a processor, the rotating machinery fault diagnosis method based on physical information enhancement is executed.

[0021] Compared with the prior art, the present disclosure has the following beneficial technical effects:

[0022] In response to the above problems, the present disclosure combines a probabilistic deep network with physical information to effectively extract and utilize the physical characteristic frequencies in the mechanical system, thereby significantly improving the accuracy of fault diagnosis. By introducing physical information, the present disclosure makes the output characteristics of the model closely related to the actual physical characteristics of the mechanical system, which not only improves the diagnostic performance of the model, but also enhances the interpretability of the model, making the diagnostic results more credible. The present disclosure also proposes a new probabilistic deep neural network model with high credibility and strong explanatory power. By combining physical information with deep learning, the probabilistic deep neural network model can not only provide accurate fault diagnosis results, but also generate intelligent data and clearly express fault mode information, greatly improving the interpretability and credibility of fault diagnosis, and laying a solid foundation for the construction of large-scale AI models. BRIEF DESCRIPTION OF THE DRAWINGS

[0023] The accompanying drawings, which constitute a part of the present disclosure, are used to provide a further understanding of the present disclosure. The exemplary embodiments of the present disclosure and their descriptions are used to explain the present disclosure and do not constitute an improper limitation to the present disclosure.

[0024] Figure 1 A flowchart of a rotating machinery fault diagnosis method based on physical information enhancement is provided in accordance with a specific embodiment of the present disclosure;

[0025] Figure 2 A schematic diagram of a rotating machinery fault diagnosis device based on physical information enhancement is provided for a specific embodiment of the present disclosure. DETAILED DESCRIPTION

[0026] The following will be combined with the accompanying drawings in the embodiments of the present disclosure to clearly and completely describe the technical solutions in the embodiments of the present disclosure. Obviously, the embodiments described are only part of the embodiments of the present disclosure, not all of the embodiments. Based on the embodiments of the present disclosure, all other embodiments obtained by ordinary technicians in this field without making any creative efforts are within the scope of protection of the present disclosure.

[0027] It should be noted that the following detailed description is illustrative and is intended to provide further explanation of the present disclosure. Unless otherwise specified, all technical and scientific terms used in this embodiment have the same meaning as commonly understood by those skilled in the art to which the present disclosure belongs.

[0028] It should be noted that the terms used herein are only for describing specific embodiments and are not intended to limit the exemplary embodiments according to the present disclosure. As used herein, unless the context clearly indicates otherwise, the singular form is also intended to include the plural form. In addition, it should be understood that when the terms include and / or include are used in this specification, they indicate the presence of features, steps, operations, devices, components and / or combinations thereof.

[0029] Figure 1 A flowchart of a rotating machinery fault diagnosis method based on physical information enhancement is provided for a specific embodiment of the present disclosure.

[0030] like Figure 1 As shown, in a specific embodiment of the present disclosure, a rotating machinery fault diagnosis method based on physical information enhancement is provided, comprising the following steps:

[0031] Collect the original vibration signal of rotating machinery;

[0032] generating a fault characteristic frequency based on the original vibration signal of the rotating machinery;

[0033] Based on the fault characteristic frequency, converting the original vibration signal of the rotating machinery into a physically enhanced characteristic signal;

[0034] A probabilistic deep neural network (PIPDN) model is used to generate fault probabilities from physically enhanced feature signals and to generate intelligent data to visually interpret physical information.

[0035] The present disclosure is based on physical information enhancement technology, which improves the accuracy and reliability of fault diagnosis.

[0036] In a specific embodiment of the present disclosure, in the rotating machinery fault diagnosis method based on physical information enhancement, the step of collecting the original vibration signal of the rotating machinery further includes the following steps:

[0037] Data acquisition step S100: Acquire the raw time-domain vibration signal of the rotating machinery; collect vibration signal data from the mechanical system using an ICP accelerometer (623C01) with a sensor sensitivity of 100 mV / g and a frequency response range of 0.5 Hz to 10 kHz. The acquisition device is mounted radially on the bearing seat of the rotating machinery and secured with a magnetic base to ensure the reliability and stability of the raw vibration signal acquisition. The data sampling frequency of the acquisition device is 42,000 Hz, with a sampling duration of 10 seconds per group to ensure the frequency resolution and time-domain information integrity of the signal analysis.

[0038] Data preprocessing step S200: The original vibration signal X0(t) is preprocessed, including bandpass filtering and wavelet transform, to remove noise and enhance the time-frequency characteristics of the signal, and obtain the preprocessed time domain signal X(t); the original vibration signal is bandpass filtered using a 6th-order Butterworth filter to remove noise; the passband range of the filter is 10Hz to 1,000Hz. The design of the filter is based on the analysis of the typical fault frequency band (bearing characteristic frequency interval) and the actual working condition noise frequency band. Then, a wavelet transform with the Db4 wavelet as the wavelet basis is used to perform a 5-layer decomposition to further process the signal. The Db4 wavelet basis is selected because its symmetry and tight support are suitable for capturing transient fault signals. The wavelet scale of the wavelet transform is selected to be 1 to 8. The wavelet coefficients of the wavelet transform are used to remove noise, retain signal feature details, and enhance fault information in the signal.

[0039] The transfer function of the Butterworth filter is:

[0040]

[0041] Where: w c is the cutoff frequency; n is the filter order.

[0042] Frequency domain conversion step S300: Based on the pre-processed time domain signal, perform a fast Fourier transform (FFT) to obtain a frequency domain signal; convert the processed time domain signal X(t) into a frequency domain signal X(f) through the fast Fourier transform to extract the fault characteristic frequency. The formula is as follows:

[0043] X(f)=FFT(x(t)).

[0044] In a specific embodiment of the present disclosure, in the rotating machinery fault diagnosis method based on physical information enhancement, the step of collecting the original vibration signal of the rotating machinery further includes the following steps:

[0045] Calculate the normalized frequency f of the original vibration signal norm , which is used to standardize the frequency characteristics at different speeds.

[0046] The normalized frequency formula is:

[0047] Where: f is the frequency component; f r is the rotation frequency. The normalization operation makes the frequency characteristics at different rotation speeds comparable.

[0048] The signal preprocessing step further includes normalizing the vibration signal to reduce the impact of different sensors and environmental conditions on signal analysis.

[0049] In a specific embodiment of the present disclosure, in the rotating machinery fault diagnosis method based on physical information enhancement, the step of generating a fault characteristic frequency based on the original vibration signal of the rotating machinery includes the following steps:

[0050] Characteristic frequency calculation step S350: Calculate the fault characteristic frequency based on the structural parameters of the rotating machinery to extract frequency information related to the physical characteristics of the rotating machinery.

[0051] The fault characteristic frequencies include: inner race fault frequency BPFI, outer race fault frequency BPFO and rolling element fault frequency BSF;

[0052] The formula for calculating the fault characteristic frequency based on the structural parameters of the rotating machinery is:

[0053] Inner race fault frequency

[0054] Outer race fault frequency

[0055] Rolling element failure frequency

[0056] The structural parameters of the rotating machine include: the number of rolling elements Z, the rotation frequency of the rolling elements f r The rolling element diameter d, the pitch circle diameter D, and the contact angle θ are also included. The pitch circle is the circular trajectory of the rolling element node center in the motion plane. The structural parameters of the rotating machine are information about the rotating machine and can be obtained through measurement and calculation.

[0057] These formulas can accurately calculate the fault characteristic frequency that is closely related to the physical structure of the rotating machinery, providing key physical information for subsequent fault diagnosis.

[0058] In a specific embodiment of the present disclosure, in the rotating machinery fault diagnosis method based on physical information enhancement, the step of converting the original vibration signal of the rotating machinery into a physically enhanced characteristic signal based on the fault characteristic frequency includes the following steps:

[0059] S400: Based on the fault characteristic frequency, extract the harmonic components related to the fault characteristic frequency from the preprocessed frequency domain signal X(f); that is, select the fault characteristic frequency and its first five harmonic components and perform signal enhancement processing; identify important harmonics related to the fault frequency based on an amplitude threshold (e.g., greater than 5% of the maximum amplitude); and generate a frequency domain characteristic signal X(f) enhanced with physical information. physics (f);

[0060] S430: Introduce the frequency domain attention mechanism. The core of the self-attention mechanism is to automatically assign weights by calculating the correlation between different components in the input signal. Specifically, it can identify which frequency components are more important for fault diagnosis and assign higher weights to these frequency components. Through an attention layer, the frequency domain feature signal X enhanced by physical information is calculated. physics The attention weight A(f) of each frequency component f in (f) is used to enhance the focus on the key frequency band; wherein the frequency domain feature signal X enhanced by physical information is physics (f) After linear transformation through the weight matrix W and the bias term b, the Softmax function is applied to the linear transformation output to perform normalization processing and generate the attention weight A(f);

[0061] The formula is A(f)=Softmax(W·X physics (f)+b).

[0062] The calculation of the attention weight A(f) is realized by the fully connected layer, which uses the fully connected layer to compress the dimension of the feature vector of each scale in the multi-scale feature to generate the initial weight value. W and b are learnable parameters, W is the weight matrix, and b is the bias term; the frequency domain feature signal X enhanced by physical information is processed by the weight matrix W and the bias term b. physics (f) Perform linear transformation. Construct an attention layer and use the learnable parameters W and b to analyze the frequency domain feature signal X enhanced by physical information. physics (f) is linearly transformed and normalized by the Softmax function to obtain the attention weight A(f). The weight matrix W and the bias term b are optimized during the training process through the back propagation algorithm in the deep learning model training process.

[0063] The Softmax function maps the output of the linear transformation into a probability distribution. It uses exponential operations to enhance the differences between values, making larger values ​​dominate the probability distribution. The Softmax function is used to normalize the weights and map the output of the linear transformation into a probability distribution.

[0064] The Softmax function converts multiple real-valued vectors into a probability distribution. The Softmax function enhances the differences between values ​​through exponential operations, making larger values ​​dominate the probability distribution while suppressing the influence of smaller values. The input value can be positive, negative, zero, or a number greater than 1, but the Softmax function converts the input value into a probability value between 0 and 1. The Softmax function maps the output of the linear transformation into a probability distribution, satisfying ∑A(f) = 1.

[0065] S460: Using the attention weight A(f) to analyze the frequency domain feature signal X enhanced by physical information physics (f) Weighted to obtain the attention-weighted frequency domain feature signal X based on physical information enhancement att (f), to highlight the fault characteristics of the critical frequency band;

[0066] The formula is X att (f) = A(f)·X physics (f).

[0067] By frequency-wise multiplication, the attention weight A(f) is multiplied by the frequency domain feature signal X enhanced by physical information. physics (f) Multiply frequency by frequency to obtain the attention-weighted frequency domain feature signal X based on physical information enhancement att (f) The frequency domain feature signal X that is enhanced based on physical information and weighted by attention att (f) Input to the subsequent fault diagnosis model as input for subsequent processing to highlight the characteristic signals of the key frequency bands.

[0068] In a specific embodiment, taking bearing fault detection as an example, BPFI is set as the characteristic frequency of the bearing inner ring fault. After extracting its harmonic components, the fault-sensitive frequency band (such as around 3BPFI) is enhanced through the attention mechanism, ultimately improving the fault classification accuracy.

[0069] S480: The frequency domain feature signal X enhanced based on physical information and weighted by attention att (f) Perform inverse Fourier transform (IFFT) to obtain the time domain characteristic signal x after physical information enhancement physics (t);

[0070] S500: Using an autoencoder, from the time domain feature signal x enhanced by physical information physics (t), extract the amplitude information a after physical information enhancement physics (f) = |X att (f)| and phase information enhanced by physical information

[0071] In the frequency domain characteristic signal generation step S400: based on the fault characteristic frequency, the harmonic components related to the fault characteristic frequency are extracted and retained in the preprocessed frequency domain signal X(f), and the harmonic components unrelated to the fault characteristic frequency are marked as zero, so as to generate a frequency domain characteristic signal X(f) enhanced by physical information. physics (f) Enhance the physical information of the fault characteristic frequency and its related harmonic components in the signal, reduce noise interference, and improve the ability to identify fault characteristics.

[0072] In the time domain signal restoration step S480: the frequency domain feature signal X enhanced based on physical information and weighted by attention is att (f) Perform inverse Fourier transform (IFFT) to obtain the time domain characteristic signal x after physical information enhancement physics (t). Steps S400 and S480 are used to enhance the physical information and generate a time domain characteristic signal x after physical information enhancement. physics (t).

[0073] x physics (t) = IFFT(X physics (f))

[0074] In the step S500 of extracting amplitude information and phase information: the autoencoder includes a multi-layer convolutional neural network, including a convolution layer, a pooling layer and an activation function. The autoencoder is used to extract the time domain feature signal x from the physical information enhanced time domain feature signal x. physics (t), extract the amplitude information α enhanced by physical information physics (f) = |X att (f)| and phase information enhanced by physical information

[0075] In a specific embodiment of the present disclosure, in the rotating machinery fault diagnosis method based on physical information enhancement, the step of using a probabilistic deep neural network (PIPDN) model to generate fault probabilities from physically enhanced characteristic signals and generating intelligent data to visually interpret the physical information includes the following steps:

[0076] High-level feature extraction step S550: Using the first independent encoder Ψ1 (Encoder1), the phase information enhanced by the physical information is extracted from the Extract phase high-level features And use the second independent encoder Ψ2 (Encoder2) to obtain the amplitude information α enhanced by the physical information physics (f) = |X att In (f), the amplitude high-level feature Ψ2(α)=Ecoder2(α) is extracted.

[0077] The encoder structure consists of three convolutional layers, with the following configuration:

[0078] First convolutional layer: convolution kernel size: 5×5; step size: 2; activation function: ReLU; number of output channels: 32.

[0079] Second convolutional layer: convolution kernel size: 4×4; stride: 2; activation function: ReLU; number of output channels: 64.

[0080] The third convolutional layer: convolution kernel size: 3×3; step size: 2; activation function: ReLU; number of output channels: 128.

[0081] From the physically enhanced amplitude and phase information, high-level features are extracted to generate a latent space feature representation. This further mines and enhances the physical feature information in the signal, providing rich physical feature input for subsequent fault diagnosis models. This step reconverts frequency domain features into time domain signals, preserving critical fault information.

[0082] In a specific embodiment of the present disclosure, in the rotating machinery fault diagnosis method based on physical information enhancement, the step of using a probabilistic deep neural network (PIPDN) model to generate fault probabilities from physically enhanced characteristic signals and generating intelligent data to visually interpret the physical information further includes the following steps:

[0083] Latent space feature extraction step S600: extracting the phase high-level features Perform inverse Fourier transform (IFFT) on the amplitude high-level feature Ψ2(α) to obtain the latent space feature Based on the amplitude high-level features and phase high-level features, the latent space feature representation is generated to further enhance the physical feature information of the signal.

[0084] This process achieves a joint time-frequency representation of features by reducing frequency-domain features to time-domain features, significantly improving the expressive power of the latent space. The latent space features z(t) contain signal features enhanced with physical information, including both the amplitude and phase information of the vibration signal and physical features strongly correlated with the fault mode.

[0085] In a specific embodiment of the present disclosure, in the rotating machinery fault diagnosis method based on physical information enhancement, the step of using a probabilistic deep neural network (PIPDN) model to generate fault probabilities from physically enhanced characteristic signals and generating intelligent data to visually interpret the physical information further includes the following steps:

[0086] S650: Use the first independent decoder (Decoder1), reconstructs the latent space feature z(t) into a phase output feature Use a second independent decoder (Decoder2), reconstructs the latent space feature z(t) into an amplitude output feature The decoder structure is similar to the encoder above, except that the convolutional layer is replaced by a deconvolutional layer. This step ensures that the reconstructed phase output features and the reconstructed amplitude output characteristics Can accurately reflect the physical information in the input signal.

[0087] The structure of the decoder is similar to that of the encoder, except that the convolutional layers are replaced by deconvolutional layers.

[0088] The first deconvolution layer: convolution kernel size: 5×5; step size: 2; activation function: ReLU; number of output channels: 64.

[0089] Second deconvolution layer: convolution kernel size: 4×4; step size: 2; activation function: ReLU; number of output channels: 32.

[0090] The third deconvolution layer: convolution kernel size: 3×3; step size: 2; activation function: Sigmoid; number of output channels: 1.

[0091] The decoder gradually reconstructs the latent space features Z(t) into output features through fully connected layers and deconvolution layers.

[0092] The first layer is a fully connected layer that linearly transforms Z(t) into an intermediate feature representation. The second layer is a deconvolution operation that gradually amplifies the feature dimensions and reconstructs the signal.

[0093] Reconstructing latent space features into phase output features and amplitude output characteristics

[0094] In a specific embodiment of the present disclosure, in the rotating machinery fault diagnosis method based on physical information enhancement, the step of using a probabilistic deep neural network (PIPDN) model to generate fault probabilities from physically enhanced characteristic signals and generating intelligent data to visually interpret the physical information further includes the following steps:

[0095] PIPDN model optimization step S700: Based on the physical information-enhanced and attention-weighted frequency domain feature signal X att (f) Amplitude output characteristics with the reconstruction The Euclidean distance between them, and the phase information enhanced by physical information With the reconstructed phase output characteristic The Euclidean distance between them is used to optimize the probabilistic deep neural network (PIPDN) model;

[0096] By minimizing this loss function, the output features of the Probabilistic Deep Neural Network (PIPDN) model can be made as close as possible to the physical labels, improving the interpretability of the PIPDN model. Through physical annotation, the PIPDN model is reconstructed.

[0097] The loss function is defined as the Euclidean distance between the output feature and the physical label. The following formula is the loss function:

[0098]

[0099] In a specific embodiment of the present disclosure, in the rotating machinery fault diagnosis method based on physical information enhancement, the step of using a probabilistic deep neural network (PIPDN) model to generate fault probabilities from physically enhanced characteristic signals and generating intelligent data to visually interpret the physical information further includes the following steps:

[0100] Diagnosis and interpretation step S800: Based on the optimized probabilistic deep neural network (PIPDN) model, the time domain characteristic signal x enhanced by physical information is obtained. physics (t), generate failure probability and generate intelligent data Explain physical information visually.

[0101] The final fault diagnosis result includes an assessment of the health status of each component in the mechanical system and provides maintenance recommendations.

[0102] The intelligent data output includes explanations of key decision points in the diagnostic process and generates corresponding reports to assist in troubleshooting.

[0103] According to a specific embodiment of the present disclosure, the rotating machinery fault diagnosis method based on physical information enhancement further includes the following steps:

[0104] A multiscale extractor is used to capture multiscale features from the raw vibration signal and extract fault information at different scales. The goal of multiscale feature extraction is to capture signal characteristics at different time scales. Rotating machinery fault signals typically contain multiple frequency components, and these frequency components may be distributed at different time scales. By performing multiscale analysis on rotating machinery fault signals, it is possible to more comprehensively capture the fault characteristics of rotating machinery.

[0105] The present disclosure utilizes a multi-scale extractor to capture fault characteristics in different frequency ranges, and through multi-scale information fusion, realizes comprehensive analysis and diagnosis of multi-level fault characteristics of mechanical systems.

[0106] To capture the characteristic performance of different fault modes in multiple frequency bands, the decoder outputs the reconstructed amplitude output features. and the reconstructed phase output characteristics It is further processed and spectral analyzed by multi-scale feature extraction method:

[0107] Multi-scalefeatures=AveragePool(x(t),τ)

[0108] Where: τ is the pooling kernel size, representing the multi-scale factor.

[0109] When using a multi-scale extractor to capture fault information at different scales, a multi-scale attention mechanism is introduced to automatically determine the importance of each scale and calculate the weight of each scale. The specific implementation is as follows:

[0110] The original vibration signal is averaged and pooled using pooling kernels of different sizes. The pooling kernel size can be 2, 4, or 8. When the pooling kernel size is 2, high-frequency features (such as impact signals) on a shorter time scale are captured; when the pooling kernel size is 8, low-frequency features (such as periodic fault signals) on a longer time scale are captured. The original vibration signal is averaged and pooled using pooling kernels of different sizes to extract multi-scale features F. scale , to achieve a comprehensive analysis of the multi-level fault characteristics of rotating machinery.

[0111] A scale =Softmax(W scale ·F scale +b scale )

[0112] Among them, the weight matrix W scale and the bias term b scale is a learnable parameter, and the Softmax function ensures weight normalization. A fully connected layer is used to compress the feature vectors of each scale in the multi-scale features to generate initial weight values. The initial weight values ​​are then normalized using the Softmax function to generate the final attention weights.

[0113] Use the attention weight A calculated by the above formula scale For multi-scale features F scale Perform weighting to obtain the weighted multi-scale feature F att .

[0114] F att =A scale ·F scale

[0115] The weighted multi-scale feature F att As the input of subsequent processing, it can highlight the scale features that contribute most to fault diagnosis.

[0116] By extracting features at different scales, we can capture fault characteristics within different frequency ranges and improve the ability to identify complex fault modes. Multi-scale analysis is performed, extracting multi-band features of the signal through average pooling operations at different scales. The weighted features of channels at different scales are then fed into the corresponding probabilistic deep neural network (PIPDN) model to achieve the fusion and enhancement of physical information at different scales.

[0117] According to a specific embodiment of the present disclosure, the rotating machinery fault diagnosis method based on physical information enhancement further includes the following steps:

[0118] Uncertainty quantification: For each scale of the multi-scale features, a probabilistic deep neural network (PIPDN) model is used to predict faults. The diagnostic results at each scale are calculated, along with the uncertainty of the diagnostic results. This uncertainty quantification step provides a basis for evaluating the diagnostic credibility of each PIPDN model in the subsequent decision fusion step, thereby helping to select the more reliable model prediction results for fusion.

[0119] Fault diagnosis models based on traditional deep neural networks still have shortcomings in terms of confidence assessment of model predictions. Because the models typically perform point estimates, they cannot provide confidence assessments of the prediction results. This lack of uncertainty assessment can seriously affect the reliability and effectiveness of intelligent diagnostic models.

[0120] To make up for this shortcoming, current research has gradually realized that it is becoming increasingly important to introduce an uncertainty quantification (UQ) mechanism for deep learning models to provide a range estimate of fault modes rather than a point estimate. The probabilistic neural network (PNN) framework based on Bayesian Neural Networks (BNN) provides richer fault information than traditional point estimates through uncertainty prediction. The purpose of decision fusion is to integrate the prediction results of multiple models and select the most credible result as the final output. In this disclosure, decision fusion is performed based on uncertainty quantification, and the weights are dynamically adjusted by evaluating the credibility of the prediction results of each model.

[0121] Specifically, uncertainty quantification (UQ) can quantify and evaluate two types of uncertainty: aleatoric uncertainty (AU) and epistemic uncertainty (EU). Aleatoric uncertainty (AU) refers to the randomness caused by noise in the data itself, representing the noise or randomness of the input data itself. If the input data is noisy, the aleatoric uncertainty of the input data is high. Epistemic uncertainty (EU) refers to the model's confidence in its predictions. If the model has low confidence in a prediction result, the epistemic uncertainty of the prediction result is high. This uncertainty quantification information can effectively enhance the interpretability and robustness of fault diagnosis models.

[0122] However, although existing probabilistic neural network-based models have been successful in providing interpretability for deep learning fault diagnosis, they still face the challenge that the underlying mechanisms of fault diagnosis are difficult to explain through the model. In addition, the fault diagnosis capabilities of traditional purely data-driven neural network models are limited due to the lack of physical information related to mechanical system faults. To make up for this shortcoming, researchers have begun to explore various methods of embedding physical information into neural networks. For example, some methods enhance the model's understanding of mechanical faults by inputting fault feature information and its harmonic components as additional features into the network. Although these methods help improve the model's ability to identify faults, they usually rely on manual extraction of physical features, lack sufficient flexibility, and are not conducive to model maintenance and updating.

[0123] This disclosure utilizes a decision fusion module based on uncertainty quantification to effectively reduce the false alarm rate during fault diagnosis, thereby improving the overall performance and safety of the diagnostic system. To address common fault diagnosis issues in mechanical systems, a rotating machinery fault diagnosis method based on physical information enhancement is proposed. This method combines physical feature information with deep learning techniques, significantly improving the accuracy and reliability of fault diagnosis through multi-scale analysis and uncertainty quantification.

[0124] This paper adopts uncertainty quantification technology to select the most credible fault diagnosis results by analyzing the epistemic uncertainty and inherent uncertainty of the diagnosis results, thereby further improving the robustness and reliability of the model in complex working conditions and noisy environments.

[0125] The total uncertainty at each scale is the sum of epistemic uncertainty EU and aleatoric uncertainty AU. The formula for calculating the total uncertainty is:

[0126] Total Uncertainty=EU+AU

[0127] Where Total Uncertainty is the total uncertainty, EU is the epistemic uncertainty calculated using the Monte Carlo Dropout method, and AU is the aleatory uncertainty calculated from the variance of the input data. Epistemic uncertainty EU represents the uncertainty of the model parameters, while aleatory uncertainty AU represents the uncertainty of the input data itself.

[0128] The results of multi-scale feature extraction are input into multiple models for prediction, and the uncertainty of each model's prediction results is quantified to assess the diagnostic credibility of each model. This process is carried out through the uncertainty quantification (UQ) method.

[0129] Uncertainty quantification steps: Use the Monte Carlo Dropout method to quantify the uncertainty of the output results of each model and calculate the epistemic uncertainty EU of the model; calculate the accidental uncertainty AU of the input data from the variance of the input data.

[0130] Epistemic uncertainty EU represents the confidence of the model in its prediction results. Aleatoric uncertainty AU reflects the impact of the noise or randomness of the input data itself on the model prediction. This uncertainty AU is usually represented by the variance of the data. The formula is as follows:

[0131]

[0132] Where: T is the number of Monte Carlo Dropout sampling; is the prediction result of the t-th sampling; is the mean of all sampling results; For accidental uncertainty.

[0133] By superimposing epistemic uncertainty EU and aleatoric uncertainty AU, the total uncertainty TotalUncertainty is calculated:

[0134] Total Uncertainty=EU+AU

[0135] By calculating the total uncertainty for each model, we can quantify how confident the model is in its predictions.

[0136] According to a specific embodiment of the present disclosure, the rotating machinery fault diagnosis method based on physical information enhancement further includes the following steps:

[0137] Through the decision fusion module based on uncertainty quantification, the most credible diagnosis result is selected as the final fault identification output.

[0138] Fault diagnosis results generated through decision fusion: Based on the uncertainty quantification results of each model, the diagnostic credibility of each model is evaluated to achieve a quantitative assessment of the physical information enhancement effect. The model results with the smallest epistemic uncertainty EU are preferentially selected for fusion to improve the accuracy and reliability of fault diagnosis.

[0139] Based on the uncertainty quantification results of each of the aforementioned models, the Uncertainty Quantification (UQ) decision fusion module performs a weighted fusion of the prediction results from multiple models. The criterion for selecting a reliable diagnosis result is the minimum epistemic uncertainty, ensuring the highest confidence in the final fault diagnosis. By selecting the result with the minimum epistemic uncertainty, the accuracy and confidence of the final diagnosis are ensured, improving the accuracy and robustness of the final diagnosis, reducing the false alarm rate, and enhancing the overall performance and safety of the fault diagnosis system.

[0140] The prediction results of each model Assign a weight w according to its uncertainty quantification result i The weight calculation formula is:

[0141] Of which: TotalUncertanty i =EU i +AU i is the total uncertainty of the i-th model; ∈ is the regularization term.

[0142] Use the calculated weight w i Prediction results for all models Perform weighted summation to obtain the final diagnosis result

[0143]

[0144] Where: n is the total number of models; is the prediction result of the i-th model; w i is the weight of the ith model.

[0145] The final fault diagnosis result is generated through multi-scale feature fusion and uncertainty weighted fusion Generates intelligent diagnostic reports. The reports include the fault type, the probability of each fault mode, characteristic performance in different frequency bands, and comparative analysis with physical information, demonstrating the impact of physical characteristics such as amplitude, frequency, and phase on the fault mode.

[0146] Table 1 below shows the fault dataset grouping.

[0147] Table 2 below shows the training hyperparameter settings.

[0148] Table 3 below shows the diagnostic accuracy of the model.

[0149] Table 1 Fault dataset grouping

[0150]

[0151]

[0152] Table 2 Training hyperparameter settings

[0153]

[0154] Table 3 Model diagnosis accuracy

[0155]

[0156] The diagnostic accuracy of MS-PIPDN in all fault modes is higher than that of traditional DCNN and SVM methods. Especially in the diagnosis of rolling element faults, MS-PIPDN shows extremely high accuracy (98.70%).

[0157] Compared with PIPDN, MS-PIPDN further improves the fault diagnosis accuracy by introducing multi-scale feature extraction and uncertainty quantification modules, especially in the cases of noise robustness and equipment aging.

[0158] The present disclosure is particularly suitable for the diagnosis of complex fault modes in mechanical systems, and compared with traditional fault diagnosis methods, it can significantly improve the accuracy and reliability of diagnosis. The proposed method analyzes the physical feature information in the vibration signal and combines multi-scale feature extraction and uncertainty quantification to perform fault diagnosis. When a fault occurs, different fault modes will show different feature changes in multi-scale signal analysis. Therefore, the method of the present disclosure can effectively monitor and identify the feature changes of the fault signal. Since it only involves the processing and analysis of vibration signals, the method does not require additional hardware resources or system parameters, and does not increase additional hardware costs and computational complexity. The method does not rely on the certainty of system parameters and therefore has higher robustness. Compared with other methods based on deep learning and big data analysis, the present disclosure adopts physical information enhancement and multi-scale analysis, which has the advantages of efficient computing performance, low data requirements, simple calculation process and low computing cost.

[0159] In a specific embodiment of the present disclosure, the present disclosure further provides a rotating machinery fault diagnosis device based on physical information enhancement, comprising a memory, a processor, and a computing program, wherein the memory is connected to the processor:

[0160] a memory configured to store the computing program;

[0161] a processor configured to run the computing program;

[0162] When the computing program is executed by the processor, the rotating machinery fault diagnosis method based on physical information enhancement is executed.

[0163] Please refer to Figure 2 , a schematic diagram of a rotating machinery fault diagnosis device based on physical information enhancement, provided in accordance with a specific embodiment of the present disclosure. The rotating machinery fault diagnosis device based on physical information enhancement is preferably a computing device comprising a memory, a processor, and a computer program stored in the memory and executable by the processor. When the processor executes the computer program, the rotating machinery fault diagnosis method based on physical information enhancement is implemented. The computer program is stored in a program code space within the memory.

[0164] In a specific embodiment of the present disclosure, the present disclosure also provides a non-transitory computer-readable storage medium, which is configured to store computer-readable instructions, wherein when the computer-readable instructions are executed by a processor, the rotating machinery fault diagnosis method based on physical information enhancement is executed.

[0165] The computer-readable storage medium comprises a storage unit for program code, wherein the storage unit is provided with a program for executing the method steps according to the present invention, and the program is executed by a processor.

[0166] The various techniques described herein may be implemented in conjunction with hardware or software, or a combination thereof. Thus, the methods and apparatus of the present invention, or certain aspects or portions of the methods and apparatus of the present invention, may be implemented in the form of program codes (i.e., instructions) embedded in a tangible medium, such as a removable hard disk, a USB flash drive, a floppy disk, a CD-ROM, or any other machine-readable storage medium, wherein when the program is loaded into a machine such as a computer and executed by the machine, the machine becomes an apparatus for practicing the present invention.

[0167] When the program code is executed on a programmable computer, the electronic device generally includes a processor, a storage medium readable by the processor (including volatile and non-volatile memory and / or storage elements), at least one input device, and at least one output device. The memory is configured to store the program code; the processor is configured to execute the target tracking method of the present invention according to the instructions in the program code stored in the memory.

[0168] By way of example and not limitation, readable media include readable storage media and communication media. Readable storage media store information such as computer-readable instructions, data structures, program modules, or other data. Communication media typically embody computer-readable instructions, data structures, program modules, or other data in a modulated data signal such as a carrier wave or other transport mechanism, and include any information delivery medium. Combinations of any of the above are also included within the scope of readable media.

[0169] Throughout this specification, references to terms such as "one embodiment," "example," or "specific example" indicate that a specific feature, structure, material, or characteristic described in conjunction with that embodiment or example is included in at least one embodiment or example of the present disclosure. In this specification, schematic representations of these terms do not necessarily refer to the same embodiment or example. Furthermore, the specific features, structures, materials, or characteristics described may be combined in any suitable manner in any one or more embodiments or examples.

[0170] The above shows and describes the basic principles, main features and advantages of the present disclosure. Those skilled in the art should understand that the present disclosure is not limited to the above embodiments. The above embodiments and descriptions are merely illustrative of the principles of the present disclosure. Various changes and improvements may be made to the present disclosure without departing from the spirit and scope of the present disclosure, and such changes and improvements fall within the scope of the present disclosure as claimed.

Claims

1. A rotating machinery fault diagnosis method based on physical information enhancement, characterized in that: The following steps are involved: Collect the original vibration signal of rotating machinery; generating a fault characteristic frequency based on the original vibration signal of the rotating machinery; Based on the fault characteristic frequency, converting the original vibration signal of the rotating machinery into a physically enhanced characteristic signal; Use probabilistic deep neural network models to generate fault probabilities from physically enhanced feature signals and generate intelligent data to visually interpret physical information.

2. The method according to claim 1, characterized in that The step of collecting the original vibration signal of the rotating machinery includes the following steps: S100: collecting data on the original vibration signal X0(t) of the rotating machinery; S200: Preprocessing the original vibration signal X0(t) to obtain a preprocessed time domain signal X(t); S300: Performing a fast Fourier transform on the preprocessed time domain signal X(t) to obtain a preprocessed frequency domain signal X(f).

3. The method according to claim 1, characterized in that The step of generating a fault characteristic frequency based on the original vibration signal of the rotating machinery includes the following steps: S350: Calculating a fault characteristic frequency based on structural parameters of the rotating machinery; The fault characteristic frequencies include: inner race fault frequency, outer race fault frequency and rolling element fault frequency; The structural parameters of the rotating machine include: the number of rolling elements Z, the rotation frequency of the rolling elements f r As well as the rolling element diameter d and pitch circle diameter D and contact angle θ.

4. The method according to claim 2, characterized in that The step of converting the original vibration signal of the rotating machinery into a physically enhanced characteristic signal based on the fault characteristic frequency includes the following steps: S400: Based on the fault characteristic frequency, extract the harmonic components related to the fault characteristic frequency from the preprocessed frequency domain signal X(f) to generate a frequency domain characteristic signal X(f) enhanced by physical information. physics (f); S430: The frequency domain characteristic signal X enhanced by the physical information physics (f) After linear transformation through the weight matrix W and the bias term b, the Softmax function is applied to the linear transformation output to perform normalization processing and generate the attention weight A(f); S460: Using the attention weight A(f) to analyze the frequency domain feature signal X enhanced by physical information physics (f) Weighted to obtain the attention-weighted frequency domain feature signal X based on physical information enhancement att (f); S480: The frequency domain feature signal X enhanced based on physical information and weighted by attention att (f) Perform inverse Fourier transform to obtain the time domain characteristic signal x after physical information enhancement physics (f); S500: Using an autoencoder, from the time domain feature signal x enhanced by physical information physics In (f), the amplitude information enhanced by physical information and the phase information enhanced by physical information are extracted.

5. The method according to claim 4, characterized in that The steps of using a probabilistic deep neural network model to generate a fault probability from a physically enhanced characteristic signal and generating intelligent data to visually interpret the physical information include the following steps: S550: Using the first independent encoder Ψ1, extracting phase high-level features from the phase information enhanced by the physical information A second independent encoder Ψ2 is used to extract the amplitude high-level feature Ψ2(α) from the amplitude information enhanced by the physical information.

6. The method according to claim 5, characterized in that The step of using a probabilistic deep neural network model to generate a fault probability from the physically enhanced characteristic signal and generating intelligent data to visually interpret the physical information further includes the following steps: S600: Advanced features of the phase The inverse Fourier transform is performed on the amplitude high-level feature Ψ2(α) to obtain the latent space feature 7. The method according to claim 6, characterized in that The step of using a probabilistic deep neural network model to generate a fault probability from the physically enhanced characteristic signal and generating intelligent data to visually interpret the physical information further includes the following steps: S650: Use the first independent decoder Reconstruct the latent space features into phase output features Use a second independent decoder Reconstruct the latent space feature z(t) into an amplitude output feature 8. The method according to claim 7, characterized in that The step of using a probabilistic deep neural network model to generate a fault probability from the physically enhanced characteristic signal and generating intelligent data to visually interpret the physical information further includes the following steps: S700: Based on the physical information-enhanced and attention-weighted frequency domain feature signal X att (f) Amplitude output characteristics with the reconstruction The Euclidean distance between them, and the phase information enhanced by physical information With the reconstructed phase output characteristic The Euclidean distance between them is used to optimize the probabilistic deep neural network model.

9. The method according to claim 8, characterized in that The step of using a probabilistic deep neural network model to generate a fault probability from the physically enhanced characteristic signal and generating intelligent data to visually interpret the physical information further includes the following steps: S800: Based on the optimized probabilistic deep neural network model, the time domain feature signal x enhanced by the physical information is obtained. physics (t), generate failure probability and generate intelligent data Explain physical information visually.

10. The method according to claim 9, characterized in that The following steps are also included: A multi-scale extractor is used to capture multi-scale features from the original vibration signal and extract fault information at different scales.

11. The method according to claim 10, characterized in that The following steps are also included: For each scale feature in the multi-scale features, an optimized probabilistic deep neural network model is used to predict faults, and the diagnosis results at each scale are calculated, and the uncertainty of the diagnosis results is calculated.

12. The method according to claim 11, characterized in that The following steps are also included: Through the decision fusion module based on uncertainty quantification, the most credible diagnosis result is selected as the final fault identification output.

13. The method according to claim 11, wherein: The formula for calculating the total uncertainty is: Total Uncertainty=EU+AU; Among them, Total Uncertainty is the total uncertainty, EU is the epistemic uncertainty calculated using the Monte Carlo Dropout method; AU is the aleatory uncertainty calculated from the variance of the input data.

14. The method according to claim 13, wherein: The weights w of each optimized probabilistic deep neural network model are calculated based on the total uncertainty and the ∈ regularization term i , Use the calculated weight w i Perform weighted summation on the prediction results of all models to obtain the final fault diagnosis result 15. A rotating machinery fault diagnosis device based on physical information enhancement, characterized in that: The system comprises a memory, a processor and a computing program, wherein the memory is connected to the processor: a memory configured to store the computing program; a processor configured to run the computing program; Wherein, when the computing program is executed by the processor, the rotating machinery fault diagnosis method based on physical information enhancement according to any one of claims 1 to 14 is executed.

16. A non-transitory computer-readable storage medium, characterized in that The non-transitory computer-readable storage medium is configured to store computer-readable instructions, wherein when the computer-readable instructions are executed by a processor, the rotating machinery fault diagnosis method based on physical information enhancement according to any one of claims 1 to 14 is performed.

Images