A method for identifying early failure of a rotating machine

Abstract

This invention belongs to the technical field of fault prediction and health management, and discloses a method for identifying early faults in rotating machinery, including the following steps: (1) collecting signal time series data; (2) processing the obtained time series signal data to convert the time domain signal to the frequency domain and time-frequency domain; (3) constructing an unsupervised learning neural network model described by Bayesian deep support vector data; (4) training the neural network model based on the obtained feature map using a combination of unsupervised and Bayesian backpropagation methods; (5) determining the indicator scale for early fault identification of the neural network model; (6) calculating the distance between the sample of the rotating machinery at each moment and the center of the hypersphere, and then identifying the early fault points of the rotating machinery by comparing it with the upper confidence limit of the hypersphere radius, and indicating the health level information. This invention solves the problems of difficulty in identifying early fault points, low detection accuracy, and unreliable results.

Description

Technical Field

[0001] This invention belongs to the technical field of fault prediction and health management, and more specifically, relates to a method for identifying early faults in rotating machinery. Background Technology

[0002] Failure is both a state and a process. The entire process from failure symptoms to degradation and failure involves multiple states, and the transitions between states are random. Any operating rotating machinery will inevitably experience failure or breakdown as its service life increases. Early failures in rotating machinery can be understood as the signs of impending failure, from which point its function begins to deteriorate slowly or rapidly. The main tasks of early failure identification in rotating machinery include identifying early failure points and indicating the health level information of the rotating machinery throughout its journey from healthy operation to early failure and eventual failure. The early failure point is the starting point of the early failure state and can be defined as the boundary between the healthy stage and the failure degradation stage of the rotating machinery. Detecting early failure points as early as possible can prevent further degradation and failure, serving as an early warning and indication of the rotating machinery's health status. Furthermore, early identification of the rotating machinery entering an early failure state and indicating its health level information facilitates further failure mode identification and remaining service life prediction, contributing to subsequent maintenance decisions and operational optimization.

[0003] Methods for early fault identification in rotating machinery mainly include those based on mechanistic models, statistical and signal analysis, and machine learning. Mechanism-based methods require establishing models based on the mathematical or physical laws governing the degradation process of rotating machinery. However, the mechanisms of most rotating machinery are highly complex, making it difficult to establish effective mechanistic models. Statistical and signal analysis methods combine prior knowledge with signal processing techniques. These methods often require significant expert experience and theoretical foundations, and lack strong generalization capabilities. Furthermore, due to the complex structure of rotating machinery, numerous fault causes, background noise interference in fault feature extraction, and difficulties in analyzing sensor data, early faults exhibit strong randomness, difficulty in capturing weak faults, and severe aliasing of operational data features. These challenges make it difficult to generalize methods based solely on mechanistic modeling or signal analysis to practical production. Machine learning methods are no longer limited to fixed prior knowledge. They can utilize a large amount of time-series sensor signals such as vibration, temperature, and pressure to learn different levels of features by training data models. These methods show good identification results, but often require a large amount of historical data as training samples. When the actual problem is complex and the sample size is small, the random uncertainty of the sample data itself and the uncertainty of model cognition caused by the difficulty in modeling accurately often lead to low accuracy and poor usability of early fault identification results, and there are also certain limitations. Summary of the Invention

[0004] To address the aforementioned deficiencies or improvement needs of existing technologies, this invention provides a method for identifying early faults in rotating machinery. The method first converts the time-series signals of the initial healthy phase of rotating machinery operation, collected by sensors, into frequency domain and time-frequency domain information, and obtains the feature images required for the model. Then, it trains a network model described by a Bayesian deep support vector network using a small number of normal sample feature images. Finally, it uses the trained model to complete high-confidence detection of early fault points and indication of health level information for rotating machinery, solving the problems of difficulty in identifying early fault points, low detection accuracy, and unreliable results.

[0005] To achieve the above objectives, according to one aspect of the present invention, a method for identifying early faults in rotating machinery is provided, the method mainly comprising the following steps:

[0006] (1) Collect signal time series data of a single or multiple similar rotating machines during the initial health phase or throughout their entire life cycle to obtain a data sample set;

[0007] (2) The obtained time-series signal data is processed to convert the time-domain signal to the frequency domain and time-frequency domain, thereby obtaining the frequency domain and time-frequency domain feature maps corresponding to the time-series signal. The feature maps include the single-sided amplitude-frequency feature map and the VMD-Hilbert time-frequency spectrum feature map.

[0008] (3) Construct an unsupervised learning neural network model for describing Bayesian deep support vector data, wrap the obtained normal samples in a hypersphere that is as small as possible, and use the Bayesian neural network structure to obtain uncertainty information of the model and data.

[0009] (4) Based on the obtained feature maps, the neural network model is trained using a combination of unsupervised and Bayesian backpropagation methods;

[0010] (5) Determine the indicator scale for early fault identification of the neural network model. The scale comprehensively represents the uncertainty of the sample being wrapped by the hypersphere after being mapped to the high-dimensional space of the model output.

[0011] (6) Calculate the distance between the sample of the rotating machinery and the center of the hypersphere at each moment, and then identify the early failure point of the rotating machinery by comparing it with the upper confidence limit of the hypersphere radius, and provide health level information indication.

[0012] Furthermore, by analyzing the obtained time-series data samples After performing a fast Fourier transform to obtain a one-dimensional single-sided amplitude spectrum, it is reconstructed into a two-dimensional feature map, thus obtaining the single-sided amplitude-frequency feature map.

[0013] Furthermore, by selecting the number of modesK Variational mode decomposition is performed on the time-series signal to obtain multiple sets of intrinsic mode components. Hilbert transform is then applied to each mode component to obtain the instantaneous frequency that varies with time. and instantaneous amplitude Then, discretize at equal time intervals. and To obtain the signal's time spectrum;

[0014] The time-frequency expression of the signal is:

[0015]

[0016] In the formula, Represents the real part of a complex number; Indicates the first One mode; The modal number.

[0017] Furthermore, the loss function of the neural network model is:

[0018]

[0019] In the formula, This represents the square of the distance between the network output and the center c of the hypersphere; s The number of sampling times for the model distribution layer. k The length of the training set samples; These are non-negative correction hyperparameter terms; It is a variable approximate probability distribution; This represents the true posterior distribution; This represents the Kullback-Leibler divergence operator; For expectation operator, express Expectations; This is normal sample data; The weights are unknown parameters that can be trained in a neural network. For given parameters Post-data The likelihood function.

[0020] Furthermore, the neural network model assumes that all trainable unknown parameters *k* are considered as independent random variables following a Gaussian distribution, and thus infers the corresponding posterior distribution from the data through training; given data and network model structure Unknown parameters The posterior distribution is as follows:

[0021]

[0022] In the formula, It is the likelihood function; This is the prior probability; For marginal likelihood, the prediction process can be derived from the posterior distribution. Sampling parameters Through computational models To obtain the prediction results Expected value:

[0023]

[0024] The Bayes by backprop approximation inference method is used to approximate the posterior distribution, and a set of data is established. For parameter mean, The parameter for standard deviation Variable approximate probability distribution of control .

[0025] Furthermore, through optimization The parameters make the following possible The KL divergence with the true posterior distribution p should be minimized as much as possible:

[0026]

[0027] In the formula, Represents the cost of complexity. Representing the likelihood cost, a local reparameterization operation is performed on , making , express It follows a standard Gaussian distribution; and Let represent the mean of the i-th network parameter. is the standard deviation; to achieve the approximate sampling of random variables from a standard Gaussian distribution to obtain 𝜔.

[0028] Furthermore, for Bayesian convolutional layers, two convolutional kernels are used to make... , ; To experience the wild; As a variance operator, Gaussian distribution is used to perform local reparameter sampling on the feature map to obtain activation values. .

[0029] Furthermore, the FlipOut operation is used to correct the Bayesian variational inference, and the basic perturbation is updated as follows:

[0030]

[0031] In the formula, These are random variables that follow a Rademacher distribution, and each was sampled independently. and The basic perturbation is used. By calculating an unbiased estimate of the loss gradient, quasi-independent weight sampling and backpropagation are achieved, thereby reducing the number of iterations.

[0032] Furthermore, based on the 3σ criterion, the early failure point indication scale is defined as follows:

[0033]

[0034] In the formula, sample points Category A value of 1 indicates a normal sample. A value of 0 indicates a normal sample; For the center of the super ball; for Distance from the center of the supersphere; The model's prediction results; The distance between the predicted result and the center of the supersphere; The prediction results after inputting training data into the model; Let be the radius of the hypersphere, and take... The maximum value of 0; This represents the number of sampling times for the model's distribution layer.

[0035] Furthermore, the health level information of rotating machinery is indicated as follows:

[0036]

[0037] In the formula, This serves as the baseline for health status.

[0038] In summary, compared with the prior art, the method for identifying early faults in rotating machinery provided by the present invention has the following advantages:

[0039] 1. By adopting a neural network model with Bayesian deep support vector data description, which is an unsupervised single-class anomaly detection model, we can well meet the problem of training data without fault labels in actual early fault identification tasks. Bayesian deep support vector data description combines Bayesian variational inference and the method of minimizing the hypersphere construction, which reduces the impact of random uncertainty of data, ensures the accuracy of early fault identification structure, and provides cognitive uncertainty information of the model.

[0040] 2. The dual-domain feature map construction provided in the data preprocessing extracts the feature information of the time series signal from the frequency domain and the time-frequency domain through Fourier transform and VMD-Hilbert time-frequency decomposition. The frequency domain feature map can more intuitively obtain the information hidden in the time series signal. The time-frequency feature map reflects the characteristics of the frequency domain information of the non-stationary signal taken by the rotating machinery changing with time. The VMD-Hilbert method used can effectively suppress the boundary effect and mode mixing phenomenon in the time-frequency decomposition and better reflect the trend of the signal changing with time.

[0041] 3. In training a Bayesian deep support vector data description neural network model, the feature maps of the convolutional layers and the weights in the fully connected layers are changed from fixed parameters to distributed parameters for Bayesian backpropagation. To address the high variance in gradient estimation caused by local reparameter sampling in small sample cases, the FlipOut operation is adopted, which can accelerate the convergence speed of the model and enhance its generalization ability.

[0042] 4. When obtaining the network early fault point identification results and health level information indicator scale described by the final Bayesian deep support vector data, the mean and variance uncertainty information of the hypersphere radius output distribution are comprehensively considered. This achieves high-confidence detection and identification of early fault points and health level information indication, avoids the overall detection problem caused by the detection error of a single sample, and enhances the robustness of early fault identification. Attached Figure Description

[0043] Figure 1 This is a flowchart illustrating the method for identifying early faults in rotating machinery provided by the present invention.

[0044] Figure 2 This is a schematic diagram of the structure of the constructed Bayesian deep support vector data description neural network model;

[0045] Figures 3-1 to 3-6 The images show the early fault identification results and health level information indication diagrams for bearings 1_1, 1_2, 1_4, 1_5, 2_1, and 3_1, respectively. (a) shows the bearing vibration signal data, (b) shows the early fault identification results, and (c) shows the health level information indication.

[0046] Figures 4-1 to 4-4 These are the vibration signal envelope spectra of bearing 1_1 from the 77th to the 80th cycle; Figure 4-5 The envelope spectrum peak value and corresponding frequency diagram of bearing 1_1 from the 77th to the 80th cycle are shown. (a) is the bearing vibration signal data, and (b) is the vibration signal envelope spectrum. Detailed Implementation

[0047] To make the objectives, technical solutions, and advantages of this invention clearer, the invention will be further described in detail below with reference to the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are merely illustrative and not intended to limit the invention. Furthermore, the technical features involved in the various embodiments of this invention described below can be combined with each other as long as they do not conflict with each other.

[0048] Please see Figure 1 This invention provides a method for identifying early faults in rotating machinery. The method is based on Bayesian deep support vector data and mainly includes the following steps:

[0049] S1, collect signal time series data of a single or multiple similar rotating machines during the initial health phase or throughout their entire life cycle to obtain a data sample set.

[0050] Time-series signal data from the operation of rotating machinery is collected. Sampling is performed at equal time intervals according to a certain sampling frequency. Offline sampling yields data for the entire lifecycle of the rotating machinery. This lifecycle data is then divided into training and test sets according to a specific time scale. Alternatively, only a small amount of historical signal data from the initial healthy phase of the rotating machinery can be sampled as the training set. Depending on the research objective, time-series signal data from a single rotating machine or multiple similar rotating machines during their operation can be collected as the data sample set. Online sampling yields data for the current moment of the rotating machinery, which is the data actually used for early fault identification.

[0051] In this embodiment, the time-series sample length is N ,sample From the sampling interval is T Sampling time is t The sampling frequency is f s The timing signal constitutes, The training set was a subset of samples collected during the initial healthy operation phase of the machinery. The remaining samples are used as the test set. .

[0052] S2, the obtained time-series signal data is processed to convert the time-domain signal to the frequency domain and time-frequency domain, thereby obtaining the frequency domain and time-frequency domain feature maps corresponding to the time-series signal. These feature maps include single-sided amplitude-frequency feature maps and VMD-Hilbert time-frequency spectrum feature maps.

[0053] Feature maps are constructed from the collected time-series signals of rotating machinery. The collected signals are a set of time-domain signals, which are time-dependent non-steady-state signals. Signal processing methods are used to convert the time-domain signals to the frequency domain and time-frequency domain, and then a dual-channel feature map spectrum in the frequency domain and time-frequency domain is constructed. The feature map spectrum is then used to complete early fault identification.

[0054] The feature maps include single-sided amplitude-frequency feature maps and VMD-Hilbert time-spectrum feature maps. These are obtained by analyzing time-series data samples. After performing a Fast Fourier Transform to obtain a one-dimensional amplitude-frequency spectrum, it is reconstructed into a two-dimensional feature map, thus obtaining the one-dimensional amplitude-frequency feature map; by selecting an appropriate number of modes... K Variational mode decomposition is performed on the time-series signal to obtain multiple sets of intrinsic mode components. Hilbert transform is then applied to each mode component to obtain the instantaneous frequency that varies with time. and instantaneous amplitude Then, discretize at equal time intervals. and Obtain the time-frequency spectrum of the signal, where the time-frequency expression of the signal is:

[0055]

[0056] In the formula, Represents the real part of a complex number; Indicates the first One mode; The modal number.

[0057] In this example, the choice of feature map construction takes into account the general characteristics of rotating machinery. For different rotating machinery and their sensor signals, it is possible to construct only the frequency domain feature map or construct multiple sets of features such as the envelope spectrum of the signal. The adaptive adjustment of the network model can be completed by simply modifying the number of channels of the input feature map of the network model.

[0058] S3. Construct an unsupervised learning neural network model for describing Bayesian deep support vector data. Wrap the obtained normal samples in a hypersphere that is as small as possible, and use the Bayesian neural network structure to obtain uncertainty information of the model and data.

[0059] A Bayesian Deep Support Vector Data Description (BD-SVDD) unsupervised learning neural network model is constructed. BD-SVDD is a single-class anomaly detection model. Instead of using the kernel function for feature extraction in Support Vector Data Description (SVM), BD-SVDD utilizes a neural network to extract features. A hypersphere is constructed within BD-SVDD, enclosing normal samples within a small enough hypersphere to allow early faults or faulty samples to fall outside the hypersphere, thus achieving early fault identification by combining deep learning methods with BD-SVDD. The Bayesian neural network structure is used to capture uncertainty information from the model and data, reducing cognitive uncertainty in the model output and the false positive rate for early faults. The results of the BD-SVDD neural network are shown below. Figure 2 As shown, specifically, the loss function of the network model is:

[0060]

[0061] In the formula, This represents the square of the distance between the network output and the center c of the hypersphere; s The number of sampling times for the model distribution layer. k The length of the training set samples; These are non-negative correction hyperparameter terms; It is a variable approximate probability distribution; This represents the true posterior distribution; This represents the Kullback-Leibler divergence operator; For expectation operator, express Expectations; This is normal sample data; The weights are unknown parameters that can be trained in a neural network. For given parameters Post-data The likelihood function.

[0062] loss function The first requirement is that all features mapped from training samples to the high-dimensional space should be as close to the center as possible. The latter two are Bayesian regularization terms, which reduce the cognitive uncertainty of the model itself by optimizing the complexity cost, and prevent overfitting by balancing the uncertainty of the model and the data. Since there is a certain difference in data scale between the minimization of the hypersphere loss term and the Bayesian regularization term, a non-negative correction hyperparameter term needs to be added. This ensures that both parts of the loss are effectively updated during gradient backpropagation.

[0063] Among them, the difference from general neural networks is that... As fixed parameters, the Bayesian neural network used assumes all trainable unknown parameters. These are treated as independent random variables following a Gaussian distribution, and their corresponding posterior distributions are inferred from the data through training. Formally, given data... And the network model structure f, unknown parameters The posterior distribution is as follows:

[0064]

[0065] The prediction process can be derived from the posterior distribution. Sampling parameters Through computational models To obtain the prediction results Expected value:

[0066]

[0067] Due to the posterior distribution The denominator needs to be integrated over the entire parameter space, but for the nonlinear, high-dimensional parameter space of neural networks, it is difficult to obtain directly. To approximate the posterior distribution, the Bayes by backprop approximation method must be used. Therefore, a set of... For the mean, The parameter for standard deviation Variable approximate probability distribution of control Through optimization The parameters are such that the KL divergence between q and the true posterior distribution p is minimized:

[0068]

[0069] The first term, representing the complexity cost, indicates the similarity between the neural network parameter weights and the prior information; the second term, representing the likelihood cost, indicates the degree of fit to the samples. Parameter optimization is achieved by striking a balance between these two cost functions. Cost updates require... Perform Monte Carlo sampling: Because sampling directly from the distribution would prevent differentiation during backpropagation, the parameters would become... It is not differentiable, therefore it is necessary to... Perform local reparameterization.

[0070] In Bayesian networks, the local reparameterization operation makes , express It follows a standard Gaussian distribution; and Indicates the firsti The mean of each network parameter, The standard deviation is used to obtain the approximate sampling random variable from the standard Gaussian distribution. In the above formula, Indicates the first layer in the network The weights of the parameters are used to obtain 𝜔 by approximating random variables from a standard Gaussian distribution.

[0071] In Bayesian convolutional layers, instead of placing probability distributions on the weights, the feature maps are treated as random variables, and... ; To experience the wild; This is the variance operator. The feature map is locally reparameterized using a Gaussian distribution through two convolutional kernels to obtain activation values. .

[0072] When the training set sample size is small, the Bayesian neural network randomly samples the weight perturbations during each gradient update, with each mini-batch's perturbation sampled only once, and all samples in the mini-batch sharing a common basic perturbation. Because the high correlation of weight perturbations makes it difficult to eliminate the variance of the gradient estimate through averaging, this results in high variance in the gradient estimate and slow convergence. To address this issue, a FlipOut operation is used to update the basic perturbation. The method is as follows:

[0073]

[0074] In the formula, These are random variables that follow a Rademacher distribution, and each is sampled independently. Therefore and They are identically distributed. By calculating an unbiased estimate of the loss gradient, quasi-independent weight sampling and backpropagation are achieved, thereby reducing the number of iterations.

[0075] S4. Based on the obtained feature maps, the neural network model is trained using a combination of unsupervised and Bayesian backpropagation methods.

[0076] Specifically, the extracted dual-domain feature maps are used to train the Bayesian deep support vector data description network. The training method differs from that of supervised neural networks with labeled data. The training of the network combines unsupervised and Bayesian backpropagation methods.

[0077] Training a neural network model includes the following sub-steps:

[0078] (a) Initialize the parameters and iteration counter of the Bayesian deep support vector data description neural network, with the parameters being twice the number of weight parameters of the neural network;

[0079] (b) Input the samples into the network for training in batches. The last layer is a high-dimensional feature distribution layer, which receives the output 2n from the previous fully connected layer to obtain the distribution information of the n-dimensional feature output. This information is used to characterize the impact of the random uncertainty of the input data on the output. Considering that the variance is non-negative, softplus is introduced as the activation function for the n-dimensional variance feature input. Set the number of sampling times s for the distribution layer, and use the sample mean and variance as the mean and variance of the feature output.

[0080] (c) Update the hypersphere radius R, calculated using the following formula:

[0081]

[0082] (d) Calculate the loss function The gradient and network parameters are updated using the Bayesian backpropagation algorithm, and the network is optimized using Adam. The convergence of the loss or the arrival of the required number of iterations is determined by checking if the standard deviation of the loss calculated in the last 5 iterations is higher than the early stopping standard deviation. If the number of iterations has not been reached, return to step (b). Otherwise, consider the network to have converged and training to end.

[0083] S5 determines the indicator scale for early fault identification in the neural network model. The scale comprehensively represents the uncertainty of the sample being wrapped by the hypersphere after being mapped to the high-dimensional space of the model output.

[0084] Among them, the hypersphere radius obtained by the Bayesian deep support vector data description early fault identification model is represented in the form of a distribution. To achieve high confidence in early fault identification, the early fault point indicator scale is defined based on the 3σ criterion as follows:

[0085]

[0086] In the formula, sample points Category A value of 1 indicates a normal sample. A value of 0 indicates a normal sample; For the center of the super ball; The number of sampling times for the model distribution layer; for Distance from the center of the supersphere; The model's prediction results; The distance between the predicted result and the center of the supersphere; The prediction results after inputting training data into the model; Let be the radius of the hypersphere, and take... The maximum value in 0 and 0.

[0087] When the distribution of the distance between a sample and the center of the hypersphere exceeds the upper confidence limit of the hypersphere radius, the sample is considered to be an early fault sample.

[0088] S6 calculates the distance between the sample of the rotating machinery and the center of the hypersphere at each moment, and then identifies the early failure point of the rotating machinery by comparing it with the upper confidence limit of the hypersphere radius, and provides health level information indication.

[0089] Specifically, the rotating machinery health level information indication indicates for:

[0090]

[0091] In the formula, This serves as the baseline for health status.

[0092] The present invention will be further described in detail below with reference to specific embodiments.

[0093] In this specific embodiment, a rotating machine called a bearing is taken as the specific object, and the corresponding early fault identification method mainly includes the following steps:

[0094] (1) Obtain vibration signal data of the bearing throughout its entire life cycle. The XJTU-SY test dataset was used for training and testing. The sampling frequency was set to 25.6 kHz, the sampling interval was 1 min, and the single sampling duration was 1.28 s. Three bearing operating conditions were used: 2100 r / min with radial force of 12 kN, 2250 r / min with radial force of 11 kN, and 2400 r / min with radial force of 10 kN. The bearing parameters used are shown in Table 1. A portion of the samples collected during the initial healthy operation phase of the bearing (less than the first 25%) was used as the training set. The training set sample length was set for the case of small samples of a single bearing. It is 30.

[0095] Table 1: Bearing Parameters in the XJTU-SY Dataset

[0096]

[0097] (2) Extract the dual-domain feature map from the horizontal vibration signal of the bearing. The signal is converted into a one-dimensional single-amplitude spectrum with a width of 32768 by Fourier transform and recombined into a frequency domain feature map of 128×128. Set the number of VMD modal components. K =10, performing variational mode decomposition on the time-series signal yields multiple sets of intrinsic mode components, and performing Hilbert transform on each mode component yields the instantaneous frequency that varies with time. and instantaneous amplitude Then, discretize at time intervals of 0.01s. and A 128×128 time-spectrum feature map was obtained.

[0098] (3) Establish a Bayesian deep support vector data description neural network, the network structure of which is as follows: Figure 2 As shown in Table 2, the hyperparameters are as follows. During training, Adam was used for parameter optimization. The learning rate of the optimization algorithm was set to lr=0.001, the maximum number of iterations was 500, and the early stopping standard deviation was [not specified]. Each iteration uses batch training, which divides the total training samples into batches, with each batch containing the same number of samples.

[0099] Table 2: Hyperparameter Table of Bayesian Deep Support Vector Data Description Neural Network

[0100]

[0101] (4) Early fault identification and health level information visualization of bearings are performed. A trained Bayesian deep support vector data description model is used, and the detection method described in the specific implementation method is employed to perform early fault identification on actual bearing data. The original vibration signal, early fault identification results, and health level information indicators are visualized, where the health status baseline is... ,like Figures 3-1 to 3-6 The results of early fault identification in different bearings are shown.

[0102] To verify the performance of the proposed early fault identification method for rotating machinery based on a Bayesian deep support vector data description neural network, this invention takes bearing 1_1 as an example. It is known that the bearing fault is an outer ring fault, which will excite periodic impact signals during bearing operation. These signals will modulate with high-frequency natural vibrations. Based on expert knowledge and experience, and considering that envelope spectrum analysis can effectively demodulate and extract such low-frequency impact signals, the original vibration signals near the early fault identification results are analyzed using the Hilbert envelope spectrum of the signals to demonstrate the effectiveness of the early fault identification results. Figure 4-1 , Figure 4-2 , Figure 4-3 , Figure 4-4 and Figure 4-5 As shown. The formula for calculating the failure frequency of the outer ring of a rolling bearing is:

[0103]

[0104] Where r is the bearing speed, the fault characteristic frequency is calculated. .from Figures 4-1 to 4-5 As can be seen, the highest frequency of the envelope spectrum peak of the 78th period sample is 34.375Hz, which is very close to the bearing's rotational frequency of 35Hz. The highest frequency of the envelope spectrum peak of the 79th sample is 108.59375Hz, which is very close to the bearing's outer ring failure frequency. Therefore, it can be proved that an early failure of the outer ring occurred at the 78th to 79th samples, verifying the effectiveness of the present invention.

[0105] Those skilled in the art will readily understand that the above description is merely a preferred embodiment of the present invention and is not intended to limit the present invention. Any modifications, equivalent substitutions, and improvements made within the spirit and principles of the present invention should be included within the scope of protection of the present invention.

Claims

1. A method for identifying early faults in rotating machinery, characterized in that, The identification method includes the following steps: (1) Collect signal time series data of a single or multiple similar rotating machines during the initial health phase or throughout their entire life cycle to obtain a data sample set; (2) The obtained time-series signal data is processed to convert the time-domain signal to the frequency domain and time-frequency domain, thereby obtaining the frequency domain and time-frequency domain feature maps corresponding to the time-series signal. The feature maps include the single-sided amplitude-frequency feature map and the VMD-Hilbert time-frequency spectrum feature map. (3) Construct an unsupervised learning neural network model for describing Bayesian deep support vector data, wrap the obtained normal samples in a hypersphere that is as small as possible, and use the Bayesian neural network structure to obtain uncertainty information of the model and data. (4) Based on the obtained feature maps, the neural network model is trained using a combination of unsupervised and Bayesian backpropagation methods; (5) Determine the indicator scale for early fault identification of the neural network model. The scale comprehensively represents the uncertainty of the sample being wrapped by the hypersphere after being mapped to the high-dimensional space of the model output. (6) Calculate the distance between the sample of the rotating machinery and the center of the hypersphere at each moment, and then identify the early failure point of the rotating machinery by comparing it with the upper confidence limit of the hypersphere radius, and provide health level information indication; The loss function of the neural network model is: In the formula, This represents the square of the distance between the network output and the center c of the hypersphere; s The number of sampling times for the model distribution layer. k The length of the training set samples; These are non-negative correction hyperparameter terms; It is a variable approximate probability distribution; This represents the true posterior distribution; This represents the Kullback-Leibler divergence operator; For expectation operator, express Expectations; This is normal sample data; The weights are unknown parameters that can be trained in a neural network. For given parameters Post-data The likelihood function.

2. The method for identifying early faults in rotating machinery as described in claim 1, characterized in that: By analyzing the obtained time series data samples After performing a fast Fourier transform to obtain a one-dimensional single-sided amplitude spectrum, it is reconstructed into a two-dimensional feature map, thus obtaining the single-sided amplitude-frequency feature map.

3. The method for identifying early faults in rotating machinery as described in claim 2, characterized in that: By selecting the number of modes K Variational mode decomposition is performed on the time-series signal to obtain multiple sets of intrinsic mode components. Hilbert transform is then applied to each mode component to obtain the instantaneous frequency that varies with time. and instantaneous amplitude Then, discretize at equal time intervals. and To obtain the signal's time spectrum; The time-frequency expression of the signal is: In the formula, Represents the real part of a complex number; Indicates the first One mode; The modal number.

4. The method for identifying early faults in rotating machinery as described in claim 3, characterized in that: The neural network model assumes that all trainable unknown parameters... These are treated as independent random variables following a Gaussian distribution, and their corresponding posterior distributions are inferred from the data through training; given data and network model structure Unknown parameters The posterior distribution is as follows: In the formula, It is the likelihood function; This is the prior probability; For marginal likelihood, the prediction process starts from the posterior distribution. Sampling parameters Through computational models To obtain the prediction results Expected value: The Bayes by backprop approximation inference method is used to approximate the posterior distribution, and a set of data is established. For parameter mean, The parameter for standard deviation Variable approximate probability distribution of control .

5. The method for identifying early faults in rotating machinery as described in claim 3, characterized in that: Through optimization The parameters make the The KL divergence with the true posterior distribution p should be minimized as much as possible: In the formula, Represents the cost of complexity. Representing the likelihood cost, a local reparameterization operation is performed on , making , express It follows a standard Gaussian distribution; and Let represent the mean of the i-th network parameter. is the standard deviation; to achieve the approximate sampling of random variables from a standard Gaussian distribution to obtain 𝜔.

6. The method for identifying early faults in rotating machinery as described in claim 3, characterized in that: For a Bayesian convolutional layer, through two convolutional kernels, let , ; To experience the wild; As a variance operator, Gaussian distribution is used to perform local reparameter sampling on the feature map to obtain activation values. .

7. The method for identifying early faults in rotating machinery as described in claim 3, characterized in that: The FlipOut operation is used to correct the Bayesian variational inference and update the basic perturbation. The method is as follows: In the formula, These are random variables that follow a Rademacher distribution, and each was sampled independently. This is the basic perturbation.

8. The method for identifying early faults in rotating machinery as described in claim 4, characterized in that: Based on the 3σ criterion, the early failure point indicator scale is defined as follows: In the formula, sample points Category A value of 1 indicates a normal sample. A value of 0 indicates a normal sample; For the center of the super ball; for Distance from the center of the supersphere; The model's prediction results; The distance between the predicted result and the center of the supersphere; The prediction results after inputting training data into the model; Let be the radius of the hypersphere, and take... The maximum value of 0; This represents the number of sampling times for the model's distribution layer.

9. The method for identifying early faults in rotating machinery as described in claim 5, characterized in that: The health level information of rotating machinery is indicated as follows: In the formula, This serves as the baseline for health status.

Images