Failure prediction method for vulnerable parts based on sensor signals and deep migration learning

Abstract

The application discloses a kind of based on sensor signal and depth migration learning's vulnerable part failure prediction method, including S1, sensor is installed on different positions of vulnerable part;S2, different sensors are received to the signal collected on vulnerable part;The data processing and frequency spectrum analysis are carried out to the received signal, determine the best installation position of sensor;S3, signal is collected to the best position and is processed by abnormal value and noise reduction, then the features of different scope are extracted;S4, using CNN convolution neural network, and combining DANN and MMD feature migration method constructs the deep learning prediction model of vulnerable part;S5, using the deep learning prediction model of vulnerable part carries out the failure prediction of vulnerable part.The application can improve the prediction accuracy, reduce the data requirement, accelerate model training using migration learning to carry out the failure prediction of vulnerable part, while realizing knowledge sharing, with great application prospect and practical value.

Description

Technical Field

[0001] This invention belongs to the technical field of signal processing and machine learning, specifically relating to a method for predicting the failure of vulnerable components based on sensor signals and deep transfer learning. Background Technology

[0002] As one of the core components of mechanical equipment, the failure of vulnerable parts can lead to machine downtime, production interruptions, and equipment maintenance problems. Therefore, failure prediction of vulnerable parts has always been a research hotspot in the mechanical field. Traditional methods for predicting the failure of vulnerable parts usually require a large amount of data and the experience and knowledge of domain experts, and have certain limitations in practical applications. In recent years, with the development and application of sensor technology and deep transfer learning technology, they have broad application prospects in solving problems in areas such as small sample sizes and multi-task scenarios. Transfer learning uses knowledge from the source domain to assist the learning of the target domain, which can effectively improve the performance and generalization ability of the model. Therefore, failure prediction methods for vulnerable parts based on transfer learning have gradually become a research hotspot. In practical applications, the task of predicting the failure of vulnerable parts usually faces the following challenges: on the one hand, actual data is often relatively scarce, which makes it difficult for traditional machine learning methods to achieve good results; on the other hand, the failure prediction of vulnerable parts needs to combine information from multiple aspects, such as vibration signals, temperature, and rotational speed. These information come from different sources and contain different noise and interference, requiring reasonable signal processing and feature extraction. Summary of the Invention

[0003] The purpose of this invention is to address the above-mentioned shortcomings in the prior art by providing a method for predicting the failure of vulnerable parts based on sensor signals and deep transfer learning, thereby solving the problem that existing traditional machine learning methods are difficult to accurately predict the failure of vulnerable parts.

[0004] To achieve the above objectives, the technical solution adopted by the present invention is as follows:

[0005] A method for predicting the failure of vulnerable components based on sensor signals and deep transfer learning includes the following steps:

[0006] S1. Install different sensors at different locations on vulnerable parts;

[0007] S2. Receive signals collected by different sensors on vulnerable parts, perform data processing and spectrum analysis on the received signals, and determine the optimal installation position of the sensor.

[0008] S3. The signal acquired at the optimal location is processed for outlier and noise reduction, and its features in different domains are extracted.

[0009] S4. A deep learning prediction model for vulnerable parts is constructed by using a CNN convolutional neural network and combining DANN and MMD feature transfer methods.

[0010] S5. Use a deep learning prediction model for vulnerable parts to predict their failure.

[0011] Furthermore, in step S2, the time-domain amplitude of the sensor signals at different locations is compared and the spectrum of Fourier transform is analyzed to determine the optimal installation location of the sensor.

[0012] Furthermore, step S3 specifically includes:

[0013] Outliers in the signal are identified and removed using the mean and standard deviation.

[0014] Fourier transform is used to perform spectral analysis on the signal;

[0015] Wavelet semi-soft thresholding denoising is used to denoise the signal after outlier processing. The threshold function is as follows:

[0016]

[0017] Where λ is a given threshold, w j,k These are wavelet coefficients. The estimated wavelet coefficients are α, where α is the correction coefficient.

[0018] Extract the signal features of the denoised signal in the time domain, frequency domain, and time-frequency domain.

[0019] Furthermore, the deep transfer learning model in step S4 specifically includes:

[0020] S4.1 Feature Extractor: The source domain signal and the target domain signal are processed by a CNN convolutional neural network to extract deep features, and the obtained features are flattened.

[0021] S4.2 Predictor: The source domain signal features obtained by the feature extractor are substituted into multiple fully connected neural networks for model pre-training. The root mean square error between the output value and the target value of the fully connected neural network is calculated and used as the loss function. Minimizing the loss function is used as the training objective to update the network parameters θ of the predictor. P With the network parameters θ of the feature extractor F ;

[0022] S4.3, Domain Classifier: The extracted source domain signal features and target domain signal features are input into the domain classification model for classification training. Its loss function, the sampling cross-entropy function, is:

[0023]

[0024] in, The classification loss is for the source domain signal and the target domain signal. These represent the source domain signal and the target domain signal, respectively, where m is the batch number. The output is a domain classification for the source domain signal. The output is a domain classification for the target domain signal. These are the features extracted from the source domain signal by the feature extractor. Features extracted from the target domain signal by the feature extractor;

[0025] Update the network parameters θ of the domain classifier using this loss function. D ;

[0026] After assigning the target domain signal features to the opposite domain labels, they are input into the domain classifier for training, and their loss function is calculated:

[0027]

[0028] in, Classification loss for the target domain;

[0029] The neighborhood classifier parameters θ are updated using a neighborhood segmentation loss function based on the features of the target domain signal. D With feature extractor parameters θ F ;

[0030] Use the normal source and target domain classification loss. Target domain classification loss with inverse domain labels Adversarial training of the obfuscated feature extractor, which is essentially a classification loss between the normal source and target domains. Reaching the maximum;

[0031] S4.4, Calculation of Maximum Mean Difference in MMD: Map the source domain signal features and target domain signal features to a Hilbert high-dimensional space, and calculate their maximum mean difference:

[0032]

[0033] in, n is the maximum mean-based difference between the source domain features and the target domain features calculated in a high-dimensional Hilbert space. S and n t These represent the number of samples in the source domain and the target domain, respectively. For mapping functions, To map the source domain samples, To map the samples in the target domain, ||·|| is the L2 norm, and H is the Hilbert space;

[0034] And minimize For the network parameters θ of the target feature extractor F Update.

[0035] Further, in step S4.2, the source domain signal features obtained by the feature extractor are substituted into multiple fully connected neural networks for model pre-training. The root mean square error between the output value and the target value of the fully connected neural network is calculated and used as the loss function. The loss function is as follows:

[0036]

[0037] in, The output of the predictor With actual label The result is the error value calculated using the root mean square error function.

[0038] The method for predicting the failure of vulnerable components based on sensor signals and deep transfer learning provided by this invention has the following beneficial effects:

[0039] The transfer learning method of this invention can make full use of the data and knowledge of the source domain to improve the learning effect of the target domain. At the same time, it can effectively utilize information from multiple information sources to improve the predictive performance and generalization ability of the model. It has important practical application value for improving the reliability and safety of mechanical equipment.

[0040] This invention employs a CNN convolutional neural network and combines DANN and MMD feature transfer methods to construct a deep learning prediction model for vulnerable components. It also introduces the ideas of domain adversarial and domain adaptation, which improves the transfer effect and generalization of the model.

[0041] This invention can reduce data requirements: Traditional machine learning methods for predicting failure of vulnerable parts require a large amount of sample data, while transfer learning methods can utilize data from other related fields, thereby reducing the need for experimental data on vulnerable parts and lowering the cost of data collection and labeling.

[0042] This invention can improve prediction accuracy: Since predicting the failure of vulnerable parts is a complex task, transfer learning can utilize knowledge from other fields to improve the generalization ability of the model, thereby improving the accuracy and precision of the prediction.

[0043] This invention can accelerate model training: using pre-trained models for transfer learning can reduce model training time and improve training efficiency.

[0044] This invention enables knowledge sharing: through transfer learning, knowledge in the same field can be mutually referenced and shared, thereby promoting cross-disciplinary integration and improving innovation capabilities. Attached Figure Description

[0045] Figure 1 This is a flowchart of a method for predicting the failure of vulnerable parts based on sensor signals and deep transfer learning.

[0046] Figure 2 This describes the model structure for transfer learning methods.

[0047] Figure 3 This provides an accelerated experimental platform for bearing degradation and vibration signals under different failure modes and three different operating conditions.

[0048] Figure 4 These are the prediction results from a traditional CNN deep learning model.

[0049] Figure 5 These are the prediction results of the transfer learning model proposed in this invention. Detailed Implementation

[0050] The specific embodiments of the present invention are described below to enable those skilled in the art to understand the present invention. However, it should be understood that the present invention is not limited to the scope of the specific embodiments. For those skilled in the art, various changes are obvious as long as they are within the spirit and scope of the present invention as defined and determined by the appended claims. All inventions utilizing the concept of the present invention are protected.

[0051] Example 1

[0052] This embodiment provides a method for predicting the failure of vulnerable components based on sensor signals and deep transfer learning. This transfer learning-based method can fully utilize data and knowledge from the source domain to improve the learning effect in the target domain. Simultaneously, it effectively utilizes information from multiple sources to improve the model's predictive performance and generalization ability. (Reference...) Figure 1 Specifically, it includes the following steps:

[0053] Step S1: Install different sensors at different locations on the vulnerable parts;

[0054] The sensor installation method in this step can be determined based on the structure and material characteristics of the vulnerable parts, and a trial-and-error method can be used to install the sensor in different positions.

[0055] Step S2: Receive the signals collected by the sensor on the vulnerable parts and perform signal analysis;

[0056] This step involves comparing the time-domain amplitude and performing spectrum analysis on the acquired signals. Finally, the optimal installation position of the sensor is determined based on the natural frequency and vibration amplitude of the vulnerable components.

[0057] Step S3: Process the sensor signal received at the optimal position and extract its signal features, which specifically includes the following:

[0058] Outlier handling: During the sensor signal acquisition process, human factors or other factors may occasionally cause some obvious outliers in the acquired signals. This step identifies and removes outliers based on the mean and standard deviation, so that the collected signals are more conducive to reflecting the true operating condition of vulnerable parts.

[0059] Spectrum analysis: Using Fourier transform, observe the frequency range of the acquired signal and determine the key frequency bands that need to be analyzed and processed based on existing experience;

[0060] Wavelet Denoising: During the operation of vulnerable components, the signals collected by sensors are not entirely caused by the operation of the vulnerable components; they contain interference from some noise signals. Therefore, this invention employs a wavelet semi-soft thresholding denoising method to reduce noise in the signal after outlier processing. The threshold function is as follows:

[0061]

[0062] Where λ is a given threshold, w j,k These are wavelet coefficients. The estimated wavelet coefficients are α, where α is the correction coefficient.

[0063] When α = 0, the hard thresholding function has a significant denoising effect, but the continuity of the processed signal is poor.

[0064] When α = 1: The soft thresholding function has good continuity, but leaves relatively more noise; a semi-soft thresholding method is used to select an appropriate α to balance the noise reduction effect.

[0065] Feature extraction: This mainly reflects the changing state of vulnerable components throughout their entire operating cycle. Signal features are extracted in three domains: statistical domain, frequency domain, and time-frequency domain. The specific extraction methods are shown in the table below:

[0066] Feature extraction table

[0067]

[0068]

[0069] In the table above, x represents the number of time-domain signals; i Represents the time-domain signal; μ represents the time-domain mean; σ represents the standard deviation; E[·] represents the expected value; k represents the batch size of the frequency-domain signal; f i This represents the amplitude corresponding to the i-th frequency. denoted as the frequency domain mean; S(·) is the power spectral density obtained by the Welch method; wtA(i) and wtD(i) represent the first-order approximation coefficients and first-order detail coefficients after wavelet decomposition, respectively.

[0070] Step S4: Construct a deep learning prediction model for vulnerable parts using a CNN convolutional neural network and combining DANN and MMD feature transfer methods;

[0071] DANN (Domain-Adversarial Neural Networks) is a transfer learning method that enables knowledge transfer between a source and target domain to improve learning performance and generalization ability in the target domain. Its basic idea is to use a domain classifier to learn to distinguish between source and target domain data, and then integrate this domain classifier with the original task classifier to achieve knowledge sharing between the source and target domains. Generally, the core of the DANN method lies in the adversarial interaction between the domain classifier and the feature extractor. The feature extractor is a model used to extract useful features from the raw data; the domain classifier is a model used to distinguish between source and target domain data. For the same input data, the classifier outputs the probability that the corresponding data comes from the source or target domain. During the training of the domain classifier, DANN uses the backpropagation algorithm to minimize the error of the domain classifier and achieves the transfer learning effect by minimizing the difference in domain distribution between the source and target domains. In this way, DANN can leverage the similarity between the source and target domains to improve learning performance and generalization ability in the target domain, thereby achieving cross-domain transfer learning.

[0072] Maximum Mean Discrepancy (MMD) is a kernel-based transfer learning method that measures the similarity between different domains and achieves transfer learning by minimizing the differences between them. Its basic idea is to evaluate the differences between the feature distributions of the source and target domains by comparing them, and to achieve knowledge transfer by minimizing these differences. The MMD method calculates the difference in feature distributions between the source and target domains using the following formula:

[0073]

[0074] Among them, X s and X t Let n represent the sample sets of the source domain and the target domain, respectively. S and n t Let represent the number of samples in the source domain and the target domain, respectively. Let ||·|| represent the function that maps sample x to a high-dimensional space, and ||·|| represent the L2 norm. Since... Since MMD is difficult to compute directly, kernel methods are typically used for approximation. MMD-based transfer learning methods achieve transfer learning effects by minimizing the feature distribution differences between the source and target domains. Specifically, MMD can be defined by minimizing the risk function in the target domain using the following formula:

[0075] miθ_θL(X s ,Y s ,θ)+λMMD(X s ,X t )

[0076] Wherein, L(X) s ,Y s Let ,θ) represent the classification loss function in the source domain, θ represent the model parameters, and λ represent the hyperparameters controlling the weights of the MMD terms. By minimizing this risk function, MMD can simultaneously optimize classification performance in the target domain and the feature distribution difference between the source and target domains, thereby achieving the effect of transfer learning. In summary, the MMD method measures the similarity between the feature distributions of the source and target domains by comparing them, and achieves knowledge transfer by minimizing this difference. This method can achieve good performance in transfer learning tasks, especially when the feature spaces of the source and target domains are highly similar.

[0077] refer to Figure 2 This step combines the two methods mentioned above, MMD and DANN, and introduces the ideas of domain adversarial and domain adaptation to improve the model's transfer performance and generalization ability. It includes the following four parts:

[0078] Step S4.1, Feature Extractor: Deep feature extraction is performed on the source domain signal and the target domain signal through a multi-layer convolutional neural network, and finally the features are flattened for subsequent analysis.

[0079] Step S4.2, Predictor: Substitute the source domain signal features obtained by the feature extractor into multiple fully connected neural networks for model pre-training, calculate the root mean square error between the network output value and the target value, and use it as the loss function.

[0080]

[0081] Minimizing the loss function is used as the training objective to update the network parameters θ of the predictor. P With the network parameters θ of the feature extractor F This gives the model a certain degree of convergence.

[0082] Step S4.3, Domain Classifier:

[0083] The first step is to input the extracted source domain signal features and target domain signal features into the domain classification model for classification training. The loss function and sampling cross-entropy function are:

[0084]

[0085] This loss function is used to update the network parameters θ of the domain classifier. D ;

[0086] The second step involves assigning the target domain signal features to opposite domain labels and then inputting them into the domain classifier for training. The loss function is:

[0087]

[0088] The domain classifier parameters θ are updated using a neighborhood partitioning loss function based on the target domain signal features. D With feature extractor parameters θ F .

[0089] Using the normal source and target domain classification loss Target domain classification loss with inverse domain labels Adversarial training between the two systems is used to confuse the feature extractor, making it impossible for the domain classifier to accurately distinguish whether the features extracted by the feature extractor come from the source domain or the target domain, thus affecting the domain classification loss. Reach the maximum.

[0090] MMD maximum mean difference calculation: The maximum mean difference is calculated by mapping the source domain features and target domain features to a Hilbert high-dimensional space using the following formula:

[0091]

[0092] The network parameters θ for feature extraction are optimized by minimizing MMD. F The feature distributions of the extracted source and target domains are updated to be more consistent, thus achieving the goal of transfer learning.

[0093] Step S5: Use a deep learning prediction model for vulnerable parts to predict their failure.

[0094] In traditional technologies, failure prediction typically uses a rated threshold for failure determination. The rated threshold varies depending on the application scenario. For example, for tool wear failure prediction in machining, an upper limit for wear value is usually set. When the model's predicted value exceeds the upper limit of the threshold, it is determined that the currently collected signal corresponds to the tool failure state. Another example is common bearing parts, whose failure mode is not just a single wear behavior, making it difficult to use specific physical quantities as failure thresholds. Generally, failure data for the entire cycle is collected, the remaining life value corresponding to each cycle is calculated, and then a rated remaining life lower limit is set as the failure determination threshold.

[0095] The prediction method of this invention is universal for different vulnerable parts. Therefore, this embodiment uses the publicly available bearing failure dataset from the 2012 PHM Failure Prediction and Health Management Challenge in the United States to test the transfer learning model of this invention.

[0096] refer to Figure 3 This paper demonstrates an accelerated experimental platform for bearing degradation and vibration signals under different failure modes and three different operating conditions. The signal spectrum shows that the vibration signal form of the bearing under different operating conditions has certain differences throughout the entire cycle. In transfer learning, this difference is usually called feature distribution inconsistency. The necessary condition for traditional machine learning to work is the consistency of feature distribution. Therefore, this invention aims to use transfer learning to bring the feature distribution of signals from different fields closer together, so that the machine learning model can achieve better predictive performance.

[0097] First, bearing data under the third operating condition is selected as the source domain (its vibration signal and corresponding life label need to be imported into the model). Then, a regular CNN convolutional neural network model is built, and the model is trained. Finally, the trained model is tested under the first operating condition, and its prediction results are as follows. Figure 4 As shown, the prediction results are very poor, indicating that the model has weak generalization ability.

[0098] The transfer learning prediction model proposed in this invention was tested. Bearing data under the third working condition was used as the source domain (its vibration signal and corresponding life label needed to be imported into the model), and the first working condition was selected as the target domain (only its vibration signal was imported into the model). The convolutional module settings of the model were consistent with the previous CNN model. Then, the model was trained and tested. The prediction results were... Figure 5 As shown in the figure, the prediction performance of the transfer learning model proposed in this invention is significantly better than that of the traditional CNN network model, which also verifies the superiority of transfer learning in different working conditions or different domain problems.

[0099] Although specific embodiments of the invention have been described in detail with reference to the accompanying drawings, this should not be construed as limiting the scope of protection of this patent. Various modifications and variations that can be made by a person skilled in the art without inventive effort within the scope described in the claims still fall within the scope of protection of this patent.

Claims

1. A method for predicting the failure of vulnerable components based on sensor signals and deep transfer learning, characterized in that, Includes the following steps: S1. Install different sensors at different locations on vulnerable parts; S2. Receive signals collected by different sensors on vulnerable parts, perform data processing and spectrum analysis on the received signals, and determine the optimal installation position of the sensor. S3. The signal acquired at the optimal location is processed for outlier and noise reduction, and its features in different domains are extracted. S4. A deep learning prediction model for vulnerable parts is constructed by using a CNN convolutional neural network and combining DANN and MMD feature transfer methods. Deep learning prediction models, specifically including: S4.1 Feature Extractor: The source domain signal and the target domain signal are processed by a CNN convolutional neural network to extract deep features, and the obtained features are flattened. S4.2 Predictor: The source domain signal features obtained by the feature extractor are substituted into multiple fully connected neural networks for model pre-training. The root mean square error between the output value and the target value of the fully connected neural network is calculated and used as the loss function. Minimizing the loss function is used as the training objective to update the network parameters of the predictor. Network parameters of the feature extractor ; S4.3, Domain Classifier: The extracted source domain signal features and target domain signal features are input into the domain classification model for classification training. Its loss function, the sampling cross-entropy function, is: in, The classification loss is for the source domain signal and the target domain signal. , These are the source domain signal and the target domain signal, respectively. For batch number of times, The output is a domain classification for the source domain signal. The output is a domain classification for the target domain signal. These are the features extracted from the source domain signal by the feature extractor. Features extracted from the target domain signal by the feature extractor; Update the network parameters of the domain classifier using this loss function. ; After assigning the target domain signal features to the opposite domain labels, they are input into the domain classifier for training, and their loss function is calculated: in, Classification loss for the target domain; The parameters of the neighborhood classifier are updated using a neighborhood segmentation loss function based on the features of the target domain signal. With feature extractor parameters ; Use the normal source and target domain classification loss. Target domain classification loss with inverse domain labels Adversarial training of the obfuscated feature extractor, which is essentially a classification loss between the normal source and target domains. Reaching the maximum; S4.4, Calculation of Maximum Mean Difference in MMD: Map the source domain signal features and target domain signal features to a Hilbert high-dimensional space, and calculate their maximum mean difference: in, This represents the maximum mean-based difference between the source domain features and the target domain features calculated in a high-dimensional Hilbert space. These represent the number of samples in the source domain and the target domain, respectively. For mapping functions, To map the source domain samples, To map the target domain samples, It is the L2 norm. For Hilbert space; And minimize Network parameters for the target feature extractor Update; In step S4.2, the source domain signal features obtained by the feature extractor are substituted into multiple fully connected neural networks for model pre-training. The root mean square error between the output value and the target value of the fully connected neural network is calculated and used as the loss function. The loss function is as follows: in, The output of the predictor With actual label The result is the error value calculated using the root mean square error function; S5. Use a deep learning prediction model for vulnerable parts to predict their failure.

2. The method for predicting the failure of vulnerable components based on sensor signals and deep transfer learning according to claim 1, characterized in that, In step S2, the sensor signals from different locations are compared in the time domain and subjected to Fourier transform spectrum analysis to determine the optimal installation location of the sensor.

3. The method for predicting the failure of vulnerable components based on sensor signals and deep transfer learning according to claim 2, characterized in that, Step S3 specifically includes: Outliers in the signal are identified and removed using the mean and standard deviation. Fourier transform is used to perform spectral analysis on the signal; Wavelet semi-soft thresholding denoising is used to denoise the signal after outlier processing. The threshold function is as follows: in, These are wavelet coefficients. The estimated wavelet coefficients, The correction factor; Extract the signal features of the denoised signal in the time domain, frequency domain, and time-frequency domain.

Images