Multi-similarity feature fusion rotating machine fault prediction method
Through the multi-similarity feature fusion method, the problem of unbalanced and computational randomness of fault prediction data in the prior art is solved, and the stable and accurate Failure prediction results.
Patent Information
- Application Number
- CN202411014941.0
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2024-07-26
- Publication Date
- 2024-11-22
AI Technical Summary
The existing rotary mechanical fault prediction methods have difficulty in obtaining label data, imbalance in health and fault data, supervised learning methods have poor diagnosis results on new equipment, and unsupervised fault prediction has a large amount of calculation and randomness. Problem of unstable performance.
By using the multi-similarity feature fusion method, time-frequency feature extraction is performed on one-dimensional vibration data, and the healthy sample data manifold learned by the generative adversarial network is used to determine the generated samples of the input samples based on multiple similarity indicators, and calculate the abnormality. score, obtain the final anomaly score to achieve fault prediction.
Achieve stable and accurate mechanical fault prediction, overcome the problem of unstable fault prediction performance caused by high stochasticity in the calculation process of a single similarity index. The final abnormal score obtained by the fusion of multiple similarity indexes can better reflect the real fault problem. .
Smart Images

Figure CN119004314A_ABST
Abstract
Description
Technical Field
[0001] The present application relates to the technical field of mechanical fault prediction, and in particular to a method for predicting rotating mechanical faults by fusing multiple similarity features. Background Art
[0002] Mechanical failure prediction can help companies take appropriate maintenance and repair measures before workshop equipment fails, thus avoiding production interruptions and unnecessary repair costs. Rotating machinery is widely used and is one of the mechanical structures with the highest failure rate. It is very important to predict the failure of rotating machinery and perform early maintenance.
[0003] Existing fault prediction methods mainly include machine learning methods based on physical properties or physical models, and data-driven fault prediction methods. However, physical modeling methods require domain prior knowledge or the characteristics of rotating machinery to carry out complex modeling or design of a large number of rules. Although data-driven fault prediction methods have good prediction effects, most of them are based on supervised learning methods, and a large amount of labeled data is required during model training. In actual industrial production scenarios, the degree of mechanical equipment failure is difficult to judge, and the failure time only accounts for a very small part of the full life cycle. This makes it difficult to obtain labeled data for mechanical equipment, and there is a serious imbalance between health and fault data. In addition, the individual differences between equipment make the supervised method less effective in diagnosing new equipment, and it is impossible to obtain fault data on new equipment to correct the model. These factors limit the application of supervised learning methods in the field of fault prediction.
[0004] In addition, the unsupervised fault prediction method used in the field of rotating machinery health monitoring only uses one similarity index information of the input sample, and does not fully utilize the multiple similarity index information of the input sample. It also has the problems of large amount of calculation, large randomness of the calculation process, and unstable prediction performance. Therefore, how to achieve stable and accurate mechanical fault prediction is a technical problem that needs to be solved urgently. Summary of the invention
[0005] In view of the above problems, an embodiment of the present application provides a method for predicting rotating machinery faults by fusing multiple similarity features, so as to overcome the above problems or at least partially solve the above problems.
[0006] The present application embodiment discloses a method for predicting rotating machinery faults by fusing multiple similarity features, the method comprising:
[0007] Extract time-frequency features from one-dimensional vibration data to obtain input samples;
[0008] Based on multiple similarity indicators, determining generated samples similar to the input sample on the healthy sample data manifold learned by the generative adversarial network, and calculating the abnormality score of the input sample under each similarity indicator according to the generated samples;
[0009] A final anomaly score is obtained according to the anomaly score of the input sample under each of the similarity indicators, and a fault prediction result is obtained according to the final anomaly score, wherein the final anomaly score represents the degree to which the input sample deviates from the healthy sample data.
[0010] Optionally, the generative adversarial network includes a discriminator and a generator, and the generative adversarial network is trained according to the following steps:
[0011] The generator generates sample data according to random noise;
[0012] According to the generated sample data and the healthy sample data, construct a discriminator optimization objective function, wherein the discriminator optimization objective function includes: a score of the discriminator on the healthy sample data, a score of the discriminator on the generated sample data, and a gradient penalty term;
[0013] According to the generated sample data and the healthy sample data, a generator optimization objective function is constructed, wherein the generator optimization objective function includes: a score of the generated sample data by the discriminator, and a matching loss term between features of the generated sample data on the discriminator feature layer and features of the healthy sample data;
[0014] Fixing the generator, updating the parameters of the discriminator according to the discriminator optimization objective function, and fixing the discriminator, updating the parameters of the generator according to the generator optimization objective function;
[0015] After the training end condition is met, a trained generative adversarial network is obtained, and the trained generative adversarial network learns the mapping relationship from the distribution obeyed by random noise to the distribution of healthy sample data.
[0016] Optionally, based on multiple similarity indicators, a generated sample similar to the input sample is determined on the healthy sample data manifold learned by the generative adversarial network, and an abnormal score of the input sample under each similarity indicator is calculated according to the generated sample, including:
[0017] Using a distribution estimation algorithm, searching for samples in the latent space of the generative adversarial network based on each similarity index, and determining the sample with the greatest similarity to the input sample as a generated sample similar to the input sample;
[0018] The inverse number of the similarity between the generated sample and the input sample under each similarity index is used as the anomaly score of the input sample under each similarity index.
[0019] Optionally, a distribution estimation algorithm is used to search for samples in the latent space of the generative adversarial network based on each similarity index, and a sample with the greatest similarity to the input sample is determined as a generated sample similar to the input sample, including:
[0020] Sample N×M sampling points from the distribution of random noise to form a first set;
[0021] The generative adversarial network generates first samples according to the sampling points in the first set in sequence, and calculates the similarity between the first sample and the input sample according to the similarity index;
[0022] According to the magnitude of the similarity, first M sampling points with large similarity are selected from the first set to form a second set;
[0023] Sampling N sampling points from the distribution of the random noise to form a third set;
[0024] Calculating the target variable according to the second set and the third set to obtain a fourth set, iteratively updating the fourth set, and obtaining a fifth set according to the updated fourth set;
[0025] According to the fifth set obtained by the last iterative update, a generated sample similar to the input sample is determined.
[0026] Optionally, calculating the target variable according to the second set and the third set to obtain a fourth set, iteratively updating the fourth set, and obtaining a fifth set according to the updated fourth set includes:
[0027] Adding the fifth set obtained by the last iterative update to the fourth set, and constraining the target variable in the fourth set to be within the distribution interval of the random noise, to obtain an updated fourth set, wherein the first iterative update is to add the second set to the fourth set;
[0028] The generative adversarial network generates second samples according to the target variables in the updated fourth set in sequence, and calculates the similarity between the second samples and the input sample according to the similarity index;
[0029] According to the magnitude of the similarity, the top M target variables with large similarity are selected from the updated fourth set to obtain the fifth set updated in this iteration.
[0030] Optionally, the multiple similarity indicators include: 2-norm, and the anomaly score of the input sample under each similarity indicator includes: the anomaly score of the input sample under the 2-norm;
[0031] The inverse number of the similarity between the generated sample and the input sample under each similarity index is used as the anomaly score of the input sample under each similarity index, including:
[0032] Calculating a 2-norm between the generated sample and the input sample to obtain a first similarity;
[0033] The inverse of the first similarity is used as an anomaly score of the input sample under the 2-norm.
[0034] Optionally, the multiple similarity indicators include: a feature matching loss item, and the anomaly score of the input sample under each similarity indicator includes: an anomaly score of the input sample under the feature matching loss item;
[0035] The inverse number of the similarity between the generated sample and the input sample under each similarity index is used as the anomaly score of the input sample under each similarity index, including:
[0036] Calculating a matching loss term between the feature of the input sample on the discriminator feature layer and the generated sample to obtain a second similarity;
[0037] The opposite number of the second similarity is used as the anomaly score of the input sample under the feature matching loss item.
[0038] Optionally, the multiple similarity indicators include: cosine similarity of high-level features of the discriminator, and the anomaly score of the input sample under each similarity indicator includes: the anomaly score of the input sample under the cosine similarity of high-level features of the discriminator;
[0039] The inverse number of the similarity between the generated sample and the input sample under each similarity index is used as the anomaly score of the input sample under each similarity index, including:
[0040] Calculating the cosine similarity of the input sample and the generated sample at the high-level features of the discriminator to obtain a third similarity;
[0041] The inverse of the third similarity is used as the anomaly score of the generated sample under the cosine similarity of the discriminator high-level features.
[0042] Optionally, obtaining a fault prediction result according to the final anomaly score includes:
[0043] When the final anomaly score is greater than an anomaly score threshold, determining that the fault prediction result is a fault;
[0044] When the final anomaly score is not greater than an anomaly score threshold, determining that the fault prediction result is healthy;
[0045] The anomaly score threshold is determined based on the anomaly score sequence of the healthy sample data used for training.
[0046] Optionally, time-frequency feature extraction is performed on the one-dimensional vibration data to obtain input samples, including:
[0047] Cutting the one-dimensional vibration data into sample bands according to a preset number of points;
[0048] Performing a one-dimensional continuous wavelet transform on the sample band, and taking a modulus of the transform result to obtain a two-dimensional time-frequency feature map;
[0049] The two-dimensional time-frequency feature map is sequentially subjected to bicubic interpolation resampling processing and normalization processing to obtain input samples.
[0050] The embodiments of the present application include the following advantages:
[0051] In an embodiment of the present application, time-frequency features are extracted from one-dimensional vibration data to obtain an input sample; based on multiple similarity indicators, a generated sample similar to the input sample is determined on the healthy sample data manifold learned by a generative adversarial network, and the anomaly score of the input sample under each similarity indicator is calculated; based on the anomaly score of the input sample under each of the similarity indicators, a final anomaly score is obtained, and a fault prediction result is obtained based on the final anomaly score, and the final anomaly score represents the degree to which the input sample deviates from the healthy sample data.
[0052] This method converts the mechanical fault prediction task into a single classification problem of anomaly detection. Based on the similarity index, the difference between the input sample and the generated sample that meets the distribution of healthy sample data is judged, and the degree of deviation of the input sample from the healthy sample data is measured, so as to obtain the mechanical fault prediction result. Since it is based on multiple similarity indicators, the generated sample similar to the input sample is determined from the healthy sample data manifold learned by the generative adversarial network, thus making full use of the anomaly detection ability of the generative adversarial network, and solving the problem of unstable fault prediction performance caused by the large randomness of the calculation process of a single similarity indicator; and the final anomaly score obtained according to the anomaly scores under multiple similarity indicators contains the anomaly information of different similarity indicators, so the final anomaly score is a score that better reflects the real fault problem. In this way, stable and accurate mechanical fault prediction is achieved. BRIEF DESCRIPTION OF THE DRAWINGS
[0053] In order to more clearly illustrate the technical solutions of the embodiments of the present application, the drawings required for use in the description of the embodiments of the present application will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present application. For ordinary technicians in this field, other drawings can be obtained based on these drawings without paying creative labor.
[0054] Figure 1 This is a flowchart of the steps of a method for predicting rotating machinery faults by fusing multiple similarity features provided in an embodiment of the present application;
[0055] Figure 2 It is a flowchart of the steps of a method for determining a generated sample similar to an input sample based on a distribution estimation algorithm provided by an embodiment of the present application;
[0056] Figure 3 This is an overall flow chart of a method for predicting rotating machinery faults by fusing multiple similarity features provided in an embodiment of the present application;
[0057] Figure 4 It is a structural schematic diagram of a rotating machinery fault prediction device using multi-similarity feature fusion provided in an embodiment of the present application. DETAILED DESCRIPTION
[0058] In order to make the above-mentioned purposes, features and advantages of the present application more obvious and understandable, the technical solutions in the embodiments of the present application will be clearly and completely described below in conjunction with the drawings in the embodiments of the present application. Obviously, the described embodiments are part of the embodiments of the present application, not all of the embodiments. Based on the embodiments in the present application, all other embodiments obtained by ordinary technicians in this field without creative work are within the scope of protection of this application.
[0059] Reference Figure 1 As shown, Figure 1 1 is a flowchart of the steps of a method for predicting rotating machinery faults by fusing multiple similarity features provided in an embodiment of the present application. Figure 1 As shown, a method for predicting rotating machinery faults by fusing multiple similarity features provided in an embodiment of the present application may include steps S110 to S130:
[0060] Step S110: extracting time-frequency features from one-dimensional vibration data to obtain input samples.
[0061] In the embodiment of the present application, the one-dimensional vibration data is a one-dimensional time domain data sequence of the collected rotating mechanical vibration, and the input sample is a two-dimensional time-frequency characteristic diagram. In order to better predict mechanical failures, the one-dimensional vibration data is converted into an input sample of a two-dimensional time-frequency characteristic diagram.
[0062] In an optional embodiment, time-frequency features are extracted from one-dimensional vibration data to obtain input samples, including: truncating the one-dimensional vibration data into sample bands according to a preset number of points; performing a one-dimensional continuous wavelet transform on the sample bands, and modulo the transform results to obtain a two-dimensional time-frequency feature map; and sequentially performing bicubic interpolation resampling processing and standardization processing on the two-dimensional time-frequency feature map to obtain input samples.
[0063] For example, 4096 (i.e., the number of preset points) points are sequentially intercepted from the one-dimensional vibration data as a sample band, and then a one-dimensional continuous wavelet transform with a wavelet scale of 32 is performed on each sample band, and the transformation result is modulo to obtain a 4096×32 two-dimensional time-frequency feature map, and then the two-dimensional time-frequency feature map is bicubic resampled to obtain a 32×32 feature map. Finally, the 32×32 feature map is standardized to obtain the input sample.
[0064] For example, the input sample is represented as:
[0065]
[0066] Among them, x i represents the input sample, min(·) and max(·) represent the minimum and maximum values of the input variable elements respectively.
[0067] In the embodiment of the present application, the value range of the input sample is limited to the interval [-0.5, 0.5] through standardization processing, so that the sample value falls within the output range (-1, 1) of the last tanh layer of the generator of the generative adversarial network. The reason for not directly scaling the sample value range to [-1, 1] is that the tanh layer is prone to gradient vanishing near -1 and 1, so the input sample obtained by this method can be better processed by the generative adversarial network.
[0068] Step S120: Based on multiple similarity indicators, a generated sample similar to the input sample is determined on the healthy sample data manifold learned by the generative adversarial network, and an abnormality score of the input sample under each similarity indicator is calculated based on the generated sample.
[0069] In the embodiment of the present application, multiple similarity indices refer to a method of using multiple similarities to describe the similarity between two sample data; the generative adversarial network is a probabilistic generation model, and the generative adversarial network is trained based on healthy sample data. After the training is completed, the generative adversarial network learns the distribution law of healthy sample data (i.e., the healthy sample data manifold), and can continuously generate samples that obey the distribution law of healthy sample data. Therefore, when predicting faults for input samples, it is only necessary to evaluate whether the input samples obey the distribution law of healthy sample data, or the probability of the input samples obeying the distribution law of healthy sample data.
[0070] Based on the similarity index, the similarity between the input sample and the sample randomly generated by the generative adversarial network based on the healthy sample data manifold is compared; the higher the similarity, the greater the possibility that the input sample is a healthy sample, and the lower the similarity, the greater the possibility that the input sample is a faulty sample. Since the samples randomly generated by the generative adversarial network are random, which leads to differences between different samples, the embodiment of the present application measures the similarity between the input sample and the sample randomly generated by the generative adversarial network based on multiple similarity indicators, thereby determining the generated sample similar to the input sample. Finally, based on the similarity between the input sample and the generated sample, the abnormal score of the input sample under each similarity indicator is calculated.
[0071] In an optional embodiment, based on multiple similarity indicators, a generated sample similar to the input sample is determined on the healthy sample data manifold learned by the generative adversarial network, and an abnormal score of the input sample under each similarity indicator is calculated according to the generated sample, including:
[0072] Using a distribution estimation algorithm, searching for samples in the latent space of the generative adversarial network based on each similarity index, and determining the sample with the greatest similarity to the input sample as a generated sample similar to the input sample;
[0073] The inverse number of the similarity between the generated sample and the input sample under each similarity index is used as the anomaly score of the input sample under each similarity index.
[0074] In the embodiment of the present application, in order to reduce the amount of calculation and randomness, the distribution estimation algorithm (EDA) is used to search for samples in the latent space of the generative adversarial network, where the latent space refers to the distribution interval obeyed by the random noise used to generate the sample. In order to reduce the influence of the randomness of the generated samples of the generative adversarial network, the random noise z is searched in the latent space to maximize the similarity of the generated samples under a certain similarity index, thereby obtaining a generated sample similar to the input sample; the opposite number of the similarity between the generated sample and the input sample under the similarity index is used as the abnormal score of the input sample under the similarity index.
[0075] For example, the generated samples similar to the input samples and the anomaly scores of the input samples under the similarity index are expressed as:
[0076]
[0077] in, represents the input sample, G(z) represents the sample generated by the generative adversarial network based on random noise z, Represents random noise searched in the latent space based on similarity metrics. It is possible to obtain generated samples similar to the input samples. SIM(·) represents the similarity under the similarity index. represents the anomaly score of the input sample under the similarity index, Represents the random noise that satisfies the maximum similarity between sample G(z) and the input sample D represents the latent space.
[0078] In addition, in order to match the input range of the generator of the generative adversarial network during training, the range of random noise z is limited to the central area with a radius of r0, so that the probability that the random variable satisfying the normal distribution N(0,I) falls in the area of the latent space D is 0.99, that is, the latent space is expressed as:
[0079] stP{X∈D}=0.99when X~N(0,I)
[0080] Among them, P represents the probability of a random variable X falling in the latent space D.
[0081] In the embodiment of the present application, the smooth transition of the latent space is utilized, and a distribution estimation algorithm is used to search for generated samples similar to the input samples in the latent space. At the same time, multiple similarity indicators are set to make the samples G(z) generated by the generative adversarial network similar to the input samples. The similarity under the corresponding similarity index is the largest, and then the anomaly score of the input sample under the similarity index is determined to improve the fault prediction effect based on the anomaly scores under multiple similarity indexes.
[0082] Step S130: Obtain a final anomaly score according to the anomaly score of the input sample under each of the similarity indicators, and obtain a fault prediction result according to the final anomaly score, wherein the final anomaly score represents the degree to which the input sample deviates from the healthy sample data.
[0083] In an embodiment of the present application, in order to make full use of the abnormal information of different similarity indicators, the abnormal scores of the input samples under different similarity indicators can be fused to obtain the final abnormal score. In a specific implementation, the abnormal scores under each similarity indicator are fused based on the abnormal detection algorithm to obtain the final abnormal score. Among them, the abnormal detection algorithms include: Isolation Forest (IF), K-Nearest Neighbor (KNN), One-Class Support Vector Machine (OCSVM), Minimum Covariance Determinant (MCD), taking the maximum value, taking the average, taking the median, and voting for the majority.
[0084] For example, the final anomaly score is expressed as:
[0085]
[0086] Where AD(·) represents the anomaly detection algorithm; is the anomaly score of the input sample under the first similarity index, SIM1 is the first similarity index, is the anomaly score of the input sample under the second similarity index, SIM2 is the second similarity index, and so on, there are s similarity indexes in total. is the anomaly score of the input sample under the sth similarity index, SIM s is the sth similarity index.
[0087] Finally, the fault prediction result is obtained based on the final anomaly score. The final anomaly score represents the degree to which the input sample deviates from the healthy sample data. If the final anomaly score is larger, it means that the input sample deviates more from the healthy sample data, and the probability that the input sample is a faulty sample is greater. If the final anomaly score is smaller, it means that the input sample is closer to the healthy sample data, and the probability that the input sample is a healthy sample is greater. Therefore, the fault prediction result can be obtained based on the final anomaly score.
[0088] Specifically, obtaining a fault prediction result according to the final anomaly score includes: determining that the fault prediction result is a fault when the final anomaly score is greater than an anomaly score threshold; determining that the fault prediction result is healthy when the final anomaly score is not greater than the anomaly score threshold; wherein the anomaly score threshold is determined according to an anomaly score sequence of healthy sample data used for training. In a specific implementation, the maximum anomaly score in the anomaly score sequence of the healthy sample data is determined as the anomaly score threshold.
[0089] In the embodiment of the present application, the mechanical fault prediction task is converted into a single classification problem of anomaly detection. The difference between the input sample and the generated sample that meets the distribution of healthy sample data is judged based on the similarity index, and the degree of deviation of the input sample from the healthy sample data is measured, so as to obtain the mechanical fault prediction result. Based on the generative adversarial network, the healthy sample data is used for training without the need for fault data to learn the distribution characteristics of the healthy data, so that the healthy sample data manifold learned by the generative adversarial network is used. In order to make full use of the network anomaly detection ability, based on multiple similarity indicators, the generated sample similar to the input sample is determined from the healthy sample data manifold learned by the generative adversarial network, which solves the problem of unstable fault prediction performance caused by the large randomness of the calculation process of a single similarity indicator. The final anomaly score is obtained according to the anomaly scores under different similarity indicators. The final anomaly score contains the abnormal information of different similarity indicators, so the final anomaly score is a score that better reflects the real fault problem. In this way, stable and accurate mechanical fault prediction is achieved.
[0090] The following is an explanation of the generative adversarial network and the training method of the generative adversarial network.
[0091] In an optional embodiment, the generative adversarial network includes a discriminator and a generator. The generative adversarial network is an unsupervised generative model that can simultaneously train the discriminator and the generator using the idea of dynamic game. The discriminator is used to identify the sample data generated by the generator that does not belong to the healthy sample data, and the generator is used to generate sample data close to the healthy sample data based on random noise to deceive the discriminator. Finally, through training, the generative adversarial network learns the mapping relationship from the distribution obeyed by random noise to the distribution of healthy sample data.
[0092] In addition, considering that the generative adversarial network is prone to gradient vanishing, mode collapse and instability during the training process, the generative adversarial network in the embodiment of the present application adopts the following improvements: 1) The generative adversarial network adopts Wasserstein distance instead of Jensen-Shannon divergence, which improves the stability of adversarial training in principle; 2) A gradient penalty term is introduced in the discriminator optimization objective function for optimization to overcome the problem of gradient vanishing during training; 3) Spectral normalization is adopted, and the spectral norm of the weight matrix is used to normalize the weights of each layer of the generative adversarial network to meet the Lipschitz coefficient constraint of the discriminator, and global regularization is applied to the discriminator; 4) Feature matching is introduced to match the features of the generator on the discriminator feature layer with the healthy sample data when training the generator to improve the training stability.
[0093] Specifically, the generative adversarial network is trained according to the following steps A1 to A5:
[0094] Step A1: The generator generates sample data according to random noise.
[0095] Among them, the random noise obeys the normal distribution N(0,1), and the generator generates sample data based on the random noise, so as to transform the random noise distribution into a probability distribution close to the healthy sample data (ie, training data) through training learning.
[0096] Step A2: Based on the generated sample data and the healthy sample data, a discriminator optimization objective function is constructed, wherein the discriminator optimization objective function includes: the score of the discriminator on the healthy sample data, the score of the discriminator on the generated sample data, and a gradient penalty term.
[0097] For example, the discriminator optimization objective function is expressed as:
[0098]
[0099] in, It means that it is implicitly defined between x and p. data and The uniform sampling distribution along the straight line between represents the distribution of healthy sample data, Is z~p z Implicit definition, z~p z represents the distribution of random noise; E represents expectation; represents the score of the discriminator on the healthy sample data; Represents the score of the discriminator on the generated sample data; represents the gradient penalty term, λ1 represents the weight parameter of the gradient penalty term, and min represents the minimum value.
[0100] Step A3: Based on the generated sample data and the healthy sample data, a generator optimization objective function is constructed, wherein the generator optimization objective function includes: a score of the generated sample data by the discriminator, and a feature matching loss term between the features of the generated sample data on the discriminator feature layer and the features of the healthy sample data.
[0101] For example, the generator optimization objective function is expressed as:
[0102]
[0103] in, represents the score of the discriminator on the generated sample data; ||f D (x)-f D (G(z))||2 represents the feature matching loss term between the generated sample data on the discriminator feature layer and the healthy sample data, λ2 represents the weight parameter of the feature matching loss term, and f D (x) represents the feature of the healthy sample on the discriminator feature layer, f D (G(z)) represents the features of the generated sample data on the discriminator feature layer, and ||·||2 represents the 2-norm calculation.
[0104] Step A4: fix the generator, and update the parameters of the discriminator according to the discriminator optimization objective function, and fix the discriminator, and update the parameters of the generator according to the generator optimization objective function.
[0105] Step A5: After the training end condition is met, a trained generative adversarial network is obtained, wherein the trained generative adversarial network has learned the mapping relationship from the distribution obeyed by random noise to the distribution of healthy sample data.
[0106] In the embodiment of the present application, the generator and the discriminator update parameters according to their respective optimization objective functions and confront each other, so as to learn the distribution law of the training data (i.e., healthy sample data) under unsupervised conditions. The trained generative adversarial network learns the mapping relationship from the distribution obeyed by random noise to the distribution of healthy sample data. When fault detection is performed based on the trained generative adversarial network in the future, the generative adversarial network can continuously generate generated samples that meet the distribution of healthy sample data, thereby evaluating the similarity difference between the generated samples and the input samples, and realizing the fault judgment of the input samples.
[0107] The following describes the process of searching for samples in the latent space of the generative adversarial network based on similarity indicators, as well as the calculation of anomaly scores under each similarity indicator.
[0108] In an optional embodiment, if Figure 2 As shown, a distribution estimation algorithm is used, and a distribution estimation algorithm is used to search for samples in the potential space of the generative adversarial network based on each similarity index, and the sample with the greatest similarity to the input sample is determined as a generated sample similar to the input sample, including steps B1 to B6:
[0109] Step B1: Sample N×M sampling points from the distribution of random noise to form a first set.
[0110] Specifically, the distribution of random noise is normal distribution N(0,I), and N×M sampling points are randomly sampled from the normal distribution N(0,I) to obtain the first set. (0) It is expressed as:
[0111]
[0112] in, represents the first sampling point, Indicates the N×Mth sampling point, n c Represents the distribution of random noise.
[0113] Step B2: The generative adversarial network generates a first sample according to the sampling points in the first set in turn, and calculates the similarity between the first sample and the input sample according to a similarity index.
[0114] For example, the similarity between the first sample and the input sample is calculated according to the similarity index and expressed as:
[0115]
[0116] in, Represents the similarity between the jth first sample and the input sample; Represents the generative adversarial network based on the jth sampling point The jth first sample generated; SIM(·) represents the similarity under the similarity index.
[0117] Step B3: According to the magnitude of the similarity, the first M sampling points with the largest similarity are selected from the first set to form a second set.
[0118] For example, the second set z' (0) It is expressed as:
[0119]
[0120] in, represents the first sampling point in the second set, represents the Mth sampling point in the second set.
[0121] Step B4: Sample N sampling points from the distribution of the random noise to form a third set.
[0122] For example, the third set is represented as:
[0123]
[0124] Among them, e1 represents the first sampling point in the third set, e N represents the Nth sampling point in the third set.
[0125] Step B5: Calculate the target variable according to the second set and the third set to obtain a fourth set, iteratively update the fourth set, and obtain a fifth set according to the updated fourth set.
[0126] In this embodiment of the present application, the fourth set is updated T times by iteration. For the tth iteration update, the fourth set z calculated is (t) It is expressed as:
[0127]
[0128] in, Indicates z (t) The first target variable in Indicates z (t) The N×Mth target variable in .
[0129] Specifically, z (t) The jth target variable in It is expressed as:
[0130]
[0131] in, represents the fifth set z' obtained by the t-1th iteration update (t-1) The sampling points; (jmodN)+1 represents the (jmodN)+1th sampling point in the third set; Indicates rounding down; mod indicates remainder operation.
[0132] Specifically, the target variable is calculated according to the second set and the third set to obtain a fourth set, and the fourth set is iteratively updated to obtain a fifth set according to the updated fourth set, including steps B51 to B53:
[0133] Step B51: Add the fifth set obtained from the last iterative update to the fourth set, and constrain the target variables in the fourth set to be within the distribution interval of the random noise to obtain an updated fourth set, wherein the first iterative update is to add the second set to the fourth set.
[0134] For example, for the t-th iteration update, the fifth set z' obtained by the t-1-th iteration update (t-1) Add a fourth set to the calculation In the fourth set, we get the updated If it is the first iteration update, this is to set the second set z' (0) Add to the calculated fourth set. And, for the updated fourth set The target variable in is constrained, when the fourth set is updated When the target variable in does not fall into the latent space D, the target variable is multiplied by a certain coefficient to make it fall within the boundary of the latent space D.
[0135] Step B52: The generative adversarial network generates second samples according to the target variables in the updated fourth set in turn, and calculates the similarity between the second samples and the input sample according to the similarity index.
[0136] For example, the similarity between the second sample and the input sample is calculated according to the similarity index and expressed as:
[0137]
[0138] in, Indicates the similarity between the jth second sample and the input sample; Represents the fourth set z that the generative adversarial network updates according to (t) The jth target variable in The j-th second sample generated; SIM(·) represents the similarity under the similarity index.
[0139] Step B53: Select the top M target variables with large similarity from the updated fourth set according to the similarity to obtain the fifth set updated in this iteration.
[0140] Specifically, according to the size of the similarity, from the updated fourth set Select the largest M target variables from the set to form the fifth set obtained by this iteration update. For example, the fifth set z' (t) It is expressed as:
[0141]
[0142] in, Represents the fifth set z' (t) The first target variable in Represents the fifth set z' (t) is the Mth target variable in .
[0143] Step B6: Determine, based on the fifth set obtained by the last iterative update, a generated sample similar to the input sample.
[0144] Specifically, the generative adversarial network generates the third sample according to the target variable in the fifth set obtained by the last iterative update, calculates the similarity between the third sample and the input sample according to the similarity index, and finally determines the third sample with the greatest similarity as the generated sample similar to the input sample.
[0145] For example, the ordinal number k of the target variable corresponding to the maximum similarity is expressed as:
[0146]
[0147] in, represents the similarity between the jth third sample and the input sample, Indicates the sequence number of the target variable corresponding to the maximum similarity.
[0148] Finally, the target variable z SIM , and the anomaly score A of the input sample under the similarity index SIM (x test ) are respectively expressed as:
[0149]
[0150] in, Indicates the similarity between the generated sample and the input sample under the similarity index.
[0151] In the embodiment of the present application, in order to reduce the amount of calculation and randomness, the latent variable corresponding to the input sample is searched in the latent space of the generative adversarial network, so that the similarity SIM(·) between the sample generated by the latent variable after the generator and the input sample under the similarity index is as large as possible, so as to find a point with the highest corresponding similarity with the input sample on the healthy sample data manifold learned by the generative adversarial network. Based on multiple similarity indicators, the difference between the input sample and the healthy sample data distribution is comprehensively measured, and the similarity under multiple similarity indicators is used as an evaluation index for input sample fault prediction, thereby achieving accurate fault prediction.
[0152] In an embodiment of the present application, different anomaly scores can be obtained by setting different similarity indicators and using a distribution estimation algorithm. The multiple similarity indicators include: 2-norm, feature matching loss term and cosine similarity of high-level features of the discriminator, 1-norm, structural similarity (Structural Similarity, SSIM) and other indicators that can measure the similarity of two variables.
[0153] In an optional embodiment, the multiple similarity indicators include: a 2-norm, and the anomaly score of the input sample under each similarity indicator includes: an anomaly score of the input sample under the 2-norm.
[0154] The method further comprises: calculating a 2-norm between the generated sample and the input sample to obtain a first similarity; and taking the inverse of the first similarity as the anomaly score of the input sample under the 2-norm.
[0155] For example, when the 2-norm is the similarity index, the similarity between the input sample and the generated sample is It is expressed as:
[0156]
[0157] Therefore, the anomaly score of the input sample under the 2-norm, and the generated sample similar to the input sample are expressed as:
[0158]
[0159] in, represents the anomaly score of the input sample under the 2-norm, represents random noise determined in the latent space under the 2-norm, according to The generated samples are similar to the input samples under the 2-norm, and ||·||2 represents the 2-norm calculation.
[0160] In an optional embodiment, the multiple similarity indicators include: a feature matching loss item, and the anomaly score of the input sample under each similarity indicator includes: the anomaly score of the input sample under the feature matching loss item.
[0161] The method further comprises: calculating a matching loss item between a feature of the input sample on a discriminator feature layer and the generated sample to obtain a second similarity; and taking the opposite number of the second similarity as the anomaly score of the input sample under the feature matching loss item.
[0162] In the embodiment of the present application, it is considered that the feature matching technology forces the generated samples to match the high-level features of the real samples (i.e., healthy sample data) on the discriminator when training the generator. When the training is completed, the high-level features of the generated samples on the discriminator are slightly different from the healthy sample data, while the faulty samples have not participated in the feature matching training, and the features on the discriminator are generally quite different from the healthy sample data. Therefore, the feature matching loss term is used as the similarity indicator.
[0163] For example, when the feature matching loss term is a similarity indicator, the similarity between the input sample and the generated sample It is expressed as:
[0164]
[0165] in, represents the features of the input sample on the discriminator feature layer, f D (G(z)) represents the features of the generated sample data on the discriminator feature layer.
[0166] Therefore, the anomaly score of the input sample under the feature matching loss, and the generated sample similar to the input sample are expressed as:
[0167]
[0168] in, represents the anomaly score of the input sample under the 2-norm feature matching loss, represents the random noise determined in the latent space under the 2-norm feature matching loss, according to Get generated samples that are similar to the input samples under the 2-norm feature matching loss.
[0169] In an optional embodiment, the multiple similarity indicators include: cosine similarity of discriminator high-level features, and the anomaly score of the input sample under each similarity indicator includes: the anomaly score of the input sample under the cosine similarity of the discriminator high-level features.
[0170] The method uses the inverse of the similarity between the generated sample and the input sample under each similarity index as the anomaly score of the input sample under each similarity index, including: calculating the cosine similarity of the input sample and the generated sample at the high-level features of the discriminator to obtain a third similarity; and using the inverse of the third similarity as the anomaly score of the generated sample under the cosine similarity of the high-level features of the discriminator.
[0171] For example, when the cosine similarity of the discriminator's high-level features is used as the similarity metric, the similarity between the input sample and the generated sample is It is expressed as:
[0172]
[0173] Among them, cosSIM represents cosine similarity, represents the features of the input sample on the discriminator feature layer, f D (G(z)) represents the features of the generated sample on the discriminator feature layer.
[0174] Therefore, the anomaly score of the input sample under the cosine similarity of the discriminator's high-level features, and the generated sample similar to the input sample are expressed as:
[0175]
[0176] in, represents the anomaly score of the input sample under the cosine similarity of the high-level features of the discriminator, represents the random noise determined in the latent space under the cosine similarity of the high-level features of the discriminator, according to Get generated samples that are similar to the input samples under the cosine similarity of the discriminator's high-level features.
[0177] For example, Figure 3This is an overall flow chart of a method for predicting rotating machinery faults by fusion of multiple similarity features provided in an embodiment of the present application. Specifically, time-frequency features are extracted from one-dimensional vibration data to obtain input samples; a distribution estimation algorithm is used to search for samples in the latent space of the generative adversarial network based on each similarity index, and the sample with the greatest similarity to the input sample is determined as a generated sample similar to the input sample; at the same time, the opposite number of the similarity between the generated sample and the input sample under each similarity index is used as the anomaly score of the input sample under each similarity index. The anomaly scores of the input samples under each similarity index are fused to obtain a final anomaly score; finally, a fault prediction result is obtained based on the final anomaly score. Specifically, the final anomaly score characterizes the degree to which the input sample deviates from the healthy sample data. When the final anomaly score is greater than the anomaly score threshold, the fault prediction result is determined to be a fault; when the final anomaly score is not greater than the anomaly score threshold, the fault prediction result is determined to be healthy.
[0178] This method converts the mechanical fault prediction task into a single classification problem of anomaly detection. Based on the similarity index, the difference between the input sample and the generated sample that meets the distribution of healthy sample data is judged, and the degree of deviation of the input sample from the healthy sample data is measured, so as to obtain the mechanical fault prediction result. Since it is based on multiple similarity indicators, the generated sample similar to the input sample is determined from the healthy sample data manifold learned by the generative adversarial network, thus making full use of the anomaly detection ability of the generative adversarial network, and solving the problem of unstable fault prediction performance caused by the large randomness of the calculation process of a single similarity indicator; and the final anomaly score obtained according to the anomaly scores under multiple similarity indicators contains the anomaly information of different similarity indicators, so the final anomaly score is a score that better reflects the real fault problem. In this way, stable and accurate mechanical fault prediction is achieved.
[0179] The present application also provides a multi-similarity feature fusion rotating machinery fault prediction device, referring to Figure 4 As shown, Figure 4 : is a schematic diagram of the structure of a multi-similarity feature fusion rotating machinery fault prediction device provided in an embodiment of the present application, the device comprising:
[0180] The feature extraction module 410 is used to extract time-frequency features from the one-dimensional vibration data to obtain input samples;
[0181] A sample determination module 420 is used to determine, based on multiple similarity indicators, generated samples similar to the input sample on the healthy sample data manifold learned by the generative adversarial network, and calculate the abnormality score of the input sample under each similarity indicator according to the generated samples;
[0182] The prediction result module 430 is used to obtain a final anomaly score according to the anomaly score of the input sample under each similarity index, and obtain a fault prediction result according to the final anomaly score, wherein the final anomaly score represents the degree to which the input sample deviates from the healthy sample data.
[0183] In an optional embodiment, the generative adversarial network includes a discriminator and a generator, and the device includes a training module for training the generative adversarial network, and the training module includes:
[0184] A sample generation module, used for the generator to generate sample data according to random noise;
[0185] A first construction module is used to construct a discriminator optimization objective function according to the generated sample data and the healthy sample data, wherein the discriminator optimization objective function includes: a score of the discriminator on the healthy sample data, a score of the discriminator on the generated sample data, and a gradient penalty term;
[0186] A second construction module is used to construct a generator optimization objective function according to the generated sample data and the healthy sample data, wherein the generator optimization objective function includes: a score of the generated sample data by the discriminator, and a feature matching loss term between the features of the generated sample data on the discriminator feature layer and the features of the healthy sample data;
[0187] A parameter updating module, used to fix the generator and update the parameters of the discriminator according to the discriminator optimization objective function, and to fix the discriminator and update the parameters of the generator according to the generator optimization objective function;
[0188] The training end module is used to obtain a trained generative adversarial network after the training end condition is met. The trained generative adversarial network has learned the mapping relationship from the distribution obeyed by random noise to the distribution of healthy sample data.
[0189] In an optional embodiment, the sample determination module includes:
[0190] A sample search module, configured to use a distribution estimation algorithm to search for samples in the latent space of the generative adversarial network based on each similarity index, and determine the sample with the greatest similarity to the input sample as a generated sample similar to the input sample;
[0191] The anomaly score module is used to use the inverse of the similarity between the generated sample and the input sample under each similarity index as the anomaly score of the input sample under each similarity index.
[0192] In an optional embodiment, the sample search module includes:
[0193] A first sampling module, used for sampling N×M sampling points from the distribution of random noise to form a first set;
[0194] A first calculation module, configured for the generative adversarial network to generate first samples according to the sampling points in the first set in sequence, and to calculate the similarity between the first samples and the input sample according to a similarity index;
[0195] A first selection module, configured to select first M sampling points with large similarity from the first set according to the similarity to form a second set;
[0196] A second sampling module, used for sampling N sampling points from the distribution of the random noise to form a third set;
[0197] an iterative updating module, configured to calculate a target variable according to the second set and the third set to obtain a fourth set, and iteratively update the fourth set to obtain a fifth set according to the updated fourth set;
[0198] The first determination module is used to determine a generated sample similar to the input sample according to the fifth set obtained by the last iterative update.
[0199] In an optional embodiment, the iterative update module includes:
[0200] A first updating module, configured to add the fifth set obtained in the last iterative update to the fourth set, and constrain the target variable in the fourth set to be within the distribution interval of the random noise, to obtain an updated fourth set, wherein the first iterative update is to add the second set to the fourth set;
[0201] A second calculation module, configured for the generative adversarial network to sequentially generate second samples according to the target variables in the updated fourth set, and calculate the similarity between the second samples and the input sample according to a similarity index;
[0202] The second selection module is used to select the top M target variables with large similarity from the updated fourth set according to the size of the similarity, so as to obtain the fifth set updated in this iteration.
[0203] In an optional embodiment, the multiple similarity indicators include: 2-norm, the anomaly score of the input sample under each similarity indicator includes: the anomaly score of the input sample under the 2-norm; the anomaly score module includes:
[0204] A first similarity module, used to calculate the 2-norm between the generated sample and the input sample to obtain a first similarity;
[0205] The first anomaly score submodule is used to use the opposite of the first similarity as the anomaly score of the input sample under the 2-norm.
[0206] In an optional embodiment, the multiple similarity indicators include: a feature matching loss item, and the anomaly score of the input sample under each similarity indicator includes: an anomaly score of the input sample under the feature matching loss item; the anomaly score module includes:
[0207] A second similarity module, used to calculate a matching loss term between the feature of the input sample on the discriminator feature layer and the generated sample to obtain a second similarity;
[0208] The second anomaly score submodule is used to use the opposite number of the second similarity as the anomaly score of the input sample under the feature matching loss item.
[0209] In an optional embodiment, the multiple similarity indicators include: cosine similarity of high-level features of the discriminator, and the anomaly score of the input sample under each similarity indicator includes: the anomaly score of the input sample under the cosine similarity of the high-level features of the discriminator; the anomaly score module includes:
[0210] A third similarity module, used to calculate the cosine similarity of the input sample and the generated sample at the high-level features of the discriminator to obtain a third similarity;
[0211] The third anomaly score submodule is used to use the inverse of the third similarity as the anomaly score of the generated sample under the cosine similarity of the discriminator high-level features.
[0212] In an optional embodiment, the prediction result module includes:
[0213] A second determination module, configured to determine that the fault prediction result is a fault when the final anomaly score is greater than an anomaly score threshold;
[0214] A third determination module, configured to determine that the fault prediction result is healthy if the final anomaly score is not greater than an anomaly score threshold;
[0215] The anomaly score threshold is determined based on the anomaly score sequence of the healthy sample data used for training.
[0216] In an optional embodiment, the feature extraction module includes:
[0217] An interception module, used for intercepting the one-dimensional vibration data into sample bands according to a preset number of points;
[0218] A transformation module, used for performing a one-dimensional continuous wavelet transformation on the sample band, and performing modulo operation on the transformation result to obtain a two-dimensional time-frequency feature map;
[0219] The processing module is used to perform bicubic interpolation resampling processing and standardization processing on the two-dimensional time-frequency feature map in sequence to obtain input samples.
[0220] The various embodiments in this specification are described in a progressive manner, and each embodiment focuses on the differences from other embodiments. The same or similar parts between the various embodiments can be referenced to each other.
[0221] The embodiments of the present application are described with reference to the flowcharts and / or block diagrams of the methods and devices according to the embodiments of the present application. It should be understood that each process and / or block in the flowchart and / or block diagram, as well as the combination of the processes and / or blocks in the flowchart and / or block diagram, can be implemented by computer program instructions. These computer program instructions can be provided to a processor of a general-purpose computer, a special-purpose computer, an embedded processor, or other programmable data processing terminal device to generate a machine, so that the instructions executed by the processor of the computer or other programmable data processing terminal device generate instructions for implementing the processes in the flowchart and / or block diagram. Figure 1 A process or multiple processes and / or boxes Figure 1 A device that provides the functions specified in a block or multiple blocks.
[0222] These computer program instructions may also be stored in a computer-readable memory capable of directing a computer or other programmable data processing terminal device to operate in a specific manner, so that the instructions stored in the computer-readable memory produce a manufactured product including an instruction device, which implements the process Figure 1 A process or multiple processes and / or boxes Figure 1 A function specified in one or more boxes.
[0223] These computer program instructions can also be loaded onto a computer or other programmable data processing terminal device so that a series of operating steps are executed on the computer or other programmable terminal device to produce a computer-implemented process, thereby providing instructions for executing on the computer or other programmable terminal device to implement the process. Figure 1 A process or multiple processes and / or boxes Figure 1 The steps for the functions specified in one or more boxes.
[0224] Although the preferred embodiments of the present application have been described, those skilled in the art may make additional changes and modifications to these embodiments once they have learned the basic creative concept. Therefore, the appended claims are intended to be interpreted as including the preferred embodiments and all changes and modifications that fall within the scope of the embodiments of the present application.
[0225] Finally, it should be noted that, in this article, relational terms such as first and second, etc. are only used to distinguish one entity or operation from another entity or operation, and do not necessarily require or imply any such actual relationship or order between these entities or operations. Moreover, the terms "include", "comprise" or any other variants thereof are intended to cover non-exclusive inclusion, so that a process, method, article or terminal device including a series of elements includes not only those elements, but also other elements not explicitly listed, or also includes elements inherent to such process, method, article or terminal device. In the absence of further restrictions, the elements defined by the sentence "comprise a ..." do not exclude the existence of other identical elements in the process, method, article or terminal device including the elements.
[0226] The above is a detailed introduction to a multi-similarity feature fusion rotating machinery fault prediction method provided by the present application. This article uses specific examples to illustrate the principles and implementation methods of the present application. The description of the above embodiments is only used to help understand the method of the present application and its core idea; at the same time, for general technical personnel in this field, based on the ideas of the present application, there will be changes in the specific implementation methods and application scope. In summary, the content of this specification should not be understood as a limitation on the present application.
Claims
1. A method for predicting rotating machinery faults by fusion of multiple similarity features, characterized in that: The method comprises: Extract time-frequency features from one-dimensional vibration data to obtain input samples; Based on multiple similarity indicators, determining generated samples similar to the input sample on the healthy sample data manifold learned by the generative adversarial network, and calculating the abnormality score of the input sample under each similarity indicator according to the generated samples; A final anomaly score is obtained according to the anomaly score of the input sample under each of the similarity indicators, and a fault prediction result is obtained according to the final anomaly score, wherein the final anomaly score represents the degree to which the input sample deviates from the healthy sample data.
2. The method for predicting rotating machinery faults by fusion of multiple similarity features according to claim 1 is characterized in that: The generative adversarial network includes a discriminator and a generator, and the generative adversarial network is trained according to the following steps: The generator generates sample data according to random noise; According to the generated sample data and the healthy sample data, construct a discriminator optimization objective function, wherein the discriminator optimization objective function includes: a score of the discriminator on the healthy sample data, a score of the discriminator on the generated sample data, and a gradient penalty term; According to the generated sample data and the healthy sample data, a generator optimization objective function is constructed, wherein the generator optimization objective function includes: a score of the generated sample data by the discriminator, and a matching loss term between features of the generated sample data on the discriminator feature layer and features of the healthy sample data; Fixing the generator, updating the parameters of the discriminator according to the discriminator optimization objective function, and fixing the discriminator, updating the parameters of the generator according to the generator optimization objective function; After the training end condition is met, a trained generative adversarial network is obtained, and the trained generative adversarial network learns the mapping relationship from the distribution obeyed by random noise to the distribution of healthy sample data.
3. The method for predicting rotating machinery faults by fusion of multiple similarity features according to claim 1 is characterized in that: Based on multiple similarity indicators, a generated sample similar to the input sample is determined on the healthy sample data manifold learned by the generative adversarial network, and an abnormal score of the input sample under each similarity indicator is calculated according to the generated sample, including: Using a distribution estimation algorithm, searching for samples in the latent space of the generative adversarial network based on each similarity index, and determining the sample with the greatest similarity to the input sample as a generated sample similar to the input sample; The inverse number of the similarity between the generated sample and the input sample under each similarity index is used as the anomaly score of the input sample under each similarity index.
4. The method for predicting rotating machinery faults by fusion of multiple similarity features according to claim 3 is characterized in that: Using a distribution estimation algorithm, searching for samples in the latent space of the generative adversarial network based on each similarity index, and determining the sample with the greatest similarity to the input sample as a generated sample similar to the input sample, including: Sample N×M sampling points from the distribution of random noise to form a first set; The generative adversarial network generates first samples according to the sampling points in the first set in sequence, and calculates the similarity between the first sample and the input sample according to the similarity index; According to the magnitude of the similarity, first M sampling points with large similarity are selected from the first set to form a second set; Sampling N sampling points from the distribution of the random noise to form a third set; Calculating the target variable according to the second set and the third set to obtain a fourth set, iteratively updating the fourth set, and obtaining a fifth set according to the updated fourth set; According to the fifth set obtained by the last iterative update, a generated sample similar to the input sample is determined.
5. The method for predicting rotating machinery faults by fusion of multiple similarity features according to claim 4 is characterized in that: Calculating the target variable according to the second set and the third set to obtain a fourth set, iteratively updating the fourth set, and obtaining a fifth set according to the updated fourth set, including: Adding the fifth set obtained by the last iterative update to the fourth set, and constraining the target variable in the fourth set to be within the distribution interval of the random noise, to obtain an updated fourth set, wherein the first iterative update is to add the second set to the fourth set; The generative adversarial network generates second samples according to the target variables in the updated fourth set in sequence, and calculates the similarity between the second samples and the input sample according to the similarity index; According to the magnitude of the similarity, the top M target variables with large similarity are selected from the updated fourth set to obtain the fifth set updated in this iteration.
6. The method for predicting rotating machinery faults by fusion of multiple similarity features according to claim 3 is characterized in that: The multiple similarity indicators include: 2-norm, and the anomaly score of the input sample under each similarity indicator includes: the anomaly score of the input sample under the 2-norm; The inverse number of the similarity between the generated sample and the input sample under each similarity index is used as the anomaly score of the input sample under each similarity index, including: Calculating a 2-norm between the generated sample and the input sample to obtain a first similarity; The inverse of the first similarity is used as an anomaly score of the input sample under the 2-norm.
7. The method for predicting rotating machinery faults by fusion of multiple similarity features according to claim 3 is characterized in that: The multiple similarity indicators include: a feature matching loss item, and the anomaly score of the input sample under each similarity indicator includes: an anomaly score of the input sample under the feature matching loss item; The inverse number of the similarity between the generated sample and the input sample under each similarity index is used as the anomaly score of the input sample under each similarity index, including: Calculating a matching loss term between the feature of the input sample on the discriminator feature layer and the generated sample to obtain a second similarity; The opposite number of the second similarity is used as the anomaly score of the input sample under the feature matching loss item.
8. The method for predicting rotating machinery faults by fusion of multiple similarity features according to claim 3 is characterized in that: The plurality of similarity indicators include: cosine similarity of high-level features of the discriminator, and the anomaly score of the input sample under each similarity indicator includes: an anomaly score of the input sample under the cosine similarity of high-level features of the discriminator; The inverse number of the similarity between the generated sample and the input sample under each similarity index is used as the anomaly score of the input sample under each similarity index, including: Calculating the cosine similarity of the input sample and the generated sample at the high-level features of the discriminator to obtain a third similarity; The inverse of the third similarity is used as the anomaly score of the generated sample under the cosine similarity of the discriminator high-level features.
9. The method for predicting rotating machinery faults by fusion of multiple similarity features according to claim 1, characterized in that: A fault prediction result is obtained according to the final anomaly score, including: When the final anomaly score is greater than an anomaly score threshold, determining that the fault prediction result is a fault; When the final anomaly score is not greater than an anomaly score threshold, determining that the fault prediction result is healthy; The anomaly score threshold is determined based on the anomaly score sequence of the healthy sample data used for training.
10. The method for predicting rotating machinery faults by fusion of multiple similarity features according to claim 1, characterized in that: Extract time-frequency features from one-dimensional vibration data to obtain input samples, including: Cutting the one-dimensional vibration data into sample bands according to a preset number of points; Performing a one-dimensional continuous wavelet transform on the sample band, and taking a modulus of the transform result to obtain a two-dimensional time-frequency feature map; The two-dimensional time-frequency feature map is sequentially subjected to bicubic interpolation resampling processing and normalization processing to obtain input samples.
Citation Information
Patent Citations
Aero-engine fault diagnosis method based on variational self-encoding generative adversarial network
CN115587290A
Rotary machinery intelligent fault diagnosis method and system oriented to sensor data abnormity
CN117407784A
Mechanical fault data generation method based on prior knowledge fusion generative adversarial network
CN117992830A
Cited By
High-precision and high-reliability harmonic reducer system testing method
CN120277474A
Cross-domain fault diagnosis method for coal mine underground rotary mechanical equipment bearing
CN120763687A
