Construction method and system of rolling bearing anomaly detection model

Abstract

The invention belongs to the technical field related to bearing abnormity, and discloses a construction method and system of a rolling bearing abnormity detection model. The method comprises the following steps: collecting vibration signals of bearing vibration to form an original vibration signal set A, and endowing a fault label with the original vibration signal set A; performing data processing on the original vibration signal set by using the data generation model to generate a new vibration signal, and forming a new vibration signal set B; the vibration signals in the vibration signal set B and the vibration signals in the original vibration signal set A are spliced to form spliced vibration signals, the fault labels of the spliced vibration signals are the same as the vibration signals in the vibration signal set B, and the spliced vibration signals are added to the original vibration signal set A to form a vibration signal set C; the vibration signal set C is used for training an anomaly detection model to obtain a needed anomaly detection model of the rolling bearing, the input of the anomaly detection model is vibration signals, and the output of the anomaly detection model is a fault label. According to the invention, the technical problem of difficulty in bearing abnormity detection is solved.

Description

Technical Field

[0001] The present invention belongs to the technical field related to bearing anomalies, and more specifically, relates to a method and system for constructing a rolling bearing anomaly detection model. Background Art

[0002] As the core components of large-scale machinery in the fields of petrochemical, wind power generation, high-speed railway, industrial machine tools, aerospace, etc., rolling bearings play a key role in the stable operation of rotating machinery. However, under long-term high-speed rotation, rolling bearings are very prone to failure, threatening the safety of machinery and workers, and putting machinery and workers in a potential safety hazard. In addition, the consequences of accidents caused by rolling bearing failures are serious, causing unpredictable casualties and economic losses. Therefore, it is of great significance to timely detect the faults of rolling bearings in use and ensure the safe operation of bearing machinery.

[0003] In addition, missed detection during bearing fault detection may cause damage to mechanical facilities, and in serious cases may cause accidents. Therefore, improving the accuracy of fault detection and avoiding misdiagnosis and missed diagnosis are extremely important to ensure the normal operation of rotating machinery systems. Summary of the invention

[0004] In view of the above defects or improvement needs of the prior art, the present invention provides a method and system for constructing a rolling bearing anomaly detection model to solve the problem of small number of data samples in the anomaly model training process in the prior art.

[0005] To achieve the above object, according to one aspect of the present invention, a method for constructing a rolling bearing abnormality detection model is provided, the method comprising the following steps:

[0006] The vibration signals of the bearing vibration are collected to form an original vibration signal set A and a fault label is assigned to each vibration signal in the original vibration signal set; the data generation model is used to perform data processing on each vibration signal in the original vibration signal set to generate a new vibration signal, thereby forming a new vibration signal set B; the vibration signal set B is spliced ​​with the vibration signals in the original vibration signal set A to form a spliced ​​vibration signal, the fault label of the spliced ​​vibration signal is the same as the vibration signal in the vibration signal set B, and the spliced ​​vibration signal is added to the original vibration signal set A to form a vibration signal set C; wherein, the fault labels remain unchanged during the process of performing data processing to generate a new vibration signal and splicing the vibration signal using the data generation model;

[0007] The anomaly detection model is trained using the vibration signal set C to obtain the required anomaly detection model for the rolling bearing. The input of the anomaly detection model is the vibration signal, and the output is the fault label.

[0008] Further preferably, the original vibration signal set A includes a normal vibration signal set A1 and an abnormal vibration signal set A2, the fault label of the normal vibration signal set A1 is normal, and the fault label of the abnormal vibration signal in the abnormal vibration signal set A2 is set according to the abnormal type of the bearing when the abnormal vibration signal is collected.

[0009] Further preferably, when the data generation model generates a new vibration signal, it generates a new abnormal vibration signal using the abnormal vibration signal in the abnormal vibration signal set A2.

[0010] Further preferably, the vibration signal set B is spliced ​​with the vibration signals in the original vibration signal set A by splicing the abnormal vibration signals in the vibration signal set B with the normal signals in the normal vibration signal set A1.

[0011] Further preferably, the relationship of the spliced ​​vibration signal is as follows:

[0012]

[0013] Where t = 1, 2, ..., N r +N g is the time index of the concatenated vibration signal Y(t), X r (t) is the original vibration signal, Nr is the length of the original vibration signal, Xg(t) is the new vibration signal in the vibration signal set B, N g is the length of the new vibration signal generated, and t=p is the random insertion point.

[0014] Further preferably, the data generation model includes a frequency domain conversion unit, a spectrum filtering unit, and a signal generation unit, wherein the frequency domain conversion unit is used to perform discrete Fourier transform on the time series data of the vibration signal and convert it into a frequency domain representation, while introducing a slight disturbance to the input vibration signal in the frequency domain to generate a new frequency domain vibration signal; the spectrum filtering unit is used to filter the new frequency domain vibration signal generated by the frequency domain conversion unit; and the signal generation unit converts the filtered new frequency domain vibration signal into a time domain representation through an inverse Fourier transform.

[0015] Further preferably, the anomaly detection model adopts a time domain convolutional network model.

[0016] Further preferably, the loss function of the anomaly detection model is as follows:

[0017]

[0018] Among them, ψ c is the prediction function of the positive sample, which is used to classify the abnormal samples, y + =1 and y -= -1 is used to classify abnormal samples and normal samples, c∈R H It is the feature space The center of the hypersphere on is the expectation of all samples, S is the positive sample set, S′ is the negative sample set, and L s∈S is the indicator function, d(s) is the joint decision result of multiple distribution distances, and φ(d(s) is the characteristic representation of the decision result.

[0019] Further preferably, in training the anomaly detection model, a random search method is used to optimize parameters in the anomaly detection model.

[0020] According to another aspect of the present invention, a system for constructing a rolling bearing abnormality detection model is provided. The system includes an actuator, and the actuator executes the above-mentioned method for constructing a rolling bearing abnormality detection model.

[0021] In general, the above technical solutions conceived by the present invention have the following beneficial effects compared with the prior art:

[0022] 1. The present invention expands the bearing vibration signal sample set through the proposed data generation and enhancement method, including both positive samples and negative samples, solves the problem of small samples, fits the normal data distribution through the bearing anomaly detection model, achieves better recognition of anomalies, and achieves better feature enhancement through data generation, expands the sample space, and realizes the problem of abnormal detection of mechanical signals;

[0023] 2. The present invention can significantly increase the amount of data, expand the sample set, and increase the diversity of the sample set through the technology of data generation and sample splicing. At the same time, the data generation process increases the module's anti-disturbance ability by adding random noise, which can help explain and understand the model's decision-making process, especially in how the model distinguishes between normal and abnormal behaviors. Therefore, this method can significantly improve the performance of the anomaly detection model, especially when the data is limited, the anomalies are rare, or the diversity is insufficient. This method provides a flexible and powerful tool for model training, which helps to build a more accurate and robust anomaly detection system;

[0024] 3. The present invention introduces a joint decision-making mechanism into the loss function, that is, the final decision is made by integrating the information of multiple distribution distances. For each distribution distance, a weight parameter can be introduced to perform weighted calculations. In this way, the influence of different distribution distances can be adjusted according to their importance in capturing data features, so as to better adapt to the diversity of data. Further, for the key hyperparameters in model training, random search is used to update and iterate, which can more efficiently find high-performance parameter combinations in the parameter space that conforms to physical laws, avoid over-optimization of parameters, reduce the risk of overfitting to a certain extent, and improve the detection performance of the model. BRIEF DESCRIPTION OF THE DRAWINGS

[0025] Figure 1 It is a flow chart of a method for constructing a variable operating condition bearing fault detection model based on time domain data generated according to a preferred embodiment of the present invention;

[0026] Figure 2 It is a flow chart of a data generation model constructed according to a preferred embodiment of the present invention;

[0027] Figure 3 It is a schematic diagram of the training process of the bearing anomaly detection model constructed according to the preferred embodiment of the present invention. DETAILED DESCRIPTION

[0028] In order to make the purpose, technical solutions and advantages of the present invention more clearly understood, the present invention is further described in detail below in conjunction with the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are only used to explain the present invention and are not intended to limit the present invention. In addition, the technical features involved in the various embodiments of the present invention described below can be combined with each other as long as they do not conflict with each other.

[0029] A method for constructing a variable operating condition bearing fault detection model based on time domain data generation includes the following steps:

[0030] (1) Constructing a training dataset

[0031] The collected time-domain vibration signals of the bearings are called set A. A bearing vibration signal generation model is built based on the time-domain data. The data generation model consists of the following modules: frequency domain conversion module, spectrum filtering module, and signal generation module. The frequency domain conversion unit is used to perform discrete Fourier transform on the time series data of the bearing vibration signal, convert it into frequency domain representation, and introduce slight disturbances to the input data in the frequency domain to generate new frequency domain vibration signals. The spectrum filtering unit can adjust and generate signals according to the spectral characteristics and frequency components of the signal through spectrum filtering operation to capture the spectral characteristics and structure of the signal. The signal generation unit converts the frequency domain representation data back into time domain representation through inverse Fourier transform to generate set B of bearing vibration signals with spectral characteristics.

[0032] The bearing vibration signal generation model contains three modules, namely, frequency domain conversion module, spectrum filtering module, and signal generation module. Before using the data generation model, the input data needs to be arranged in a specified format, so the original data needs to be processed in a certain data format through the data loader. The time series data such as vibration is used as the time series part, and the data loader is used to input it into the data generation neural network. In order to prevent the neural network from model divergence when the input data frequency is too high, the original data will be downsampled in the data loader.

[0033] S1.1 Frequency Domain Conversion Module

[0034] In the first step of the process, the present invention performs a time series x=[x0,...,x T-1 ] is converted into its spectrum representation by Fourier transform - the present invention is implemented by applying discrete Fourier operation to each feature dimension d, assuming x d =[x0,d[r d ],...,x T-1 ,d[r d ]] is a time series related to feature dimension d, and the Fourier transform layer calculates x for all d∈{1,...,D} d The n-point DFT is obtained by the following three steps:

[0035] Temporal zero padding: ∨t,t mod r d ≠0,

[0036] Fourier Transform:

[0037] Spectrum Clipping:

[0038] Here, 0 N-T represents a set of NT zeros, and the union ∪ represents the addition of 0 N-TAppend zeros to the time series x d The padding operation on the DFT further utilizes the characteristics of DFT interpolation and sampling frequency to ensure that the time series is filled Its spectrum has a fixed length N, regardless of T and sampling period r d The length of the interval, because the DFT coefficients are complex numbers, is a tensor with dimensions 2×1×N and the Fourier transform of d is the collection of feature dimensions, is a tensor of dimension 2×D×N, that is, the DFT layer transforms each time series x into a two-channel "image-like" D×N matrix and The stream with n-point DFT is an n-point Fourier stream. In order to ensure lossless recovery of the time series x by inverse DFT, the present invention ensures that N≥T.

[0039] At the same time, in order to enhance the anti-disturbance ability of the data generation model, a slight disturbance is added to the frequency domain data after the frequency domain conversion. The specific method is: Each point of the frequency domain data can be expressed in the form of amplitude and phase: X(f) = A(f)e jφ(f) , where A(f) is the amplitude, φ(f) is the phase, and j is the imaginary unit. The present invention simulates the disturbance that may occur in actual situations by changing the amplitude and phase: Amplitude change: multiply the original amplitude A(f) by a random factor 1+βM(f) close to 1, where M(f) is a randomly generated array of the same size as A(f), and β controls the degree of change: A′(f)=A(f)·(1+βM(f)); Phase change: add a small random disturbance γP(f) to the original phase φ(f), where P(f) is a randomly generated array of the same size as φ(f), and γ controls the degree of change: φ′(f)=φ(f)+γP(f); Therefore, the perturbed frequency domain signal can be expressed as:

[0040] X′(f)=A′(f)e jφ′(f) =A(f)·(1+βM(f))·e j(φ(f)+γP(f))

[0041] Where A′(f) and φ′(f) are the amplitude and phase after slight changes, respectively, and β and γ are small positive coefficients used to control the degree of amplitude and phase changes.

[0042] S1.2 Spectrum filtering module

[0043] The second layer of the Fourier flow is an affine coupling layer, but applied in the frequency domain to the time series x as follows:

[0044]

[0045] Wherein, H and μ are D×N matrices, BiRNN represents a bidirectional recurrent neural network, ⊙ represents a Hadamard (element) product, and here, the present invention The real channel and the virtual channel are divided in .

[0046] Affine transformations in the frequency domain can be considered as a spectral filtering operation, where the frequency transformation of the time series part is applied to a filter with transfer function H. The transfer function itself is data-dependent: it depends on In other words, the frequency transformation of the time series X. The mapping to the transfer function H is achieved by using an RNN parameter in all different frequency components, because N can be very long for the time series T. Since the present invention uses a bidirectional RNN, all frequency components can be generated at the same time. The bearing vibration signal data generation process is as follows Figure 1 shown.

[0047] S1.3 Signal generation module

[0048] The frequency domain data after spectrum filtering is converted back to time domain representation through inverse Fourier transform to generate a new bearing vibration signal. The inverse Fourier transform formula is: The inverse Fourier transform converts the data represented in the frequency domain back into the time domain to generate a bearing vibration signal with spectral characteristics.

[0049] The training data building module is used to generate data X obtained through the data generation model g (t) part X gpart (t), at the randomly inserted point p and the original real data X r (t) According to the normal sample size, the spliced ​​data Y(t) is used as a supplement to the data set to form a training sample set. Replace the i elements in A with the i elements in B to form a new data set. The details are as follows:

[0050] A) Define the signal and ratio, let X r (t) is the real signal, with length N r ,X g (t) is the generated signal, with a length of N g .

[0051] B) Determine the splicing ratio, from X g Take out 4 / 10 of it for splicing, that is, N gpart =4 / 10N g This part is recorded as X gpart . Define a new signal Y(t), where X gpart (t) will be randomly inserted into X r (t), keep the total length Nr +N gpart .

[0052] C) Describe the random insertion process, randomly select X r (t). Let the insertion point be p, where p is a random number that satisfies 1≤p≤N r . Then Y(t) can be constructed as follows:

[0053] First p-1 points: Y(t) = X r (t), for t=1,2,...,p-1.

[0054] Insert X gpart Part: Y(t) = X gpart (tp), for t=p,p+1,...,p+N g .

[0055] The remaining X r Part (t): Y(t) = X r (tN g -p) for t = p + N g +1,p+N g +2,...,N r +N g .

[0056] Right now:

[0057] Y(t)=X r (t) t﹤p

[0058] Y(t)=X gpart (tp) p≤t≤p+N g

[0059] Y(t)=X r (tN g -p) p+N g <t≤N r +N g

[0060] Here, t = 1, 2, ..., N r +N g is the time index of the new signal Y(t).

[0061] (2) Anomaly Detection Model

[0062] Construct a bearing anomaly detection model, including a sequence partitioning module and a classifier module; the sequence partitioning module uses a sliding window of length l and span r to convert the training set into a subsequence S = {s1,s 1+r ,...};

[0063] The classifier module models the time axis dependency and the interaction between variables through the temporal convolutional network TCN, and further maps the data into the H-dimensional feature space through a lightweight projection head variable ψ of a multi-layer perceptron network structure. Set the center of the hypersphere c as the feature space The distance between the data map and the center of the hypersphere c represents the abnormality of the sample. This method is used to construct a calibrated single-class classifier, and the iterative parameters are updated by random search.

[0064] Based on the above anomaly detection methods, the distance between the evaluation sample and the center of the hypersphere directly determines the detection effect of the model. Different distribution distances have their own advantages. Combining multiple distance metrics for joint decision-making can obtain a more stable and accurate anomaly detection structure. Therefore, in order to optimize the anomaly detection process and improve the performance of the classifier, a decision-making mechanism for multiple distribution distances is proposed. The key to the loss function is to evaluate the distribution distance from the sample to the center of the hypersphere. Three distribution distances suitable for anomaly detection are selected: Euclidean distance L2, KL divergence D KL , cross entropy H is used as the evaluation index, and the three distribution distances are simultaneously involved in the decision-making through parameter weighting.

[0065] In the subsequent process of establishing a bearing abnormality detection model through positive sample modeling, the present invention inputs the generated signal together with the normal signal into the model as a supplement to the data set, and inputs the normal bearing vibration data into the data neural network to train the corresponding data generation model. Different models are used to generate data under different speed conditions. These generated data and the original normal bearing data are used as training sets for fault detection, which can expand the training set, provide more types of positive samples for the detection model, and improve the accuracy of the detection model.

[0066] Before model training, the time series must be preprocessed. Let Y = <Y1,Y2,...Y N > is a generated time series signal, where each observation in Y is represented by D variables (X t ∈R D ,∨X t ∈Y). When D>1, the dataset Y is called a multivariate time series. If D=1, the dataset is simplified to a univariate setting. Unsupervised time series anomaly detection f measures the degree of anomaly of each observation without obtaining any label information and gives an anomaly score, i.e., f:Y→r. The higher the anomaly score, the higher the possibility of anomaly.

[0067] The present invention considers the local context window of each observation to model their temporal dependency. Specifically, a sliding window of length l and span r is used to convert the training set into a subsequence S = {s1,s 1+r,...}, where S t = <x t ,x t+1 ,...,x t+l-1 >. In the inference phase, the test set is also split into subsequences using the same window length of 1, and the sliding stride is set to l. The anomaly detection model evaluates the degree of abnormality of each subsequence and assigns an anomaly score to the last timestamp of each subsequence. The present invention uses 0 to fill the starting l-1 timestamps to obtain the final anomaly score list.

[0068] In model training, a temporal convolutional network (TCN) is used as a temporal modeling network to map data to a feature space F. A lightweight projection head with a multi-layer perceptron network structure is used to map data to an H-dimensional feature space. Therefore, in this method, the loss function of the model mainly includes two parts: a partial loss function of a class of learning objectives is used to calibrate the model to adapt to the uncertainty of abnormal samples and encourage confident predictions at the same time, thereby shielding abnormal contamination in the training set; a partial loss function of the classification learning objective is used to classify the original abnormal samples through the classification learning branch to distinguish the real abnormal behavior. The calculation formula of the loss function is as follows:

[0069]

[0070] Among them, ψ c is the prediction function of the positive sample, which is used to classify the abnormal samples, y + =1 and y - = -1 is used to classify abnormal samples and normal samples, c∈R H It is the feature space The center of the hypersphere on is the expectation of all samples, S is the positive sample set, S′ is the negative sample set, and L s∈S is an indicator function, which takes the value of 1 when the sample belongs to S, otherwise it is 0. d(s) is the joint decision result of multiple distribution distances, and φ(s) is the feature representation of the input sample s.

[0071] In order to improve the selection of metric distance and thus optimize the performance of anomaly detection, a method for joint decision-making of multiple mapping distances is proposed. During the research process, the present invention finds that the performance of the model is closely related to the selection of distribution distance. The role of distribution distance is to quantify the gap between the model prediction distribution and the real data distribution. Each distribution distance has its specific advantages and usage scenarios. The present invention finds that the model performance can be further improved by combining multiple distribution distances through weighted decision-making. This method can combine the advantages of different metrics to achieve better results. When using the TCN model for anomaly detection, it is crucial to select a suitable distribution distance to measure the difference between the model output and the actual data. Several distribution distances suitable for the TCN anomaly detection model will be introduced below, and multiple distribution distances will be combined to improve the loss function for joint decision-making.

[0072] Euclidean distance Where y is the true value, is the predicted value, KL divergence Where P is the true distribution, Q is the model prediction distribution, and cross entropy H(P,Q) = -∑ x P(x)log(Q(x)), these are all distribution distances widely used in classification models. The present invention makes decisions by combining these distribution distances to improve the prediction accuracy of the model. The following is the implementation process:

[0073] The three distribution distances are represented as d1(s), d2(s), and d3(s). For each distribution distance d i (s), introduce the weight parameter y i (y1=1), indicating its influence in the loss function. Then the weighted distance calculation result of each distribution distance is w i d i (s), where w i =y i / ∑y i The weighted distance calculation results of multiple distribution distances can be used for joint decision making by weighted summation. The final joint decision result is: d(s) = ∑iw i d i (s), the final loss function can be calculated by the joint decision result d(s), which is defined as:

[0074]

[0075] Through the above methods, it is possible to achieve improvements by combining multiple distribution distances, making joint decisions, and performing weighted calculations on multiple distance spaces. Such improvements can improve the model's ability to characterize data features, thereby improving the accuracy and robustness of anomaly detection.

[0076] A random search method is proposed to update the iterative model. For the key hyperparameters in model training, namely the sliding window length l and span r, and the distribution distance weights y2 and y3, the present invention hopes to automatically update the iteration by random search to find the parameter combination with the best model training effect. For each parameter pi∈{l,r,y2,y3}, the search space Spi=[p imin ,p imax ]. Perform N experiments, in each experiment: randomly select each parameter p from its search space Spi i The value of v pi , using the parameter combination (v l ,v r ,v y2 ,v y3 ) Train the model and evaluate its performance P.

[0077] In the model training, a temporal convolutional network (TCN) is used as a temporal modeling network to map the data to the feature space F. A lightweight projection head with a multi-layer perceptron network structure is used to map the data to the H-dimensional feature space. The model learns the normal characteristics of the data through training data. In the process of testing the model, the present invention finds that the key parameters that affect the accuracy of the model are as follows: the length l and span r of the sliding window, and the distribution distance weight parameters y2 and y3. Therefore, the present invention hopes to quickly find the most suitable parameter combination through a certain parameter setting method.

[0078] The implementation method of the present invention is to use a random search method to set parameters, and the specific implementation method is as follows:

[0079] 1. Define parameter space: First, define a search space S for each parameter. This space contains the range of possible values ​​of the parameter. The search space of length 1 is S l =[l min ,l max ]=, where l min and l max are the minimum and maximum values ​​of l respectively, and the search space of span r is S r =[r min ,r max ], the distribution distance weight parameter is S y2 =[y 2min ,y 2max ],S y3 =[y 3min ,y 3max ], combining the search ranges of each hyperparameter to form the entire search space. Such a search space can be a hyperparameter grid or a randomly sampled set of hyperparameter spaces.

[0080] 2. Determine the number of searches: Determine the number of random samplings, that is, the number of hyperparameter configurations to try. This number of searches needs to be determined based on computing resources and time constraints.

[0081] 3. Random sampling: In the random sampling stage, a sample is randomly selected from the defined hyperparameter search space as the hyperparameter configuration. This usually involves uniformly distributed random sampling within the range of values ​​of each hyperparameter. Specifically, the following formula is used for random sampling: For the hyperparameter λ i , from the range [a i ,b i ] randomly select a value: λ i =a i +(b i -a i )random(), where random() is a random number uniformly drawn between [0,1].

[0082] 4. Evaluate performance: In the performance evaluation phase, the model is trained using the selected hyperparameter combination (l, r, y2, y3), and the performance of the model is evaluated in the validation set or cross-validation. The performance of the model is usually measured using indicators such as loss function or accuracy. Specifically, the performance of the model on the validation set can be calculated using the following formula:

[0083] For the hyperparameter configuration λ, calculate the loss function value on the validation set: ψ valid (λ)=mean x∈Xvaild L(x; A λ (X train )), where L(x; A λ (X train )) represents the loss function value of the model on the validation set sample x.

[0084] 5. Select the best configuration: After all searches are completed, the best performing hyperparameter combination (l, r, y2, y3) is selected as the final model configuration. This is usually the configuration that performs best on the validation set.

[0085] 6. Repeat the experiment: Repeat the above steps multiple times to ensure robust results. Each random search may result in a different optimal configuration, so multiple experiments can improve the reliability of the results.

[0086] The present invention will be further described below with reference to specific embodiments.

[0087] like Figure 3As shown, the bearing vibration signal data collected by the vibration sensor is sliced, and the samples are marked as positive samples and abnormal samples, respectively. The samples are then input into the trained variable operating condition bearing fault diagnosis model provided in Example 1 of the present invention, and the corresponding predicted fault category label is obtained through the abnormal detection model.

[0088] The following are the specific implementation steps, including:

[0089] 1) Prepare data sets: Vibration acceleration sensors are used to collect raw data of bearings under different working conditions. Each working condition may include multiple types of faults, such as normal, outer ring fault, inner ring fault, and rolling element fault. Signals are intercepted by setting a sliding window with an appropriate window width, and the signals are marked and divided into normal samples and abnormal samples. Long time series are segmented into several equal long time series segments to form the target domain data.

[0090] 2) Bearing anomaly detection: Target domain data Input the trained variable-condition bearing anomaly detection model, the model automatically performs anomaly detection and outputs the predicted fault category label Realize bearing abnormality detection under variable working conditions.

[0091] It will be easily understood by those skilled in the art that the above description is only a preferred embodiment of the present invention and is not intended to limit the present invention. Any modifications, equivalent substitutions and improvements made within the spirit and principles of the present invention should be included in the protection scope of the present invention.

Claims

1. A method for constructing a rolling bearing abnormality detection model, characterized in that: The method comprises the following steps: The vibration signals of the bearing vibration are collected to form an original vibration signal set A and a fault label is assigned to each vibration signal in the original vibration signal set; the data generation model is used to perform data processing on each vibration signal in the original vibration signal set to generate a new vibration signal, thereby forming a new vibration signal set B; the vibration signal set B is spliced ​​with the vibration signals in the original vibration signal set A to form a spliced ​​vibration signal, and the spliced ​​vibration signal is added to the original vibration signal set A to form a vibration signal set C; wherein, the fault labels remain unchanged during the process of performing data processing to generate a new vibration signal and splicing the vibration signal using the data generation model; The anomaly detection model is trained using the vibration signal set C to obtain the required anomaly detection model for the rolling bearing. The input of the anomaly detection model is the vibration signal, and the output is the fault label.

2. The method for constructing a rolling bearing abnormality detection model according to claim 1, characterized in that: The original vibration signal set A includes a normal vibration signal set A1 and an abnormal vibration signal set A2. The fault label of the normal vibration signal set A1 is normal, and the fault label of the abnormal vibration signal in the abnormal vibration signal set A2 is set according to the abnormal type of the bearing when the abnormal vibration signal is collected.

3. The method for constructing a rolling bearing abnormality detection model according to claim 2, characterized in that: When the data generation model generates a new vibration signal, it uses the abnormal vibration signals in the abnormal vibration signal set A2 to generate a new abnormal vibration signal.

4. A method for constructing a rolling bearing abnormality detection model as described in claim 2 or 3, characterized in that: The vibration signal set B is spliced ​​with the vibration signals in the original vibration signal set A, that is, the vibration signals in the vibration signal set B are spliced ​​with the normal signals in the normal vibration signal set A1.

5. The method for constructing a rolling bearing abnormality detection model according to claim 4, characterized in that: The relationship of the spliced ​​vibration signal is as follows: Where t = 1, 2, ..., N r +N g is the time index of the concatenated vibration signal Y(t), X r (t) is the original vibration signal, Nr is the length of the original vibration signal, Xg(t) is the new vibration signal in the vibration signal set B, N g is the length of the new vibration signal generated, and t=p is the random insertion point.

6. A method for constructing a rolling bearing abnormality detection model as claimed in claim 1 or 2, characterized in that: The data generation model includes a frequency domain conversion unit, a spectrum filtering unit, and a signal generation unit, wherein the frequency domain conversion unit is used to perform discrete Fourier transform on the time series data of the vibration signal to convert it into a frequency domain representation, and at the same time introduce disturbances into the input vibration signal in the frequency domain to generate a new frequency domain vibration signal; the spectrum filtering unit is used to filter the new frequency domain vibration signal generated by the frequency domain conversion unit; and the signal generation unit converts the filtered new frequency domain vibration signal into a time domain representation through an inverse Fourier transform.

7. A method for constructing a rolling bearing abnormality detection model as described in claim 1 or 2, characterized in that: The anomaly detection model adopts a time domain convolutional network model.

8. The method for constructing a rolling bearing abnormality detection model according to claim 7, characterized in that: The loss function of the anomaly detection model is as follows: Among them, ψ c is the prediction function of the positive sample, which is used to classify the abnormal samples, y + =1 and y - = -1 is used to classify abnormal samples and normal samples, c∈R H It is the feature space The center of the hypersphere on is the expectation of all samples, S is the positive sample set, S′ is the negative sample set, and L s∈S is the indicator function, d(s) is the joint decision result of multiple distribution distances, and φ(d(s) is the characteristic representation of the decision result.

9. The method for constructing a rolling bearing abnormality detection model according to claim 8, characterized in that: In training the anomaly detection model, a random search method is used to optimize the parameters in the anomaly detection model.

10. A system for constructing a rolling bearing abnormality detection model, characterized in that: The system includes an actuator, which executes the method for constructing a rolling bearing abnormality detection model as described in any one of claims 1-9.

Images