A method and system for constructing an abnormality detection model for rolling bearings

Abstract

This invention belongs to the technical field of bearing anomaly related fields and discloses a method and system for constructing a rolling bearing anomaly detection model. The method includes the following steps: collecting vibration signals from the bearing to form an original vibration signal set A and assigning fault labels; using a data generation model to process the original vibration signal set to generate new vibration signals, forming a new vibration signal set B; splicing the vibration signal set B with the vibration signals in the original vibration signal set A to form a spliced ​​vibration signal, the fault labels of the spliced ​​vibration signal being the same as those in the vibration signal set B; adding the spliced ​​vibration signal to the original vibration signal set A to form a vibration signal set C; training an anomaly detection model using the vibration signal set C to obtain the required rolling bearing anomaly detection model, the input of which is the vibration signal and the output is the fault label. This invention solves the technical problem of difficulty in bearing anomaly detection.

Description

Technical Field

[0001] This invention belongs to the technical field of bearing anomaly-related technologies, and more specifically, relates to a method and system for constructing a rolling bearing anomaly detection model. Background Technology

[0002] Rolling bearings are core components of large machinery in fields such as petrochemicals, wind power generation, high-speed railways, industrial machine tools, and aerospace, playing a crucial role in the stable operation of rotating machinery. However, under long-term high-speed rotation, rolling bearings are highly susceptible to failure, threatening the safety of the machinery and personnel, and placing them at risk of accidents. Furthermore, accidents caused by rolling bearing failures can have severe consequences, resulting in unpredictable personal injury and economic losses. Therefore, timely fault detection of rolling bearings in use and ensuring the safe operation of bearing-based machinery are of paramount importance.

[0003] Furthermore, missed faults during bearing fault detection can damage mechanical equipment, and in severe cases, lead to accidents. Therefore, improving the accuracy of fault detection and avoiding misdiagnosis and missed diagnoses are of paramount importance to ensuring the normal operation of rotating machinery systems. Summary of the Invention

[0004] In view of the above-mentioned defects or improvement needs of the existing technology, the present invention provides a method and system for constructing a rolling bearing anomaly detection model, which solves the problem of insufficient data samples in the training process of the anomaly model in the prior art.

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

[0006] Vibration signals from bearing vibration are collected to form an original vibration signal set A, and each vibration signal in the original vibration signal set is assigned a fault label. A data generation model is used to process the vibration signals in the original vibration signal set to generate new vibration signals, thus forming a new vibration signal set B. This vibration signal set B is then spliced ​​with the vibration signals in the original vibration signal set A to form a spliced ​​vibration signal. The fault labels of the spliced ​​vibration signal are the same as those of the vibration signals in vibration signal set B. The spliced ​​vibration signal is then added to the original vibration signal set A to form a vibration signal set C. During the data processing to generate new vibration signals and the splicing of vibration signals, the fault labels remain unchanged.

[0007] The required anomaly detection model for rolling bearings is obtained by training an anomaly detection model using a vibration signal set C. The input of the anomaly detection model is the vibration signal, and the output is the fault label.

[0008] More 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] More preferably, 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.

[0010] More 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] More preferably, the relationship between the spliced ​​vibration signals is as follows:

[0012]

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

[0014] More preferably, the data generation model includes a frequency domain conversion unit, a spectrum filtering unit, and a signal generation unit. The frequency domain conversion unit is used to convert the time series data of the vibration signal into a frequency domain representation by performing a discrete Fourier transform, and at the same time, introduces a slight disturbance 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. The signal generation unit converts the filtered new frequency domain vibration signal into a time domain representation by performing an inverse Fourier transform.

[0015] More preferably, the anomaly detection model employs a temporal convolutional network model.

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

[0017]

[0018] Where, ψ c It is the prediction function for positive samples, used to classify outlier samples, y + =1 and y -=-1 is used to classify abnormal samples and normal samples, c∈R H It is the feature space The Super Ball Center on the top, Let S be the expectation of all samples, S' be the set of positive samples, and S' be the set of negative samples. s∈S It is an 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] More preferably, in training the anomaly detection model, a random search method is used to optimize the parameters in the anomaly detection model.

[0020] According to another aspect of the present invention, a system for constructing a rolling bearing anomaly detection model is provided, the system including an actuator that performs the above-described method for constructing a rolling bearing anomaly detection model.

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

[0022] 1. This invention expands the bearing vibration signal sample set by proposing a data generation and enhancement method, including both positive and negative samples, thus solving the problem of small sample size. It fits the normal data distribution through a bearing anomaly detection model to achieve better anomaly identification. Through data generation, it achieves better feature enhancement, expands the sample space, and solves the problem of anomaly detection in mechanical signals.

[0023] 2. This invention, through data generation and sample concatenation techniques, can significantly increase the amount of data, expand the sample set, and increase its diversity. Simultaneously, the data generation process adds random noise to enhance the module's robustness against disturbances, aiding in the interpretation and understanding of the model's decision-making process, particularly in how the model distinguishes between normal and anomalous behavior. Therefore, this approach can significantly improve the performance of anomaly detection models, especially when data is limited, anomalies are rare, or diversity is insufficient. This method provides a flexible and powerful tool for model training, contributing to the construction of more accurate and robust anomaly detection systems.

[0024] 3. This invention introduces a joint decision-making mechanism into the loss function, that is, making the final decision by integrating information from multiple distribution distances. For each distribution distance, a weight parameter can be introduced to perform a weighted calculation. This allows the influence of different distribution distances to be adjusted according to their importance in capturing data features, thereby better adapting to the diversity of data. Furthermore, for key hyperparameters in model training, iterative updates are performed through random search, which can more efficiently find high-performance parameter combinations in a parameter space that conforms to physical laws. This avoids over-optimization of parameters, mitigates the risk of overfitting to a certain extent, and improves the model's detection performance. Attached Figure Description

[0025] Figure 1 This is a flowchart of a method for constructing a variable-condition bearing fault detection model based on time-domain data, according to a preferred embodiment of the present invention.

[0026] Figure 2 This is a flowchart illustrating the data generation model constructed according to a preferred embodiment of the present invention;

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

[0028] To make the objectives, technical solutions, and advantages of this invention clearer, the invention will be further described in detail below with reference to the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are merely illustrative and not intended to limit the invention. Furthermore, the technical features involved in the various embodiments of this invention described below can be combined with each other as long as they do not conflict with each other.

[0029] A method for constructing a bearing fault detection model under varying operating conditions based on time-domain data includes the following steps:

[0030] (1) Constructing the training dataset

[0031] The collected bearing time-domain vibration signals are referred to as 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 transformation module, spectrum filtering module, and signal generation module. The frequency domain transformation unit performs discrete Fourier transform on the time-series data of the bearing vibration signals, converting it into a frequency domain representation, and introduces slight perturbations into the input data in the frequency domain to generate new frequency-domain vibration signals. The spectrum filtering unit adjusts and generates signals according to their spectral characteristics and frequency components to capture their spectral features and structure. The signal generation unit converts the frequency-domain data back into a time-domain representation through inverse Fourier transform, generating bearing vibration signals with spectral characteristics, thus generating set B.

[0032] The bearing vibration signal generation model comprises three modules: a frequency domain conversion module, a spectrum filtering module, and a signal generation module. Before using the data generation model, the input data needs to be arranged according to a specified format. Therefore, a data loader is required to perform certain data format processing on the raw data. Vibration and other time-series data are considered as the time-series component. The data loader inputs this data into the neural network for data generation. To prevent the neural network from diverging when the input data frequency is too high, the data loader performs downsampling processing on the raw data.

[0033] S1.1 Frequency Domain Conversion Module

[0034] In the first step of this process, the present invention performs a time series x = [x0,...,x] T-1 The invention transforms the feature into its spectral representation through Fourier transform. This is achieved by applying a discrete Fourier operation to each feature dimension d, where x... d =[x0,d[r d ],...,x T-1 ,d[r d For a time series with feature dimension d, the Fourier transform layer computes x for all d∈{1,...,D}. d The n-point DFT is obtained through the following three steps:

[0035] Zero-fill time: ∨t,t mod r d ≠0,

[0036] Fourier transform:

[0037] Spectrum trimming:

[0038] Here, 0 N-T Let N be the set of N zeros, and let ∪ denote the union of N zeros. N-TZeros are appended to the time series x d The padding operation further utilizes the characteristics of DFT interpolation and sampling frequency to ensure the padding of the time series. and its spectrum It has a fixed length N, regardless of T and the sampling period r d Interval length, because DFT coefficients are complex numbers. It is a set of characteristic dimensions of a tensor and its Fourier transforms of dimensions 2×1×N and d. It is a 2×D×N tensor, meaning that the DFT layer transforms each time series x into a two-channel "image-like" D×N matrix. and A flow with an n-point DFT is an n-point Fourier flow. To ensure lossless recovery of the time series x using the inverse DFT, this invention guarantees N ≥ T.

[0039] To enhance the robustness of the data generation model, a slight perturbation is added to the frequency domain data after frequency domain transformation. Specifically, each point in the frequency domain data can be represented as 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, this invention simulates possible disturbances in real-world situations by changing the amplitude and phase: Amplitude change: The original amplitude A(f) is multiplied 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: A small random disturbance γP(f) is added 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 frequency domain signal after disturbance 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 change.

[0042] S1.2 Spectrum Filtering Module

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

[0044]

[0045] Where H and μ are D×N matrices, BiRNN represents a bidirectional recurrent neural network, and ⊙ represents the Hadamard (element-wise) product. Here, the present invention... It is divided into a real channel and a virtual channel.

[0046] Affine transformation in the frequency domain can be considered a spectral filtering operation, where the frequency transformation of the time series portion... This is applied to filters with a transfer function H. The transfer function itself is data-dependent: this depends on... Or, in other words, the frequency transformation X of the time series. From Mapping to the transfer function H is achieved through an RNN parameter across all different frequency components, since N can be a very long time series T. This invention uses a bidirectional RNN so that all frequency components can be generated simultaneously. The process of generating bearing vibration signal data is as follows: Figure 1 As shown.

[0047] S1.3 Signal Generation Module

[0048] The frequency domain data, after spectral filtering, is converted back to its time domain representation using the inverse Fourier transform, generating a new bearing vibration signal. The formula for the inverse Fourier transform is: The inverse Fourier transform converts the frequency domain data back into the time domain, generating a bearing vibration signal with spectral characteristics.

[0049] The training data construction module is used to process the generated data X obtained through the data generation model. g X (t) gpart (t), at the random insertion point p and the original true data X r (t) The data is concatenated according to a certain proportion according to the normal sample size. The concatenated data Y(t) is used as a supplement to the dataset to form the training sample set. The i elements in B are replaced with the i elements in A to form a new dataset. The details are as follows:

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

[0051] B) Determine the splicing ratio, from X g Take 4 / 10 of (t) for splicing, that is, N gpart =4 / 10N g This part is denoted as X. gpart Define a new signal Y(t), where X gpart (t) will randomly insert X r In (t), the total length is kept to be N.r +N gpart .

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

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

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

[0055] The remaining X r (t) part: 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 It is the time index of the new signal Y(t).

[0061] (2) Anomaly Detection Model

[0062] The bearing anomaly detection model is constructed, 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 subsequences S = {s1, s2}. 1+r A set of ,...};

[0063] The classifier module models temporal dependencies and interactions between variables using a Temporal Convolutional Network (TCN), and further maps the data to an H-dimensional feature space using a lightweight projection head variable ψ from a multilayer perceptron network structure. Let the hypersphere center c be the feature space. The distance c from the center of the hypersphere represents the degree of anomaly 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 sample and the center of the hypersphere directly determines the detection performance of the model. Different distribution distances have their own advantages, and joint decision-making using multiple distance metrics can yield a more stable and accurate anomaly detection structure. Therefore, to optimize the anomaly detection process and improve the performance of the classifier, a decision-making mechanism involving multiple distribution distances is proposed. The key to the loss function is evaluating 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 an evaluation index, and the three distribution distances are simultaneously involved in the decision-making process through parameter weighting.

[0065] In the subsequent process of establishing a bearing anomaly detection model through positive sample modeling, this invention inputs the generated signal along with the normal signal into the model as a supplement to the dataset. Normal bearing vibration data is input into the data neural network to train the corresponding data generation model. Different models are used to generate data under different rotational speeds. These generated data, along with the original normal bearing data, are used as the training set for fault detection. This expands the training set, providing the detection model with more types of positive samples and improving the model's accuracy.

[0066] Before training the model, the time series data must first be preprocessed. Let Y = <Y1,Y2,...Y N > is the generated time series signal, where each observation in Y is generated by the D variable (X). t ∈R D ,∨X t The vector ∈Y) describes the dataset. 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 for each observation without acquiring any label information, giving an anomaly score, i.e., f:Y→r. The higher the anomaly score, the higher the probability of an anomaly.

[0067] This invention considers a local context window for each observation to model their temporal dependencies. Specifically, a sliding window of length l and span r is used to transform the training set into subsequences S = {s1, s2}. 1+rA set S, ..., ... t = <x t ,x t+1 ,...,x t+l-1 During the inference phase, the test set is also divided into subsequences with the same window length of 1 and a sliding step of l. The anomaly detection model evaluates the anomaly level of each subsequence and assigns an anomaly score to the last timestamp of each subsequence. This invention uses 0-padded first l-1 timestamps to obtain the final anomaly score list.

[0068] In model training, a Temporal Convolutional Network (TCN) is used as the temporal modeling network to map the data to a feature space F. A lightweight projection head with a multilayer perceptron network structure is employed to map the data to an H-dimensional feature space. Therefore, in this method, the model's loss function mainly consists of two parts: a partial loss function for the first-class learning objective is used to calibrate the model to adapt to the uncertainty of anomalous samples while encouraging confident predictions, thereby shielding the training set from anomalous contamination; the partial loss function for the classification learning objective is used to classify the original anomalous samples through a classification learning branch to distinguish true anomalous behavior. The formula for calculating the loss function is as follows:

[0069]

[0070] Where, ψ c It is the prediction function for positive samples, used to classify outlier samples, y + =1 and y - =-1 is used to classify abnormal samples and normal samples, c∈R H It is the feature space The Super Ball Center on the top, Let S be the expectation of all samples, S' be the set of positive samples, and S' be the set of negative samples. s∈S is an indicator function that takes the value 1 when the sample belongs to S and 0 otherwise. d(s) is the joint decision result of multiple distribution distances, and φ(s) is the feature representation of the input sample s.

[0071] To improve the selection of distance metrics and thus optimize anomaly detection performance, this invention proposes a method for joint decision-making using multiple mapping distances. During the research process, this invention discovered that the performance of the model is closely related to the choice of distribution distance. The role of distribution distance is to quantify the difference between the model's predicted distribution and the actual data distribution. Each distribution distance has its specific advantages and application scenarios. This invention finds that combining multiple distribution distances through weighted decision-making can further improve model performance. This method can combine the advantages of different metrics to achieve better results. When using the TCN model for anomaly detection, selecting an appropriate distribution distance to measure the difference between the model output and the actual data is crucial. The following sections will introduce several distribution distances suitable for TCN anomaly detection models and discuss how to combine multiple distribution distances to improve the loss function for joint decision-making.

[0072] European distance Where y is the true value. These are predicted values, KL divergence. Where P is the true distribution, Q is the model-predicted distribution, and the cross-entropy H(P,Q) = -∑ x P(x)log(Q(x)), these are all distribution distances widely used in classification models. This invention improves the prediction accuracy of the model by combining these distribution distances for decision-making. The implementation process is as follows:

[0073] The three distribution distances are denoted as d1(s), d2(s), and d3(s), respectively. For each distribution distance d i (s), introducing weight parameter y i (y1=1), representing its influence in the loss function. The weighted distance calculation result for each distribution distance is w. i d i (s), where w i =y i / ∑y i The weighted distance calculation results from 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 The final loss function can be calculated from the joint decision result d(s), and is defined as:

[0074]

[0075] The above methods enable improvements by combining multiple distribution distances, making joint decisions, and weighting calculations across multiple distance spaces. Such improvements enhance the model's ability to represent data features, thereby increasing the accuracy and robustness of anomaly detection.

[0076] This paper proposes a random search method to update and iterate the model. For the key hyperparameters in model training, namely the sliding window length *l* and span *r*, and the distributed distance weights *y2* and *y3*, this invention aims to automatically update and iterate through random search to find the optimal parameter combination for model training. For each parameter *pi* ∈ {l, r, y2, y3}, its search space *Spi* = [p...] is defined. imin ,p imax Perform N experiments. In each experiment: randomly select each parameter p from its search space Spi. i The value v pi Using parameter combinations (v l ,v r ,v y2 ,v y3 Train the model and evaluate its performance P.

[0077] In model training, a Temporal Convolutional Network (TCN) is used as the temporal modeling network to map the data to a feature space F. A lightweight projection head with a multilayer perceptron network structure is employed to map the data to an H-dimensional feature space. The model learns the normal features of the data through training data. During model testing, this invention found that the key parameters affecting the model's accuracy are: the length l and span r of the sliding window, and the distribution distance weight parameters y2 and y3. Therefore, this invention aims to quickly find the most suitable parameter combination through a specific parameter setting method.

[0078] The present invention is implemented by using a random search method for parameter setting, and the specific implementation method is as follows:

[0079] 1. Define the parameter space: First, define a search space S for each parameter. This space contains the range of possible values ​​for the parameter. A search space of length 1 is S. l =[l min ,l max ] =, where l min and l max Let S be the minimum and maximum values ​​of l, and let S be the search space of the span r. r =[r min ,r max The distribution distance weight parameter is S. y2 =[y 2min ,y 2max ], S y3 =[y 3min ,y 3max The search range of each hyperparameter is combined 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 to be performed, i.e., the number of hyperparameter configurations to be tried. This number of searches needs to be determined based on computing resources and time constraints.

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

[0082] 4. Performance Evaluation: In the performance evaluation phase, the model is trained using the selected hyperparameter combination (l, r, y2, y3), and its performance is evaluated on the validation set or cross-validation. Metrics such as loss function or accuracy are typically used to measure model performance. Specifically, the model's performance 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 optimal configuration: After all searches are completed, select the best-performing combination of hyperparameters (l, r, y2, y3) 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 yield 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 labeled as positive samples and abnormal samples. Then, the samples are input into the variable working condition bearing fault diagnosis model provided in Embodiment 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) Dataset Preparation: Raw data of the bearing under different operating conditions is collected using a vibration acceleration sensor. Each condition may include multiple types of faults, such as normal, outer ring fault, inner ring fault, and rolling element fault. Signals are extracted using a sliding window with an appropriate window width and labeled as normal and abnormal samples. The long-term series is then segmented into several equally long-term series fragments to form the target domain data.

[0090] 2) Bearing anomaly detection: Target domain data Input a pre-trained variable-condition bearing anomaly detection model, the model will automatically perform anomaly detection and output predicted fault category labels. Enables abnormal detection of bearings under varying operating conditions.

[0091] Those skilled in the art will readily understand that the above description is merely a preferred embodiment of the present invention and is not intended to limit the present invention. Any modifications, equivalent substitutions, and improvements made within the spirit and principles of the present invention should be included within the scope of protection of the present invention.

Claims

1. A method for constructing a rolling bearing anomaly detection model, characterized in that, The method includes the following steps: Vibration signals from bearing vibration are collected to form an original vibration signal set A, and each vibration signal in the original vibration signal set is assigned a fault label. A data generation model is used to process the vibration signals in the original vibration signal set to generate new vibration signals, thus forming a new vibration signal set B. This vibration signal set B is then spliced ​​with the vibration signals in the original vibration signal set A to form a spliced ​​vibration signal. The spliced ​​vibration signal is then added back to the original vibration signal set A to form a vibration signal set C. During the data processing to generate new vibration signals and the splicing of vibration signals, the fault labels remain unchanged. The data generation model includes a frequency domain conversion unit, a spectrum filtering unit, and a signal generation unit. The frequency domain conversion unit converts the time-series data of the vibration signal into a frequency domain representation using a discrete Fourier transform, while simultaneously introducing perturbations into the input vibration signal in the frequency domain to generate a new frequency domain vibration signal. The spectrum filtering unit filters the new frequency domain vibration signal generated by the frequency domain conversion unit. The signal generation unit converts the filtered new frequency domain vibration signal into a time domain representation using an inverse Fourier transform. An anomaly detection model is trained using a vibration signal set C to obtain the required anomaly detection model for the rolling bearing. The input of this anomaly detection model is the vibration signal, and the output is a fault label. The loss function of the anomaly detection model evaluates the distribution distance of the vibration signal to the center of the supersphere, employing a decision-making mechanism that considers multiple distribution distances.

2. The method for constructing a rolling bearing anomaly detection model as described in 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 anomaly detection model as described in claim 2, characterized in that, The data generation model generates new abnormal vibration signals by using the abnormal vibration signals in the abnormal vibration signal set A2.

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

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

6. A method for constructing a rolling bearing anomaly detection model as described in claim 1 or 2, characterized in that, The anomaly detection model employs a temporal convolutional network model.

7. The method for constructing a rolling bearing anomaly detection model as described in claim 1, characterized in that, In training the anomaly detection model, a random search method is used to optimize the parameters of the anomaly detection model.

8. A system for constructing a rolling bearing anomaly detection model, characterized in that, The system includes an actuator that performs a method for constructing a rolling bearing anomaly detection model as described in any one of claims 1-7.

Images