An adaptive strategy-based noisy data anomaly detection method

Abstract

The application relates to a noise-containing data anomaly detection method based on an adaptive strategy, which comprises the following steps: obtaining a vibration signal and constructing a data sample; performing training set and test set division; generating a no-label subset Ut from K samples randomly sampled from a training set no-label data set; merging a positive example set P and the no-label subset Ut to serve as a training subset D; using a grid search algorithm to find optimal initial model parameters, setting a batch size, a learning rate and an iteration number N according to the result; training a neural network through a training set, extracting frequency domain feature information through an FCN model, learning a time sequence relationship state feature through an LSTM module, predicting a class of the training set no-label data, dividing the final prediction result by n, calculating an evaluation index, and outputting a training set calculation result. The method can reduce the influence of label noise, can also perform signal noise reduction, and solves the problem that signal noise interferes with an anomaly detection result.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of intelligent data monitoring, and particularly relates to a noisy data anomaly detection method based on an adaptive strategy. BACKGROUND

[0002] Mechanical equipment failure often leads to factory shutdown, resulting in large economic losses, and even personnel casualties, environmental pollution and other major accidents. Therefore, the research on mechanical equipment anomaly detection technology is the key to ensure the safe and reliable operation of equipment, and has practical significance in protecting life and property safety and social stability.

[0003] Currently, industrial field monitoring data has the characteristics of large data volume, high dimension, fast generation speed and low value density. For example, a set of cooling equipment in an ultra-high voltage converter station has 402 measuring points, which generates about 3GB of monitoring data per minute, and manual data interpretation has a large workload and cannot be implemented on site. Therefore, how to identify the data state of these large amounts of monitoring data has become a difficulty and hotspot for equipment operation and maintenance.

[0004] The data state identification of monitoring data is essentially anomaly detection. In the face of the situation that there is a large amount of signal noise in industrial field data, which interferes with the anomaly detection result, the traditional anomaly detection method generally has two steps. First, the signal is preprocessed, including signal denoising, feature representation, feature dimension reduction and other operations. Then, the processed signal features are input into the decision model, and the decision model gives the final result. However, an end-to-end anomaly detection algorithm that embeds signal denoising structure into network structure has not been proposed, resulting in different target functions in each stage, which makes the anomaly detection model not good. And the specific method of preprocessing needs to be designed by professional technical personnel, and combined with effective anomaly discrimination algorithm to show significant effect. Even so, engineers are difficult to propose an algorithm process that always maintains good performance under various noise conditions.

[0005] The semi-supervised learning anomaly detection method is easily affected by two factors. (1) The probability of obtaining positive and negative samples by randomly and uniformly sampling each time from the unlabeled data set is determined by the pollution rate. However, in the training process, they are all trained as negative samples, which are called label noise in the present application, and will definitely affect the classification accuracy of the base learner; (2) The industrial field mechanical equipment state monitoring data itself is easily affected by noise, which is called signal noise in the present application, and will make the learner unable to correctly learn the feature representation of the monitoring signal. Therefore, a noisy data anomaly detection based on an adaptive strategy can be proposed for the two types of noise conditions, label noise and signal noise, that occur in the anomaly detection process. SUMMARY

[0006] The application aims to provide a noisy data anomaly detection method based on an adaptive strategy, so as to realize adaptive monitoring of noisy data, reduce the requirement for user's own experience, and reduce the consumption of human resources and time cost.

[0007] To achieve the above-mentioned purpose, the application adopts the following technical solutions:

[0008] A noisy data anomaly detection method based on an adaptive strategy comprises the following steps:

[0009] (1) Obtain a vibration signal, arrange, standardize and segment the vibration signal into fixed segments to construct a data sample;

[0010] (2) Divide the vibration signal sequentially collected in step (1) into a training set and a test set;

[0011] (3) Randomly sample K samples from the training set unlabeled data set to generate an unlabeled subset Ut, wherein K is the size of the unlabeled subset;

[0012] (4) Merge the positive example set P and the unlabeled subset Ut as a training subset D;

[0013] (5) Use a grid search algorithm to find the optimal initial model parameters, and set the batch size, learning rate and iteration number N according to the results;

[0014] (6) Train the neural network with the training set, extract frequency domain feature information through the FCN model, learn the state characteristics of the time sequence relationship through the LSTM module, perform softmax classification, and update the model parameters using the back propagation algorithm;

[0015] (7) Determine whether the network training number m reaches the preset iteration number N, if yes, proceed to the next step, otherwise repeat step (5);

[0016] (8) Predict all unlabeled data U to generate a class probability P, save the prediction class s(x) corresponding to the data, and increase the counter n by 1;

[0017] (9) Determine whether the difference between the prediction probabilities before and after is greater than the set threshold value δ, if yes, proceed to the next step, otherwise generate an unlabeled subset with 1-P as the unlabeled set sampling probability, and repeat step (4);

[0018] (10) Divide the prediction class of the training set unlabeled data by n to obtain the final prediction result, calculate the evaluation index, and output the training set calculation result;

[0019] (11) Test the performance of the trained model, calculate the evaluation index, and output the test set calculation result.

[0020] Furthermore, vibration signals are acquired in real time using an accelerometer to obtain vibration signals over several cycles; specifically, the following steps are included:

[0021] 1) Install the hardware system containing the acceleration vibration sensor onto the drive end of the motor housing and fix it to the housing using a magnetic base. Use the hardware system to collect vibration signals.

[0022] 2) The system begins to periodically collect and store vibration signals throughout the entire bearing rotation process.

[0023] Furthermore, in step (5), the adaptive sampling method initially samples from the unlabeled dataset D with uniform probability 1 / n, and trains the given classification model h. θ，0 (D, 0), then by using the mislabeling probability of each sample Update the sampling probabilities of samples in D, iteratively update the training dataset (D, k), where k is the number of iterations. The sampling probabilities of positive and negative instances in the unlabeled samples are standardized.

[0024]

[0025] In the formula: n+ is the number of positive instances in D; n- is the number of negative instances in D.

[0026] Furthermore, in step (8), different sampling probabilities are applied to the unlabeled dataset based on the predicted probability given by the previous classifier. The higher the predicted probability, the more likely the sample is to be a positive sample. Therefore, a smaller sampling probability is required to prevent it from being selected again. The lower the predicted probability, the more likely the sample is not a positive sample. A larger sampling probability is required to increase the probability of negative samples in the next unlabeled subset, thus alleviating the label noise problem.

[0027] Furthermore, in step (9), attention layers are added to the LSTM module and the FCN module respectively.

[0028] Furthermore, while adding an attention layer to the FCN module, an adaptive soft thresholding module was designed to denoise the extracted data features.

[0029] The beneficial effects of this invention are:

[0030] The present invention provides an adaptive strategy-based anomaly detection method for noisy data. In the face of the problem that a large amount of signal noise in industrial field data interferes with the anomaly detection results, an adaptive sampling method is proposed for anomaly detection of unlabeled data in noisy environments. This method can effectively reduce the contamination rate of unlabeled subsets, thereby overcoming the influence of centralized label noise in unlabeled data.

[0031] The adaptive soft threshold denoising method can automatically learn the denoising threshold of different channel signals, and embed the denoising module into the network architecture to complete signal denoising. The blank of the end-to-end anomaly detection algorithm of embedding the signal denoising structure into the network structure is filled. Compared with the current research which usually adopts the anomaly detection method of pre-processing, denoising and anomaly discrimination, the method has higher convenience and significant economy, and is convenient, safe and reliable to operate. BRIEF DESCRIPTION OF DRAWINGS

[0032] Figure 1 The application is an adaptive strategy-based noisy data anomaly detection method. DETAILED DESCRIPTION

[0033] The technical solutions in the embodiments of the application will be clearly and completely described below with reference to the drawings in the embodiments of the application.

[0034] Embodiment:

[0035] As shown in the figure, an adaptive strategy-based noisy data anomaly detection method comprises the following steps: Figure 1

[0036] (1) Obtain a vibration signal, arrange, standardize and segment into fixed segments to construct a data sample;

[0037] (2) Divide the vibration signal sequentially collected in step (1) into a training set and a test set;

[0038] (3) Randomly sample K samples from the training set unlabeled data set to generate an unlabeled subset Ut, wherein K is the size of the unlabeled subset;

[0039] (4) Merge the positive example set P and the unlabeled subset as a training subset D;

[0040] (5) Use a grid search algorithm to find the optimal initial model parameters, and set the batch size, learning rate, iteration number N and other parameters according to the results;

[0041] The adaptive sampling method initially samples from the unlabeled data set D with a uniform probability of 1 / n, and trains a given classification model h θ，0 (D, 0), and then updates the sampling probability of the samples in D by the mislabeling probability of each sample , and iteratively updates the training data set (D, k) (k is the number of iterations). In addition, the sampling probabilities of the positive and negative instances in the unlabeled samples are normalized

[0042]

[0043] In the formula: n + — The number of positive instances in D; n​- The number of negative instances in D. The probability normalization makes the model select similar numbers of positive and negative instances, thus being robust to datasets with imbalanced class distribution. The weighted sampling of normalized probabilities reduces the risk of selecting potentially mislabeled instances without instance filtering.

[0044] (6) The training set is used to train the neural network. The frequency domain feature information is extracted by the FCN model, and the LSTM module learns the state features of the time sequence relationship. After softmax classification, the model parameters are updated using the back propagation algorithm.

[0045] (7) Determine whether the number of network training times m reaches the pre-set iteration number N. If yes, proceed to the next step, otherwise repeat step (5);

[0046] (8) Predict all unlabeled data U to generate class probability P, and save the predicted class s(x) corresponding to the data. The counter n is incremented by 1.

[0047] According to the prediction probability given by the last classifier, different sampling probabilities are applied to the unlabeled data set. The greater the prediction probability, the more likely the sample is a positive sample, so it needs to be given a smaller sampling probability to prevent being selected again. The smaller the prediction probability, the more likely the sample is not a positive sample, so it is given a larger sampling probability to increase the probability of negative samples in the next unlabeled subset, thereby alleviating the label noise problem.

[0048] (9) Determine whether the difference between the prediction probabilities before and after is greater than the set threshold δ. If yes, proceed to the next step, otherwise generate an unlabeled subset with a sampling probability of 1-P, and repeat step (4);

[0049] Based on the attention mechanism and the idea of wavelet soft threshold, a soft threshold denoising module is designed. Attention layers are added to the LSTM module and the FCN module respectively, so that different features of the sample have different weights, and the main resources of the model are concentrated on the features that contribute more to the final result. In addition, an adaptive soft threshold module is designed to denoise the extracted data features while adding an attention layer to the FCN module.

[0050] (10) Divide the predicted class of the training set unlabeled data by n to obtain the final prediction result, calculate the evaluation index, and output the training set calculation result.

[0051] (11) Test the performance of the trained model on the test set, calculate the evaluation index, and output the test set calculation result.

[0052] The adaptive strategy-based noisy data anomaly detection method of the application uses a data acquisition hardware system to collect bearing vibration signals; the prediction result probability of the last sub-model is regarded as the prior probability of generating the next training subset, different sampling probabilities are applied to the unlabeled data set, and the influence of label noise on the detection accuracy of the model is reduced; further, for signal noise reduction, a soft threshold noise reduction module is designed by using the attention mechanism and the wavelet soft threshold idea, and the attention layer is added to the LSTM and FCN modules respectively, so that different features of the sample have different weights, and thus the main resources of the model are concentrated on the features that contribute more to the final result.

[0053] The application is not limited to the above best embodiment, and anyone can derive other various forms of products under the inspiration of the application, but regardless of any change in shape or structure, any technical solution with the same or similar technical solution as the application falls within the protection scope of the application.

Claims

1. A method for detecting anomalies in noisy data based on an adaptive strategy, characterized in that: Includes the following steps: (1) Obtain vibration signals, organize, standardize and divide them into fixed segments to construct data samples; (2) Divide the vibration signals collected in step (1) into training set and test set; (3) Randomly sample K samples from the unlabeled training dataset to generate an unlabeled subset Ut, where K is the size of the unlabeled subset; (4) Merge the set of positive examples P and the unlabeled subset Ut as the training subset D; (5) Use the grid search algorithm to find the optimal initial model parameters, and set the batch size, learning rate, and number of iterations N according to the results; (6) The training set is used to train the neural network. Frequency domain feature information is extracted through the FCN model, and the temporal relationship state features are learned through the LSTM module. After softmax classification, the model parameters are updated using the backpropagation algorithm. (7) Determine whether the number of network training times m has reached the preset number of iterations N. If yes, proceed to the next step; otherwise, repeat step (5). (8) Predict all unlabeled data U, generate class probabilities P, and save the predicted class s(x) corresponding to the data, and increment the counter n by 1; (9) Determine whether the difference between the two prediction probabilities reaches the set threshold δ. If yes, proceed to the next step. Otherwise, use 1-P as the sampling probability of the unlabeled set to generate an unlabeled subset and repeat step (4). (10) Divide the predicted category of the unlabeled data in the training set by n to obtain the final prediction result, calculate the evaluation index, and output the calculation result of the training set. (11) Test the performance of the trained model on the test set, calculate the evaluation index, and output the test set calculation results.

2. The method for detecting anomalies in noisy data according to claim 1, characterized in that: Vibration signals are acquired in real time using an accelerometer to obtain vibration signals over several cycles; specifically, the following steps are included: 1) Install the hardware system containing the acceleration vibration sensor onto the drive end of the motor housing and fix it to the housing using a magnetic base. Use the hardware system to collect vibration signals. 2) The system begins to periodically collect and store vibration signals throughout the entire bearing rotation process.

3. The method for detecting anomalies in noisy data according to claim 1, characterized in that, In step (5), the adaptive sampling method initially samples from the unlabeled dataset D with uniform probability 1 / n and trains the given classification model h. θ,0 (D,0), then by using the mislabeling probability of each sample Update the sampling probabilities of samples in D, iteratively update the training dataset (D, k), where k is the number of iterations. The sampling probabilities of positive and negative instances in the unlabeled samples are standardized. In the formula: n+ is the number of positive instances in D; n- is the number of negative instances in D.

4. The method for detecting anomalies in noisy data according to claim 1, characterized in that: In step (8), based on the predicted probability given by the previous classifier, different sampling probabilities are applied to the unlabeled dataset. The higher the predicted probability, the more likely the sample is to be a positive sample. Therefore, a smaller sampling probability is required to prevent it from being selected again. The lower the predicted probability, the more likely the sample is not a positive sample. A larger sampling probability is required to increase the probability of negative samples in the next unlabeled subset, thus alleviating the label noise problem.

5. The method for detecting anomalies in noisy data according to claim 1, characterized in that: In step (9), attention layers are added to the LSTM module and the FCN module respectively.

6. The method for detecting anomalies in noisy data according to claim 5, characterized in that: While adding an attention layer to the FCN module, an adaptive soft thresholding module was designed to denoise the extracted data features.

Images