Unsupervised classification method for time series based on adversarial joint maximum mean difference

Abstract

The present invention proposes an unsupervised classification method for time series data based on adversarial joint maximum mean difference, which belongs to the field of time series data classification. In view of the dependence of high-dimensional time series data on labels and the problem that traditional domain adaptation methods only perform domain adaptation from one side, an unsupervised classification method for time series data based on adversarial joint maximum mean difference is proposed. The present invention innovatively combines the two domain adaptation methods based on feature mapping and adversarial, uses the joint maximum mean difference distance metric to map the source domain and the target domain to another space for alignment, and also conducts adversarial training through the domain adversarial model, which can better extract the common features of the source domain time series data and the target domain time series data for domain adaptation. Applying the proposed network model to the unsupervised classification of time series data can achieve the purpose of improving the accuracy, and the experimental results verify the effectiveness of the proposed method.

Description

Technical field:

[0001] The present invention relates to a technology of unsupervised classification of data, and in particular to an unsupervised classification method of time series data based on adversarial joint maximum mean difference. Background technology:

[0002] Time series data is ubiquitous, and time series analysis has many applications in other related fields in the real world, such as healthcare, industrial diagnosis, and financial forecasting. Deep learning models have become successful models for time series analysis, but existing deep supervised models are not suitable for high-dimensional time series data with a limited number of training samples, because these data-driven methods rely on finding true labels for supervised training, but for some time series data, collecting data labels is a labor-intensive and time-consuming process, and sometimes even impossible. One solution is to learn useful representations from unlabeled data, which can greatly reduce the reliance on expensive manually annotated data.

[0003] Unsupervised Domain Adaptation is a type of transfer learning. Its source domain has labels, while the target domain does not. The labeled data in the source domain is used to assist the training of the unlabeled data in the target domain to achieve the purpose of unsupervised learning. Unsupervised domain adaptation methods are mainly divided into two categories: feature mapping-based methods and adversarial methods. Feature mapping-based methods map the features extracted by deep neural networks to another feature space, then calculate the distance between the source domain and the target domain, and narrow this difference to achieve the purpose of aligning the source domain and the target domain. This difference metric can be MK-MMD, CORAL, JMMD, Wasserstein distance, etc. Adversarial methods train a feature extractor to extract feature vectors on the one hand, and train a domain classifier to distinguish whether the feature vector is from the source domain or the target domain on the other hand. The two learn adversarially to achieve the purpose of extracting common features of the source domain and the target domain, thereby narrowing the distance between the two domains. Adversarial methods include DANN, CDA, ADDA, GTA, etc.

[0004] Traditional unsupervised domain adaptation methods only use feature mapping-based methods or adversarial methods alone, and can only extract effective features from one aspect. This study innovatively combines the two, mapping the features to another space for alignment while also performing adversarial training. Combining these two methods and applying them to the unsupervised classification task of time series data can better extract the common features of source domain time series data and target domain time series data for domain adaptation, ultimately achieving the goal of improving the accuracy of unsupervised time series classification.

[0005] The applications and advantages of the time series unsupervised classification method based on adversarial joint maximum mean difference are listed as follows.

[0006] Case 1: Applied to electrocardiogram (ECG) diagnosis to help diagnose arrhythmia, myocardial ischemia, myocardial infarction and location, determine the impact of drugs or electrolyte conditions on the heart, and promptly identify the patient's heart health status.

[0007] Case 2: Applied to human activity recognition. After collecting human motion data through sensors, the model is used for training to identify people's specific movements or actions and classify them. It can be used in the health detection function of smartphones.

[0008] Case 3: Application to rotating machinery fault diagnosis. Since the load and movement speed of mechanical equipment are constantly changing, the model trained under a certain load and speed often cannot achieve a good diagnostic effect when applied to a machine in another working environment. It may be difficult to collect data again for labeled training. This method can be used to migrate well, shorten the distance between the source domain and the target domain, and make full use of the labeled data information in the source domain. The target domain does not need labeled data for training. As long as the original data is migrated to the target domain data using the model, the purpose of fault diagnosis can be achieved.

[0009] In summary, the time series unsupervised classification method based on adversarial joint maximum mean difference can effectively align the source domain and the target domain from two aspects based on feature mapping and adversarial. Applying it to the scenario of time series classification can achieve the purpose of improving the accuracy of unsupervised classification. Summary of the invention:

[0010] The purpose of the present invention is to solve the problem of high-dimensional time series data's dependence on labels and the problem that traditional domain adaptation methods only perform domain adaptation from one aspect, and to provide a time series unsupervised classification method based on adversarial joint maximum mean difference.

[0011] (I) Technical solution

[0012] The basic idea of ​​the time series unsupervised classification method based on adversarial joint maximum mean discrepancy is: on the one hand, the joint distribution of multiple specific cross-domain layers is adjusted, that is, the joint maximum mean discrepancy (JMMD) is used to learn the migration network, and on the other hand, the DANN (Domain Adversarial Neural Network) model is used to train the feature extractor so that the domain discriminator cannot distinguish the difference between the two domains. By combining the JMMD distance and the DANN model, the purpose of more accurately aligning the distribution of the two domains can be achieved. It includes the following specific steps:

[0013] 1. Data preprocessing

[0014] like Figure 1As shown in the method flow chart, the first step of the method is to input data and perform data preprocessing. The process of data preprocessing is mainly to standardize the data. In order to ensure that the network can converge well, before the relative importance of each dimension is clear, standardization makes the distribution of each input data similar, so that we do not overly bias towards data with large variance during network training.

[0015] Step 1_1 Load the source domain and target domain data files.

[0016] and The definition of is as follows:

[0017]

[0018]

[0019] Where s represents the source domain and t represents the target domain. represents the i-th sample data in the source domain, Represents the corresponding label value, n s represents the number of sample data in the source domain, X s Represents the set of all source domain data, Y s Represents the set of all source domain data labels; represents the i-th sample data in the target domain, n t represents the number of sample data in the target domain, X t Represents the set of all target domain data. It can be seen that the target domain data is unlabeled. Since this article discusses the case where the source domain and target domain data have the same label, the same classifier can be used to classify data from both domains, that is, the classifier of the source domain can be used to classify the target domain unlabeled data.

[0020] Step 1_2 Data normalization. Use the z-score normalization method to normalize the data of the source domain and the target domain. The z-score is normalized based on the mean and standard deviation of the original data. The formula is as follows:

[0021]

[0022] where x i represents the i-th value of the time series data, represents the mean of the time series data, and s represents the variance of the time series data. This process is performed on each time series data in the source domain and the target domain.

[0023] Step 1_3 divides the source domain and target domain into training set, validation set, and test set in a ratio of 6:2:2. The training set is used to train the model, the validation set is used to adjust the model parameters, and the test set is used to test the final performance of the model.

[0024] 2. Source domain data model pre-training stage

[0025] Pre-trained models such as Figure 2 As shown in the figure, it consists of a feature extractor f(·) and a classifier c(·) of a four-layer convolutional neural network (CNN). The purpose of inputting the source domain data into CNN for pre-training is to initialize the network parameters so that the model has a certain feature extraction capability for the subsequent adversarial training. At the same time, a classifier for the source domain data can be trained so that the model can classify the source domain data well.

[0026] Step 2_1 Input the source domain training set data into the feature extractor for feature extraction, for example, the i-th sample in the source domain After feature extraction, the feature vector is obtained

[0027] Step 2_2 Input the extracted source domain features into the classifier for label classification:

[0028] Step 2_3 Use the true labels of source domain samples Calculate classification loss with cross entropy loss function The formula is as follows:

[0029]

[0030] Among them l ce represents the cross entropy loss function.

[0031] Step 2_4 Backpropagate classification loss gradient.

[0032] Step 2_5 calculates the accuracy of the source domain validation set. If there is an improvement, calculate the accuracy of the source domain test set and record the maximum source domain test accuracy.

[0033] Step 2_6 determines whether the current number of iterations is less than or equal to 50. If so, jump to step 2_1 to continue training. Otherwise, exit pre-training and enter the formal training stage.

[0034] 3. Model formal training stage

[0035] The formal training model is as follows Figure 3As shown in the figure, the formal training phase uses data from both the source domain and the target domain to input the model for training. The figure mainly contains three important modules: feature extractor f(·), classifier c(·), and domain discriminator d(·). The feature extractor and classifier in the figure share features and are essentially the same. f(·) and c(·) have been preliminarily trained in the pre-training phase and have certain feature extraction and classification capabilities. The domain discriminator consists of three fully connected layers to distinguish whether the extracted features are from the source domain or the target domain. In this stage, not only the JMMD distance is calculated, but also adversarial training is performed.

[0036] Step 3_1 Input the source domain and target domain training set data.

[0037] Step 3_2 extracts features from the source domain and target domain training data. For example, extract the i-th sample vector of the source domain and the target domain i-th sample vector enter Figure 3 The feature extractor f(·) in the source domain is used to extract features and obtain the source domain feature vectors. and the target domain feature vector The specific process is shown in the following formula:

[0038]

[0039] Step 3_3 Input the extracted source domain and target domain features into the classifier for label classification:

[0040] Step 3_4 is the same as step 2_3 of pre-training, using the true labels of the source domain samples Calculate classification loss with cross entropy loss function

[0041] Step 3_5 Calculate the joint maximum mean difference loss like Figure 3 As shown, The calculation requires the output f of the feature extractor s 、f t and the output of the classifier c s 、c t , the calculation formula is as follows:

[0042]

[0043] in represents reproducing kernel Hilbert space (RKHS), Represents the kernel function, select the Gaussian kernel function The Gaussian kernel can map infinite-dimensional space, and z is a collection of specific layers, including the output z of the last layer of the feature extraction layer and the classification layer. f and z c .

[0044] Step 3_6 Calculate domain adversarial loss The source domain and target domain features f extracted in step 3_2 are s 、f t Input domain discriminator d(·), perform domain classification, and get domain classification loss The domain classification loss is calculated using the binary cross entropy function, as follows:

[0045]

[0046] Step 3_7 Calculate the total loss The formula is as follows:

[0047]

[0048] Where λ represents the trade-off parameter of JMMD loss and its range is [0,1].

[0049] Step 3_8 back-propagates the gradient according to the total loss value in step 3_7, where and Normal back propagation, The purpose of adversarial training is achieved by inverting the gradient through the Gradient Reversal Layer (GRL) and then back-propagating it.

[0050] 4. Model testing phase

[0051] The model in the testing phase is Figure 4 As shown in the figure, the test uses the data of the target domain. The validation set and test set data of the target domain are sequentially passed through the feature extractor f(·) and classifier c(·) trained in the formal training phase, and finally the true label of the target domain is obtained to achieve the effect of unsupervised classification. The purpose of the validation set is to adjust the hyperparameters of the model, and the division of the test set can enable the model to have a better generalization effect.

[0052] Step 4_1 Load the target domain validation set and test set data.

[0053] Step 4_2 Pass the target domain verification data through Figure 4 For example, for the i-th target domain validation set sample First, the data is classified by the feature extractor f(·) and then by the classifier c(·) to obtain the predicted label The formula is as follows:

[0054]

[0055] Step 4_3 compares the predicted label with the true label and calculates the accuracy of the target domain validation set. If there is an improvement, use the same operation as step 4_2 to calculate the accuracy of the target domain test set and use the current test accuracy to update the final classification accuracy.

[0056] Step 4_4 determines whether the current number of iterations is less than or equal to 250. If so, jump to step 3_1 to continue training, otherwise end.

[0057] (II) Beneficial effects

[0058] 1. The present invention solves the problem of dependence on a large number of data labels during high-dimensional time series data training, avoids tedious data labeling work and improves training efficiency. At the same time, it can also utilize a large amount of original source domain labeled data to assist in training new target domain unlabeled data to obtain an effective unsupervised classification model.

[0059] 2. Compared with traditional unsupervised domain adaptation methods, the present invention extracts features more effectively. The temporal unsupervised classification method based on adversarial joint maximum mean difference proposed in the present invention aligns the source domain and the target domain from two aspects. Adversarial training is added in the process of aligning the two domains using a feature mapping-based method. At the same time, the feature mapping-based method can also make the adversarial training process more stable. The collaborative training of these two methods can better narrow the distance between the two domains, so that the target domain can make full use of the labeled data of the source domain and improve the accuracy of unsupervised classification. Description of the drawings:

[0060] Figure 1 The figure is a flow chart of the method of the present invention.

[0061] Figure 2 This is the pre-training model diagram described in the present invention.

[0062] Figure 3 This is the formal training model diagram described in the present invention.

[0063] Figure 4 This is a diagram of the test model described in the present invention. Specific implementation method:

[0064] In order to explain the technical solution of the present invention more clearly and completely, the present invention is further explained below in conjunction with the accompanying drawings and examples.

[0065] The following is a bearing data set PU obtained from the University of Paderborn. The source domain is data with a speed of 1500 rpm, a load torque of 0.7 Nm, and a radial force of 1000 N. The target domain is data with a speed of 900 rpm, a load torque of 0.7 Nm, and a radial force of 1000 N. This is used as an example to explain. Figure 1 As shown, the present invention provides a temporal unsupervised classification method based on adversarial joint maximum mean difference, comprising the following steps:

[0066] Step 1 Figure 1 As shown in the method flow chart, the first step of the method is to input data and perform data preprocessing. The specific steps of data preprocessing are as follows:

[0067] Step 1_1 Load the mat data files corresponding to the source domain and target domain in the PU folder. After loading, there are 3255 data in the source domain and 3257 data in the target domain.

[0068] Step 1_2: Use the z-score standardization method to normalize each data in the source domain and the target domain. For example, for the time series data x=(-0.00367,-0.00016,-0.00005.-0.00020,0.00163,...) with a length of 1024, its mean is 0.0008 and the standard deviation is 0.0019. Substitute each value of the x sequence into the formula Then we get a new sequence x=(-2.37456,-0.49505,-0.43354,-0.51591,0.46342,...) with a length of 1024.

[0069] Step 1_3 divides the source domain and target domain into training set, validation set, and test set. The number of data items in the source domain training set is 1953, the validation set is 651, and the test set is 651. The number of data items in the target domain training set is 1954, the validation set is 651, and the test set is 651. The ratio is 6:2:2, and the length of each data item is 1024.

[0070] Step 2: Source domain data model pre-training phase. The pre-trained model is as follows: Figure 2 As shown, the source domain data is input into CNN for pre-training to initialize the network parameters, and a classifier about the source domain data is trained so that the model can classify the source domain data well. The specific steps are as follows:

[0071] Step 2_1 Input 1953 data of the source domain training set into the feature extractor for feature extraction.

[0072] Step 2_2: Input the extracted source domain features into the classifier for label classification.

[0073] Step 2_3 Use the true labels of source domain samples Calculate classification loss with cross entropy loss function

[0074] Step 2_4 Backpropagate classification loss gradient.

[0075] Step 2_5 calculates the source domain verification set accuracy to be 0.2719, which is an improvement over the previous verification accuracy of 0. 0.2719 is taken as the current maximum source domain test accuracy.

[0076] Step 2_6 determines whether the current number of iterations is less than or equal to 50. If so, jump to step 2_1 to continue training. Otherwise, exit pre-training and enter the formal training stage.

[0077] Step 3: Model formal training phase. The formal training model is as follows Figure 3 As shown, the data from the source domain and the target domain are used to input the model training at the same time. At this stage, the JMMD distance calculation and adversarial training are performed. The specific steps are as follows:

[0078] Step 3_1 Input 1953 source domain and 1954 target domain training set data.

[0079] Step 3_2 extracts features from the source domain and target domain training data. For example, extract the i-th sample vector of the source domain and the target domain i-th sample vector enter Figure 3 The feature extractor f(·) in the source domain is used to extract features and obtain the source domain feature vectors. and the target domain feature vector

[0080] Step 3_3 Input the extracted source domain and target domain features into the classifier for label classification.

[0081] Step 3_4 is the same as step 2_3 of pre-training, using the true labels of the source domain samples Calculate classification loss with cross entropy loss function

[0082] Step 3_5 Calculate the joint maximum mean difference loss like Figure 3 As shown, z is a collection of specific layers, including the output z of the last layer of the feature extraction layer and the classification layer. f and z c . The calculation formula is as follows:

[0083]

[0084] Step 3_6 Calculate domain adversarial loss The source domain and target domain features f extracted in step 3_2 are s 、f t Input domain discriminator d(·), perform domain classification, and get domain classification loss The domain classification loss is calculated using the binary cross entropy function, as follows:

[0085]

[0086] Step 3_7 Calculate the total loss The formula is as follows:

[0087]

[0088] Where λ represents the trade-off parameter of JMMD loss, ranging from [0,1], and λ gradually increases from 0 to 1 according to the following formula:

[0089]

[0090] epoch represents the current iteration number, and its value range is [0,250].

[0091] Step 3_8 back-propagates the gradient according to the total loss value in step 3_7, where and Normal back propagation, The purpose of adversarial training is achieved by inverting the gradient through the Gradient Reversal Layer (GRL) and then back-propagating it.

[0092] Step 4: Model testing phase. The model in the testing phase is as follows: Figure 4 As shown in Figure 1, the test uses the data from the target domain. The validation set and test set data of the target domain are sequentially passed through the feature extractor f(·) and classifier c(·) trained in the formal training phase, and then the accuracy is calculated. The specific steps are as follows:

[0093] Step 4_1 loads 651 validation set and 651 test set data of the target domain.

[0094] Step 4_2 Pass the target domain verification data through Figure 4 For example, for the i-th target domain validation set sample First, the data is classified by the feature extractor f(·) and then by the classifier c(·) to obtain the predicted label The formula is as follows:

[0095]

[0096] Step 4_3 compares the predicted label with the true label and calculates the accuracy of the target domain validation set. If there is an improvement, use the same operation as step 4_2 to calculate the accuracy of the target domain test set and use the current test accuracy to update the final classification accuracy.

[0097] Step 4_4 determines whether the current number of iterations is less than or equal to 250. If so, jump to step 3_1 to continue training, otherwise end.

[0098] The data set used in this example comes from the PU data example of the University of Paderborn in Germany. The unsupervised classification accuracy is selected as the classification performance index of the algorithm. Compared with the traditional unsupervised domain adaptation method, the classification accuracy of the present invention is improved. The accuracy comparison is shown in Table 1, where JDA represents the temporal unsupervised classification method based on the adversarial joint maximum mean difference proposed by the present invention. The numbers in the first row of the table represent different operating conditions. By changing the speed of the drive system, the radial force on the test bearing, and the load torque on the transmission system, the PU data set includes four operating conditions, as shown in Table 2. For example, for task 0→1, it means that the source domain is data with a speed of 1500rpm, a load torque of 0.7Nm, and a radial force of 1000N, and the target domain is data with a speed of 900rpm, a load torque of 0.7Nm, and a radial force of 1000N. There are 12 transfer learning settings in the PU data set.

[0099] Table 1

[0100] Task 0-1 0-2 0-3 1-0 1-2 1-3 2-0 2-1 2-3 3-0 3-1 3-2 AVG Basic 0.2165 0.7719 0.4208 0.3406 0.3409 0.2412 0.7771 0.2150 0.4206 0.4087 0.2900 0.4042 0.4040 DAN 0.3325 0.8024 0.4372 0.3742 0.3884 0.2711 0.7926 0.3503 0.4272 0.4178 0.3034 0.4418 0.4449 AdaBN 0.3582 0.7540 0.4862 0.4509 0.4244 0.3273 0.7375 0.3582 0.4951 0.4587 0.3253 0.4795 0.4713 CDA 0.3853 0.8171 0.5213 0.3828 0.4605 0.3156 0.8169 0.3975 0.4989 0.4507 0.2972 0.4669 0.4842 JDA 0.3874 0.7997 0.5661 0.3785 0.4223 0.3222 0.8101 0.4181 0.4998 0.5214 0.3067 0.5185 0.4959

[0101] Table 2

[0102] Task 0 1 2 3 Load torque(Nm) 0.7 0.7 0.1 0.7 Radial force(N) 1000 1000 1000 400 Speed ​​(rpm) 1500 900 1500 1500

[0103] The above is a detailed description of the embodiments of the present invention in conjunction with the accompanying drawings, and the specific implementation methods herein are only used to help understand the method of the present invention. For those of ordinary skill in the art, according to the idea of ​​the present invention, changes and modifications can be made in the specific implementation methods and application scopes, so the present invention should not be understood as limiting the present invention.

Claims

1. An unsupervised classification method for time series data based on adversarial joint maximum mean difference, characterized by The method comprises the following steps: Step 1: Data preprocessing, including data standardization and data set division; The data is a bearing data set, the source domain is data with a rotation speed of 1500 rpm, a load torque of 0.7 Nm, and a radial force of 1000 N, and the target domain is data with a rotation speed of 900 rpm, a load torque of 0.7 Nm, and a radial force of 1000 N; Step 2: In the model pre-training phase, a classifier is trained to effectively classify source domain data; Step 3: In the formal training phase of the model, the domain adaptation method based on adversarial joint maximum mean difference is used for training; Step 4: Model testing phase: Use the trained model to test the data in the target domain; The pre-training module in step 2 includes the following specific steps: Step 2_1 Input the source domain training set data into the CNN feature extractor for feature extraction; Step 2_2: Input the extracted source domain features into the classifier for label classification; Step 2_3: Compute the classification loss using the cross entropy loss function by comparing the predicted label with the true label of the source domain sample. Step 2_4 Backpropagate classification loss The gradient of Step 2_5: Calculate the accuracy of the source domain validation set. If there is an improvement, calculate the accuracy of the source domain test set and record the maximum source domain test accuracy. Step 2_6 determines whether the current number of iterations is less than or equal to 50. If so, jump to step 2_1 to continue training. Otherwise, exit pre-training and enter the formal training stage. The formal training module in step 3 includes the following specific steps: Step 3_1 Input source domain and target domain training set data; Step 3_2: Input the source domain and target domain training data into the CNN feature extractor for feature extraction; Step 3_3 Input the extracted source domain and target domain features into the classifier for label classification; Step 3_4 is the same as step 2_3 of pre-training. The classification loss is calculated using the true label of the source domain sample and the cross entropy loss function. Step 3_5 Calculate the joint maximum mean difference loss The calculation requires the output f of the feature extractor s 、f t and the output of the classifier c s 、c t , the calculation formula is as follows: in represents reproducing kernel Hilbert space (RKHS), Represents the kernel function, select the Gaussian kernel function Gaussian kernel maps infinite dimensional space, z is a collection of specific layers, including the output z of the last layer of the feature extraction layer and the classification layer f and z c , f(·) is the feature extractor, x s represents the source domain, x t represents the target domain; Step 3_6 Calculate domain adversarial loss The source domain and target domain features f extracted in step 3_2 are s 、f t Input domain discriminator d(·), perform domain classification, and get domain classification loss The domain classification loss is calculated using the binary cross entropy function, as follows: is the domain classification loss, n s Represents the number of sample data in the source domain, represents the i-th sample data in the source domain, represents the i-th sample data in the target domain, n t Represents the number of sample data in the target domain Step 3_7 Calculate the total loss The formula is as follows: in is the joint maximum mean difference loss, Calculate the classification loss for the cross entropy loss function, λ represents the trade-off parameter of the JMMD loss, and the range is [0,1]; Step 3_8 Based on the total loss value in step 3_7 Perform gradient back propagation, where and Normal back propagation, The purpose of adversarial training is achieved by inverting the gradient through the Gradient Reversal Layer (GRL) and then back-propagating it.

2. The unsupervised classification method for time series data based on adversarial joint maximum mean difference according to claim 1, characterized in that: The data preprocessing module in step 1 comprises the following specific steps: Step 1_1 Load the source domain and target domain Data files; Step 1_2 data normalization; The data of the source domain and the target domain are normalized using the z-score standardization method; the z-score is standardized based on the mean and standard deviation of the original data, and the formula is as follows: where x i represents the i-th value of the time series data, represents the mean of the time series data, and s represents the variance of the time series data. This processing is performed on each time series data in the source domain and the target domain. Step 1_3 divides the source domain and target domain into training set, validation set, and test set in a ratio of 6:2:

2.

3. The unsupervised classification method for time series data based on adversarial joint maximum mean difference according to claim 1, characterized in that: The model testing module in step 4 includes the following specific steps: Step 4_1 Load the target domain validation set and test set data; Step 4_2 extracts features and classifies the target domain verification data through the test model; Step 4_3 compares the predicted label with the true label and calculates the accuracy of the target domain validation set. If there is an improvement, use the same operation as step 4_2 to calculate the accuracy of the target domain test set and use the current test accuracy to update the final classification accuracy. Step 4_4 determines whether the current number of iterations is less than or equal to 250. If so, jump to step 3_1 to continue training, otherwise end.

Images