A distributed fiber optic sensing event recognition method based on open set unsupervised domain adaptation
By adopting an open set unsupervised domain adaptation method in distributed fiber sensing technology, using circular generation adversarial networks and adversarial learning for knowledge transfer, the problem of difficulty in identifying unknown categories and high data collection costs is solved, and the model is improved by improving the robustness of the model and cross-domain adaptability.
Patent Information
- Application Number
- CN202411052331.X
- Authority / Receiving Office
- CN · China
- Patent Type
- Patents(China)
- Current Assignee / Owner
- Filing Date
- 2024-08-02
- Publication Date
- 2024-11-12
- Estimated Expiration
- 2044-08-02
AI Technical Summary
Existing deep learning methods are difficult to effectively identify and classify vibration events in unknown categories in distributed fiber sensing technology, resulting in high false alarm rates and the need to collect large amounts of labeled data for each deployment environment, which is costly.
Using an open set unsupervised domain adaptation method, the source domain data is converted into target domain data through a circular generation adversarial network, a new data set is constructed, and an adversarial learning method is used to transfer knowledge, identify unknown categories, and reduce false positive rates.
It improves the cross-domain generalization capability and robustness of the model, reduces the labor cost of recollecting labeled data, can effectively identify unknown categories in the deployment environment, and reduces the false positive rate.
Smart Images

Figure CN118568550B_ABST
Abstract
Description
Technical Field
[0001] The present invention relates to the technical field of signal detection, and in particular to a distributed optical fiber sensing event recognition method based on open set unsupervised domain adaptation. Background Art
[0002] Distributed fiber optic sensing technology is a sensing technology that uses optical fiber as a sensing medium and measures physical parameters (such as temperature, strain and vibration) by analyzing the changes in backscattered light in the optical fiber. It can perform long-distance, high-precision continuous monitoring and has the advantages of high sensitivity, anti-electromagnetic interference, and no need for power supply. Distributed fiber optic acoustic wave sensing technology based on phase-sensitive optical time domain reflectometry (φ-OTDR) can effectively detect vibration signals. By identifying and classifying different vibration events, it can monitor whether the event is dangerous. It can be used in perimeter security, pipeline intrusion monitoring and other fields.
[0003] At present, researchers mainly use deep learning methods to identify and classify vibration signals. Deep learning methods are highly dependent on large-scale labeled data, and require the collection of exclusive data for each new deployment environment for training to ensure the effectiveness of the model. However, it is very difficult or even impossible to collect enough vibration event labeled data for each new deployment environment. Unsupervised domain adaptation is a branch of transfer learning, which can solve this problem by transferring the knowledge of existing vibration event labeled data (source domain) to new scene data (target domain). The problem is that existing deep learning methods usually assume that the event categories in the deployment environment are exactly the same as the categories of the training data. However, in actual scenarios, there are usually many unknown classes that are not included in the training data. Existing methods will forcibly classify these unknown categories into pre-specified categories (known classes), resulting in false alarms and reducing the robustness and accuracy of the event classification model. Summary of the invention
[0004] In view of the shortcomings of the above-mentioned prior art, the present application provides a distributed fiber optic sensing event recognition method based on open set unsupervised domain adaptation, which can fully utilize the data in the source domain for knowledge transfer and effectively identify unknown event categories in the target domain.
[0005] The technical solution adopted by the present invention to solve the above technical problems is as follows:
[0006] A distributed optical fiber sensing event recognition method based on open set unsupervised domain adaptation comprises the following steps:
[0007] S1: The distributed fiber optic acoustic wave sensing system collects vibration event signals, converts the vibration event signals into two-dimensional image data, and constructs the source domain image dataset and the target domain dataset respectively;
[0008] S2: Use a cyclic generative adversarial network to convert source domain samples into corresponding target domain versions and construct a new source domain dataset;
[0009] S3: using the source domain dataset converted in step S2 to train a classification model, calculating a normal score based on the output of the model for the target domain image, and dividing the target domain dataset into a known class target domain dataset and an unknown class target domain dataset according to the normal score;
[0010] S4: Use adversarial learning methods to align the features of known class data in the source domain and the target domain to transfer knowledge. At the same time, use the source domain data and unknown class data in the target domain to train the classification model so that it can correctly classify known class samples and identify unknown class samples.
[0011] S5: Deploy the model in a real environment, collect vibration signals using a distributed fiber optic acoustic wave sensing system, classify the converted vibration signals through a classification model, and determine whether to generate an alarm or take action.
[0012] Furthermore, in step S1, two-dimensional image data is obtained by performing Gram angular field transformation on the vibration signal, and a labeled source domain vibration event image dataset and an unlabeled target domain vibration event image dataset are constructed.
[0013] Furthermore, in step S2, in order to convert the image from the source domain to the target domain, a generator and a domain discriminator need to be trained. The generator generates similar target domain samples from the source domain samples, and the domain discriminator distinguishes the converted samples from the real target domain samples. The generative adversarial network is trained using the objective function.
[0014] Furthermore, a cycle consistency method is used to impose constraints on the transformed images to ensure that the semantic information of the source domain samples does not change after the transformation. The cycle consistency method is to transform the transformed samples back to the source domain, and the reconstructed source domain images should also be similar to the original samples at the pixel level. In the process, additional generators and domain discriminators are used for training to allow reverse transformation.
[0015] The trained generator can generate a target domain converted version for each source domain sample. This version of the source domain sample is more similar to the sample in the target domain and is used in the subsequent domain adaptation process. All converted source domain samples constitute a new source domain dataset.
[0016] Furthermore, in step S3, a classification model is trained using the converted source domain dataset. The model includes two parts: a feature extractor and a classifier. The output dimension of the classifier is the number of categories in the source domain. The cross entropy loss function is used to train the model. The trained classification model is used to calculate a normal score for each target domain sample. According to the size of the normal score, the target domain samples are preliminarily divided into known classes and unknown classes.
[0017] Furthermore, after separating the known class target domain dataset and the unknown class target domain dataset, knowledge transfer is performed by aligning the features of the known class data of the source domain and the target domain using the adversarial learning method, and the model is trained using the unknown class samples of the target domain to identify the unknown class samples.
[0018] Furthermore, in order to simplify the training process, a gradient reversal layer is added between the feature extractor and the domain discriminator to implement adversarial training in one back propagation.
[0019] Furthermore, the distributed fiber optic acoustic wave sensing system collects vibration event signals, performs Gram angle field conversion on the original vibration signals to obtain two-dimensional image data, and inputs the image data into the model to obtain the output result. The category corresponding to the largest item in the vector of the output result is the predicted event category of the vibration signal. The dangerousness is judged according to the event category, and a decision is made whether to generate an alarm message or take action.
[0020] The beneficial effects of the present invention are as follows:
[0021] Compared with the prior art, the present invention transfers the knowledge of distributed optical fiber sensing source domain data to the target domain through a domain adaptation method, making full use of the existing labeled data, thereby greatly reducing the labor cost of re-collecting labeled data;
[0022] Compared with the prior art, the difference in the distribution of training data and test data will cause the existing distributed optical fiber sensing method to fail. The present invention uses the domain adaptation method to improve the cross-domain generalization ability of the model and ensure the effectiveness of the model in different environments.
[0023] Compared with the prior art, the present invention can identify unknown categories in the deployment environment, reduce the false alarm rate of the distributed optical fiber sensing event recognition and classification system, and thus improve the robustness of the model. BRIEF DESCRIPTION OF THE DRAWINGS
[0024] Figure 1 is a flow chart of the present invention;
[0025] Figure 2 This is a schematic diagram of event classification based on oil and gas pipeline intrusion events according to the present invention;
[0026] Figure 3 This is a system block diagram of the cyclic generative adversarial network in the present invention. DETAILED DESCRIPTION
[0027] The specific implementation of the present invention will be described below in conjunction with the accompanying drawings.
[0028] The present invention proposes a distributed fiber optic sensing event recognition method based on open set unsupervised domain adaptation, which migrates the knowledge of labeled source domain data to unlabeled target domain data, improves the classification effect of the model in the new environment, and can also identify unknown categories in the deployment environment to avoid model false alarms.
[0029] The present invention takes oil and gas pipeline intrusion detection as an example. Figures 1 to 3 As shown:
[0030] In a preferred embodiment of the present invention, the specific implementation method of the distributed optical fiber sensing data preparation stage is as follows:
[0031] The distributed fiber optic acoustic wave sensing system is used to collect vibration event signals, and the original one-dimensional time series data (vibration signal) is converted into two-dimensional image data by Gram angle field conversion. A labeled source domain vibration event image dataset is constructed ( ) and an unlabeled target domain vibration event image dataset ( ), where the source domain refers to the auxiliary data set that provides label knowledge, and the target domain refers to the data set that currently needs to be classified for events. The data distribution of the two is different. Source domain label set ( ) is the target domain label set ( ), that is, , The categories in are called known categories, Private classes are called unknown classes.
[0032] In a preferred embodiment of the present invention, the specific implementation method of the source domain data conversion stage is:
[0033] Use a cyclic generative adversarial network to convert source domain samples into corresponding target domain versions, reduce the inter-domain distribution differences between source domain samples and target domain samples, and avoid the interference of inter-domain differences on the model's distinction between known and unknown target domain samples. The specific steps are as follows:
[0034] To convert images from the source domain to the target domain, we train a generator ( ) and a domain discriminator ( ), the corresponding loss function is:
[0035] ;
[0036] Generator( ) aims to generate realistic target domain samples using source domain samples ( ), making it look similar to the image in the target domain, while the domain discriminator ( ) aims to distinguish the transformed samples ( ) and the real target domain samples ( ). The generative adversarial network is trained using the following minimax objective function, namely:
[0037] ;
[0038] In addition, in order to ensure that the semantic information of the source domain samples does not change after conversion, it is necessary to use the cycle consistency method to impose constraints on the converted images. The idea of cycle consistency is to convert ( ) is converted back to the source domain, and the reconstructed source domain image ( ) should also be consistent with the original sample at the pixel level ( ). This requires the use of an additional generator ( ) and domain discriminator( ) to allow reverse conversion, which is still trained using the generative adversarial loss function:
[0039] ;
[0040] The cycle consistency loss is then defined as:
[0041] ;
[0042] The image similarity is calculated by The schematic diagram of the structure of the cyclic generative adversarial network is as follows: Figure 2 shown.
[0043] Generator completed through training , a target domain conversion version can be generated for each source domain sample , this version of the source domain samples is more similar to the samples in the target domain and is used in the subsequent domain adaptation process. All converted source domain samples constitute a new source domain dataset .
[0044] The target domain dataset is initially divided into a target domain known class dataset and a target domain unknown class dataset. The specific steps are as follows:
[0045] Leverage the transformed source domain dataset Train a classification model that includes a feature extractor and classifier Two parts, classifier The output dimension of is the number of categories in the source domain. The model is trained using the cross entropy loss function, which is as follows:
[0046] ;
[0047] Next, the trained classification model is used to calculate the normal score for each target domain sample. The larger the normal score, the greater the probability that the target domain sample belongs to the known category, and vice versa. Therefore, the target domain samples can be preliminarily divided into known categories and unknown categories based on the size of the normal score. The specific steps are as follows:
[0048] The target domain sample Input the above trained classification model, the output of the model is , the confidence value of the target domain sample is calculated through the output of the model and entropy , the specific calculation methods of the two are as follows:
[0049] ;
[0050] ;
[0051] Represents the probability value of the sample belonging to the kth class, and the max operation represents the acquisition The largest item in the vector. Using the maximum and minimum normalization method, and Normalized to Within the range.
[0052] The larger the confidence value of the target domain sample and the smaller the entropy value, the greater the probability that the sample belongs to the known class, and vice versa. According to this principle, for each target domain sample Calculate its normal score :
[0053] ;
[0054] Take the max operation and Larger values of , yield more reliable normal scores.
[0055] The average normal score of the entire target domain dataset is used as the discrimination threshold, and all samples with normal scores greater than the threshold are classified as the target domain known class sample dataset. , all samples with normal scores less than or equal to the threshold are divided into the unknown class dataset of the target domain .
[0056] After separating the known class target domain dataset and the unknown class target domain dataset, we can use the adversarial learning method to align the known class data of the source domain and the target domain to transfer knowledge. At the same time, the unknown class samples of the target domain can be used to train the model to identify the unknown class samples. The model at this stage consists of three components: feature extractor , classifier and domain discriminator . Feature Extractor The initial parameters of the feature extractor are the same as those of the known and unknown class separation stage. The output dimension is , and the last dimension represents the unknown class. The specific training steps of the model are as follows:
[0057] Using adversarial learning methods to align source domain samples and known class target domain samples, the loss function is as follows:
[0058] ;
[0059] At the same time, the source domain samples and the unknown class target domain dataset are used to train the classification model to identify the known class samples and distinguish the unknown class samples. The classification loss of the source domain samples And the classification loss of unknown class samples in the target domain The loss function is as follows:
[0060] ;
[0061] ;
[0062] To sum up, the overall loss function is as follows:
[0063] ;
[0064] The above loss function involves adversarial learning process and needs to be trained in a maximum-minimum optimization manner. In order to simplify the training process, in the feature extractor and domain discriminator A gradient reversal layer is added in between to implement one back propagation to complete the adversarial training.
[0065] The distributed fiber optic acoustic wave sensing system collects vibration event signals, performs Gram angle field conversion on the original one-dimensional time series data (vibration signal) to obtain two-dimensional image data, and converts the image data into Input model can get output results , The category corresponding to the largest item in the vector is the predicted event category of the test data. According to the event category, its danger can be judged and whether to generate an alarm message or take action can be decided.
[0066] The above description is an explanation of the present invention, not a limitation of the present invention, and any form of modification may be made within the protection scope of the present invention.
Claims
1. A distributed optical fiber sensing event recognition method based on open set unsupervised domain adaptation, characterized by: The steps include: S1: The distributed fiber optic acoustic wave sensing system collects vibration event signals, converts the vibration event signals into two-dimensional image data, and constructs the source domain image dataset and the target domain dataset respectively; S2: Use a cyclic generative adversarial network to convert source domain samples into corresponding target domain versions and construct a new source domain dataset; To convert from the source domain to the target domain, a generator and a domain discriminator need to be trained. The generator generates similar target domain samples from source domain samples, and the domain discriminator distinguishes the converted samples from the real target domain samples. The generative adversarial network is trained using the objective function; S3: using the source domain dataset converted in step S2 to train a classification model, calculating a normal score based on the output of the model for the target domain image, and dividing the target domain dataset into a known class target domain dataset and an unknown class target domain dataset according to the normal score; S4: Use adversarial learning methods to align the features of known class data in the source domain and the target domain to transfer knowledge. At the same time, use the source domain data and unknown class data in the target domain to train the classification model so that it can correctly classify known class samples and identify unknown class samples. S5: Deploy the model in a real environment, collect vibration signals using a distributed fiber optic acoustic wave sensing system, classify the converted vibration signals through a classification model, and determine whether to generate an alarm or take action; In step S2, in order to ensure that the semantic information of the source domain sample does not change after conversion, a constraint is imposed on the converted image by a cycle consistency method, wherein the cycle consistency method converts the converted target domain sample back to the source domain, and the reconstructed source domain image is also similar to the original sample at the pixel level; In step S3, the classification model includes two parts: a feature extractor and a classifier. The output dimension of the classifier is the number of categories in the source domain. The model is trained using a cross entropy loss function. The trained classification model is used to calculate a normal score for each target domain sample. According to the size of the normal score, the target domain samples are divided into known classes and unknown classes.
2. The method for identifying distributed optical fiber sensing events based on open set unsupervised domain adaptation according to claim 1, characterized in that: In step S1, two-dimensional image data is obtained by performing Gram angular field transformation on the vibration event signal, and a labeled source domain vibration event image dataset and an unlabeled target domain vibration event image dataset are constructed.
3. The method for identifying distributed optical fiber sensing events based on open set unsupervised domain adaptation according to claim 1, characterized in that: In order to simplify the training process, a gradient reversal layer is added between the feature extractor and the domain discriminator to implement adversarial training in one back propagation.
4. The method for identifying distributed optical fiber sensing events based on open set unsupervised domain adaptation according to claim 1, characterized in that: The distributed fiber optic acoustic wave sensing system collects vibration event signals, performs Gram angle field conversion on the original vibration signals to obtain two-dimensional image data, and inputs the image data into the model to obtain the output result. The category corresponding to the largest item in the vector of the output result is the predicted event category of the vibration signal. The dangerousness is judged according to the event category, and it is decided whether to generate an alarm message or take action.
