A Distributed Optical Fiber Sensing Cross-Domain Event Classification Method Based on Model Migration

Abstract

The present invention provides a distributed fiber optic sensing cross-domain event classification method based on model migration, aiming to solve the problem of event classification methods based on knowledge migration in the prior art, which are inconvenient to transmit and share when accessing source domain and target domain data at the same time. The method mainly includes the steps of constructing a data set, training a source domain model, training a target domain model, and deploying a target domain model. The present invention realizes cross-domain event classification when the source data is inaccessible by introducing a passive domain adaptation method. The knowledge aggregation and pseudo-label optimization of neighborhood samples can effectively reduce the noise in the pseudo-labels, generate more accurate pseudo-labels, and thus more accurately classify events. At the same time, compared with the existing methods that rely on complex generative models, the present invention also realizes a low-computational-cost pseudo-label optimization strategy based on neighborhood sample knowledge aggregation and a self-supervised comparative learning framework, reducing the demand for computing resources and being more suitable for large-scale data processing and real-time monitoring in practical applications.

Description

Technical Field

[0001] The present invention relates to the field of signal detection technology, and in particular to a distributed optical fiber sensing cross-domain event classification method based on model migration. Background Art

[0002] In terms of event classification, distributed fiber optic sensing distinguishes different types of events by detecting and identifying features in sensor signals, such as pipeline leak detection, earthquake monitoring, border security, tunnel and bridge structural health monitoring, etc. This requires the analysis and processing of large amounts of sensor data, and machine learning and deep learning algorithms are usually used to automate this process.

[0003] Existing distributed fiber optic sensing event classification technology mainly relies on supervised machine learning methods, which require a large amount of well-labeled training data. In practical applications, obtaining sufficient and accurate labeled data is both time-consuming and expensive. In particular, distributed fiber optic sensing systems are usually deployed in different physical environments, which leads to differences in data distribution, the so-called domain shift problem. To solve the domain shift problem, researchers have proposed domain adaptation (Domain Adaptation, DA) technology attempts to solve this problem by transferring knowledge between the source domain (labeled data) and the target domain (unlabeled data); although this method improves the effect of cross-domain migration to a certain extent, there are still at least the following problems: 1) A major data security issue: Many existing cross-domain event classification methods require access to both source and target domain data. In practical applications, source domain data often involves privacy and security issues, making it inconvenient to transmit and share; 2) It is also found that there is a pseudo-label noise problem: unsupervised domain adaptation methods usually rely on pseudo-labels to label target domain data; however, there is inevitably noise in pseudo-labels, which will cause the model to be disturbed during training, thereby affecting classification performance; 3) It is also found that there is a problem of high computational cost: although some methods based on generative models perform well in cross-domain migration, their computational complexity is high, and the training process requires a lot of computing resources, making it difficult to promote in practical applications; Therefore, in response to the above pain points, it is necessary to propose further improvement solutions for the distributed fiber optic sensing cross-domain event classification method based on knowledge transfer. Summary of the Invention

[0004] In response to the shortcomings of the above-mentioned prior art, the present application provides a distributed fiber optic sensing cross-domain event classification method based on model migration. By introducing a passive domain adaptation method, cross-domain event classification is achieved when the source data is inaccessible. At the same time, through the knowledge aggregation and pseudo-label optimization of neighborhood samples, the noise in the pseudo-label can be reduced to more accurately classify events, and a low-computational-cost pseudo-label optimization strategy is implemented, reducing the demand for computing resources.

[0005] The technical solutions adopted in the present invention are as follows:

[0006] A distributed optical fiber sensing cross-domain event classification method based on model migration comprises the following steps:

[0007] S1. Distributed fiber optic sensors are deployed in the monitoring area to construct a labeled source domain vibration event dataset and an unlabeled target domain vibration event dataset;

[0008] S2, train a source domain model using a labeled source domain vibration dataset;

[0009] S3, transfer the knowledge of the source domain model to the target domain, and train the target domain model based only on the source domain model and the unlabeled target domain vibration event dataset;

[0010] S4. Deploy the trained target domain model into the distributed fiber optic sensing system to perform real-time event classification on new data in the target domain.

[0011] Furthermore, the target domain model training method in step S3 includes the following steps:

[0012] A1. Build pseudo labels and optimize them;

[0013] A2. Measure the uncertainty of the pseudo-label and calculate the corresponding weight;

[0014] A3. To reduce the negative impact of incorrect labels on model training, a self-training loss function for target domain data is constructed.

[0015] A4. Use multi-cycle historical pseudo-label information to exclude negative sample pairs that may belong to the same category;

[0016] A5. Promote the diversity of classifier predictions through regularization methods and construct an overall loss function.

[0017] Furthermore, in step A1, the trained source domain model is used to generate preliminary pseudo labels for the unlabeled data in the target domain. Based on the neighborhood knowledge aggregation strategy, the pseudo labels are optimized by integrating the information of similar samples in the target domain. For each sample in the target domain, the cosine distance between samples is calculated to find K similar samples in the feature space, and the average score is calculated using the following soft voting strategy calculation formula:

[0018] ;

[0019] represents the nearest neighbor sample set, Represents the probability output, and the superscript c indicates that the averaging operation is performed by class;

[0020] Use the argmax operation to get the corrected pseudo-label:

[0021] .

[0022] Furthermore, in step A2, the uncertainty estimation method based on the consistency of neighborhood predictions measures the uncertainty of the pseudo-label by calculating the entropy value of the prediction results of the neighborhood samples, and reweights the classification loss according to the uncertainty. The pseudo-labels with low uncertainty (high confidence) are given higher weights, and the pseudo-labels with high uncertainty (low confidence) are given lower weights. The steps are as follows: For a target domain sample , calculate its average prediction score according to formula (1), which is defined as a probability distribution, and calculate the entropy of the probability distribution:

[0023] ;

[0024] Where C is the number of categories, and the entropy value is scaled by the maximum value. The scaling formula is as follows:

[0025] ;

[0026] The weight of the sample is calculated by the normalized entropy value. The calculation formula is as follows:

[0027] .

[0028] Furthermore, in step A3, a negative learning strategy is introduced into the classification loss, that is, when calculating the classification loss, complementary labels are selected instead of correct labels, and the self-training loss function of the target domain data is obtained as follows:

[0029] ;

[0030] in is a label randomly selected from a label set that does not contain the pseudo label of the current sample, that is, , It is a strongly enhanced image The probability output of The above-mentioned estimate Uncertainty weight.

[0031] Furthermore, in step A4, the target domain feature space is regularized using a contrastive learning framework to force positive sample pairs in the feature space to be close and negative sample pairs to be far apart. A time queue is introduced to store the prediction results of multiple past cycles, and negative sample pairs that are labeled as the same class in historical predictions are excluded. The steps are as follows: the pseudo labels of each sample in the past T cycles are stored in the queue, and then all sample pairs that share the same pseudo label at least once in the past T cycles are excluded from the negative sample pair list. Accordingly, the following loss function is optimized:

[0032] ;

[0033] in is the index set of samples in the storage queue that have never shared the same pseudo label with the query sample in the past T cycles, q is the current query value, k+ is the positive sample key value, kj is the negative sample key value, is the temperature coefficient.

[0034] Furthermore, in step A5, in order to avoid posterior collapse, that is, the classifier tends to predict all samples as the same class, the following regularization method is used to promote the diversity of classifier predictions:

[0035] ;

[0036] And construct the overall loss function:

[0037] .

[0038] Furthermore, in step S2, the source domain model is a deep neural network including a feature extractor and classifier Two parts.

[0039] Furthermore, in step S1, after the distributed optical fiber sensor collects environmental vibration data, the original vibration data is first filtered and denoised for data preprocessing, and then labeled source domain vibration event data and unlabeled target domain vibration event data are constructed.

[0040] The beneficial effects of the present invention are as follows:

[0041] 1) Compared to existing technologies, this method improves the accuracy of cross-domain event classification. By aggregating knowledge from neighboring samples and optimizing pseudo-labels, the proposed method can effectively reduce the noise in pseudo-labels and generate more accurate pseudo-labels. Combined with a self-supervised contrastive learning framework, this method can more accurately classify events in the target domain, improving the overall performance of cross-domain event classification.

[0042] 2) Compared to existing technologies, this invention achieves model migration even when source data is inaccessible. By introducing a passive domain adaptation method, it enables cross-domain event classification even when source data is inaccessible. This approach is particularly suitable for applications with high data privacy and security requirements, such as underground pipeline intrusion detection and urban geological disaster monitoring.

[0043] 3) Compared to existing technologies, the computational cost is reduced. Compared to existing methods that rely on complex generative models, this invention achieves a low-computational-cost pseudo-label optimization strategy through neighborhood sample knowledge aggregation and a self-supervised contrastive learning framework. This method reduces the demand for computing resources and is more suitable for large-scale data processing and real-time monitoring in practical applications.

[0044] 4) Compared to existing techniques, this method enhances model robustness. By estimating the uncertainty of pseudo-labels and performing loss reweighting, the method can effectively address the noise in pseudo-labels and improve model robustness. Furthermore, the negative sample exclusion strategy further reduces the negative impact of pseudo-label noise during training, making the model more stable across different application scenarios. BRIEF DESCRIPTION OF THE DRAWINGS

[0045] Figure 1 A flow chart of event classification for a distributed optical fiber sensing cross-domain event classification method based on model migration according to the present invention;

[0046] Figure 2 This is a flowchart of the target domain model training in the present invention. DETAILED DESCRIPTION

[0047] The specific embodiments of the present invention will be described below with reference to the accompanying drawings.

[0048] The present invention discloses a distributed fiber optic sensing cross-domain event classification method based on model migration, aiming to solve the problems in the prior art of event classification methods based on knowledge migration that are inconvenient to transmit and share when accessing source domain and target domain data at the same time, as well as the problems of pseudo-label noise and high computational cost of generating models.

[0049] The present invention achieves efficient classification of vibration events in the target domain by training the model in the source domain and migrating it to the target domain. The overall flow chart of the present invention is as follows: Figure 1 shown.

[0050] The main steps of this method are as follows:

[0051] Specific implementation of step S1, data set construction:

[0052] First, distributed fiber optic sensors are arranged in the monitoring area to collect environmental vibration data and perform data preprocessing such as filtering and denoising on the raw vibration data to construct a labeled source domain vibration event dataset and an unlabeled target domain vibration event dataset.

[0053] Specific implementation of step S2: source domain model training:

[0054] Use the labeled source domain dataset to train a source domain model, which is a deep neural network including a feature extractor. and classifier Two parts.

[0055] Specific implementation of step S3, target domain model training:

[0056] The knowledge of the source domain model is transferred to the target domain. Only the source domain model and the unlabeled target domain vibration event data are used to train the target domain model. The flowchart of the target domain model training is shown in Figure 2 .

[0057] Among them, such as Figure 2 As shown, the implementation method of step A1 pseudo-labeling and optimizing it is as follows:

[0058] Pseudo-label generation: Use the trained source domain model to generate preliminary pseudo-labels for the unlabeled data in the target domain.

[0059] Pseudo-label optimization: Due to the difference in the distribution of target domain data and source domain data, the initial pseudo-labels may contain noise. A neighborhood knowledge aggregation strategy is introduced to optimize pseudo-labels by integrating information from similar samples in the target domain. Specifically, for each sample in the target domain, the cosine distance between samples is calculated to find the K closest samples in the feature space. The average score is then calculated using a soft voting strategy:

[0060] ;

[0061] represents the nearest neighbor sample set, Represents the probability output, and the superscript c indicates that the averaging operation is performed by class. Next, the argmax operation is used to obtain the corrected pseudo-label:

[0062] .

[0063] Step A2 measures the uncertainty of the pseudo-label and uses it to calculate the corresponding weight implementation:

[0064] To improve the reliability of these pseudo-labels, we employ an uncertainty estimation method based on the consistency of neighborhood predictions. The uncertainty of the pseudo-labels is measured by calculating the entropy of the predictions of the neighborhood samples, and the classification loss is reweighted based on the uncertainty. Pseudo-labels with low uncertainty (high confidence) are given higher weights, while pseudo-labels with high uncertainty (low confidence) are given lower weights, thereby reducing the impact of noise on model training. The specific steps are as follows:

[0065] For a target domain sample , calculate its average prediction score according to the formula, which can be regarded as a probability distribution. Then, calculate the entropy of this probability distribution:

[0066] ;

[0067] Where C is the number of categories. The entropy value is also scaled by the maximum value as follows:

[0068] ;

[0069] The weight of the sample can be calculated through the normalized entropy value:

[0070] .

[0071] Step A3: Implementation of constructing the self-training loss function:

[0072] To reduce the impact of noisy pseudo-labels, a negative learning strategy is introduced into the classification loss. That is, when calculating the classification loss, complementary labels are selected instead of correct labels, thereby reducing the negative impact of incorrect labels on model training. Therefore, the self-training loss function of the target domain data is as follows:

[0073] ;

[0074] in is a label randomly selected from a label set that does not contain the pseudo label of the current sample, that is, , It is a strongly enhanced image The probability output of The above-mentioned estimate Uncertainty weight.

[0075] Step A4, contrastive learning, uses multi-cycle historical pseudo-label information to exclude negative sample pairs that may belong to the same category:

[0076] The contrastive learning framework is used to regularize the target domain feature space, that is, to make the positive sample pairs in the feature space as close as possible and the negative sample pairs as far apart as possible, so as to ensure that samples of the same category are clustered together in the feature space; by introducing a time queue to store the prediction results of multiple cycles in the past and excluding those negative sample pairs that were marked as the same category in historical predictions, the accuracy of pseudo-labeling is further improved. The specific steps are as follows:

[0077] Store the pseudo-labels of each sample in the past T cycles in the queue, and then exclude all sample pairs that share the same pseudo-label at least once in the past T cycles from the negative sample pair list. Accordingly, optimize the following loss function:

[0078] ;

[0079] in is the index set of samples in the storage queue that have never shared the same pseudo label with the query sample in the past T cycles, q is the current query value, k+ is the positive sample key value, kj is the negative sample key value, is the temperature coefficient.

[0080] Step A5, constructing an implementation method of the overall loss function:

[0081] To prevent posterior collapse, where the classifier tends to predict all samples as the same class, the following regularization method is used to promote the diversity of classifier predictions:

[0082] ;

[0083] Self-training, contrastive learning, regularization and other loss functions constitute the overall loss function:

[0084] .

[0085] Specific implementation of step S4, target domain model deployment:

[0086] The trained target domain model is deployed in a distributed fiber optic sensing system for real-time event classification. By classifying newly collected target domain vibration event data, accurate identification and early warning of events such as underground pipeline intrusion, leak monitoring, road subsidence, and geological disasters can be achieved.

[0087] 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 scope of protection of the present invention.

Claims

1. A distributed optical fiber sensing cross-domain event classification method based on model migration, characterized by: The steps include: S1. Distributed fiber optic sensors are deployed in the monitoring area to construct a labeled source domain vibration event dataset and an unlabeled target domain vibration event dataset; S2, train a source domain model using a labeled source domain vibration dataset; S3, transfer the knowledge of the source domain model to the target domain, and train the target domain model based only on the source domain model and the unlabeled target domain vibration event dataset; S4, deploy the trained target domain model into the distributed fiber optic sensing system to perform real-time event classification on new data in the target domain; The method for training the target domain model in step S3 includes the following steps: A1. Build pseudo labels and optimize them; A2. Measure the uncertainty of the pseudo-label and calculate the corresponding weight; A3. To reduce the negative impact of incorrect labels on model training, a self-training loss function for target domain data is constructed. A4. Use multi-cycle historical pseudo-label information to exclude negative sample pairs that may belong to the same category; A5. Promote the diversity of classifier predictions through regularization methods and construct an overall loss function; In step A2, the uncertainty estimation method based on the consistency of neighborhood predictions measures the uncertainty of the pseudo-label by calculating the entropy of the prediction results of the neighborhood samples, and reweights the classification loss according to the uncertainty. The pseudo-labels with low uncertainty are given higher weights, and the pseudo-labels with high uncertainty are given lower weights. The steps are as follows: For a target domain sample , calculate its average prediction score according to the formula, which is defined as a probability distribution, and calculate the entropy of the probability distribution: ； Where C is the number of categories, and the entropy value is scaled by the maximum value. The scaling formula is as follows: ； The weight of the sample is calculated by the normalized entropy value. The calculation formula is as follows: ； In step A3, a negative learning strategy is introduced into the classification loss. That is, when calculating the classification loss, complementary labels are selected instead of correct labels. The self-training loss function of the target domain data is obtained as follows: ； in is a label randomly selected from a label set that does not contain the pseudo label of the current sample, that is, , It is an enhanced image The probability output of It is estimated Uncertainty weight.

2. The distributed optical fiber sensing cross-domain event classification method based on model migration according to claim 1, characterized in that: In step A1, the trained source domain model is used to generate preliminary pseudo labels for the unlabeled data in the target domain. Based on the neighborhood knowledge aggregation strategy, the pseudo labels are optimized by integrating the information of similar samples in the target domain. For each sample in the target domain, the cosine distance between samples is calculated to find K similar samples in the feature space, and the average score is calculated using the following soft voting strategy calculation formula: ； represents the nearest neighbor sample set, Represents the probability output, and the superscript c indicates that the averaging operation is performed by class; Use the argmax operation to get the corrected pseudo-label: 。 3. The distributed optical fiber sensing cross-domain event classification method based on model migration according to claim 1, characterized in that: In step A4, the contrastive learning framework is used to regularize the target domain feature space, so that positive sample pairs in the feature space are forced to be close and negative sample pairs are forced to be far apart. A time queue is introduced to store the prediction results of multiple past cycles, and negative sample pairs that are labeled as the same class in historical predictions are excluded. The steps are as follows: the pseudo labels of each sample in the past T cycles are stored in the queue, and then all sample pairs that share the same pseudo label at least once in the past T cycles are excluded from the negative sample pair list. Accordingly, the following loss function is optimized: ； in is the index set of samples in the storage queue that have never shared the same pseudo label with the query sample in the past T cycles, q is the current query value, k+ is the positive sample key value, kj is the negative sample key value, is the temperature coefficient.

4. The distributed optical fiber sensing cross-domain event classification method based on model migration according to claim 3, characterized in that: In step A5, to avoid posterior collapse, that is, the classifier tends to predict all samples as the same class, the following regularization method is used to promote the diversity of classifier predictions: ； And construct the overall loss function: 。 5. The distributed optical fiber sensing cross-domain event classification method based on model migration according to claim 1, characterized in that: In step S2, the source domain model is a deep neural network, including a feature extractor and classifier Two parts.

6. The distributed optical fiber sensing cross-domain event classification method based on model migration according to claim 1, characterized in that: In step S1, after the distributed optical fiber sensor collects environmental vibration data, the original vibration data is first filtered and denoised for data preprocessing, and then labeled source domain vibration event data and unlabeled target domain vibration event data are constructed.

Images