ADVERSARIAL-NETWORK (GAN)

Field of the Invention

The present invention relates to a method for analyzing Distributed Acoustic Sensing (DAS) data, and particularly to methods and systems for the generation of training sets for training a deep-learning-based DAS analysis system.

Background of the Invention

In recent years significant progress has taken place in fiber optic Distributed Acoustic Sensing (DAS), which use a fiber-optic cables to provide distributed strain sensing. DAS systems are capable of continuously monitoring acoustic signals and vibrations in harsh environments along tens of kilometers with high sensitivity and high update rate. This is accomplished by using a conventional telecom-type optical fiber whose Rayleigh backscatter profile is measured repeatedly by a reflectometric method. A DAS system typically acquires ~1000 complex backscatter profiles per second and processes them in real-time. The data rates of optical data are in the order of lOMbyte/s, which can accumulate to roughly 1Tbyte of data per day. In order for this huge amount of data to be useful, it is imperative to develop automatic, efficient and accurate tools for processing the recorded signal. Specifically, detection, classification, and localization of recorded events are of the utmost importance.

Due to the complexity of the optical data, processing of the data and detection of the events of interest can be time and computational capacity consumer. However, so far the art has failed to provide an efficient method and system that can grant a sufficient solution which will improve DAS analysis systems ("DAS classifiers"). It is therefore clear that it would be highly desirable to provide a training system, which may improve the performance of a DAS system while saving significant resources of time and labor.

It is a purpose of the present invention to provide a method and system, which is effective, fast and low-cost, which generates a data train-set for the effective training phase of deep- learning-based DAS analysis systems. It is another purpose of the invention to provide a novel hybrid method that uses a second modality (such as a camera), to aid in the DAS analysis system training phase in order to increase the robustness of DAS classification.

It is a further purpose of the invention to provide a hybrid-modal architecture for a long- deployed DAS system, which can be utilized in far-reaching system operative over tens of kilometers in changing terrains.

Additional advantages and purposes of the invention will become apparent as the description proceeds.

Summary of the Invention

The present invention relates to a Distributed Acoustic Sensing (DAS) system employing a Generative Adversarial Net (GAN) process to generate train-sets from a computer simulation of the DAS system, said train-set being adapted to train an Artificial Neural Network (ANN) to classify events taking place in the vicinity of the fiber optic of said DAS, and to a method for performing said training.

In one embodiment, the system employs a test generation apparatus comprises a memory and a processor. The memory is configured to store actual data sets acquired from an optical fiber by a Distributed Acoustic Sensing (DAS) system. The processor is configured to use a Generative Adversarial Network (GAN) for generating data sets for training the DAS analysis system ("DAS classifier"). The processor carries out a two-stage process, comprising a GAN training stage followed by a test generation stage.

In the GAN training stage, the processor trains the GAN to modify the first computer- simulated data sets, so as to mimic the actual data sets acquired by the DAS system. In the test generation stage, the processor generates second computer-simulated data sets, modifies the second computer-simulated data sets using the trained GAN, and outputs the modified second computer-simulated data sets for training the analysis system. Like the original simulated data sets, each of the modified second computer-simulated data sets is labeled with one or more simulated events-of-interest. When operating according to the invention, the GAN produces simulated data sets that appear realistic as if acquired by the DAS system, and are labeled with eve nts-of-inte rest. Such data sets are able to train the DAS analysis system to achieve high quality and accuracy, e.g., high detection probability and low false-alarm probability.

In some embodiments, additional labeling of events-of-interest is provided, originating from a different modality. For example, for a given segment of the fiber, events-of-interest may be identified reliably in video images of the vicinity of the segment, or in audio recorded in the vicinity of the segment. Such events of interest can be used for labeling actual data sets acquired by the DAS system for that segment, thereby generating additional training data for the DAS analysis system.

In some embodiments, for an intermediate segment located between first and second segments, the processor labels an actual data set acquired by the DAS system for the intermediate segment, based on labeled data sets available for the first and second segments. For example, the processor may perform this task by applying a triple cycle-GAN to the first segment, the intermediate segment, and the second segment.

All the above and additional characteristics and advantages of the invention will be further understood from the following illustrative description of embodiments thereof.

Brief Description of the Drawings

In the drawings:

Fig. 1 shows and examples of an input image for a 5km sensing fiber (a) vehicle and (b) footsteps;

Fig. 2 shows and examples of an input image for a 20km sensing fiber (a) vehicle and (b) footsteps;

Fig. 3 schematically shows an optical setup according to one embodiment of the invention;

Fig. 4 is a simulation of a seismic footstep signature from a 20km DAS system (a) and its refined image (b); performed in a system according to an embodiment of the invention;

Fig. 5 schematically illustrates a Video-DAS hybrid modal according to another embodiment of the invention; Fig. 6 illustrates the "blind spot" along a DAS route with cameras; and

Fig. 7 illustrates Triple cycle-GAN setup according to still another embodiment of the invention, including six generators and three discriminators.

Detailed Description of the Invention

Broadly speaking, the invention employs a Generative Adversarial Network (GAN) for generating data-sets for training a Distributed Acoustic Sensing (DAS) analysis system (also referred to herein as "DAS classifier"), which receives data acquired from an optical fiber by a DAS system. This is performed in two steps, i.e., performing a GAN training stage followed by a test generation stage. According to one embodiment of the invention the DAS classifier on simulated data that were generated by GAN.

DAS data complexity

Experimental data needed for the DAS system to operate efficiently are difficult to acquire and/or simulatey Many current DAS systems are able to measure both the amplitude and phase of the backscatter signal, and types of data can be used as input to the detection and classification signal processing layer, both having their respective advantages and disadvantages. The use of the amplitude is simple and robust. The use of phase, on the other hand, requires more elaborate processing but can provide linearity of the signal with concerning the excitation and enhanced signal-to-noise ratio (SNR). A disadvantage of phase processing is the risk of phase wrapping and phase ambiguity when the excitation is strong.

Using Artificial Neural Networks for processing of the raw data according to the invention enables a straightforward fusion of both modalities. The complex fiber profile is parsed to segments and the amplitude in each segment is saved as an image. Another image is generated for each segment from the phase difference between consecutive resolution cells (differential phase). The two images are normalized and appended to create a 'two-channel image' (similarly to RGB images but with two channels instead of three). The two-channel image, corresponding to a ~90meter fiber segment and 0.5s time interval, is the input to the network. With update rates of 2000 scans/s for the 5km and 1000 scans/s for the 20km experiments, the corresponding input image size 50x500 and 32x250 pixels, respectively. Having short time intervals reduces the processing latency and makes the method compatible with the real-time operation. It enables capturing fast-occurring details of the seismic excitations but does not require substantial human intervention in each image. Accordingly, the tested network does not enjoy longer time-scale-features, primarily cadence rate. An additional classification layer, which uses this information, significantly improves its performance.

GAN specifications

There is obviously a difference between the simulated data and the true data. To overcome problems that may arise as a result of simulated data, the invention uses a GANs neural network that adjusts the simulated data such that it becomes more similar to the true data. This is performed in two stages, comprising a GAN training stage followed by a test generation stage.

The GAN architecture used according to one embodiment of the invention is the Conditional GAN (also referred to hereinafter as "C-GAN" and described in L Shiloh, A. Eyal, and R. Giryes, "Deep Learning Approach for Processing Fiber-Optic DAS Seismic Data," in Optical Fiber Sensors International Conference (2018).). This version of GAN connects the generator and discriminator through some extra information in a conditional manner. In this case, the generator learns to generate from a fake sample with a specific condition rather than a generic sample from unknown noise distribution. Of course, the C-GAN is only one example of a suitable GAN architecture and the invention is not limited to it or to any specific GAN architectures.

A specific C-GAN variant combines simulation and experimental data (as described, for instance in A. Shrivastava, T. Pfister, O. Tuzel, J. Susskind, W. Wang, and R. Webb, "Learning from Simulated and Unsupervised Images through Adversarial Training," CVPR 2017 (2016)) and is referred to herein as "SimGAN." In the SimGAN methodology, the generator is trained to transform simulation data to appear realistic, while the discriminator is trained to differentiate between generated data and real data. It is very useful for applications that suffer from an insufficient number of training examples that cannot be artificially increased using simulations (due to modeling complexity). It enables the increase of the size of the database for further analysis. In the description to follow the generator is denoted G _g , and the discriminator D , where Q and f are their weights, respectively. The real experimental data is denoted as y Y , and the simulation data as xe . The cross-entropy loss used to optimize the discriminator is formulated as follows:

The generator is trained to accurately mimic genuine data using two weighted loss functions:

L ₀ (Q) = -å log (i - o, (a, (*))) + l ||o _B (*) - x| li 2

The first term is the cross-entropy term that aims at "fooling" the discriminator, i.e., making the generated data look like the experimental one. The second term is a regularization term weighted by a scalar A that does not allow the generated data to deviate much from the input simulated data. The purpose of this term is to keep the class information in the generated sample, i.e., if the input corresponds to a step, we want the generated realistic sample to contain a step as otherwise we will not be able to use the same labels of the simulated data. We use the U norm (M. Schmidt, G. Fung and R. Rosales, "Fast Optimization Methods for LI Regularization: A Comparative Study and Two New Approaches", in Machine Learning: ECML, (2007), pp. 286-297.), to allow robustness in the deviation from the input as is common in image rendering.

After converging to an optimal generator, the classification network is initially trained using the refined simulation data. The next phase is finetuning the network using a smaller experimental data-set. The purpose of this second training stage is to finetune the network to a more optimal working point for the system's data.

Network architecture

The following description refers to an architecture according to one embodiment of the invention, which is detailed for illustrative purposes and which is not intended to limit the invention in any way. As will be apparent to the skilled person, various elements can be replaced with alternative ones in alternative embodiments, and therefore this description should be understood to be provided solely to exemplify the invention.

The task of classifying seismic events from the input DAS is essentially composed of classification image segments. According to one embodiment of the invention the network is chosen to be a convolutional neural network-based (CNN). The Artificial Neural Networks (ANN) model employed according to one embodiment hereof is the commonly used network architecture for classification, segmentation and image denoising. In one embodiment of the invention the architecture is based on Oxford's VGG16 (K. Simonyan and A. Zisserman, "Very Deep Convolutional Networks for Large-Scale Image Recognition," in ICLR (2015), pp. 1-14.), which is known for its success in classifying the ImageNet dataset. An additional convolutional layer is added as an input layer to match picture size, and another fully- connected layer at the end (with ReLU ( G. E. Dahl, T. N. Sainath and G. E. Hinton, "Improving deep neural networks for LVCSR using rectified linear units and dropout.", in ICASSP (2013), pp. 8609- 8613.) as the activation function and batch normalization ( S. Ioffe and C. Szegedy, " Batch Normalization: Accelerating Deep Network Training by Reducing Internal Covariate Shift", in CoRR (2015).) for regularization).

The full architecture of what is referred to as FiberNet is detailed in Table 1 below, and the optical setup is shown in Fig. 3: Optical Interrogator 30, Waveform Generator 31, Digital Storage Oscilloscope 32, Sensing Fiber 34, buried fiber in the field 33.

Table 1

A Convolutional Neural Network, that is specifically designed to deal with the variability of 2D shapes (Conv2D), is described in a paper by Y. LeCun, L. Bottou, Y. Bengio and P. HAffner, "Gradient-Based Learning Applied to Document Recognition", in Proc. Of the IEEE (1998), pp. 2278-2324.

FC1, FC2 and FC3 referred to in Table 1 above, are described in a paper by D. W. Ruck, S. K. Rogers and M. Kabrisky, "Feature Selection Using a Multilayer Perceptron", in Journal of Neural Network Computing (1990), pp. 40-48.

The VGG16 convolution layers' weights are initialized by the ImageNet pre-trained values, and the other layers are initialized with random weights. The input images are normalized ±1.

The GAN's discriminator model has FiberNet architecture. To train a multi-class GAN, the final fully-connected layer is implemented with 1+N _c | _as ses outputs to differentiate between noise, footstep, vehicles, and simulations. The generator consists of a 6-layer residual network, inspired by its performance in image denoising. A 3x3 convolution kernel with 65 filters is used for each layer when one of them is used to calculate the residual image. The weight l from equation (2), is set to 10 ⁵ . In each GAN training step, the generator is trained on 12 batches and the discriminator is trained on one batch. Each batch consisted of 30 signals. An example of the refiner effect on a footstep simulation is shown in Fig. 4. In this figure the two image channels are concatenated for visibility: the upper image is the power and the lower image is the differential phase. Fig 4(a) is a footstep simulation, and Fig. 4(b) shows its refined state after inference through the generator. The change in signature and background noise is notable.

Training the FiberNet classification network consists of two phases: first using refined simulation data and then fine-tuning with experimental data. All training is also done on 30- signal batch size, using data augmentation for generalization. The augmentation includes flipping the image along the fiber distance axis and random translation on both axes. Optimization is performed using Stochastic Gradient Descent (SGD). Simulation dataset

In the illustrative example discussed herein two datasets were used to train the GAN and the classification networks: datasets recorded in field experiments and datasets generated by computer simulations. The simulation datasets comprised ~70k images for the 5km fiber and ~150k for the 20km fiber. The experimental datasets comprise ~49k images taken with the 5km sensing fiber and ~10k images of the 20km fiber. The smaller number of signals in the 20km case is due to uncontrolled factors in the experimental field. These induced multiple noise sources that degraded the signal's SNR and therefore additional simulations where added to the training phase.

The test sets were recorded on different days than the train datasets and generated by different subjects. Specifically, the footsteps correspond to a ~75kg person in the test set as opposed to a ~60kg person in the train set. The vehicle was a small campus car, as opposed to Renault Kangoo in the train set. Test sets consisted of 135 images per class for the 5km fiber and 224 images per class for the 20km fiber.

Experiments

Field experiments were conducted at the backyard of Tel-Aviv university school of Electrical Engineering. 200m of telecom-type optical cable were buried ~0.5m below the ground surface. The rest of the cable was deployed to a lab were additional fiber spools could be concatenated to it. Interrogation was performed by a self-built OFDR DAS system, comprising an ultra-coherent laser with central wavelength of 1550.12nm. Two sensing fibers were used in this experiment - one with a 5km spooled fiber preceding the buried cable and the other with a 20km spooled fiber. The optical signal was sampled at 1.25GSamples/s, and the processing windows was 13.1ps. The scan rates were 2kHz and 1kHz for the 5km and 20km fibers respectively .

As mentioned above, the lengths of the fiber segments used as inputs to the ANNs corresponded to ~90m for the 5km fiber and ~170m for the 20km one. Examples of an input image for 5km sensing fiber and 20km sensing fiber are presented in Fig. 1 and Fig. 2 respectively. In both figures, (a) corresponds to a recording of a vehicle at the vicinity of the fiber and (b) corresponds to a footstep. The low SNR in the case of the 20km fiber, which resulted mainly from very noisy campus environment and non-linearity in the frequency scan of the laser, is evident.

To generate synthetic datasets of events of interest, a computer simulation was performed. The computer simulation had two main parts:

1. An optical part - repeatedly produces a complex backscatter profile of a synthetic sensing fiber (similar to the description in H. Gabai and A. Eyal, "On the sensitivity of distributed acoustic sensing," Opt. Let. 41, 5648-5651 (2016).). After each cycle, the phases of the reflection coefficients were updated to simulate the externally induced acoustical perturbations. To generate the complex backscatter profile, the fiber was represented by its impulse response. The impulse response was generated by dividing the fiber into small sections of length 8cm. For each section, a backscattering coefficient was drawn from a complex normal distribution. The optical loss of the fiber (-0.2dB/km) was taken into consideration by multiplying each coefficient with its appropriate decay term. The backscatter signal, at the input of the coherent receiver, was obtained by convolving the fiber's impulse response with the input waveform. Once the backscatter signal is known it is added to the reference (the input waveform). The square of the magnitude of the resulting signal was calculated to yield the detector output. Finally, the Fourier transform of the detector output yielded the simulated complex backscatter profile of the fiber. Two types of noise signals were introduced into the optical part: additive detection noise and phase noise. The additive detection noise, which represents shot noise and thermal noise, was added to the generated detector output. Laser phase noise was introduced by adding a noise term to the, otherwise linearly swept, instantaneous frequency of the laser.

2. A seismic part- considered seismic excitations such as human footsteps, vehicles et cetera at the fiber vicinity. The excitations were introduced into the simulation as modifications to the phase response of the sensing fiber. The modifications were made in each scan period. The temporal and spectral signatures of the seismic events were determined empirically from the analysis of the experimental results. For example, footsteps were modeled as wavelets with center frequencies uniformly distributed in the range 55-60Hz and with duration uniformly distributed between 13.3ps and 13.5ps. Synthetic vehicles' signals were made by generating white Gaussian noise and filtering it to the range 150-270Hz, according to the power spectral density of the experimental samples. An example of a footstep simulation at the vicinity of 20km fiber is shown in Fig. 4(a).

Tables 2-5 summarize the classification results. An initial attempt to detect human footsteps on a 5km long fiber, based on the experimental dataset alone, reaches an accuracy of ~70% (Table 2). That was due to the fact that the available experimental dataset was not large and diverse enough. In addition, the data suffers from label noise, which in some cases degrades the DNN performance. In our binary classification case, where the noise factor is not known, it may corrupt the classification accuracy. Analyzing the events based on the simulation dataset alone resolves in a close 50-50 accuracy. This demonstrates that an attempt to simulate the fiber sensor response to a seismic event was not accurate. When classifying with the refined version of the simulated dataset, the accuracy of the test set increases to ~67%. Interestingly, increasing the simulation dataset by a factor of 1.4 increases the accuracy to ~78%. By this logic, increasing the simulations dataset even more may be advantageous. Fine-tuning this network with the experimental data increased the accuracy to 94%.

Table 2: 5km Detection results under different training setups

The confusion matrices at the right column of Tables 3 and 4 show cross-class classification between the three classes. The rows (from left to right) and columns (from top to bottom) in the matrices correspond to noise, footsteps and vehicles respectively.

Table 3: Classification accuracy for a 5km fiber for 3 classes (noise, footsteps and vehicles)

Table 4: 20km Detection results under different training setups

The confusion matrix of Tables 2 and 4 have two classes for noise and footsteps respectively. The confusion matrix in Table 2 suggests that the refiner accomplishes a better realization of the fiber background texture while the finetune phase gives an advantage to the detection of the steps. Adding a third class (vehicles) produced the results summarized in Table 3. Despite the limited scale of the experimental dataset, the invention leads to a remarkable increase in classification accuracy. Training the network on experimental data only classifies correctly with high probability the footsteps and the vehicles (96% and 99.3% respectively) but the false alarm rate is very high (54%). Training on simulation data alone yields poor performance. Once again, an increase in accuracy is observed when refined simulation data is used. Further improvement is obtained by fine-tuning this network with experimental data. Classification accuracy for footsteps achieves 94% and 100% for vehicles. The false alarm rate is reduced to 45%. While a low false alarm rate is desired, the current improvement shows the potential of the proposed method and its ability to produce valid train sets.

The detection and classification results for the 20km sensing fiber experiment are summarized in Tables 4 and 5, respectively.

Table 5: Classification accuracy for a 20km fiber for 3 classes (noise, footsteps and vehicles)

The impact of the smaller experimental dataset and decreased SNR is evident. The detection accuracy of footsteps from ambient noise is increased from 83.7, based on experimental data alone, to 89.3% when training on refined simulation data and fine-tuned with a small experimental dataset. It can be seen that training on simulations alone is irrelevant and achieves random accuracy. The classification accuracy of 3 classes, noise footsteps and vehicles, summarized in Table 4, has also improved from 42% to 80.2% when operating according to the invention.

Hybrid network

In another embodiment thereof, the invention can include more information from additional sensors in a hybrid modal architecture. According to this embodiment, the additional sensors add a more robust modality to help train the DAS classifier. For example, it can utilize a video camera as the second modality (in conjunction with DAS). The video modality facilitates the "self" -training ability of the network and helps with the insufficient tagged database challenge.

A schematic illustration of a specific hybrid network embodiment of the invention is shown in Fig. 5: video sensor 50, video network 51, DAS network 52, DAS Interrogator 53. The first modality is the optical fiber sensor, the second modality is a video camera, provided with a very advanced algorithms for classification of its output. For example, deep learning architectures which detect and recognize pedestrians and vehicle, such as the current state- of-the-art architecture for object detection in the Kitty database for car view scenes- MMLab PV-RCNN (S. Shi, C. Guo, L Jiang, Z. Wang, J. Shi, X. Wang and H. Li, "PV-RCNN: Point-Voxel Feature Set Abstraction for 3D Object Detection", in arXiv preprint arXiv:1912.13192 (2019)). Hence, according to a particular embodiment of the invention a camera is directed to a segment of the buried fiber. In this embodiment the classification network is a "self learning" one, in which labels from the video camera stream are used to train the DAS net. The confidence levels of video-based classification algorithms are high, and so they can be used as a reliable source of tagging for the DAS data. Together, these two modalities act like a semi-automatic system for training a classification network based on DAS data.

By using only a few cameras the DAS net may be trained at a given location and the local results may be propagated to the entire length of the fiber. Besides, this can be performed with other modalities that have known good classifiers and is not limited to the image modality. For example, a microphone that distinguish events based on their sound signatures can be used.

To illustrate the above in greater detail, denote the pre-trained second modality and confident model, by W. The input detected by this sensor is denoted by x and its estimated label as y— W(c). The DAS model is denoted by F with weights f, and its input image is of size M X N X 2, denoted by x. Its classification is denoted by y = F c ). The training phase of the DAS classification model consists of minimizing the following supervised loss function:

The confident classification network, W, is not trained and can be of any architecture. For example, computer vision architectures for video processing have shown excellent object detection and classification. Architectures such as Fully Convolutional Networks (FCN) (A. Krizhevsky, I. Sutskever, and G. E. Hinton, "ImageNet Classification with Deep Convolutional Neural Networks," Adv. Neural Inf. Process. Syst 1-9 (2012).) , U-Nets (O. Ronneberger, P. Fischer, and T. Brox, "U-Net: Convolutional Networks for Biomedical Image Segmentation," Med. Image Comput. Comput. Interv. - MICCAI 2015 234-241 (2015).) etc. If an acoustic sound is used as a second modality, Wavenet has been shown to have good classification abilities (A. Van den Oord, S. Dieleman, H. Zen, K. Simonyan, O. Vinyals, A. Graves, N. Kalchbrenner, and K. Senior, Andrew W. KorayKavukcuoglu, "WaveNet: A Generative Model for Raw Audio," in SSI/I (2016), pp. 1-15.) . Of course, the above are only illustrative examples and the skilled person will recognize alternative useful architectures.

At the end of the training phase, a trained DAS classification network, F, can be used in real time without the use of the second modality.

At the end of the training phase, a trained DAS classification network can be used in realtime without the use of the second modality.

Triple-cvcle-GAN architecture

According to a further embodiment of the invention, DAS systems with tens of kilometers in changing terrains for a long-deploy needs can be used, with discrete sensors which must cover the whole route. This significantly increases the complexity of the system, due to the numerous sensors needed and their power, network and analysis requirements. To mitigate this demand, one embodiment of the invention combines the two methods described above:

1. GAN architecture to enable the training of a DAS classifier based on computer simulation.

2. Hybrid modal architecture to increase DAS classification's robustness.

According to this embodiment, the fiber route is not entirely covered by complementary sensors and the classifier of the uncovered fiber segments, or "blind spots", are trained using a triple-cycle-GAN architecture. The cycle-GAN architecture, which has been introduced in 2017 by Efros et al (J. Zhu, T. Park, I. Phillip, and A. A. Efros, "Unpaired Image- to-lmage Translation using Cycle-Consistent Adversarial Networks," in ICCV (2017).) has shown astonishing results in the fields of image rendering and generation, object configuration and more. According to this embodiment of the invention it is combined with the fiberGAN described above, to generate new-labeled data along a fiber, given that some annotated data at a few spots of the fiber are available.

This embodiment is advantageous for two main reasons: First, different seismic wave propagation properties are expected when comparing the ground of the covered fiber segment and the ground in the blind spot. This is likely to affect the event's seismic signature and make the trained classifier in the covered segments irrelevant for the blind spots. A second reason is the optical signal degradation originated by the propagation in the fiber. This may cause an increase in noise sources and a decrease in SNR as the distance increases.

In an embodiment in which the discrete sensors are cameras, this is illustrated by Fig. 6: DAS Interrogator 60, video sensor segment A 61, video sensor segment C 62, buried optic fiber 63, which schematically shows covered and uncovered fiber segments. Fiber segments A and C are covered by the camera's field of view, while segment B is not. The covered A and C segments' classifiers are optimally trained using the hybrid-modal method described above, while segment B cannot be. Therefore, in order to train a reliable classifier for the uncovered fiber segment, the trained hybrid-modal networks (which provide a confident class estimation for the DAS data) from adjacent covered segments are used in a triple cycle-GAN architecture, the architecture of which is schematically shown in Fig.7: Discriminator A 70, data from segment A 71, data from segment B 72, Discriminator B 73, data from segment C 74, Discriminator C 75. In this network, the GAN is used to generate annotated data for segment B from the annotated data provided in segments A and C. In the training of this network, unlabeled data was used for segment B. When annotated data were available from neighboring segments A and C, then both were used to generate labeled data for B. If only labeled data from segment A was available, then computer simulated data (that are labeled) was used instead of the data of segment C. In other words, in case one of the two adjacent segments' data is not available with the hybrid-modal data (e.g., the one of segment C), it may be replaced by computer simulations of the blind spot segment. This is the same data used in the "Simulation dataset" for generating labeled data from the simulation. The generator GCB, detailed hereinafter, will act as refiner similar to the proposed architecture, and learn to transform the simulations to appear more realistic. The difference here is that both simulated and real data from a neighbor segment are used to improve the data generation.

When operating according to the triple cyclic GAN embodiment, the network learns to transfer features from DAS data measured from covered adjacent segments and/or computer simulation data, to the blind spot. At the end of the training phase, two refiners are optimized to create a realistic dataset that is then used for training a classifier for blind spot segments .

For each segment /, / e {A, B, C}, denote data from segment / as r _j G Jl _j , and its discriminator as D _j . The generators are denoted as GJJ, where the data is transformed from distribution / to distribution ].

Training an original cycle GAN consists of minimizing three loss components: two adversarial GAN losses similar to Eq. (l)+(2) above, and another cycle consistency term formulated as:

When operating according to the triple cycle GAN architecture embodiment, six generators are trained with combined loss terms. A complete cycle is achieved through three generators, thus the cycle consistency term is formulated as: The discriminators, detailed in the paper by J. Zhu, T. Park, I. Phillip, and A. A. Efros, ("Unpaired Image-to-lmage Translation using Cycle-Consistent Adversarial Networks," in ICCV (2017)),, differentiate between real and fake data. In this case, each discriminator also differentiates between real and fake data, however, the fake data consists of two groups. For example, D _B will differentiate between real DAS data from segment B, r _b , and fake data {^ _CB ( ^r c) _> 6AB (. ^T _O )}· In addition, thanks to the hybrid modality and computer simulations, the classes in the cycle GAN between AC (G _AC , G _CA , D _A , D _C ) are known. Therefore, a conditional loss term can be used, as explained above.

Hence, the full objective is a weighted sum of all the terms:

The workflow for the triple-cycleGAN-hybrid combination is as follows:

1. Train a hybrid-modal network for the covered segments;

2. Train the triple-cycle-GAN architecture, using the networks from stage 1;

3. Create a dataset for the DAS data originated from the blind spot segments based on the relevant two generators from stage 2.

4. Train a DAS classifier for the blind spots using the dataset from stage 3.

All the above description and examples have been provided for the purpose of illustration and are not intended to limit the invention in any way. The skilled person appreciates from the description above, that a GAN-trained network has significantly improved performance with respect to classifiers that were trained with simulation data only or experimental data only, as illustrated by the field experiments with 5km and 20km long sensing fibers described above, with footsteps and vehicles excitations at their vicinity. As it will further be appreciated by the man of the art, short time windows and narrow spatial segments make the invention compatible with real time operation and localization abilities.