Technical Field

The present invention is a method for fusion of spatio-temporal unsupervised anomaly detection and supervised classification for high performance DAS systems.

Background

Supervised learning provides performance to a certain level in distributed acoustic systems due to geographical differences. In addition, system that can act adaptively according to geographical and other factors such as weather, soil type, urban activities, etc. are needed.

The application numbered W02020096565A1 is directed to a method of utilizing an acoustic sensing cable, such as a fiber optic DAS cable, in a borehole to detect microseismic events and to generate three dimensional fracture plane parameters utilizing the detected events. Alternatively, the method can use various categorizations of microseismic data subsets to generate one or more potential fracture planes. Also disclosed is an apparatus utilizing a single acoustic sensing cable capable of detecting microseismic events and subsequently calculating fracture geometry parameters. Additionally disclosed is a system utilizing a processor to analyze collected microseismic data to generate one or more sets of fracture geometry parameters.

In the state of art, the machine learning model exhibits a non-adaptive behavior depending on the data learned during the training phase. Since the usage areas of the DAS sensor cover different geographical areas, this approach is not sufficient in areas with many variable factors (weather, soil type, urban activities, etc.).

Summary The invention presents fusion of spatio-temporal unsupervised anomaly detection and supervised classification methods for high performance DAS systems. Supervised techniques based on CNN or RNN have shown good generalization performance in classification of DAS signals where spatio-temporal data come from a wide variety of geographic regions and events. On the other hand, these models require a dataset sampled from different conditions that include different soil types, weather effects, background noise sources. Modelling the effect of these conditions on training data is very complex. Instead, application of unsupervised techniques allows modelling these conditions in an online manner. Spatial and temporal nature of the DAS data can be modelled with unsupervised learning techniques and the different behaviors with respect to adjacent channels and past behaviours can be highlighted as an anomaly. In addition to classification probability, anomaly score of the signal can be evaluated the improve performance.

Brief Description of the Figures

Figure 1 shows an autoencoder architecture wherein input of the autoencoder is fed to output.

Figure 2 shows a diagram of fully independent classification and anomaly detection. Figure 3 show a diagram of anomaly detection and classification based on autoencoder embeddings.

Detailed Description

Each field has its norms, therefore, besides the classification of signals sensed by DAS, deciding the anomaly level of the detected activity with respect to the field conditions is required to improve detection and classification performance without false positive. Supervised learning provides reliable generalization for activity classification unless suffers field variations like soil types, weather conditions, or background noise. For sake of the specialized activity detection and classification, an online learning model, that learns field conditions and scores anomaly in conditions of the field, is proposed for DAS.

With the help of the spatio-temporal features, events are evaluated differently for each field and activity region of DAS. Thus, some events can be ignored in conditions of a specific region or field even they are threats to other areas. For instance, an expected event in the mining area is unexpected for agricultural areas. The model that specialized in an agricultural area, detects that event as a threat, on the other hand, the mining area model evaluates that event as usual. Consequently, custom models, that trained by unsupervised online learning, for each field capable of assessment of events on their conditions are obtained.

DAS sensor converts standard fiber optic cables to a linearly spaced acoustic array. Data collected by DAS sensors are in spatio-temporal form. These sensors are used in activity detection for long range up to 50 km. In such large areas, encountered different situations poses a challenging problem in activity detection.

In recent times, machine learning techniques, especially deep learning, overperforms other algorithms in sound activity recognition task. In general, these algorithms use 2D spectrogram features for classification with Convolutional Neural Networks (CNN). Variants of these approach provides high accuracy in sound classification. In DAS systems, this approach provides high accuracy if the training data is able to generalize possible conditions in the field. The performance of this approach highly depends on the consistency between training data distribution and test data distribution. Due to diversity of deployment fields of DAS sensors, this approach is insufficient in modelling new conditions. So an adaptive algorithm is required to capture changing conditions.

The main problem with the changing conditions in large outdoor fields is to model all possible conditions to train a supervised system. So unsupervised techniques are required to model these conditions in the field. In contrary to classification algorithm, this algorithm must be trained in an online manner to follow changing distribution of the input data.

In order to capture the shift in data distribution, it is necessary to model normal conditions. For a point sensor it is harder to verify the shift in conditions but DAS sensor acts like a sensor network that provides data from neighbour regions so that the results of sensor fusion provide consistent results. As a result, spatio-temporal form of DAS data allows fast and consistent adaptation to changing conditions.

In order to model the distribution of data, a function that transforms the high dimensional input data to low dimensional distribution is needed. Autoencoders are widely used to compress data. The purpose of the training is to reconstruct an input from a low dimensional latent space. In the training phase, input of the autoencoder is fed to output and the reconstruction error is used as loss function (Figure 1 ). For different applications structured latent space is required for regularization so that the latent space is forced to be sparse (sparse autoencoders) or the latent space is forced to fit a probability distribution (variational autoencoders). So variational autoencoders are able to map the input data to a probability distribution. For the anomaly detection task, the latent space of the variational autoencoders presents a low dimensional probability distribution.

Anomaly detection part fits a joint distribution to inputs where the distribution models spatial and temporal information in data. If a channel behaves different from its neighbours and its past data, anomaly score will be higher at that channel.

Two algorithms to fuse the classification and anomaly detection outputs are proposed.

Fully Independent Classification and Anomaly Detection: Classification and anomaly detection task can be implemented independently. In this case, anomaly detection algorithm needs to be designed independently and the outputs must be fused (Figure 2).

Anomaly Detection and Classification based on Autoencoder Embeddings: Both of the classification and anomaly detection algorithms maps the input data to a low dimensional space so the dimensionality reduction parts can be used jointly to reduce computational load (Figure 3). In this case:

• Train an autoencoder for dimensionality reduction.

• For classification, fit a supervised algorithm that uses latent space of input data and labels.

• For the anomaly detection, use latent space as probability distribution and generate anomaly score.