Detection method and system for distributed optical fibre sensing measurements

Abstract

The invention provides a method and system for detecting a signal response in distributed optical fibre sensing data. The method comprises receiving, from a distributed optical fibre sensing system, distributed optical fibre sensing data, where the distributed optical fibre sensing system comprises a plurality of spatial channels, each spatial channel associated with a respective scattering location along an optical fibre of the distributed optical fibre sensing system. The distributed optical fibre sensing data comprises a plurality of measurement signals as a function of time, each of the plurality of measurement signals corresponding to a respective one of the spatial channels. The method further comprises providing the distributed optical fibre sensing data to a machine learning model, where the machine learning model is configured to detect in the distributed optical fibre sensing data a signal response corresponding to an event of a predetermined type. The machine learning model was trained using training data comprising simulated signal responses for events of the predetermined type. If the machine learning model detects a signal response corresponding to an event of the predetermined type, the method comprises outputting, by the machine learning model, information indicative of the detected signal response. The invention further provides a method and system for training the machine learning model using simulated distributed acoustic sensing data.

Description

[0001] DETECTION METHOD AND SYSTEM FOR DISTRIBUTED OPTICAL FIBRE SENSING MEASUREMENTS

[0002] Field of the Invention

[0003] The present invention relates to a method of, and system for, analysing data received from a distributed optical fibre sensing (DOFS) system. In particular, the invention provides a method of, and system for, detecting signal responses in DOFS data which correspond to an event of a predetermined type. Also provided is a method of, and system for, training a machine learning model for detecting signal responses in DOFS data which correspond to an event of a predetermined type.

[0004] Background

[0005] There exist various distributed optical fibres sensing (DOFS) techniques for measuring properties of an optical fibre by interrogating the optical fiber with a transmitted optical signal.

[0006] Distributed Acoustic Sensing (DAS) is an established DOFS technique with several commercial systems available. In these systems, a pulse or pulses of laser light are launched into a length of optical fibre and the light that is scattered within the fibre is analysed in order to derive the nature of the acoustic environment, i.e. any physical vibrations, of the fiber transducer. In particular, these systems typically make a measurement of the acoustic strain environment of an optical fibre transducer using an optical time domain reflectometer (OTDR) approach. This gives a differential strain measurement as a function of position along the optical fibre.

[0007] As an optical fibre is manufactured it is cooled or quenched from a high temperature as it is drawn. This process leads to the presence of small variations in the density of the optical fibre. These tiny variations in density equate to variations in the effective refractive index of the fibre. These discontinuities lead to scattering of laser light passing through the optical fibre, particularly by Rayleigh scattering. The amplitude of the scattering follows a Rayleigh distribution, but the phase angle of the scattering is uniformly distributed around a unit circle, i.e. — TT < <t> < TT where <t> is the phase angle.

[0008] In the OTDR approach, a coherent light pulse is sent along the optical fibre. As the pulse travels along the optical fibre, it is scattered at scattering locations along the optical fibre, and backscattered signals are received at a detector stage. Scattered (i.e. backscattered) signals arising from different scattering locations along the optical fibre will be received at the detector stage at different times, such that each scattered signal can be assigned to its corresponding scattering location (or “spatial channel”) based on its time of receipt at the detector stage. In this manner, measuring the scattered signals over time enables a strain, and hence an acoustic field, at the different scattering locations along the optical fibre to be monitored over time.

[0009] Another example of a DOFS techniques is Distributed Strain and Temperature Sensing (DSTS). DSTS enables simultaneous measurement of temperature and strain in an optical fibre, using a laser pulse which is transmitted along the optical fibre. The scattering mechanism of interest for DSTS is Brillouin scattering.

[0010] DOFS techniques provide the ability to monitor optical fibre properties over long distances, for example over distances of 10 km to 80 km or more. As a result, DOFS techniques find applications in a wide range of fields, such as pipeline monitoring, train track monitoring, and road traffic monitoring, among others.

[0011] The present invention has been devised in light of the above considerations.

[0012] Summary of the Invention

[0013] At its most general, the present invention provides a method of (and system for) analysing distributed optical fibre sensing (DOFS) data to detect a signal response (or signature) in the DOFS data which corresponds to an event of a predetermined type. In particular, the DOFS data is provided to a machine learning model, which is trained to detect signal responses in the DOFS data corresponding to events of the predetermined type. Advantageously, compared to conventional analysis techniques which can be highly computationally intensive, the invention enables DOFS data to be analysed substantially in real-time, thus enabling substantially realtime event detection based on received DOFS data. The invention enables sensitive and accurate detection and classification of signal responses in DOFS data, which can be implemented using relatively inexpensive computer hardware, thus facilitating deployment of detection techniques in the field.

[0014] According to a first aspect of the invention, there is provided a method of detecting a signal response in distributed optical fibre sensing data, the method comprising: receiving, from a distributed optical fibre sensing system, distributed optical fibre sensing data, wherein the distributed optical fibre sensing system comprises a plurality of spatial channels, each spatial channel associated with a respective scattering location along an optical fibre of the distributed optical fibre sensing system, and wherein the distributed optical fibre sensing data comprises a plurality of measurement signals as a function of time, each of the plurality of measurement signals corresponding to a respective one of the spatial channels; providing the distributed optical fibre sensing data to a machine learning model, wherein the machine learning model is configured to detect in the distributed optical fibre sensing data a signal response corresponding to an event of a predetermined type, wherein the machine learning model was trained using training data comprising simulated signal responses for events of the predetermined type; and if the machine learning model detects a signal response corresponding to an event of the predetermined type, outputting, by the machine learning model, information indicative of the detected signal response.

[0015] The method is a computer-implemented method, which can be implemented using any suitable computing system. For example, the method may be implemented by an analysis system (e.g. as described in the second aspect of the invention, below), which includes a computer memory storing instructions for performing the method, and a processing device (e.g. including one or more processors) configured to execute the instructions to perform the method.

[0016] The DOFS data may be received from a detector stage of the DOFS system. The DOFS system may comprise any suitable type of optical fibre sensing system, such as an optical time domain reflectometer (OTDR), a distributed acoustic sensing (DAS) system, or a distributed strain and temperature sensing (DSTS) system.

[0017] The DOFS system may be configured to launch a test signal along the optical fibre, and to receive at the detector stage a plurality of scattered signals, each of which was scattered at a respective location along the optical fibre. The test signal may be a pulsed test signal. Each of the scattered signals thus corresponds to a respective “spatial channel” of the DOFS system. Scattered signals corresponding to different scattering locations (spatial channels) can be distinguished on the basis of their time of receipt at the detector stage, e.g. by comparing a time at which the test signal (e.g. a pulse of the test signal) was launched along the optical fibre and the time of receipt of the scattered signal at the detector stage, taking into account the speed of light along the optical path. The detector stage may then be configured to output a respective measurement signal for each of the plurality of spatial channels, the measurement signal being a function of the scattered signal for that spatial channel.

[0018] The DOFS data includes a respective measurement signal as a function of time for each of the plurality of spatial channels. Thus, each measurement signal may correspond to a time-series of data points.

[0019] The measurement signal for each spatial channel may be derived from (i.e. be a function of) the scattered signal received at the detector stage forthat spatial channel. The specific nature of the measurement signal may depend on a type of DOFS system used and / or a type of measurement performed. For example, each measurement signal may be a function of a phase and / or amplitude of the scattered signal received for the corresponding spatial channel. In some cases, the detector stage may be configured to interfere the scattered signals with a local oscillator signal, in which case the measurement signal may be a function of a phase difference between the local oscillator signal and the scattered signal for the corresponding channel.

[0020] In some cases, the DOFS system may be a Distributed Acoustic Sensing (DAS) system, in which case the DOFS data comprises DAS data. Each of the plurality of measurement signals in the DAS data may be a function of amplitude, phase, and / or phase difference of the scattered signal for the corresponding spatial channel. In some cases, each of the plurality of measurement signals in the DAS data may be indicative of a strain in the corresponding spatial channel.

[0021] The method of the first aspect may include steps performed with the DOFS system to acquire the DOFS data. For example, the method may include steps of launching a test signal (e.g. a pulsed test signal) along the optical fibre, receiving at the detector stage a plurality of scattered signals, each of which was scattered at a respective location along the optical fibre, and outputting, by the detector stage, DOFS data comprising a plurality measurement signals, each measurement signal being a function of a respective one of the scattered signals.

[0022] The received DOFS data is provided to a machine learning model, which is configured to detect a signal response corresponding to an event of a predetermined type in the DOFS data. Here, a signal response can refer to a response (or signature) in the DOFS data caused by an acoustic signal (acoustic wave) in an environment of the optical fibre of the DOFS system, where the acoustic signal results from an event of the predetermined type. For example, an acoustic signal may be generated by an event of the predetermined type. As the acoustic signal propagates through the environment, it will impinge on the optical fibre and thus affect scattering of the test signal (e.g. pulsed test signal) to create a signal response in the DOFS data.

[0023] In more detail, an arrival time of the acoustic signal at each spatial channel (scattering locations) in the optical fibre may vary, e.g. depending on a position of a source of the acoustic signal and a shape of a wavefront of the acoustic signal. This may result in the measurement signal for each spatial channel responding at a different time to the acoustic signal. In this manner, one or more of the plurality of measurement signals may include a response to the acoustic signal at a respective timestamp. Thus, looking at the DOFS data across the plurality of spatial channels, the signal response in the DOFS data resulting from the event of the predetermined type may comprise a response across one or more (e.g. multiple) spatial channels, i.e. the signal response may be a function of spatial channel and time (e.g. time of arrival). In other words, the signal response may comprise a response to the acoustic signal in one or more of the plurality of measurement signals at a respective timestamp.

[0024] Accordingly, a signal response may comprise a distribution (or pattern) of arrival times of the acoustic signal across the plurality of spatial channels. The predetermined type of event may be associated with a predetermined distribution (or pattern) of arrival times across the plurality of spatial channels, e.g. based on known characteristics of acoustic signals caused by events of the predetermined type. Thus, the machine learning model may be configured to detect a predetermined distribution (or pattern) of arrival times across the plurality of spatial channels.

[0025] The signal response may also be referred to as a signature of the event of the predetermined type. The machine learning model may be implemented using various types of suitable model (algorithm), examples of which are provided below. For example, the machine learning model may comprise an artificial neural network (ANN), such as a fully convolutional network (FCN)

[0026] The machine learning model may comprise a classifier model, which is configured to classify the DOFS data to indicate any portions of the DOFS data which correspond to a signal response resulting from an event of the predetermined type. Thus, the DOFS data may be divided into a plurality of portions (or data elements), and the classifier may be configured to assign probability to each portion, the probability being indicative of a likelihood of the portion corresponding to a signal response. In some cases, the machine learning model may comprise a semantic segmentation model, which is configured to categorise each data element (e.g. each data point, or each pixel) in the DOFS data according to a probability of whether that data element corresponds to a signal response. The semantic segmentation model may be a FCN.

[0027] If the machine learning model detects in the DOFS data a signal response corresponding to an event of the predetermined type, then the machine learning model outputs information indicative of the detected signal response. Various types of information may be output, and the information may be in any suitable format. In this manner, a user may be notified of the detection of the signal response, and provided with details of the signal response.

[0028] In some cases, the information indicative of the detected signal response may comprise a notification or an alert, in order to notify or alert a user to detection of an event of the predetermined type.

[0029] Use of the machine learning model to detect the signal response in DOFS data may significantly reduce an amount of processing required to detect the signal response, thus enabling rapid detection of event signatures in DOFS data. In contrast, conventional techniques for detecting signal responses typically involve brute force analysis for determining correlations between measurement signals from different spatial channels, in order to detect the presence of a signal response which extends across multiple spatial channels and which is indicative of the occurrence of an event of a predetermined type. However, in view of the high density of DOFS data (for example, some DOFS systems collect over 50 Megabytes of data per second), such brute force techniques are extremely computationally expensive. As a result, they usually require high-speed and high-capacity data storage devices for subsequent off-line analysis of the DOFS data. In contrast, using a machine learning model that is specifically trained to analyse DOFS data makes the detection of signal responses in DOFS data much less computationally intensive. As a result, the method of the invention can be implemented using relatively inexpensive computer hardware, thus facilitating use of the method in the field. Additionally, due to the reduced processing requirements of the invention, the method can be applied to “live” DOFS data received from the DOFS system, thus enabling rapid (e.g. substantially real-time) detection of event signatures. Such rapid detection of event signatures may be advantageous in many applications, where it is desirable to be alerted of an event as it occurs, or shortly after it has occurred.

[0030] The machine learning model is trained using training data comprising simulated (e.g. computer- simulated) signal responses for events of the predetermined type. This allows for effective training of the machine learning model, to provide accurate signal response detection. In particular, the inventors have found that the high density and often low signal-to-noise ratio of DOFS data can limit the accuracy with which a human can label DOFS data for training a machine learning model. By using simulated signal responses to train the machine learning model, issues associated labelling accuracy can be avoided. For example, simulated DOFS data can be automatically labelled based on parameters of the simulated signal responses. Additionally, parameters of the simulated signal responses can easily be varied, to generate a training dataset with a large number of example signal responses corresponding to a type of event. Thus, a large and accurately labelled training dataset can be used to train the machine learning model, to provide a machine learning model having a high level of detection accuracy.

[0031] The training data may thus comprise computer-generated, or synthetic, data.

[0032] The training data may comprise simulated DOFS data which includes the simulated signal responses. Examples of training data and training of the machine learning model are described in more detail below, in relation to the fourth aspect of the invention.

[0033] The DOFS data may be received as a data stream from the DOFS system. In other words, the DOFS data may be output as a data stream by the DOFS system (e.g. by the detector stage of the DOFS system), such that the DOFS data is received over time. Thus, the DOFS data may be received during operation of the DOFS system, i.e. while the DOFS data is being acquired. This enables the DOFS data to be analysed substantially in real-time (“on the go”), i.e. whilst the DOFS system is acquiring new data, thus allowing for substantially real-time detection of event signatures in the DOFS data. The data stream may be in the form of a sequence of data elements (e.g. data points) for each of the plurality of measurement signals.

[0034] As an example, the received DOFS data can be provided as a continuous stream to the machine learning model. Alternatively, received DOFS data may be provided to the machine learning model at regular intervals, i.e. so that the machine learning model sequentially analyses DOFS data corresponding to individual time intervals. The machine learning model can be configured to sequentially analyse batches (or portions) of the received DOFS data. For example, a batch of the DOFS data may correspond to DOFS data acquired over a (predetermined) time period (interval). Then, for each time period, the batch of data can be provided to the machine learning model for analysis. Assuming the machine learning model can process the batch of data over a shorter period than the time period, the data can be continuously processed in real-time. In some cases, the received DOFS data can be portioned into different spatial windows, e.g. where each spatial window corresponds to one or more of the plurality of spatial channels. Then, for each time period, the spatial windows can be batched together to form a batch for analysis by the machine learning model.

[0035] The plurality of measurement signals may be raw time-series signals. Thus, raw time-series signals from the DOFS system may be provided directly to the machine learning model for analysis. This can improve a speed with which the DOFS data can be analysed by the machine learning model, as the machine learning model can be applied directly to the raw data, e.g. without having to perform pre-processing steps on the received data. This can also serve to ensure that measurement signals with a high sample rate are provided to the machine learning model. The inventors have found that use of raw (e.g. high sample rate) data facilitates detecting response signals of interest, particularly in view of a high transience and low signal-to- noise ratio of the response signals of interest. In particular, raw time-series data may provide high spatial and temporal resolution, so that the arrival times of acoustic signals of interest in each of the spatial channels can be accurately detected. The use of simulated DOFS data for training the machine learning model is particularly beneficial where such high-resolution timeseries signals are used, as it becomes increasingly difficult for a human to accurately label DOFS data at higher resolutions.

[0036] Here, the time-series signals being raw may mean that no pre-processing or signal-modifying techniques are applied to the time-series signals before they are provided to the machine learning model. For example, no noise reduction and / or signal transformation steps are applied to the time-series signals. Thus, data points which are provided to the machine learning model can correspond to data points that were acquired by the DOFS system.

[0037] In some cases, sampling or decimation of the raw time-series may be performed, to achieve a desired sample rate of the time-series signals. The sample rate of the time-series signals may be selected based on an intended application of the method, e.g. to provide sufficient resolution for detecting signal responses of interest in the DOFS data.

[0038] The raw time-series data is in contrast to other forms of data which may involve averaging or otherwise combining multiple data points together. For example, a known way of representing DOFS data involves

[0039] Each of the plurality of measurement signals may have a sample rate of at least 100 Hz. For example, a sample rate of the raw time-series signals provided to the machine learning model may be of at least 100 Hz. This can provide the DOFS data with a high level of temporal resolution to allow for accurate signal response detection in the DOFS data. In some cases, each of the plurality of measurement signals may have a sample rate of 500 Hz or more.

[0040] Although raw time-series signals are mentioned above, in some cases DOFS data may be provided to the machine learning model in other forms. As an example, the DOFS data provided to the machine learning model may comprise an indication of signal power level (e.g. as a function of time and spatial channel). Such DOFS data may be obtained by applying a fast Fourier transform (FFT) to the raw time-series data and averaging the power level over multiple time samples. This form of DOFS data may be represented in a so-called waterfall plot. The raw time-series discussed above is in contrast to such forms of data which may involve averaging or otherwise combining multiple data points together, as the raw time-series does not involve averaging of data.

[0041] The predetermined type of event may correspond to any type of event that generates an acoustic signal that is detectable with a DOFS system, and where a signal response in the DOFS data to the acoustic signal can be accurately simulated (modelled). For example, a type of event that generates an acoustic signal having a wavefront of a known shape may be a suitable type of event, as a signal response in DOFS data corresponding to the acoustic signal can readily be modelled based on the known shape of the wavefront. For instance, where the wavefront shape can be approximated by a parametric function, the parametric function can be used to simulate signal responses for DOFS data. Thus, the invention is particularly suited to detecting types of events which generate an acoustic signal with a known or predictable wavefront shape. Accordingly, the invention can be used to detect signal responses associated with a large number of different types of events, various examples of which are provided below. The machine learning model is specifically trained to detect signal responses associated with a particular type of event (i.e. the predetermined type of event), using simulated signal responses corresponding to that particular type of event.

[0042] The predetermined type of event may correspond to a vehicle motion relative to the distributed optical fibre sensing system. Here, motion of the vehicle relative to the DOFS system may refer to motion of the vehicle relative to the optical fibre of the DOFS system. Thus, the machine learning model can detect signatures in the received DOFS data of a vehicle moving relative to the DOFS system. As a vehicle moves relative to the DOFS system, an acoustic signal caused by motion of the vehicle along the ground will reach the spatial channels in the optical fibre at different times, e.g. depending on a velocity of the vehicle, a direction of travel of the vehicle, and an angle of travel of the vehicle relative to the optical fibre. Accordingly, acoustic signals resulting from motion of a vehicle relative to the DOFS system can readily be modelled, e.g. for different vehicle speeds, directions and angles of travel, to simulate DOFS data with signal responses indicative of vehicle motion. When a signal response corresponding to vehicle motion in the DOFS data is detected, the detected signal response can be further analysed, e.g. to determine a velocity, direction of travel, and or angle of travel of the vehicle relative to the DOFS system.

[0043] The DOFS system may comprise an optical fibre which extends along a vehicle path (e.g. a road, or track). In this manner, the method of the invention can be used to detect vehicle motion (vehicle traffic) along the vehicle path. For example, the method could be used to detect vehicle (e.g. road vehicle) motion along a road, or vehicle (e.g. train) motion along a train track.

[0044] The predetermined type of event may correspond to a seismic event. Thus, the machine learning model can detect signatures in the received data of a seismic event, to detect when a seismic event takes place. When a seismic event occurs, an acoustic signal is generated which will reach the spatial channels in the optical fibre a different times, depending on a shape of a wavefront of the acoustic signal and an epicentre of the seismic event. Thus, by modelling acoustic signals generated by seismic events, it is possible to simulate signal responses for DOFS data corresponding to seismic events, to enable training of the machine learning model. In particular, various types of seismic events generate acoustic signals with wavefronts that can be modelled with a parametric function.

[0045] As an example, the predetermined type of event may correspond to a microseismic event. A microseismic event is a seismic event which is caused as a result of human activity, such as mining or oil and gas production. A wavefront of an acoustic signal caused by a microseismic event can be modelled as a spherical surface which expands outwards from a source (source) of the microseismic event, thus enabling arrival time of the acoustic signal at the spatial channels of the optical fibre to be simulated, e.g. based on a location of the microseismic source relative to the optical fibre and a speed of propagation of the acoustic signal.

[0046] In general, the method of the invention can be used to detect event types which result in an acoustic signal whose arrival time in the spatial channels of the optical fibre can be modelled (e.g. via a parametric function or other type of model), to allow signal responses in DOFS data to be simulated.

[0047] The information indicative of the signal response output by the machine learning model may comprise an indication of a detection time of the signal response in one or more of the spatial channels. In this manner, an indication of portions of the DOFS data which correspond to the signal response is provided. Details of the signal response can then be further analysed, e.g. to determine characteristics of the event. For example, where a microseismic event is detected, the output information may be used to detect a time and location of the microseismic event. The output from the machine learning model may be used as an input for another algorithm or model, in order to further characterise the event. As a result of the high level of accuracy with which the simulated data can be labelled for training the machine learning model, the machine learning model can accurately identify specific portions of the DOFS data which correspond to the signal response. For example, the machine learning model may identify coordinates in the DOFS data corresponding to the signal response. Thus, the information indicative of the signal response may comprise coordinates of the signal response in the DOFS data. The coordinates may, for example be provided in terms of detection time and spatial channel. The detection time of the signal response in the one or more spatial channels may correspond to an arrival time in each of the one or more channels of the acoustic signal generated by the event of the predetermined type. Thus, the output information can be used, for example, to determine a shape of a wavefront of the acoustic signal, e.g. as a function of arrival time and spatial channel, which can then be used to further characterise the event.

[0048] The method may further comprise generating a representation of the detected signal response. In this manner, a visual representation of the signal response (event signature) can be produced, which may facilitate monitoring events by a user. For example, the representation may comprise a plot of detection time (and optionally amplitude) of the signal response in each of the spatial channels. Thus, the representation may enable a user to visualise a shape of a wavefront of the acoustic signal caused by the event. As the method of the invention enables DOFS data to be analysed substantially in real-time, this may allow the user to view acoustic signals associated with events of interest substantially in real-time.

[0049] In some cases, the representation of the detected signal response may comprise a mask which is overlaid onto a visualisation of the received (e.g. raw) DOFS data, highlighting portions of the DOFS data corresponding to the detected signal response.

[0050] The training data may further comprise one or more noise models representing background noise in a distributed optical fibre sensing system. In this manner, the training data may comprise a realistic simulation of DOFS data, including the simulated signal responses in combination with the one or more noise models. This may improve a detection accuracy of the machine learning model. In particular, this may enhance the machine learning model’s ability to distinguish signal responses of interest from background noise.

[0051] According to a second aspect of the invention, there is provided an analysis system for detecting a signal response in distributed optical fibre sensing data, the analysis system comprising a processing device, and a memory storing instruction which, when executed by the processing device, cause the processing device to perform the method of the first aspect of the invention. Thus, the analysis system can be used to implement the method of the first aspect of the invention. Accordingly, any features described above in relation to the first aspect of the invention can be shared with the second aspect of the invention (and vice versa).

[0052] The analysis system is a computer-implemented system, which can be implemented using any suitable computer system or network of computer systems. The processing device may correspond to one or more computer processors which are coupled to the memory and configured to execute the instructions stored in the memory. The memory may comprise a nonvolatile storage medium, such as a local memory drive and / or a cloud-based storage system. The trained machine learning model is stored in the memory. According to a third aspect of the invention, there is provided a distributed optical fibre sensing system comprising: a test signal generator configured to transmit a test signal along an optical fibre; a detector stage configured to receive a plurality of scattered signals from the optical fibre, wherein each scattered signal corresponds to a respective spatial channel associated with a respective scattering location along the optical fibre, and wherein the detector stage is further configured to output a respective measurement signal as a function of each of the plurality of scattered signals; and an analysis system configured to receive the respective measurement signals from the detector stage, and to perform the method of the first aspect of the invention. The DOFS system of the third aspect may be used to implement the method of the first aspect of the invention. Accordingly, features described above in relation to the first aspect may be shared with the third aspect of the invention (and vice versa).

[0053] The analysis system in the DOFS system of the third aspect may correspond to the analysis system described in relation to the second aspect of the invention above. Accordingly, any features described in relation to the second aspect may be shared with the third aspect (and vice versa).

[0054] The test signal generator may comprise a pulse generator configured to transmit a pulsed test signal along the optical fibre. The pulse generator may comprise an optical modulator for generating the pulsed test signal from a received light signal. For example, the DOFS system may comprise a coherent light source (e.g. laser), which provides a continuous wave light signal to the pulse generator, which generates the pulsed test signal. The pulse generator may be coupled to an end of the optical fibre to launch the pulsed test signal into the optical fibre.

[0055] The detector stage may be coupled to the optical fibre to receive the scattered signals. The detector stage may comprise an optical detector for detecting the scattered signals. For example, a square law detector may be used.

[0056] In some cases, the detector stage may further be configured to receive a local oscillator signal, and to interfere the local oscillator signal with the received scattered signals on the detector. In this manner, a signal output by the detector may be indicative of an interference between the local oscillator signal and the scattered signals.

[0057] According to a fourth aspect of the invention, there is provided a method of training a machine learning model for detecting a signal response in distributed optical fibre sensing data, the method comprising: generating a plurality of training datasets, wherein each training dataset comprises simulated distributed optical fibre sensing data for a distributed optical fibre sensing system having a plurality of spatial channels, each spatial channel associated with a respective scattering location along an optical fibre of the distributed optical fibre sensing system, wherein, for each of the plurality of training datasets, simulating the distributed acoustic sensing data comprises: providing a model of an acoustic signal corresponding to an event of a predetermined type, and generating a simulated signal response to the acoustic signal in each of the spatial channels; and combining the simulated signal response with a noise model to provide simulated distributed optical fibre sensing data, the noise model being representative of background noise in the distributed optical fibre sensing system; and using the plurality of training datasets to train a machine learning model, wherein the machine learning model is configured to detect a signal response corresponding to an event of the predetermined type in distributed optical fibre sensing data.

[0058] The method of the fourth aspect may be used to train the machine learning model which is used in the first aspect of the invention. Accordingly, any features described in relation to the fourth aspect of the invention are applicable to the first method of the invention (and vice versa).

[0059] The method is a computer-implemented method, which can be implemented using any suitable computing system. For example, the method may be implemented by a training system (e.g. as described in the fifth aspect of the invention, below), which includes a computer memory storing instructions for performing the method, and a processing device (e.g. including one or more processors) configured to execute the instructions to perform the method.

[0060] In line with the discussion above in relation to the first aspect of the invention, using simulated DOFS data to train the machine learning model enables the machine learning model to accurately detect signal responses resulting from events of interest in DOFS data. For instance, simulating DOFS data enables a large number of training datasets to be rapidly generated, e.g. by varying (sampling) parameters of the simulation. This can provide coverage of a wide range of realistic scenarios, thus improving an ability of the machine learning model to recognise and detect signal responses of interest in real-world DOFS data. In particular, the simulation of DOFS data facilitates obtaining a larger number of training datasets compared to techniques involving manual labelling of measurement data. Moreover, using simulated DOFS data can greatly increase the accuracy with which training data can be labelled, in turn improving a detection accuracy of the machine learning model. For example, as discussed above, it may be difficult for a human to accurately label signal responses of interest in DOFS data, in view of the generally low signal-to-noise ratio, high density of the DOFS data, and highly transient nature of the signal responses of interest. In contrast, as parameters of the simulated signal responses are known, the signal responses can be labelled to a high degree of accuracy in the simulated data.

[0061] Each generated training dataset includes simulated DOFS data for a DOFS system with a plurality of spatial channels. In other words, DOFS data is simulated for a model DOFS system having an optical fibre with a plurality of spatial channels. Thus, the simulated DOFS data in each training dataset includes a plurality of simulated measurement signals as a function of time, each simulated measurement signal corresponding to a respective spatial channel. In other words, the simulated DOFS data may have a similar (or same) format to real-world DOFS data, e.g. such that the simulated measurement signals have an analogous format to the received measurement signals discussed in the first aspect of the invention. For example, the simulated measurement signals may correspond to “raw” time-series signals, analogous to those discussed above. A number of training datasets used for training the machine learning model may vary, e.g. depending on a complexity of the simulation and the signal response to be detected. For example, where a simulation for a signal response has a large number of variable parameters, a larger number of training datasets may be required in order to provide suitable coverage of realistic scenarios.

[0062] The simulated DOFS data may have a time-resolution and a spatial-resolution similar to (corresponding to) real-world DOFS with which the machine learning model is to be used. Accordingly, the simulated DOFS data may have a realistically high density, with its resolution being similar to that of actual DOFS data. This serves to improve an accuracy of the machine learning model. For example, a sample rate of the plurality simulated measurement signals may correspond to 100 Hz or more (e.g. 500 Hz or more), which may correspond to realistic DOFS data sample rates used with the machine learning model. A spatial resolution of the simulated DOFS data may correspond to a realistic spatial resolution of a DOFS system, e.g. of the order of 1 m to 15 m, e.g. about 6 m. The simulated DOFS data may have a spatial extent (i.e. number of spatial channels) corresponding to N times the spatial resolution, where N is chosen such that a signal response of interest can be resolved within the chosen spatial extent of the DOFS data.

[0063] The time-resolution (sample rate) and spatial-resolution of the simulated DOFS data may be adapted to (e.g. correspond to) a time-resolution (sample rate) and spatial-resolution of a DOFS system with which the machine learning model will be used.

[0064] Here, a spatial resolution of a DOFS system or of DOFS data refers to a spacing between adjacent spatial channels in the DOFS system.

[0065] In order to simulate the DOFS data, a model of an acoustic signal corresponding to an event of a predetermined type is provided. The model may describe (define) one or more characteristics of the acoustic signal, such as a shape (e.g. waveform) of the signal, frequency of the signal, amplitude of the signal, origin of the signal, velocity of the signal, and / or wavefront shape of the signal.

[0066] As an example, where the machine learning model is trained to detect signal responses arising from seismic (e.g. microseismic) events, the acoustic signal may be modelled as a Ricker wavelet, with a wavefront having a spherical shape. An arrival time of the acoustic signal at the spatial channels of the optical fibre may then be modelled using a hyperbolic curve.

[0067] The model of the acoustic signal can be used to determine (e.g. calculate, model) an arrival time of the acoustic signal in each of the spatial channels of the DOFS system. For example, the arrival time in each of the spatial channels can be calculated based on an origin, velocity and wavefront shape of the acoustic signal. This can then be used to simulate a response signal to the acoustic wave for each of the spatial channels in the model DOFS system. For instance, when the acoustic signal reaches a spatial channel, it may cause an amplitude of the measurement signal for that channel to vary in response to the acoustic signal. In other words, the simulated signal response in each spatial channel may comprise a variation in signal amplitude at an arrival time of the acoustic signal in that spatial channel. The variation in signal amplitude could, for example, comprise a peak and / or trough in the measurement signal for the spatial channel.

[0068] In order to provide a realistic simulation of DOFS data, the simulated signal response is combined with a noise model which is representative of background noise in the distributed optical fibre sensing system. In this manner, the machine learning model is trained using realistic data, to enable it to effectively distinguish signals of interest from background noise. Thus, the DOFS data includes, for each spatial channel, a simulated measurement signal comprising a combination of the noise model and the signal response simulated for that spatial channel.

[0069] The noise model and the simulated signal response can be combined in any suitable way. For example, the noise model may be added (e.g. superimposed) onto the simulated signal response. This may also include setting a signal-to-noise ratio between the noise model and the simulated signal response, e.g. to achieve a realistic measurement signal.

[0070] The noise model represents background (measurement) noise in a DOFS system. For example, background noise may comprise fluctuations in measurement signal due to noise-generating effects in the DOFS system and along the optical fibre. Such noise-generating effects could, for instance, include thermal fluctuations and / or strain fluctuations along the optical fibre. The noise model may also include noise arising from other noise sources in the environment, and which could interfere with detection of signal responses of interest. For example, the noise model could include noise from generators, pumps, or other machinery which might be expected to be in the environment where the measurements are to be performed. Noise from such other sources can be simulated and / or based on real-world noise measurements.

[0071] The machine learning model is trained, using the plurality of trained datasets, to detect a signal response corresponding to an event of the predetermined type in DOFS data. In this manner, the machine learning model can be fed new (i.e. previously unseen) DOFS data, to determine if the DOFS data includes any signal responses corresponding to events of the predetermined type. The machine learning model may be as described above in relation to the first aspect of the invention. Various different training techniques may be used for training the machine learning model, examples of which are provided below. When generating the plurality of training datasets, one or more parameters of the acoustic signal model and / or the simulated signal response may be varied across the plurality of training datasets. In other words, a set of one or more variable parameters may be used for the acoustic signal model and / or for simulating the signal response. This enables the plurality of training datasets to cover a wide range of different scenarios, e.g. where details of the acoustic signal model and / or simulated signal response are varied. This serves to improve an ability of the machine learning model to reliably detect signal responses arising from events of the predetermined type. For instance, whilst events of the predetermined type may generate acoustic signals having a common set of characteristics, there may be variations in acoustic signals and hence in signal responses arising from different events. This may be due, for example, to differences in location (origin) of the event relative to the optical fibre, differences in strength (magnitude) of the event, time evolution of the event, etc.

[0072] The one or more parameters of the acoustic signal and / or simulated response signal may be randomly sampled (selected) for each of the plurality of training datasets. In other words, each time a training dataset is generated, values for each of the one or more parameters may be randomly sampled (selected). As an example, a respective value range may be associated with each of the one or more parameters, and a value for each of the one or more parameters may be randomly selected from the associated value range. This enables DOFS data to be simulated for a wide range of scenarios.

[0073] The acoustic signal model may comprise a parametric function, and the one or more parameters may comprise a parameter (or parameters) of the parametric function. For example, a parametric function may be used to define a wavefront shape of the acoustic signal, and / or a parametric function may be used to define a waveform of the acoustic signal.

[0074] The noise model may be obtained from a measurement performed with a distributed optical fibre sensing system. In other words, the noise model used to produce the simulated DOFS data may be obtained from a measurement performed with a real-world (i.e. physical, nonsimulated) DOFS system. In this manner, the simulated DOFS data can include a real noise model, thus enhancing the machine learning model’s ability to distinguish signal responses of interest from background noise. The noise model can be obtained from a measurement performed in an environment where the signal responses of interest are expected to be measured.

[0075] The method may further comprise providing a plurality of noise models, wherein simulating the distributed acoustic sensing data further comprises selecting one of the plurality of noise models. In other words, for each of the plurality of training datasets, simulating the DOFS data further comprises selecting one of the plurality of noise models, and combining the selected noise model with the simulated signal response to produce the simulated DOFS data. In this manner, different noise models may be used for generating the plurality of training datasets. This may improve the machine learning model’s ability to recognise background noise, thus enhancing a level of generality of the model and an accuracy with which it can detect signal responses of interest.

[0076] In line with the above, each of the plurality of noise models may be obtained from a respective measurement performed with a (physical, non-simulated) DOFS system, so that the machine learning model is trained with real background noise models.

[0077] The method may further comprise, for each of the plurality of training datasets, (automatically) labelling portions of the simulated distributed optical fibre sensing data where the simulated signal response has a higher amplitude than the noise model. In this manner, the machine learning model can learn, based on the labelled portions of simulated DOFS data, to recognise portions of DOFS data which correspond to signal responses of interest. As a simulated response is used, portions of the simulated DOFS data corresponding to the simulated response signal can be automatically labelled with a high degree of accuracy, e.g. based on known locations where the simulated response signal was added into the data. Due to the high labelling accuracy of the training data, detection accuracy of the machine learning model can be improved. Advantageously, labelling can be performed automatically as part of generating the training datasets, such that no human labelling is required.

[0078] In some cases, the method may comprise, for each of the plurality of training datasets, (automatically) labelling portions of the simulated DOFS data corresponding to an arrival of the acoustic signal (e.g. a wavefront of the acoustic signal) in each of the spatial channels. In line with the discussion above, the arrival time of the acoustic signal in each of the spatial channels can be readily determined from the model of the acoustic signal. In this manner, the machine learning model can learn to recognise features of DOFS data which are indicative of an acoustic signal being incident on a spatial channel.

[0079] The machine learning model may be trained using a supervised learning process. In particular, supervised learning may be performed using the labelled DOFS data, to train the machine learning model to detect (classify) portions of DOFS data corresponding to a signal response resulting from an event of the predetermined type.

[0080] According to a fifth aspect of the invention, there is provided a training system for training a machine learning model for detecting a signal response in distributed optical fibre sensing data, the training system comprising a processing device, and a memory storing instructions which, when executed by the processing device, cause the processing device to perform the method of the fourth aspect of the invention. The training system of the fifth aspect is used to perform the training method of the fourth aspect of the invention. Therefore, any features described in relation to the fourth aspect are applicable to the fifth aspect of the invention (and vice versa). The training system is a computer-implemented system, which can be implemented using any suitable computer system or network of computer systems. The processing device may correspond to one or more computer processors which are coupled to the memory and configured to execute the instructions stored in the memory. The memory may comprise a nonvolatile storage medium, such as a local memory drive, a distributed (e.g. networked) storage system, and / or a cloud-based storage system. Data and algorithms used for simulating the DOFS data, and / or the simulated DOFS data may be stored in the memory. Likewise, the machine learning model may be stored in the memory.

[0081] The invention includes the combination of the aspects and preferred features described except where such a combination is clearly impermissible or expressly avoided.

[0082] Summary of the Figures

[0083] Embodiments and experiments illustrating the principles of the invention will now be discussed with reference to the accompanying figures in which:

[0084] Fig. 1 shows a schematic diagram of a distributed optical fibre sensing (DOFS) system and analysis system according to an embodiment of the invention;

[0085] Fig. 2 shows a diagram of a method of detecting a signal response in DOFS data according to an embodiment of the invention;

[0086] Fig. 3 shows a schematic diagram of an example use of a DOFS system for detecting a microseismic event;

[0087] Fig. 4 shows an example graph of arrival time as a function of spatial channel in a DOFS system, for an acoustic signal resulting from a microseismic event;

[0088] Fig. 5 shows an example of detection of signal responses corresponding to a microseismic event in DOFS data;

[0089] Fig. 6 shows a schematic diagram of an example use of a DOFS system for detecting vehicle motion relative to the DOFS system;

[0090] Fig. 7 shows an example graph of arrival time as a function of spatial channel in a DOFS system, for an acoustic signal resulting from vehicle motion relative to the DOFS system;

[0091] Fig. 8 shows an example of detection of signal responses corresponding to vehicle motion in DOFS data;

[0092] Fig. 9 shows a diagram of a method of training a machine learning model according to an embodiment of the invention; Fig. 10 shows an example of simulating a measurement signal for a spatial channel of a DOFS system; and

[0093] Fig. 11 shows a schematic diagram of a training system according to an embodiment of the invention.

[0094] Detailed Description of the Invention; Further Optional Features

[0095] Aspects and embodiments of the present invention will now be discussed with reference to the accompanying figures. Further aspects and embodiments will be apparent to those skilled in the art. All documents mentioned in this text are incorporated herein by reference.

[0096] Fig. 1 shows a schematic diagram of a distributed optical fibre sensing (DOFS) system 10 and analysis system 60 according to an embodiment of the invention. In the example shown, the DOFS system 10 comprises distributed acoustic sensing (DAS) system. However, in other examples, alternative DOFS system configurations may be used to perform other types of optical fibres sensing measurements. The system 10 is arranged to interrogate an optical path, in particular an optical fiber 1000, which may be of any desirable length for a given purpose.

[0097] The system 10 comprises a light source which produces coherent light, which is given here as a laser 12, and is used in continuous wave (CW) operation. The light produced by the laser 12 can be directed into an optical isolator to ensure that light is not passed back to the laser 12. Light from the laser 12 is split into two paths by an optical coupler 16 or beam splitter. The first path, from which light is directed into the fiber 1000 is known as the launch path. The second path, from which light is passed directly to a detector stage 50 (discussed below), is known as the local oscillator path. The light is split between the two paths by the optical coupler 16, for example such that 90% of the incoming light is directed into the launch path, and 10% of the incoming light is directed into the local oscillator path. Of course, the ratio of incoming light directed into each path may be chosen by the operator depending on the nature of the operation for which the OTDR system 10 is used.

[0098] The laser light which is directed into the launch path then passes through a pulse generator 18, such as an acousto-optic modulator (AOM) 18. The AOM 18 is a device which can simultaneously generate an optical pulse as well as upshift or downshift the frequency of light by an amount equal to the radiofrequency which drives the AOM 18. This frequency shift, F, may be known as the intermediate frequency or the difference frequency. In this way, the AOM 18 is able to generate a pulsed test signal which may be between 5 ns and 100 ns in duration, but not limited to this range. Of course, any preferred method of generating a pulse of light may be used, such as an electro-optic modulator (EOM). The pulsed test signal may also be referred to herein as a launch pulse. The pulse of light can then be amplified using an optical amplifier. The light pulse is then introduced to the optical fiber 1000 via an optical circulator 22, which has three ports. The amplified light pulse enters the circulator 22 through a first port, where it is passed to a second port in order to enter the optical fiber 1000. As the pulse of light passes through the fiber 1000, a fraction of the light is backscattered from the fiber 1000, e.g. by Rayleigh scattering, and a further fraction captured and guided back towards the circulator 22. The scattered light, which may be referred to herein as a scattered signal, enters the circulator 22 at the second port, and leaves the circulator 22 to enter a detector stage 50 via a third port.

[0099] The detector stage 50 has two inputs. The first input is the scattered laser light from the third port of the circulator 22. The second input is the laser light taken directly from the local oscillator (LO) path mentioned above. The scattered light is then mixed with the LO light at an optical coupler 28. The light the optical coupler 28 is then allowed to interfere on an optical detector 30 (e.g. square law detector). An output signal from the detector is then taken and measured at an analog-digital-converter (ADC) 32, which outputs a corresponding measurement signal. The measurement signal can then be further analysed, e.g. to determine a strain and / or acoustic field at the corresponding scattering location along the optical fibre 1000. It will be appreciated that the invention is not limited to use of the specific DAS system shown in Fig. 1, and that various alternative systems may be used. For example, additional or alternative components may be included in the DAS system. In some cases, a polarisation diverse detector stage 50 is used, e.g. where the scatter signal and the LO signal are each split into horizontal and vertical polarised states, to enable polarisation diverse detection.

[0100] The system 10 described above can make use of a heterodyne sensing approach, wherein the frequency of the local oscillator and of the launch pulse are shifted relative to one another by the AOM 18. The difference in these two frequencies should be larger than the bandwidth required to represent the scattering without allowing crosstalk between the carrier and the DC terms which are also generated, allowing the phase and amplitude information of the scattering to be recovered using a real carrier. Another method employs a complex carrier detector stage, replicating the polarisation diverse detector stage for two copies of the local oscillator shifted by 90 degrees relative to each other. This allows detection via a complex carrier, allowing either the positive sidelobe or the negative sidelobe of the resulting interference signal to be recovered independently. This allows homodyne operation whereby the local oscillator signal and launch pulse operate at the same optical frequency.

[0101] As the pulsed test signal travels along the optical fibre 1000, the test signal will be scattered at a plurality of scattering locations distributed along a length of the optical fibre 1000. Thus, for each pulse launched along the optical fibre, a plurality of scattered signals with be received at the detector stage 50, each of the scattered signals corresponding to a respective scattering location in the optical fibre. The scattered signals are received sequentially in time, with the time of receipt for a given scattered signal depending on its scattering location in the optical fibre 1000. Thus, the scattering location corresponding to a received scattered signal can be determined based on its time of receipt at the detector stage 50, e.g. comparing a time at which a pulse of the test signal was launched along the optical fibre 1000 and the time of receipt, taking into account the speed of light along the optical path. Accordingly, the detector stage 50 can output a respective measurement signal for each of the plurality of scattering locations along the optical fibre 1000, the measurement signal being derived from interference of the scattered signal for that scattering location with the local oscillator signal. Each of the scattering locations in the optical fibre 1000 for which a measurement signal is output by the detector stage 50 may be referred to as a spatial channel of the system 10.

[0102] Returning to Fig. 1, an analysis system 60 is connected to an output of the detector stage 50, to receive DOFS data comprising the plurality of measurement signals output by the detector stage 50, i.e. by the ADC 32. The analysis system 60 may be directly coupled to the ADC 32 as shown in Fig. 1. Alternatively, the analysis system 60 may be arranged to received the plurality of measurement systems from the detector stage 50 via a computer network or intermediate storage system which is configured to store data output by the detector stage 50. The analysis system 60 is configured to analyse the DOFS data received from the detector stage 50, in order to detect signal responses in the DOFS data corresponding to events of the predetermined type, as discussed in more detail below. The analysis system 60 is a computer-implemented system which includes a memory 62 that stores a machine learning model, and a processor 64 configured to analyse the received DOFS data using the machine learning model. The memory 62, or a separate memory (not shown), may be configured to store the received DOFS data for analysis by the machine learning model. In practice, the analysis system 60 can be implemented using any suitable arrangement of computer hardware, such as a personal computer (e.g. desktop or laptop computer), computer server (e.g. cloud-based server), or a network of computing devices.

[0103] The analysis system 60 can receive the DOFS data from the detector stage 50 as a “live” stream of data. In other words, the analysis system 60 can receive and record the DOFS data as it is output by the detector stage 50. Thus, for example, the analysis system 60 can receive and record (e.g. in its memory) the measurement signals output for each of the spatial channels of the system 10 as a function of time. This enables the analysis system 60 to analyse DOFS data as it is received, thus allowing for substantially real-time data analysis and detection of signal responses of interest.

[0104] Fig. 2 shows a diagram illustrating a method 200 according to an embodiment of the invention. The method 200 is for detecting a signal response of interest in DOFS data. For example, the method 200 can be implemented by the analysis system 60 described above, in order to analyse the DOFS data received from the detector stage 50. For convenience, the method is described in the context of the example of Fig. 1, although it will be appreciated that the method 200 can be used with other types of DOFS system.

[0105] In line with the discussion above relating to Fig. 1, a pulsed test signal is launched along the optical fibre 1000. This results in a plurality of measurement signals being output by the detector stage 50, each measurement signal corresponding to a respective spatial channel (i.e. scattering location) of the system 10. Accordingly, in step 202, the method 200 comprises receiving (e.g. by the analysis system 60), DOFS data comprising the plurality of measurement signals as a function of time. For example, as noted above, the DOFS data can be received as a data stream, i.e. as it is being output by the detector stage 50. The received DOFS data can be stored (or cached) in a memory of the analysis system 60, such as the memory 62 or another dedicated memory. The received data can correspond to the raw data which is output by the detector stage 50, e.g. such that no down-sampling or other pre-processing steps are performed on the data. In step 204, the DOFS data is provided to a machine learning model, e.g. the machine learning model stored in the memory 62 mentioned above. The machine learning model is configured to detect in the DOFS data a signal response corresponding to an event of a predetermined type. The machine learning model was trained using training data comprising simulated signal responses for events of the predetermined type, as described in more detail below. In line with the above, the DOFS data provided to the machine learning model can correspond to the raw data output by the detector stage 50. Thus, the DOFS data may be of relatively high density, with a spatial resolution and time-resolution of the data corresponding to a measurement resolution of the system 10. Then, in step 206, if the machine learning model detects in the DOFS data a signal response corresponding to an event of the predetermined type, the machine learning model outputs information indicative of the detected signal response.

[0106] For illustration purposes, an example of signal response detection using the method 200 will be described in the context of detecting a microseismic event in DOFS data. Fig. 3 shows a schematic diagram of an OTDR system 300 (which may correspond, for example, to the system 10 described above) having an optical fibre 302 that extends below a ground surface 304. The optical fibre 302 may, for example, extend downwards through the ground in a substantially vertical direction. The system 300 is used for detecting acoustic signals arising from microseismic events, which may occur as a result of human activities such as mining or oil and gas extraction. Fig. 3 shows of an acoustic wave (signal) 306 which is propagating outwards from a source (origin) 308 of a microseismic event occurring below the ground surface. As shown, the acoustic wave 306 may have a substantially spherical wavefront which expands outwards from the source 308. As the acoustic wave 306 propagates outwards, it will impinge on the optical fibre 302, with different arrival times in each of the spatial channels of the system 300 due to its spherical wavefront. This is depicted in Fig. 4, which shows an arrival time of the acoustic wave 306 in each of a plurality of spatial channels (scattering locations) of the system 300. The horizontal axis in Fig. 4 corresponds to position along the length of the optical fibre 302, with the vertical dashed lines indicating positions of spatial channels (scattering locations) measured by the system 300. The vertical axis in Fig. 4 represents a measurement time. The points 400 shown for each spatial channel correspond to an arrival time of the acoustic wave 306 at that spatial channel. As can be see, the arrival times of the outward-expanding acoustic wave 306 in the different spatial channels follow a substantially hyperbolic curve 402. When the acoustic wave 306 reaches a spatial channel, this will result in a modulation of strain in the optical fibre 302 at that spatial channel (scattering location), resulting in a modulation (variation) of the measurement signal corresponding to that spatial channel. For example, the measurement signal may display a peak and / or trough resulting from arrival of the acoustic wave 306 at the spatial channel. Such a variation in the measurement signal caused by the acoustic wave 306 may be considered as a signal response to the acoustic wave 306. Looking across the plurality of spatial channels shown in Fig. 4, a signal response corresponding to the acoustic wave 306 can be considered as the hyperbolic distribution (pattern) of arrival times. Fig. 3 further indicates a position of an “apex” channel 310 of the optical fibre 302, which corresponds to location of the optical fibre 302 that is normal to an incident wavefront of the acoustic signal 306, and thus where the acoustic signal 306 will first impinge on the optical fibre 302.

[0107] Taking the example of Figs. 3 and 4, at step 204 of the method 200, DOFS data measured by the system 300 is provided to the machine learning model. The DOFS data includes measurement signals as a function of time for each of the spatial channels measured by the system 300. In this example, the machine learning model is specifically trained to detect signal responses in the DOFS data corresponding to microseismic events. Thus, for example, the machine learning model can be trained to look for a hyperbolic distribution (or pattern) of arrival times across the plurality of spatial channels. If such a hyperbolic distribution of arrival times is detected by the machine learning model in the DOFS data, e.g. as shown in Fig. 4, the machine learning model can determine the detected distribution of arrival times as a response signal which is indicative of a microseismic event.

[0108] Fig. 5 shows an example of using a machine learning model to detect a signal response corresponding to a microseismic event in DOFS data. Panel (a) of Fig. 5 shows example DOFS data 500 as a function of time and spatial channel acquired using a DOFS system, e.g. system 10 or 300 discussed above. Thus, each vertical pixel row in the DOFS data 500 may correspond to a measurement signal for a respective spatial channel of the DOFS system. In this example, shading in the graph of panel (a) of Fig. 5 represents an amplitude of the raw time-series signals. The DOFS data 500 is provided to the machine learning model, which is configured to detect signal responses corresponding to microseismic events. In particular, as discussed above, the machine learning model is trained to detect hyperbolic distributions of arrival times in DOFS data. Panel (b) of Fig. 5 shows an example of an output 502 from the machine learning model, which may correspond to an output from step 206 of the method 200. In the output 502, the machine learning model has isolated portions of the DOFS data 500 which it has determined as corresponding to signal responses indicative of microseismic events. Thus, the machine learning model outputs a representation of the detected signal response(s). As can be seen, the detected signal responses correspond generally hyperbolic distributions of arrival times across the spatial channels, and are thus indicative of the occurrence of a microseismic event. The signal responses shown in the output 502 can then be further analysed, e.g. to determine a time and source location of the microseismic event, for example by calculating properties of the acoustic wave based on the detected signal responses.

[0109] Figs. 6 and 7 illustrate another example application of the method 200, where a DOFS system 600 (e.g. system 10) is used to monitor vehicle traffic along a path (e.g. road). The DOFS system 600 includes an optical fibre 602 which extends along the path, under a surface 604 of the path. As a vehicle 606 travels along the path, it generates an acoustic wave 608 which propagates in the ground to the optical fibre 602. In the example shown, the vehicle is a road vehicle moving along a road or a path. However, the method is equally applicable to other types of vehicle, e.g. a train moving along a track. The vehicle 606 acts as a moving source for the acoustic wave 608, such that progress of the vehicle 606 along the path can be detected by monitoring arrival of the acoustic wave in the spatial channels of the optical fibre 602. An example of this is shown in Fig. 7, which is a graph of measurement time as a function of spatial channel. As in Fig. 4, the vertical dashed lines in Fig. 7 indicate locations of spatial channels that are measured by the DOFS system 600. The points 700 in Fig. 7 indicate an arrival time of the acoustic wave 608 generated by the vehicle 606 in each spatial channel. As can be seen, an arrival time of the acoustic wave increases for increasing spatial channel number, indicating that the vehicle 606 is moving along the path (and hence along the optical fibre 602), e.g. in a positive direction. The arrival times of the acoustic wave 608 follow a curve 702, which is indicative of a speed and direction of motion of the vehicle 606 relative to the optical fibre 602. When the acoustic wave 608 reaches a spatial channel, this will result in a modulation of strain in the optical fibre 602 at that spatial channel (scattering location), resulting in a modulation (variation) of the measurement signal corresponding to that spatial channel. For example, the measurement signal may display a peak and / or trough resulting from arrival of the acoustic wave 608 at the spatial channel. Such a variation in the measurement signal caused by the acoustic wave 608 may be considered as a signal response to the acoustic wave 608. Thus, by analogy to the discussion in relation to Fig. 4 above, a signal response in DFOS data corresponding to vehicle motion can be considered as a distribution (pattern) of arrival times which is indicative of the source of the acoustic wave (i.e. the vehicle) moving relative to the optical fibre 602 over time. Taking the example of Figs. 6 and 7, at step 204 of the method 200, DOFS data measured by the system 600 is provided to the machine learning model. The DOFS data includes measurement signals as a function of time for each of the spatial channels measured by the system 600. In this example, the machine learning model is specifically trained to detect signal responses in the DOFS data corresponding to vehicle motion relative to the optical fibre 602. Thus, for example, the machine learning model can be trained to look for a distribution (or pattern) of arrival times across the plurality of spatial channels which is indicative a vehicle motion. If such a distribution of arrival times is detected by the machine learning model in the DOFS data, e.g. as shown in Fig. 7, the machine learning model can determine the detected distribution of arrival times as a response signal which is indicative of a vehicle motion.

[0110] Fig. 8 shows an example of using a machine learning model to detect a signal response corresponding to vehicle motion in DOFS data. Panel (a) of Fig. 8 shows example DOFS data 800 as a function of time and spatial channel acquired using a DOFS system, e.g. system 10 or 600 discussed above. Thus, each vertical pixel row in the DOFS data 800 may correspond to a measurement signal for a respective spatial channel of the DOFS system. The DOFS data in panel (a) of Fig. 8 represents power level in a particular energy band, obtained by applying a fast Fourier transform (FFT) to the raw time-series data and averaging the power level over multiple time samples. The DOFS data 800 is provided to the machine learning model, which is configured to detect signal responses corresponding to vehicle motion. In particular, as discussed above, the machine learning model is trained to detect distributions of arrival times in DOFS data consistent with vehicle motion relative to the optical fibre. Panel (b) of Fig. 8 shows an example of an output 802 from the machine learning model, which may correspond to an output from step 206 of the method 200. In the output 802, the machine learning model has isolated portions of the DOFS data 800 which it has determined as corresponding to signal responses indicative of vehicle motion. As can be seen, the machine learning model has detected a number of signal responses corresponding to vehicle motion, indicating that a number of vehicles are travelling relative to the optical fibre in different directions. This may correspond, for example, to vehicle traffic along a two-way road. Signal responses where the spatial channel number increases with time correspond to vehicles travelling in a first (e.g. positive) direction, whilst signal responses where the spatial channel number decreases with time correspond to vehicles travelling in a second (e.g. negative) direction. Properties of vehicle motion can then be calculated from the detected signal responses, such as number of vehicles, speed and direction of travel of the vehicles.

[0111] Fig. 9 shows a flow diagram of a method 900 of training a machine learning model for detecting a signal response in distributed optical fibre sensing data, according to an embodiment of the invention. The method 900 can be used to train the machine learning model used in method 200, and the machine learning model stored in memory 62 of the analysis system 60 discussed above. The machine learning model is trained to detect signal responses in DOFS data corresponding to events of a predetermined type.

[0112] In step 902, the method 900 comprises generating a plurality of training data sets, each training dataset comprising simulated DOFS data for a model DOFS system having a plurality of spatial channels. For example, the simulated DOFS data may correspond to a simulation of data that could be obtained from a system such as DOFS system 10 described above. The simulated DOFS data can thus have a similar format to real (i.e. acquired) DOFS data, i.e. the simulated DOFS data can have a plurality of measurement signals as a function of time, each measurement signal corresponding to a respective spatial channel.

[0113] In more detail, simulating the DOFS data in step 902 comprises providing (or defining) a model of an acoustic signal that is generated by an event of the predetermined type. This could include, for example, a functional or parametric definition of the acoustic signal. In some cases, the model could be based on a measurement of an acoustic signal resulting from an event of the predetermined type. The model of the acoustic signal may define one or more characteristics of the acoustic signal, such as a shape (e.g. waveform) of the signal, frequency of the signal, amplitude of the signal, origin of the signal, velocity of the signal, and / or wavefront shape of the signal. The model of the acoustic signal can further define, or be used to define (or calculate), a model of arrival time of the acoustic signal in each of the spatial channels of the modelled system.

[0114] Then, using the model of the acoustic signal, a signal response to the acoustic signal is simulated for each spatial channel in the modelled DOFS system. In particular, the model of arrival time is used to determine the arrival time of the acoustic signal at each spatial channel, and a signal response corresponding to the acoustic signal is added to each spatial channel at the determined arrival time. The added signal response may have a shape correspond to a waveform of the acoustic signal. Thus, each spatial channel includes a simulated signal response corresponding to the acoustic signal, at an arrival time which is determined based on the arrival time model for the acoustic signal.

[0115] A noise model is then added the combined with the simulated signal response for each spatial channel, in order to provide a simulated measurement signal for each channel. The noise model is representative of background noise in a DOFS system, and serves to provide more realistic simulated DOFS data. The noise model can, for example, be derived from a noise measurement performed with a real (physical) DOFS system. Alternatively, the noise model can be simulated, e.g. using suitable noise simulation software.

[0116] The simulated DOFS data resulting from the combination of the simulated signal responses for the spatial channels and the noise model forms a training dataset. The same process is followed for generating each of the plurality of training datasets. However, parameters of the acoustic model and / or the arrival time model are varied between the plurality of training datasets, so that the training datasets cover a range of different scenarios. In particular, the model of the acoustic signal and / or arrival time model may have one or more variable parameters, which can be randomly sampled in order to generate each training dataset.

[0117] For the sake of example, simulation of DOFS data with signal responses corresponding a microseismic event is described in relation to Fig. 10. An acoustic signal resulting from a microseismic event can be modelled as a Ricker wavelet which propagates with a spherical wavefront. Additionally, as discussed above in relation to Fig. 4, a distribution of arrival times for an acoustic signal resulting from a microseismic event follows a substantially hyperbolic function. Therefore, an acoustic signal resulting from a microseismic event can be modelled as a Ricker wavelet, with arrival times at the spatial channels of the DOFS system being modelled by a hyperbolic function. Variable parameters for such a model can include, for example, an eccentricity of the hyperbolic function, a spatial channel corresponding to an apex of the hyperbolic function, a volume of the hyperbolic function. Other variable parameters of the model can include, for example, parameters of a function used to define the Ricker wavelet, such as an amplitude of the Ricker wavelet. Each of the variable parameters can be randomly sampled (e.g. from a predetermined range for each variable), to determine an arrival time of the Ricker wavelet at each spatial channel. The Ricker wavelet can then be added to each spatial channel at the arrival time determined for that channel. Panel (a) in Fig. 10 shows an example of simulated signal response for a spatial channel, the simulated signal response being a Ricker wavelet 110 which was inserted at an arrival time T determined for that channel based on a hyperbolic arrival time function (which is a hyperbolic function of arrival time vs. spatial channel position). Similarly, a Ricker wavelet is added to each spatial channel at a respective arrival time determined based on the hyperbolic arrival time function. Panel (b) of Fig. 10 shows an example of a noise model 120 that can be combined with the simulated signal responses to provide simulated DOFS data. The noise model was obtained from measurements performed with a real DOFS system. Panel (c) of Fig. 10 shows the simulated signal response 110 in combination with the noise model 120 to produce a simulated measurement signal 130 for the spatial channel. The simulated measurement signal can be automatically labelled to indicate a location of the simulated signal response, based on the known arrival time T at which the simulated signal response was inserted into the signal. For example, as shown in panel (c) of Fig. 10, a peak 132 and troughs 134, 136 of the simulated signal response can be labelled. The labels can then be used in training the machine learning model. A simulated measurement signal is produced in an analogous manner for each of the spatial channels of the modelled DOFS signal, to produce a set of simulated DOFS data. Further sets of simulated DOFS data are generated, each time randomly sampling one or more of the variable parameters of the acoustic signal and / or arrival time model mentioned above, in order to produce a plurality of training datasets. As an example, approximately 5000 training datasets may be generated, using the described techniques. Each training dataset can include N time samples and M spatial channels (in a particular example, N and M can be 128). Such a number of training datasets may provide suitable coverage of a wide variety of different realistic scenarios, enabling the machine learning model to accurately detect microseismic events.

[0118] Regarding the example of detecting vehicle motion, similar principles can be used to generate a plurality of training datasets including simulated response signals corresponding to vehicle motion. For example, arrival time models can be defined for ranges of speed, acceleration, and direction of travel of the vehicle, which can then be used to add corresponding signal responses in the spatial channels of the modelled DOFS system. Parameters of the arrival time models can be randomly sampled, to generate a large number of training datasets.

[0119] Returning to Fig. 9, in step 904 the plurality of training datasets generated at step 902 are used to train the machine learning model to detect signal responses corresponding to events of the predetermined type in distributed optical fibre sensing data. In more detail, each set of simulated DOFS data generated in step 902 is labelled to indicate portions of the simulated data which correspond to simulated signal responses, and where the simulated signal responses have a higher amplitude than the noise model. Such labelling may be performed automatically to a high degree of accuracy, based on the known arrival times at which the signal responses were inserted into the data. Accordingly, locations in the simulated DOFS corresponding to arrival times where simulated signal responses were added are reviewed, to determine if the simulated signal responses are detectable over the noise model. Where the simulated signal responses are detectable over the noise model, a corresponding portion of simulated DOFS data is labelled accordingly. For instance, in the example where the training datasets include signal responses corresponding to microseismic activity, a peak and / or troughs of the Ricker wavelet can be labelled. Labelling of both peaks and troughs of the wavelets may serve to increase a likelihood of detection of signal responses by the machine learning model. Using the labelled training datasets, a supervised learning process can be performed to train the machine learning model detect (classify) portions of DOFS data corresponding to a signal response resulting from an event of the predetermined type.

[0120] In some cases, the output created by the machine learning model at step 206 of the method 200 can be used as training data for further training the machine learning model. In particular, as discussed above, at step 206 the machine learning model outputs details of locations of the detected signal responses in the DOFS data. Accordingly, the DOFS data can be labelled to indicate the locations of the detected response signals, and used as a training dataset in step 904 to further train the machine learning model.

[0121] As an example, the machine learning model used in the invention can be a semantic segmentation model which can be implemented using a fully convolutional network (FCN). For instance, a U-Net model can be used for this purpose, e.g. as described in the article U-Net: Convolutional Networks for Biomedical Image Segmentation by Olaf Ronneberger et al. (arXiv: 1505.04597). The model can be configured to take as an input the raw time-series data, or a suitable representation of the data. The machine learning model is configured to output a classification of each data element (e.g. each pixel or data point) in the input data, e.g. to indicate a likelihood that the data element corresponds to a signal response of interest. The output can, for example, be in the form of a mask which can be overlaid onto the original input data, to highlight regions of the data that may correspond to signal responses of interest.

[0122] Fig. 11 shows a schematic diagram of a training system 1100 according to an embodiment of the invention. The training system 1100 may be used, for example, to implement the method 100 described above. The training system 1100 includes a first memory 1102 which stores the machine learning model to be trained, and a second memory 1104 which stores instructions for generating a plurality of training datasets and for training the machine learning model with the plurality of training datasets. The generated training datasets may also be stored in the second memory 1104, or in a separate memory (not shown). In practice, the first and second memory can implemented by a same memory device or by separate memory devices. The training system 1100 further includes a processor 1106, which is configured to execute the instructions stored in the second memory 1104, to generate the plurality of training datasets and train the machine learning model. In practice, the training system 1100 can be implemented using any suitable arrangement of computer hardware, such as a personal computer (e.g. desktop or laptop computer), computer server (e.g. cloud-based server), or a network of computing devices.

[0123] The features disclosed in the foregoing description, or in the following claims, or in the accompanying drawings, expressed in their specific forms or in terms of a means for performing the disclosed function, or a method or process for obtaining the disclosed results, as appropriate, may, separately, or in any combination of such features, be utilised for realising the invention in diverse forms thereof.

[0124] While the invention has been described in conjunction with the exemplary embodiments described above, many equivalent modifications and variations will be apparent to those skilled in the art when given this disclosure. Accordingly, the exemplary embodiments of the invention set forth above are considered to be illustrative and not limiting. Various changes to the described embodiments may be made without departing from the spirit and scope of the invention.

[0125] For the avoidance of any doubt, any theoretical explanations provided herein are provided for the purposes of improving the understanding of a reader. The inventors do not wish to be bound by any of these theoretical explanations. Any section headings used herein are for organizational purposes only and are not to be construed as limiting the subject matter described.

[0126] Throughout this specification, including the claims which follow, unless the context requires otherwise, the word “comprise” and “include”, and variations such as “comprises”, “comprising”, and “including” will be understood to imply the inclusion of a stated integer or step or group of integers or steps but not the exclusion of any other integer or step or group of integers or steps.

[0127] It must be noted that, as used in the specification and the appended claims, the singular forms “a,” “an,” and “the” include plural referents unless the context clearly dictates otherwise. Ranges may be expressed herein as from “about” one particular value, and / or to “about” another particular value. When such a range is expressed, another embodiment includes from the one particular value and / or to the other particular value. Similarly, when values are expressed as approximations, by the use of the antecedent “about,” it will be understood that the particular value forms another embodiment. The term “about” in relation to a numerical value is optional and means for example + / - 10%.

Claims

Claims:

1. A method of detecting a signal response in distributed optical fibre sensing data, the method comprising: receiving, from a distributed optical fibre sensing system, distributed optical fibre sensing data, wherein the distributed optical fibre sensing system comprises a plurality of spatial channels, each spatial channel associated with a respective scattering location along an optical fibre of the distributed optical fibre sensing system, and wherein the distributed optical fibre sensing data comprises a plurality of measurement signals as a function of time, each of the plurality of measurement signals corresponding to a respective one of the spatial channels; providing the distributed optical fibre sensing data to a machine learning model, wherein the machine learning model is configured to detect in the distributed optical fibre sensing data a signal response corresponding to an event of a predetermined type, wherein the machine learning model was trained using training data comprising simulated signal responses for events of the predetermined type; and if the machine learning model detects a signal response corresponding to an event of the predetermined type, outputting, by the machine learning model, information indicative of the detected signal response.

2. A method according to claim 1, wherein the distributed optical fibre sensing data is received as a data stream from the distributed acoustic sensing system.

3. A method according to claim 1 or 2, wherein the plurality of measurement signals are raw time-series signals.

4. A method according to any preceding claim, wherein each of the plurality of measurement signals has a sample rate of at least 100 Hz.

5. A method according to any preceding claim, wherein the predetermined type of event corresponds to a vehicle motion relative to the distributed optical fibre sensing system.

6. A method according to any of claims 1 to 4, wherein the predetermined type of event corresponds to a seismic event.

7. A method according to any preceding claim, wherein the information indicative of the signal response comprises an indication of a detection time of the signal response in one or more of the spatial channels.

8. A method according to any preceding claim, wherein the method further comprises generating a representation of the detected signal response.

9. A method according to any preceding claim, wherein the training data further comprises one or more noise models representing background noise in a distributed optical fibre sensing system.

10. An analysis system for detecting a signal response in distributed optical fibre sensing data, the analysis system comprising a processing device, and a memory storing instruction which, when executed by the processing device, cause the processing device to perform the method of one of claims 1 to 9.

11. A distributed optical fibre sensing system comprising: a test signal generator configured to transmit a test signal along an optical fibre; a detector stage configured to receive a plurality of scattered signals from the optical fibre, wherein each scattered signal corresponds to a respective spatial channel associated with a respective scattering location along the optical fibre, and wherein the detector stage is further configured to output a respective measurement signal as a function of each of the plurality of scattered signals; and an analysis system configured to receive the respective measurement signals from the detector stage, and to perform the method of one of claims 1 to 9.

12. A method of training a machine learning model for detecting a signal response in distributed optical fibre sensing data, the method comprising: generating a plurality of training datasets, wherein each training dataset comprises simulated distributed optical fibre sensing data for a distributed optical fibre sensing system having a plurality of spatial channels, each spatial channel associated with a respective scattering location along an optical fibre of the distributed optical fibre sensing system, wherein, for each of the plurality of training datasets, simulating the distributed acoustic sensing data comprises: providing a model of an acoustic signal corresponding to an event of a predetermined type, and generating a simulated signal response to the acoustic signal in each of the spatial channels; and combining the simulated signal response with a noise model to provide simulated distributed optical fibre sensing data, the noise model being representative of background noise in the distributed optical fibre sensing system; andusing the plurality of training datasets to train a machine learning model, wherein the machine learning model is configured to detect a signal response corresponding to an event of the predetermined type in distributed optical fibre sensing data.

13. A method according to claim 12, wherein one or more parameters of the acoustic signal model and / or the simulated signal response signal are varied across the plurality of training datasets.

14. A method according to claim 12 or 13, wherein the noise model is obtained from a measurement performed with a distributed optical fibre sensing system.

15. A method according to one of claims 12 to 14, further comprising providing a plurality of noise models, wherein simulating the distributed acoustic sensing data further comprises selecting one of the plurality of noise models.

16. A method according to one of claims 12 to 15 further comprising, for each of the plurality of training datasets, labelling portions of the simulated distributed optical fibre sensing data where the simulated signal response has a higher amplitude than the noise model.

17. A method according to claim 16, wherein using the plurality of training datasets to train the machine learning model comprises performing a supervised learning process.

18. A training system for training a machine learning model for detecting a signal response in distributed optical fibre sensing data, the training system comprising a processing device, and a memory storing instructions which, when executed by the processing device, cause the processing device to perform the method of one of claims 12 to 17.

Images