WO2013104888A1

WO2013104888A1 - Signature identification

Info

Publication number: WO2013104888A1
Application number: PCT/GB2013/000009
Authority: WO
Inventors: Suresh George JACOB
Original assignee: The Secretary Of State For Defence
Priority date: 2012-01-10
Filing date: 2013-01-09
Publication date: 2013-07-18
Also published as: GB201300384D0; GB2499502B; GB201200313D0; GB2499502A

Abstract

There is provided a method of identifying the presence of an emission signature. The method comprises: forming a first model of the ambient electromagnetic spectrum by adapting coefficients of an adaptive filter so that the impulse response of the adaptive filter resembles the ambient electromagnetic spectrum as closely as possible; and comparing the first model of the ambient electromagnetic spectrum to a second model of the ambient electromagnetic spectrum that was formed when the ambient electromagnetic spectrum was known to contain the emission signature. The deployment of a change point algorithm in comparing the first and second models may be used to quickly identify such emissions. The deployment of devices using the method allows the location of signals to be geo-located. The invention is primarily used for detecting specific emitters, such as those worn by personnel operating in avalanche or earthquake zones, that produce signals that are known.

Description

SIGNATURE IDENTIFICATION

The present invention is in the field of emission detection, in particular when seeking to identify the presence of an emitter by "listening" to the ambient electromagnetic spectrum (or a recording thereof) in an attempt to determine if the specific electromagnetic emission such an emitter produces (typically referred to as its "signature"), is present.

If it can be determined that the specific signature of an emitter (or an approximation that is sufficiently similar) is present, the user can have a reasonable degree of confidence that the emitter is present. In some cases such emitters are specific items that a user is searching for, such as a source of interference, or a particular device. For example, such devices are often carried by persons in avalanche or earthquake prone areas, and after an avalanche or earthquake search teams may attempt to identify the presence of such emitters by attempting to identify their signatures in order to locate survivors. Such emitters may produce a constant tone signature, a fast moving or non-stationary signature, or a combination of emissions thereof. ·

The art discloses numerous methods of attempting to identify the presence of a signature. The first is for' a human being to listen to one or more specified frequencies in one or more specified frequency bands (after an emission has been converted in to an audible representation b a known process). Typically this requires such a person^'to be in close proximity to receiving means, and for the person to be able to listen to the audible representation and to try to determine that an emission is being produced at (or sufficiently similar to) one or more specific frequencies. Listening for a signature requires a person to be trained to identify its characteristics (and they may be required to listen for more than one signature at a time), and/ or for a signal's strength to be sufficiently strong to enable its \ signature to be heard, and/ or for there to be little or no interference which may obfuscate the signature, making it difficult to be certain that a signature is genuinely present. If such a person has to move within or around an area where an emitter is suspected of being, changes to the environment may not be conducive to allowing such signatures to be heard, and typically listening takes place in near-real time, which requires concentration: a person listening may suffer from fatigue, and if they think there may have been a signal, but would like to listen for a second or further time to the emissions in a time frame that has now passed, they are likely to be able to do so only by interrupting their continued monitoring of near-real time emissions, which increases the risk of missing the presence of a signature. Any or all of these factors can act individually or in combination resulting in reduced certainty that a signal can be accurately and/ or consistently identified by a person.

As such, machines have been employed to take the burden away from human beings and to provide greater consistency of identification and verification, and techniques have evolved to suit. It is common for automated means to attempt to scan, recursively or otherwise, across a frequency band and to attempt to identify the presence of signals. In this way emissions can be hear-continuously monitored, a machine can store a record of emissions over time, and can perform any number of analyses on stored information after it has been received.

' The most common method, known jn the art, of analysing signals for signature identification is Fourier analysis. Originally the Discrete Fourier Transform (DFT) was used, and more recently with advances to this method typically the Fast Fourier Technique (FFT) is used.

Fourier showed that any function that changes in time can be divided in to single period signals, so that a representation of the frequency domain can be created. FFT

representations can be achieved in many ways, FFT is based on a certain number of points against an emission having been chosen during initial analysis, an increase in the number of points increases the overall definition of the representation of a signal, ie, the number of Hz represented by each specific sampling point, typically being represented graphically as individual bars, with such bars being referred to as "bins", emphasising that each contains the energy (effectively the voltage) from a frequency range, and not a single frequency. Where an increased number of FFT points leads to a finer definition of a signal, this has to be traded off against the time it takes to process an increased volume of information. As such, a conventional approach to speeding up an FFT is to increase the processing capacity and/ or the speed of hardware in use. This can be appropriate if there is sufficient available powe and space for such hardware to operate effectively, however it does not lend itself well to every circumstance, in particular that of a requirement for portable systems (for example those being deployed for avalanche or earthquake rescue teams), where any additional size and weight of a device to deplo the invention would be undesirable. It also does not address the fundamental problem of an FFT requiring time to function, it simply masks the fundamental problem by speeding up the number of calculations possible in any given time frame.

Because of energy spillage into (and from) adjacent bins, a signal can be distorted or obfuscated, thus.FFT techniques are prone to false detection, and it is only with further processing that the user can be provided with a sufficient degree of confidence that the presence of a signal can be confirmed. For example if a signal is suspected of being present in a particular bin, processes can increase the number of FFT points in that (and perhaps adjacent) bins, re-running analysis to see if it is possible to identify the presence of a signal with more certainty. Naturally this involves additional processing, after the presence of a signal has initially been suspected, which delays the message to the user that a signal, is (or is sufficiently suspected of being) present. -

In an attempt to address some of the shortcomings of the art in a first aspect of the present invention there is provided; A method of identifying the presence of an emission signature comprising the following steps:

- forming a first model of the ambient electromagnetic spectrum by adapting

coefficients of an adaptive filter so that the impulse response of the adaptive filter resembles the ambient electromagnetic spectrum as closely as possible, the first model comprising the adapted coefficients of the adaptive filter;

- comparing the first model of the ambient electromagnetic spectrum to a second model of the ambient electromagnetic spectrum that was formed when the ambient electromagnetic spectrum was known to contain the emission signature; and

- if the first mode) of the ambient electromagnetic spectrum is sufficiently similar to the second model of the electromagnetic spectrum, determine that the emission signature is present.

The invention looks at the problem the art understands-from a completely different perspective. Rather than attempting to identify a specific signal's presence by trying to identify the emission from a)) other electromagnetic activity, the invention seeks to monitor ambient electromagnetic activity and form a model thereof, known as the "first model". The invention then identifies the variance between the first model, and what is expected to happen to that model if a specific signature is present, by comparing the first model to one or more second models that may have been pre-recorded.

In this way, the processing to determine what a model looks like if a signal is present (i.e. if the first model exhibits sufficient properties to one or more such second models), can be formed in advance of the monitoring of ambient emissions. Thus, when the presence of a signal is determined among those emissions, no further processing is required. Thus, by expending effort up-front, a signal can be more clearly identified in near-real time; no further processing of data is required and the user can be informed with more certainty more quickly that an emitter is present.

Optionally, the first model and the second model of the ambient electromagnetic spectrum are of one band. This does away with the FFT method's requirement to perform further analysis on data once a signal is detected. By forming a model of only one band, there is no energy spill over (in either direction) as with the frequency bins of an FFT, and thus the band as analysed is as true a representation of ambient emissions as is possible, and comparison between the models is relatively straightforward. Preferably, the first and/or second models are formed in the time domain so that Fourier analysis in the frequency domain is avoided.

The first model of the ambient electromagnetic spectrum is formed using an adaptive filter. Adaptive filters are known in the art, and self adjust according to optimisation algorithms driven by error signals>This allows for filters with smaller rejection ranges that can respond to_. interference and compensate for it, to enhance the quality of the output signal.

The second model is preferably formed in a similar way to the first model, by adapting coefficients of an adaptive filter so that the impulse response of the adaptive filter resembles the ambient electromagnetic spectrum as closely as possible, at a time when the ambient electromagnetic spectrum is known to contain the emission signature. The same adaptive filter may be used to form the first and second models, or different adaptive filters may be used, for example if there is some time overlap between the time period when the second . model is formed and the time period when the first model is formed. .

The adapted co-efficients of each model represent an occurrence probability distribution of the ambient electromagnetic spectrum that was used to create the model. Accordingly, the adapted co-efficients of each model indicate the frequencies that the_. ambient

electromagnetic spectrum contained over the period of time when the model was formed.

The coefficients of the adaptive filter may be adapted by minimising an error signal between the output of the adaptive filter and the ambient electromagnetic spectrum, to make the impulse response of the adaptive filter match the ambient electromagnetic spectrum as closely as possible. The ambient electromagnetic spectrum may be used as the input to the adaptive filter.

Optionally, the adaptive filter.uses a least mean squares algorithm to adapt the coefficients of the adaptive filter. This is a technique known in the art to be an efficient method for defining filter co-efficient updates by producing the least mean squares of an error signal typically adapted based on the error experienced at the time.

Optionally determination that the emission signature is present is performed by change point detection. Change point detection is known in the art, but not for this ^'application. A change point is desirable because this method requires comparison between two models, and it may be the case that interference experienced by the ambient spectrum (or other factors, weak signal strength for example), results in the first model not directly mapping to one or more of the second models. As such by setting a change point, comparison is not required directly between the first model and one or more of the second models (when more than one is recorded), but between the first model and one or more change points that are defined according to the second model.. . ,

Optionally change point detection comprises a. log likelihood test. This test compares the fit of a first model and a second model based on the logarithm of the likelihood ratio being compared to a critical value to determine if the null model should be rejected in favour of the alternative model. This enhances the accuracy of the comparison to reduce the likelihood of false positives. " -

Optionally at least one of the method steps -is conducted by a Very Large Scale Integration device (VLSI), which may be a Field Programmable Gate Array (FPGA) or a Digital Signal Processor (DSP). VLSI's have low purchase costs, are typically robust and are easily replaceable. FPGA's and DSP's are examples of VLSI's that lend themselves well to conducting one or more of the invention's method steps.

Optionally the signal produced by the ambient electromagnetic spectrum is collected using a receiver. This allows the ambient electromagnetic spectrum to be monitored in near-real time, allowing the first model to be formed from near-real time data.

In a second aspect of the invention, there is provided a method of identifying the location of one or more emitters comprising the following steps:

- deploying at least two devices employing the above described method of identifying the presence of an emission signature; ·

- recording the time and location of each device when deployed;

- overlaying data collected by the devices to a common time frame: and/ or a common location map; and

- using time of arrival and/ or the relative strengths of a signal to determine the location of an emitter. '

This allows the user to capitalise on the strengths of the method to be able to geoTlocate the source of a signal, as collating the strength of a signal from multiple points at the same time allows a user to plot on a map where an emitter is likely to be. The present invention will now be described, by way of example only and with reference to the accompanying figures, in which;

Fig 1 .shows a block diagram of an example architecture in accordance with a first embodiment of the invention;

Fig. 2 shows a block diagram of an adaptive fitter included within of the Fig. 1 block diagram; Fig. 3 shows a block diagram of a comparator included within a DSP of the Fig. 1 block diagram; and

Fig. 4 shows a block diagram of an architecture according to a second embodiment of the invention.

In the figure, an example architecture is shown as 1 , a receiver antenna as 2, a tuner IC as 3, an analogue to digital converter as 4, a DSP as 5, and an adaptive filter stage as 6.

In a first embodiment of the invention a known radio frequency tuner integrated circuit 3 (a "tuner IC" being operable across a plurality of channels) is connected to a known VLSI (not shown in isolation), in this case a DSP 5. It should be noted that so ingrained in the art is the use of FFT techniques for this task, many commercially available DSP's are provided with an FFT core processor, optimised to run FFT algorithms. This invention can be used with a DSP 5 that has an FFT processor, however it would not be optimised for the invention's deployment. This theorem is proven with the DSP 5 in the architecture shown in Figure 1 used in this embodiment being one with a core FFT processor. In this first embodiment the tuner IC 3 is analogue and comprises. a commercial off-the-shelf radio tuner IC, which is a Silicon Labs ® Si 4735: This specific analogue tuner IC 3 is connected to a receiver antenna 2 and is able to tune from 1 MHz to 30MHz; a conversion process is undertaken by an analogue to digital converter 4 to convert the analogue signal the tuner IC 3 detects to a digital one as input for the DSP 5. The skilled- person will appreciate that if a digital tuner IC was selected in lieu of the analogue tuner IC 3, it could be deployed and thus the conversion step from an analogue signal to a digital one could be avoided. In this embodiment the DSP 5 comprises a commercial off-the-shelf unit (a Texas Instruments ® T S320C5XX) which comprises the aforementioned FFT core processor. The skilled person will appreciate that any commercial off the shelf (or bespoke) product with characteristics suited to dentifying and interpreting emissions in the desired frequency range could be used, in particular a ■ bespoke FPGA (not shown). In this embodiment, one specific frequency at a time-is selected for analysis.

Using this specific architecture for the sake of example only, the given tuner IC 3 is able to tune/ hop at the rate of 65ms per frequency with a resolution of 1kHz. When deployed with a DFT method for the same 1kHz resolution over the frequency range 1 Hz to 30MHz, it takes -14 seconds on a 150MHz low-powered DSP (not shown).

If the known FFT method was used on this hardware for emission detection, then the FFT step itself would take 3.51 ms. However this alone would not be sufficient to provide a meaningful answer to the user, since once the FFT step was complete there would be at least two more steps required to determine an emission; the peak corresponding to the emission would need to be located in the FFT, and then interpolated to find the exact .frequency. An optional and common third step would be to use some form of known threshold detection to detect the signal above any present noise. Experiments with all three steps have pushed the total time for the FFT to produce a result to 6ms.

According to this first embodiment of the invention, an adaptive filter stage 6 connected to the DSP 5 comprises gradient estimation and least mean squares algorithms to form a model of the input, such that the impulse response of the adaptive filter resembles the input signal as closely as is possible. ^' .

In order to determine what signal is being searched for (and hence the formation and/ or characteristics. of the second model), a variety of techniques can be deployed. In this first embodiment specific measurement has been taken of the signature of a known emitter which has been recorded in a model format, which is stored on the DSP 5, comprising the second model used by the invention.

In particular, the second model is generated from the input signal when the input signal is known to contain the signature of the known emitter. The coefficients of the adaptive filter are adapted to make the output of the adaptive filter match the input signal. During adaption, the coefficient values gradually converge to values which give the best representation of the input signal, and the final coefficient values constitute_. the second.model. Since the coefficient values are adapted to make the output of the adaptive filter match the input signal, the coefficient values reflect an occurrence probability function of the input signal.

The input signal can later be tested to see if it still contains the signature of the known emitter, by again adapting coefficient values of the adaptive filter to match the input signal, the coefficient values constituting the first model. The first model is then compared to the second model, by comparing the coefficient values of each model together. If the coefficient values of the first model are sufficiently close to the coefficient values of the second model,, then the known emitter is considered to still be present.

A more detailed diagram of the adaptive filter stage 6 is shown in Fig. 2. The DSP 5 outputs a signal SI to the adaptive filter stage 6, the signal S1 representing the ambient

electromagnetic spectrum that is detected by the IC 3 and digitised by the ADC 5. The signal S1 is received at the adaptive filter 20, and is filtered in accordance with the coefficients d - C_N to produce an output signal S2. In this embodiment, the adaptive filter 20 is an FIR filter, and the coefficients d - C_N are applied to successive samples of the input signal. S1 and summated to produce the. output signal S2, as will be apparent to those skilled in the art.

A summation block 22 receives both the input signal S1 and the output signal S2, and . calculates the difference (error signal) ES between them. The error signal ES is fed to a least mean square optimiser block LMS, which optimises the adaptive filter coefficients - CN to minimise the error signal ES so that the output S2 closely follows the input S 1. At first sight, the transfer function, that the adaptive filter converges to should be 1 , although in practice the adaptive filter coefficients converge to reflect an occurrence probability distribution of the frequencies present in the input signal S1.

In particular, when the adaptive filter has converged onto a signal, the spectrum of the signal can be obtained by taking the Fourier Transform (FT) of the filter co-efficients (impulse response). This is in fact an approximation of the signature of the signal under investigation.

Once optimisation has taken place, the optimised (adapted) filter coefficients d C_N are sent to the DSP 5 within a signal CF. The adapted filter coefficients d - C_N are stored by the DSP 5 as a model representing the ambient electromagnetic ^"spectrum. In this embodiment, the adaptive filter stage 6 is used to form both the first arid second models of the ambient electromagnetic spectrum.

As shown in the diagram of Fig. 3, the first and second models MOD1 and MOD2 are stored within a comparator 30 of the DSP, and are compared COMP to one another to determine whether they differ from one another by more than a threshold amount TH. If the first and second models differ by less than the threshold amount, then the ambient electromagnetic spectrum is considered to include the emission signature that was present when the second model was formed. If the first and second models differ by more than the threshold amount, then the ambient electromagnetic spectrum is not considered to include the emission signature.

The skilled person will appreciate that the second model can be refined over time, or a plurality of second models can be formed, and the invention is capable of recursively checking the first model against one or more second models as the user may require. A second model can be used to derive either a single threshold (that may comprise a plurality of specific points) that can be checked against the first model that allows a user to determine the presence of a- specific emitter (or class of emitters). Alternatively, second models can be combined to derive a composite threshold (again comprising a plurality of specific points) that may indicate the presence of a plurality of emitters where, for example, such emitters share common characteristics.

In this first embodiment of the invention, the first model is recursively checked against the second model, using a change point algorithm. The change point algorithm defines the threshold at which the first model is considered to no longer correspond to the second model. The change point is calculated as the threshold that indicates a significant change between the first model and second model. The change point can be fixed or can be adapted over time as experience suggests the most appropriate threshold level to use to minimise false positives and consistently identify the presence of a signal. Using the same hardware as in the examples for DFT and FFT above, the total time taken to produce a meaningful answer to a user is 2ms for a specific band, contrasting with the 6ms it would take FFT to achieve the same. This is possible because the change point detection is a single step that yields the user a result, there is no need for the additional post-processing steps FFT needs (as above this comprises locating the peak, interpolating it to find the exact frequency, and optionally using some form of known threshold detection to detect the signal above any present noise) to get the user an answer. -

Machine learning can be used to determine modifications to an initial change point, where for example in this first embodiment, the occurrence probability distribution of a sequentially inputted data series is learnt as a first model, which is defined by a finite number of variables. Then, an "outlying" score is derived that identifies, the degree of difference between actual data as recorded, and data predicted from the learned first statistical model for each data in the series, and the moving average of the outlying score is recorded. From then on, the occurrence probability distribution of the moving average series of the outlying scores is learnt as a second model that is also defined by a finite number of variables: Each moving average outlying score is derived from the second model and the moving average of outlying scores, and produces a result that is the change-degree score of the original data. Then, the result is compared with a threshold value to detect a change point. If the user so requires, further steps can be used to refine the accuracy of the change point.

In a second embodiment of the invention shown in Fig. 4, an identical method and architecture is used as to the first embodiment, however at least two of the Fig. devices 1 embodying the method of the invention are deployed, to allow the user to identify signals in order to geo-locate an emitter. Initially such devices 1 are either synchronised and/ or the time of deployment is noted such that a common timeframe can be established during which time the devices analyse emissions. The location of each device is known, and at a central station 40 the output from each device is compared and overlaid to a common time frame , and to a common map. Then, time of arrival estimations are used at the central station 40 not only to identify the likely presence of specific emitters, but also to geo-locate such emitters using known techniques. This is possible with the deployment of at least two devices 1 , noting that with two devices it is possible to identify perhaps more than one likely location where an emitter may be (depending on a range of factors, such as the specific geography being assessed, the location of the devices, etc). However, this may be sufficient for a user's needs as for example, if two devices are placed on the sea shore a distance from each other, if time distance of arrival techniques suggest one possible location for an emitter is on land, and another possible location is in the sea, the user may be able to deduce that only one of these locations could possibly host an emitter that produces the signal being searched for, so only one of the possible locations is relevant. If three (or more) such devices are deployed, it will aid the specific identification of the location of an emitter as three data sources allows for triangulation to be used to identify the source. of a signal. It is acknowledged that the increased ability to more precisely identify the location of an emitter using triangulation is traded-off with a greater amount of data that has to be combined and overlaid before conclusions are drawn. The fast transmission of data from each device (using wired . or wireless methods) or to one of the devices (acting as a central "controller" to collate information from all devices), combined with the use of fast hardware to analyse data received by the devices, allows the user to identify the presence and activity of a signal (such as its movements) in near real time.

Claims

1 A method of identifying the presence of an emission signature comprising the following steps: ., ^'

forming a first mode) of the ambient electromagnetic spectrum by adapting coefficients of an adaptive filter so that the impulse response of the adaptive filter resembles the ambient electromagnetic spectrum as closely as possible, the first model comprising the adapted coefficients of the adaptive filter;

comparing the first model of the ambient electromagnetic spectrum to a second model of the ambient electromagnetic spectrum that was formed when the ambient electromagnetic spectrum was known to contain the emission signature; and if the first model of the ambient electromagnetic spectrum is sufficiently similar to the second model of the electromagnetic spectrum, determine that the emission signature is present. A method of identifying the presence of an emission signature according to claim 1 , wherein the first mode) and. the second model of the ambient electromagnetic spectrum is of one band. A method of identifying the presence of an emission signature according to claim 1 or 2, wherein adapting the coefficients of the adaptive filter so that the impulse response of the adaptive filter resembles the ambient electromagnetic spectrum as closely as possible, comprises adapting the coefficients of the adaptive filter to minimise an error signal between the output of the adaptive filter and the ambient electromagnetic spectrum. A method of identifying the presence of an emission signature according to any previous claim, wherein the ambient electromagnetic spectrum is used as the input to the adaptive filter. .·

A method of identifying the presence of an emission signature according to any previous claim, wherein the adaptive filter comprises a least mean squares algorithm. A method of identifying the presence of an emission signature according to any previous claim, wherein determination that the signature is present is performed by change point detection. A method of identifying the presence of an emission signature according to claim 6 wherein change point detection comprises a log likelihood test. A method of identifying the presence of an emission signature according to any previous claim wherein at least one of the method steps is conducted by a Very Large Scale Integration device. A method of identifying the presence of an emission signature according to claim 8 wherein the Very Large Scale Integration device comprises an FPGA. A method of identifying the presence of ah emission signature according to claim 8 wherein the Very Large Scale Integration device comprises a DSP.

A method of identifying the presence of an emission signature according to any previous claim wherein the signal produced by the ambient electromagrietic spectrum is collected using a receiver. ) '

A method of identifying the location of one or more emitters comprising the following steps: .

deploying at least two devices employing the method of -identifying the presence of an emission signature according to any previous claim;

recording the time and location of each device when deployed;

overlaying data collected by the devices to a common time frame and/ or a common location map; and

using time of arrival and/ or the relative strengths of a signal to determine the location of an emitter. . ^~