Method for analyzing an input signal and device integrating an electronic circuit for implementing such a method

Abstract

Title: Method for analyzing an input signal and device integrating an electronic circuit for implementing such a method. The invention relates to a method for analyzing an input signal comprising: - a parameterization step; - a step for obtaining a filtered signal by filtering said input signal; and - a step for analyzing said filtered signal, characterized in that the method comprises, between the obtaining step and the analysis step, a characterization step of the filtered signal, parameterized to allow or prohibit, based on a predetermined setpoint, the analysis step of said filtered signal. Figure for the abstract: Fig. 1

Description

Title of the invention: Method for analyzing an input signal and device integrating an electronic circuit for implementing such a method Scope of the invention

[0001] The field of the invention is that of signal processing.

[0002] More particularly, the invention relates to the field of signal processing in a constrained system, particularly in a system constrained in terms of energy consumption. State of the art

[0003] Recent years have seen the revolution of the Internet of Things (IoT) and that of remote data processing (known by its English expression "cloud computing").

[0004] The number of IoT devices continues to grow annually and it is predicted that at least 30 billion devices will be operational worldwide within a few years.

[0005] Most of these systems transmit raw data for remote processing.

[0006] Often, remote processing methods use energy-intensive Artificial Intelligence (AI) algorithms: for example, the convolutional neural network (CNN) which can have tens of thousands of neurons and a few million connections.

[0007] Since most of the energy consumption is devoted to data transmission (and not to processing), the overall scheme for managing and processing remote data is very energy-intensive.

[0008] It is therefore necessary to increase the semantic level of the information obtained locally within the IoT devices.

[0009] For example, in most cases of thermo-industrial monitoring, the temperature can normally be constant: it is not necessary to transfer these low-level data because the relevant events only occur when there is a significant change in the measurement.

[0010] The detection of this event should be carried out locally, by increasing the semantic level of the signal to be transferred remotely.

[0011] In this case, an alert can be sent remotely, which avoids having to transfer raw low-level data.

[0012] This need for local processing raises a problem however: how to implement complex processing and in particular artificial intelligence-based processing in embedded architectures of the IoT type?

[0013] Indeed, avoiding the continuous transfer of all sensor data implies their local processing.

[0014] This is necessary and has been emphasized in the European Commission's text on communication and digital guidelines, setting a target of 80% of data processing carried out locally.

[0015] However, this raises several technical problems: local data processing must be adapted to the hardware available in embedded systems, particularly in terms of memory and computing power. Furthermore, the introduction of signal processing (or artificial intelligence) increases average power consumption, especially if advanced algorithms such as deep learning are used, even if it also reduces power consumption due to data transmission. Given that battery capacity is limited in many industrial or everyday products, using a significant portion of the energy for artificial intelligence tasks also leads to a reduction in product lifespan in non-rechargeable applications, or the need to recharge the battery or use an external power supply.

[0016] There is therefore a need for a solution to have autonomous devices that can process signals locally, with a significant reduction in the energy consumption of such devices.

[0017] It is within this framework that the technical solution disclosed in the patent document published under number FR3136573 was developed.

[0018] However, this technique can still be improved.

[0019] Indeed, although its energy consumption remains low, the analysis results of this solution often lack precision.

[0020] However, the increase in precision is generally correlated with the increase in computing and / or analysis means which are generally energy-intensive.

[0021] There is therefore a need for a solution to have autonomous devices that can process signals locally, with an increase in their accuracy while minimizing their energy consumption.

[0022] The disclosed technique improves the situation. Objectives of the invention

[0023] The invention aims in particular to overcome the drawbacks of the prior art.

[0024] More specifically, the invention aims to provide a signal analysis solution that allows for high analysis accuracy with lower energy consumption.

[0025] The invention also aims to provide such a solution enabling mixed (analog / digital) processing of an input signal. Description of the invention

[0026] These objectives, as well as others which will appear subsequently, are achieved thanks to the invention which relates to a method for analyzing an input signal comprising: - a parameterization step; - a step of obtaining a filtered signal by filtering said input signal, and - an analysis step of said filtered signal, characterized in that the process includes, between the obtaining step and the analysis step, a characterization step of the filtered signal, parameterized to allow or prohibit, from a predetermined setpoint, the analysis step of said filtered signal.

[0027] This characterization step thus makes it possible to limit the detailed analysis of the input signal to only those cases where the signal presents a signal of interest.

[0028] Thus, the energy consumption of a device implementing the process according to the invention is limited.

[0029] Indeed, energy consumption, which experiences a peak during the precise analysis of an input signal, is only significant when the input signal is of interest to be analyzed, this interest being detected during the characterization step.

[0030] According to an advantageous aspect, the method also includes a step of updating the parameterization step from output data of the characterization step.

[0031] This update step allows for refining the filtering of the input signal and therefore refining the results obtained in the characterization step.

[0032] Consequently, this makes it possible to further limit the energy consumption of a device implementing the process.

[0033] According to another advantageous aspect, the update step of the parameterization step also includes the integration of data from the analysis step of said filtered signal.

[0034] The refinement of the input signal processing can be modified according to the classification of an output signal, i.e. a signal processed precisely during the analysis step.

[0035] This results in a triple learning process to refine the processing of the input signal, namely learning through initial instructions for processing the input signal, learning at the output of the characterization step and learning at the output of the analysis step.

[0036] According to another advantageous aspect, the method also includes a step of temporarily recording an input signal over a predetermined period.

[0037] The recording allows us to go back upstream of the processing point to analyze a signal longer than the portion integrating the signal of interest.

[0038] In addition, the recording makes it possible to guarantee processing of any signal even when the ongoing processing is unintentionally interrupted since it is possible to restart the processing from the recorded input signal.

[0039] According to another advantageous aspect, said predetermined duration is less than or equal to 1 second.

[0040] Such a delay makes it possible to limit the size of the memory for recording the signal, while minimizing the energy consumption of a device implementing the process.

[0041] The invention also relates to an electronic circuit for analyzing an input signal, comprising: - a main electronic unit; - a bank of bandpass filters receiving filtering instructions issued by the electronic unit to obtain a filtered signal from the input signal, and - a convolutional neural network for analyzing said filtered signal, characterized in that the filters of the bank of bandpass filters are each coupled to envelope detectors, and in that the electronic circuit includes, between the bank of bandpass filters and the convolutional neural network, an analysis unit configured to allow analysis of the filtered signal by the convolutional neural network when at least one frequency of interest is detected in the filtered signal or, conversely, to prohibit analysis of the filtered signal by the convolutional neural network when no frequency of interest is detected in the filtered signal.

[0042] The analysis unit thus makes it possible to limit the detailed analysis of the input signal to only those cases where the signal presents a signal of interest.

[0043] Thus, the energy consumption of said device is limited.

[0044] Indeed, energy consumption, which experiences a peak during precise analysis of an input signal by the convolutional neural network, is only important when the input signal has an interest to be analyzed, this interest being detected by the analysis unit.

[0045] According to an advantageous aspect, the analysis unit comprises a simplified neural network and secondary electronic unit.

[0046] Such a simplified neural network can evolve over time in order to detect new signals of interest or, in other words, in order to detect updated signals of interest based on input signals, for example.

[0047] Furthermore, the secondary electronic unit ensures, in a simple and low-energy-consuming manner, the choice of allowing or prohibiting the fine analysis of the input signal by the convolutional neural network.

[0048] According to another advantageous aspect, the analysis unit is configured to be continuously active.

[0049] This makes it possible to limit the detailed analysis to only the input signals presenting a signal of interest, for example a target frequency corresponding to a particular event to be identified.

[0050] According to another advantageous aspect, the analysis unit is parameterized to transmit analysis data to the main electronic unit for the modification of existing filtering instructions or the creation of new filtering instructions.

[0051] This allows for fine-tuning the filtering of the input signal and therefore for refining the results obtained at the output of the analysis unit.

[0052] Consequently, this makes it possible to further limit the energy consumption of said device.

[0053] According to another advantageous aspect, the convolutional neural network is parameterized to transmit to the main electronic unit data modifying the filtering instructions.

[0054] The refinement of the input signal processing can be modified according to the classification of an output signal, i.e. a signal processed precisely by the convolutional neural network.

[0055] This results in a triple learning process to refine the processing of the input signal, namely learning through initial instructions for processing the input signal, learning at the output of the analysis unit and learning at the output of the convolutional neural network.

[0056] According to another advantageous aspect, the filters in the bandpass filter bank are each coupled to envelope detectors.

[0057] Envelope detectors allow for a moment extraction operation on the input signal, that is, converting a high-frequency signal (acoustic signal) to a very low-frequency signal (limited by the time constant of the envelope detector's low-pass filter). Coupled with a single bandpass filter, each envelope detector provides the average energy present in a frequency band defined by the bandpass filter. Integrated into a filter bank, the set of N (a bandpass filter paired with an envelope detector) allows for determining the average energy present in each of the frequency bands selected by the different bandpass filters. This output quantity evolves slowly (due to the time constant of the envelope detector's low-pass filter), which allows for (in analog form) a dual time-frequency analysis of a transform. Fast Fourier (FFT) in digital, while guaranteeing much lower energy consumption.

[0058] The invention further relates to an electronic device incorporating a circuit as previously described.

[0059] The invention finally relates to a computer program product downloadable from a communication network and / or stored on a computer-readable medium and / or executable by a microprocessor, characterized in that it includes program code instructions for the execution of a method for analyzing an input signal as previously presented, when executed. Figures

[0060] Other features and advantages of the invention will become more apparent from the following description of a preferred embodiment of the invention, given by way of illustrative and non-limiting example, and the accompanying drawings described below.

[0061] [Fig.1] Fig.1 is a schematic representation of a device integrating an electronic circuit according to the invention, for the analysis of an input signal.

[0062] [Fig.2] Fig.2 is a schematic representation of a method for analyzing a input signal according to the invention. Detailed description of the invention

[0063] The same elements have been designated by the same reference numerals in the different figures. In particular, the structural and / or functional elements common to the different embodiments may have the same reference numerals and may have identical structural, dimensional and material properties.

[0064] For the sake of clarity, only the steps and elements useful for understanding the described embodiments have been shown and are detailed. In particular, circuits for generating a signal and controlling the frequency or intensity of that signal, as well as circuits for controlling and receiving values ​​provided by sensors, are not described in detail, as the described embodiments are compatible with such common circuits.

[0065] Unless otherwise specified, when referring to two elements connected together, this means directly connected without intermediate elements other than conductors, and when referring to two elements linked or coupled together, this means that these two elements can be connected or linked or coupled through one or more other elements.

[0066] In the following description, when referring to absolute positional qualifiers, such as the terms "front", "back", "top", "bottom", "left", "right", etc., or relative positional qualifiers, such as the terms "above", "below", "superior", "inferior", etc., or to Orientation qualifiers, such as "horizontal," "vertical," etc., refer, unless otherwise specified, to the orientation of the figures. Unless otherwise specified, the expressions "approximately," "roughly," "about," and "in the order of" mean within 10%, preferably within 5%.

[0067] As explained previously, the disclosure relates, in a first aspect, to a device 100 integrating an electronic circuit which includes an analog portion and a digital portion for the processing of an input signal S.

[0068] The analog portion is notable in that it includes configurable functionalities (FPAA, FPMA). This analog portion comprises a plurality of potentially usable functionalities (filters in particular), called "features" or "primitives." These functionalities can be independently activated and configured so that their power consumption is limited to their actual use. The potential number of implementable configurable functionalities depends essentially on the complexity of the analog portion.

[0069] The circuit also includes a digital portion. This digital portion is interfaced with the analog portion via at least one analog / digital converter 110 (a single converter 110 being illustrated by the schematic representation in [Fig.1]).

[0070] The digital portion, for its part, includes in particular decision-making means, in this case a convolutional neural network 120.

[0071] The convolutional neural network 120 typically comprises a few hundred or more neurons depending on the intended application and the intended power consumption.

[0072] In such an electronic circuit, as explained in more detail below, the input signal S is therefore processed using a mixed architecture in which the analog portion is implemented to perform a first series of data processing (frequency processing, moment extraction, etc.) and this data obtained from the input signal S is provided to the convolutional neural network 120, on the digital portion for processing at lower frequencies and obtaining results (for example a classification of the input signal S).

[0073] This architecture allows for a first limitation of energy consumption.

[0074] Indeed, an important feature of the disclosure relates to the general configuration of the proposed electronic circuit.

[0075] The learning of the analog and digital parameters of such a circuit is carried out globally and includes both the learning of the parameters of the digital neural network and, at the same time, the learning of the parameters of the primitives of the analog portion.

[0076] Thus, the circuit is usable in many situations and can have a single certification, which also reduces implementation costs.

[0077] In other words, the primitive network constitutes a mixed electronic architecture (analog and digital) intended to implement embedded artificial intelligence for decision-making or pattern identification, targeting ultra-low power applications with joint optimization of energy consumption and classification accuracy, with the following particularities:

[0078] The proposed technique differs from other classic architectures for embedded 1TA comprising a fixed analog front end circuit and a digital portion implementing neural networks, in that it allows joint end-to-end learning of the parameters of the analog and digital parts.

[0079] The proposed technique makes it possible to obtain ultra-low power consumption systems and networks of limited size.

[0080] By way of indication and not limitation, the technique allows in particular the processing of an input signal S in a frequency range close to that of speech, or even of ultrasound, i.e. from 100Hz up to 100Hz.

[0081] For this purpose, with reference to [Fig. 1], the electronic circuit comprises, in its analog part: - a main electronic unit 130; - a bank of 140 bandpass filters, and - an analysis unit of 150.

[0082] The main electronic unit 130 includes a set of configurable pre-filters for amplifying or attenuating frequency areas of the input signal S. These pre-filters are configurable in frequency and gain, as will be explained below.

[0083] The bandpass filter bank 140 receives filtering instructions from the electronic unit 130 to obtain a filtered signal from the input signal S. The filters are coupled to envelope detectors. The bandpass filter bank 140 (with the envelope detectors) allows for a subsequent time-frequency analysis, which is dual, of a Fast Fourier Transform (FFT). As explained previously, the bandpass filter bank coupled to envelope detectors (each bandpass filter is coupled to a single envelope detector) allows for real-time, dual time-frequency analysis of an FFT.

[0084] By way of example and not limitation, the bandpass filter bank comprises sixty-four filters that can be used simultaneously or not. As will be explained later, the boosting or attenuation of certain filters is the result of continuous learning.

[0085] The analysis unit 150 is positioned between the bandpass filter bank 140 and the convolutional neural network 120. More precisely, the analysis unit 150 is located upstream of the converter 110.

[0086] The analysis unit 150 is configured to allow analysis of the filtered signal by the convolutional neural network 120 when at least one frequency of interest is detected in the filtered signal or, conversely, to prohibit analysis of the filtered signal by the convolutional neural network 120 when no frequency of interest is detected in the filtered signal.

[0087] In more detail, the analysis unit 150 comprises a simplified neural network 151 and a secondary electronic unit 152.

[0088] The filtered signal is analyzed by the simplified neural network 151 to detect a signal of interest, for example a frequency peak or a predetermined sequence. The use of frequency peaks in a chronological manner can, in particular, make it possible to obtain a "snapshot" of the past, that is to say, of the evolution of the signal over time, for a short duration, for example on the order of a few milliseconds.

[0089] When a signal of interest is detected by the simplified neural network 151, the secondary electronic unit 152 transmits to the converter 110 a command to convert the input signal S and / or the filtered signal so that it can be analyzed by the convolutional neural network 120.

[0090] Furthermore, the secondary electronic unit 152 transmits analysis data to a supervisor 160 which transforms said analysis data into filtering conditions to be integrated by the electronic unit 130 to generate new filtering instructions or modify existing filtering instructions.

[0091] This is reflected for example by the increase or decrease of frequency bands taken into account by the bandpass filter bank 140.

[0092] In order to further limit the energy consumption of the electronic circuit, and therefore of the device 100, the analysis unit 150 is configured to be continuously active.

[0093] In other words, as soon as the device is switched on or the electronic circuit is powered, the analysis unit 150 is active.

[0094] The device 100 also includes a classification tool 170 and a buffer memory 180.

[0095] The classification tool can, for example, take the form of a command line executed by a dedicated computer to classify the analysis results output from the convolutional neural network 120.

[0096] As for the buffer memory 180, it allows recording for a duration less than or equal to 1 second, and preferably less than or equal to 0.6 seconds, the input signal or each signal S for its analysis as explained below.

[0097] The device 100 is used for the implementation of a method for analyzing an incoming signal S as illustrated by [Fig.2].

[0098] The method for analyzing an input signal S includes a parameterization step 210 during which filtering parameters are executed to generate filtering instructions for each input signal S.

[0099] This parameterization step 210 can be carried out by a technician who loads said parameters into the electronic unit 130 during a first use of the device 100, or by the electronic unit 130 which then executes lines of code, during continuous use.

[0100] The process then includes a step of obtaining 220 a filtered signal by filtering said input signal S. In the device of [Fig.1], it is the bank of bandpass filters 140 which makes it possible to obtain the filtered signal from the filtering instructions transmitted by the electronic unit 130.

[0101] In other words, the coupling between the bandpass filter bank 140 and the electronic unit 130 allows obtaining a filtered signal.

[0102] The filtered signal can then be analyzed during an analysis step 240 to allow classification of the input signal S for example.

[0103] In device 100 of [Fig. 1], the analysis is performed by the convolutional neural network 120

[0104] As explained previously, in order to limit the energy consumption related to its implementation, the process includes, between the obtaining step 220 and the analysis step 240, a characterization step 230 of the filtered signal.

[0105] This step is configured to allow or prohibit, based on a predetermined instruction, the analysis step 240 of said filtered signal.

[0106] The predetermined setpoint can be modulated according to the use of the process, in particular according to the type of input signal and / or the type of analysis to be performed.

[0107] In the device 100 of [Fig.1], this characterization step 230 is carried out by the analysis unit 150, and more particularly, by the simplified neural network 152 which detects a signal of interest in the filtered signal, then by the secondary electronic unit 152 which authorizes or not the analysis of the input signal S during the analysis step 240.

[0108] Furthermore, as illustrated by [Fig.2], the method also includes an update step 250 of the parameterization step 210 from output data of the characterization step 230.

[0109] This then allows learning by the analog part of the electronic circuit of device 100.

[0110] Learning by the analog part of the electronic circuit of the device 100 is also carried out by integrating data from the analysis step 240 of said device filtered signal, or more generally of the input signal S. It should be noted that the filtered signal is a refined signal of the input signal S.

[0111] During the analysis step 240, the analyzed signal is either the filtered signal or the input signal S in its raw version. In the device 100 of [Fig. 1], the analyzed signal comes either from the bandpass filter bank 140, in which case it is the filtered signal, or from the buffer memory 180, in which case it is the input signal S in its raw version.

[0112] In order for the signal analyzed during the analysis step 240 to be the input signal S in its raw version, the method also includes a temporary recording step 260 of the input signal S over a predetermined duration.

[0113] Said predetermined duration is for example less than or equal to 1 second and preferably less than or equal to 0.6 seconds.

[0114] Through its implementation by the illustrated embodiment using device 100, the method just described makes it possible to perform a precise analysis of an input signal S with limited energy consumption. The target consumption of device 100 in one year is, for example, equivalent to the electrical energy contained in a button cell battery, for example, known by the trade name LR44.

[0115] This is made possible thanks to the triple learning offered by the process and its implementation.

[0116] Indeed, a first learning is carried out by the implementation during the parameterization phase 210 of filtering setpoint which makes it possible to reduce the noise of the signal, that is to say the frequencies which are not of interest.

[0117] This parameterization phase 210 reduces noise or amplifies certain frequency ranges of the input signal S. The parameterization phase 210 is dynamically adjusted during operation. Therefore, there is a coupling of initial learning and dynamic adjustment. The dynamic adjustment is based on criteria derived from the use of the analysis unit 150 and the classification tool 170.

[0118] A second learning process takes place in the update step 250 during which data from the analysis step 240 are used to modify the filtering instructions or generate new ones.

[0119] In device 100 this translates into the enhancement or attenuation of certain filters from the bandpass filter bank 140 from the data provided by the classification tool 170.

[0120] Finally, a third learning process takes place in the characterization step 230. Indeed, the data from this step are also used in the update step 250.

[0121] More specifically, the data relating to the detection of a signal of interest are used to perform additional learning enabling the supervisor 160 to transform said analysis data into filtering conditions to be integrated by the electronic unit 130 in order to generate new filtering instructions or modify existing filtering instructions.

[0122] Thus, thanks to the first analysis step by the analysis unit 150, the input signal S is analyzed accurately only when it has certain characteristics corresponding for example to a signal of interest such as a target frequency.

[0123] Therefore, the energy consumption related to the implementation of the process by the device 100 remains limited and is increased only for certain input signals S which prove to be relevant.

Claims

Demands

1. Method for analyzing an input signal comprising: - a parameterization step; - a step for obtaining a filtered signal by filtering said input signal, and - a step for analyzing said filtered signal, characterized in that the method comprises, between the obtaining step and the analysis step, a filtered signal characterization step, parameterized to allow or prohibit, based on a predetermined setpoint, the analysis step of said filtered signal.

2. A method according to the preceding claim, characterized in that it also includes a step of updating the parameterization step from output data of the characterization step.

3. A method according to the preceding claim, characterized in that the update step of the parameterization step also includes the integration of data from the analysis step of said filtered signal.

4. A method according to any one of the preceding claims, characterized in that it also includes a step of temporarily recording an input signal over a predetermined period.

5. Method according to the preceding claim, characterized in that said predetermined duration is less than or equal to 1 second.

6. An electronic circuit for analyzing an input signal, comprising: - a main electronic unit; - a bandpass filter bank receiving filtering instructions issued by the electronic unit to obtain a filtered signal from the input signal; and - a convolutional neural network for analyzing said filtered signal, characterized in that the filters in the bandpass filter bank are each coupled to envelope detectors, and in that the electronic circuit comprises, between the bandpass filter bank and the convolutional neural network, an analysis unit configured to allow analysis of the filtered signal by the convolutional neural network when at least one frequency of interest is detected in the filtered signal or, conversely, to prohibit analysis of the signal filtered by the convolutional neural network when no frequency of interest is detected in the filtered signal.

7. Circuit according to the preceding claim, characterized in that the analysis unit comprises a simplified neural network and secondary electronic unit.

8. Circuit according to any one of claims 6 or 7, characterized in that the analysis unit is configured to be continuously active.

9. Circuit according to any one of claims 6 to 8, characterized in that the analysis unit is parameterized to transmit analysis data to the main electronic unit for the modification of existing filtering instructions or the creation of new filtering instructions.

10. Circuit according to any one of claims 6 to 9, characterized in that the convolutional neural network is parameterized to transmit to the main electronic unit data modifying the filtering instructions.

11. Electronic device incorporating a circuit according to any one of claims 6 to 10.

12. Product computer program downloadable from a communication network and / or stored on a computer-readable medium and / or executable by a microprocessor, characterized in that it includes program code instructions for the execution of a method for analyzing an input signal according to any one of claims 1 to 5, when executed.

Images