Resonant MEMS sensor for generating a pulse output signal

Description

Technical field

[0001] The present invention relates to resonant MEMS (Microelectromechanical Systems) sensors. These resonant MEMS sensors form a large family of MEMS sensors: mass sensors, gas sensors, accelerometers, etc. These sensors operate around a resonant frequency, which depends on the parameter of interest. In the case of a gas sensor, the resonant frequency depends directly on the mass of the mobile structure, which itself depends on the species adsorbed on the surface and therefore on the gas concentration. In the case of a resonant accelerometer, the frequency of a resonant mobile structure depends on the acceleration via the quasistatic displacement of a mass subjected to inertial forces. For all of these sensors, the mechanical signal of interest is an alternating signal, even when the signal to be measured is a DC signal (constant acceleration, or constant gas concentration).

[0002] The signal from the MEMS is processed by an interface electronics, referred to as the "analog front-end" (AFE) hereinafter. At the output of the AFE, the signal can be processed in different ways, by on-board electronics, or by remote processing units, in real time or delayed, in order to deliver the relevant information.

[0003] The invention relates to all resonant MEMS sensors. It is particularly relevant when the signal processing is carried out in real time by real-time embedded neuromorphic type processing, i.e. processing imitating the functioning of neurons. Prior art

[0004] US 2007 / 152682 A1 discloses a capacitive sensor.

[0005] In recent years, sensors have multiplied in many fields (IoT, automotive, etc.). Data from these sensors are increasingly processed by "machine learning" type processing, which allows relevant information to be extracted from a large amount of data. These processes can be done on classic digital units, i.e. microprocessors whose transistors reproduce the structure of a neural network, or, advantageously, with circuits whose hardware structure mimics as closely as possible the operation of a neural network. This type of circuit has several advantages: Reduction of consumption by several orders of magnitude No intermediate storage of generated data No data transfer to an external control unit Real-time processing.

[0006] This type of circuit is based on spiking neural networks (SNNs). However, optimal coupling with the sensors generating the pulses is rarely achieved. We define as "spike-based sensor" any type of sensor to which is added electronics processing the information as a sequence of events or pulses without going through a classic analog-to-digital converter. In the article "Towards spike-based machine intelligence with neuromorphic computing" by K. Roy, A. Jaiswal, and P. Panda, Nature, vol. 575, no. 7784, pp. 607-617, Nov. 2019, doi: 10.1038 / s41586-019-1677-2, a state of the art of neuromorphic processing in integrated circuits is presented. These spike-based sensors are used for vision, as reported in the paper “A Spike-Based Neuromorphic Architecture of Stereo Vision” by N. Risi, A. Aimar, E. Donati, S. Solinas, and G. Indiveri in Frontiers in Neurorobotics, vol.14, 2020, Accessed: Jul. 07, 2022. [Online]. Available: https: / / www.frontiersin.org / articles / 10.3389 / fnbot.2020.568283. These spike-based sensors are also used for audio, as reported in the paper "AER EAR: a matched silicon cochlea pair with address event representation interface," by A. van Schaik and S.-C. Liu in 2005 IEEE International Symposium on Circuits and Systems (ISCAS), May 2005, pp. 4213-4216 Vol. 5. doi: 10.1109 / ISCAS.2005.1465560.

[0007] Regarding odor sensors, the paper "Rapid online learning and robust recall in a neuromorphic olfactory circuit" by N. Imam and TA Cleland in Nature Machine Intelligence, vol. 2, no. 3, pp. 181-191, Mar. 2020, doi: 10.1038 / s42256-020-0159-4 proposes a real-time conversion into pulse of the output voltage after digital conversion of a "chemo-sensor" measuring changes in conductivity. This digital conversion step makes the system suboptimal but it allows the conversion into pulses to be easily carried out with commercial electronic components.

[0008] There is a lot of work on resonant sensors and how to read the resonant frequency with the desired sampling and resolution.

[0009] One example is approaches using a PLL (for "Phase-Locked Loop"). In this type of architecture, the resonant structure is excited at its resonant frequency via phase control. The excitation frequency therefore reproduces a variable of interest as explained in the document "Frequency-addressed NEMS arrays for mass and gas sensing applications" by E. Sage et al. in 2013 Transducers & Eurosensors XXVII: The 17th International Conference on Solid-State Sensors, Actuators and Microsystems (TRANSDUCERS & EUROSENSORS XXVII), Jun. 2013, pp. 665-668. doi: 10.1109 / Transducers.2013.6626854.

[0010] Other approaches involve a self-oscillation loop comprising the MEMS and amplifying electronics as shown in the paper "Improved Interface Circuits for CMUT Chemical Sensors" by Q. Stedman, JD Fox, and BT Khuri-Yakub in 2019 IEEE International Ultrasonics Symposium (IUS), Oct. 2019, pp. 989-992. doi: 10.1109 / ULTSYM.2019.8926206. With proper sizing, this type of architecture naturally starts to oscillate at a frequency close to the resonant frequency. A second electronic stage acts as a frequency counter, which gives a digital output signal proportional to the self-oscillation frequency. A special type of sensor is the electronic nose. This is a set of sensors, for example "gravimetric", each with a particular chemical affinity towards the different gases likely to be measured. Thus, the response of each sensor to a given set of gases is particular.Processing data from the various sensors allows the composition of the gas to be determined. This principle, which mimics the functioning of the human nose, is called an electronic nose.

[0011] In the case of an electronic nose based on resonant gravimetric sensors (cMUT, SAW, NEMS, quartz microbalance, etc.), the signal from each sensor is a resonance or self-oscillation frequency.

[0012] Information processing therefore consists of detecting the resonance frequency value for each sensor, then combining the different frequencies to extract information about the nature (and quantity) of the gas. This processing is done via a reference database. In the case of "machine learning" processing, this database is used for learning, which will enable classification.

[0013] There is a need to propose a generic architecture for resonant MEMS sensors suitable for delivering signals in the form of pulses. Statement of the invention

[0014] The present invention aims to at least partially address this need.

[0015] More particularly, the present invention aims to cover a resonant MEMS sensor adapted to generate a pulsed output signal from a signal of interest, said signal of interest being a square signal having a frequency oscillating around a carrier frequency, said MEMS sensor comprising at least one channel for processing the signal of interest, each processing channel comprising: a block for demodulating the signal of interest to form a demodulated signal, said demodulation block comprising a frequency mixer between said signal of interest and a reference signal, said demodulated signal having a low-frequency component and a high-frequency component; a block for filtering the demodulated signal to form a filtered signal, the filtering block being adapted to allow the low-frequency component of the demodulated signal to pass; a block for comparing the filtered signal with a fixed threshold signal to form a comparison signal, said comparison signal comprising rising edges and falling edges; a block for detecting the rising edges, each rising edge corresponding to a pulse of the output signal.

[0016] Thus, the invention makes it possible to obtain a generic architecture for resonant MEMS sensors whose carrier frequency, also called resonance frequency, is measured, so that these sensors are "spike-based sensors", delivering signals in the form of pulses. In terms of electronics, the advantages are numerous, particularly for integration. In addition, the binary nature of the signal makes it very robust to electromagnetic disturbances. In addition, the quantity of pulses is proportional to the activity / intensity of the signal. This allows for reduced consumption when the signal is little present or absent. It is the resonant MEMS sensor, via its conditioning electronics, which is adapted to generate the pulsed output signal.

[0017] In a particular embodiment, the sensor comprises a CMUT transducer adapted to generate the signal of interest from a variation of a physical characteristic studied.

[0018] In a particular embodiment, the physical characteristic studied from which the signal of interest is generated is selected from a group of physical characteristics comprising at least: a gas; a mass; an acceleration.

[0019] In a particular embodiment, the sensor comprises at least two processing channels, each processing channel having a CMUT transducer determined to generate a pulsed output signal particular to said processing channel.

[0020] In a particular embodiment, each processing channel is coupled at output to a classifier, said classifier being adapted to classify the physical characteristic studied.

[0021] In a particular embodiment, the classifier is adapted to process digital data and in that each processing channel comprises a block for converting pulses into digital data.

[0022] In a particular embodiment, the classifier is a spiking neural network.

[0023] The use of such a classifier is particularly relevant in the case of matrix or multi-channel sensors which are as many parallel inputs for the neuromorphic circuit.

[0024] In a particular embodiment, the sensor comprises a transition detector, said transition detector being adapted to put all or part of the processing channels into standby mode if the physical characteristic studied does not vary over a certain period, said transition detector being adapted to take all or part of the processing channels out of standby mode if the physical characteristic studied varies rapidly.

[0025] In a particular embodiment, the MEMS sensor comprises a reference signal calibration block in the demodulation block.

[0026] The present invention will be better understood upon reading the detailed description of embodiments taken as non-limiting examples and illustrated by the appended drawings in which:

[0027] [ Fig 1 ] there figure 1 illustrates a resonant MEMS sensor according to a first embodiment of the invention;

[0028] [ Fig 2 ] there figure 2 illustrates a resonant MEMS sensor according to a second embodiment of the invention comprising a plurality of processing channels;

[0029] [ Fig 3 ] there figure 3 illustrates the principle of converting pulses generated by MEMS sensors from figures 1 And 2 to a digital signal;

[0030] [ Fig 4 ] there figure 4 illustrates a resonant MEMS sensor according to a third embodiment of the invention;

[0031] [ Fig 5 ] there figure 5 illustrates the principle of detecting transitions on each processing channel of the resonant MEMS sensor of the figure 2 ;

[0032] [ Fig 6 ] there figure 6 illustrates a resonant MEMS sensor according to a fourth embodiment of the invention.

[0033] The invention is not limited to the embodiments and variations presented and other embodiments and variations will become apparent to those skilled in the art.

[0034] There figure 1 illustrates a resonant MEMS sensor 10 according to a first embodiment of the invention. In this embodiment the resonant MEMS sensor comprises: a set 100; a demodulation block 101; a filtration block 102; a comparison block 103; a detection block 104.

[0035] The assembly 100 is adapted to generate a signal of interest S Int from a variation of a physical characteristic studied. Such a physical characteristic studied is, for example, a gas, a mass or an acceleration. It will be noted that the signal of interest S Int is here a periodic signal, for example a square signal. By square signal, we mean a signal alternating regularly and instantaneously between two levels.

[0036] More particularly, the assembly 100 comprises a CMUT sensor 1000 and a self-oscillator 1001. This self-oscillation electronics is designed so that the MEMS spontaneously starts to oscillate when the electronics are powered, in particular by a DC voltage. The oscillation frequency of a resonant MEMS is dictated by its mechanical structure, designed to obtain the desired performance. For example, in the case of a cMUT transducer, this frequency can be between 10 and 50 MHz for a bandwidth of 100 kHz.

[0037] The demodulation block 101 is adapted to demodulate the signal of interest S Int to form a demodulated analog signal S Dem . This demodulation block 101 comprises a frequency mixer 1010 between the signal of interest S Int and a reference signal V demod . This demodulated analog signal S Dem has a low frequency component and a high frequency component. The demodulation electronics thus have the role of bringing the signal back to a low frequency between 0 Hz and the value of its bandwidth, 100 kHz in our example. In the absence of a physical signal of interest, for example in the event of no acceleration or no gas, the demodulation electronics bring the signal back to a zero frequency. This makes it possible to have a DC signal in the absence of a physical signal of interest, which makes it possible to limit the consumption of certain subsequent blocks.

[0038] The processing carried out by the demodulation block 101 is based on the principle of frequency mixing between a signal of interest sin(w 0 + dw)t called S Int and a reference signal sin w 0 t called (V demod ). This is the heterodyne method, based on a multiplication of several frequencies illustrated by the following equation: sin w 0 t * sin w 0 + dw t = 1 2 * cos dwt − 1 2 * cos 2 w 0 + dw t in which w 0 corresponds to a pulsation and dw corresponds to a variation of pulsation.

[0039] Frequency mixing here involves switches. Using an on / off switch on the signal of interest S Int is equivalent to mixing it with a square wave.

[0040] The filtering block 102 of the demodulated signal S Dem is adapted to form a filtered analog signal S Fil . It thus allows the low-frequency component of the demodulated analog signal S Dem to pass. This filtering block 102 is also called a “Low Pass Filter”.

[0041] The comparison block 103 is adapted to compare the filtered signal S Fil with a fixed threshold signal Vth to form a comparison signal S Comp . This comparison signal S Comp comprises rising edges and falling edges. This comparison block 103 is also called a “comparator”.

[0042] The detection block 104 is adapted to detect rising edges. Each rising edge will correspond to a pulse of the output signal S Imp . This detection block 104 is also called “Rising Edge”.

[0043] The CMUT transducer 100, the demodulation block 101, the filtration block 102, the comparison block 103, the detection block 104 form a processing channel C or path of the MEMS sensor 10.

[0044] There figure 2 illustrates a resonant MEMS sensor according to a second embodiment of the invention comprising a plurality of processing channels C 0 , .. , C i , ..., CN . Such a sensor is called cMUT or multi-channel cMUT sensor. Each processing channel C 0 ,.., C i , ..., CN comprises a CMUT transducer, a demodulation block, a filtration block, a comparison block and a detection block. All of the processing channels C 0 , .., C i , ..., CN are output coupled to the same classifier, for example a classifier of the SNN type of spiking neural network. Such a classifier is suitable for classifying the physical characteristic studied.

[0045] It should be noted that the coupling between a cMUT and the SNN is not obvious because the expertise is carried by different actors, from different communities.

[0046] It should also be noted that in the case of a resonant accelerometer, a single-channel system can be designed that delivers temporal information in the form of pulses, which is processed to analyze the type of movements. For comparison, data fusion algorithms within a smartphone can determine whether the smartphone user is walking, in a train, or in a car.

[0047] A system formed by several resonant accelerometers measuring acceleration along the 3 axes can also be proposed.

[0048] Generally speaking, any combination with other types of resonant sensors (gyrometers, pressure sensors, etc.) can lead to the creation of a “spike-based sensor” delivering relevant signals to the SNN.

[0049] In case the signals are not directly used by an SNN, an electronic block allows the pulse signal to be converted into a digital signal containing information on the oscillation frequency of the MEMS. figure 3 shows the principle of such a block with as input a series of asynchronous pulses (upper line in the figure), an asynchronous pulse counter (middle line) and a reference signal (lower line) which sends a counter reset signal with a fixed frequency, allowing the measurement of a pulse density.

[0050] The electronic block for converting the signal is illustrated in the figure 6 and positioned at the output of a detection block. In this figure, each processing channel C 0 , .., C i , ...CN includes a conversion block Bc 0 , .., Bc i , ...Bc N of pulses into digital data. These conversion blocks Bc 0 , , Bc i , .., Bc N are intended to feed a classification system which is not an SNN. For heterodyne demodulation to work optimally, with a pulse density equal to 0 when the CMUT resonates at its no-load frequency f 0 =w 0 / 2π, the demodulation signal at the mixer input must have a frequency equal to this no-load frequency.

[0051] There Figure 4 presents a possibility of calibration circuit 105 where the CMUT resonates at w 0 . The reference signal Vdemod is generated by a frequency controlled oscillator (VCO). Its voltage to phase transfer function is an integrator. The phase difference between Vdemod and the output of the auto-oscillator is made with a phase comparator. A low-pass filter is added to stabilize the VCO control voltage.

[0052] In normal operation, the feedback loop of the calibration circuit is cut and the calibrated voltage is maintained at the VCO input.

[0053] This circuit can also be used to compensate for drifts in operation with an appropriate algorithm, for example a minimum detection over a large time window.

[0054] Some odor classifiers may operate only on stabilized sensor states. If this is the case, a transition detector can be used to inhibit the classifier when the sensor state has not reached stability.

[0055] There Figure 5 shows that during transitions, the demodulated signal becomes high frequency. This is explained by the fact that the frequency step contains all the frequencies that are demodulated in the baseband. At the top, the modulation frequency varying over time, in the middle the output of the low-pass filter after demodulation and at the bottom the pulse train.

[0056] On the Figure 6 , at least one transition detector 30 connected to each channel is added. This transition detector 30 can thus transmit to the classifier a control message called “inference_enable” message to indicate to said classifier whether the measurement is stabilized or not. These transitions can be detected using a frequency counter as in the Figure 3 , or with a “Leaky Integrate and Fire” neuron implemented in an analog or digital way. Note that on this figure 6 , the processing channels C 1 to CN are transparent to the processing channel C 0 so as to indicate that these channels C 1 to CN are in standby.

[0057] By observing that the channels are correlated to each other, it is possible for the odor sensor to operate in a low-power mode with at least one channel turned on to detect the presence of a gas with the transition detector described above (see the figure 6). When a gas is detected, the other channels light up to allow classification.

[0058] This architecture is obviously particularly relevant if the signal processing is done by "machine learning", especially when coupled with a neuromorphic pulse processing circuit. In particular, it applies well to matrix or multi-channel sensors, which are all parallel inputs for the neuromorphic circuit. The case of the electronic nose is the most interesting target case since the very principle of the sensor is based on a classification, sometimes carried out with "machine learning".

[0059] This invention is part of the low-power systems framework, which implies particular architectures (no ADC, no frequency counter). However, the invention also applies to systems that are not low-power.

Claims

1. A resonant MEMS sensor adapted to generate a pulse output signal (SImp) from a signal of interest (SInt), said signal of interest (SInt) being a signal having a frequency oscillating around a carrier frequency, said MEMS sensor (10) comprising at least one processing channel (C) for processing the signal of interest (SInt), each processing channel (C) containing: - a demodulation unit (101) for demodulating the signal of interest (SInt) to form a demodulated signal (SDem), said demodulation unit (101) comprising a frequency mixer (1010) between said signal of interest (SInt) and a reference signal (Vdemod), said demodulated signal (SDem) having a low-frequency component and a high-frequency component; - a filtration unit (102) for filtering the demodulated signal (SDem) to form a filtered signal (SFil), the filtration unit (12) being adapted to allow the low-frequency component of the demodulated signal (SDem) to pass through; - a comparison unit (103) for comparing the filtered signal (SFil) with a fixed threshold signal (Vth) to form a comparison signal (SComp), said comparison signal (SComp) comprising rising edges and falling edges; - a detection unit (104) for detecting rising edges, each rising edge corresponding to a pulse of the output signal (SImp).

2. The MEMS sensor according to claim 1, said sensor (10) being adapted to generate the signal of interest (SInt) from a variation in a studied physical feature.

3. The MEMS sensor according to claim 2, the studied physical feature from which the signal of interest (SInt) is generated being selected from a group of physical features comprising at least: - a gas; - a mass; - an acceleration.

4. The MEMS sensor according to any one of claims 1 to 3, said sensor (10) comprising at least two processing channels (C0, ..., Ci, ..., CN), each processing channel having a CMUT transducer (100) determined in order to generate a pulse output signal (SImp0, ..., SImpi, ..., SImpN) specific to said processing channel (C0, ..., Ci, ..., CN).

5. The MEMS sensor according to claim 4, wherein each processing channel (C0, ..., Ci, ..., CN) is coupled, at its output, to a classifier, said classifier being adapted to classify the studied physical feature.

6. The MEMS sensor according to claim 5, wherein the classifier is adapted to process digital data and wherein each processing channel (C0, ..., Ci, ..., CN) comprises a conversion unit (Bc0, ..., Bci, ..., BcN) for converting pulses to digital data.

7. The MEMS sensor according to any one of claims 5 or 6, wherein the classifier is a pulse neural network.

8. The MEMS sensor according to any one of claims 5 to 7, wherein said sensor comprises a transition detector (30), said transition detector (30) being adapted to set all or some of the processing channels (C0, ..., Ci, ..., CN) to standby if the studied physical feature does not vary over a certain period.

9. The MEMS sensor according to any one of claims 1 to 8, wherein said MEMS sensor comprises a calibration unit (105) for calibrating the reference signal (Vdemod) in the demodulation unit (101).