CN112163358B - Method and device for realizing reserve pool computing hardware based on coupling MEMS resonator - Google Patents
Method and device for realizing reserve pool computing hardware based on coupling MEMS resonator Download PDFInfo
- Publication number
- CN112163358B CN112163358B CN202011069927.2A CN202011069927A CN112163358B CN 112163358 B CN112163358 B CN 112163358B CN 202011069927 A CN202011069927 A CN 202011069927A CN 112163358 B CN112163358 B CN 112163358B
- Authority
- CN
- China
- Prior art keywords
- signal
- output signal
- resonator
- mems resonator
- coupled
- Prior art date
- Legal status (The legal status is an assumption and is not a legal conclusion. Google has not performed a legal analysis and makes no representation as to the accuracy of the status listed.)
- Active
Links
Classifications
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06F—ELECTRIC DIGITAL DATA PROCESSING
- G06F30/00—Computer-aided design [CAD]
- G06F30/20—Design optimisation, verification or simulation
- G06F30/23—Design optimisation, verification or simulation using finite element methods [FEM] or finite difference methods [FDM]
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06F—ELECTRIC DIGITAL DATA PROCESSING
- G06F17/00—Digital computing or data processing equipment or methods, specially adapted for specific functions
- G06F17/10—Complex mathematical operations
- G06F17/11—Complex mathematical operations for solving equations, e.g. nonlinear equations, general mathematical optimization problems
- G06F17/12—Simultaneous equations, e.g. systems of linear equations
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06F—ELECTRIC DIGITAL DATA PROCESSING
- G06F30/00—Computer-aided design [CAD]
- G06F30/20—Design optimisation, verification or simulation
- G06F30/27—Design optimisation, verification or simulation using machine learning, e.g. artificial intelligence, neural networks, support vector machines [SVM] or training a model
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06F—ELECTRIC DIGITAL DATA PROCESSING
- G06F2111/00—Details relating to CAD techniques
- G06F2111/10—Numerical modelling
Landscapes
- Engineering & Computer Science (AREA)
- Physics & Mathematics (AREA)
- Theoretical Computer Science (AREA)
- General Physics & Mathematics (AREA)
- Mathematical Physics (AREA)
- General Engineering & Computer Science (AREA)
- Evolutionary Computation (AREA)
- Pure & Applied Mathematics (AREA)
- Computational Mathematics (AREA)
- Mathematical Analysis (AREA)
- Mathematical Optimization (AREA)
- Computer Hardware Design (AREA)
- Data Mining & Analysis (AREA)
- Geometry (AREA)
- Software Systems (AREA)
- Medical Informatics (AREA)
- Computer Vision & Pattern Recognition (AREA)
- Operations Research (AREA)
- Algebra (AREA)
- Artificial Intelligence (AREA)
- Databases & Information Systems (AREA)
- Piezo-Electric Or Mechanical Vibrators, Or Delay Or Filter Circuits (AREA)
- Measurement Of Resistance Or Impedance (AREA)
Abstract
A method and a device for realizing reservoir computing hardware based on a coupled MEMS resonator, the method comprises the following steps: preprocessing the time sequence signal to be detected so that the time sequence signal to be detected corresponds to a virtual node coupled with the MEMS resonator; designing a nonlinear vibration equation of the coupled MEMS resonator, and regulating and controlling the coupled MEMS resonator to a preset nonlinear working point according to the equation; respectively detecting two signal testing ends of the MEMS coupled resonator to obtain a first output signal and a second output signal corresponding to the signal to be tested corresponding to each moment, wherein the output signals corresponding to the signal to be tested at the current moment are fed back to the virtual nodes corresponding to the next moment through a bidirectional delay feedback loop; and carrying out regression training on the first output signal and the second output signal corresponding to the preset target value and the signal to be detected corresponding to each time, and obtaining a weight coefficient required by reservoir calculation. The method enhances the data mapping dimension and memory performance and provides richer nonlinear characteristics for reservoir calculation.
Description
Technical Field
The disclosure relates to the field of neural network computation, and is characterized by relating to a method and a device for realizing reserve pool computing hardware based on a coupled MEMS resonator.
Background
Reservoir computation (Reservoir Computing, RC) is a neural network algorithm model modified on the basis of a recurrent neural network (Recurrent Neural Network, RNN) and is generally composed of an input layer, a reservoir consisting of a large number of nonlinear nodes randomly interconnected, and an output layer. During training, the input connection weight and the internal connection weight are randomly generated and kept unchanged, and only the output connection weight is required to be trained. RC has wide application in many scenarios due to its simple training and superior performance in time series signal prediction, speech recognition and classification tasks.
The coupled Micro-electromechanical system (Micro-electromechanical Systems, MEMS for short) resonator is composed of two resonators which are manufactured by Micro-and nano-processing technology and are connected through a mechanical or electrostatic coupling structure, and mainly works by utilizing the mechanical resonance characteristics of the internal structure of the resonator and the related mechanism of the vibration energy dissipation and transfer between modes. Besides structural nonlinearity and material nonlinearity, the coupled MEMS resonator can introduce more nonlinearity phenomena due to the coupling of the internal mode and the physical field, so that the nonlinear dynamics of the coupled MEMS resonator is richer. Therefore, by utilizing the rich nonlinear dynamics characteristics of the coupled MEMS resonator, the reserve tank calculated by the coupled MEMS resonator can not only save calculation resources and improve processing speed and calculation capacity, but also can meet the actual requirements of sensing, storage and calculation integration in the scene of the Internet of things.
At present, related technology discloses a storage pool computing hardware implementation method based on MEMS resonators, in the method, a storage pool computing system consists of a single MEMS resonator and a set of time delay feedback circuits, classification capability of simple voice signals is preliminarily realized through tests on Parity check (Parity benchmark) and TI-46 isolated voice digital data sets, and feasibility of the hardware implementation scheme is verified. However, the method does not optimize the design of the resonator, so that the power consumption of a circuit is too high in practical application, and the method does not perform optimization processing such as feature extraction and the like on a voice signal to be tested, so that the precision is low, the test precision of the TI-46 isolated voice digital data set is preferably 78+/-2%, and more complex tasks cannot be met.
Disclosure of Invention
First, the technical problem to be solved
In view of the above technical problems, the present disclosure proposes a method and an apparatus for implementing reservoir computing hardware based on a coupled MEMS resonator, for at least partially solving the technical problems described above.
(II) technical scheme
According to a first aspect of the present disclosure, there is provided a method for implementing pool computing hardware based on coupled MEMS resonators, comprising: preprocessing a time sequence signal to be detected so that the signal to be detected corresponding to each moment in the time sequence signal to be detected corresponds to the dimension of a virtual node of the coupled MEMS resonator one by one; designing a nonlinear vibration equation of the coupled MEMS resonator:
wherein m is the equivalent mass of the resonator; gamma is damping ratio, and is determined by parameters such as quality factor of the system; k is a linear mechanical stiffness coefficient, and is determined by the materials, the size and the like of the coupled MEMS resonator; k (k) 3 Is a nonlinear mechanical stiffness coefficient; k (k) c Is a linear coupling stiffness coefficient; k (k) c3 Is a nonlinear coupling stiffness coefficient; f is the amplitude of the driving signal, ω is the eigen-resonance frequency of the coupled MEMS resonator, x is the amplitude, x 1 Amplitude, x of one end beam of double-end clamped beam resonator 2 Is the amplitude of the supporting beam at the other end of the double-end clamped beam resonator,to determine the first derivative of the amplitude, +.>To obtain the second derivative of the amplitude; regulating and controlling the coupled MEMS resonator to a preset nonlinear working point according to the nonlinear vibration equation to realize nonlinear response to the time sequence signal to be tested; detecting two signal testing ends of the MEMS coupling resonator respectively to obtain a first output signal and a second output signal corresponding to the signals to be tested corresponding to each moment, wherein the first output signal and the second output signal corresponding to the signals to be tested corresponding to the current moment are fed back to a virtual node corresponding to the signals to be tested corresponding to the next moment in a bidirectional time delay feedback mode for the signals to be tested corresponding to the current moment; will preset the targetAnd carrying out regression training on the first output signal and the second output signal corresponding to the signals to be detected, the values of which correspond to the moments, and obtaining the weight coefficient required by the calculation of the reserve pool.
Optionally, the preprocessing the signal to be sequenced includes: aiming at a signal to be detected corresponding to a certain moment, extracting the characteristics of the signal to be detected corresponding to the moment to obtain a first signal; randomly generating a mask sequence of a signal to be detected corresponding to the moment, and multiplying the mask sequence with the first signal to obtain a second signal; the second signal is converted from a digital signal to an analog signal.
Optionally, the converting the second signal from a digital signal to an analog signal further includes: and carrying out amplitude modulation on the converted analog signal and the excitation signal, and inputting the analog signal and the excitation signal into the coupling MEMS resonator.
Optionally, the detecting the two signal testing ends of the MEMS coupled resonator includes: carrying out primary amplification on the current signal output by the test end through the group-crossing amplifier, and converting the current signal into a voltage signal; performing secondary amplification on the voltage signal, and inputting the signal subjected to secondary amplification into a band-pass filter to acquire a specific frequency waveform; demodulating the envelope signal of the waveform with the specific frequency; and converting the digital signal from the analog signal to the digital signal after demodulation to obtain the first output signal and the second output signal.
Optionally, the feeding back, by a bidirectional delay feedback manner, the first output signal and the second output signal corresponding to the signal to be detected corresponding to the current moment to the virtual node corresponding to the signal to be detected corresponding to the next moment includes: multiplying a first output signal corresponding to a signal to be detected corresponding to the current moment by a first feedback intensity value and feeding back the first output signal to a virtual node corresponding to the signal to be detected corresponding to the next moment; and multiplying a second output signal corresponding to the signal to be detected corresponding to the current moment by a second feedback intensity value and feeding back the second output signal to the virtual node corresponding to the signal to be detected corresponding to the next moment.
Optionally, performing regression training on the first output signal and the second output signal corresponding to the signal to be tested corresponding to the preset target value and each time includes: training by adopting a ridge regression algorithm, wherein the ridge regression algorithm is as follows:
W＝y′X T (XX T +λI) -1
wherein W is the matrix corresponding to the weight coefficient, X is the matrix corresponding to the first output signal or the second output signal, X T And the transposed matrix of X, the unit matrix of I, the regularization coefficient of lambda and the target vector corresponding to the preset target value of y'.
Optionally, the calculation of the required weight coefficient by the obtained reserve pool further comprises: acquiring a test time sequence signal; and calculating a reserve pool according to the weight coefficient, and adjusting the related parameters of the coupled MEMS resonator according to a calculation result until the weight coefficient obtained by training according to the preset target value and the first output signal and the second output signal output by the coupled MEMS resonator is an optimal weight coefficient.
Optionally, the related parameter includes at least one of a driving signal frequency, a driving signal amplitude, a feedback intensity value, a signal time interval to be measured, a feedback time, and a virtual node number of the coupled MEMS resonator.
According to a second aspect of the present disclosure there is provided a pool computing device based on a coupled MEMS resonator, comprising: the integrated circuit board is coupled with the MEMS resonator, the training unit and the FPGA chip; the integrated circuit board is integrated with a detection circuit and a bidirectional delay feedback circuit, the coupled MEMS resonator is packaged on the integrated circuit board, wherein the detection circuit is used for detecting two signal testing ends of the MEMS coupled resonator to obtain a first output signal and a second output signal corresponding to signals to be tested corresponding to each moment, and the bidirectional delay feedback circuit is used for feeding back the first output signal and the second output signal corresponding to the signals to be tested corresponding to the current moment to a virtual node of the coupled MEMS resonator corresponding to the signals to be tested corresponding to the next moment; the training unit is used for carrying out regression training on a first output signal and a second output signal which correspond to the to-be-detected signals corresponding to the preset target values at all times to obtain a weight coefficient required by the calculation of the reserve pool; the FPGA chip is used for controlling the detection circuit, the bidirectional delay feedback circuit and classifying and identifying the time sequence signals to be detected according to the weight coefficient; wherein the nonlinear vibration equation of the coupled MEMS resonator:
wherein m is the equivalent mass of the resonator; gamma is damping ratio, and is determined by parameters such as quality factor of the system; k is a linear mechanical stiffness coefficient, and is determined by the materials, the size and the like of the coupled MEMS resonator; k (k) 3 Is a nonlinear mechanical stiffness coefficient; k (k) c Is a linear coupling stiffness coefficient; k (k) c3 Is a nonlinear coupling stiffness coefficient; f is the amplitude of the driving signal, ω is the eigen-resonance frequency of the coupled MEMS resonator, x is the amplitude, x 1 Amplitude, x of one end beam of double-end clamped beam resonator 2 Is the amplitude of the supporting beam at the other end of the double-end clamped beam resonator,to determine the first derivative of the amplitude, +.>To take the second derivative of the amplitude.
Optionally, the detection circuit comprises a group-crossing amplifier, a secondary amplifier, a band-pass filter, an envelope detector and an ADC/DAC conversion module which are sequentially connected.
(III) beneficial effects
The invention provides a method and a device for realizing reservoir computing hardware based on a coupling MEMS resonator, which have the beneficial effects that:
1. according to the method, the coupled MEMS resonator is used as a reserve tank calculation core unit, device parameters required by reserve tank calculation are met through finite element simulation and optimization of design, and a nonlinear vibration equation of the MEMS resonator is further coupled according to mechanical resonance characteristics of a weakly coupled internal structure and a related mechanism of vibration energy dissipation and transfer among modes, so that nonlinear vibration states of the coupled MEMS resonator are regulated and controlled, dynamic mapping capacity of the reserve tank to input data can be improved by utilizing rich nonlinear dynamics characteristics of the coupled MEMS resonator, and further the capacity of the reserve tank calculation for classifying and identifying time sequence signals is improved.
2. According to the method, a bidirectional delay feedback control loop is adopted, and the test result of the nonlinear working state of the coupled MEMS resonator corresponding to the signal to be tested corresponding to the current moment is fed back to the virtual node corresponding to the signal to be tested corresponding to the next moment in a bidirectional delay feedback mode, so that the design of multiple ports of the coupled MEMS resonator is fully utilized, the response generated by the earlier input signal can be input into the reserve pool, the dynamics of the current reserve pool is affected to different degrees by the earlier input signal, and the memory capacity of the reserve pool is improved.
3. The method combines the advantages of MEMS devices in small size, low power consumption, easy integration and batch production, and the advantages of a reserve pool calculation training mode in tasks such as time sequence signal prediction, mode recognition and voice classification, so as to meet the requirements of sensing and calculating integration in the scene of the Internet of things. And the compatibility problem of the device volume, the power consumption and the accurate classification capability of the traditional hardware implementation method of the reserve pool calculation is solved.
Drawings
For a more complete understanding of the present disclosure and the advantages thereof, reference is now made to the following descriptions taken in conjunction with the accompanying drawings, which are incorporated in and constitute a part of this specification, illustrate embodiments consistent with the disclosure and together with the description serve to explain the principles of the disclosure. Wherein:
FIG. 1 schematically illustrates a flow chart of an implementation method of pool computing hardware based on coupled MEMS resonators, according to an embodiment of the disclosure;
FIG. 2 schematically illustrates a coupled MEMS resonator bi-directional delay feedback circuit and detection circuit diagram of an embodiment of the present disclosure;
FIG. 3 is a schematic diagram of a lumped model of a coupled MEMS resonator in accordance with an embodiment of the disclosure;
FIG. 4 is a schematic diagram of a coupled MEMS resonator according to an embodiment of the disclosure;
FIG. 5 is a schematic diagram of a pool computing device based on coupled MEMS resonators according to an embodiment of the present disclosure;
fig. 6A is a schematic illustration of an original signal speech signal graph and a signal graph after feature extraction via a mel-down spectral coefficient model, according to an embodiment of the present disclosure.
Fig. 6B schematically illustrates a paul signal diagram according to an embodiment of the disclosure;
fig. 6C schematically illustrates an output signal diagram according to an embodiment of the present disclosure.
Detailed Description
Hereinafter, embodiments of the present disclosure will be described with reference to the accompanying drawings. It should be understood that the description is only exemplary and is not intended to limit the scope of the present disclosure. In the following detailed description, for purposes of explanation, numerous specific details are set forth in order to provide a thorough understanding of the embodiments of the present disclosure. It may be evident, however, that one or more embodiments may be practiced without these specific details. In addition, in the following description, descriptions of well-known structures and techniques are omitted so as not to unnecessarily obscure the concepts of the present disclosure.
The terminology used herein is for the purpose of describing particular embodiments only and is not intended to be limiting of the invention. The terms "comprises," "comprising," and/or the like, as used herein, specify the presence of stated features, steps, operations, and/or components, but do not preclude the presence or addition of one or more other features, steps, operations, or components.
Fig. 1 schematically illustrates a flow chart of an implementation method of pool computing hardware based on coupled MEMS resonators, according to an embodiment of the disclosure.
As shown in fig. 1, the method may include, for example, operations S101 to S106.
S101, preprocessing the time sequence signals to be detected so that the signals to be detected corresponding to each moment in the time sequence signals to be detected correspond to the dimensions of the virtual nodes of the coupled MEMS resonator one by one.
In operation S101, through a preprocessing procedure, a mutual mapping relationship of data in the present time interval may be ensured.
Referring to fig. 2, the specific pretreatment process is as follows:
for a signal to be detected u (t) corresponding to a certain moment, extracting features of the signal to be detected u (t) corresponding to the moment to obtain a first signal, wherein the specific technical method is MFCC (mel-frequency coefficient) or a cochlea model.
And randomly generating a mask sequence m (t) of the signal to be detected corresponding to the moment, and multiplying the mask sequence m (t) with the first signal to obtain a second signal.
The second signal is converted from a digital signal to an analog signal by the DAC module.
Wherein T represents the corresponding time of the signal to be measured, the input signal time interval tau, the virtual node number N, the virtual node time interval theta, the resonator decay time T, the quality factor Q and the resonator eigenfrequency w n The conditions required to be met for realizing the functions are tau=nθ, θ≡t=2q/w n 。
S102, the coupled MEMS resonator performs nonlinear response on the preprocessed input signal.
With continued reference to fig. 2, before the analog signal obtained after the preprocessing is input into the coupled MEMS resonator, the converted analog signal and the excitation signal are subjected to amplitude modulation, that is, the modulated signal is used as the input of the coupled MEMS resonator.
The coupled MEMS resonator is regulated and controlled to a preset nonlinear working point by using the detection circuit and the bidirectional delay feedback circuit shown in fig. 2, so that nonlinear response of a signal to be sequenced is realized. The implementation of this process depends on the choice of the nonlinear operating point of the resonator, first by theoretical derivation and modeling analysis of the forced vibration equation of the double-ended clamped beam, the basic characteristics of the nonlinear response are studied, then, the coupled resonator forced vibration equation under ideal conditions is derived, as in fig. 3, which is a lumped model of the coupled resonator, assuming that the mass m of the two resonators is the same as the linear mechanical stiffness coefficient k, and the change in displacement can be obtained by solving the differential equation:
the formula (1) is a forced vibration equation of the double-end clamped beam resonator, and the formula (2) is a forced nonlinear vibration equation of the coupled MEMS resonator. Wherein m is the equivalent mass of the resonator; gamma is damping ratio, and is determined by parameters such as quality factor of the system; k is a linear mechanical stiffness coefficient, and is determined by the materials, the size and the like of the coupled MEMS resonator; k (k) 3 Is a nonlinear mechanical stiffness coefficient; k (k) c Is a linear coupling stiffness coefficient; k (k) c3 Is a nonlinear coupling stiffness coefficient; f is the amplitude of the driving signal, ω is the eigen-resonance frequency of the coupled MEMS resonator, x is the amplitude, x 1 Amplitude, x of one end beam of double-end clamped beam resonator 2 Is the amplitude of the supporting beam at the other end of the double-end clamped beam resonator,to determine the first derivative of the amplitude, +.>To take the second derivative of the amplitude.
The constraint condition of the formula is that the resonator is driven by distributed load, the related parameter value under the action of the mechanical force is deduced, and the linear mechanical stiffness coefficient k= (32 Edω) 3 )/(L 3 ) Nonlinear mechanical stiffnessE is young's modulus of monocrystalline silicon, and the value e=170×10 9 P a D is the thickness of the resonator beam, W is the width of the beam, and L is the length of the beam.
The design of the nonlinear vibration equation (equation 2) of the coupled MEMS resonator described above can be accomplished by finite element simulation design (as shown in fig. 4). The method for designing the nonlinear vibration equation of the MEMS resonator is due to the new addition of the coupling coefficient term k c 、k c3 The coupling MEMS resonator can obtain richer nonlinear states relative to two independent double-end clamped beam resonators, which is equivalent to increasing the dimension of a neuron network mapping unit, and greatly improves the performance of a reserve pool computing system.
S103, respectively detecting two signal testing ends of the MEMS coupled resonator to obtain a first output signal and a second output signal corresponding to the signals to be tested, which correspond to each moment.
With continued reference to fig. 2, parallel plate capacitance detection is implemented on two signal testing terminals of the coupled resonator through the detection circuit designed in fig. 2, so as to convert a current signal into a voltage signal and amplify a weak signal, thereby completing sampling and conversion of an output signal.
Specifically, first, a current signal output from a test terminal is amplified in one stage by a group-crossing amplifier, and the current signal is converted into a voltage signal. Secondly, carrying out secondary amplification on the voltage signal, and inputting the signal subjected to secondary amplification into a band-pass filter to acquire a waveform with a specific frequency, wherein the specific frequency can comprise a driving frequency, a frequency doubling or a frequency quadrupling, for example. Again, the waveform of the specific frequency is demodulated for the envelope signal. Finally, the digital signal conversion from the analog signal to the demodulated signal is carried out to obtain a first output signal x 1 (t) and a second output signal x 2 (t)。
At the time of obtaining the first output signal x 1 (t) and a second output signal x 2 And (t) feeding back the first output signal and the second output signal corresponding to the signal to be detected corresponding to the current moment to the virtual node corresponding to the signal to be detected corresponding to the next moment in a bidirectional delay feedback mode. Specifically, a first output signal corresponding to a signal to be detected corresponding to the current moment is multiplied by a first feedback intensity value and then fed back to a virtual node corresponding to the signal to be detected corresponding to the next moment; and multiplying a second output signal corresponding to the signal to be detected corresponding to the current moment by a second feedback intensity value and feeding back the second output signal to the virtual node corresponding to the signal to be detected corresponding to the next moment. At the virtual node corresponding to the next moment, the signal fed back by the bidirectional delay and the signal to be detected corresponding to the next moment are fed backThe signals are added, amplitude-modulated with the excitation signal and then input into the coupled MEMS resonator. Repeating the feedback operation for the signals to be tested (time sequence signals to be tested) arranged according to the time sequence to obtain a first output signal x corresponding to the signals to be tested at each moment 1 (t) and a second output signal x 2 (t)。
Wherein the first output signal x 1 (t) and a second output signal x 2 (t) multiplying a feedback intensity value respectively to realize certain attenuation memory efficiency, feeding back the feedback to a corresponding node of a signal to be measured at the next moment and adding the feedback to the signal to be measured at the next moment, realizing richer dynamic influence between signal contexts, and improving the memory capacity of the calculation of the reserve pool, wherein the specific feedback form is as follows:
V d ＝F(u(t)m(t)+α 1 x 1 (t-τ 1 )+α 2 x 2 (t-τ 2 ))cosωt (3)
wherein V is d For the actual driving signal of the resonator, the amplitude V of the excitation signal ac The current output signal u (t), the mask sequence m (t) and the two feedback signals at the last moment are determined together; alpha 1 、α 2 Determining the memory depth between contexts for the two-way time delay feedback intensity; τ 1 、τ 2 For the two-way delay feedback time, the memory capacity of the system is determined.
S104, carrying out regression training on the first output signal and the second output signal corresponding to the to-be-detected signal corresponding to the preset target value and each time, and obtaining the weight coefficient required by the calculation of the reserve pool.
And (3) carrying out regression training on the first output signal and the second output signal corresponding to the signal to be tested corresponding to each moment obtained in the operation S103 and a preset target value. Regression training may employ, for example, a ridge regression algorithm:
W＝y′X T (XX T +λI) -1
wherein W is a matrix corresponding to the weight coefficient, X is a matrix corresponding to the first output signal or the second output signal, X T Is the transposed matrix of X, y' is the target corresponding to the preset target valueThe standard vector, I is the identity matrix, lambda is the regularization coefficient, is used for preventing data from overflowing, and generally takes a value of 0.1. The loss function is utilized to realize the training of the output weight, the robustness of the system is enhanced, and the classification or prediction function of the time sequence signals is better realized.
S105, testing the obtained weight coefficient, and judging whether the weight coefficient is the optimal weight coefficient according to a test result.
After the training phase is finished, the weight coefficient obtained in the obtained operation S104 needs to be tested, which specifically includes: and acquiring a test time sequence signal. And carrying out reserve pool calculation on the test time sequence signals according to the weight coefficients to obtain test results. If the test result meets the requirement of practical application, the weight coefficient at the moment is considered to be the optimal weight coefficient, and training is finished. The specific criteria for judging whether the weight coefficient is the optimal weight coefficient are as follows: and carrying out the same processing on the test set, carrying out normalized root mean square error calculation on the predicted value and the true value obtained by multiplying the test set by the weight coefficient, judging the system performance quality by the quantization performance index, and determining whether the system performance quality is the optimal weight coefficient. The performance index of the quantitative classification accuracy of the Winner-Takes-All algorithm can be added, and the system performance can be evaluated more intuitively.
For example, the signal to be tested is a voice signal, so that the purpose is to identify and classify the voice signal, and if the identification and classification are accurate and clear, the test result can be considered to meet the requirements of practical application. If the test result does not meet the requirement of the practical application, operation S106 is executed.
And S106, adjusting the related parameters of the coupled MEMS resonator according to the test result.
In operation S106, the adjustment process is performed until the weight coefficient obtained by training according to the preset target value and the first output signal and the second output signal output by the coupled MEMS resonator is an optimal weight coefficient, where the relevant parameter may include at least one of a driving signal frequency, a driving signal amplitude, a feedback intensity value, a signal time interval to be measured, a feedback time, and a virtual node number of the coupled MEMS resonator. That is, the above operations are repeatedly performed after the parameters are adjusted, and the output signals are adjusted by adjusting the parameters, so as to finally obtain the output signals corresponding to the optimal weight coefficients.
The training stage uses the upper computer and the corresponding data set to obtain the weight coefficient. The test stage copies the written control program to a field programmable gate array (Field Programmable Gate Array, abbreviated as FPGA) system board to realize overall program control, and the test stage comprises the functions of preprocessing operation of signals to be tested, downsampling of output signals, bidirectional feedback control, storage, result display and the like.
The embodiment of the disclosure also provides a reservoir computing device based on the coupled MEMS resonator, referring to fig. 2 and 5, which may include an integrated circuit board, the coupled MEMS resonator, a training unit, and an FPGA chip.
The integrated circuit board is integrated with a detection circuit, a bidirectional delay feedback circuit and a DAC module, and the coupled MEMS resonator is vacuum packaged according to the required Q value and is arranged on the integrated circuit board to realize the integration of the MEMS chip and the CMOS circuit. The detection circuit comprises a transimpedance amplifier, a secondary amplification circuit, a band-pass filter circuit, an envelope detector circuit, an ADC module, a power module and the like, and is used for detecting two signal testing ends of the MEMS coupled resonator to obtain a first output signal and a second output signal corresponding to signals to be detected, which correspond to each moment. The bidirectional delay feedback circuit is used for feeding back a first output signal and a second output signal corresponding to a signal to be detected corresponding to the current moment to a virtual node of the coupled MEMS resonator corresponding to the signal to be detected corresponding to the next moment.
The training unit is used for carrying out regression training on a first output signal and a second output signal which correspond to the to-be-detected signals corresponding to the preset target values at all times to obtain an optimal weight coefficient required by the calculation of the reserve pool;
and the FPGA chip is used for controlling the detection circuit and the bidirectional delay feedback circuit and classifying and identifying the time sequence signals to be detected according to the optimal weight coefficient.
The details of the device embodiment are not fully the same as those of the method embodiment, please refer to the method embodiment, and are not repeated here.
In order to facilitate better understanding of the technical solution of the present embodiment, the present invention further describes the above device and solution by taking recognition of a simple voice signal of 0-9 as an example.
The experimental parameter value selected in this example is V ac ＝220mV、V dc ＝30V、f n ＝202KHz、Q＝1000、T＝1.58mS、N＝400。
The specific process is as follows:
operation a: the device arrangement is completed and the parameters are adjusted. And finishing the preprocessing of the original voice signal according to the given parameters and adjusting the upper computer program of the digital feedback control loop.
The pretreatment comprises the following steps: feature extraction is implemented by constructing Mel-frequency cepstrum coefficient (Mel-Frequency Cepstrum, MFCC) model for the voice signal, the extraction result is shown in fig. 6A, and the process of multiplying the signal to be measured with a random mask in operation S101.
Operation b: the digital-to-analog conversion is realized through a DAC module by utilizing a control program in the FPGA, and the digital-to-analog conversion is coupled with an excitation signal after amplitude modulation; regulating and controlling the coupled MEMS resonator to a specified nonlinear working state by using a driving circuit shown in FIG. 2, wherein a signal to be tested firstly passes through a single resonator at a driving end to finish nonlinear mapping once; and meanwhile, the mapping is realized by transmitting the mapping data to the coupling end. The two detection ends output corresponding time domain amplitude variation signals at the same time.
Operation c: the MEMS interface circuit is used to collect two real-time outputs of operation B, so as to complete the function in operation S103, that is, performing the group-crossing first-stage amplification, the second-stage method, the band-pass filtering, and the envelope signal demodulation in sequence, wherein the signal diagram is shown in fig. 6B. Analog-to-digital conversion and downsampling output x by using FPGA to control ADC module 1 (t)、x 2 (t); fig. 6C shows the actual signal extracted by the module.
Operation d: the FPGA control program realizes time delay feedback of the real-time signals and stores the time delay feedback.
And (e) operation step e: after the training step is completed, an optimized output weight coefficient matrix is obtained by utilizing the constructed ridge regression training model; and then respectively selecting the classification precision of the system constructed by the test of the voice signals in the training set and the test set, and adjusting the relevant parameters of the model according to the result to obtain the optimal performance.
Operation f: and copying the complete program and the original voice signal to be tested to the FPGA according to the parameters such as the optimal weight coefficient and the like, obtaining the identification result after operation, and displaying the identification result on a display screen of the FPGA.
Thus, operations a through f implement the simple speech signal classification recognition function based on the method and apparatus of the present embodiment.
It will be understood by those skilled in the art that while the present disclosure has been shown and described with reference to particular exemplary embodiments thereof, it will be understood by those skilled in the art that various changes in form and details may be made therein without departing from the spirit and scope of the present disclosure as defined by the appended claims and their equivalents. The scope of the disclosure should, therefore, not be limited to the above-described embodiments, but should be determined not only by the following claims, but also by the equivalents of the following claims.
Claims (10)
1. A method for implementing pool computing hardware based on coupled MEMS resonators, comprising:
preprocessing a time sequence signal to be detected so that the signal to be detected corresponding to each moment in the time sequence signal to be detected corresponds to the dimension of a virtual node of the coupled MEMS resonator one by one;
designing a nonlinear vibration equation of the coupled MEMS resonator:
wherein m is the equivalent mass of the resonator; gamma is the damping ratio, determined by parameters including the figure of merit of the system; k is a linear mechanical stiffness coefficient, determined by the material and dimensions of the coupled MEMS resonator; k (k) 3 Is a nonlinear mechanical stiffness coefficient; k (k) c Is a linear coupling stiffness coefficient; k (k) c3 Is a non-linearA coefficient of sexual coupling stiffness; f is the amplitude of the driving signal, ω is the eigen-resonance frequency of the coupled MEMS resonator, x is the amplitude, x 1 Amplitude, x of one end beam of double-end clamped beam resonator 2 Is the amplitude of the supporting beam at the other end of the double-end clamped beam resonator,to determine the first derivative of the amplitude, +.>To obtain the second derivative of the amplitude;
regulating and controlling the coupled MEMS resonator to a preset nonlinear working point according to the nonlinear vibration equation to realize nonlinear response to the time sequence signal to be tested;
detecting two signal testing ends of the MEMS coupling resonator respectively to obtain a first output signal and a second output signal corresponding to the signals to be tested corresponding to each moment, wherein the first output signal and the second output signal corresponding to the signals to be tested corresponding to the current moment are fed back to a virtual node corresponding to the signals to be tested corresponding to the next moment in a bidirectional time delay feedback mode for the signals to be tested corresponding to the current moment;
and carrying out regression training on the first output signal and the second output signal corresponding to the preset target value and the signal to be detected corresponding to each time, and obtaining a weight coefficient required by reservoir calculation.
2. The method of claim 1, wherein the preprocessing the signal to be sequenced comprises:
aiming at a signal to be detected corresponding to a certain moment, extracting the characteristics of the signal to be detected corresponding to the moment to obtain a first signal;
randomly generating a mask sequence of a signal to be detected corresponding to the moment, and multiplying the mask sequence with the first signal to obtain a second signal;
the second signal is converted from a digital signal to an analog signal.
3. The method of claim 2, wherein the converting the second signal from a digital signal to an analog signal further comprises:
and carrying out amplitude modulation on the converted analog signal and the excitation signal, and inputting the analog signal and the excitation signal into the coupling MEMS resonator.
4. The method for implementing the pool computing hardware based on the coupled MEMS resonator according to claim 2, wherein the detecting the two signal test terminals of the MEMS coupled resonator respectively includes:
carrying out primary amplification on the current signal output by the test end through the group-crossing amplifier, and converting the current signal into a voltage signal;
performing secondary amplification on the voltage signal, and inputting the signal subjected to secondary amplification into a band-pass filter to acquire a specific frequency waveform;
demodulating the envelope signal of the waveform with the specific frequency;
and converting the digital signal from the analog signal to the digital signal after demodulation to obtain the first output signal and the second output signal.
5. The method for implementing the reservoir computing hardware based on the coupled MEMS resonator according to claim 1, wherein the feeding back the first output signal and the second output signal corresponding to the signal to be tested corresponding to the current moment to the virtual node corresponding to the signal to be tested corresponding to the next moment by using a bidirectional delay feedback method includes:
multiplying a first output signal corresponding to a signal to be detected corresponding to the current moment by a first feedback intensity value and feeding back the first output signal to a virtual node corresponding to the signal to be detected corresponding to the next moment;
and multiplying a second output signal corresponding to the signal to be detected corresponding to the current moment by a second feedback intensity value and feeding back the second output signal to the virtual node corresponding to the signal to be detected corresponding to the next moment.
6. The method for implementing the reservoir computing hardware based on the coupled MEMS resonator according to claim 1, wherein performing regression training on the first output signal and the second output signal corresponding to the signal to be measured corresponding to each time of the preset target value includes:
training by adopting a ridge regression algorithm, wherein the ridge regression algorithm is as follows:
W＝y′X T (XX T +λI) -1
wherein W is the matrix corresponding to the weight coefficient, X is the matrix corresponding to the first output signal or the second output signal, X T And the transposed matrix of X, the unit matrix of I, the regularization coefficient of lambda and the target vector corresponding to the preset target value of y'.
7. The method of claim 1, wherein the obtaining the weight coefficients required for the reservoir calculation further comprises:
acquiring a test time sequence signal;
and calculating a reserve pool according to the weight coefficient, and adjusting the related parameters of the coupled MEMS resonator according to a calculation result until the weight coefficient obtained by training according to the preset target value and the first output signal and the second output signal output by the coupled MEMS resonator is an optimal weight coefficient.
8. The method of claim 7, wherein the related parameters include at least one of a driving signal frequency, a driving signal amplitude, a feedback intensity value, a signal time interval to be measured, a feedback time, and a number of virtual nodes of the coupled MEMS resonator.
9. A pool computing device based on a coupled MEMS resonator, comprising:
the integrated circuit board is coupled with the MEMS resonator, the training unit and the FPGA chip;
the integrated circuit board is integrated with a detection circuit and a bidirectional delay feedback circuit, the coupled MEMS resonator is packaged on the integrated circuit board, wherein the detection circuit is used for detecting two signal testing ends of the MEMS coupled resonator to obtain a first output signal and a second output signal corresponding to signals to be tested corresponding to each moment, and the bidirectional delay feedback circuit is used for feeding back the first output signal and the second output signal corresponding to the signals to be tested corresponding to the current moment to a virtual node of the coupled MEMS resonator corresponding to the signals to be tested corresponding to the next moment;
the training unit is used for carrying out regression training on a first output signal and a second output signal which correspond to the to-be-detected signals corresponding to the preset target values at all times to obtain a weight coefficient required by the calculation of the reserve pool;
the FPGA chip is used for controlling the detection circuit, the bidirectional delay feedback circuit and classifying and identifying the time sequence signals to be detected according to the weight coefficient;
wherein the nonlinear vibration equation of the coupled MEMS resonator:
wherein m is the equivalent mass of the resonator; gamma is the damping ratio, determined by parameters including the figure of merit of the system; k is a linear mechanical stiffness coefficient, determined by the material and dimensions of the coupled MEMS resonator; k (k) 3 Is a nonlinear mechanical stiffness coefficient; k (k) c Is a linear coupling stiffness coefficient; k (k) c3 Is a nonlinear coupling stiffness coefficient; f is the amplitude of the driving signal, ω is the eigen-resonance frequency of the coupled MEMS resonator, x is the amplitude, x 1 Amplitude, x of one end beam of double-end clamped beam resonator 2 Is the amplitude of the supporting beam at the other end of the double-end clamped beam resonator,to determine the first derivative of the amplitude, +.>To take the second derivative of the amplitude.
10. The coupled MEMS resonator-based reservoir computing device of claim 9, wherein the detection circuit comprises a cross-bank amplifier, a two-stage amplifier, a band pass filter, an envelope detector, and an ADC/DAC conversion module connected in sequence.
Priority Applications (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN202011069927.2A CN112163358B (en) | 2020-09-30 | 2020-09-30 | Method and device for realizing reserve pool computing hardware based on coupling MEMS resonator |
Applications Claiming Priority (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN202011069927.2A CN112163358B (en) | 2020-09-30 | 2020-09-30 | Method and device for realizing reserve pool computing hardware based on coupling MEMS resonator |
Publications (2)
| Publication Number | Publication Date |
|---|---|
| CN112163358A CN112163358A (en) | 2021-01-01 |
| CN112163358B true CN112163358B (en) | 2023-09-26 |
Family
ID=73861135
Family Applications (1)
| Application Number | Title | Priority Date | Filing Date |
|---|---|---|---|
| CN202011069927.2A Active CN112163358B (en) | 2020-09-30 | 2020-09-30 | Method and device for realizing reserve pool computing hardware based on coupling MEMS resonator |
Country Status (1)
| Country | Link |
|---|---|
| CN (1) | CN112163358B (en) |
Families Citing this family (1)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN114781636B (en) * | 2022-06-21 | 2022-09-06 | 太原理工大学 | Integrated photon storage pool computing device |
Citations (2)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN109655674A (en) * | 2019-02-27 | 2019-04-19 | 西北工业大学 | Faint electrostatic field measurement device and method based on weak coupling micromechanical resonator |
| CN109828141A (en) * | 2019-02-27 | 2019-05-31 | 西北工业大学 | Highly sensitive voltage measuring apparatus and measurement method based on weak coupling micromechanical resonator |
Family Cites Families (1)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| FR2929775B1 (en) * | 2008-04-08 | 2011-12-09 | Commissariat Energie Atomique | RESONANT FILTER BASED ON N / MEMS MATRIX |
-
2020
- 2020-09-30 CN CN202011069927.2A patent/CN112163358B/en active Active
Patent Citations (2)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN109655674A (en) * | 2019-02-27 | 2019-04-19 | 西北工业大学 | Faint electrostatic field measurement device and method based on weak coupling micromechanical resonator |
| CN109828141A (en) * | 2019-02-27 | 2019-05-31 | 西北工业大学 | Highly sensitive voltage measuring apparatus and measurement method based on weak coupling micromechanical resonator |
Also Published As
| Publication number | Publication date |
|---|---|
| CN112163358A (en) | 2021-01-01 |
Similar Documents
| Publication | Publication Date | Title |
|---|---|---|
| CN103900610B (en) | MEMS gyro random error Forecasting Methodology based on Lycoperdon polymorphum Vitt wavelet neural network | |
| Cao et al. | Design and test of the MEMS coupled piezoelectric–electromagnetic energy harvester | |
| US9372111B2 (en) | Acoustic sensor | |
| CN112362756A (en) | Concrete structure damage monitoring method and system based on deep learning | |
| CN112163358B (en) | Method and device for realizing reserve pool computing hardware based on coupling MEMS resonator | |
| CN111106783A (en) | Signal manufacturing method, signal manufacturing device, vibration motor and touch screen equipment | |
| Tanaka et al. | A novel mechanical cochlea" Fishbone" with dual sensor/actuator characteristics | |
| CN105892293A (en) | Silicon micro-machined gyroscope digital driving closed loop control system | |
| WO2022000600A1 (en) | Motor parameter test method | |
| CN110631792A (en) | Seismic hybrid test model updating method based on convolutional neural network | |
| CN112906282A (en) | Method and system for identifying working mode parameters extracted by parallel principal component of Sanger neural network | |
| CN115798515A (en) | Transform-based sound scene classification method | |
| Givois et al. | Experimental analysis of nonlinear resonances in piezoelectric plates with geometric nonlinearities | |
| CN102647657A (en) | Monolithic integrated MEMS (Micro-electromechanical Systems) piezoresistive ultrasonic sensor | |
| Iannacci et al. | A novel MEMS-based piezoelectric multi-modal vibration energy harvester concept to power autonomous remote sensing nodes for Internet of Things (IoT) applications | |
| Dion et al. | Reservoir computing in MEMS | |
| McNamara et al. | Novel relationship between shear and energy density at the bounce in nonsingular Bianchi I spacetimes | |
| CN114417529A (en) | Method and device for adjusting equivalent Q value of silicon micro-resonator RC system by mechanical pumping | |
| CN105956294A (en) | Fruit fly algorithm-based structure damage identification method | |
| CN112992182B (en) | Vehicle wind noise level testing system and testing method thereof | |
| CN112710382B (en) | Micro tuning fork resonator vibration sensitivity on-chip test structure and method | |
| Kuntzman et al. | Network modeling of multiple-port, multiple-vibration-mode transducers and resonators | |
| US20230140456A1 (en) | Parameter setting method and control method for reservoir element | |
| Van Drieënhuizen et al. | Integrated micromachined electrostatic true RMS-to-DC converter | |
| Colin et al. | Identification-based approach for electrical coupling compensation in a MEMS gyroscope |
Legal Events
| Date | Code | Title | Description |
|---|---|---|---|
| PB01 | Publication | ||
| PB01 | Publication | ||
| SE01 | Entry into force of request for substantive examination | ||
| SE01 | Entry into force of request for substantive examination | ||
| GR01 | Patent grant | ||
| GR01 | Patent grant |