Method and device for mechanically pumping to adjust equivalent Q value of silicon microresonator RC system

Abstract

A method and device for adjusting the equivalent Q value of a silicon microresonator RC system by mechanical pumping, the method comprising: determining the modal frequency of the resonator and the mechanical pumping signal; performing a verification experiment of dynamically adjusting the equivalent Q value by mechanical pumping; further, based on the verification experiment, performing a complex nonlinear response on the pre-processed signal to be tested by using a coupled MEMS resonator RC system under specified conditions; filtering and collecting the resonator response signal; precise control of the digital delay feedback control program, data detection, downsampling, and logical operations, etc.; training and testing of the test data set. This method uses mechanical pumping to realize energy exchange between different modes of the resonator, and realizes dynamic adjustment of the equivalent Q value of the used mode while ensuring that the system works under strong nonlinear rich conditions (higher vacuum environment). This method not only ensures the nonlinear mapping performance of the reserve pool calculation, but also improves its calculation speed.

Description

Technical Field

[0001] The present invention relates to the field of neural network calculation, and is characterized by a method and a device for dynamically adjusting the equivalent Q value of a silicon microresonator RC system by using a mechanical pump. Background Art

[0002] Reservoir Computing (RC) is a neural network algorithm model improved on the basis of Recurrent Neural Network (RNN). It is usually composed of an input layer, a reservoir composed of a large number of randomly fixed-connected nonlinear nodes, and an output layer. The input connection weights and the internal connection weights of the reservoir are randomly generated and remain unchanged, and only the output connection weights need to be trained. RC has a wide range of applications in many scenarios due to its simple training method and excellent performance in classification tasks such as time series signal prediction and speech recognition, especially driving a research boom in the field of hardware algorithm implementation.

[0003] Micro-electromechanical systems (MEMS) resonators can be used as a typical carrier that meets the hardware implementation of RC algorithms due to their rich nonlinear vibration characteristics. They are generally manufactured through micron and nanometer silicon processing technology and consist of two or more single resonators connected by mechanical or electrostatic coupling structures. They mainly work by using the mechanical resonance characteristics of the internal structure of the resonator and the vibration energy dissipation transfer mechanism between modes. In addition to structural nonlinearity and material nonlinearity, coupled MEMS resonators will introduce more nonlinear phenomena due to the coupling of their internal modes and physical fields, making their nonlinear dynamic characteristics richer.

[0004] Mechanical pumping is a method to achieve energy dissipation transfer between different modes of MEMS resonators. The principle is similar to optical pumping. The energy of a mechanical mode of the resonator is coupled to another mode to adjust the equivalent Q value of the mode. The mechanical mode (also known as the phonon cavity) is usually higher in frequency than the mode that undergoes equivalent Q value adjustment. The choice of mechanical pumping signal for adjusting energy transfer is generally divided into "frequency sum--Stoke" and "frequency difference--anti-Stoke", which respectively achieve the increase and decrease of the equivalent Q of the first mode.

[0005] Therefore, based on the practical application of the silicon microresonator (including coupling devices) reserve pool computing hardware system, the present invention proposes to use mechanical pumping to realize the dynamic adjustment of its inter-modal energy, so that the system can achieve self-adjustment of the equivalent Q value according to different needs, in order to meet the actual needs of "perception + computing integration" in the Internet of Things scenario and flexibly adjust the working state according to different task requirements.

[0006] At present, the relevant technology discloses a hardware implementation method for reserve pool calculation based on MEMS "coupled" resonators. The method is composed of a single nonlinear MEMS resonator with a unidirectional time-delay feedback circuit. The regression and classification capabilities are preliminarily verified by testing on the parity check (Paritybenchmark) and TI-46 isolated voice digital data sets, proving the feasibility of the resonator reserve pool calculation hardware implementation scheme. However, this method does not optimize the resonator system design, the selection of nonlinear working points, etc., resulting in excessive circuit power consumption in actual application and low classification accuracy. The test accuracy of the TI-46 isolated voice digital data set is at best 78±2%, which cannot meet more complex tasks; and according to different types of prediction tasks, the overall parameters can only be adjusted by recalibrating the nonlinear working point, which cannot meet the needs of actual sensor-level application scenarios, and can only be used as a laboratory research solution. Summary of the invention

[0007] 1. Technical issues to be resolved

[0008] In view of the above technical problems, the present disclosure proposes a method for dynamically adjusting the equivalent Q value of a silicon microresonator RC system based on mechanical pumping, which is used to at least partially solve the above technical problems.

[0009] (II) Technical solution

[0010] According to a first aspect of the present disclosure, a method for dynamically adjusting the equivalent Q value of a silicon microresonator RC system based on mechanical pumping is provided, comprising: preprocessing a timing signal to be measured so that the signal to be measured corresponding to each moment in the timing signal to be measured corresponds one-to-one to the virtual node dimension of the coupled MEMS resonator; and designing a nonlinear vibration equation of the coupled MEMS resonator:

[0011]

[0012] Where m is the equivalent mass of the resonator; γ is the damping term, which is determined by the quality factor and other parameters of the system; k is the linear mechanical stiffness coefficient, which is determined by the material and size of the coupled MEMS resonator; k3 is the nonlinear mechanical stiffness coefficient; k c is the linear coupling stiffness coefficient; k c3 is the nonlinear coupling stiffness coefficient; F drive is the driving signal amplitude, ω d is the resonant frequency of the driving signal coupled to the MEMS resonator, F pump is the driving signal amplitude, ω pis the resonant frequency of the driving signal of the coupled MEMS resonator, x1 is the amplitude of one end beam of the double-end clamped beam coupled resonator, x2 is the amplitude of the other end beam of the double-end clamped beam coupled resonator, To find the first derivative of the amplitude, The second-order derivative of the amplitude is calculated; the coupled MEMS resonator is adjusted to a preset nonlinear operating point according to the nonlinear vibration equation, and the equivalent Q value of the mechanical pump signal dynamic adjustment system obtained in the preliminary experiment is used to adjust the optimal operating point of the timing signal to be tested of the different task types, and the nonlinear dynamic response of RC under different tasks is obtained; the signal test end of the MEMS coupled resonator is detected to obtain the output signal corresponding to the signal to be tested at each moment, wherein, for the signal to be tested corresponding to the current moment, the output signal 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 through a one-way time delay feedback method; the preset target value and the output signal corresponding to the signal to be tested corresponding to each moment are regressed and trained to obtain the weight coefficient required for the reserve pool to calculate the output layer.

[0013] Optionally, the preprocessing of the timing signal to be tested includes: for the signal to be tested corresponding to a certain moment, performing feature extraction on the signal to be tested corresponding to the moment to obtain a first signal; randomly generating a mask sequence of the signal to be tested corresponding to the moment, and multiplying it with the first signal to obtain a second signal; and converting the second signal from a digital signal to an analog signal.

[0014] Optionally, after converting the second signal from a digital signal to an analog signal, the method further includes: performing amplitude modulation on the converted analog signal and the excitation signal and then inputting the modulated signals into the coupled MEMS resonator.

[0015] Optionally, the state of the preset nonlinear working point is dynamically adjusted by an equivalent Q value through a mechanical pump signal, so that the state of the RC system is more suitable for the current task to be tested. The equivalent Q value can be adjusted by applying an anti-Stokes pump frequency. can be reduced, making the decay time T of the RC system smaller; applying the Stokes pump frequency The size of the kernel is increased, making the system suitable for processing regression tasks that require strong memory capabilities.

[0016] Optionally, the detection of the signal test end of the MEMS coupled resonator includes: performing a primary amplification on the current signal output by the test end through a cross-group amplifier to convert the current signal into a voltage signal; performing a secondary amplification on the voltage signal, and inputting the secondary amplified signal into a bandpass filter to collect a specific frequency waveform; demodulating the envelope signal of the waveform of the specific frequency; and converting the demodulated signal from an analog signal to a digital signal to obtain the first output signal and the second output signal.

[0017] Optionally, the method of feeding back the 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 through a one-way time-delay feedback method includes: multiplying the output signal corresponding to the signal to be tested corresponding to the current moment by a first feedback strength value and then feeding back the result to the virtual node corresponding to the signal to be tested corresponding to the next moment.

[0018] Optionally, the performing regression training on the preset target value and the output signal corresponding to the signal to be measured corresponding to each moment includes: using a ridge regression algorithm for training, and the ridge regression algorithm is:

[0019] W=y′X T (XX T +λI) -1

[0020] 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, and X T is the transposed matrix of X, I is the identity matrix, λ is the regularization coefficient, and y' is the target vector corresponding to the preset target value.

[0021] Optionally, after obtaining the weight coefficient required for the reserve pool calculation, the method also includes: obtaining a test timing signal; performing a reserve pool calculation on the test timing signal according to the weight coefficient, and adjusting relevant parameters of the coupled MEMS resonator according to the calculation result until the weight coefficient of the output signal output by the preset target value and the coupled MEMS resonator is the optimal weight coefficient.

[0022] Optionally, the relevant parameters include at least one of a driving signal frequency, a driving signal amplitude, a feedback strength value, a time interval of a signal to be measured, a feedback time, and a number of virtual nodes of the coupled MEMS resonator.

[0023] (III) Beneficial effects

[0024] The present disclosure proposes a method for calculating the equivalent Q value of a hardware system based on a mechanical pump to dynamically adjust a coupled MEMS resonator reserve pool, and the beneficial effects are:

[0025] 1. The present invention adopts coupled silicon micro-resonators as the core unit of reserve pool calculation. Through finite element simulation and optimization, its design meets the device parameters required for reserve pool calculation, and derives the nonlinear vibration equation of the coupled MEMS resonator based on the mechanical resonance characteristics of its weakly coupled internal structure, so as to regulate the nonlinear vibration state of the coupled MEMS resonator. Therefore, its rich nonlinear dynamic characteristics can be used to improve the dynamic mapping ability of the reserve pool to the input data, thereby improving the ability of the reserve pool calculation to classify and identify time series signals.

[0026] 2. The present invention adopts mechanical pumping to dynamically adjust the equivalent Q value of the silicon microresonator system, and utilizes the dissipative transfer mechanism of energy between different modes of the coupled MEMS resonator. In the classification task, the rich nonlinearity requires high Q value conditions, which leads to an increase in decay time and a significant decrease in calculation speed. The anti-Stoke and Stoke signals of the mechanical pump are used to dynamically reduce / increase the first-order mode equivalent Q value, so that it can achieve optimal performance in different types of prediction tasks; for classification prediction tasks, it not only guarantees its high classification accuracy, but also greatly improves the calculation speed of the RC system; for regression prediction tasks, further improving the decay time can meet the needs of tasks with stronger memory capacity.

[0027] 3. The present disclosure combines the characteristics of MEMS devices, such as small size, low power consumption, easy integration, and easy mass production, with the advantages of simple training methods for reserve pool calculation, in tasks such as timing signal prediction, pattern recognition, and speech classification, thereby meeting the needs of perception and computing integration in the Internet of Things scenario. It also solves the compatibility issues of device size, power consumption, and accurate classification capabilities of traditional hardware implementation methods for reserve pool calculation. BRIEF DESCRIPTION OF THE DRAWINGS

[0028] For a more complete understanding of the present disclosure and its advantages, reference will now be made to the following description in conjunction with the accompanying drawings, which are incorporated into and constitute a part of the specification, illustrate embodiments consistent with the present disclosure, and together with the description are used to explain the principles of the present disclosure. Among them:

[0029] Figure 1 A flowchart of a method for dynamically adjusting the equivalent Q value of a silicon microresonator RC system based on mechanical pumping according to an embodiment of the present disclosure is schematically shown;

[0030] Figure 2 The schematic diagram of the one-way time delay feedback test principle of dynamically adjusting the equivalent Q value of the silicon microresonator RC system based on mechanical pumping in the embodiment of the present disclosure is schematically shown;

[0031] Figure 3 Schematic diagram of a coupled MEMS resonator SEM (Scan Electron Microscope) according to an embodiment of the present disclosure;

[0032] Figure 4 Schematic diagram of COMSOL simulation of different mode shapes of coupled MEMS resonators and schematic diagram of mechanical pump Q-value modulation principle according to an embodiment of the present disclosure;

[0033] Figure 5 A schematic diagram of a device of a reservoir computing system based on coupled MEMS resonators according to an embodiment of the present disclosure;

[0034] Figure 6 The optimal classification accuracy result diagram of handwritten digit recognition under different equivalent Q values ​​according to an embodiment of the present disclosure is schematically shown. DETAILED DESCRIPTION

[0035] Hereinafter, embodiments of the present disclosure will be described with reference to the accompanying drawings. However, it should be understood that these descriptions are exemplary only and are not intended to limit the scope of the present disclosure. In the following detailed description, for ease of explanation, many specific details are set forth to provide a comprehensive understanding of the embodiments of the present disclosure. However, it is apparent that one or more embodiments may also be implemented without these specific details. In addition, in the following description, descriptions of known structures and technologies are omitted to avoid unnecessary confusion of the concepts of the present disclosure.

[0036] The terms used herein are only for describing specific embodiments and are not intended to limit the present invention. The terms "comprise", "include", etc. used herein indicate the existence of the features, steps, operations and / or components, but do not exclude the existence or addition of one or more other features, steps, operations or components.

[0037] Figure 1 The flowchart schematically shows a method for calculating the equivalent Q value of a hardware system based on mechanical pumping to dynamically adjust a coupled MEMS resonator reserve pool.

[0038] like Figure 1 As shown, the method may include, for example, operations S101 to S107.

[0039] S101, performing a frequency sweep experiment on the coupled MEMS resonator to determine the first-order and second-order modal frequencies of the device. The equivalent Q value of the resonator is adjusted by applying a specific frequency pump signal.

[0040] In operation S101, a preliminary experiment of adjusting the equivalent Q value of the coupled resonator by mechanical pumping can accurately achieve the control of the Q value in the later experiment.

[0041] Please combine Figure 3 , the specific pre-experimental process is:

[0042] An open-loop frequency sweep experiment is performed on the coupled MEMS resonator to determine the first-order modal angular frequency w1 and the second-order modal angular frequency w2, and further obtain two mechanical pump signals: the angular frequency of the anti-Stoke signal is the difference between the two frequencies w p =w2-w1; the angular frequency value of the Stoke signal is the sum of the two frequencies w p =w2+w1.

[0043] Then different pump signals are applied to adjust the equivalent Q value, and its mechanism is:

[0044] When anti-Stoke is applied, the larger the excitation amplitude, the smaller the system equivalent Q value; when Stoke signal is applied, the larger the excitation amplitude, the larger the system equivalent Q value.

[0045] S102, preprocessing the timing signal to be measured, so that the signal to be measured corresponding to each moment in the timing signal to be measured corresponds to the dimension of the virtual node coupled to the MEMS resonator one by one.

[0046] In operation S102, through a preprocessing process, the mutual mapping relationship of data within the time interval can be guaranteed.

[0047] Please combine Figure 2 and Figure 5 , the specific preprocessing process is:

[0048] For a signal u(t) to be tested corresponding to a certain moment, a redundancy removal operation is performed on the signal u(t) to be tested corresponding to the moment to obtain a first signal. The specific technical method is as follows.

[0049] A mask sequence m(t) of the signal to be tested corresponding to the moment is randomly and fixedly generated, and is multiplied with the first signal and a sinusoidal signal of a specified frequency to obtain a second signal.

[0050] The second signal is converted from a digital signal to an analog signal through a DAC module.

[0051] S103, applying a selected pump signal to the coupled MEMS resonator, adjusting the system to work at a specified equivalent Q value, and then the resonator performs a nonlinear response to the preprocessed input signal.

[0052] Please continue to combine Figure 2 and Figure 5 The equivalent Q value at the first-order mode frequency under different intensity pump signals is used as the change of the reserve pool calculation system to measure its effect on the nonlinear mapping ability.

[0053] use Figure 2The detection circuit and unidirectional time-delay feedback control loop shown in the figure adjust the coupled MEMS resonator to the preset nonlinear operating point, and realize the nonlinear response to the timing signal to be measured under the specified system equivalent Q value. Specifically, the realization of this process depends on the selection of the nonlinear operating point of the resonator. First, the forced vibration equation of the double-clamped beam is theoretically derived and modeled to study the basic characteristics of the nonlinear response. Then, the forced vibration equation of the coupled resonator under ideal conditions is derived, as shown in Figure 3 This is the COMSOL simulation model of the coupled resonator. Assuming that the equivalent mass m of the two resonators is the same as the linear mechanical stiffness coefficient k, the change in displacement can be obtained by solving the differential equation:

[0054]

[0055] The above formula is the forced nonlinear vibration equation of the coupled MEMS resonator. Among them, m is the equivalent mass of the resonator; γ is the damping term, which is determined by parameters such as the quality factor of the system; k is the linear mechanical stiffness coefficient, which is determined by the material and size of the coupled MEMS resonator; k3 is the nonlinear mechanical stiffness coefficient; k c is the linear coupling stiffness coefficient; k c3 is the nonlinear coupling stiffness coefficient; F drive is the driving signal amplitude, ω d is the driving signal frequency of the coupled MEMS resonator, F pump is the pump signal amplitude, ω p is the pump signal frequency of the coupled MEMS resonator, x1 is the amplitude of the beam at one end of the double-clamped beam resonator, x2 is the amplitude of the beam at the other end of the double-clamped beam resonator, To find the first derivative of the amplitude, To find the second derivative of the amplitude.

[0056] The constraint condition of the above formula is that the resonator is driven by a distributed load. The relevant parameter values ​​under the action of mechanical force can be derived. The linear mechanical stiffness coefficient k = (32EdW 3 ) / (L 3 ), nonlinear mechanical stiffness E is the Young's modulus of single crystal silicon, and its value is E=170*10 6 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.

[0057] The design of the nonlinear vibration equation (Formula 2) of the above coupled MEMS resonator can be completed by finite element simulation design (such as Figure 3 As shown in Figure 2, this method of designing the nonlinear vibration equation of the MEMS resonator adds a new coupling coefficient term k. c , k c3, so that the coupled MEMS resonator can obtain richer nonlinear states than two independent double-ended fixed beam resonators, which is equivalent to increasing the dimension of the neural network mapping unit, greatly improving the performance of the reserve pool computing system. At the same time, the equivalent Q value of the system can be adjusted by adjusting the size of the applied pump signal, which is suitable for dynamic parameter control of packaged devices.

[0058] S104, coupling the driving end of the MEMS resonator to the symmetrical testing end to obtain the response signal at each moment, and performing delayed feedback, and performing a logic operation with the signal to be tested at the next moment.

[0059] Please continue to combine Figure 2 ,pass Figure 2 The designed detection circuit realizes parallel plate capacitance detection on the coupled resonator and the symmetrical signal test end of the driving end, converts the current signal into a voltage signal and amplifies the weak signal, completing the sampling and conversion of the output signal.

[0060] Specifically, first, the current signal output from the test end is amplified by the cross-group amplifier to convert the current signal into a voltage signal. Secondly, the voltage signal is amplified by the second stage, and the signal after the second stage amplification is input into the bandpass filter to collect the waveform of a specific frequency. Thirdly, the analog signal is converted to a digital signal, and the ADC bit number is 12 bits and the sampling rate meets the Nyquist sampling theorem. Finally, the envelope signal of the waveform of the specific frequency is demodulated, and the first response signal is obtained from the demodulated signal.

[0061] After obtaining the response signal, the response signal 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 through a one-way time-delay feedback method. Specifically, the response signal corresponding to the signal to be tested corresponding to the current moment is multiplied by an appropriate feedback strength value and then fed back to the virtual node corresponding to the signal to be tested corresponding to the next moment. At the virtual node corresponding to the next moment, the signal of the one-way time-delay feedback is added to the signal to be tested corresponding to the next moment, and after amplitude modulation with the excitation signal, it is input into the coupled MEMS resonator through the DAC module. For the signals to be tested (timing signals to be tested) arranged in time sequence, the above feedback operation is repeated to obtain the response signal corresponding to the signal to be tested at each moment.

[0062] Among them, multiplying the response signal x1(t) by a feedback strength value can change the attenuation strength to a certain extent, and feeding back the corresponding node of the signal to be tested at the next moment and adding it to the signal to be tested at the next moment can achieve richer dynamic influences between signal contexts and enhance the memory capacity of the reserve pool calculation. At the same time, the equivalent Q value of the system can be adjusted to achieve a more direct change in the attenuation strength. This parameter directly affects the memory performance and computing performance.

[0063] Specifically, for regression tasks, the reserve pool system is required to have higher memory performance, that is, the decay time of the resonator is longer; for classification tasks, the ball reserve pool system has stronger nonlinear mapping capabilities, that is, the system Q value is smaller, and the final classification accuracy will be reduced due to the low system signal-to-noise ratio. Therefore, under high vacuum packaging conditions, mechanical pumping is used to reduce its equivalent Q value, which can ensure classification accuracy and low memory strength.

[0064]

[0065] S105, performing regression training on the preset target value and the response signal of the signal to be measured corresponding to each moment, to obtain the weight coefficient required for the reserve pool calculation.

[0066] The response corresponding to the signal to be tested at each moment obtained in operation S104 is subjected to regression training together with the preset target value. The regression training may adopt, for example, the ridge regression algorithm:

[0067] W=y′X T (XX T +λI) -1

[0068] Where W is the matrix corresponding to the weight coefficient, X is the matrix corresponding to the first output signal or the second output signal, and X T is the transposed matrix of X, y' is the target vector corresponding to the preset target value, I is the unit matrix, and λ is the regularization coefficient, which is used to prevent data overflow and is generally set to 0.1. The loss function is used to train the output weights, which enhances the robustness of the system and better realizes the classification or prediction function of time series signals.

[0069] S106, testing the obtained weight coefficient, and judging whether the weight coefficient is an optimal weight coefficient according to the test result.

[0070] After the training phase, the weight coefficient obtained in operation S105 needs to be tested, specifically including: obtaining a test timing signal. The test timing signal is calculated as a reserve pool according to the weight coefficient to obtain a test result. If the test result has met the needs of the actual application, the weight coefficient at this time can be considered to be the optimal weight coefficient, and the training is completed. Among them, the specific standard for judging whether the weight coefficient is the optimal weight coefficient is: by performing the same processing on the test set, the predicted value multiplied by the weight coefficient and the true value are calculated by normalizing the root mean square error, and the system performance is judged by the quantitative performance index to determine whether it is the optimal weight coefficient. The Winner-Takes-All algorithm can also be added to quantify the classification accuracy performance index to more intuitively evaluate the system performance.

[0071] For example, the signal to be tested is a handwritten digital signal, and the purpose is to identify and classify the handwritten digital signal. If the identification and classification are accurate and clear, it can be considered that the test result has met the requirements of the actual application. If the test result does not meet the requirements of the actual application, operation S107 is performed.

[0072] S107, adjusting relevant parameters of the coupled MEMS resonator according to the test results and optimizing the experiment.

[0073] In operation S107, the adjustment process is performed until the weight coefficient obtained by training according to the preset target value and the response signal output by the coupled MEMS resonator is the optimal weight coefficient, wherein the relevant parameters may include at least one of the driving signal frequency, driving signal amplitude, feedback strength value, time interval of the signal to be measured, feedback time and number of virtual nodes of the coupled MEMS resonator. That is, after adjusting the parameters, the above operations are repeatedly performed, and the output signal is adjusted by adjusting the parameters, and finally the output signal corresponding to the optimal weight coefficient is obtained.

[0074] In the training phase, the weight coefficients are obtained using the host computer and the corresponding data set. In the testing phase, the written control program is copied to the Field Programmable Gate Array (FPGA) system board to realize the overall program control, including the preprocessing operation of the signal to be tested, the downsampling of the output signal, the output of the pump signal, the one-way feedback control, the storage and the result display.

[0075] The disclosed embodiment also provides a device for calculating the equivalent Q value of the hardware system by dynamically adjusting the coupled MEMS resonator reserve pool based on mechanical pumping. Figure 2 and Figure 5 , the device may include an integrated circuit board, a coupled MEMS resonator and an FPGA chip.

[0076] The integrated circuit board integrates the MEMS interface circuit module, ADC and DAC modules. The coupled MEMS resonator is vacuum packaged according to the required Q value and installed on the integrated circuit board after wire bonding to realize the integration of the MEMS chip and the CMOS circuit. Among them, the detection circuit includes a transimpedance amplifier, a secondary amplifier circuit, a bandpass filter circuit and a power module, etc., which are used to detect and detect the MEMS coupled resonator and obtain the response signal corresponding to the signal to be tested at each moment.

[0077] The FPGA chip is used to control the detection circuit to realize DAC signal pickup, and implement envelope detection, downsampling and other operations on the digital end, perform data storage and shift register operations, realize the association between contexts (complex logical operations such as multiplication and addition), and control the ADC to achieve synchronous output of unidirectional delayed feedback signals; and perform regression training on the timing signal to be tested according to the optimal weight coefficient, obtain the optimal weight coefficient required for the reserve pool calculation and perform classification and identification.

[0078] The details not included in the device embodiment are the same as those in the above method embodiment. Please refer to the above method embodiment and will not be repeated here.

[0079] In order to facilitate a better understanding of the technical solution of this embodiment, the present invention takes the recognition of handwritten digital images of 0-9 as an example to further illustrate the above-mentioned device and solution.

[0080] The experimental parameter values ​​selected in this example are: V ac =1V, V dc =30V, f d =195KHz, Q=1500, T=2Q / w n ; Q eff ≈100, 15T eff ≈T.

[0081] Under the above parameter settings, while ensuring high signal-to-noise ratio and high classification accuracy of the system, the processing speed is 15 times that when no mechanical pump is applied. Under other high vacuum conditions, the speed can be increased by dozens or even hundreds of times. Figure 6 The red triangles show the classification accuracy of the coupled resonator RC system with different Q values ​​for handwritten digit recognition without mechanical pumping. The green ellipse shows the classification accuracy of the coupled resonator RC system with different Q values ​​under the above conditions. eff ≈100, which preliminarily shows the advantage of using mechanical pumping to adjust the equivalent Q value.

[0082] The specific process is as follows:

[0083] Operation a: Complete the device layout and adjust the above parameters. Complete the preprocessing of the original handwritten digital image signal and the adjustment of the upper computer program of the digital feedback control loop according to the given parameters.

[0084] Preprocessing includes: de-redundancy operation on handwritten digital signal construction, extraction results such as Figure 5 Shown in the dashed box.

[0085] Operation b: Use the control program in the FPGA to realize digital-to-analog conversion through the DAC module after amplitude modulation with the excitation signal, and input the coupled MEMS resonator; the other end outputs the required pump signal, and also realizes digital-to-analog conversion through the DAC module. Figure 2 As shown, the coupled MEMS resonator is adjusted to a nonlinear working state under a specified system equivalent Q value, and the measured signal first passes through the single resonator at the driving end to complete the nonlinear mapping. The detection end symmetrical to the driving end synchronously outputs the corresponding time domain amplitude change signal.

[0086] Operation c: Use the MEMS interface circuit to collect the real-time output signals of operation b respectively, and complete the functions in operations S103-S104, that is, perform cross-group primary amplification, secondary amplification, bandpass filtering, etc. in sequence, including using FPGA to control the ADC module to realize analog-to-digital conversion, envelope detection and down-sampling output and other operations.

[0087] Operation d: The FPGA control program realizes the time delay feedback of the real-time signal and stores it.

[0088] Operation step e: After completing the above training steps, use the constructed ridge regression training model to obtain the optimized output weight coefficient matrix; next, select several groups of handwritten digital signals in the training set and the test set to test the classification accuracy of the constructed system, and adjust the model-related parameters according to the results to obtain the optimal performance.

[0089] Operation f: Based on the optimal weight coefficient and other parameters obtained above, copy the complete program and the handwritten digital signal to be tested to the FPGA. After running, the recognition result is obtained and displayed on the FPGA display.

[0090] So far, operations a to f have realized the handwritten digital image signal classification and recognition function based on the method and device of this embodiment.

[0091] It will be appreciated by those skilled in the art that, although the present disclosure has been shown and described with reference to specific exemplary embodiments of the present disclosure, it will be appreciated by those skilled in the art that, without departing from the spirit and scope of the present disclosure as defined by the appended claims and their equivalents, various changes in form and detail may be made to the present disclosure. Therefore, the scope of the present disclosure should not be limited to the above-described embodiments, but should be determined not only by the appended claims, but also by the equivalents of the appended claims.

Claims

1. A method for calculating the equivalent Q value of a hardware system by dynamically adjusting a coupled MEMS resonator reserve pool using a mechanical pump, characterized in that: include: Preprocessing the timing signal to be measured so that the signal to be measured corresponding to each moment in the timing signal to be measured corresponds to the virtual node dimension of the coupled MEMS resonator one by one; Dynamically adjust the resonator by mechanical pumping to reduce the system equivalent quality factor Q value; The resonator is designed to have two modal frequencies with similar frequencies that can realize energy transfer, and has the characteristics of multiple ports. The resonator is a mechanically coupled multi-port coupled MEMS resonator. The nonlinear vibration equation of the MEMS resonator is: Where m is the equivalent mass of the resonator; γ is the damping ratio, which is determined by the quality factor Q parameter of the system; k is the linear mechanical stiffness coefficient, which is determined by the material and size of the coupled MEMS resonator; k3 is the nonlinear mechanical stiffness coefficient; k c is the linear coupling stiffness coefficient; k c3 is the nonlinear coupling stiffness coefficient; F drive is the driving signal amplitude, F pump is the pump signal driving signal amplitude, ω d is the single-ended driving angular frequency of the coupled MEMS resonator, ω p is the driving angular frequency of the pump signal applied to one end of the coupled MEMS resonator, x is the amplitude, x1 is the amplitude of the support beam at one end of the coupled MEMS resonator, x2 is the amplitude of the support beam at the other end of the coupled MEMS resonator, To find the first derivative of the amplitude, To find the second derivative of the amplitude; According to the nonlinear vibration equation, the coupled MEMS resonator is adjusted to a preset nonlinear working point, and a mechanical pump signal f is used to generate a nonlinear vibration signal. pump The preliminary experimental results of the equivalent Q value of the control system are used to determine the corresponding parameters of the required pump signal and the control law; according to the preliminary experiment, a stable mechanical pump signal f is applied at one end of the coupled MEMS resonator. pump , so as to improve the processing speed or memory capability of the nonlinear RC response of the timing signal to be measured; Detecting the signal test end of the coupled MEMS resonator to obtain an output signal corresponding to the signal to be tested at each moment, wherein, for the signal to be tested corresponding to the current moment, the output signal 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 through a time-delay feedback method; The preset target value and the output signal corresponding to the signal to be tested at each moment are regressed to obtain the optimal weight coefficient required for the reserve pool calculation.

2. The method for calculating the equivalent Q value of a hardware system by dynamically adjusting the coupled MEMS resonator reserve pool using a mechanical pump according to claim 1, characterized in that: The preprocessing of the timing signal to be measured includes: For a signal to be tested corresponding to a certain moment, a feature extraction operation is performed on the signal to be tested corresponding to the moment to obtain a first signal; Randomly generate a fixed mask sequence of the signal to be tested corresponding to the moment, and perform multiplication operation on the fixed mask sequence and the first signal to obtain a second signal; The second signal is converted from a digital signal to an analog signal.

3. The method for calculating the equivalent Q value of a hardware system by dynamically adjusting the coupled MEMS resonator reserve pool using a mechanical pump according to claim 2, characterized in that: After converting the second signal from a digital signal to an analog signal, the method further includes: The converted analog signal and the excitation signal are amplitude modulated and then input into the driving end of the coupled MEMS resonator.

4. The method for realizing the calculation of the equivalent Q value of the hardware system by dynamically adjusting the coupled MEMS resonator reserve pool using a mechanical pump according to claim 3, characterized in that: Performing dual-port driving and single-port detection operations on the coupled MEMS resonator includes: The converted analog signal and the excitation signal are used as a driving end after the amplitude modulation of the measured signal to achieve the driving of the specific nonlinear working point of the coupled MEMS resonator and the input of the measured signal; By generating a stable mechanical pump signal and inputting it to the other driving end, the equivalent Q value of the system at the working modal frequency is adjusted; Single-ended detection uses a cross-group amplifier to amplify the current signal output by the test end and convert the current signal into a voltage signal; Performing secondary amplification on the voltage signal, and inputting the secondary amplified signal into a bandpass filter to collect a specific frequency waveform; Demodulating the envelope signal of the waveform of the specific frequency; The demodulated signal is converted from an analog signal to a digital signal to obtain the output signal.

5. The method for calculating the equivalent Q value of a hardware system by dynamically adjusting a coupled MEMS resonator reserve pool using a mechanical pump according to claim 1, characterized in that: The method of feeding back the 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 means of time delay feedback includes: The output signal corresponding to the signal to be tested corresponding to the current moment is multiplied by the feedback strength value and then fed back to the virtual node corresponding to the signal to be tested corresponding to the next moment.

6. The method for realizing the calculation of the equivalent Q value of the hardware system by dynamically adjusting the coupled MEMS resonator reserve pool using a mechanical pump according to claim 1, characterized in that: The step of performing regression training on the preset target value and the output signal corresponding to the signal to be tested at each moment includes: The ridge regression algorithm is used for training, and its equation is: 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 output signal, and X T is the transposed matrix of X, I is the identity matrix, λ is the regularization coefficient, and y' is the target vector corresponding to the preset target value.

7. The method for calculating the equivalent Q value of a hardware system by dynamically adjusting a coupled MEMS resonator reserve pool using a mechanical pump according to claim 1, characterized in that: After obtaining the weight coefficient required for the reserve pool calculation, the following steps are also included: Get test timing signal; A reserve pool calculation is performed on the test timing signal according to the weight coefficient, and relevant parameters of the coupled MEMS resonator are adjusted according to the calculation result, until the weight coefficient trained according to the preset target value and the output signal output by the coupled MEMS resonator is the optimal weight coefficient.

8. The method for realizing the calculation of the equivalent Q value of the hardware system by dynamically adjusting the coupled MEMS resonator reserve pool using a mechanical pump according to claim 7, characterized in that: The related parameters include at least one of a driving signal frequency, a driving signal amplitude, a feedback strength value, a time interval of a signal to be measured, a feedback time and a number of virtual nodes of the coupled MEMS resonator.

9. The method for calculating the equivalent Q value of a hardware system by dynamically adjusting a coupled MEMS resonator reserve pool using a mechanical pump according to claim 4, characterized in that: The applying of the stable mechanical pump signal includes determining the modal frequency of the coupled MEMS resonator and a mechanical pumping preliminary experiment, including: Determine the first-order modal frequency w1 and the second-order modal frequency w2 of the coupled MEMS resonator at the required nonlinear operating point by using corresponding simulation or experimental methods; Using the two modal frequencies, the anti-Stokes pumping frequency can be obtained according to the mechanical pumping principle. Stokes pump frequency Two different signals, through experimental results it can be determined that when the pump frequency is selected at a certain frequency With amplitude The sinusoidal signal can reduce the equivalent Q value of the coupled MEMS resonator system, thereby improving the processing speed of the reserve pool calculation.

10. A device for dynamically adjusting the equivalent Q value of a coupled MEMS resonator reservoir system using a mechanical pump, characterized in that: include: Integrated circuit board, coupled MEMS resonator and FPGA chip; The integrated circuit board is integrated with a MEMS interface circuit module, an ADC and a DAC module, and the coupled MEMS resonator is vacuum packaged according to the required Q value, wherein the Q value is obtained according to the method for realizing the calculation of the equivalent Q value of the hardware system by dynamically adjusting the coupled MEMS resonator reserve pool by mechanical pumping according to any one of claims 1 to 9, and is installed on the integrated circuit board after wire bonding to realize the integration of the MEMS chip and the CMOS circuit, wherein the detection circuit is used to detect the coupled MEMS resonator to obtain a response signal corresponding to the signal to be measured corresponding to each moment; The FPGA chip is used to control the detection circuit to realize DAC signal pickup, and implement envelope detection and downsampling operations on the digital end, perform data storage and shift register operations, realize the association between contexts, and control the ADC to realize the synchronous output of unidirectional delayed feedback signals; and perform regression training on the timing signal to be tested according to the optimal weight coefficient, obtain the optimal weight coefficient required for the reserve pool calculation and perform classification and recognition.

11. The device for dynamically adjusting the equivalent Q value of a coupled MEMS resonator reserve pool system by using a mechanical pump according to claim 10, characterized in that: The detection circuit includes a transimpedance amplifier, a secondary amplifier circuit, a bandpass filter circuit and a power supply module which are connected in sequence.

Images