A calibration method and system for chip dynamic balance testing

Abstract

The present invention relates to the technical field of dynamic balance correction, and specifically relates to a correction method and system for chip dynamic balance testing. A correction system for chip dynamic balance testing includes: an unbalance frequency domain feature vector construction module, an unbalance amount analysis module, a dynamic balance correction module, and a reinforcement learning module. The present invention performs frequency domain analysis on the electrical signals collected by the gyroscope and accelerometer, then determines the vibration signal characteristics corresponding to the chip unbalance through a peak search algorithm, and further determines the dynamic balance correction parameter set corresponding to the chip dynamic balance correction through the cooperation of the unbalance amount analysis model and the dynamic balance correction model. The driving voltage of the IPMC array unit corresponding to the IPMC thin film layer installed on the chip is adjusted through the dynamic balance correction parameter set to realize the adjustment of the chip mass distribution, and further realize the dynamic balance correction of the chip.

Description

Technical Field

[0001] The present invention relates to the technical field of dynamic balance correction, and particularly relates to a correction method and system for chip dynamic balance testing. Background Art

[0002] Chip dynamic balance technology is an important technology in semiconductor manufacturing and packaging processes, aiming to solve problems such as vibration and uneven heat distribution in chips during high-speed operation or high-precision applications. With the continuous progress of chip manufacturing processes, the integration and operating speed of chips have been significantly improved, but this has also made the requirements for chip dynamic balance higher and higher. Summary of the Invention

[0003] The present invention performs frequency-domain analysis on the electrical signals collected by gyroscopes and accelerometers, then determines the vibration signal characteristics corresponding to chip imbalance through a peak-finding algorithm, and further determines the dynamic balance correction parameter set corresponding to chip dynamic balance correction through the cooperation of an imbalance amount analysis model and a dynamic balance correction model. The driving voltage of the IPMC array units corresponding to the IPMC thin film layer installed on the chip is adjusted through the dynamic balance correction parameter set to adjust the mass distribution of the chip, thereby achieving dynamic balance correction of the chip; and during the iterative execution of dynamic balance correction, the dynamic balance correction model is adjusted through reinforcement learning, so that the dynamic balance correction process can better fit the individual differences of chips and improve the dynamic balance correction efficiency.

[0004] The present invention provides a correction method for chip dynamic balance testing, including:

[0005] Attach an IPMC thin film layer to the chip, and the IPMC thin film layer includes a number of IPMC array units;

[0006] Obtain the rotation state data of the chip during rotation through gyroscopes and accelerometers, preprocess the rotation state data to obtain standard rotation state data, then perform a fast Fourier transform on the standard rotation state data to construct a rotation state frequency spectrum diagram. In the rotation state frequency spectrum diagram, with frequency as the horizontal axis and amplitude as the vertical axis, determine the unbalance frequency domain feature vector from the rotation state frequency spectrum diagram through a peak-finding algorithm;

[0007] Send the unbalance frequency domain feature vector into the imbalance amount analysis model for processing, output the imbalance state vector, send the imbalance state vector into the dynamic balance correction model for processing, output the dynamic balance correction parameter set, and the dynamic balance correction parameter set includes the control parameters of all IPMC array units. Control the IPMC thin film layer through the dynamic balance correction parameter set to achieve dynamic balance correction of the chip;

[0008] Repeat the dynamic balance correction of the chip, and adjust the dynamic balance correction model through reinforcement learning; stop the dynamic balance correction of the chip until the peak search algorithm cannot determine the unbalanced frequency domain eigenvector.

[0009] Preferably, determine the unbalanced frequency domain eigenvector from the rotation state spectrogram through the peak search algorithm, specifically including the following steps:

[0010] Arrange all the frequencies in the rotation state spectrogram in ascending order to construct a rotation state frequency set, and divide the rotation state frequency set through a sliding window to obtain several rotation state frequency set segments. The size of the sliding window is C, and the step size is S. The size and step size of the sliding window are both set by the operator, and generally C = S - 2;

[0011] For each rotation state frequency set segment, perform the following: select the frequency with the largest amplitude, the corresponding amplitude, and the corresponding phase in the rotation state frequency set segment to form a candidate eigenvector;

[0012] For each candidate eigenvector, record the frequency corresponding to the candidate eigenvector as the candidate frequency. Based on the candidate frequency, traverse the frequencies to the left in the rotation state spectrogram until a frequency with an amplitude higher than the amplitude E corresponding to the candidate frequency or reaching the boundary of the rotation state spectrogram is found. Record the lowest amplitude during the left traversal as the left lowest amplitude L. Traverse the frequencies to the right in the rotation state spectrogram until a frequency with an amplitude higher than the amplitude corresponding to the candidate frequency or reaching the boundary of the rotation state spectrogram is found. Record the lowest amplitude during the right traversal as the right lowest amplitude R. Calculate the significance P = E - max(L, R). Compare the significance P corresponding to the candidate eigenvector with the significance threshold T. If the significance P corresponding to the candidate eigenvector is higher than the significance threshold T, send the candidate eigenvector to the candidate set; if the significance P corresponding to the candidate eigenvector is not higher than the significance threshold T, do not perform any operation and consider it as unable to determine the unbalanced frequency domain eigenvector. The significance threshold T is adjusted according to the noise floor level corresponding to the rotation state spectrogram;

[0013] Traverse the candidate set. For each selected candidate eigenvector, record the candidate frequency corresponding to the selected candidate eigenvector as f, and determine whether "f ∈ (kg - μ, kg + μ)" holds, where k is a positive integer, g is the rotation frequency of the chip, and μ is the tolerance. If "f ∈ (kg - μ, kg + μ)" holds, record the selected candidate eigenvector as the unbalanced frequency domain eigenvector; if "f ∈ (kg - μ, kg + μ)" does not hold, do not perform any operation and consider it as unable to determine the unbalanced frequency domain eigenvector.

[0014] Preferably, the significance threshold T is adjusted according to the noise floor level corresponding to the rotation state spectrogram, which specifically includes the following steps:

[0015] Denote the average value of the amplitudes corresponding to all frequencies except the candidate frequencies in the rotation state spectrogram as the noise floor level m, and adjust the significance threshold T through the formula T = U0exp(α(m - H0)), where U0 is the initial value of the significance threshold, α is the adjustment coefficient, and H0 is the initial value of the noise level.

[0016] Preferably, the dynamic balance correction model is adjusted through reinforcement learning, which specifically includes the following steps:

[0017] Send the unbalance state vector into the target dynamic balance correction model for processing, output the target dynamic balance correction parameter set, then send the unbalance state vector and the target dynamic balance correction parameter set into the dynamic balance correction parameter evaluation network for processing, output the dynamic balance correction parameter evaluation value, calculate the policy value for the unbalance state vector based on the dynamic balance correction parameter evaluation value, and denote the difference between the unbalance amplitude corresponding to the current unbalance frequency domain feature vector and the unbalance amplitude corresponding to the previous unbalance frequency domain feature vector as the reward value. Calculate the sum of the policy value and the reward value as the gradient value, send the unbalance state vector into the dynamic balance correction model for processing, and adjust the dynamic balance correction model based on the gradient value through the gradient ascent method in the direction of maximizing the gradient value to achieve the adjustment of the dynamic balance correction model; the initial state of the target dynamic balance correction model is the same as that of the dynamic balance correction model, and each time the dynamic balance correction model is adjusted, the target dynamic balance correction model is adjusted.

[0018] Preferably, the target dynamic balance correction model is adjusted, which specifically includes the following steps:

[0019] Adjust the parameters of the target dynamic balance correction model through the following formula: θ(new) = θ(bef) + βb, where θ(new) is the parameter of the target dynamic balance correction model after adjustment, θ(bef) is the parameter of the target dynamic balance correction model before adjustment, β is the learning rate, and b is the parameter of the dynamic balance correction model, and θ(new), θ(bef), and b correspond to the same parameter item.

[0020] Preferably, the unbalance amount analysis model is trained, which specifically includes the following steps:

[0021] Obtain the unbalance amount analysis training samples, where the unbalance amount analysis training samples include unbalance frequency domain feature vectors and corresponding unbalance state vectors. Combine all the unbalance amount analysis training samples into an unbalance amount analysis training set, and train the unbalance amount analysis model through the unbalance amount analysis training set. The training target is the unbalance state vectors in the unbalance amount analysis training samples.

[0022] Preferably, the dynamic balance correction model and the dynamic balance correction parameter evaluation network are trained, specifically including the following steps:

[0023] Obtain dynamic balance correction training samples. The dynamic balance correction training samples include unbalance state vectors and corresponding dynamic balance correction parameter sets. All the dynamic balance correction training samples are combined into a dynamic balance correction training set, and the dynamic balance correction model is trained through the dynamic balance correction training set. The training objective is the dynamic balance correction parameter set in the dynamic balance correction training samples;

[0024] Obtain dynamic balance correction parameter evaluation training samples. The dynamic balance correction parameter evaluation training samples include unbalance state vectors and corresponding dynamic balance correction parameter sets. The dynamic balance correction parameter evaluation training samples are labeled through policy values. All the labeled dynamic balance correction parameter evaluation training samples are combined into a dynamic balance correction parameter evaluation training set, and the dynamic balance correction parameter evaluation network is trained through the dynamic balance correction parameter evaluation training set. The training objective is the labeled policy values.

[0025] The present invention also provides a correction system for chip dynamic balance testing, including:

[0026] An unbalance frequency domain feature vector construction module, configured to obtain rotation state data when the chip rotates through a gyroscope and an accelerometer, preprocess the rotation state data to obtain standard rotation state data, then perform a fast Fourier transform on the standard rotation state data to construct a rotation state frequency spectrum diagram. In the rotation state frequency spectrum diagram, the frequency is used as the horizontal axis and the amplitude is used as the vertical axis, and the unbalance frequency domain feature vector is determined from the rotation state frequency spectrum diagram through a peak search algorithm;

[0027] An unbalance amount analysis module, configured to send the unbalance frequency domain feature vector into an unbalance amount analysis model for processing and output an unbalance state vector;

[0028] A dynamic balance correction module, configured to send the unbalance state vector into the dynamic balance correction model for processing and output a dynamic balance correction parameter set. The dynamic balance correction parameter set includes control parameters of all IPMC array units, and the IPMC thin film layer is controlled through the dynamic balance correction parameter set to achieve dynamic balance correction of the chip;

[0029] A reinforcement learning module, configured to perform reinforcement learning on the dynamic balance correction model during the iterative process of dynamic balance correction.

[0030] The present invention has the following advantages:

[0031] The present invention performs frequency-domain analysis on the electrical signals collected by the gyroscope and accelerometer, determines the vibration signal characteristics corresponding to the chip imbalance through the peak search algorithm, and then determines the dynamic balance correction parameter set corresponding to the chip dynamic balance correction through the cooperation of the unbalance amount analysis model and the dynamic balance correction model. The driving voltage of the IPMC array unit corresponding to the IPMC thin film layer installed on the chip is adjusted through the dynamic balance correction parameter set to adjust the mass distribution of the chip, and then the dynamic balance correction of the chip is realized. Moreover, during the iterative execution of the dynamic balance correction process, the dynamic balance correction model is adjusted through reinforcement learning, so that the dynamic balance correction process can better fit the individual differences of the chip and improve the dynamic balance correction efficiency. Description of the Drawings

[0032] Figure 1 It is a schematic structural diagram of the correction system for chip dynamic balance testing adopted in the embodiment of the present invention. Detailed Embodiments

[0033] In order to enable those skilled in the art to better understand the technical solutions in the present invention, the following will clearly and completely describe the technical solutions in the embodiments of the present invention with reference to the accompanying drawings in the embodiments of the present invention.

[0034] Embodiment 1, a correction method for chip dynamic balance testing, includes:

[0035] Attach an IPMC thin film layer to the chip. The IPMC thin film layer includes a number of IPMC array units. It should be noted that IPMC is an ionic polymer metal composite material, which can bend under the action of voltage. Acting on the chip, it can adjust the mass distribution of the chip, and then realize the dynamic balance correction of the chip. Electrodes are set for each IPMC array unit, and control can be achieved through the driving voltage.

[0036] The rotation state data of the chip during rotation is obtained through a gyroscope and an accelerometer. The rotation state data here consists of electrical signals collected by the gyroscope and the accelerometer. These electrical signal data can reflect the balance state of the chip during rotation. And it can be expected that if an imbalance occurs during the rotation of the chip, the chip will generate periodic vibrations during rotation, and the center of gravity of the chip will shift, which is reflected in the electrical signals collected by the gyroscope and the accelerometer. The electrical signals collected by the gyroscope and the accelerometer will show obvious peaks in the frequency domain; the rotation state data is preprocessed to obtain standard rotation state data. The preprocessing operations include filtering operations, de-mean operations, and window function processing operations, etc. Among them, the filtering operations include low-pass filtering (removing high-frequency noise and retaining low-frequency imbalance signal components), high-pass filtering (removing low-frequency drift or DC components), band-pass filtering (if the approximate range of the imbalance frequency is known, a band-pass filter can be used to retain only the signals within the frequency range of interest). The imbalance frequency corresponds to the rotation frequency of the chip or its harmonic frequencies; the de-mean operation is to remove the DC component of the signal, making the signal fluctuate around the zero axis, which is beneficial for subsequent FFT analysis; the window function operation is to apply a window function to the time-domain signal, such as a Hanning window, a Hamming window, etc. The window function can smooth the edges of the signal, reduce spectral leakage, and improve frequency resolution and amplitude accuracy; then perform a fast Fourier transform (FFT) on the standard rotation state data to construct a rotation state spectrogram. In the rotation state spectrogram, the frequency is the horizontal axis and the amplitude is the vertical axis; determine the imbalance frequency domain feature vector from the rotation state spectrogram through a peak search algorithm. The imbalance frequency domain feature vector generally includes data such as the imbalance frequency, imbalance amplitude, and imbalance phase corresponding to the imbalance vibration, which are used to describe the imbalance state of the chip during rotation;

[0037] The imbalance frequency domain feature vector is sent into an imbalance analysis model for processing, and an imbalance state vector is output. The imbalance state vector includes the magnitude of the imbalance mass and the angle of the imbalance mass. The imbalance analysis model is established based on a BP neural network. The imbalance analysis model realizes the mapping from the imbalance amplitude to the magnitude of the imbalance mass and the mapping from the imbalance phase to the angle of the imbalance mass; the magnitude of the imbalance mass here is the mass corresponding to the shift of the chip's center of gravity, and the angle of the imbalance mass is the angle corresponding to the shift of the chip's center of gravity;

[0038] The unbalanced state vector is sent to the dynamic balance correction model for processing, and a dynamic balance correction parameter set is output. The dynamic balance correction parameter set includes control parameters of all IPMC array units. The control parameter here is the driving voltage corresponding to the IPMC array unit. The IPMC film layer is controlled by the dynamic balance correction parameter set to achieve dynamic balance correction of the chip. It should be noted that in the process of controlling the IPMC film layer by the dynamic balance correction parameter set, the driving voltage of all IPMC array units is actually adjusted, so that each IPMC array unit produces corresponding deformation, and then the mass distribution of the entire chip is fine-tuned to complete the correction of the chip dynamic balance; the dynamic balance correction model is established based on the BP neural network;

[0039] During the dynamic balancing test of the chip, the chip will be rotated to analyze the imbalance phenomenon of the chip in the rotation test. These imbalance phenomena are usually caused by material non-uniformity and process errors in the chip preparation process. Based on the analyzed imbalance phenomenon, the IPMC film layer is fine-tuned for correction. However, this dynamic balancing correction process is not completed in one go, but requires iterative correction. The dynamic balancing correction of the chip is repeated, and the dynamic balancing correction model is adjusted through reinforcement learning. When the peak search algorithm cannot determine the imbalance frequency domain feature vector, the dynamic balancing correction of the chip is stopped.

[0040] The present application performs frequency domain analysis on the electrical signals collected by the gyroscope and the accelerometer, and then determines the vibration signal characteristics corresponding to the chip imbalance through a peak search algorithm, and then determines the dynamic balance correction parameter set corresponding to the chip dynamic balance correction through the coordination of the imbalance analysis model and the dynamic balance correction model. The driving voltage of the IPMC array unit corresponding to the IPMC thin film layer installed on the chip is adjusted through the dynamic balance correction parameter set to adjust the chip mass distribution, thereby achieving dynamic balance correction of the chip; and in the process of iterative execution of dynamic balance correction, the dynamic balance correction model is adjusted through reinforcement learning, so that the dynamic balance correction process can better fit the individual differences of the chip and improve the efficiency of dynamic balance correction.

[0041] Determining the unbalanced frequency domain feature vector from the rotation state spectrum diagram by using a peak search algorithm specifically includes the following steps:

[0042] Arrange all the frequencies in the rotation state spectrum diagram in ascending order to construct a rotation state frequency set, and segment the rotation state frequency set through a sliding window to obtain several rotation state frequency set fragments. The size of the sliding window is C, and the step length is S. The size and step length of the sliding window are set by the operator, generally C=S-2;

[0043] For each segment of the rotational state frequency set, the following operations are performed: Select the frequency with the largest amplitude, the corresponding amplitude, and the corresponding phase in the segment of the rotational state frequency set to form a candidate feature vector. By analyzing the local maximum amplitude of the rotational state spectrogram, features that may correspond to the unbalanced state can be preliminarily screened out. This is because when there is an imbalance in the chip, one or more obvious peaks will appear at positions corresponding to the unbalanced frequency (usually the rotational frequency of the chip or its multiple frequencies) on the spectrogram. The frequencies of these peaks indicate the frequency components of the imbalance, and the magnitude of the amplitude is related to the degree of imbalance;

[0044] For each candidate feature vector, denote the frequency corresponding to the candidate feature vector as the candidate frequency. Based on the candidate frequency, traverse the frequencies to the left in the rotational state spectrogram until a frequency with an amplitude higher than the amplitude E corresponding to the candidate frequency is found or the boundary of the rotational state spectrogram is reached. Denote the lowest amplitude during the left traversal as the left lowest amplitude L. Traverse the frequencies to the right in the rotational state spectrogram until a frequency with an amplitude higher than the amplitude corresponding to the candidate frequency is found or the boundary of the rotational state spectrogram is reached. Denote the lowest amplitude during the right traversal as the right lowest amplitude R. Calculate the significance P = E - max(L, R). A peak with a high significance indicates that the energy of this frequency component is significantly higher than the noise level of the surrounding frequencies and is more likely to be a true signal peak (such as a peak caused by imbalance). A peak with a low significance may be just noise fluctuations or unimportant frequency components. By judging the significance, the frequencies corresponding to noise and the frequencies corresponding to the unbalanced state can be distinguished. Compare the significance P corresponding to the candidate feature vector with the significance threshold T. If the significance P corresponding to the candidate feature vector is higher than the significance threshold T, send the candidate feature vector into the candidate set; if the significance P corresponding to the candidate feature vector is not higher than the significance threshold T, no operation is performed, and it is regarded as unable to determine the unbalanced frequency domain feature vector. The significance threshold T is adjusted according to the noise floor level corresponding to the rotational state spectrogram;

[0045] Traverse the candidate set. For each selected candidate feature vector, denote the candidate frequency corresponding to the selected candidate feature vector as f, and judge whether "f ∈ (kg - μ, kg + μ)" holds, where k is a positive integer, g is the rotational frequency of the chip, and μ is the tolerance value set by the operator. The judgment of "f ∈ (kg - μ, kg + μ)" is because the unbalanced frequency coincides with the rotational frequency of the chip or its harmonic frequency. If "f ∈ (kg - μ, kg + μ)" holds, denote the selected candidate feature vector as the unbalanced frequency domain feature vector; if "f ∈ (kg - μ, kg + μ)" does not hold, no operation is performed, and it is regarded as unable to determine the unbalanced frequency domain feature vector.

[0046] Adjust the significance threshold T according to the noise floor level corresponding to the rotation state spectrogram, specifically including the following steps:

[0047] Denote the average value of the amplitudes corresponding to all frequencies except the candidate frequencies in the rotation state spectrogram as the noise floor level m. Adjust the significance threshold T through the formula T = U0exp(α(m - H0)), where U0 is the initial value of the significance threshold determined by the operator, α is the adjustment coefficient, and H0 is the initial value of the noise level obtained by the operator based on experiments. α is used to map exp(α(m - H0)) to the range [0, 2] to adjust the significance threshold T. When the noise level is high, increasing the significance threshold can reduce false detections; when the noise level is low, decreasing the significance threshold can improve sensitivity.

[0048] Adjust the dynamic balance correction model through reinforcement learning, specifically including the following steps:

[0049] Send the unbalanced state vector into the target dynamic balance correction model for processing, and output the target dynamic balance correction parameter set. Then send the unbalanced state vector and the target dynamic balance correction parameter set into the dynamic balance correction parameter evaluation network for processing. The dynamic balance correction parameter evaluation network is also established based on the BP neural network and outputs the dynamic balance correction parameter evaluation value. Calculate the policy value for the unbalanced state vector based on the dynamic balance correction parameter evaluation value, specifically calculate the derivative of the unbalanced state vector based on the dynamic balance correction parameter evaluation value, and this derivative is the policy value. Denote the difference between the unbalanced amplitude corresponding to the current unbalanced frequency domain feature vector and the unbalanced amplitude corresponding to the previous unbalanced frequency domain feature vector as the reward value. Calculate the sum of the policy value and the reward value as the gradient value. Send the unbalanced state vector into the dynamic balance correction model for processing, and adjust the dynamic balance correction model through the gradient ascent method in the direction of maximizing the gradient value based on the gradient value, so as to realize the adjustment of the dynamic balance correction model. Each adjustment process is executed based on the effect of the dynamic balance correction feedback, which can make the dynamic balance correction process more suitable for the individual differences of the chip; the initial state of the target dynamic balance correction model is the same as that of the dynamic balance correction model. After each adjustment of the dynamic balance correction model, adjust the target dynamic balance correction model.

[0050] Adjust the target dynamic balance correction model, specifically including the following steps:

[0051] Adjust the parameters of the target dynamic balance correction model through the following formula: θ(new) = θ(bef) + βb, where θ(new) is the parameter of the target dynamic balance correction model after adjustment, θ(bef) is the parameter of the target dynamic balance correction model before adjustment, β is the learning rate, generally 0.01, to ensure the stability of the target dynamic balance correction model and avoid over-adjustment of the dynamic balance correction model, b is the parameter of the dynamic balance correction model, and θ(new), θ(bef), and b correspond to the same parameter item.

[0052] Train the unbalance analysis model, which specifically includes the following steps:

[0053] Obtain unbalance analysis training samples. The unbalance analysis training samples include unbalance frequency domain feature vectors and corresponding unbalance state vectors. Both the unbalance frequency domain feature vectors and the corresponding unbalance state vectors here are obtained by the operator according to actual experiments, generally assisted by professional software. Combine all unbalance analysis training samples into an unbalance analysis training set, and train the unbalance analysis model through the unbalance analysis training set. The training target is the unbalance state vector in the unbalance analysis training samples, and judge whether the corresponding training conditions are met. The training conditions generally mean that the accuracy of the unbalance analysis model meets the expectations. If the corresponding training conditions are met, output the trained unbalance analysis model; otherwise, continue to train the unbalance analysis model through the unbalance analysis training set.

[0054] Train the dynamic balance correction model and the dynamic balance correction parameter evaluation network, which specifically includes the following steps:

[0055] Obtain dynamic balance correction training samples. The dynamic balance correction training samples include unbalance state vectors and corresponding dynamic balance correction parameter sets. The dynamic balance correction parameters here are the optimal correction schemes set by the operator for the unbalance state vectors. Combine all dynamic balance correction training samples into a dynamic balance correction training set, and train the dynamic balance correction model through the dynamic balance correction training set. The training target is the dynamic balance correction parameter set in the dynamic balance correction training samples, and judge whether the corresponding training conditions are met. The training conditions generally mean that the accuracy of the dynamic balance correction model meets the expectations. If the corresponding training conditions are met, output the trained dynamic balance correction model; otherwise, continue to train the dynamic balance correction model through the dynamic balance correction training set;

[0056] Obtain the training samples for evaluating the dynamic balance correction parameters. The training samples for evaluating the dynamic balance correction parameters include the unbalance state vectors and the corresponding sets of dynamic balance correction parameters. Label the training samples for evaluating the dynamic balance correction parameters by the policy values. Since the dynamic balance correction parameters here are the optimal correction schemes set by the operator for the unbalance state vectors, the labeled policy values are generally recorded as 1. Combine all the labeled training samples for evaluating the dynamic balance correction parameters to form the training set for evaluating the dynamic balance correction parameters. Train the network for evaluating the dynamic balance correction parameters with the training set for evaluating the dynamic balance correction parameters. The training objective is the labeled policy value, and determine whether the corresponding training conditions are met. The training conditions generally mean that the accuracy of the network for evaluating the dynamic balance correction parameters meets the expectation. If the corresponding training conditions are met, output the trained network for evaluating the dynamic balance correction parameters; otherwise, continue to train the network for evaluating the dynamic balance correction parameters with the training set for evaluating the dynamic balance correction parameters.

[0057] Example 2. A correction system for chip dynamic balance testing, see Figure 1 , including:

[0058] An unbalanced frequency domain feature vector construction module is used to obtain the rotation state data of the chip during rotation through a gyroscope and an accelerometer. The rotation state data here consists of electrical signals collected by the gyroscope and the accelerometer. These electrical signal data can reflect the balance state of the chip during rotation. And it can be expected that if there is an unbalanced phenomenon during the rotation of the chip, the chip will generate periodic vibrations during rotation, and the center of gravity of the chip will shift, which is reflected in the electrical signals collected by the gyroscope and the accelerometer. The electrical signals collected by the gyroscope and the accelerometer will show obvious peaks in the frequency domain; the rotation state data is preprocessed to obtain standard rotation state data. The preprocessing operations include filtering operations, de-mean operations, and window function processing operations, etc. Among them, the filtering operations include low-pass filtering (removing high-frequency noise and retaining low-frequency unbalanced signal components), high-pass filtering (removing low-frequency drift or DC components), and band-pass filtering (if the approximate range of the unbalanced frequency is known, a band-pass filter can be used to retain only the signals within the frequency range of interest). The unbalanced frequency corresponds to the rotation frequency of the chip or its harmonic frequencies; the de-mean operation is to remove the DC component of the signal, making the signal fluctuate around the zero axis, which is beneficial for subsequent FFT analysis; the window function operation is to apply a window function to the time-domain signal, such as a Hanning window, a Hamming window, etc. The window function can smooth the edges of the signal, reduce spectral leakage, improve frequency resolution and amplitude accuracy; then perform a fast Fourier transform (FFT) on the standard rotation state data to construct a rotation state spectrogram. In the rotation state spectrogram, the frequency is used as the horizontal axis and the amplitude is used as the vertical axis; determine the unbalanced frequency domain feature vector from the rotation state spectrogram through a peak search algorithm. The unbalanced frequency domain feature vector generally includes data such as the unbalanced frequency, unbalanced amplitude, and unbalanced phase corresponding to the unbalanced vibration, which are used to describe the unbalanced state of the chip during rotation;

[0059] An unbalance analysis module is used to send the unbalanced frequency domain feature vector into an unbalance analysis model for processing and output an unbalanced state vector. The unbalanced state vector includes the magnitude of the unbalanced mass and the angle of the unbalanced mass. The unbalance analysis model is established based on a BP neural network. The unbalance analysis model realizes the mapping from the unbalanced amplitude to the magnitude of the unbalanced mass and the mapping from the unbalanced phase to the angle of the unbalanced mass; the magnitude of the unbalanced mass here is the mass corresponding to the shift of the center of gravity of the chip, and the angle of the unbalanced mass is the angle corresponding to the shift of the center of gravity of the chip;

[0060] The dynamic balance correction module is used to send the unbalanced state vector into the dynamic balance correction model for processing and output a set of dynamic balance correction parameters. The set of dynamic balance correction parameters includes the control parameters of all IPMC array units. Here, the control parameter is the driving voltage corresponding to the IPMC array unit. The IPMC thin film layer is controlled through the set of dynamic balance correction parameters to achieve the dynamic balance correction of the chip. It should be noted that in the process of controlling the IPMC thin film layer through the set of dynamic balance correction parameters, the driving voltages of all IPMC array units are actually adjusted, so that each IPMC array unit generates a corresponding deformation, and then the mass distribution of the entire chip is finely adjusted to complete the correction of the chip's dynamic balance; the dynamic balance correction model is established based on the BP neural network;

[0061] The reinforcement learning module is used to perform reinforcement learning on the dynamic balance correction model during the iteration of dynamic balance correction.

[0062] It should be understood that those of ordinary skill in the art can make improvements or transformations according to the above description, and all such improvements and transformations should fall within the protection scope of the appended claims of the present invention. The parts not detailedly described in this specification belong to the prior art well known to those of ordinary skill in the art.

Claims

1. A calibration method for chip dynamic balance testing, characterized in that Including: Attach an IPMC thin film layer on the chip, and several IPMC array units are included on the IPMC thin film layer; Obtain the rotation state data when the chip rotates through a gyroscope and an accelerometer, preprocess the rotation state data to obtain standard rotation state data, then perform a fast Fourier transform on the standard rotation state data to construct a rotation state frequency spectrum diagram. In the rotation state frequency spectrum diagram, the frequency is used as the horizontal axis and the amplitude is used as the vertical axis. Determine the unbalance frequency domain feature vector from the rotation state frequency spectrum diagram through a peak search algorithm; Send the unbalance frequency domain feature vector into an unbalance amount analysis model for processing, output an unbalance state vector, send the unbalance state vector into a dynamic balance correction model for processing, output a set of dynamic balance correction parameters. The set of dynamic balance correction parameters includes the control parameters of all IPMC array units, and control the IPMC thin film layer through the set of dynamic balance correction parameters to achieve dynamic balance correction of the chip; Repeat the dynamic balance correction of the chip, and the dynamic balance correction model is adjusted through reinforcement learning; stop the dynamic balance correction of the chip until the peak search algorithm cannot determine the unbalance frequency domain feature vector; Determine the unbalance frequency domain feature vector from the rotation state frequency spectrum diagram through a peak search algorithm, specifically including the following steps: Arrange all the frequencies in the rotation state frequency spectrum diagram in ascending order to construct a rotation state frequency set, and divide the rotation state frequency set through a sliding window to obtain several rotation state frequency set segments. The size of the sliding window is C and the step size is S; For each rotation state frequency set segment, perform the following: select the frequency with the largest amplitude, the corresponding amplitude and the corresponding phase in the rotation state frequency set segment to form a candidate feature vector; For each candidate feature vector, record the frequency corresponding to the candidate feature vector as the candidate frequency. Based on the candidate frequency, traverse the frequencies to the left in the rotation state frequency spectrum diagram until a frequency with an amplitude higher than the amplitude E corresponding to the candidate frequency is queried or the boundary of the rotation state frequency spectrum diagram is reached. Record the lowest amplitude during the left traversal as the left lowest amplitude L. Traverse the frequencies to the right in the rotation state frequency spectrum diagram until a frequency with an amplitude higher than the amplitude corresponding to the candidate frequency is queried or the boundary of the rotation state frequency spectrum diagram is reached. Record the lowest amplitude during the right traversal as the right lowest amplitude R. Calculate the significance P = E - max(L, R), and compare the significance P corresponding to the candidate feature vector with the significance threshold T. If the significance P corresponding to the candidate feature vector is higher than the significance threshold T, send the candidate feature vector into the candidate set; if the significance P corresponding to the candidate feature vector is not higher than the significance threshold T, do not perform any operation and consider it as unable to determine the unbalance frequency domain feature vector. The significance threshold T is adjusted according to the noise floor level corresponding to the rotation state frequency spectrum diagram; Traverse the candidate set. For each selected candidate feature vector, denote the candidate frequency corresponding to the selected candidate feature vector as f, and determine whether "f ∈ (kg - μ, kg + μ)" holds, where k is a positive integer, g is the rotation frequency of the chip, and μ is the tolerance. If "f ∈ (kg - μ, kg + μ)" holds, denote the selected candidate feature vector as the unbalanced frequency domain feature vector; if "f ∈ (kg - μ, kg + μ)" does not hold, no operation is performed and it is regarded as unable to determine the unbalanced frequency domain feature vector.

2. The calibration method for chip dynamic balance testing according to claim 1, wherein, Adjust the significance threshold T according to the noise floor level corresponding to the rotation state spectrogram, which specifically includes the following steps: Denote the average value of the amplitudes corresponding to all frequencies except the candidate frequency in the rotation state spectrogram as the noise floor level m, and adjust the significance threshold T through the formula T = U0exp(α(m - H0)), where U0 is the initial value of the significance threshold, α is the adjustment coefficient, and H0 is the initial value of the noise level.

3. The calibration method for chip dynamic balance testing according to claim 2, wherein Adjust the dynamic balance correction model through reinforcement learning, which specifically includes the following steps: Send the unbalanced state vector into the target dynamic balance correction model for processing, output the target dynamic balance correction parameter set, then send the unbalanced state vector and the target dynamic balance correction parameter set into the dynamic balance correction parameter evaluation network for processing, output the dynamic balance correction parameter evaluation value, calculate the policy value based on the dynamic balance correction parameter evaluation value for the unbalanced state vector, and denote the difference between the unbalanced amplitude corresponding to the current unbalanced frequency domain feature vector and the unbalanced amplitude corresponding to the previous unbalanced frequency domain feature vector as the reward value. Denote the sum of the policy value and the reward value as the gradient value. Send the unbalanced state vector into the dynamic balance correction model for processing, and adjust the dynamic balance correction model based on the gradient value through the gradient ascent method in the direction of maximizing the gradient value to achieve the adjustment of the dynamic balance correction model; the initial state of the target dynamic balance correction model is the same as that of the dynamic balance correction model. After each adjustment of the dynamic balance correction model, adjust the target dynamic balance correction model.

4. A calibration method for chip dynamic balance testing according to claim 3, characterized in that, Adjust the target dynamic balance correction model, which specifically includes the following steps: Adjust the parameters of the target dynamic balance correction model through the following formula: θ(new) = θ(bef) + βb, where θ(new) is the parameter of the target dynamic balance correction model after adjustment, θ(bef) is the parameter of the target dynamic balance correction model before adjustment, β is the learning rate, and b is the parameter of the dynamic balance correction model, and θ(new), θ(bef), and b correspond to the same parameter item.

5. A calibration method for chip dynamic balance testing according to claim 4, characterized in that, Train the unbalance amount analysis model, which specifically includes the following steps: Obtain the unbalance amount analysis training samples. The unbalance amount analysis training samples include unbalanced frequency domain feature vectors and corresponding unbalanced state vectors. Combine all unbalance amount analysis training samples into an unbalance amount analysis training set, and train the unbalance amount analysis model through the unbalance amount analysis training set. The training target is the unbalanced state vector in the unbalance amount analysis training samples.

6. A calibration method for chip dynamic balance testing according to claim 5, characterized in that, Train the dynamic balance correction model and the dynamic balance correction parameter evaluation network, which specifically includes the following steps: Obtain dynamic balance correction training samples. The dynamic balance correction training samples include unbalance state vectors and corresponding dynamic balance correction parameter sets. Combine all the dynamic balance correction training samples to form a dynamic balance correction training set, and use the dynamic balance correction training set to train the dynamic balance correction model. The training objective is the dynamic balance correction parameter set in the dynamic balance correction training samples; Obtain dynamic balance correction parameter evaluation training samples. The dynamic balance correction parameter evaluation training samples include unbalance state vectors and corresponding dynamic balance correction parameter sets. Label the dynamic balance correction parameter evaluation training samples through policy values. Combine all the labeled dynamic balance correction parameter evaluation training samples to form a dynamic balance correction parameter evaluation training set, and use the dynamic balance correction parameter evaluation training set to train the dynamic balance correction parameter evaluation network. The training objective is the labeled policy values.

7. A calibration system for chip dynamic balance testing, characterized in that, The system applies the calibration method for chip dynamic balance testing described in any one of claims 1-6 above, including: An unbalance frequency domain feature vector construction module, which is used to obtain the rotation state data of the chip during rotation through a gyroscope and an accelerometer, preprocess the rotation state data to obtain standard rotation state data, then perform a fast Fourier transform on the standard rotation state data to construct a rotation state frequency spectrum diagram. In the rotation state frequency spectrum diagram, with frequency as the horizontal axis and amplitude as the vertical axis, determine the unbalance frequency domain feature vector from the rotation state frequency spectrum diagram through a peak search algorithm; An unbalance amount analysis module, which is used to send the unbalance frequency domain feature vector into the unbalance amount analysis model for processing and output the unbalance state vector; A dynamic balance correction module, which is used to send the unbalance state vector into the dynamic balance correction model for processing and output a dynamic balance correction parameter set. The dynamic balance correction parameter set includes the control parameters of all IPMC array units. Control the IPMC thin film layer through the dynamic balance correction parameter set to achieve dynamic balance correction of the chip; A reinforcement learning module, which is used to perform reinforcement learning on the dynamic balance correction model during the iteration of dynamic balance correction.

Images