Torsion test method based on Kalman filtering algorithm

Abstract

The invention discloses a torsion test method based on a Kalman filtering algorithm, and relates to the technical field of torsion test.According to the torsion test method, a dynamic noise model and a material feature library are combined to enable a filter to automatically adapt to acoustic characteristics of different substrates, and the problem of misjudgment caused by fluctuation of material parameters in a traditional method is solved; secondly, a sub-band processing mechanism can restrain high-frequency noise and meanwhile completely reserve transient characteristics before the material is broken, and effective signal loss is avoided; the transfer learning model eliminates batch differences through adversarial training, and the generalization ability of dynamic threshold judgment is ensured; the real-time compensation strategy generates a refastening scheme considering both mechanical safety and assembly precision in emergency.

Description

Technical Field

[0001] The present invention relates to the technical field of torque testing, and in particular to a torque testing method based on a Kalman filter algorithm. Background Art

[0002] Currently, the assembly of microelectronic components is evolving towards ultra-precision and the integration of heterogeneous materials. Brittle substrates such as aluminum nitride ceramics and ultra-thin glass are widely used in fields such as smart wearable devices and micro-medical devices. In such scenarios, the tightening torque of M0.6-class micro screws often needs to be controlled at the order of 10-50mN·m, equivalent to an accuracy range of one percent of the pinching force of a human finger.

[0003] Although the recently emerging micro-servo electric screwdriver system can achieve a resolution of 0.1mN·m, its torque detection module still faces fundamental challenges when dealing with brittle materials: the critical difference between the material's fracture strength and the optimal tightening torque is approaching the sensor's noise tolerance, and traditional signal processing methods based on time-domain averaging are prone to misjudgment.

[0004] Current solutions mainly focus on two dimensions: on the one hand, some manufacturers have introduced deep learning models, such as LSTM time series prediction, to identify torque curve characteristics in real time. However, in the early stage of micro-screw insertion, the non-stationary noise caused by debris accumulation will cause the network weights to oscillate violently. On the other hand, resonance suppression algorithms based on frequency domain analysis, such as wavelet threshold denoising, can weaken high-frequency interference but will erase the transient mutation characteristics before the material breaks. Such characteristics often only last for 2-5ms and the frequency band overlaps with the noise. More seriously, existing systems generally rely on preset material parameter libraries. In actual production, dynamic changes within ±15% of parameters such as substrate thickness and sintering density will directly lead to pre-trained model mismatch. Therefore, a torque testing method is urgently needed to solve such problems. Summary of the Invention

[0005] In view of the above existing problems, the present invention is proposed.

[0006] The present invention provides a torque testing method based on the Kalman filter algorithm to solve the problems of noise interference and dynamic parameter mismatch faced by micro-brittle material fastening, and the existing filtering method is prone to losing key transient characteristics and leading to misjudgment.

[0007] In order to solve the above technical problems, the present invention provides the following technical solutions:

[0008] The embodiment of the present invention provides a torque testing method based on a Kalman filter algorithm, which includes:

[0009] Step S1, obtaining the original torque signal during the screw-in process in real time through a micro torque sensor;

[0010] Step S2, dynamically adjusting the noise covariance matrix parameters of the Kalman filter according to the material feature library of the currently assembled substrate;

[0011] Step S3, performing wavelet transform on the original torque signal to extract sub-signals containing at least a low frequency band, a mid-frequency band, and a high frequency band;

[0012] Step S4: constructing independent Kalman filter channels for each frequency band sub-signal, wherein a dynamic notch operator is introduced into the high frequency band channel;

[0013] Step S5: reconstructing the multi-channel filtering results and inputting them into the transfer learning model to match the optimal torque judgment threshold of the current material in real time;

[0014] Step S6: When the reconstructed signal exceeds the judgment threshold, an emergency stop command is triggered and a compensation torque recommendation value is output.

[0015] As a preferred embodiment of the torque testing method based on the Kalman filter algorithm described in the present invention, the dynamic adjustment of the noise covariance matrix parameters in step S2 includes:

[0016] Based on the offline established mapping relationship library between the acoustic emission spectrum characteristics of the substrate material and the optimal filtering parameters, the filtering parameters at least include the process noise covariance matrix and the measurement noise covariance matrix, similarity matching is performed through the online collected acoustic emission signal spectrum characteristics, and the corresponding initial noise covariance matrix is ​​automatically loaded;

[0017] During the filtering process, the process noise covariance matrix is ​​updated in real time using an exponentially weighted moving average algorithm based on the statistical characteristics of the signal residual.

[0018] The similarity matching specifically includes:

[0019] Extract the Mel-frequency cepstral coefficients of the acoustic emission signal in the 50-150kHz frequency band as the feature vector;

[0020] The cosine similarity algorithm is used to calculate the similarity value between the current feature and each sample in the material library;

[0021] When the maximum similarity value is lower than a preset threshold, the parameter interpolation mode is started to generate transition parameters.

[0022] As a preferred solution of the torque testing method based on the Kalman filter algorithm described in the present invention, in step S2, the material feature library is constructed as follows:

[0023] Collect acoustic emission signals and corresponding fracture torque data of at least aluminum nitride ceramics, ultra-thin glass, and silicon-based composite materials;

[0024] The material characteristics are divided into several subclasses through spectral clustering algorithm;

[0025] Each type of material is associated with a set of Kalman filter parameter combinations and safe torque boundary conditions that have been optimized by particle swarm optimization.

[0026] As a preferred embodiment of the torque testing method based on the Kalman filter algorithm described in the present invention, in step S2, a residual-driven exponential weighting strategy is used to dynamically correct the process noise covariance matrix so that the filter maintains adaptability in a multi-material assembly environment. The update formula is:

[0027]

[0028] Among them, Q k+1 represents the process noise covariance matrix at discrete time k+1, α k is the exponential decay coefficient of the current time step k, Q k is the process noise covariance matrix at time step k, r k is the filter residual vector, For r k The transposed matrix of

[0029] The exponential decay coefficient is adaptively adjusted according to the residual energy ratio:

[0030]

[0031] Among them, β is the smoothing adjustment coefficient, ζ k is the residual energy ratio, γ is the threshold constant;

[0032] The residual energy ratio is defined as:

[0033]

[0034] Among them, ε k is the instantaneous residual energy, is the exponentially weighted residual energy mean;

[0035] Exponentially weighted residual energy mean recursion:

[0036]

[0037] Among them, ρ is the residual energy smoothing coefficient, represents the exponentially weighted residual energy mean of the previous time step;

[0038] Residual vector source: Among them, z k is the observation vector, H is the observation matrix, is the state prediction vector for time step k.

[0039] As a preferred solution of the torque testing method based on the Kalman filter algorithm described in the present invention, the frequency band processing in step S3 is specifically as follows:

[0040] The db4 wavelet basis is used to perform stationary wavelet transform and decompose the signal into three layers of detail coefficients and approximate coefficients;

[0041] The pre-trained frequency band recognition model is used to determine whether each sub-band belongs to the noise-dominated frequency band;

[0042] Applying adaptive notch filtering to the frequency band identified as noise-dominated, preserving its phase information;

[0043] The frequency band identification model is constructed as follows:

[0044] Collect a noise sample library containing metal debris, ceramic fragments, and plastic debris;

[0045] Extract the wavelet packet energy distribution characteristics of each sample as training data;

[0046] Supervised training is performed using a one-dimensional convolutional neural network classifier.

[0047] As a preferred embodiment of the torque testing method based on the Kalman filter algorithm of the present invention, the high-frequency channel processing in step S4 further includes:

[0048] In the Kalman filter prediction stage, controllable white noise is injected, and the resonance peak frequency is shifted by adjusting the noise intensity;

[0049] A sliding time window is used to calculate the signal kurtosis index, and it automatically switches to the robust gain matrix when a transient pulse is detected.

[0050] The controllable white noise injection method is:

[0051] Superimpose a Gaussian white noise sequence on the state variables of the Kalman filter prediction equation;

[0052] The noise intensity is proportional to the derivative of the current signal amplitude;

[0053] Automatically resets the noise generator when a resonant frequency shift exceeding 10% is detected.

[0054] As a preferred embodiment of the torque test method based on the Kalman filter algorithm described in the present invention, in step S4, in the process of calculating the signal kurtosis index using a sliding time window, a sliding window method is used to calculate the signal kurtosis and drive the gain switching:

[0055]

[0056] Among them, κ mis the kurtosis of the mth sliding window, L is the window length, s i is the i-th signal sample in the window, μ m is the window mean, and the constant 3 is used to align with the normal distribution benchmark;

[0057] The window mean is defined as:

[0058]

[0059] In order to reduce the amount of calculation, the central moment recursion is introduced:

[0060]

[0061] M p,m+1 =M p,m +(s m+L -μ m+1 ) p -(s m -μ m ) p ,

[0062]

[0063] Among them, M p,m is the p-order central moment of the m-th window, m is the window number, s m+L Indicates the new samples when the window moves right, s m represents the sample removed, μ m+1 is the new window mean.

[0064] As a preferred embodiment of the torque testing method based on the Kalman filter algorithm described in the present invention, the transfer learning model in step S5 includes:

[0065] The material feature extraction module is built based on a convolutional neural network, and its input is the time-frequency diagram of the wavelet reconstructed signal;

[0066] A cross-material domain adversarial training module to eliminate feature distribution differences between different substrate batches;

[0067] The output layer generates dynamic torque thresholds and confidence assessment values.

[0068] As a preferred solution of the torque testing method based on the Kalman filter algorithm described in the present invention, the method further includes the step of generating a compensation strategy:

[0069] When the emergency stop command is triggered, the optimal torque path is reversely deduced based on the historical trajectory of the current filtered state variable;

[0070] The torque compensation curve is generated by combining the screw penetration depth prediction model;

[0071] The output includes a control parameter package including compensation torque value, re-tightening angle and recommended speed.

[0072] As a preferred solution of the torque test method based on the Kalman filter algorithm described in the present invention, in the compensation strategy generation step, in order to provide a feasible and safe re-tightening solution after triggering the emergency stop, according to the filter state history before time T Reverse deduction of an optimal torque path The steps include:

[0073] First, perform Rauch-Tung-Striebel smoothing on the filtered trajectory:

[0074]

[0075] in, represents the state vector at step t after smoothing, is the filter estimate, represents the smoothed state vector at step t+1, is the one-step prediction state, A t is the smoothed gain matrix;

[0076] The smoothed gain matrix is ​​defined as:

[0077]

[0078] Among them, P t is the filtered covariance, F is the state transfer matrix, is the one-step forecast covariance;

[0079] The smoothed torque component As a benchmark, construct a secondary objective including tracking error, depth error and smoothness:

[0080]

[0081] Among them, J is the cost function, τ t is the torque to be optimized, is the smooth torque reference, d t is the real-time screw depth prediction value, d T is the depth at the time of triggering, are the first-order and second-order differences of torque, Δt is the sampling period, w1, w2, w3 are weight coefficients, acting on torque tracking, depth error and smoothness terms respectively;

[0082] Weight adaptation definition:

[0083]

[0084] Among them, ξ and η are constant coefficients, cT The confidence level output by the transfer learning model;

[0085] Minimize the discrete target J so that:

[0086]

[0087] The torque optimality condition in differential form is obtained:

[0088]

[0089] in, Represents the gradient of the tracking error term, and sets the boundary Recursively from t = T to 0, we can obtain:

[0090]

[0091] in, is the previous moment value of the optimal torque sequence, Its first-order difference, symbol (·) ★ represents the optimal solution, initial value Get the real-time torque when triggering;

[0092] If the final depth deviation exceeds the limit after reverse deduction, use linear correction:

[0093]

[0094] in, is the corrected torque, d0 is the starting depth, and the correction makes the compensation curve meet the depth constraint at the end of the path.

[0095] The beneficial effects of the present invention are as follows: the combination of the dynamic noise model of the present invention and the material feature library enables the filter to automatically adapt to the acoustic characteristics of different substrates, overcoming the misjudgment problem caused by material parameter fluctuations in traditional methods; secondly, the frequency band processing mechanism suppresses high-frequency noise while completely retaining the transient characteristics of the material before rupture, avoiding the loss of effective signals; the transfer learning model eliminates batch differences through adversarial training, ensuring the generalization ability of dynamic threshold judgment; the real-time compensation strategy generates a re-tightening plan that takes into account both mechanical safety and assembly accuracy in emergency situations. BRIEF DESCRIPTION OF THE DRAWINGS

[0096] In order to more clearly illustrate the technical solutions of the embodiments of the present invention, the following briefly introduces the drawings required for use in the description of the embodiments. Obviously, the drawings described below are only some embodiments of the present invention. For ordinary technicians in this field, other drawings can be obtained based on these drawings without paying any creative work.

[0097] Figure 1Schematic diagram of the process of a torque testing method based on the Kalman filter algorithm in Example 1. DETAILED DESCRIPTION

[0098] In order to make the above-mentioned objects, features and advantages of the present invention more obvious and easy to understand, the specific embodiments of the present invention are described in detail below with reference to the accompanying drawings.

[0099] In the following description, many specific details are set forth to facilitate a full understanding of the present invention. However, the present invention may also be implemented in other ways different from those described herein. Those skilled in the art may make similar generalizations without violating the connotation of the present invention. Therefore, the present invention is not limited to the specific embodiments disclosed below.

[0100] Secondly, the term "one embodiment" or "embodiment" herein refers to a specific feature, structure, or characteristic that may be included in at least one implementation of the present invention. The phrase "in one embodiment" appearing in various places throughout this specification does not necessarily refer to the same embodiment, nor does it refer to a separate or selective embodiment that is mutually exclusive of other embodiments.

[0101] Example 1, with reference to Figure 1 , this embodiment provides a torque testing method based on the Kalman filter algorithm, comprising the following steps:

[0102] Step S1, obtaining the original torque signal during the screw-in process in real time through a micro torque sensor;

[0103] Step S2, dynamically adjusting the noise covariance matrix parameters of the Kalman filter according to the material feature library of the currently assembled substrate;

[0104] Dynamically adjusting the noise covariance matrix parameters in step S2 includes:

[0105] Based on the offline mapping relationship library between the acoustic emission spectrum characteristics of the substrate material and the optimal filtering parameters, the filtering parameters at least include the process noise covariance matrix and the measurement noise covariance matrix. Similarity matching is performed through the spectral characteristics of the acoustic emission signal collected online, and the corresponding initial noise covariance matrix is ​​automatically loaded;

[0106] During the filtering process, the process noise covariance matrix is ​​updated in real time using an exponentially weighted moving average algorithm based on the statistical characteristics of the signal residual.

[0107] Similarity matching specifically includes:

[0108] Extract the Mel-frequency cepstral coefficients of the acoustic emission signal in the 50-150kHz frequency band as the feature vector;

[0109] The cosine similarity algorithm is used to calculate the similarity value between the current feature and each sample in the material library;

[0110] When the maximum similarity value is lower than a preset threshold, the parameter interpolation mode is started to generate transition parameters;

[0111] In step S2, the material feature library is constructed as follows:

[0112] Collect acoustic emission signals and corresponding fracture torque data of at least aluminum nitride ceramics, ultra-thin glass, and silicon-based composite materials;

[0113] The material characteristics are divided into several subclasses through spectral clustering algorithm;

[0114] Each type of material is associated with a set of Kalman filter parameter combinations and safety torque boundary conditions that have been optimized by particle swarm optimization;

[0115] In step S2, the residual-driven exponential weighting strategy is used to dynamically modify the process noise covariance matrix so that the filter maintains adaptability in a multi-material assembly environment. The update formula is:

[0116]

[0117] Among them, Q k+1 represents the process noise covariance matrix at discrete time k+1, α k is the exponential decay coefficient of the current time step k, Q k is the process noise covariance matrix at time step k, r k is the filter residual vector, For r k The transposed matrix of

[0118] The exponential decay coefficient is adaptively adjusted according to the residual energy ratio:

[0119]

[0120] Among them, β is the smoothing adjustment coefficient, ζ k is the residual energy ratio, γ is the threshold constant;

[0121] The residual energy ratio is defined as:

[0122]

[0123] Among them, ε k is the instantaneous residual energy, is the exponentially weighted residual energy mean;

[0124] Exponentially weighted residual energy mean recursion:

[0125]

[0126] Among them, ρ is the residual energy smoothing coefficient, represents the exponentially weighted residual energy mean of the previous time step;

[0127] Residual vector source: Among them, z k is the observation vector, H is the observation matrix, is the state prediction vector for time step k;

[0128] Specifically, the update formula projects the residual energy into the covariance space and uses the adaptive α k Control matrix expansion rate, when the working condition suddenly causes ε k When the surge k Rising makes α k Reduce, introduce larger Weights, covariances are quickly relaxed, and the filter retains sensitivity to anomalies. When the system stabilizes, ε k and Close, α k Approximate to 1, matrix regression smoothing, avoid gain oscillation, exponential mean suppresses isolated spikes, ensures the continuity of the adjustment process, and takes into account both fast response and stable convergence;

[0129] Step S3, performing wavelet transform on the original torque signal to extract sub-signals containing at least a low frequency band, a mid-frequency band, and a high frequency band;

[0130] The frequency band division process in step S3 is specifically as follows:

[0131] The db4 wavelet basis is used to perform stationary wavelet transform and decompose the signal into three layers of detail coefficients and approximate coefficients;

[0132] The pre-trained frequency band recognition model is used to determine whether each sub-band belongs to the noise-dominated frequency band;

[0133] Applying adaptive notch filtering to the frequency band identified as noise-dominated, preserving its phase information;

[0134] The frequency band identification model is constructed as follows:

[0135] Collect a noise sample library containing metal debris, ceramic fragments, and plastic debris;

[0136] Extract the wavelet packet energy distribution characteristics of each sample as training data;

[0137] Supervised training using a one-dimensional convolutional neural network classifier;

[0138] Step S4: constructing independent Kalman filter channels for each frequency band sub-signal, wherein a dynamic notch operator is introduced into the high frequency band channel;

[0139] The high frequency channel processing in step S4 further includes:

[0140] In the Kalman filter prediction stage, controllable white noise is injected, and the resonance peak frequency is shifted by adjusting the noise intensity;

[0141] A sliding time window is used to calculate the signal kurtosis index, and it automatically switches to the robust gain matrix when a transient pulse is detected.

[0142] The controllable white noise injection method is:

[0143] Superimpose a Gaussian white noise sequence on the state variables of the Kalman filter prediction equation;

[0144] The noise intensity is proportional to the derivative of the current signal amplitude;

[0145] Automatically reset the noise generator when a resonant frequency shift exceeding 10% is detected;

[0146] In step S4, when calculating the signal kurtosis index using a sliding time window, a sliding window method is used to calculate the signal kurtosis and drive the gain switching:

[0147]

[0148] Among them, κ m is the kurtosis of the mth sliding window, L is the window length, s i is the i-th signal sample in the window, μ m is the window mean, and the constant 3 is used to align with the normal distribution benchmark;

[0149] The window mean is defined as:

[0150]

[0151] In order to reduce the amount of calculation, the central moment recursion is introduced:

[0152]

[0153] M p,m+1 =M p,m +(s m+L -μ m+1 ) p -(s m -μ m ) p ,

[0154]

[0155] Among them, M p,m is the p-order central moment of the m-th window, m is the window number, s m+L Indicates the new samples when the window moves right, s m represents the sample removed, μ m+1 is the new window mean;

[0156] Specifically, the kurtosis measures the fourth-order statistical characteristics and is extremely sensitive to high-amplitude spikes. The recursive method only needs to update two samples and the mean to complete the central moment iteration. The amount of computation scales linearly with the sampling rate, meeting the hard real-time constraints. The sliding window length is 3-5 times the main period of the resonance, which can take into account both transient capture and background stabilization. The real-time κ m After entering the high-frequency channel logic, when the value exceeds the set threshold, the robust gain matrix is ​​triggered. The filter responds to impact interference at the millisecond level, keeping the high-frequency information intact and providing reliable input for subsequent torque threshold judgment;

[0157] Step S5: reconstructing the multi-channel filtering results and inputting them into the transfer learning model to match the optimal torque judgment threshold of the current material in real time;

[0158] The transfer learning model in step S5 includes:

[0159] The material feature extraction module is built based on a convolutional neural network, and its input is the time-frequency diagram of the wavelet reconstructed signal;

[0160] A cross-material domain adversarial training module to eliminate feature distribution differences between different substrate batches;

[0161] The output layer generates dynamic torque thresholds and confidence assessment values;

[0162] Step S6, when the reconstructed signal exceeds the judgment threshold, an emergency stop command is triggered and a compensation torque recommendation value is output;

[0163] The torque test method also includes the compensation strategy generation step:

[0164] When the emergency stop command is triggered, the optimal torque path is reversely deduced based on the historical trajectory of the current filtered state variable;

[0165] The torque compensation curve is generated by combining the screw penetration depth prediction model;

[0166] Output control parameter package including compensation torque value, re-tightening angle and recommended speed;

[0167] In the compensation strategy generation step, in order to provide a feasible and safe re-tightening solution after triggering the emergency stop, the filtered state history before time T is used. Reverse deduction of an optimal torque path The steps include:

[0168] First, perform Rauch-Tung-Striebel smoothing on the filtered trajectory:

[0169]

[0170] in, represents the state vector at step t after smoothing, is the filter estimate, represents the smoothed state vector at step t+1, is the one-step prediction state, A t is the smoothed gain matrix;

[0171] The smoothed gain matrix is ​​defined as:

[0172]

[0173] Among them, P t is the filtered covariance, F is the state transfer matrix, is the one-step forecast covariance;

[0174] The smoothed torque component As a benchmark, construct a secondary objective including tracking error, depth error and smoothness:

[0175]

[0176] Among them, J is the cost function, τ t is the torque to be optimized, is the smooth torque reference, d t is the real-time screw depth prediction value, d T is the depth at the time of triggering, are the first-order and second-order differences of torque, Δt is the sampling period, w1, w2, w3 are weight coefficients, acting on torque tracking, depth error and smoothness terms respectively;

[0177] Weight adaptation definition:

[0178] w1=1+ξc T ,

[0179] Among them, ξ and η are constant coefficients, c T The confidence level output by the transfer learning model;

[0180] Minimize the discrete target J so that:

[0181]

[0182] The torque optimality condition in differential form is obtained:

[0183]

[0184] in, Represents the gradient of the tracking error term, and sets the boundary Recursively from t = R to 0, we can obtain:

[0185]

[0186] in, is the previous moment value of the optimal torque sequence, Its first-order difference, symbol (·) ★ represents the optimal solution, initial value Get the real-time torque when triggering;

[0187] If the final depth deviation exceeds the limit after reverse deduction, use linear correction:

[0188]

[0189] in, is the torque after correction, d0 is the starting depth, and the correction makes the compensation curve meet the depth constraint at the end of the path;

[0190] Specifically, a smoother is first used to suppress state estimation noise to ensure the continuity of subsequent optimization benchmarks. Then, torque tracking, depth maintenance, and curve smoothing are unified into a single objective function. The weights are adjusted automatically with the confidence level to achieve robust and delicate control priority allocation. The recursive formula derived from Euler-Lagrange only involves local gradient calculation and constant coefficient update, which is suitable for real-time reverse stepping. If there is a systematic deviation in the final depth, the linear correction term can fine-tune the full-segment torque without destroying the curve smoothness, thereby meeting both mechanical safety and assembly geometry requirements.

[0191] It should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention and are not intended to limit the present invention. Although the present invention has been described in detail with reference to the preferred embodiments, those skilled in the art should understand that the technical solutions of the present invention may be modified or replaced by equivalents without departing from the spirit and scope of the technical solutions of the present invention, which should all be included in the scope of the claims of the present invention.

Claims

1. A torque testing method based on Kalman filter algorithm, characterized in that: include, Step S1, obtaining the original torque signal during the screw-in process in real time through a micro torque sensor; Step S2, dynamically adjusting the noise covariance matrix parameters of the Kalman filter according to the material feature library of the currently assembled substrate; Step S3, performing wavelet transform on the original torque signal to extract sub-signals containing at least a low frequency band, a mid-frequency band, and a high frequency band; Step S4: constructing independent Kalman filter channels for each frequency band sub-signal, wherein a dynamic notch operator is introduced into the high frequency band channel; Step S5: reconstructing the multi-channel filtering results and inputting them into the transfer learning model to match the optimal torque judgment threshold of the current material in real time; Step S6: When the reconstructed signal exceeds the judgment threshold, an emergency stop command is triggered and a compensation torque recommendation value is output.

2. A torque testing method based on Kalman filter algorithm as claimed in claim 1, characterized in that: Dynamically adjusting the noise covariance matrix parameters in step S2 includes: Based on the offline established mapping relationship library between the acoustic emission spectrum characteristics of the substrate material and the optimal filtering parameters, the filtering parameters at least include the process noise covariance matrix and the measurement noise covariance matrix, similarity matching is performed through the online collected acoustic emission signal spectrum characteristics, and the corresponding initial noise covariance matrix is ​​automatically loaded; During the filtering process, the process noise covariance matrix is ​​updated in real time using an exponentially weighted moving average algorithm based on the statistical characteristics of the signal residual. The similarity matching specifically includes: Extract the Mel-frequency cepstral coefficients of the acoustic emission signal in the 50-150kHz frequency band as the feature vector; The cosine similarity algorithm is used to calculate the similarity value between the current feature and each sample in the material library; When the maximum similarity value is lower than a preset threshold, the parameter interpolation mode is started to generate transition parameters.

3. A torque testing method based on Kalman filter algorithm as claimed in claim 2, characterized in that: In step S2, the material feature library is constructed as follows: Collect acoustic emission signals and corresponding fracture torque data of at least aluminum nitride ceramics, ultra-thin glass, and silicon-based composite materials; The material characteristics are divided into several subclasses through spectral clustering algorithm; Each type of material is associated with a set of Kalman filter parameter combinations and safe torque boundary conditions that have been optimized by particle swarm optimization.

4. A torque testing method based on Kalman filter algorithm as claimed in claim 3, characterized in that: In step S2, the residual-driven exponential weighting strategy is used to dynamically modify the process noise covariance matrix so that the filter maintains adaptability in a multi-material assembly environment. The update formula is: Among them, Q k+1 represents the process noise covariance matrix at discrete time k+1, α k is the exponential decay coefficient of the current time step k, Q k is the process noise covariance matrix at time step k, r k is the filter residual vector, For r k The transposed matrix of The exponential decay coefficient is adaptively adjusted according to the residual energy ratio: Among them, β is the smoothing adjustment coefficient, ζ k is the residual energy ratio, γ is the threshold constant; The residual energy ratio is defined as: Among them, ε k is the instantaneous residual energy, is the exponentially weighted residual energy mean; Exponentially weighted residual energy mean recursion: Among them, ρ is the residual energy smoothing coefficient, represents the exponentially weighted residual energy mean of the previous time step; Residual vector source: Among them, z k is the observation vector, H is the observation matrix, is the state prediction vector for time step k.

5. The torque testing method based on Kalman filter algorithm according to claim 1, characterized in that: The frequency band division process in step S3 is specifically as follows: The db4 wavelet basis is used to perform stationary wavelet transform and decompose the signal into three layers of detail coefficients and approximate coefficients; The pre-trained frequency band recognition model is used to determine whether each sub-band belongs to the noise-dominated frequency band; Applying adaptive notch filtering to the frequency band identified as noise-dominated, preserving its phase information; The frequency band identification model is constructed as follows: Collect a noise sample library containing metal debris, ceramic fragments, and plastic debris; Extract the wavelet packet energy distribution characteristics of each sample as training data; Supervised training is performed using a one-dimensional convolutional neural network classifier.

6. The torque testing method based on the Kalman filter algorithm according to claim 1, characterized in that: The high frequency channel processing in step S4 further includes: In the Kalman filter prediction stage, controllable white noise is injected, and the resonance peak frequency is shifted by adjusting the noise intensity; A sliding time window is used to calculate the signal kurtosis index, and it automatically switches to the robust gain matrix when a transient pulse is detected. The controllable white noise injection method is: Superimpose a Gaussian white noise sequence on the state variables of the Kalman filter prediction equation; The noise intensity is proportional to the derivative of the current signal amplitude; Automatically resets the noise generator when a resonant frequency shift exceeding 10% is detected.

7. A torque testing method based on Kalman filter algorithm as claimed in claim 6, characterized in that: In step S4, when calculating the signal kurtosis index using a sliding time window, a sliding window method is used to calculate the signal kurtosis and drive the gain switching: Among them, κ m is the kurtosis of the mth sliding window, L is the window length, s i is the i-th signal sample in the window, μ m is the window mean, and the constant 3 is used to align with the normal distribution benchmark; The window mean is defined as: In order to reduce the amount of calculation, the central moment recursion is introduced: M p,m+1 =M p,m +(s m+L -m m+1 ) p -(s m -m m ) p , Among them, M p,m is the p-order central moment of the m-th window, m is the window number, s m+L Indicates the new samples when the window moves right, s m represents the sample removed, μ m+1 is the new window mean.

8. The torque testing method based on Kalman filter algorithm according to claim 1, characterized in that: The transfer learning model in step S5 includes: The material feature extraction module is built based on a convolutional neural network, and its input is the time-frequency diagram of the wavelet reconstructed signal; A cross-material domain adversarial training module to eliminate feature distribution differences between different substrate batches; The output layer generates dynamic torque thresholds and confidence assessment values.

9. The torque testing method based on Kalman filter algorithm according to claim 1, characterized in that: Also includes the compensation strategy generation step: When the emergency stop command is triggered, the optimal torque path is reversely deduced based on the historical trajectory of the current filtered state variable; The torque compensation curve is generated by combining the screw penetration depth prediction model; The output includes a control parameter package including compensation torque value, re-tightening angle and recommended speed.

10. A torque testing method based on Kalman filter algorithm as claimed in claim 9, characterized in that: In the compensation strategy generation step, in order to provide a feasible and safe re-tightening solution after triggering the emergency stop, the filtered state history before time T is used. Reverse deduction of an optimal torque path The steps include: First, perform Rauch-Tung-Striebel smoothing on the filtered trajectory: in, represents the state vector at step t after smoothing, is the filter estimate, represents the smoothed state vector at step t+1, is the one-step prediction state, A t is the smoothed gain matrix; The smoothed gain matrix is ​​defined as: Among them, P t is the filtered covariance, F is the state transfer matrix, is the one-step forecast covariance; The smoothed torque component As a benchmark, construct a secondary objective including tracking error, depth error and smoothness: Among them, J is the cost function, τ t is the torque to be optimized, is the smooth torque reference, d t is the real-time screw depth prediction value, d T is the depth at the time of triggering, are the first-order and second-order differences of torque, Δt is the sampling period, w1, w2, w3 are weight coefficients, acting on torque tracking, depth error and smoothness terms respectively; Weight adaptation definition: w1=1+ξc T , Among them, ξ and η are constant coefficients, c T The confidence level output by the transfer learning model; Minimize the discrete target J so that: The torque optimality condition in differential form is obtained: in, Represents the gradient of the tracking error term, and sets the boundary Recursively from t = T to 0, we can obtain: in, is the previous moment value of the optimal torque sequence, Its first-order difference, symbol (·) ★ represents the optimal solution, initial value Get the real-time torque when triggering; If the final depth deviation exceeds the limit after reverse deduction, use linear correction: in, is the corrected torque, d0 is the starting depth, and the correction makes the compensation curve meet the depth constraint at the end of the path.

Images