Intelligent calibration system and method based on dynamic error compensation and intelligent prediction fusion

Abstract

This invention belongs to the field of calibration technology, specifically disclosing an intelligent calibration system and method based on the fusion of dynamic error compensation and intelligent prediction. The system includes a dynamic error compensation module, an intelligent prediction module, and a fusion decision control module. The dynamic error compensation module incorporates a system transfer function model containing time-varying error components, used to acquire raw measurement signals in real time, calculate and subtract system dynamic errors online, and output a first-level compensation value. The intelligent prediction module outputs the predicted drift value and confidence level of the measured parameter's future drift. The fusion decision control module is connected to both the dynamic error compensation module and the intelligent prediction module, used to convert the drift prediction value into parameter correction values ​​for the system transfer function model. The dynamic error compensation module updates the parameters based on these correction values. This invention solves the problems of existing calibration instruments' static compensation mechanisms being unable to track dynamic errors in real time, lacking parameter drift prediction capabilities, and having insufficient automation in production line calibration.

Description

Technical Field

[0001] This invention belongs to the field of calibration technology, and in particular relates to an intelligent calibration system and method based on the fusion of dynamic error compensation and intelligent prediction. Background Technology

[0002] Quartz crystal frequency components, including crystal resonators, crystal oscillators, temperature-compensated crystals, and oven-controlled crystals, are the core frequency reference sources for electronic devices and are widely used in mobile communications, the Internet of Things, automotive electronics, server storage, medical equipment, and other fields. In the production process of frequency components, the precise measurement and calibration of parameters such as crystal frequency accuracy, equivalent series resistance, load capacitance, and excitation power are crucial steps in ensuring product quality and reliability.

[0003] However, the current testing and calibration of frequency components faces the following technical challenges: First, the static compensation mechanism of traditional calibration instruments has inherent limitations. Most existing testing instruments use factory-preset static correction algorithms, which cannot track dynamic errors introduced by changes in contact resistance, environmental temperature drift, and cable impedance fluctuations in real time. This leads to systematic deviations in measured values ​​during long-term batch testing. Specifically, during continuous production line testing for several hours, the contact resistance between the probe and the test piece gradually changes due to wear and oxidation, and fluctuations in ambient temperature also cause drift in the parameters of various components in the measurement chain. These time-varying errors cannot be eliminated through a single factory calibration.

[0004] Secondly, there is a lack of predictive assessment methods for device aging and temperature characteristics. Currently, the industry mainly relies on high and low temperature test chambers for sampling inspection, lacking intelligent methods to predict the long-term drift trend of devices online. When devices exhibit parameter deviations or performance degradation due to temperature changes or aging, they often need to be removed and sent for testing, which seriously affects production line efficiency and product yield.

[0005] Third, mass production line testing places higher demands on calibration efficiency and automation. Existing instruments require manual intervention or rely on standard parts for comparison during the calibration process, making it impossible to achieve fully automated adaptive calibration. Furthermore, there is a lack of real-time closed-loop communication between calibration data and intelligent analysis systems, which makes it difficult to support the data-driven decision-making needs of smart factories. Summary of the Invention

[0006] The purpose of this invention is to provide an intelligent calibration system and method based on the fusion of dynamic error compensation and intelligent prediction, so as to solve the problems that the static compensation mechanism of existing calibration instruments cannot track dynamic errors in real time, lacks parameter drift prediction capability, and has insufficient automation of production line calibration.

[0007] To achieve the above objectives, one technical solution of the present invention is: an intelligent calibration system based on the fusion of dynamic error compensation and intelligent prediction, comprising a dynamic error compensation module, an intelligent prediction module, and a fusion decision control module; The dynamic error compensation module has a built-in system transfer function model containing time-varying error components, which is used to acquire the original measurement signal in real time, calculate and subtract the system dynamic error online, and output a first-level compensation value. The intelligent prediction module takes historical compensation value sequence and synchronous environmental data as input and outputs the drift prediction value and confidence level of the future drift of the measured parameter. The fusion decision control module is connected to the dynamic error compensation module and the intelligent prediction module respectively. It is used to convert the drift prediction value output by the intelligent prediction module into the parameter correction amount of the system transfer function model in the dynamic error compensation module. The dynamic error compensation module updates the parameters according to the parameter correction amount, forming a dual-loop closed-loop calibration architecture.

[0008] Furthermore, the dynamic error compensation module calculates the time-varying parameter vector online using a recursive least squares algorithm. The time-varying parameter vector includes at least one of the temperature drift coefficient and the change in contact impedance.

[0009] Furthermore, the intelligent prediction module incorporates a trained parameter drift prediction model, which is a temporal-feature fusion neural network, including a temporal coding layer, an environmental attention layer, and a prediction output layer. The temporal coding layer is used to extract features from the input joint historical sequence and output a temporal latent state containing historical environmental features; the joint historical sequence consists of a historical compensation value sequence and synchronization environment data. The environment attention layer is used to calculate the attention weight between the current environment vector acquired in real time and the historical environment features output by the temporal coding layer, and to generate a context vector; The prediction output layer is used to fuse the temporal hidden state with the context vector and output the drift prediction value and confidence level.

[0010] Furthermore, the fusion decision control module calculates the parameter update amount based on the drift prediction value, confidence level, and sensitivity of the time-varying parameter vector to the compensation output, limits the parameter update amount through a saturation limiting function, and then uses the limited parameter update amount to update the time-varying parameter vector of the system transfer function model.

[0011] Furthermore, the intelligent prediction module can calculate the residual between the current primary compensation value and the current primary compensation reference value. The current primary compensation reference value is predicted based on the historical compensation value sequence and synchronous environmental data before the current time. When the residual exceeds the dynamic threshold, it is determined to be a calibration abnormality or device mutation, triggering an alarm and recording the complete data chain. The recorded data chain is used for subsequent incremental retraining of the intelligent prediction module to form an iteratively optimized anomaly tracing closed loop.

[0012] Furthermore, it also includes an automatic parameter type identification module; the automatic parameter type identification module outputs a scanning excitation signal to the device under test and collects the response characteristics of the device under test, identifies the current parameter type based on the response characteristics, and retrieves the system transfer function model parameters that are pre-stored in the system and match the parameter type.

[0013] Furthermore, it also includes a control host, an environmental monitoring unit, an electrical parameter calibrator, a triaxial positioning testing mechanism, a test probe, and a signal conditioning and control unit; the dynamic error compensation module, the intelligent prediction module, and the fusion decision control module are configured within the control host; the control host is electrically connected to the electrical parameter calibrator, the triaxial positioning testing mechanism, and the signal conditioning and control unit respectively; the test probe is electrically connected to the signal conditioning and control unit; the test probe is mounted on the triaxial positioning testing mechanism, which is used to move the test probe to contact the test piece and complete alignment; The electrical parameter calibrator outputs a standard excitation signal to a signal conditioning and control unit. The signal conditioning and control unit adapts the standard excitation signal and applies the conditioned excitation signal to the test probe. After the excitation signal is applied to the device under test through the test probe, a response signal is generated. The response signal is sent back to the signal conditioning and control unit for preprocessing. The signal conditioning and control unit sends the preprocessed signal to the electrical parameter calibrator. The electrical parameter calibrator uploads a timestamped measurement feedback signal to the control host. The dynamic error compensation module in the control host receives the measurement feedback signal and uses it as the original measurement signal. The signal conditioning and control unit is connected to the environmental detection unit and is used to collect environmental data in real time and upload it to the control host.

[0014] Another technical solution of the present invention is: an intelligent calibration method based on the fusion of dynamic error compensation and intelligent prediction, comprising the following steps: S1. Retrieve the initial baseline parameters of the dynamic error compensation module stored in the equipment's factory calibration, and load the locally pre-trained intelligent prediction model. S2. The dynamic error compensation module acquires the original measurement signal, calculates and subtracts the system dynamic error online, and outputs the first-level compensation value. S3. Synchronously collect environmental data, input the compensation value sequence and environmental data within the current time window into the intelligent prediction model, and output the drift prediction value and confidence level of the future drift of the measured parameter. S4. Update the parameters of the dynamic error compensation module based on the drift prediction value and confidence level; S5. Generate the final calibration result using the updated dynamic error compensation module.

[0015] The beneficial effects of this technical solution are as follows: First, this invention introduces a deep dual-loop fusion mechanism of a real-time dynamic compensation loop and an intelligent predictive optimization loop. The inner loop uses a system transfer function model that includes time-varying error components and a recursive least squares algorithm to suppress instantaneous errors such as contact resistance and temperature drift online; the outer loop analyzes historical drift trends through an intelligent prediction module to predict and proactively compensate for long-term errors such as device aging. The collaborative operation of the two loops achieves a technological leap from passive correction to active pre-compensation, improving calibration accuracy by more than an order of magnitude compared to traditional static compensation instruments.

[0016] Secondly, the intelligent prediction model automatically identifies the type of the test component, automatically matches calibration parameters, and automatically determines whether it is qualified or not, which greatly reduces the reliance on the professional experience of operators and is especially suitable for testing scenarios of multiple varieties and small batches of precision components.

[0017] Third, the intelligent prediction module can perform periodic incremental learning based on accumulated data, and the system can continuously optimize and evolve over long-term use.

[0018] Fourth, by setting an automatic parameter type identification module, the system can be expanded to cover the calibration and measurement of various electrical parameters such as resistance, current, voltage, capacitance and AC impedance without changing the hardware. Attached Figure Description

[0019] Figure 1 This is a schematic diagram of the intelligent calibration system based on the fusion of dynamic error compensation and intelligent prediction according to the present invention. Figure 2 This is a schematic diagram of a fully automatic electrical parameter calibrator. Figure 3 A schematic diagram of the dual-layer closed-loop fusion architecture of dynamic error compensation and intelligent prediction; Figure 4 This is a block diagram of the signal conditioning and control unit. Figure 5 This is a flowchart of the intelligent calibration method. Detailed Implementation

[0020] The following detailed description illustrates the specific implementation method: The reference numerals in the accompanying drawings include: 1. Workbench; 2. Electrical parameter calibrator; 3. Display screen; 4. Control host; 5. Test probe; 6. Triaxial positioning test mechanism; 7. Signal conditioning and control unit; 8. Insulation resistance tester.

[0021] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention. Example 1

[0022] The basics are as follows: Figure 1-4 As shown: The intelligent calibration system based on the fusion of dynamic error compensation and intelligent prediction includes a workbench 1, a control host 4 (specifically a computer host), a display screen 3, an environmental detection unit, an electrical parameter calibrator 2, a three-axis positioning test mechanism 6, a test stylus 5, and a signal conditioning and control unit 7; the dynamic error compensation module, the intelligent prediction module, and the fusion decision control module are configured in the control host 4.

[0023] Workbench 1 is used to support and install all other modules, providing a stable working platform.

[0024] The electrical parameter calibrator 2 (specifically, the fully automatic electrical parameter calibrator 2) is set on workbench 1. It serves as a standard source, outputting precise current, voltage, and multi-waveform dynamic signals. It incorporates a high-precision readback measurement module, environmental compensation algorithm, and self-test function, and is the core standard source and measurement unit of the system. The fully automatic electrical parameter calibrator 2 has a resistance measurement range of 0.01Ω to 10MΩ with an accuracy of ±0.01%FS; a current measurement range of 1μA to 10A with an accuracy of ±0.01%FS; an AC excitation frequency range of 10Hz to 100kHz; a sampling frequency of 1kHz; a dynamic compensation update rate of 1ms / time; and an intelligent coprocessor inference cycle of 100ms / time.

[0025] Display screen 3 is used to display the test interface, real-time data, and test reports.

[0026] The control host 4 is equipped with test control software, a dynamic error compensation module, an intelligent prediction module, and a fusion decision control module, which are responsible for the overall test process control and data processing. The control host 4 is electrically connected to the fully automatic electrical parameter calibrator 2, the three-axis positioning test mechanism 6, the signal conditioning and control unit 7, and the insulation resistance tester 8 (optionally).

[0027] Test probe 5 is used to contact the test point of the device under test (DUT) and can be automatically adjusted according to the spacing of the DUT. Test probe 5 is a gold-plated probe, and the contact force can be controlled within the range of 100gf±10gf. Test probe 5 is electrically connected to the signal conditioning and control unit 7.

[0028] The three-axis positioning testing mechanism 6 is mounted on the worktable 1 and is used to move the test probe 5 in three-dimensional space to achieve high-precision automatic alignment. The test probe 5 is mounted on the three-axis positioning testing mechanism 6, which is used to move the test probe 5 to contact the workpiece under test and complete the alignment. The X-axis and Y-axis travel of the three-axis positioning testing mechanism 6 is 300mm × 200mm, with a repeatability of ±0.01mm; the Z-axis travel is 50mm.

[0029] The signal conditioning and control unit 7 is electrically connected to the fully automatic electrical parameter calibrator 2 and the test probe 5, respectively, to achieve standard excitation signal adaptation, including level conversion, impedance matching, filtering, isolation, as well as multi-channel automatic routing switching, bidirectional signal transmission and closed-loop readback, and also has overload protection and timing synchronization control functions. The signal conditioning and control unit 7 is connected to the environmental detection unit to collect environmental data in real time and upload it to the control host 4.

[0030] The insulation resistance tester 8 is an optional module that is electrically connected to the signal conditioning and control unit 7 and is used to perform insulation resistance testing.

[0031] The fully automatic electrical parameter calibrator 2 outputs a standard excitation signal to the signal conditioning and control unit 7. The signal conditioning and control unit 7 adapts the standard excitation signal and applies the conditioned excitation signal to the test probe 5. After the excitation signal is applied to the device under test through the test probe 5, a response signal is generated. The response signal is sent back to the signal conditioning and control unit 7 for preprocessing. The signal conditioning and control unit 7 sends the preprocessed signal to the electrical parameter calibrator 2. The electrical parameter calibrator 2 uploads the measurement feedback signal with a timestamp to the control host 4. The dynamic error compensation module in the control host 4 receives the measurement feedback signal and uses it as the original measurement signal.

[0032] 1. Dynamic error compensation module The dynamic error compensation module has a built-in system transfer function model containing time-varying error components. It is used to acquire raw measurement signals in real time, and calculates and subtracts the inherent system error by solving the model online, and outputs the first-level compensation value.

[0033] The general form of this model can be expressed as:

[0034] The system transfer function model includes time-varying error components such as the calibration gain coefficient K0, the temperature drift correction term α·ΔT(t), and the contact impedance correction term β·ΔRc(t). Among these, P... comp (t) represents the parameter value after dynamic compensation (i.e., the first-level compensation value), Mraw(t) represents the original measurement signal reading, and δi(t) represents each known error component.

[0035] The dynamic error compensation module monitors the loop current and voltage in real time and uses a recursive least squares algorithm to estimate the time-varying parameter vector θ online. The time-varying parameter vector includes the temperature drift coefficient α(t) and the change in contact impedance β(t), etc. The recursive least squares algorithm is equipped with a forgetting factor λ, which ranges from 0.95 to 0.999, to control the rate of weight decay of historical data.

[0036] The recursive least squares algorithm is as follows:

[0037] in, Let be the time-varying parameter vector at time t. The time-varying parameter vector from the previous moment (previous sampling round), at startup. These are the initial parameters of the dynamic error compensation module (obtained from factory calibration). For residuals, Let φ(t) be the raw measurement signal actually acquired at time t, and let φ(t) be the regression vector (input feature vector), which is composed of synchronously acquired observed variables such as ambient temperature and excitation conditions. This is the model prediction output calculated using the model parameters from the previous time step. K(t) is the gain vector (Kalman gain / RLS gain), which determines... How much force should be used to correct the parameters? The larger K(t) is, the greater the modification of parameters by the new sampling data; the smaller K(t) is, the more trust is placed in historical data. P(t-1) is the covariance matrix, representing the parameter estimation. The uncertainty is denoted by P(0). The larger the diagonal, the less reliable the corresponding parameter estimate, allowing for a larger parameter update; the P matrix is ​​also updated synchronously with each iteration. P(0) is the factory-initialized covariance matrix, which is part of the initial parameters.

[0038] The process of the recursive least squares online estimation algorithm is as follows: First, initialize the parameter vector. The covariance matrix P(0) is calculated; then, at each sampling time, the input feature vector φ(t) and the measured output y(t) are acquired. The input feature vector includes temperature, frequency, excitation amplitude, etc.; next, the gain vector K(t) is calculated; and then the parameter vector is updated. Finally, update the covariance matrix P(t) and proceed to the next sampling period. The updated... It is the time-varying parameter vector of the system transfer function model, used to calculate the first-level compensation value.

[0039] 2. Intelligent Prediction Module The input of the intelligent prediction module is connected to the output of the dynamic error compensation module and the environmental detection unit (environmental sensor), and it has a built-in trained parameter drift prediction model. The input feature vector of the intelligent prediction module includes the first-level compensation value P. comp (t), temperature T(t), humidity H(t), excitation frequency f(t), excitation signal amplitude A exc (t) and the cumulative usage time of the device t age This can be represented as eigenvector = [P] comp (t),T(t),H(t),f(t),A exc (t),t age ].

[0040] The parameter drift prediction model is expressed as: ΔP(t+Δt)=FLSTM({P comp (tn:t),T(tn:t),f(tn:t),A exc (tn:t)}) Where ΔP(t+Δt) represents the predicted drift value of the future drift, indicating the amount of drift change that the measured parameter will undergo after time Δt from the current time t. FLSTM(•) is an improved LSTM network (Feature-Enhanced Long Short-Term Memory network). tn: t is the historical time window, taking historical time series data from tn to t (n+1 times); n is the length of the historical window, representing the prediction using the most recent n sets of historical data. P comp (tn:t) represents the parameter sequence after historical compensation, a time series consisting of the first-level compensation values ​​output by the system within the time period [tn~t]; reflecting the historical changes of component parameters over time. T(tn:t) is the temperature time series, the ambient temperature time series data synchronously collected within the same historical window; temperature is the most significant factor causing component drift. f(tn:t) is the excitation frequency time series, the frequency of the excitation signal applied during testing within the historical window; at high frequencies, crystal oscillator and capacitor parameters change with the excitation frequency. A exc (tn:t) represents the excitation amplitude timing sequence, where the excitation signal amplitude (voltage / current magnitude) is within the historical window. Excitation levels can cause device heating and nonlinear shift. All input sequences must be strictly time-aligned: parameters, temperature, frequency, and amplitude at the same moment must correspond one-to-one. The model uses multiple sets of synchronized historical timing sequences as input to predict future drift changes in the measured parameters and outputs the results to the fusion decision control module for outer loop pre-compensation correction.

[0041] The model uses the historical compensation value sequence {P}comp The input consists of the synchronous environmental data sequence {E}, and the output is the predicted value of the parameter drift ΔP(t+Δt) of the device under test at future time.

[0042] The parameter drift prediction model is a time-feature fusion neural network, and its forward propagation process includes: Temporal coding layer: For the input joint history sequence {P} comp Feature extraction is performed on E}, and the temporal latent state h containing historical environmental features is output. t The timing coding layer can employ a long short-term memory network or a gated recurrent unit structure.

[0043] Environment Attention Layer: Used to calculate the current environment vector e acquired in real time. t Attention weights are applied between the historical context features output by the temporal coding layer and the resulting context vector c. t The environmental attention layer enables the model to adaptively focus on historical environmental states most relevant to the current drift.

[0044] Predictive output layer: used to predict the temporal hidden state h t With context vector c t The data is fused and output through a fully connected network, showing the drift prediction value ΔP and the confidence level σ. The confidence level σ ranges from 0 to 1, reflecting the reliability of the prediction result.

[0045] 3. Integrated Decision Control Module The fusion decision control module is connected to both the dynamic error compensation module and the intelligent prediction module to execute the fusion algorithm. The fusion decision control module feeds back the drift prediction value ΔP to the dynamic error compensation module as the basis for online correction of its model parameters.

[0046] The parameterized form of the fusion algorithm is:

[0047] in, For the updated i-th time-varying parameter, The new value of the i-th component in the time-varying parameter vector after correction. For example, i=1 represents the temperature drift coefficient, and i=2 represents the contact impedance correction coefficient. is the current i-th time-varying parameter, which is the i-th component of the uncorrected current time-varying parameter vector output by the RLS recursive least squares algorithm, and is the inner loop output value. is the learning rate for the i-th parameter, ranging from 0.01 to 0.1. sat(•) is the saturation constraint function. The partial derivative of the measured output parameter with respect to the i-th time-varying parameter characterizes the parameter sensitivity. The sensitivity of the compensated output is determined by the predicted drift ΔP, the predicted confidence level σ, the learning rate γ, and the time-varying parameter vector. Calculate parameter update amount The parameter update amount is then limited using the saturation constraint function `sat(·)` to prevent over-adjustment. The time-varying parameter vector of the system transfer function model is then updated using the limited parameter update amount. ).

[0048] The time-varying parameters and sensitivities of the system transfer function model differ depending on the parameter type: For the resistance parameter, the time-varying parameter is the temperature coefficient α. R The sensitivity is R0•(T-T0), and the effect of temperature change on the resistance value is calculated.

[0049] For the current parameter, the time-varying parameter is the shunt coefficient β. I The sensitivity is -I0•(R) shunt -2 The shunt error is caused by the aging of the sampling resistor.

[0050] For voltage parameters, the time-varying parameter is the voltage divider ratio drift coefficient γ. V Sensitivity is V in •(R2 / (R1+R2) 2 ), the voltage divider resistor network is aging.

[0051] For the capacitance parameters, the time-varying parameter is the dielectric loss compensation coefficient δ. C The sensitivity is C0·tan(δ), and the capacity drift is caused by dielectric aging.

[0052] 4. Anomaly detection and source tracing closed loop The intelligent prediction module can calculate the first-level compensation value P at the current moment. comp (t) and the first-level compensation reference value P at the current time ， comp The residual between (t) and the current first-level compensation reference value is predicted based on the historical compensation value sequence before the current time and the synchronization environment data. P ， comp (t) is obtained by performing a one-step forward prediction using the FLSTM model: taking the time series of the first-level compensation value, temperature, excitation frequency, and excitation amplitude from the historical window (tn) to (t-1) before time t as the model input, the first-level compensation reference value P corresponding to the current time t is derived. ， comp (t), this prediction process does not use the actual first-level compensation value P at time t. comp (t).

[0053] When the residual exceeds the dynamic threshold, it is determined to be an abnormality in the calibration process or a sudden change in the device, triggering an alarm and recording the complete data chain of this event for subsequent model retraining, forming a self-evolving traceability loop. The dynamic threshold is adaptively adjusted based on the statistical characteristics (mean, variance, etc.) of the historical residual sequence, which can provide... , Let be the mean of the residual sequence. is the moving standard deviation of the residual sequence. Example 2

[0054] The intelligent calibration system based on the fusion of dynamic error compensation and intelligent prediction also includes an automatic parameter type identification module. The automatic parameter type identification module outputs a scanning excitation signal to the device under test and collects the response characteristics of the device under test. Based on the response characteristics, it identifies the current type of the measured parameter and retrieves the system transfer function model parameters that match the type of the measured parameter, as well as the corresponding calibration process in the control host 4.

[0055] The system supports calibration of various parameter types, including resistance, current, voltage, and capacitance. For different parameter types, the excitation signal, measurement method, and key error sources differ; the system achieves unified adaptation through hardware reuse and model generalization. For resistance parameters: a constant current source is used for excitation, and the voltage drop is measured. The key error sources are contact resistance and thermoelectric potential. The compensation model is R=(V H -V L ) / I, compensating for line resistance and thermoelectric potential.

[0056] For current parameters: a constant voltage source is used for excitation, and the voltage across the sampling resistor is measured. The key error sources are shunt error and mutual inductance. The compensation model is I=V. sense / R shunt , to compensate for frequency characteristics.

[0057] For voltage parameters: a direct measurement method is used, measuring the signal after voltage division. The key error sources are input bias and gain drift, and the compensation model is V=V in •G(T), compensates for partial voltage ratio drift.

[0058] For capacitor parameters: an AC constant current source is used for excitation, and the phase difference and amplitude ratio are measured. The key error sources are dielectric absorption and leakage current. The compensation model is C=I / (2πfV)·[1+εDA(f,T)] -1 , where εDA(f,T) is the medium absorption correction factor.

[0059] The automatic parameter type identification process is as follows: The system first outputs a scanning excitation signal to detect the response characteristics of the device under test, including the impedance amplitude and the phase variation with frequency; then it matches the detected response characteristics with the preset parameter type template; after successful matching, it automatically loads the initial parameters, calibration process and qualification criteria of the system transfer function model of the corresponding parameter type. Example 3

[0060] like Figure 5 As shown, the intelligent calibration method based on the fusion of dynamic error compensation and intelligent prediction includes the following steps: S1. Retrieve the initial reference parameters of the dynamic error compensation module stored in the equipment's factory calibration, and simultaneously load the locally pre-trained intelligent prediction model. The initial reference parameters include the initial values ​​of the calibration gain coefficient, temperature drift coefficient, contact impedance, and the forgetting factor of the recursive least squares algorithm. The pre-trained intelligent prediction model is obtained through offline training using historical test data, and the model parameters are encrypted and stored in the control host 4.

[0061] S2. The dynamic error compensation module acquires the original measurement signal, calculates and subtracts the system dynamic error online, and outputs the first-level compensation value. Specifically, the fully automatic electrical parameter calibrator 2 outputs a standard excitation signal, which is applied to the device under test (DUT) through the test probe 5 after being adapted by the signal conditioning and control unit 7. The four-wire measurement module of the fully automatic electrical parameter calibrator 2 synchronously acquires the voltage signal and loop current signal at both ends of the DUT. The dynamic error compensation module performs a recursive least squares online estimation algorithm for each sampling point, updates the time-varying parameters in real time, and calculates the compensated parameter value, i.e., the first-level compensation value.

[0062] S3. Synchronously collect environmental data, input the compensation value sequence and environmental data within the current time window into the intelligent prediction model, and output the drift prediction value and confidence level of the future drift of the measured parameter; environmental data includes temperature, humidity, excitation frequency, and excitation power, etc. The length of the time window can be adjusted according to the drift characteristics of the measured parameter, and usually includes tens to hundreds of sampling points.

[0063] S4. Update the parameters of the dynamic error compensation module based on the drift prediction value and confidence level. Specifically, the fusion decision control module calculates the parameter update amount based on the drift prediction value, prediction confidence level, learning rate, and sensitivity of the time-varying parameter vector to the compensation output. After limiting the update amount through the saturation constraint function, the updated parameters are sent to the dynamic error compensation module.

[0064] S5. The updated dynamic error compensation module generates the final calibration result. The control host 4 automatically determines whether the tested part is qualified according to the preset qualification criteria and displays the test result on the display screen 3.

[0065] S6. Calculate the first-level compensation value P at the current moment. comp (t) and the first-level compensation reference value P at the current time ， comp The residual between (t) is considered a calibration anomaly or device mutation when it exceeds a dynamic threshold, triggering an alarm and recording the complete data chain. Under normal circumstances, the complete calibration data is archived to the historical database for periodic incremental learning.

[0066] Steps S2 to S5 are continuously executed in a loop during the test, forming a dual-loop closed-loop calibration of dynamic compensation and intelligent prediction. The dynamic compensation loop operates on a millisecond-level cycle to achieve real-time error suppression; the intelligent prediction loop operates on a hundred-millisecond to second-level cycle to achieve drift trend prediction and pre-compensation. Example 4

[0067] The intelligent calibration system automatically archives complete data from each calibration, including timestamps, ambient temperature and humidity, dynamic compensation parameter trajectories, intelligent prediction records, and final test results, into a historical database.

[0068] The intelligent calibration system periodically initiates incremental learning tasks. For example, it automatically starts incremental learning once a week, using the accumulated test data of the week as new training samples to fine-tune and update the pre-trained parameter drift prediction model through online learning. The learning rate of incremental learning is lower than that of the initial training, and the number of training rounds is also reduced accordingly, in order to absorb new data features while retaining existing knowledge.

[0069] After the update is completed, the new model parameters are sent to the intelligent coprocessor of the fully automatic electrical parameter calibrator 2 via the internal communication interface, replacing the original model parameters and realizing the system's self-evolution. As usage time increases and data volume accumulates, the accuracy of the intelligent prediction model will continue to improve, and the calibration precision will also be continuously optimized.

[0070] Example 1 Taking the batch testing of quartz crystal resonators on a production line as an application scenario, the specific implementation process of this invention is described in detail: I. System Setup and Initial Configuration The main technical parameters of the fully automatic electrical parameter calibrator 2 are set as follows: Resistance measurement range: 0.01Ω-10MΩ, accuracy ±0.01%FS; Current measurement range: 1μA-10A, accuracy ±0.01%FS; AC excitation frequency range: 10Hz-100kHz; Sampling frequency: 1kHz; Dynamic compensation update rate: 1ms / time; Intelligent coprocessor inference cycle: 100ms / time.

[0071] The technical parameters of the three-axis positioning testing mechanism 6 are set as follows: X / Y axis travel: 300mm×200mm; repeatability: ±0.01mm; Z axis travel: 50mm; test probe 5 contact force: 100gf±10gf; single crystal test cycle target: ≤3 seconds.

[0072] II. Pre-training of Intelligent Prediction Model Before the system is officially put into use, the parameter drift prediction model is pre-trained on the control host 4 using more than 100,000 sets of crystal test data accumulated by the company in the past.

[0073] The training dataset includes: Input characteristics: Equivalent series resistance (ESR) sequence of the crystal after dynamic compensation, ambient temperature sequence, ambient humidity sequence, excitation frequency, and excitation power; Output labels: ESR drift and frequency deviation of the crystal at the 100th measurement in the future.

[0074] The model uses a three-layer GRU network structure with hidden layer dimensions of 64, 128, and 64, respectively. It is trained for 100 epochs with a learning rate of 0.001 and uses the Adam optimizer.

[0075] After pre-training is completed, the model parameters are encrypted and packaged, and then downloaded to the intelligent coprocessor of the fully automatic electrical parameter calibrator 2 via USB or Ethernet.

[0076] III. Batch Testing Workflow 1. Loading and positioning of the crystal under test The operator places a tray containing 100 12MHz crystals onto the test area of ​​the triaxial positioning test mechanism 6 on the workbench 1. The control host 4 sends positioning commands to the triaxial positioning test mechanism 6 through the signal conditioning and control unit 7.

[0077] The X-axis and Y-axis linear modules of the three-axis positioning testing mechanism 6 precisely move the test probe 5 above the first crystal to be tested according to the preset crystal tray coordinate mapping table (positioning error ≤ ±0.01mm). The Z-axis lifting module drives the test probe 5 to descend with a contact force of 100gf, ensuring reliable contact between the gold-plated probe of the test probe 5 and the crystal pin. After the contact resistance detection circuit in the signal conditioning and control unit 7 confirms that the contact resistance is ≤50mΩ, it sends a "contact complete" signal to the control host 4.

[0078] 2. Automatic parameter type recognition The control host 4 instructs the fully automatic electrical parameter calibrator 2 to output a scanning excitation signal (frequency range 11.8MHz-12.2MHz, power -10dBm) through the signal conditioning and control unit 7. The four-wire measurement module of the fully automatic electrical parameter calibrator 2 synchronously acquires the voltage response signal and loop current signal at both ends of the crystal, and the measurement module completes the scan within 200ms.

[0079] The control host 4 automatically determines the type of the device under test as "AT-cut quartz crystal resonator" based on the response spectrum characteristics. At the same time, it initially obtains that the series resonant frequency of the crystal is about 12.0005MHz and the equivalent series resistance ESR is about 40Ω.

[0080] The control host 4 automatically loads the initial parameters of the dynamic compensation model corresponding to the crystal resonator based on the identification results: K0=1.00000, α=1.5×10 -6 / ℃, R line =0.02Ω, β=0.01Ω / ℃, forgetting factor λ=0.995.

[0081] 3. Dynamic compensation closed-loop calibration The adjustable precision source module of the fully automatic electrical parameter calibrator 2 outputs an AC sinusoidal excitation signal with a frequency of 12.0005MHz (scanning center frequency) and a power of -10dBm. After being amplified and impedance matched by the signal conditioning and control unit 7, the signal is applied to both ends of the crystal through the test probe 5.

[0082] The four-wire measurement module of the fully automatic electrical parameter calibrator 2 synchronously acquires the voltage V across the crystal at a sampling frequency of 1 kHz. H V LO and loop current I. The dynamic error compensation module performs a recursive least squares online estimation algorithm for each sampling point: Every 1ms, the dynamic error compensation module adjusts the current collected V value. H V LO I and the real-time temperature T (provided by the environmental monitoring unit, currently 25.3°C) are substituted into the system transfer function model. The RLS algorithm iteratively updates the temperature coefficient α and the change in contact resistance ΔRc.

[0083] Taking the 10th second as an example: Original measured voltage V m = 120.5mV (attenuated after test pin 5 and cable), loop current I = 3.0mA, real-time temperature T = 25.8℃ (up 0.5℃ from the initial value), current RLS estimate: α = 1.52 × 10 -6 / ℃, β·ΔRc = 0.035Ω.

[0084] A dedicated physical model for resistance / current calibration. P = R (resistance being measured) M raw = V m / I (original measured resistance) δ 1 = α•T•V true / I (Voltage gain error introduced by temperature, referred to the resistance domain) δ2=β•ΔRc (Contact resistance error) K0 is the calibration gain, θ1=α, φ1=T, θ2=β, φ2=ΔRc; Substitute the above data into the unified model

[0085] It can be obtained

[0086] The dynamic error compensation processor substitutes the previously defined values ​​into the model and solves the problem. Original formula: 120.5×10 -3 =[1.00000+(1.52×10 -6 [(×25.8)•V] true +[0.020+0.035]×(3.0×10 -3 ) Calculate the gain coefficient: K0 + α•T = 1.00000 + 1.52 × 10 -6 ×25.8=1.0000392 Calculate the line resistance voltage drop: (R line +β•ΔRc)×I=(0.020+0.035)×3.0×10 -3 =0.055×3.0×10 -3 =1.65×10 -4 V Substitute: 120.5×10 -3 =1.0000392×V true +0.165×10 -3 Transpose terms to solve:

[0087] Equivalent series resistance on the crystal under test:

[0088] The compensated equivalent series resistance of the crystal, ESR, was calculated to be approximately 40.1Ω, while simultaneously tracking the current resonant frequency fs = 12.00048MHz. This process continues to run, enabling real-time error tracking and elimination.

[0089] 4. Intelligent prediction module intervention Every 100 accumulated sampling points (i.e. every 100ms), the control host 4 packages 100 sets of data (including ESR value, temperature, and humidity after first-level compensation) from the past second and sends them to the intelligent coprocessor.

[0090] The intelligent coprocessor performs the forward propagation of the parameter drift prediction model. After input layer normalization (mapping ESR, temperature, and humidity to the [-1,1] interval), one-dimensional convolutional layer extracts local trend features, and GRU layer captures temporal dependencies, the predicted output layer outputs: ΔXp = +0.15Ω (predicts that the crystal ESR will drift positively by 0.15Ω within the next minute), confidence level σ = 0.92.

[0091] 5. Integration of decision-making and parameter updates After receiving the prediction results, the control host 4 executes the fusion decision: Current dynamic error compensation module parameter α current =1.52×10 -6 / °C, preset fusion gain γ=0.5, maximum adjustment limit M=0.2Ω, calculated according to the formula:

[0092] The control host 4 will use the new temperature coefficient α new =1.589×10 -6 The / °C value is sent to the dynamic error compensation module. This means that for subsequent temperature changes, the instrument will compensate with a higher temperature coefficient, precisely offsetting the positive drift trend of the intelligently predicted ESR.

[0093] 6. Final Result Output and Judgment After a 3-second closed-loop optimization calibration (including approximately 100 iterations of AI prediction and parameter updates), the fully automated electrical parameter calibrator 2 outputs the final test results for the crystal: The series resonant frequency fs = 12.000480 MHz Equivalent series resistance ESR ≈ 40.1Ω Quality factor Q = 85,000 Load capacitance CL = 12.5 pF After receiving the test data, the control host 4 automatically determines that the crystal is a "qualified" product according to the preset judgment criteria (frequency deviation ≤ ±10ppm, ESR ≤ 50Ω, Q ≥ 50,000).

[0094] 7. Automatic feeding and cycle testing The control host 4 instructs the three-axis positioning test mechanism 6 to raise the test probe 5 along the Z-axis via the signal conditioning and control unit 7, and the X-axis and Y-axis linear modules move the test probe 5 to the position of the next crystal, repeating the second to sixth steps above.

[0095] The total continuous testing of 100 crystals took approximately 280 seconds (an average of 2.8 seconds per crystal). During this time, the system automatically recorded complete test data for each crystal (including timestamps, ambient temperature and humidity, dynamic compensation parameter trajectories, intelligent prediction records, etc.) and automatically generated a batch test report after the test was completed.

[0096] Example 2 The same platform, without changing the hardware, is used to calibrate a precision voltage reference source (output 10V DC).

[0097] 1. Parameter Identification: The instrument first outputs a ±1V scanning voltage. If the measured device exhibits high impedance and its response is linearly related to the excitation, it is determined to be a "voltage" parameter. The internal analog switch is automatically switched to "direct measurement" mode, and the excitation source is turned off.

[0098] 2. Dynamic compensation initialization: Loading voltage model

[0099] Initial parameters G0 = 1.00000, α V =5×10 -6 / ℃.

[0100] 3. Measurement and Compensation: The instrument measures a 10V standard source at a 100Hz sampling rate. The RLS algorithm monitors the input bias current I in real time. b (Approximately 50 pA and rising slowly), dynamically adjust this term to ensure V comp It stabilized at 10.00000±0.00005V.

[0101] 4. Intelligent prediction module intervention: After 10 seconds of learning, the intelligent prediction module model predicts that "within the next 30 seconds, due to a 0.5℃ increase in ambient temperature, α..." V The drift will be approximately 0.2 × 10⁻⁶. -6 / ℃". The integrated decision control module transforms this prediction into α. V The advance increase makes subsequent compensation more accurate.

[0102] 5. Results output: The average change in the final measured value within the 10s~40s window is only 0.00002V, which is better than traditional instruments (usually above 0.0001V).

[0103] Example 3 The same platform, used to calibrate capacitor parameters without changing the hardware.

[0104] 1. The user connects a 100nF polypropylene capacitor to the instrument. The instrument detects the capacitive response through scanning and automatically switches to "capacitive mode", outputting a 1Vrms, 1kHz sinusoidal constant current (I=1mA).

[0105] 2. The dynamic compensation algorithm calculates the initial capacitance value of 99.8nF based on the voltage amplitude (approximately 1.59V) and phase (approximately -89.5°), and simultaneously estimates the dielectric loss tangent tanδ≈0.0005.

[0106] 3. The dynamic compensation model for capacitance measurement is as follows:

[0107] Where ε DA (f,T)=k0+k1•f 0.5 +k2•T is the dielectric absorption correction factor. The coefficient is updated online using the RLS algorithm, which stabilizes the capacitance measurement accuracy to ±0.05% within a few minutes.

[0108] 4. The intelligent prediction module analyzes the historical tanδ trend of the capacitor at 1kHz, 10kHz, and 100kHz to predict the capacitance decay rate caused by aging, and automatically corrects the measurement values ​​of similar capacitors in subsequent batch tests.

[0109] It should be noted that, in this document, relational terms such as "first" and "second" are used only to distinguish one entity or operation from another, and do not necessarily require or imply any such actual relationship or order between these entities or operations. Furthermore, the terms "comprising," "including," or any other variations thereof are intended to cover non-exclusive inclusion, such that a process, method, article, or apparatus that comprises a list of elements includes not only those elements but also other elements not expressly listed, or elements inherent to such process, method, article, or apparatus.

[0110] The above descriptions are merely embodiments of the present invention. Commonly known structures and characteristics are not described in detail here. Those skilled in the art are aware of all common technical knowledge in the field prior to the application date or priority date, are aware of all existing technologies in that field, and have the ability to apply conventional experimental methods prior to that date. Those skilled in the art can, under the guidance of this application, improve and implement this solution in combination with their own capabilities. Some typical known structures or methods should not be obstacles for those skilled in the art to implement this application. It should be noted that those skilled in the art can make several modifications and improvements without departing from the structure of the present invention. These should also be considered within the scope of protection of the present invention, and will not affect the effectiveness of the implementation of the present invention or the practicality of the patent. The scope of protection claimed in this application should be determined by the content of its claims, and the specific embodiments described in the specification can be used to interpret the content of the claims.

Claims

1. An intelligent calibration system based on the fusion of dynamic error compensation and intelligent prediction, characterized in that: It includes a dynamic error compensation module, an intelligent prediction module, and a fusion decision control module; The dynamic error compensation module has a built-in system transfer function model containing time-varying error components, which is used to acquire the original measurement signal in real time, calculate and subtract the system dynamic error online, and output a first-level compensation value. The intelligent prediction module takes historical compensation value sequence and synchronous environmental data as input and outputs the drift prediction value and confidence level of the future drift of the measured parameter. The fusion decision control module is connected to the dynamic error compensation module and the intelligent prediction module respectively. It is used to convert the drift prediction value output by the intelligent prediction module into the parameter correction amount of the system transfer function model in the dynamic error compensation module. The dynamic error compensation module updates the parameters according to the parameter correction amount, forming a dual-loop closed-loop calibration architecture.

2. The intelligent calibration system based on the fusion of dynamic error compensation and intelligent prediction according to claim 1, characterized in that: The dynamic error compensation module calculates the time-varying parameter vector online using a recursive least squares algorithm. The time-varying parameter vector includes at least one of the temperature drift coefficient and the change in contact impedance.

3. The intelligent calibration system based on the fusion of dynamic error compensation and intelligent prediction according to claim 1, characterized in that: The intelligent prediction module has a built-in trained parameter drift prediction model, which is a temporal-feature fusion neural network, including a temporal coding layer, an environmental attention layer, and a prediction output layer. The temporal coding layer is used to extract features from the input joint historical sequence and output a temporal latent state containing historical environmental features; the joint historical sequence consists of a historical compensation value sequence and synchronization environment data. The environment attention layer is used to calculate the attention weight between the current environment vector acquired in real time and the historical environment features output by the temporal coding layer, and to generate a context vector; The prediction output layer is used to fuse the temporal hidden state with the context vector and output the drift prediction value and confidence level.

4. The intelligent calibration system based on the fusion of dynamic error compensation and intelligent prediction according to claim 1, characterized in that: The fusion decision control module calculates the parameter update amount based on the drift prediction value, confidence level, and sensitivity of the time-varying parameter vector to the compensation output. It then limits the parameter update amount using a saturation limiting function and uses the limited parameter update amount to update the time-varying parameter vector of the system transfer function model.

5. The intelligent calibration system based on the fusion of dynamic error compensation and intelligent prediction according to claim 4, characterized in that: The intelligent prediction module can calculate the residual between the current first-level compensation value and the current first-level compensation reference value. The current first-level compensation reference value is predicted based on the historical compensation value sequence and synchronous environmental data before the current time. When the residual exceeds the dynamic threshold, it is determined to be a calibration abnormality or device mutation, triggering an alarm and recording the complete data chain. The recorded data chain is used for subsequent incremental retraining of the intelligent prediction module to form an iteratively optimized anomaly tracing closed loop.

6. The intelligent calibration system based on the fusion of dynamic error compensation and intelligent prediction according to claim 1, characterized in that: It also includes an automatic parameter type identification module; the automatic parameter type identification module outputs a scanning excitation signal to the device under test and collects the response characteristics of the device under test, identifies the current parameter type based on the response characteristics, and retrieves the system transfer function model parameters that are pre-stored in the system and match the parameter type.

7. The intelligent calibration system based on the fusion of dynamic error compensation and intelligent prediction according to claim 1, characterized in that: It also includes a control host, an environmental monitoring unit, an electrical parameter calibrator, a triaxial positioning testing mechanism, a test probe, and a signal conditioning and control unit; the dynamic error compensation module, the intelligent prediction module, and the fusion decision-making main control module are configured in the control host; the control host is electrically connected to the electrical parameter calibrator, the triaxial positioning testing mechanism, and the signal conditioning and control unit respectively; the test probe is electrically connected to the signal conditioning and control unit; the test probe is mounted on the triaxial positioning testing mechanism, which is used to move the test probe to contact the test piece and complete the alignment; The electrical parameter calibrator is used to output a standard excitation signal to the signal conditioning and control unit. The signal conditioning and control unit performs adaptation processing on the standard excitation signal and applies the conditioned excitation signal to the test probe. After the excitation signal is applied to the device under test through the test probe, a response signal is generated. The response signal is sent back to the signal conditioning and control unit for preprocessing. The signal conditioning and control unit sends the pre-processed signal to the electrical parameter calibrator; The electrical parameter calibrator uploads the timestamped measurement feedback signal to the control host. The dynamic error compensation module in the control host receives the measurement feedback signal and uses it as the original measurement signal. The signal conditioning and control unit is connected to the environmental detection unit and is used to collect environmental data in real time and upload it to the control host.

8. An intelligent calibration method based on the fusion of dynamic error compensation and intelligent prediction, characterized in that: Includes the following steps: S1. Retrieve the initial baseline parameters of the dynamic error compensation module stored in the equipment's factory calibration, and load the locally pre-trained intelligent prediction model. S2. The dynamic error compensation module acquires the original measurement signal, calculates and subtracts the system dynamic error online, and outputs the first-level compensation value. S3. Synchronously collect environmental data, input the compensation value sequence and environmental data within the current time window into the intelligent prediction model, and output the drift prediction value and confidence level of the future drift of the measured parameter. S4. Update the parameters of the dynamic error compensation module based on the drift prediction value and confidence level; S5. Generate the final calibration result using the updated dynamic error compensation module.

Images