High-precision data compensation method and system

Abstract

The invention relates to the technical field of data compensation, in particular to a high-precision data compensation method and system. According to the technical scheme, the method comprises the following steps: data acquisition: synchronously acquiring target data and environment parameters, preprocessing the acquired original data to eliminate abnormal values and suppress random noise, and constructing a measurement matrix M and an environment matrix E; error modeling: calculating an integral nonlinear error INL and a differential nonlinear error DNL of the data based on the measurement matrix M, and establishing a nonlinear error fitting function fINL (x); and environment compensation: constructing an error equation containing the environment matrix E. Through organic combination of dynamic self-adaption, multi-source collaboration, intelligent learning and closed-loop control, a data compensation technology is upgraded from a static, isolated and passive traditional mode to a dynamic, collaborative and intelligent advanced mode, the compensation precision is extremely high, and the dynamic response speed and long-term environmental adaptability are excellent.

Description

Technical Field

[0001] This invention relates to the field of data compensation technology, and in particular to a high-precision data compensation method and system. Background Technology

[0002] In high-precision data acquisition and signal conversion systems (such as high-precision ADCs / DACs and precision sensing and measurement equipment), the final accuracy of the data is the core indicator for evaluating system performance. However, in practical applications, the measurement accuracy of the system is constrained by a variety of non-ideal factors, the most critical of which are integral nonlinearity (INL) error, differential nonlinearity (DNL) error, and systematic deviations introduced by dynamic changes in environmental parameters.

[0003] INL and DNL errors primarily stem from inherent defects in the analog-to-digital conversion process, including component matching errors, parasitic capacitance, and inductance effects. INL describes the overall deviation between the actual and ideal conversion functions, while DNL describes the inconsistency in width between adjacent code values. These nonlinear errors lead to distortion in the quantization results, significantly reducing the system's effective bits per second (ENOB), and creating a nonlinear relationship between the output data and the true value that is difficult to eliminate through simple calibration.

[0004] Fluctuations in parameters such as temperature, humidity, and air pressure in the working environment can directly cause performance drift in key components such as reference voltage sources, operational amplifiers, and sensors, resulting in systematic measurement deviations that vary over time. This error introduced by environmental interference is often coupled with nonlinear errors, making data compensation more complex. Currently, conventional compensation methods used in the industry to improve data accuracy mainly suffer from the following shortcomings:

[0005] Relying on hardware optimization and offline calibration: Many solutions suppress errors at the source by carefully selecting low-temperature drift components and improving circuit layout, or by using a one-time calibration before shipment. These methods are costly and cannot adapt to dynamic error changes caused by aging, wear, or sudden environmental changes in actual operating environments. They are slow to respond and cannot guarantee long-term accuracy.

[0006] The compensation model is singular and has weak generalization ability: Existing software compensation algorithms mostly use a single model, such as a fixed polynomial fitting, which only corrects static nonlinear errors. Such models are prone to overfitting or underfitting when faced with complex and coupled multi-source errors, resulting in poor generalization ability. When environmental conditions or the system operating point change, the compensation effect will drop sharply.

[0007] Lack of a collaborative compensation mechanism for multiple error sources: Most existing technologies treat nonlinear errors and environmental errors as independent problems, failing to establish an effective collaborative compensation model. The interaction between the two is not fully considered, resulting in significant residual errors after compensation, making it difficult to break through the part-in-a-million (ppm) level in compensation accuracy, thus failing to meet the ever-increasing demand for ultra-high precision measurements.

[0008] In summary, this application proposes a high-precision data compensation method and system. Summary of the Invention

[0009] The purpose of this invention is to address the problems in the prior art where existing data compensation methods rely on hardware optimization and offline calibration, and have weak generalization capabilities, by proposing a high-precision data compensation method and system.

[0010] In a first aspect, the present invention provides a high-precision data compensation method, comprising the following steps:

[0011] Data acquisition involves simultaneously collecting target data and environmental parameters, and preprocessing the collected raw data to remove outliers and suppress random noise, thereby constructing the measurement matrix M and the environmental matrix E.

[0012] Error modeling is performed based on the measurement matrix M. The integral nonlinear error INL and the differential nonlinear error DNL of the data are calculated, and a nonlinear error fitting function fINL(x) is established.

[0013] Environmental compensation involves constructing an error equation that includes the environmental matrix E, solving for the environmental impact coefficient matrix β, and verifying and updating the reliability of the environmental impact coefficient matrix β.

[0014] Intelligent optimization involves inputting the data corrected by the nonlinear error fitting function fINL(x) and the environmental parameters into a pre-trained neural network model, which outputs a final compensation value. The final compensation value Dcompensated is calculated as follows: Dcompensated = X - fINL(x) - E·β, where X is the true value vector.

[0015] Verification involves calculating the residual standard deviation σ of the compensated data. If the residual standard deviation σ exceeds a preset threshold, the model parameters are updated.

[0016] Optionally, in the data acquisition, synchronous acquisition involves simultaneously acquiring the target data and the environmental parameters using the same sampling clock; the preprocessing includes using the Grubbs test to remove outliers and using a moving average filter to suppress random noise.

[0017] Optionally, in the error modeling, the least squares method is used to fit the ideal transformation function Fideal(V)=kV+b, and the INL and DNL are calculated based on the ideal transformation function and the actual transformation function Factual(V); the nonlinear error fitting function fINL(x) is a polynomial function.

[0018] Optionally, the nonlinear error fitting function fINL(x) is a third-order polynomial: fINL(x) = a·x³ +b·x² + c·x + d, where x is the normalized input value, and a, b, c, and d are fitting coefficients.

[0019] Optionally, in the environmental compensation step, the error equation is: M = X + E·β + ε, where ε is the residual vector; the reliability verification and update includes: calculating the cumulative change value of environmental parameters within the observation period; if the cumulative change value is less than or equal to a preset threshold, the environmental impact coefficient matrix β is determined to be applicable; otherwise, the environmental impact coefficient matrix β is updated by gradient descent.

[0020] Optionally, in the intelligent optimization step, the neural network model is a BP neural network, where the number of input layer nodes is the sum of the number of data channels n and the number of environmental variables k, and the number of output layer nodes is the number of data channels n.

[0021] Optionally, the BP neural network includes two hidden layers, uses the ReLU activation function, and is trained using the Adam optimizer.

[0022] Optionally, in the verification step, triggering the model parameter update includes refitting the coefficients of the nonlinear error fitting function fINL(x) and updating the weights of the neural network model.

[0023] In a second aspect, the present invention provides a high-precision data compensation system for implementing the method described in the first aspect, comprising:

[0024] The data acquisition module is used to synchronously collect target data and environmental parameters, and to preprocess the collected raw data to construct the measurement matrix M and the environmental matrix E.

[0025] An error modeling module, connected to the data acquisition module, is used to calculate the integral nonlinear error INL and the differential nonlinear error DNL of the data based on the measurement matrix M, and to establish a nonlinear error fitting function fINL(x).

[0026] An environmental compensation module, connected to the data acquisition module and the error modeling module, is used to construct an error equation containing the environmental matrix E, solve for and dynamically update the environmental impact coefficient matrix β;

[0027] The intelligent optimization module, connected to the error modeling module and the environmental compensation module, is used to fuse the corrected data and environmental parameters through a neural network model and output the final compensation value.

[0028] The verification module, connected to the intelligent optimization module, is used to calculate the residual standard deviation σ of the compensated data, and triggers model parameter updates when the residual standard deviation σ exceeds a preset threshold.

[0029] Optionally, the data acquisition module includes:

[0030] The synchronous sampling unit is used to synchronously acquire target data and environmental parameters using the same sampling clock.

[0031] The filtering preprocessing unit is used to remove outliers using the Grubbs test, suppress random noise using moving average filtering, and normalize environmental parameters.

[0032] The intelligent optimization module is a processor based on a BP neural network. It receives the outputs of the error modeling module and the environmental compensation module through a network connection and performs online training and inference.

[0033] Compared with the prior art, this application includes at least one of the following beneficial technical effects:

[0034] This invention utilizes a dynamic verification and update mechanism (such as environmental change constraint testing and gradient descent method) within the environmental compensation module to sense changes in environmental parameters such as temperature, humidity, and air pressure in real time and automatically adjust compensation parameters. This completely changes the traditional static compensation mode that relies on one-time calibration or hardware optimization before delivery, solving the industry problem of delayed compensation effects and decreased accuracy due to its inability to track dynamic changes in the on-site environment.

[0035] This invention combines an error modeling module with an environmental compensation module, first handling nonlinear errors and environmental disturbances separately, and then deeply integrating them through an intelligent optimization module. This collaborative mechanism can effectively handle the complex situation where nonlinear errors and environmental drift are coupled together, avoiding the problems of weak generalization ability and easy overfitting of a single model, thus achieving more comprehensive and accurate error compensation.

[0036] By employing a backpropagation (BP) neural network as the core of the intelligent optimization module, the network can accurately learn and counteract complex residual errors that remain after initial correction and are difficult to describe using mathematical models. This ability to capture subtle error patterns is key to pushing the overall compensation accuracy of the system to the part-in-a-million (ppm) level, far exceeding the accuracy limit of traditional pure mathematical fitting methods.

[0037] The unique verification module of this invention forms a complete feedback loop. By continuously monitoring the standard deviation of the compensated residuals, it can intelligently diagnose model failures caused by equipment aging or drastic environmental changes and automatically trigger global parameter updates. This enables the system to have self-diagnostic and self-optimization capabilities, ensuring that it maintains its initial high-precision performance throughout its entire lifecycle, greatly improving the long-term reliability and service life of the product.

[0038] In the data acquisition module, a series of meticulous preprocessing operations, including synchronous sampling, Grubbs' test to remove outliers, moving average filtering, and environmental parameter normalization, ensure the synchronization, purity, and consistency of the input data. This eliminates many potential errors at the source, providing reliable data support for the establishment of a high-precision model in the backend and facilitating high-precision compensation.

[0039] In summary, this invention upgrades data compensation technology from a static, isolated, and passive traditional mode to a dynamic, collaborative, and intelligent advanced mode by organically combining dynamic adaptation, multi-source collaboration, intelligent learning, and closed-loop control. It not only has extremely high compensation accuracy but also excellent dynamic response speed and long-term environmental adaptability. Attached Figure Description

[0040] Figure 1 A flowchart of a high-precision data compensation method;

[0041] Figure 2 This is a block diagram illustrating the principle of a high-precision data compensation system. Detailed Implementation

[0042] The following specific examples illustrate the implementation of the present invention. Those skilled in the art can easily understand other advantages and effects of the present invention from the content disclosed in this specification. The present invention can also be implemented or applied through other different specific embodiments, and various details in this specification can also be modified or changed based on different viewpoints and applications without departing from the spirit of the present invention. It should be noted that, unless otherwise specified, the following embodiments and features described therein can be combined with each other.

[0043] Example 1: As Figure 1 As shown, this invention proposes a high-precision data compensation method. The following detailed description, in conjunction with the accompanying drawings and an application example of a precision ADC (Analog-to-Digital Converter) system, illustrates this high-precision data compensation method. Those skilled in the art will understand that the specific embodiments described herein are intended to explain the invention and not to limit its scope. The core of this invention lies in achieving dynamic and coordinated compensation for inherent nonlinear errors and environmental interference through five collaboratively working stages.

[0044] S1. Data Acquisition and Preprocessing Stage: The purpose is to obtain synchronized and clean raw data, specifically including:

[0045] Synchronous sampling uses the same sampling clock to synchronously trigger and acquire target data (such as the raw output code value of the ADC) and environmental parameters (including temperature, air pressure, and humidity). Environmental parameters are acquired in real time by a high-precision digital sensor (such as the DS18B20 temperature sensor) integrated on the PCB board. Within a configurable measurement time window, for example, m = 1024 consecutive samples are performed, ultimately constructing a measurement matrix M of dimension 1024 × n (where n is the number of data channels) and an environmental matrix E of dimension 1024 × k (where k = 3 is the number of environmental variables).

[0046] Filtering and normalization:

[0047] Outlier removal involves applying the Grubbs' test (with a significance level of 0.05) to each column of the measurement matrix M (i.e., each data channel) to detect and remove outliers. This effectively filters out gross errors caused by electromagnetic pulses or transmission interference, fundamentally ensuring the reliability of the input data.

[0048] Noise suppression is achieved by using a moving average filter with a window length of 32 to smooth the original data and environmental parameters in order to suppress random white noise.

[0049] Environmental parameters are normalized by normalizing the filtered environmental parameters (temperature, air pressure, and humidity) to the range of [-1, 1]. This eliminates model solution bias caused by differences in physical dimensions and numerical ranges of different environmental parameters, ensuring the accuracy and stability of coefficient calculation in the subsequent environmental compensation module.

[0050] S2. Error Modeling and Nonlinear Correction Stage: This stage aims to accurately quantify and model the inherent nonlinear error of the ADC, specifically including:

[0051] Ideal transformation function fitting: extract m measured data points from the measurement matrix M, and use the least squares method to fit the best ideal linear transformation function Fideal(V)=kV+b.

[0052] INL and DNL calculations, based on the aforementioned ideal function and actual conversion function Factual(V), calculate the integral nonlinearity error (INL) and differential nonlinearity error (DNL) corresponding to each code value:

[0053] INL(Vi) = [Factual(Vi^mid) - Fideal(Vi^mid)] / 1LSB;

[0054] DNL(Vi) = [Vi^high - Vi^low - 1LSB] / 1LSB;

[0055] Where Vi^mid is the midpoint analog value corresponding to the i-th code value, Vi^high and Vi^low are the high and low boundaries of its conversion interval, and LSB is the analog quantity represented by the least significant bit.

[0056] A nonlinear error model is established, using the normalized input value x (range [0,1]) as the independent variable and all calculated INL values ​​as the dependent variable. A third-order polynomial regression is performed to establish the nonlinear error fitting function: fINL(x) = a·x³ + b·x² + c·x + d, where a, b, c, and d are the fitting coefficients. Using a third-order polynomial is key to achieving a balance between accuracy and complexity. It can accurately fit the "S-shaped" nonlinear characteristics of most ADCs while avoiding the overfitting problem that is prone to occur with higher-order polynomials, thus giving the model both accuracy and excellent generalization ability.

[0057] S3. Dynamic Compensation Stage for Environmental Disturbances: This stage is specifically designed to eliminate systematic drift caused by fluctuations in environmental parameters, and includes:

[0058] The error equation is constructed and solved, establishing M = X + E·β + ε. Here, X is the vector of true values ​​to be determined, β is a k × n dimensional environmental influence coefficient matrix, and ε is the residual vector. The initial value of the coefficient matrix β is obtained by solving the minimum residual method.

[0059] Dynamic verification and update mechanism:

[0060] The environmental change constraint test calculates the cumulative change of each environmental parameter in real time within the current observation period. If the cumulative change of all parameters is less than a preset threshold (e.g., temperature change ΔT < 0.5°C), the environment is considered stable, and the current coefficient matrix β can continue to be used.

[0061] The residual t-test performs a statistical t-test on the residual vector ε after solving the error equation to determine whether there is a significant systematic bias. If the test fails, it indicates that the environmental influence pattern has changed.

[0062] If any of the above tests fail, the system immediately invokes the gradient descent method to iteratively update the coefficient matrix β. This mechanism endows the system with the ability to perceive environmental changes in real time and automatically adjust the compensation strategy, completely overcoming the inherent defects of traditional offline calibration technology in adapting to dynamic changes in the field environment, and realizing a leap from "static compensation" to "dynamic adaptation".

[0063] S4. The intelligent fusion and residual error optimization stage involves fusing previous results using intelligent algorithms and precisely canceling residual errors. This includes:

[0064] Neural Network Construction: Design a BP (backpropagation) neural network model. Its input layer has n + k nodes (receiving corrected data and environmental parameters), with two hidden layers, and the output layer has n nodes (outputting the final compensated value). The activation function is ReLU, the optimizer is Adam, and the initial learning rate is set to 0.001.

[0065] Training and inference are performed using X - fINL(x) (i.e., the data after nonlinear correction) and the environment matrix E as inputs, with the theoretical truth or high-precision reference value as the target, to train the network (e.g., 500 iterations). After training, the network is put into online inference. The final compensation value is calculated as: Dcompensated = X - fINL(x) - E·β. The BP neural network, with its powerful nonlinear mapping capability, can accurately learn and compensate for complex coupled residual errors missed by simple models, achieving deep fusion and synergistic cancellation of multi-source errors, thereby improving the overall system accuracy to the ppm level.

[0066] S5. Verification and Closed-Loop Update Phase M This phase ensures the accuracy and stability of the system during long-term operation, specifically including:

[0067] Performance monitoring: The system continuously calculates the standard deviation σ of the residual between the compensated data and the reference value.

[0068] Closed-loop update trigger: A precision threshold is set for the standard deviation σ. Once σ is detected to continuously exceed this threshold, it is determined that the current comprehensive compensation model (including polynomial coefficients a, b, c, d and neural network weights) has degraded in performance due to long-term system drift or drastic environmental changes. At this time, the system automatically triggers a global update mechanism to re-execute the error modeling and neural network training process. This closed-loop design endows the system with the ability to self-diagnose and self-optimize, ensuring that it can maintain a stable output with ultra-high precision throughout its entire life cycle, greatly improving product reliability and market competitiveness.

[0069] Example 2

[0070] like Figure 2 As shown in this embodiment, a high-precision data compensation system is proposed. Its core lies in the use of five collaborative modules: a data acquisition module, an error modeling module, an environmental compensation module, an intelligent optimization module, and a verification module, to achieve dynamic and coordinated compensation for nonlinear errors and environmental interference. The specific implementation of each module is explained below using an example in a precision ADC conversion system.

[0071] 1. Data Acquisition Module

[0072] This module is responsible for acquiring raw, reliable input data, which forms the basis for all subsequent compensation efforts. The specific implementation is as follows:

[0073] Synchronous Sampling Unit: Employing the same sampling clock, this unit synchronously acquires target data (such as the raw output code value of the ADC) and environmental parameters (including temperature, air pressure, and humidity). Environmental parameters are obtained from high-precision sensors integrated into the system. The acquisition process occurs within a configurable time window, for example, performing m=1000 consecutive samples, ultimately constructing a measurement matrix M of dimension 1000 x n (where n is the number of ADC channels) and an environmental matrix E of dimension 1000 x k (where k is the number of environmental variables, 3 in this example).

[0074] Filtering preprocessing unit: preprocesses the original measurement matrix M and environment matrix E.

[0075] First, the Grubbs test was used to detect outliers in each data channel with a significance level of 0.05. This identified and eliminated gross errors caused by transient interference, thus ensuring the purity of the input data and laying a reliable foundation for subsequent high-precision modeling.

[0076] Secondly, a moving average filter (e.g., with a window length of 32) is used to smooth the data, effectively suppressing random white noise. Environmental parameters are also filtered using a moving average filter and further normalized to the interval [-1, 1]. This normalization operation is crucial; it eliminates the numerical magnitude differences caused by different units such as temperature, air pressure, and humidity, avoiding the problem of variables with larger units dominating the model when solving for environmental influence coefficients, thus ensuring the accuracy and fairness of the coefficient solution.

[0077] 2. Error modeling module, implemented as follows:

[0078] Based on m measured data obtained from the data acquisition module, an ideal transformation function F_ideal(V) = k*V + b is fitted using the best straight line method (least square method).

[0079] Calculate the integral nonlinearity error INL and the differential nonlinearity error DNL for each data point:

[0080] INL(Vi) = [Factual(Vi^mid) - Fideal(Vi^mid)] / 1LSB;

[0081] DNL(Vi) = [Vi^high - Vi^low - 1LSB] / 1LSB;

[0082] Where Vi^mid is the midpoint value of the i-th data group, Vi^high and Vi^low are the highest and lowest values ​​of the i-th data group, and LSB is the least significant bit of the system.

[0083] After obtaining the INL values ​​for all data points, a third-order polynomial fitting function fINL(x) = a·x³ + b·x² + c·x + d is established with the normalized input value x (range [0,1]) as the independent variable. The coefficients a, b, c, and d are determined through regression analysis. Using a third-order polynomial effectively avoids overfitting, which is common with higher-order polynomials, while ensuring fitting accuracy, resulting in a model that is both accurate and has good generalization ability. This polynomial will be used for preliminary nonlinear correction of the original data.

[0084] 3. Environmental Compensation Module: This module is specifically designed to eliminate systematic biases introduced by dynamic environmental changes. The specific implementation is as follows:

[0085] The error equation is constructed as: M = X + E·β + ε. Here, X is the vector of true values ​​to be determined, β is the k×n dimensional environmental influence coefficient matrix, and ε is the residual vector.

[0086] The coefficient matrix β is solved using the minimum residual method.

[0087] To ensure the timeliness and reliability of the coefficient β, a dynamic verification and update mechanism is implemented:

[0088] Environmental change constraint test: Calculate the cumulative change of environmental parameters within the current observation period. If the value is less than the preset threshold (e.g., cumulative temperature change < 0.5°C), the environment is considered stable, and the existing coefficient β is used.

[0089] Residual t-test: Statistical analysis is performed on the residual vector ε to test for the presence of significant systematic errors. If the test fails, it indicates that the influence pattern of the current environment has changed, and the existing coefficient β is no longer applicable.

[0090] If any of the above tests fail, the update procedure is immediately initiated, and the coefficient matrix β is iteratively updated using the gradient descent method. This mechanism enables the system to keenly perceive environmental changes and automatically adjust the compensation strategy, thereby overcoming the fatal drawback of the traditional offline calibration method's response lag and achieving "all-weather" dynamic suppression of environmental drift.

[0091] 4. The intelligent optimization module is responsible for fusing the outputs of the preceding modules and accurately canceling residual errors. A backpropagation (BP) neural network model is constructed, with the number of input layer nodes being the sum of the number of corrected data channels (n) and the number of environmental variables (k). Two hidden layers are set, and the output layer has n nodes. The activation function is ReLU, the optimizer is Adam, the initial learning rate is set to 0.001, and the training is iterated 500 times.

[0092] The network input consists of the data after preliminary correction by the error modeling module, namely X - fINL(x), and the environmental matrix E used by the environmental compensation module.

[0093] The network's target output is the final compensation value, calculated using the formula: Dcompensated = X - fINL(x) - E·β. Through network training, the model learns complex residual errors that persist even after nonlinear correction and environmental compensation, errors that are difficult to describe with simple models. The powerful nonlinear mapping capability of the BP neural network enables it to accurately capture these subtle error patterns, achieving deep fusion and collaborative compensation of multi-source errors, thus pushing the overall compensation accuracy to a new level.

[0094] 5. Verification Module: This module ensures the long-term stable operation of the system, monitoring the compensation effect and correcting the model in a timely manner. The system calculates the residual standard deviation σ of the compensated data in real time and compares it with a preset accuracy threshold.

[0095] If the residual standard deviation σ consistently exceeds this threshold, the current compensation model is deemed to have failed due to system aging or drastic environmental changes. In this case, the system automatically triggers a model parameter update mechanism, notifying the error modeling module to refit the polynomial coefficients a, b, c, and d, and notifying the intelligent optimization module to retrain the weights of the BP neural network. This closed-loop design endows the system with self-diagnosis and self-optimization capabilities, ensuring that it maintains ultra-high compensation accuracy at the ppm level throughout its entire lifecycle, significantly improving product reliability and lifespan.

[0096] Through the precise collaboration of the above five modules, this invention realizes a high-precision data compensation method that requires no pre-calibration, can dynamically adapt to environmental changes, and can collaboratively suppress nonlinearity and environmental interference. It is especially suitable for ADC / DAC conversion systems and precision sensing and measurement equipment with stringent accuracy requirements.

[0097] The above specific embodiments are merely several optional embodiments of the present invention. Based on the technical solutions of the present invention and the relevant teachings of the above embodiments, those skilled in the art can make various alternative improvements and combinations to the above specific embodiments.

Claims

1. A high-precision data compensation method, characterized in that, The method comprises the following steps: Data acquisition, synchronously acquiring target data and environmental parameters, and pre-processing the acquired raw data to eliminate outliers and suppress random noise, to construct a measurement matrix M and an environmental matrix E; Error modeling, calculating integral nonlinear error INL and differential nonlinear error DNL of the data based on the measurement matrix M, and establishing a nonlinear error fitting function fINL(x); Environmental compensation, constructing an error equation containing the environmental matrix E, solving an environmental influence coefficient matrix β, and performing reliability verification and updating on the environmental influence coefficient matrix β; Intelligent optimization, inputting the data corrected by the nonlinear error fitting function fINL(x) and the environmental parameters into a pre-trained neural network model, and the neural network model outputs a final compensation value; wherein the calculation formula of the final compensation value Dcompensated is: Dcompensated = X - fINL(x) - E·β, X is a true value vector; Verification, calculating the residual standard deviation σ of the compensated data, and if the residual standard deviation σ exceeds a preset threshold, triggering model parameter updating.

2. The high-precision data compensation method according to claim 1, characterized in that, In the data acquisition, the target data and the environmental parameters are synchronously acquired by the same sampling clock; the pre-processing includes eliminating outliers by using Grubbs test method, and suppressing random noise by using sliding average filtering.

3. The high-precision data compensation method of claim 1, wherein, In the error modeling, an ideal conversion function Fideal(V)=kV+b is fitted by using least square method, and the INL and DNL are calculated based on the ideal conversion function and an actual conversion function Factual(V); the nonlinear error fitting function fINL(x) is a polynomial function.

4. The high-precision data compensation method according to claim 3, characterized in that, The nonlinear error fitting function fINL(x) is a third-order polynomial: fINL(x) = a·x³ + b·x² + c·x + d, wherein x is a normalized input value, a, b, c, and d are fitting coefficients.

5. The high-precision data compensation method according to claim 1, characterized in that, In the environmental compensation step, the error equation is: M = X + E·β + ε, wherein ε is a residual vector; the reliability verification and updating includes: calculating the cumulative change value of the environmental parameters within an observation period, if the cumulative change value is less than or equal to a preset threshold, it is determined that the environmental influence coefficient matrix β is applicable, otherwise the environmental influence coefficient matrix β is updated by gradient descent method.

6. The high-precision data compensation method according to claim 1, characterized in that, In the intelligent optimization step, the neural network model is a BP neural network, the number of input layer nodes is the sum of the number of data channels n and the number of environmental variables k, and the number of output layer nodes is the number of data channels n.

7. The high-precision data compensation method according to claim 6, characterized in that, The BP neural network includes two hidden layers, uses ReLU activation function, and is trained by using Adam optimizer.

8. The high-precision data compensation method according to claim 1, characterized in that, In the verification step, triggering model parameter updating includes refitting the coefficients of the nonlinear error fitting function fINL(x) and updating the weights of the neural network model.

9. A high-precision data compensation system for implementing the method according to any one of claims 1 to 8, characterized in that The method comprises: A data acquisition module is configured to synchronously acquire target data and environmental parameters, and pre-process the acquired raw data to construct a measurement matrix M and an environmental matrix E; An error modeling module, connected with the data acquisition module, is configured to calculate integral nonlinear error INL and differential nonlinear error DNL of data based on the measurement matrix M, and establish a nonlinear error fitting function fINL(x); An environment compensation module, connected with the data acquisition module and the error modeling module, is configured to construct an error equation containing the environment matrix E, solve and dynamically update an environment influence coefficient matrix β; An intelligent optimization module, connected with the error modeling module and the environment compensation module, is configured to fuse the corrected data and the environment parameters through a neural network model, and output a final compensation value; A verification module, connected with the intelligent optimization module, is configured to calculate a residual standard deviation σ of the compensated data, and trigger model parameter updating when the residual standard deviation σ exceeds a preset threshold.

10. The system of claim 8, wherein, The data acquisition module comprises: A synchronous sampling unit is configured to synchronously acquire target data and environment parameters through a same sampling clock; A filter preprocessing unit is configured to remove abnormal values by using a Grubbs test method, suppress random noise by using a sliding average filter, and normalize the environment parameters; The intelligent optimization module is a processor based on a BP neural network, which receives outputs of the error modeling module and the environment compensation module through a network connection, and performs online training and reasoning.

Images