A deep learning-based precision measurement signal noise reduction and error compensation method

Abstract

The application discloses a kind of based on deep learning's precision measurement signal noise reduction and error compensation method;Including the following steps: measurement signal generation and data set construction: the precision measurement time series signal containing random noise and instrument system error is constructed training and test data set;Time-frequency dual-channel feature extraction: measurement signal is carried out time domain normalization processing, and wavelet frequency domain feature is extracted, forms time domain-frequency domain joint input;Dual-branch deep learning model training: construct noise reduction branch and error compensation branch, using Adam optimizer and MSE loss function completes model training;Signal reconstruction and metrological accuracy output: the dual-branch feature is fused and reconstructed signal, output high-precision measurement signal after noise reduction and error compensation;The present application can simultaneously inhibit measurement random noise and compensate instrument temperature drift, non-linear system error, pure algorithm implementation, without hardware modification, suitable for displacement, vibration, mechanics and other precision testing metrological scene, effectively improve measurement accuracy and reliability.

Description

Technical Field

[0001] This invention belongs to the field of testing and metrology, and specifically relates to a deep learning processing method for precision measurement time-series signals, which is particularly suitable for noise suppression of measurement signals, error compensation of instrument systems, and improvement of metrological accuracy. Background Technology

[0002] In precision testing and metrology, various sensors, data acquisition devices, and measuring instruments are inevitably affected by environmental disturbances, circuit noise, mechanical vibrations, and temperature changes during actual operation, resulting in a large amount of random noise and systematic errors in the acquired signals. Common interferences mainly include high-frequency noise such as Gaussian white noise and impulse interference, as well as systematic deviations such as temperature drift, zero-point drift, and nonlinear hysteresis. These noises and errors directly reduce the fidelity and accuracy of the measurement signal, affecting the reliability and repeatability of the measurement results.

[0003] Currently, traditional measurement signal processing methods mainly employ wavelet transform, empirical mode decomposition, Kalman filtering, and moving average filtering to achieve noise suppression. However, these methods can only smooth high-frequency noise to a certain extent and cannot effectively compensate for inherent nonlinear errors, temperature drift errors, and other system deviations. Furthermore, traditional filtering methods are prone to over-smoothing, loss of signal features, and waveform distortion, making it difficult to meet the high-fidelity and high-precision requirements of precision metrology.

[0004] In recent years, deep learning has been widely used in the field of time-series signal processing. However, most existing methods are only aimed at a single noise reduction task and do not integrate noise suppression and system error compensation into a joint learning process, resulting in weak model generalization ability and limited improvement in measurement accuracy. Therefore, developing a precision measurement signal processing method that can simultaneously achieve noise suppression and adaptive compensation for system errors has significant engineering value for improving the accuracy of testing and measurement systems. Summary of the Invention

[0005] The purpose of this invention is to overcome the problems of traditional signal processing methods in the prior art, such as inability to compensate for instrument system errors, easy to cause measurement waveform distortion, and insufficient improvement in metrological accuracy. It provides a method for noise reduction and error compensation of precision measurement signals based on deep learning, so as to achieve high-precision repair of measurement signals and improvement of metrological performance.

[0006] To achieve the above objectives, the technical solution adopted by the present invention is summarized as follows: This invention provides a method for noise reduction and error compensation of precision measurement signals based on deep learning, comprising the following steps: (10) Measurement signal generation and dataset construction: Generate precise measurement time-series signals containing Gaussian noise, impulse interference, temperature drift, and nonlinear system errors, and construct training and test datasets; (20) Time-frequency dual-channel feature extraction: The original measurement signal is normalized and wavelet frequency domain features are extracted to form two input features in the time domain and frequency domain; (30) Training of dual-branch deep learning model: Construct a one-dimensional convolutional network containing a noise reduction branch and an error compensation branch, and train it with clean measurement signals as labels; (40) Signal reconstruction and measurement output: The features extracted by the two branches are fused and reconstructed to output the standard measurement signal after noise reduction and error compensation.

[0007] Preferably, the measurement signal in step (10) includes one or more of displacement signal, vibration signal, mechanical measurement signal, and geometric measurement signal.

[0008] Preferably, the wavelet frequency domain features in step (20) are obtained by db4 wavelet decomposition and spliced ​​with the time domain signal to form a dual-channel input.

[0009] Preferably, in step (30), the noise reduction branch uses a 3×1 small convolution kernel to extract high-frequency noise features, and the error compensation branch uses a 5×1 large convolution kernel to extract temperature drift and nonlinear error features.

[0010] Preferably, step (30) uses the MSE loss function and the Adam optimizer, with a training cycle of 1500 rounds.

[0011] Preferably, step (40) uses channel splicing to achieve feature fusion and completes signal reconstruction through a single convolution layer.

[0012] Compared with the prior art, the present invention has the following beneficial effects: This invention adopts an integrated dual-branch structure for noise reduction and error compensation. One branch focuses on high-frequency noise suppression, while the other focuses on instrument temperature drift and nonlinear error fitting correction, avoiding the accumulation of secondary errors caused by step-by-step processing, and significantly improving measurement accuracy.

[0013] This invention introduces dual input in the time and frequency domains, combining wavelet frequency domain features with time domain waveforms to significantly improve the model's ability to distinguish noise and errors, and effectively avoid waveform distortion and feature loss.

[0014] This invention is based on a one-dimensional lightweight convolutional network, which is specifically adapted to time-series measurement signals. It is stable in training, fast inference speed, and can be directly deployed on the host computer system of measurement instruments. It has strong versatility and high engineering application value.

[0015] This invention adopts a purely data-driven approach, which can adaptively compensate for various system errors without the need to establish a complex error mechanism model, and has good generalization ability in different measurement scenarios. Attached Figure Description

[0016] To facilitate the explanation of the technical solutions of the embodiments, the accompanying drawings are briefly described below. The following drawings only show some embodiments and do not constitute a limitation on the scope of protection.

[0017] Figure 1 This is an overall flowchart of the deep learning-based precision measurement signal noise reduction and error compensation method of the present invention; Figure 2 This is a sub-flowchart of the measurement signal generation and dataset construction steps of the present invention; Figure 3 This is a sub-flowchart of the time-frequency dual-path feature extraction steps of the present invention; Figure 4 This is a schematic diagram of the dual-branch deep learning model structure of the present invention; Figure 5 This is a comparison diagram of the signal effects of various algorithms in this invention. Detailed Implementation

[0018] To make the objectives, technical solutions, and advantages of the embodiments of the present invention clearer, the embodiments are described below with reference to the accompanying drawings. These embodiments are only for explaining the present invention and are not intended to limit the scope of protection.

[0019] Based on the embodiments of the present invention, other embodiments obtained by those skilled in the art without creative effort are all within the protection scope of the present invention.

[0020] The instruments, equipment, reagents, and materials used in the examples were all obtained commercially. Example

[0021] like Figure 1 As shown, a method for denoising and error compensation of precision measurement signals based on deep learning specifically includes the following steps: (10) Measurement signal generation and dataset construction, such as Figure 2 The specific steps shown are as follows: (11) Generate standard precision measurement timing signals, using a combination of sinusoidal signals and slowly varying signals to simulate real measurement waveforms such as displacement, vibration, and mechanics.

[0022] (12) Gaussian white noise and random pulse interference are superimposed on the pure signal to simulate the environmental noise in actual measurement.

[0023] (13) Introduce temperature drift, slow change error, nonlinear distortion error and zero offset into the signal to simulate the inherent system deviation of the instrument.

[0024] (14) Construct measurement samples with noise error, and use the clean signal as the label to divide the training set and the independent test set.

[0025] (20) Time-frequency dual-path feature extraction, such as Figure 3 The specific steps shown are as follows: (21) Perform maximum-minimum normalization on the original noisy signal to map the signal to the [0,1] interval and eliminate the influence of dimensions and amplitude.

[0026] (22) The normalized signal is decomposed into multiple layers using the db4 wavelet basis to obtain components of different frequency bands and construct wavelet frequency domain features.

[0027] (23) The time-domain normalized signal and wavelet frequency domain features are spliced ​​together in the channel dimension to form a dual-channel time-frequency joint input.

[0028] (30) Training of dual-branch deep learning models, such as Figure 4 The specific steps shown are as follows: (31) Construct a one-dimensional convolutional dual-branch network, and use a 3×1 small convolutional kernel to extract local high-frequency noise features in the noise reduction branch.

[0029] (32) The error compensation branch uses a 5×1 large convolution kernel to extract long-term temperature drift and nonlinear error features.

[0030] (33) Using mean squared error (MSE) as the loss function, the Adam optimizer is used for gradient update, and the training cycle is set to 1500 rounds.

[0031] (34) By learning the noise distribution and error patterns through training samples, the model can achieve joint optimization of noise suppression and error compensation.

[0032] (40) Feature fusion and signal reconstruction, specifically including the following steps: (41) The feature maps output by the noise reduction branch and the error compensation branch are spliced ​​together in the channel dimension to complete the fusion of multiple feature information.

[0033] (42) The signal is reconstructed through a convolutional layer, and the standard measurement signal after noise reduction and error compensation is output.

[0034] (43) The repair signal is evaluated using measurement indicators such as RMSE, and the final high-precision measurement results are output. Example

[0035] To verify the noise reduction and error compensation effects of this invention, simulation comparison experiments were conducted and generated. Figure 5 First, a precise measurement clean time-series reference signal is generated, and Gaussian noise, impulse interference, and systematic errors such as temperature drift and nonlinear distortion are superimposed to simulate a noisy measurement signal under actual working conditions. Then, wavelet denoising, Kalman filtering, and the method of this invention are applied to the same noisy signal respectively. Using the clean signal as the benchmark and RMSE as the evaluation index, time-domain waveform comparison curves of each algorithm are plotted to obtain... Figure 5The image shows a comparison of the effects of multiple algorithms.

[0036] Depend on Figure 5 As can be seen, the original noisy signal has an RMSE of 0.1500, indicating severe waveform distortion. After wavelet denoising and Kalman filtering, the RMSEs are 0.1424 and 0.1466, respectively, which can only slightly suppress high-frequency noise and cannot effectively compensate for temperature drift and nonlinear system errors, resulting in limited accuracy improvement. The signal processed by the method of this invention has an RMSE of only 0.0400, and the waveform closely matches the clean signal. It can simultaneously suppress random noise and compensate for instrument system errors. The measurement accuracy and signal fidelity are far superior to traditional algorithms, making it suitable for various precision measurement scenarios.

[0037] The above are merely preferred embodiments of the present invention and are not intended to limit the present invention. Various modifications and variations can be made to the present invention by those skilled in the art. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the scope of protection of the present invention.

Claims

1. A method for denoising and error compensation of precision measurement signals based on deep learning, characterized in that, Includes the following steps: (10) Measurement signal generation and dataset construction: Generate precise measurement time-series signals containing Gaussian noise, impulse interference, temperature drift, and nonlinear system errors, and construct training and test datasets; (20) Time-frequency dual-channel feature extraction: The original measurement signal is normalized and wavelet frequency domain features are extracted to form two input features in the time domain and frequency domain; (30) Training of dual-branch deep learning model: Construct a one-dimensional convolutional network containing a noise reduction branch and an error compensation branch, and train it with clean measurement signals as labels; (40) Signal reconstruction and measurement output: The features extracted by the two branches are fused and reconstructed to output the standard measurement signal after noise reduction and error compensation.

2. The method for noise reduction and error compensation of precision measurement signals according to claim 1, characterized in that, The measurement signal in step (10) includes one or more of displacement signals, vibration signals, mechanical measurement signals, and geometric measurement signals.

3. The method for noise reduction and error compensation of precision measurement signals according to claim 1, characterized in that, The wavelet frequency domain features in step (20) are obtained by db4 wavelet decomposition and spliced ​​with the time domain signal to form a dual-channel input.

4. The method for noise reduction and error compensation of precision measurement signals according to claim 1, characterized in that, In step (30), the noise reduction branch uses a 3×1 small convolution kernel to extract high-frequency noise features, and the error compensation branch uses a 5×1 large convolution kernel to extract temperature drift and nonlinear error features.

5. The method for noise reduction and error compensation of precision measurement signals according to claim 1, characterized in that, Step (30) uses the MSE loss function and the Adam optimizer, with a training cycle of 1500 rounds.

6. The method for noise reduction and error compensation of precision measurement signals according to claim 1, characterized in that, The step (40) uses channel splicing to achieve feature fusion and completes signal reconstruction through a single convolution layer.

Images