Signal dynamic denoising method based on noise depth reconstruction of star-shaped magnetic field sensor array

Abstract

The present invention discloses a signal dynamic denoising method based on deep reconstruction of star-shaped magnetic field sensor array noise, and the implementation steps include: (1) establishing a star-shaped magnetic field sensor array with a "center-satellite architecture"; (2) synchronously collecting array sensor noise in a static state; (3) building a deep network model, taking multi-channel satellite sensor noise as input vector and center sensor noise as output, and using data to conduct supervised training on the network; (4) in a state where a magnetic field is generated by an energized wire, using the trained deep network model to reconstruct the center sensor noise from the multi-channel satellite sensor noise; (5) subtracting the reconstruction noise from the noisy signal captured by the center sensor to obtain a denoised magnetic field signal. The present invention reconstructs the in-situ noise of the center sensor during the test by using the star-shaped sensor array noise, thereby realizing dynamic denoising of the signal under strong magnetic field interference and improving the accuracy of magnetic field testing.

Description

Technical Field

[0001] The present invention relates to the field of magnetic field signal denoising and precision detection technology, and in particular to a signal dynamic denoising method based on star-shaped magnetic field sensor array noise depth reconstruction. Background Art

[0002] The quality inspection of power batteries for new energy vehicles is an important guarantee for ensuring the safety of power batteries. According to the Biot-Savart law, the flow of internal current during battery charging and discharging will form a magnetic field, and the distribution of internal current in the battery can reflect the quality and health of the battery. Therefore, magnetic field detection during battery charging and discharging is an important means to achieve non-destructive testing and quality assessment of batteries.

[0003] Currently, magnetic field sensors can achieve extremely high accuracy, such as atomic magnetometers that can reach 10 -15 However, the most important factor affecting the accuracy of magnetic field measurement is background noise interference. For example, the interference intensity of the geomagnetic field is 10 -5 For battery testing on the production line, the magnetic field interference of electrical and mechanical equipment is more significant, so how to achieve fast and accurate signal denoising under strong interference is the key to magnetic field measurement. Currently, magnetic field signal denoising is mainly achieved through physical shielding and background noise cancellation.

[0004] The physical shielding method places the test battery and sensor into a multi-layer alloy shielding barrel at the same time. For example, researchers at New York University in the United States proposed a magnetic resonance imaging technology to evaluate the state of rechargeable lithium-ion batteries by measuring the tiny induced magnetic field and permanent magnetic field changes around the battery. During the test, the battery is placed in an electromagnetic shielding barrel to reduce the interference of environmental noise (Ilott et al., Nature Commun., 2018). Researchers at Gutenberg University in Germany and others designed a method to use an atomic magnetometer to map the weak induced magnetic field around lithium-ion batteries in a magnetic shielding environment (Hu et al., PNAS, 2020). The physical shielding method can shield most noise, but the shielding barrel has a complex structure, high price, complex operation, and low detection efficiency.

[0005] In the background noise cancellation method, a mixed signal of magnetic field and background noise is measured when the battery is charged and discharged, and a background noise signal is measured when the battery is at rest. The two are subtracted to obtain the magnetic field signal caused by battery charging and discharging. For example, researchers at Mercedes-Benz in Germany use anisotropic magnetoresistive sensors to measure magnetic fields to analyze the current distribution of a single battery. This method achieves denoising by measuring the mixed signal when the battery is powered on and subtracting the background magnetic field when the battery is at rest (Brauchle et al., J. Power Sources, 2021). Researchers at the University of Sheffield in the UK designed a magnetic sensor array for high-resolution spatial measurement of magnetic fields for battery current distribution. During the test, the magnetic sensor output is subtracted from the geomagnetic field output to reduce noise interference (Green et al., IEEE T. Instrum. Meas., 2015). The background noise cancellation method is easy to operate and can guarantee a certain degree of accuracy, but since the noise during the test is difficult to obtain, the existing technology uses the noisy signal during the test minus the noise during the rest period. In fact, there are obvious differences between test noise and static noise, so the current magnetic field denoising technology is not ideal, especially the elimination of strong interference, high transient and directional noise under complex working conditions is still a difficulty and challenge.

[0006] It can be seen that how to accurately obtain or reconstruct the in-situ noise during the magnetic field sensor test is the key to improving the dynamic denoising effect of the magnetic field signal. It has important research significance and broad application prospects for precise detection of battery magnetic fields and online evaluation of manufacturing quality. Summary of the invention

[0007] The technical solution of the present invention overcomes the deficiencies of the prior art and provides a signal dynamic denoising method based on deep noise reconstruction of a star-shaped magnetic field sensor array, which significantly improves the signal dynamic denoising effect and magnetic field testing accuracy under strong magnetic field interference.

[0008] The technical solution adopted by the present invention is:

[0009] A signal dynamic denoising method based on star-shaped magnetic field sensor array noise depth reconstruction comprises the following steps:

[0010] S10: establishing a star-shaped magnetic field sensor array with a “center-satellite architecture”; the star-shaped magnetic field sensor array with the “center-satellite architecture” includes a center sensor and N-channel satellite sensors installed on the same plane;

[0011] S20: synchronously collecting the noise of the star-shaped magnetic field sensor array in a static state; the static state is a state where there is no external current excitation to generate a magnetic field and only the environment interferes with the magnetic field;

[0012] S30: Build a deep network model and train it.

[0013] During training, the input of the deep network model is the matrix Π consisting of the N-channel satellite sensor noise signals of the star-shaped magnetic field sensor array in a static state = [n1(t); n2(t); ...; n N (t)] T ; The target output is set to the noise signal n actually collected by the central sensor of the star-shaped magnetic field sensor array in a static state c (t); The reconstruction noise output by the actual deep network model is:

[0014]

[0015] Where g(·) is a deep network model with the following structure:

[0016] g(Π)=g M (...g2(g1(Π,β (1) ),β (2) )...,β (M) ), (2)

[0017] where g1(·),…,g M (·) represent the 1st,…,Mth layers of the deep network model, β (1) ,…,β (M) Respectively represent the deep network model parameter matrices obtained by training in the 1st,…,Mth stages;

[0018] Calculate n c (t) and The root mean square error between Characterize the noise reconstruction accuracy of the network, and make the root mean square error Minimize, the iterative update equation of the deep network model parameter matrix is:

[0019]

[0020] Where γ∈(0, +∞) represents the learning rate, d and d+1 represent the dth and d+1th iteration training stages respectively, d and d+1 are both positive integers with a minimum value of 1 and a maximum value of M, Ξ(β (d) ;Π;n c (t)) represents the cost function;

[0021] S40: Under the test state, the energized wire is placed directly below the center sensor to generate a magnetic field; at this time, the noisy signal s captured by the center sensor c (t) is the magnetic field change signal s caused by the current cm (t) and the mapping of external noise at the center sensor n cm(t), that is, s c (t) = s cm (t)+n cm (t); The noise of the N-channel satellite sensor collected synchronously is expressed as Π m =[n 1m (t); n 2m (t); ...; n Nm (t)] T ; Use the trained deep network model to reconstruct the noise at the center sensor in the test state N-channel satellite sensor noise Π in the state to be tested m Input the deep network model g(·) to obtain the reconstruction noise of the central sensor;

[0022] S50: Noisy signal s captured by the central sensor c (t) Subtract the deep network model reconstruction noise De-noising is achieved, and a magnetic field change signal without noise interference and only caused by current is approximately obtained.

[0023] In step S10, the satellite sensors are arranged in a circular manner at equal angles, and the central sensor is located at the center of the circle.

[0024] The advantages and positive effects of the present invention are:

[0025] (1) The method of the present invention utilizes the nonlinear fitting advantages of deep networks and is based on big data training to characterize the noise mapping rules between sensor arrays and improve the noise reconstruction capability. On this basis, a supervised online learning system architecture is constructed to break through the generalization capability under dynamic scenes and random disturbances.

[0026] (2) The method of the present invention designs a deep network model to reconstruct noise by revealing the correlation mapping mechanism between the center-satellite sensor signals, ensuring that the reconstructed noise is the "in-situ real-time" noise during the test, thus overcoming the difficulty of traditional methods in effectively obtaining the noise during the test.

[0027] (3) The method of the present invention utilizes a deep noise reconstruction network trained with a star-shaped magnetic field sensor array to achieve noise synthesis under any working conditions, thereby improving the dynamic signal denoising effect under strong magnetic field interference. It has broad application prospects in the fields of precise battery magnetic field detection and online evaluation of manufacturing quality. BRIEF DESCRIPTION OF THE DRAWINGS

[0028] Figure 1 Flow chart for realizing the method of the present invention;

[0029] Figure 2 A schematic diagram of a star-shaped magnetic field sensor array with a "center-satellite architecture" is shown;

[0030] Figure 3 Schematic diagram of the training process of the deep network model of this method;

[0031] Figure 4 Schematic diagram of noise reconstruction and signal denoising of the deep network model of this method;

[0032] Figure 5 This is the central magnetic field sensor signal diagram of the current switching process under strong interference;

[0033] Figure 6 The signal diagram of the magnetic field sensors of four satellites during the current switching process under strong interference;

[0034] Figure 7 A static noise waveform captured by traditional methods;

[0035] Figure 8 A reconstructed noise waveform diagram obtained by the method of the present invention;

[0036] Fig. 9 This is the denoising signal waveform of the traditional method;

[0037] Fig.10 The figure is a waveform diagram of a denoised signal obtained by implementing the method of the present invention. DETAILED DESCRIPTION

[0038] The present invention is further described below with reference to the accompanying drawings and specific embodiments.

[0039] like Figure 1 As shown, the present invention provides a signal dynamic denoising method based on star-shaped magnetic field sensor array noise depth reconstruction, comprising the following steps:

[0040] S10: Establishing a star-shaped magnetic field sensor array with a “hub-satellite architecture”;

[0041] S20: synchronously collect the array sensor noise in a static state;

[0042] S30: Build a deep network model, use multi-channel satellite sensor noise as input vector and central sensor noise as output, and use a large amount of data to conduct supervised training on the network;

[0043] S40: When the magnetic field is generated by the energized wire, the central sensor noise is reconstructed from the multi-channel satellite sensor noise using the trained deep network model;

[0044] S50: The reconstructed noise is subtracted from the noisy signal captured by the central sensor to obtain a denoised magnetic field signal.

[0045] like Figure 2As shown, the star-shaped magnetic field sensor array with a "center-satellite architecture" constructed in step S10 includes a center sensor and multiple satellite sensors installed on the same plane, the satellite sensors are arranged at equal angles in a circular manner, and the center sensor is located at the center of the circle. The current of the magnetic field to be tested is located below the center sensor.

[0046] like Figure 3 As shown, in step S20, the array sensor noise in the static state is first synchronously collected, and the obtained N-channel satellite sensor noise Π=[n1(t); n2(t); ...; n N (t)] T and the single-channel center sensor noise n c (t).

[0047] According to step S30, a deep network model is built to train the noise data to explore the intrinsic correlation between the center-satellite sensor noise. The model can be a convolutional neural network, a long short-term memory neural network, a recursive neural network, etc. The input of the network is the one-dimensional noise signal of the N-channel satellite sensor, and the output is the one-dimensional noise signal of the single-channel center sensor. The reconstructed noise of the deep network model is expressed as:

[0048]

[0049] Where g(·) is the deep network model, β (1) ,…,β (M) Represents the model parameter matrix obtained by training in the 1st, ..., Mth stage. By calculating n c (t) and The root mean square error between To characterize the noise reconstruction accuracy of the network, the data training model is used to make Minimize, the iterative update equation of the neural network model parameter matrix is ​​expressed as:

[0050]

[0051] Where γ∈(0,+∞) represents the learning rate, Ξ(β (d) ;Π;n c (t)) represents the cost function. The supervised learning method is used to iteratively train the model. After the training is completed, the optimal noise reconstruction model is obtained, which can reconstruct the central sensor noise in the same time interval through the N-channel satellite sensor noise in any time interval.

[0052] According to step S40 of the method of the present invention, a powered wire is placed under the central magnetic field sensor. According to the Biot-Savart law, when the wire is powered, a magnetic field is generated, which is detected by the magnetic field sensor. At this time, the noisy signal s captured by the central sensor c(t) is the magnetic field change signal s caused by current cm (t) and the mapping of external noise at the center sensor n cm (t), that is, s c (t) = s cm (t)+n cm (t), noise n cm (t) is already embedded in s c (t) and difficult to separate. Therefore, it is necessary to cm (t) can be accurately estimated and reconstructed to achieve effective denoising. At this time, the N-channel satellite sensor signal collected synchronously is expressed as Π m =[n 1m (t); n 2m (t); ...; n Nm (t)] T The present invention obtains the mapping relationship between the center and satellite sensor noise through deep network model training, and uses the trained deep network model to reconstruct the noise at the center sensor during the test. The method overcomes the difficulty of the prior art in obtaining in-situ real-time noise during the central sensor test.

[0053] According to step S50 of the method of the present invention, the deep network model reconstructs the noise In fact, it is the best estimate of the embedded noise during the current magnetic field test, so the noisy signal s captured by the center sensor c (t) minus the reconstruction noise Denoising can be achieved, that is like Figure 4 This process uses a deep network model to achieve optimal estimation and reconstruction of noise during current magnetic field testing, overcoming the difficulty that existing technologies cannot effectively obtain and filter out dynamic noise during magnetic field testing.

[0054] In order to more clearly show the effect of the present invention, a denoising experiment was conducted on the actual current magnetic field signal. The star-shaped magnetic field sensor array experimental device includes 1 central sensor and 4 satellite sensors. The satellite sensors are equidistantly installed in the southwest, northwest, northeast and southeast directions of the central sensor, and the serial numbers are marked as 1-4 respectively. In order to quantitatively analyze the noise, in this experiment, a straight wire is used to load an intermittent DC current to obtain a certain measurable magnetic field, and the loading current is set to 1A. In order to quantitatively analyze the filtering effect of the method of the present invention on dynamic interference noise, a strong magnetic NdFeB permanent magnet is used to shake in the southwest direction (near satellite sensor 1) to create a magnetic field disturbance during the test.

[0055] Figure 5 is the capture signal of the center sensor, Figure 6 The captured signals are from satellite sensors in 4 different directions. Figure 5 The center sensor signal shows that the current unloading leads to a decrease in magnetic field strength. At the same time, due to the interference of the permanent magnet, the magnetic field waveform has an obvious fluctuation trend. Figure 6 In the figure, the amplitude fluctuation is also clearly visible in the signal waveform of satellite sensor 1 which is closest to the permanent magnet. The other three satellite sensors are farther away from the permanent magnet, so the magnetic field interference is smaller than that of sensor 1.

[0056] The traditional method directly measures the background noise of the central sensor under static and unpowered conditions. The results are as follows: Figure 7 As shown. The present invention builds a bidirectional long short term memory (BiLSTM) deep network model, which includes an input feature layer (feature dimension 20000×4), a BiLSTM layer (hidden nodes 200), a fully connected layer (nodes 50), a dropout layer (ratio 0.5), a fully connected layer (nodes 1), and a regression layer (feature dimension 20000×1). The original noise of the central sensor is reconstructed using the noise of the four satellite sensors. The results are shown in Figure 8 As shown. It can be seen that there are obvious differences between the noise waveforms obtained by the traditional method and the method proposed in this invention. Figure 7 The static noise measured by the traditional method shown is relatively stable, but it cannot reflect the interference of magnetic field fluctuation noise during the current switching period. Figure 8 The method proposed by the present invention can better characterize the dynamic noise interference during the test. Then, the noise signal collected by the central sensor is subtracted from the noise obtained by the traditional method and the method proposed by the present invention, and the result is as follows: Fig. 9 and 10 In order to quantitatively analyze the test results, the theoretical magnetic field change curve calculated by the Biot-Savart law is shown in the figure as a red dotted line. Fig. 9 As shown) and the method proposed by the present invention ( Fig.10 The root mean square errors obtained by the method of the present invention are 1.0716 Gauss and 0.2231 Gauss respectively. By comparison, it can be seen that the denoised signal obtained by the method of the present invention is closer to the theoretical signal. Therefore, the method of the present invention has higher noise reconstruction accuracy and better signal denoising effect. The method of the present invention uses a deep network model to achieve optimal estimation and reconstruction of noise during current magnetic field testing, overcoming the problem that the prior art cannot effectively obtain and filter out dynamic noise during magnetic field testing.

[0057] The contents not described in detail in the specification of the present invention belong to the prior art known to the professional and technical personnel in this field.

[0058] Although the embodiments and drawings of the present invention are disclosed for illustrative purposes, it will be appreciated by those skilled in the art that various substitutions, changes and modifications are possible without departing from the spirit and scope of the present invention and the appended claims. Therefore, the present invention should not be limited to the contents disclosed in the embodiments and drawings of the present invention.

Claims

1. A signal dynamic denoising method based on deep noise reconstruction of star-shaped magnetic field sensor array, characterized in that: The steps include: S10: establishing a star-shaped magnetic field sensor array with a "center-satellite architecture"; the star-shaped magnetic field sensor array with the "center-satellite architecture" includes a center sensor and N-channel satellite sensors installed on the same plane; the satellite sensors are arranged at equal angles in a circular manner, and the center sensor is located at the center of the circle; S20: synchronously collecting the noise of the star-shaped magnetic field sensor array in a static state; the static state is a state where there is no external current excitation to generate a magnetic field and only the environment interferes with the magnetic field; S30: Build a deep network model and train it. During training, the input of the deep network model is the matrix Π = [n1(t); n2(t);…; n N (t)] T ; The target output is set to the noise signal n actually collected by the central sensor of the star-shaped magnetic field sensor array in a static state c (t); The reconstruction noise output by the actual deep network model is: = g(Π),(1) Where g(·) is a deep network model with the following structure: g(Π) = ,(2) where g1(·),…, g M (·) represent the 1st,…,Mth layers of the deep network model, β (1) ,…, β (M) Respectively represent the deep network model parameter matrices obtained by training in the 1st,…,Mth stages; Calculate n c (t) and The root mean square error RMSE (n c (t), ) characterizes the noise reconstruction accuracy of the network, and the root mean square error RMSE (n c (t), ) is minimized, and the iterative update equation of the deep network model parameter matrix is: （3） In the formula represents the learning rate, d and d+1 represent the dth and d+1th iteration training stages respectively, d and d+1 are both positive integers with a minimum value of 1 and a maximum value of M. represents the cost function; S40: Under the test state, the energized wire is placed directly below the center sensor to generate a magnetic field; at this time, the noisy signal s captured by the center sensor c (t) is the magnetic field change signal s caused by current cm (t) and the mapping of external noise at the center sensor n cm (t), that is, s c (t) = s cm (t) + n cm (t); The noise of the N-channel satellite sensor collected synchronously is expressed as Π m = [n 1m (t); n 2m (t);…; n Nm (t)] T ; Use the trained deep network model to reconstruct the noise at the center sensor in the test state = g(Π m ), the N-channel satellite sensor noise Π under test m Input the deep network model g(·) to obtain the reconstruction noise of the central sensor; S50: Noisy signal s captured by the central sensor c (t) Subtract the deep network model reconstruction noise De-noising is achieved, and a magnetic field change signal without noise interference and only caused by current is approximately obtained.

Images