Temperature compensation method for gyroscope

Abstract

The invention relates to a gyroscope-oriented temperature compensation method, in particular to an MEMS gyroscope-oriented temperature compensation method, which comprises the following steps: S1, after specific noise is added, decomposing an output signal of a gyroscope into a series of intrinsic mode functions (IMF); s2, dividing the decomposed intrinsic mode function IMF into a noise segment, a mixed segment and a feature segment by using a sample entropy SE; s3, de-noising processing is carried out on the noise section by using long-window time-frequency peak filtering TFPF, and useful signals are reserved while de-noising is carried out on the mixed section by using short-window time-frequency peak filtering TFPF; s4, using an extreme learning machine ELM optimized by a non-dominated sorting genetic algorithm NSGA II to perform temperature compensation on the feature segment containing the temperature drift term; and S5, combining the de-noised noise section, the de-noised mixed section and the compensated feature section, and reconstructing a final gyroscope output signal. According to the temperature compensation method for the gyroscope, interference of noise can be well reduced, and the temperature compensation effect of the gyroscope is greatly improved.

Description

Technical Field

[0001] The invention relates to temperature compensation, and in particular to a temperature compensation method for a gyroscope. Background Art

[0002] As an important micro-inertial device, MEMS gyroscope has a wide range of applications in aerospace, intelligent electronics, virtual reality and other fields. It can measure the angular velocity of an object and provide data support for applications such as navigation and attitude control. However, temperature changes in internal thermosensitive materials can cause temperature drift, which affects the stability and accuracy of its output signal. Temperature compensation can offset the impact of temperature on MEMS gyroscopes and improve their accuracy and long-term stability. The research and development of temperature compensation methods is one of the current hot spots in MEMS gyroscope research. The compensation methods of MEMS gyroscopes can be roughly divided into two types: hardware compensation and software compensation. These methods aim to mitigate the harmful effects of temperature fluctuations on gyroscope performance and ultimately improve their accuracy and stability in different operating scenarios. Hardware compensation methods mainly involve designing and integrating physical structures or components in the gyroscope system to offset temperature-induced errors. This method usually involves embedding temperature sensors directly into gyroscope components to monitor changes in ambient temperature. By measuring these fluctuations, the sensor provides feedback to adjust the operating parameters of the gyroscope, such as bias and scale factor, in real time to offset temperature effects. Software compensation algorithms can be divided into two structures: serial and parallel. The serial structure is to denoise first and then perform temperature compensation, which will lead to the loss of useful signals. On the contrary, the parallel architecture is to decompose the signal first, and then select the corresponding processing method for different signals. Software compensation algorithms usually rely on adaptive signal decomposition methods, denoising methods, neural networks, etc.

[0003] Temperature compensation of MEMS gyroscopes can be used to improve temperature stability and wide-spectrum dynamic frequency adjustment capabilities through innovative structural design, such as the design of type I bulk acoustic resonator. For high-Q value MEMS gyroscopes, some scholars have designed a real-time temperature self-sensing hardware compensation algorithm, which greatly improves the measurement accuracy of the gyroscope. Although the hardware compensation method is more accurate and reliable, this method has a long development cycle, limited adaptability, and is not easy to promote.

[0004] Mainstream denoising algorithms include wavelet transform, Kalman filter, Fourier transform, etc., all of which have disadvantages to varying degrees. Fourier transform lacks time-domain positioning capability and is not suitable for non-stationary signals. Kalman filtering involves matrix operations, which is computationally intensive and causes severe signal distortion. Some researchers have explored a dual closed-loop control strategy, characterized by differential modulation of the resonant integrated optical gyroscope to improve detection accuracy and minimize noise interference, which significantly improves the reliability of the gyroscope and reduces control inaccuracy under different conditions. Through this method, a MEMS gyroscope with a unique design is designed to enhance its adaptability to manufacturing and environmental changes. Some scholars have alleviated previous limitations on gyroscope robustness by adopting a single-degree-of-freedom actuator and a 2-degree-of-freedom sensing mechanism to enable independent frequency and bandwidth adjustment. Due to the small degrees of freedom of the actuator and the production period, the technical requirements are high, and the use of independent frequency and bandwidth adjustment will greatly increase the cost.

[0005] It should be noted that the information disclosed in the above background technology section is only used for understanding the background of the present application, and therefore may include information that does not constitute prior art known to ordinary technicians in the field. Summary of the invention

[0006] The main purpose of the present invention is to overcome the defects of the above-mentioned background technology and provide a temperature compensation method for a gyroscope.

[0007] To achieve the above object, the present invention adopts the following technical solutions:

[0008] A temperature compensation method for a gyroscope, especially for a MEMS gyroscope, comprises the following steps:

[0009] S1. After adding specific noise, decomposing the output signal of the gyroscope into a series of intrinsic mode functions IMF; wherein the specific noise is a subset selected from a Gaussian white noise sequence according to a predetermined rule;

[0010] S2. Using sample entropy SE, the decomposed intrinsic mode function IMF is divided into a noise segment, a mixed segment and a characteristic segment; wherein the noise segment mainly contains a large amount of noise, the mixed segment contains noise and useful signals, and the characteristic segment mainly consists of drift terms caused by temperature changes;

[0011] S3, using a long window time-frequency peak filter TFPF to denoise the noise segment, and using a short window time-frequency peak filter TFPF to denoise the mixed segment while retaining useful signals;

[0012] S4, using an extreme learning machine ELM optimized by a non-dominated sorting genetic algorithm NSGA II to perform temperature compensation on the characteristic segment containing a temperature drift term;

[0013] S5. Combine the denoised noise segment and the mixed segment and the compensated feature segment to reconstruct the final gyroscope output signal.

[0014] Further, in step S1, the white noise E derived from the kth component of Gaussian white noise is selected. k [w (i) ] As a specific noise aid, the decomposition is performed by the empirical mode decomposition EMD method.

[0015] Furthermore, step S1 specifically includes:

[0016] S11. Add a specific noise w with a standard deviation of β0 to the output signal of the gyroscope. (i) ;

[0017] S12, using an empirical mode decomposition (EMD) algorithm to decompose the signal with specific noise, and sequentially calculate the first to kth IMF values;

[0018] S13, combining the k IMF values ​​obtained by decomposition calculation and the residual term obtained finally to reconstruct an approximate representation of the original signal.

[0019] Furthermore, in step S12, for the original signal x (i) , calculate its local average value as the first residual r1 through the EMD algorithm, from the original signal x (i) Subtract the residual r1 from the original signal, and the difference is the first IMF value c1; for the remaining signals, the current residual is used as the new original signal, and a new local average is calculated, and then the new local average is subtracted from the current residual to obtain a new IMF value. Repeat this process until the predetermined number of IMF values ​​k is reached.

[0020] Furthermore, in step S3, the signal is filtered and denoised using a time-frequency peak filtering TFPF algorithm based on WVD and instantaneous frequency parameter estimation principles, wherein when processing multi-component signals, a pseudo Wiener-Ville distribution PWVD is integrated for signal processing.

[0021] Furthermore, in step S4, before using the extreme learning machine ELM for temperature compensation, the number of neurons in the hidden layer, the type of activation function and the range of the input layer connection weights are optimized by using a non-dominated sorting genetic algorithm NSGA II.

[0022] Further, in step S4, using a non-dominated sorting genetic algorithm NSGA II to optimize the extreme learning machine ELM includes:

[0023] The parameters to be optimized in ELM are mapped to the population individuals in the NSGA-II algorithm, the optimization interval of hidden neurons is set, and multiple types of activation functions are selected for testing; the scaling factor λ is introduced to dynamically adjust the range of the input layer connection weights:

[0024] w ij ∈λ·[-1,1]λ∈[0,1]

[0025] Among them, w ij represents the input layer connection weight, and λ is the scaling factor.

[0026] Further, in step S4, using the non-dominated sorting genetic algorithm NSGA II to optimize the extreme learning machine ELM includes: using the mean square error RMSE and the 2-norm of the output weight as the objective function, as follows:

[0027]

[0028] Where N is the number of input samples, y p,i and t,i are the predicted value and true value of the first training sample.

[0029] Furthermore, in step S4, the algorithm process of NSGA II includes: first, randomly generating an initial population P containing N individuals t ; Then, the basic operation of genetic algorithm is used to obtain the offspring population Q t ; Then, by reconstructing P t and Q t , and obtain a new population R with a population size of 2N t ; According to the elite strategy, P is obtained by selecting the best N individuals in Rt t+1 , and obtain its offspring population Q again through genetic algorithm t+1 , construct a new population R t+1 , and this process is iterated until the stopping condition is met.

[0030] A computer-readable storage medium stores a computer program, wherein the computer program implements the temperature compensation method for a gyroscope when executed by a processor.

[0031] The present invention has the following beneficial effects:

[0032] The present invention proposes a temperature compensation method for gyroscopes, which is a new parallel temperature compensation and noise reduction scheme, and effectively solves the mixing problem of noise and temperature drift in the gyroscope output signal. First, the gyroscope output signal after adding auxiliary noise is decomposed, and then the gyroscope output signal is subjected to improved complete ensemble empirical mode decomposition using adaptive noise, thereby improving the recognition rate of the gyroscope. Then, a series of intrinsic mode functions are classified using sample entropy and divided into noise segments, mixed segments and feature segments. For noise segments and mixed segments with different noise levels, a time-frequency peak filtering (TFPF) method with different window lengths is used to achieve a compromise between noise removal and signal retention. For feature segments with temperature drift, an extreme value learning machine (ELM) is optimized using a non-dominated sorting genetic algorithm (NSGA II), and a high-precision temperature compensation model is established. Finally, the processed components are reconstructed to obtain the final output signal. Through comparative analysis of experimental results, it is proved that the method has good performance and effect in gyroscope temperature compensation.

[0033] In a preferred embodiment, the present invention proposes a gyroscope temperature compensation hybrid algorithm based on ICEEMDAN-SE-TFFPF and NSGAII-ELM. After ICEEMDAN decomposition and SE classification, the output signal of the gyroscope is divided into a noise segment, a mixed segment and a feature segment, and then the long window and short window TFPF are selected to denoise the noise segment and the mixed segment respectively. Before processing the feature segment, NSGAII is used to optimize ELM, including the number of hidden layer neurons, the type of activation function and the range of input layer connection weights. The optimization target is determined by the mean square error (RMSE) of the predicted output and the 2-norm of the output layer connection weights. This makes ELM have better generalization ability and higher prediction accuracy.

[0034] Compared with BP neural network and ELM, the prediction value of the improved ELM of the present invention is closest to the actual value, which improves the compensation accuracy. The experimental results show that the angle random walk and bias stability of the output signal after compensation are reduced from 0.531076 to 6.65894×10-3° / h / √Hz and 32.7364 to 0.259247° / h, respectively.

[0035] The advantages of the embodiments of the present invention mainly include:

[0036] 1. After ICEEMDAN decomposition and SE classification, the output signal of the gyroscope after adding specific noise is divided into noise segment, mixed segment and feature segment, and then the long window and short window TFPF are selected to denoise the noise segment and mixed segment respectively, which can effectively reduce the interference of noise;

[0037] 2. Before processing the feature fragments, use NSGAII to optimize ELM, including the number of hidden layer neurons, the type of activation function, and the range of input layer connection weights, which can greatly improve the temperature compensation effect;

[0038] 3. The optimization target is determined by the mean square error (RMSE) of the predicted output and the 2-norm of the output layer connection weight. This makes ELM have better generalization ability and higher prediction accuracy. Compared with BP neural network and ELM, the predicted value of the improved ELM is closest to the actual value, which improves the compensation accuracy.

[0039] Other beneficial effects of the embodiments of the present invention will be further described below. BRIEF DESCRIPTION OF THE DRAWINGS

[0040] Figure 1 Figure 2 is a structural diagram of a dual-mass MEMS gyroscope.

[0041] Figure 2 Schematic diagram of the four working modes of the gyroscope: (a)-(d) are ω1=2623×2πrad / s, ω2=3342×2πrad / s, ω3=3468×2πrad / s, and ω4=3484×2πrad / s respectively.

[0042] Figure 3 A description diagram of the gyroscope monitoring system.

[0043] Figure 4 FIG. 4 is a network structure diagram of the ELM according to an embodiment of the present invention.

[0044] Figure 5 It is a framework diagram of ICEEMDAN-SE-TFPF and NSGAII-ELM according to an embodiment of the present invention.

[0045] Figure 6 FIG. 4 is a diagram of a gyroscope temperature testing device according to an embodiment of the present invention.

[0046] Figure 7 Graph showing experimental temperature changes and corresponding gyroscope outputs according to an embodiment of the present invention.

[0047] Figure 8 This is a diagram of the decomposition result of the gyroscope output signal based on ICEEMDAN according to an embodiment of the present invention.

[0048] Fig. 9 A summary diagram of classifying IMFs in a gyroscope output signal using sample entropy according to an embodiment of the present invention, including (a) SE of each IMF and (b) examples of different parts.

[0049] Fig.10 This is a particle distribution diagram in the NSGAII search according to an embodiment of the present invention.

[0050] Fig.11 Graph showing noise reduction or compensation results for each segment according to an embodiment of the present invention.

[0051] Fig.12 3 is a comparison diagram of the gyroscope output before and after denoising according to an embodiment of the present invention.

[0052] Fig.13 This is a final output diagram of the gyroscope based on ICEEMDAN-SE-TFPF and NSGAII-ELM according to an embodiment of the present invention.

[0053] Fig.14 It is an Allan variance analysis diagram of the output signal of the embodiment of the present invention.

[0054] Fig.15 It is a comparison chart of the experimental results of BP neural network, ELM and NSGAII-ELM of the embodiment of the present invention. DETAILED DESCRIPTION

[0055] The following is a detailed description of the embodiments of the present invention. It should be emphasized that the following description is only exemplary and is not intended to limit the scope and application of the present invention.

[0056] The embodiment of the present invention provides a temperature compensation method for a gyroscope, especially for a MEMS gyroscope, comprising the following steps:

[0057] S1. After adding specific noise, decomposing the output signal of the gyroscope into a series of intrinsic mode functions IMF; wherein the specific noise is a subset selected from a Gaussian white noise sequence according to a predetermined rule;

[0058] S2. Using sample entropy SE, the decomposed intrinsic mode function IMF is divided into a noise segment, a mixed segment and a characteristic segment; wherein the noise segment mainly contains a large amount of noise, the mixed segment contains noise and useful signals, and the characteristic segment mainly consists of drift terms caused by temperature changes;

[0059] S3, using a long window time-frequency peak filter TFPF to denoise the noise segment, and using a short window time-frequency peak filter TFPF to denoise the mixed segment while retaining useful signals;

[0060] S4, using an extreme learning machine ELM optimized by a non-dominated sorting genetic algorithm NSGA II to perform temperature compensation on the characteristic segment containing a temperature drift term;

[0061] S5. Combine the denoised noise segment and the mixed segment and the compensated feature segment to reconstruct the final gyroscope output signal.

[0062] The present invention proposes a temperature compensation method for gyroscopes, which is a new parallel temperature compensation and noise reduction scheme, and effectively solves the mixing problem of noise and temperature drift in the gyroscope output signal. First, the gyroscope output signal after adding auxiliary noise is decomposed, and then the gyroscope output signal is subjected to improved complete ensemble empirical mode decomposition using adaptive noise, thereby improving the recognition rate of the gyroscope. Then, a series of intrinsic mode functions are classified using sample entropy and divided into noise segments, mixed segments and feature segments. For noise segments and mixed segments with different noise levels, a time-frequency peak filtering (TFPF) method with different window lengths is used to achieve a compromise between noise removal and signal retention. For feature segments with temperature drift, an extreme value learning machine (ELM) is optimized using a non-dominated sorting genetic algorithm (NSGA II), and a high-precision temperature compensation model is established. Finally, the processed components are reconstructed to obtain the final output signal. Through comparative analysis of experimental results, it is proved that the method has good performance and effect in gyroscope temperature compensation.

[0063] The specific embodiments of the present invention are further described below.

[0064] The temperature compensation method of the embodiment of the present invention is applied to a dual-mass MEMS gyroscope as an example. Figure 1 As shown in Figure 1. The gyroscope operates in drive mode and sense mode. The drive mode, which includes components such as drive combs and drive springs, promotes movement along the X-axis. Conversely, the sense mode, which consists of sense combs and sense springs, enables movement along the Y-axis. Figure 1 The mass shown oscillates in both modes along the negative X-axis. In addition, the two modes of the gyroscope are intentionally isolated to prevent displacement coupling. When an angular velocity Ωz is applied about the Z-axis, the Coriolis force generated by the vibrating mass is transferred to the frame along the Y-axis and monitored by the circuit.

[0065] The gyroscope works on the principle of tuning fork, using U-shaped connecting springs to connect the dual driving mass blocks, and the driving springs to connect the dual sensing mass blocks. In order to check its operation mode, Ansys software was used for simulation, such as Figure 2 It is noteworthy that there is a significant frequency difference of over 1000 Hz between the first and fourth operating modes, with the fourth mode displaying a quality factor of over 2000, identifying it as the drive mode of the gyroscope. Figure 2Four operating modes of the gyroscope are shown: (a)-(d) are represented as ω1=2623×2πrad / s, ω2=3342×2πrad / s, ω3=3468×2πrad / s, ω4=3484×2πrad / s. (2a)-(2d) sequentially give the simulation results of the gyroscope operating modes. The first mode, the driven in-phase mode, involves the dual mass vibration of the gyroscope in line with the X-axis. Next, the second mode, the sensed in-phase mode, sees the two masses vibrating simultaneously along the Y-axis direction. In contrast, the third mode, the sensed anti-phase mode, exhibits oscillations on the Y-axis in opposite directions of the two mass blocks. Finally, the fourth mode, the driven anti-phase mode, demonstrates the oscillation counter of the two mass blocks of the gyroscope on the Y-axis. In addition, the resonant frequencies of these four modes are observed at frequencies of 2623Hz, 3342Hz, 3468Hz and 3484Hz. It is noteworthy that the fourth mode not only acts as the driving antiphase mode, but also as the main driving mode. From this analysis, it can be seen that the two masses of the gyroscope have two degrees of freedom, while the driving mode and the frame only exhibit one degree of freedom.

[0066] Figure 3 A description of the gyro monitoring system is shown in FIG. Figure 3 In the example, the drive comb uses a split amplifier to measure the displacement of the drive frame x(t). In order to achieve the same dac Sin(ω d t) required phase alignment, a 90° phase delay is applied to the signal. Subsequently, V is extracted through a full-wave rectifier and a low-pass filter. dac and with the reference voltage V ref When the comparator output passes through the integrator controller, a control signal is generated. This control signal drives the DC signal V DC Accumulate to V dac Sin(ω d t), thereby stimulating the drive mode. A differential sense amplifier is used to capture the motion signals of the left and right sensing masses. A second differential amplifier then processes the output signals to produce a total sensed motion signal, V stotal The signal is then demodulated with V dac Sin(ω d t), and then filtered using a low-pass filter to obtain the final sensitive motion signal.

[0067] Improved adaptive noise complete ensemble empirical mode decomposition

[0068] EMD is a classic algorithm for complex signal analysis and processing. EMD decomposes the signal into an amplitude-frequency modulation function and a monotonic trend signal. Although empirical mode decomposition (EMD) has a good effect in complex signal analysis, it will lead to mode mixing due to its local characteristics. In order to solve this problem, this paper proposes to introduce noise in the decomposition process to alleviate the mode mixing phenomenon. On this basis, the CEEMDAN algorithm based on adaptive noise is proposed, which is an important improvement on the EEMD algorithm and solves the influence of auxiliary noise on the number of modes. However, compared with EMD, CEEMDAN still has some shortcomings, such as residual noise in the mode, signal characteristics appearing after the false mode, etc. The above problems are solved by ICEEMDAN. Different decomposition methods. ICEEMDAN selects a special white noise E k [w (i) ] As an auxiliary decomposition, the noise comes from the kth component of Gaussian white noise, which is decomposed by EMD. A brief introduction to ICEEMDAN:

[0069] Before describing the algorithm steps, we first define several operation symbols. M(.) is the operator that generates the local average value. The k-th IMF value after EMD decomposition can be obtained by E k (.) Get the operator. Obviously, E1(x) = xM(x).

[0070] The decomposition process is as follows:

[0071] Step 1: Add some noise: In equation (1), w (i) represents the added noise, and β0 represents the standard deviation of the added noise.

[0072]

[0073] Step 2: EMD calculation for each x (i) The local average of , and take its local average as the first residual r1, and then get the first IMF value c1:

[0074]

[0075] Step 3: Immediately calculate the second IMF value c2:

[0076]

[0077] Step 4: Similarly, the kth value of IMF is calculated according to the above steps, where:

[0078]

[0079] After obtaining the k modes and adding the residual, the original signal can be rewritten as:

[0080]

[0081] The output of the gyroscope is decomposed into a series of intrinsic mode functions (IMFs) using ICEEMDAN. The proposed hybrid algorithm requires the classification of the obtained IMFs to identify different components, so SE is used as a decision rule to distinguish these IMFs.

[0082] Sample Entropy

[0083] SE is an effective measurement method that can measure the regularity and complexity of complex signals by calculating the probability of generating different patterns in a sequence. This method does not depend on the length of the data and has a high detection accuracy. Relatively short data can get a stable estimate. The smaller the SE value, the higher the similarity of the signal sequence, which means that the sequence is regular and can be considered a useful signal, and vice versa.

[0084] Time-Frequency Peak Filter (TFPD)

[0085] TFPF is a noise elimination technology. The TFPF algorithm mainly filters and denoises the signal based on the principles of WVD and instantaneous frequency parameter estimation. Due to the appearance of cross terms, its efficacy will be reduced when processing multi-component signals, thereby affecting its time-frequency resolution. In order to address the challenges brought by cross terms in TFPF, the pseudo-Wiener-Ville distribution (PWVD) is integrated in the preferred embodiment. In essence, TFPF involves encoding the noise signal, converting it into an analytical signal of the instantaneous frequency, and estimating the effective signal through instantaneous frequency estimation.

[0086] Extreme Learning Machines (ELM)

[0087] ELM is a feedforward neural network with only one hidden layer. The key innovation is that the connection weights and biases are randomly generated, especially for the hidden layer, and do not need to be adjusted after setting. Compared with most neural networks that require constant adjustment of weights and thresholds, this method reduces the amount of computation by half. In addition, the output layer connection weight β is determined once by solving the equation and does not require iterative adjustment. ELM has a fast speed while ensuring learning accuracy. The ELM network structure is as follows Figure 4 As shown, the description is as follows:

[0088] Before using ELM, determine the number of neurons in the hidden layer, the type of activation function, and the range of the input layer connection weights. Appropriate parameter settings will affect the performance of ELM and, in turn, the effect of temperature compensation.

[0089] Non-Inferior Sorting Genetic Algorithm II

[0090] The non-dominant sorting genetic algorithm (NSGA) is an improved genetic algorithm based on Pareto optimality. The algorithm stratifies the population according to the non-dominant relationship between individuals, and then performs operations such as selection, crossover, and mutation. The NSGA II algorithm introduces the fast non-dominated sorting (FNS) method, the congestion level comparison operator, and the elite strategy to improve NSGA. This not only reduces the complexity of the NSGA algorithm, but also makes the Pareto frontier optimal solution evenly distributed over the entire Pareto domain, thereby alleviating the phenomenon of local optimal solutions and improving the algorithm's search ability. The following will introduce FNS, congestion comparison operator, and elite strategy:

[0091] A. Fast non-inferior sorting.

[0092] The essence of multi-objective optimization problem is to obtain the optimal solution of Pareto frontier. The FNS technique is used to explore the entire population, which is stratified according to the non-inferiority level of individual solutions, where the individual solutions of the next layer are dominated by any solution of the previous layer, and there is no dominant relationship between individuals in the same layer. According to this principle, the search is directed to the Pareto optimal frontier.

[0093] B. Congestion comparison operator.

[0094] Crowding degree d It is defined as the individual density of the area where a given point is located. First, the crowding degree i of each individual d is set to 0. After FNS of the population, the crowding for individuals on the boundary is symbolic infinity, while for everyone else the crowding is:

[0095]

[0096] Among them, f j i+1 is the j-th objective function value at the ith point.

[0097] C. Elite Strategy

[0098] After the FNS of the population, the crowding level of individuals within the generated non-dominant layer is calculated. “Elite” individuals are selected based on low non-dominant levels and high crowding, representing the optimal solution of the Pareto front among these elites.

[0099] The algorithm process of NSGA II: First, randomly generate an initial population P containing N individuals t Then, the basic operation of genetic algorithm is used to obtain the offspring population Q t Then, by reconstructing P t and Q t , we get a new population R with a population size of 2N t According to the elite strategy, P is obtained by selecting the best N individuals in Rt. t+1, and obtain its offspring population Q again through genetic algorithm t+1 , construct a new population R t+1 , iteratively execute the steps until the stopping condition is met.

[0100] Optimization of ELM

[0101] In order to improve the generalization and search capabilities of ELM, NSGA-II is used to optimize the number of hidden layer neurons, the type of activation function, and the range of input layer connection weights. The key technologies of NSGA II-ELM are:

[0102] 1) Definition of individuals in a population.

[0103] The mapping relationship between the parameters to be optimized in ELM and the individuals in the NSGA-II algorithm population is established. The optimization interval of hidden neurons is set to [0,50], and five commonly used activation functions such as "sigmoid", "sin", "hardlim", "tribas" and "radbas" are selected for testing. In order to optimize the input layer connection weights, the scaling factor λ is introduced to limit the range of the input layer connection weights. Therefore, the value range of the input layer connection weights can be expressed as:

[0104] w ij ∈λ·[-1,1]λ∈[0,1],(7)

[0105] where w ij Represents the input connection weight. The random setting range of the input layer weight changes with the value of the scale factor λ, thereby achieving the purpose of limiting its range.

[0106] 2) Determination of objective function.

[0107] For ELM, the smaller the output weight and prediction error, the better the network generalization ability. Therefore, this paper takes the mean square error (RMSE) between the predicted value and the actual value and the 2-norm of the output weight as two objective functions.

[0108]

[0109] Where N is the number of input samples, y p,i and t,i are the predicted value and true value of the first training sample.

[0110] ICEEMDAN-SE-TFPF and NSGAII-ELM

[0111] Based on the above algorithm, a hybrid algorithm ICEEMDAN-SE-TFPF and NSGAII-ELM for gyroscope temperature compensation is proposed. The algorithm logic diagram is shown in Figure 5As shown, the specific steps are as follows:

[0112] Step 1: First, the output noise and temperature drift of the gyroscope are decomposed into IMFs using ICEEMDAN;.

[0113] Step 2: Divide the decomposed image into noise segmentation, mixed segmentation and feature segmentation.

[0114] Step 3: Use long-window TFPF to denoise the noise segment with high noise content, short-window TFPF to denoise the useful signal in the mixed segment, and use NSGAII-ELM to compensate for the characteristic segment containing temperature drift.

[0115] Step 4: Finally, the denoised noise segment, denoised mixed segment and compensated feature segment are combined to reconstruct the final gyroscope output.

[0116] experiment

[0117] Experimental procedures

[0118] The temperature experiment of the gyroscope is as follows Figure 6 As shown. A detection circuit is distributed on three separate PCB boards. Metal pins are not only used to interconnect electronic signals in the detection circuit, but also facilitate the connection between the three PCBs, each of which is packaged in a rubber pad. Subsequently, the rubber-wrapped PCB is placed in a metal housing, a measure that effectively protects the structure of the chip while minimizing the risk of severe impact. To mitigate electromagnetic interference, the ground signal is connected to the metal housing. In addition, one of the three PCBs serves as a weak signal interface with the structural chip, while the other two serve as a sensing circuit and a driving circuit, respectively. The experimental setup includes a temperature-controlled oven, a multimeter, and a signal generator. and ±10V DC power supply. The experimental process is as follows Next: First, activate the gyroscope and allow it to stabilize at room temperature One hour. Next, the temperature-controlled oven was quickly raised to 60°C to ensure that the gyroscope housing reached a constant temperature of 60°C and was maintained at this temperature for an additional hour. Subsequently, when the temperature in the oven was lowered to 10°C, the gyroscope continued to operate under these conditions for another hour. Finally, when the temperature dropped to -40°C, the gyroscope operated for a final hour before concluding the experiment.

[0119] Experimental Results

[0120] Figure 7 The temperature experiment results of the gyroscope are shown. It can be seen that the output of the gyroscope has a nonlinear relationship with the temperature change. When the temperature changes from 60℃ to -40℃, the output of the gyroscope changes from 0.115° to 0.15°. The output signal of the gyroscope contains not only temperature drift but also noise.

[0121] According to the process of the proposed hybrid algorithm, the output signal of the gyroscope is firstly decomposed by ICEEMDAN. Figure 8 It shows that the output signal of the gyroscope is decomposed into 12 IMFs, and the SE value of each IMF is calculated and classified. The results are as follows Fig. 9 As shown, (a) SE of each IMF, (b) examples of different parts. The SE of the 1st, 2nd, and 3rd IMFs are all greater than 0.6, and these IMFs are divided into noise segments, which contain a lot of noise. Therefore, a long window TFPF with good denoising ability is selected for denoising, and the SE of the 4th, 5th, 6th, and 7th MFs is between 0.1 and 0.6, and these IMFs are divided into mixed segments containing noise and useful signals. Since the short window TFPF can well retain useful components while denoising, the SE values ​​of the 8th, 9th, 10th, 11th, and 12th IMFs are selected to be less than 0.1, and these IMFs are divided into characteristic segments, which is a drift term caused by temperature changes, and then compensated by NSGAII-ELM.

[0122] Before establishing the temperature compensation model to compensate the characteristic segment, NSGA II is used to optimize the ELM to obtain the best network parameters. As mentioned above, the best search range of hidden neurons is [0,50], the activation functions are "sigmoid", "sin", "hardlim", "tribas" and "radbas", and the connection weight range of the input layer is set to [-1,1]. In addition, the population size is set to 40, the crossover probability is 0.8, the mutation probability is 0.15, and the number of iterations is 15. After iteration, the optimization results of the NSGAII particle swarm are obtained. Fig.10 The distribution of particles in the NSGA II search in three-dimensional space is shown, and each blue circle represents a particle. The particle position marked by the red circle is the optimal network parameter of ELM, that is, the number of hidden neural networks is 11, the activation function type is the "sin" function, and the range of the input layer connection weight is [-0.65, 0.65]. Fig.11 The comparison of each segment before and after processing is given, and the noise reduction or compensation results of each segment are given. In order to study the denoising ability of ICEEMDAN-SE-TFPF, the denoised noise segment, denoised mixed segment and feature segment are reconstructed. Fig.12 In the comparison of the gyroscope output before and after denoising, we can see that the noise is basically eliminated. The final signal is obtained by compensation. Fig.13 It is shown that based on ICEEMDAN-SE-TFPF and The final output of the NSGAII-ELM gyroscope is Fig.13 It can be seen that the noise component and temperature drift of the gyroscope output signal are eliminated. To further evaluate the proposed algorithm, a quantitative analysis will be performed in the next subsection.

[0123] Comparative Analysis

[0124] Fig.14 The Allan variance analysis of the output signal is shown to illustrate the comparison results. After denoising and compensation, the angle random walk of the original output is reduced from 0.531076 to 6.65894×10 -3 ° / h / √Hz, and the bias stability is reduced from 32.7364° / h to 0.259247° / h. These findings emphasize the superior performance of the proposed hybrid algorithm. The key to improving the compensation effect is to establish a high-precision model and optimize the ELM using NSGAII. In order to verify the superiority of the improved ELM, a comparative analysis was performed using the BP neural network and the non-optimized ELM with the experimental temperature change as input and the gyroscope output drift as the training output. Fig.15 The comparison results of BP neural network, ELM and NSGA II-ELM are shown in Fig.15 In the example, the inflection point error is larger than the actual output. Compared with BP, the predicted output of the improved ELM is closest to the actual value. The input parameters of ELM are optimized through NSGAII, which makes ELM have stronger generalization ability and learning accuracy.

[0125] In summary, the gyroscope temperature compensation hybrid algorithm based on ICEEMDAN-SE-TFFPF and NSGAII-ELM proposed in the embodiment of the present invention divides the output signal of the gyroscope into noise segment, mixed segment and feature segment after ICEEMDAN decomposition and SE classification, and then selects long window and short window TFPF to denoise the noise segment and mixed segment respectively. Before processing the feature segment, NSGAII is used to optimize ELM, including the number of hidden layer neurons, the type of activation function and the range of input layer connection weights. The optimization target is determined by the mean square error (RMSE) of the predicted output and the 2-norm of the output layer connection weights. This makes ELM have better generalization ability and higher prediction accuracy. Compared with BP neural network and ELM, the predicted value of the improved ELM is closest to the actual value, which improves the compensation accuracy. Experimental results show that the angle random walk and bias stability of the output signal after compensation are reduced from 0.531076 to 6.65894×10-3° / h / √Hz and 32.7364 to 0.259247° / h, respectively.

[0126] Compared with the traditional method, the advantages of the embodiment of the present invention mainly include:

[0127] 1. After ICEEMDAN decomposition and SE classification, the output signal of the gyroscope after adding specific noise is divided into noise segment, mixed segment and feature segment, and then the long window and short window TFPF are selected to denoise the noise segment and mixed segment respectively, which can effectively reduce the interference of noise;

[0128] 2. Before processing the feature fragments, use NSGAII to optimize ELM, including the number of hidden layer neurons, the type of activation function, and the range of input layer connection weights, which can greatly improve the temperature compensation effect;

[0129] 3. The optimization target is determined by the mean square error (RMSE) of the predicted output and the 2-norm of the output layer connection weight. This makes ELM have better generalization ability and higher prediction accuracy. Compared with BP neural network and ELM, the predicted value of the improved ELM is closest to the actual value, which improves the compensation accuracy.

[0130] An embodiment of the present invention further provides a storage medium for storing a computer program, which at least performs the above method when executed.

[0131] An embodiment of the present invention further provides a control device, comprising a processor and a storage medium for storing a computer program; wherein the processor is configured to execute at least the method described above when executing the computer program.

[0132] An embodiment of the present invention further provides a processor, wherein the processor executes a computer program and at least executes the method described above.

[0133] The storage medium can be implemented by any type of non-volatile storage device, or a combination thereof. Among them, the non-volatile memory can be a read-only memory (ROM), a programmable read-only memory (PROM), an erasable programmable read-only memory (EPROM), an electrically erasable programmable read-only memory (EEPROM), a magnetic random access memory (FRAM), a ferromagnetic random access memory (Flash Memory), a magnetic surface memory, an optical disc, or a compact disc read-only memory (CD-ROM); the magnetic surface memory can be a disk memory or a tape memory. The storage medium described in the embodiments of the present invention is intended to include, but is not limited to, these and any other suitable types of memory.

[0134] In the several embodiments provided by the present invention, it should be understood that the disclosed systems and methods can be implemented in other ways. The device embodiments described above are only schematic. For example, the division of the units is only a logical function division. There may be other division methods in actual implementation, such as: multiple units or components can be combined, or can be integrated into another system, or some features can be ignored or not executed. In addition, the coupling, direct coupling, or communication connection between the components shown or discussed can be through some interfaces, and the indirect coupling or communication connection of the devices or units can be electrical, mechanical or other forms.

[0135] The units described above as separate components may or may not be physically separated, and the components displayed as units may or may not be physical units, that is, they may be located in one place or distributed on multiple network units; some or all of the units may be selected according to actual needs to achieve the purpose of the present embodiment.

[0136] In addition, all functional units in the embodiments of the present invention may be integrated into one processing unit, or each unit may be separately used as a unit, or two or more units may be integrated into one unit; the above-mentioned integrated units may be implemented in the form of hardware or in the form of hardware plus software functional units.

[0137] A person skilled in the art can understand that: all or part of the steps of implementing the above method embodiment can be completed by hardware related to program instructions, and the aforementioned program can be stored in a computer-readable storage medium. When the program is executed, it executes the steps of the above method embodiment; and the aforementioned storage medium includes: mobile storage devices, read-only memories (ROM, Read-Only Memory), random access memories (RAM, Random Access Memory), disks or optical disks, etc. Various media that can store program codes.

[0138] Alternatively, if the above-mentioned integrated unit of the present invention is implemented in the form of a software function module and sold or used as an independent product, it can also be stored in a computer-readable storage medium. Based on this understanding, the technical solution of the embodiment of the present invention can be essentially or partly reflected in the form of a software product that contributes to the prior art. The computer software product is stored in a storage medium and includes several instructions for a computer device (which can be a personal computer, a server, or a network device, etc.) to execute all or part of the methods described in each embodiment of the present invention. The aforementioned storage medium includes: various media that can store program codes, such as mobile storage devices, ROM, RAM, magnetic disks or optical disks.

[0139] The methods disclosed in the several method embodiments provided by the present invention can be arbitrarily combined without conflict to obtain new method embodiments.

[0140] The features disclosed in several product embodiments provided by the present invention can be arbitrarily combined without conflict to obtain new product embodiments.

[0141] The features disclosed in several method or device embodiments provided by the present invention can be arbitrarily combined without conflict to obtain new method embodiments or device embodiments.

[0142] The above contents are further detailed descriptions of the present invention in combination with specific preferred embodiments, and it cannot be determined that the specific implementation of the present invention is limited to these descriptions. For those skilled in the art of the present invention, several equivalent substitutions or obvious variations can be made without departing from the concept of the present invention, and the performance or use is the same, which should be regarded as belonging to the protection scope of the present invention.

Claims

1. A temperature compensation method for a gyroscope, especially for a MEMS gyroscope, characterized in that: The steps include: S1. After adding specific noise, decomposing the output signal of the gyroscope into a series of intrinsic mode functions IMF; wherein the specific noise is a subset selected from a Gaussian white noise sequence according to a predetermined rule; S2. Using sample entropy SE, the decomposed intrinsic mode function IMF is divided into a noise segment, a mixed segment and a characteristic segment; wherein the noise segment mainly contains noise, the mixed segment contains noise and useful signals, and the characteristic segment contains a drift term caused by temperature change; S3, using a long window time-frequency peak filter TFPF to denoise the noise segment, and using a short window time-frequency peak filter TFPF to denoise the mixed segment while retaining useful signals; S4, using an extreme learning machine ELM optimized by a non-dominated sorting genetic algorithm NSGA II to perform temperature compensation on the characteristic segment containing a temperature drift term; S5. Combining the denoised noise segment and the mixed segment and the temperature compensated feature segment to reconstruct the final gyroscope output signal.

2. The temperature compensation method for a gyroscope according to claim 1, characterized in that: In step S1, a white noise E derived from the kth component of Gaussian white noise is selected. k [w (i) ] As a specific noise aid, the signal decomposition is performed by the empirical mode decomposition (EMD) method.

3. The temperature compensation method for a gyroscope as claimed in claim 2, characterized in that: Step S1 specifically includes: S11. Add a specific noise w with a standard deviation of β0 to the output signal of the gyroscope. (i) ; S12, using an empirical mode decomposition (EMD) algorithm to decompose the signal with specific noise, and sequentially calculate the first to kth IMF values; S13, combining the k IMF values ​​obtained by decomposition calculation and the residual term obtained finally to reconstruct an approximate representation of the original signal.

4. The temperature compensation method for a gyroscope as claimed in claim 3, characterized in that: In step S12, for the original signal x (i) , calculate its local average value as the first residual r1 through the EMD algorithm, from the original signal x (i) Subtract the residual r1 from the original signal, and the difference is the first IMF value c1; for the remaining signals, the current residual is used as the new original signal, and a new local average is calculated, and then the new local average is subtracted from the current residual to obtain a new IMF value. Repeat this process until the predetermined number of IMF values ​​k is reached.

5. The temperature compensation method for a gyroscope according to any one of claims 1 to 4, characterized in that: In step S3, the signal is filtered and denoised using a time-frequency peak filtering TFPF algorithm based on WVD and instantaneous frequency parameter estimation principles, wherein when processing multi-component signals, a pseudo Wiener-Ville distribution PWVD is integrated for signal processing.

6. The temperature compensation method for a gyroscope according to any one of claims 1 to 5, characterized in that: In step S4, before using the extreme learning machine ELM for temperature compensation, the number of neurons in the hidden layer, the type of activation function and the range of the input layer connection weights are optimized by using the non-dominated sorting genetic algorithm NSGAII.

7. The temperature compensation method for a gyroscope according to any one of claims 1 to 6, characterized in that: In step S4, using a non-dominated sorting genetic algorithm NSGAII to optimize the extreme learning machine ELM includes: The parameters to be optimized in ELM are mapped to the population individuals in the NSGA-II algorithm, the optimization interval of hidden neurons is set, and multiple types of activation functions are selected for testing; the scaling factor λ is introduced to dynamically adjust the range of the input layer connection weights: In ij ∈λ·[-1,1]λ∈[0,1] Among them, w ij represents the input layer connection weight, and λ is the scaling factor.

8. The temperature compensation method for a gyroscope according to any one of claims 1 to 7, characterized in that: In step S4, using the non-dominated sorting genetic algorithm NSGAII to optimize the extreme learning machine ELM includes: using the mean square error RMSE and the 2-norm of the output weight as the objective function, as follows: Where N is the number of input samples, y p,i and t,i are the predicted value and true value of the first training sample.

9. The temperature compensation method for a gyroscope according to any one of claims 1 to 8, characterized in that: In step S4, the algorithm process of NSGA II includes: first, randomly generate an initial population P containing N individuals t ; Then, the basic operation of genetic algorithm is used to obtain the offspring population Q t ; Then, by reconstructing P t and Q t , and obtain a new population R with a population size of 2N t ; According to the elite strategy, P is obtained by selecting the best N individuals in Rt t+1 , and obtain its offspring population Q again through genetic algorithm t+1 , construct a new population R t+1 , and this process is iterated until the stopping condition is met.

10. A computer-readable storage medium storing a computer program, characterized in that: When the computer program is executed by a processor, the temperature compensation method for a gyroscope is implemented as described in any one of claims 1 to 9.

Images