Gyro array fusion system and method based on improved VMD and BP
By improving the combination of VMD and BP neural network, the problem of accuracy loss in angular random walk and zero-bias instability is solved, and a higher accuracy signal fusion and noise reduction effect is achieved.
Patent Information
- Application Number
- CN202411792870.7
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2024-12-08
- Publication Date
- 2025-05-06
AI Technical Summary
When gyroscopes face random angle walks and zero-bias instability, it is difficult for them to meet higher accuracy requirements, and the prior art is difficult to effectively reduce noise and improve signal fusion accuracy.
The improved variational autoencoder (VMD) is used for signal preprocessing, combined with the BP neural network for signal fusion, and the improved VMD decomposition and noise reduction are used for accurate signal prediction and fusion.
The zero-side instability and angle random walk are significantly reduced, which improves the accuracy of gyro fusion data, which is specifically manifested as zero-side instability is reduced to 3% of the original and angle random walk is reduced to 16% of the original.
Smart Images
Figure CN119935177A_ABST
Abstract
Description
Technical Field
[0001] The invention belongs to the technical field of gyroscope and data representation, and proposes a gyroscope array fusion system and method based on improved VMD and BP. Background Art
[0002] Array gyros usually use multi-sensor data fusion to improve the accuracy and stability of navigation systems. Since the information collected by various sensors has random drift errors, it will change with the changes in the external environment without clear rules, making it difficult to establish an accurate error model. Therefore, studying the random error identification and noise reduction methods of gyro output data is crucial for array gyro signal fusion.
[0003] Usually, data is preprocessed to improve the accuracy of the data as much as possible. Usually, the preprocessing method of signal decomposition and reconstruction is adopted, including empirical mode decomposition, local mean decomposition and its related derivative methods.
[0004] This application aims to explore the composition of the array gyro system and the fusion method to combat random errors and reduce noise, thereby improving the accuracy of gyro fusion data, as the gyro cannot meet higher precision requirements in terms of angle random walk and zero bias instability. Summary of the invention
[0005] The purpose of the present invention is to solve the accuracy loss of the gyroscope when facing angle random walk and zero bias instability, and propose a gyroscope array fusion system and method based on improved VMD and BP. The system can realize noise reduction processing on the output signal of the low-precision gyroscope array module by adopting improved VMD to pre-process the single-axis gyroscope signal for noise reduction, and at the same time combine the improved BP neural network to realize signal fusion to improve the design, aiming to optimize the performance of the low-precision gyroscope, so as to meet the complex high-precision navigation and stability control requirements with the help of low-cost low-precision gyroscope.
[0006] In order to achieve the above object, the present invention adopts the following technical scheme: As a first aspect of the present invention, a gyro array fusion system based on improved VMD and BP is proposed, comprising: a temperature control module, a raw data acquisition module, a preprocessing module, a decomposition and reconstruction module and a fusion model processing module; wherein the raw data acquisition module is controlled by the temperature control module, and the collected data needs to be preprocessed by the preprocessing module, and then sent to the decomposition and reconstruction module, and then sent to the fusion model processing module after the data is decomposed and reconstructed.
[0007] The raw data acquisition module includes a plurality of inertial measurement units for acquiring raw angular velocity data within a preset temperature range; The temperature control module includes one or more temperature control units, which include: a control subunit, a heating subunit, a cooling subunit, and a temperature sensor, which controls the working temperature of the raw data acquisition module within a preset temperature range, and provides a temperature feedback signal for the preprocessing module. Specifically, the temperature sensor collects the ambient temperature, and the control subunit gives a control signal based on the set temperature to control the heating subunit for heating, and / or controls the cooling subunit for cooling, thereby achieving temperature control, so that the raw data acquisition module works within the preset temperature range.
[0008] The preprocessing module includes a temperature compensation unit. The temperature control unit sends a temperature feedback signal of the temperature sensor to the temperature compensation unit. The temperature compensation unit then compensates the angular velocity signal output by the high-pass filter. The temperature compensation is specifically achieved by sending the temperature feedback signal to a temperature error neural network, and then subtracting the angular velocity error output by the temperature error neural network from the original angular velocity. The temperature error neural network is a BP model, which is used to predict the output angular velocity error based on the temperature feedback signal.
[0009] The decomposition and reconstruction module comprises decomposing the corrected angular velocity and reconstructing it according to the attribute weights determined by the information entropy of the decomposed data to obtain reconstructed angular velocity data; The fusion model processing module performs training and testing based on the reconstructed angular velocity data and relies on the BP model to obtain a predicted output fusion angular velocity signal.
[0010] The raw angular velocity data is specifically obtained by acquiring multiple channels of raw angular velocity data within a preset temperature range through a raw data acquisition module according to a set mode and filtering out low-frequency noise through a high-pass filter.
[0011] The preset temperature range is from -40 to +85 degrees Celsius.
[0012] The number of the inertial measurement units is greater than or equal to 2 and less than or equal to 200.
[0013] The training and testing are performed to obtain a predicted output fused angular velocity signal, specifically: after model training is performed according to a training set and a validation set constructed using the reconstructed angular velocity data, testing is performed based on the trained model to output a fused predicted value.
[0014] The training set and the validation set are specifically obtained through the following process: setting the turntable mode of the inertial measurement unit and collecting data through the inertial measurement unit, and then preprocessing, decomposing and reconstructing the data, and the obtained reconstructed signal is decomposed into the training set and the validation set.
[0015] As a second aspect of the present invention, a gyro array fusion method based on improved VMD and BP is proposed, comprising: Acquire multiple channels of raw angular velocity data within a preset temperature range; Performing preprocessing including filtering on each channel of the original angular velocity data to obtain corrected angular velocity data; Decomposing the corrected angular velocity data using improved VMD, and reconstructing the decomposed data according to information entropy and attribute weights to obtain reconstructed angular velocity data; The reconstructed angular velocity data is fused based on a neural network model and a fused angular velocity signal is predicted and output; the neural network model is a BP fusion model.
[0016] Beneficial Effects Compared with the prior art, the gyro array fusion system and method based on improved VMD and BP proposed in the present invention have the following beneficial effects: 1. The gyro array fusion method predicts and outputs the fusion signal based on the improved VMD decomposition and combined with the BP neural network for the original output signal of the MEMS gyro array. The improved VMD decomposition can reduce noise and the BP neural network model can accurately predict the fusion signal. 2. The system and method can be compared to see the performance of the angular velocity signal without data processing, zero bias instability: 0.67887° / h, angle random walk: 14.2189 , bias stability: 14.2189° / h; bias instability after data fusion: 0.019882° / h, angle random walk: 2.2835 , bias stability: 2.2835° / h; bias instability is reduced to 3% of the original, angle random walk is reduced to 16% of the original, and bias instability is reduced to 16% of the original. BRIEF DESCRIPTION OF THE DRAWINGS
[0017] The drawings described herein are used to provide a further understanding of the present invention and constitute a part of the present invention. The exemplary embodiments of the present invention and their descriptions are used to explain the present invention and do not constitute improper limitations on the present invention. In the drawings: Figure 1 It is a composition and connection block diagram of a gyro array fusion system based on improved VMD and BP in an embodiment of the present invention; Figure 2 It is a flowchart of a gyro array fusion method based on improved VMD and BP in an embodiment of the present invention; Figure 3 It is a schematic diagram of a temperature control unit of a gyro array fusion system based on improved VMD and BP in an embodiment of the present invention; Figure 4It is a schematic diagram of a temperature compensation unit of a gyro array fusion system based on improved VMD and BP in an embodiment of the present invention; Figure 5 It is a schematic diagram of the composition of the fusion model of the gyro array fusion system based on the improved VMD and BP in an embodiment of the present invention; Figure 6 It is a schematic diagram of a model training unit of a gyro array fusion system based on improved VMD and BP in an embodiment of the present invention; Figure 7 This is a comparison diagram of data reconstruction using information entropy and data reconstruction using classification Kalman filtering in an embodiment of the present invention; Figure 8 Schematic diagram of comparison of angular velocity signals in different temperature ranges and different stages of data fusion processing in an embodiment of the present invention. DETAILED DESCRIPTION
[0018] In order to clearly describe the technical solutions of the embodiments of the present invention, in the embodiments of the present invention, words such as "first" and "second" are used to distinguish the same items or similar items with basically the same functions and effects. For example, the first threshold and the second threshold are only used to distinguish different thresholds, and their order is not limited. Those skilled in the art can understand that words such as "first" and "second" do not limit the quantity and execution order, and words such as "first" and "second" do not necessarily limit them to be different.
[0019] It should be noted that, in the present invention, words such as "exemplary" or "for example" are used to indicate examples, illustrations or descriptions. Any embodiment or design described as "exemplary" or "for example" in the present invention should not be interpreted as being more preferred or more advantageous than other embodiments or designs. Specifically, the use of words such as "exemplary" or "for example" is intended to present related concepts in a specific way.
[0020] In the present invention, "at least one" means one or more, and "plurality" means two or more. "And / or" describes the association relationship of associated objects, indicating that three relationships may exist. For example, A and / or B can mean: A exists alone, A and B exist at the same time, and B exists alone, where A and B can be singular or plural. The character " / " generally indicates that the objects associated before and after are in an "or" relationship. "At least one of the following" or similar expressions refers to any combination of these items, including any combination of single or plural items. For example, at least one of a, b or c can mean: a, b, c, the combination of a and b, the combination of a and c, the combination of b and c, or the combination of a, b and c, where a, b, c can be single or multiple.
[0021] The technical solution of the present application is described in detail below with reference to the accompanying drawings and specific embodiments.
[0022] like Figure 1 As shown, a schematic diagram of a gyro array fusion system based on improved VMD and BP is given, wherein the system comprises: a raw data acquisition module 21, a preprocessing module 22, a decomposition and reconstruction module 23, a fusion model processing module 24, and a temperature control module 26. The temperature control module 26 is connected to the raw data acquisition module 21 and the preprocessing module 22, the preprocessing module 22 is connected to the decomposition and reconstruction module 23, and the decomposition and reconstruction module 23 is connected to the fusion model processing module 24.
[0023] The raw data acquisition module 21 is configured to acquire multiple channels of raw angular velocity data within a preset temperature range; A preprocessing module 22 is configured to perform preprocessing including filtering on each channel of the original angular velocity data to obtain corrected angular velocity data; A decomposition and reconstruction module 23 is configured to decompose the corrected angular velocity data, and reconstruct the decomposed data according to information entropy and attribute weights to obtain reconstructed angular velocity data; The fusion model processing module 24 is configured to perform fusion processing on the reconstructed angular velocity data based on a neural network model and predict and output a fusion angular velocity signal. The neural network model is a BP fusion model; The temperature control module 26 is configured to control the raw data acquisition module 21 to operate within a preset temperature range, and provide a temperature feedback signal to the preprocessing module 22. The preprocessing module 22 performs temperature compensation correction on each channel of the raw angular velocity data according to the temperature feedback signal.
[0024] Furthermore, the gyro array fusion system also includes: The fusion model training module 25 is configured to perform model training based on a training set and a validation set including the reconstructed angular velocity data and the prediction data, and provide a neural network model to the fusion model processing module; the neural network model is a BP fusion model.
[0025] The raw data acquisition module is a gyroscope array, the preprocessing module is an improved VMD preprocessing module, and the neural network model is a BP fusion model. The gyroscope array includes a plurality of inertial measurement units, and the inertial measurement unit is a gyroscope. Preferably, the gyroscope is a MEMS gyroscope.
[0026] like Figure 2 As shown, a flow chart of a gyro array fusion method based on improved VMD and BP is shown, and the gyro array fusion method comprises the following steps: Step S1, obtaining raw angular velocity data of each gyroscope in the gyroscope array within a preset temperature range; Step S2, performing preprocessing including filtering on each channel of the original angular velocity data to obtain corrected angular velocity data; Step S3, decomposing the corrected angular velocity data, and reconstructing the decomposed data according to the information entropy and attribute weights to obtain reconstructed angular velocity data; Step S4, performing fusion processing on the reconstructed angular velocity data based on a neural network model and predicting and outputting a fused angular velocity signal; the neural network model is a BP fusion model.
[0027] The temperature control module 26 includes one or more temperature control units. Figure 3 As shown, each temperature control unit includes: a control subunit, a heating subunit, a cooling subunit, and a temperature sensor. The temperature sensor is an integrated digital temperature sensor responsible for monitoring the temperature of the gyroscope array and feeding back the temperature signal to the control subunit. The heating subunit and the cooling subunit receive the control signal of the control subunit and are responsible for adjusting the temperature of the gyroscope array. The control subunit determines whether the temperature signal fed back by the temperature sensor is within the preset temperature range. If temperature adjustment is required, it outputs the current or voltage control signal of the heating subunit and the cooling subunit. The control signal is transmitted to the heating subunit and the cooling subunit through the drive circuit to adjust their cooling or heating effects to maintain the required preset temperature range.
[0028] like Figure 4 As shown, the preprocessing module includes a temperature compensation unit, and the temperature compensation unit includes a temperature error neural network and an error compensation subunit. The temperature error neural network obtains a data acquisition temperature value from a temperature sensor, and predicts an angular velocity error caused by an output temperature according to the temperature value. The error compensation subunit uses the angular velocity error caused by the temperature to compensate the original angular velocity signal to obtain an angular velocity signal after temperature compensation. The temperature error neural network is a BP model, a GRU model, or a LSTM model.
[0029] like Figure 5 As shown, a more detailed schematic diagram of the gyro array fusion system based on improved VMD and BP is further given. The temperature control unit in the temperature control module 26 is set as one, which is used to control the temperature of multiple inertial measurement units in the entire raw data acquisition module 21 to remain within the required preset temperature range.
[0030] exist Figure 5In the gyro array fusion system shown, the raw data acquisition module 21 includes multiple inertial measurement units, which are gyroscopes. The multiple gyroscopes form a gyro array as a raw data acquisition module. The gyroscopes are MEMS gyroscopes. The temperature control module 26 includes one or more temperature control units, which are used to measure the temperature of the inertial measurement unit and control the temperature within a preset temperature range by heating or cooling.
[0031] The preprocessing module 22 includes multiple high-pass filters and temperature compensation units. After the inertial measurement unit collects the original angular velocity data, it is filtered through a high-pass filter, and then the temperature compensation unit performs temperature compensation on the filtered data according to the current working temperature based on a preset temperature compensation curve, or compensates and corrects the error output according to the neural network prediction.
[0032] The pre-processed angular velocity data enters the decomposition and reconstruction module 23, which includes a plurality of decomposition units and a reconstruction unit. The decomposition unit is used to decompose the pre-processed angular velocity data to obtain a plurality of IMF components and a group of residual value components, respectively calculate the information entropy corresponding to the attribute value of each component, and determine the attribute weight vector of each component according to the information entropy of the attribute value of each component; the reconstruction unit is used to perform weighted reconstruction on the plurality of IMF components and the group of residual value components obtained by the decomposition according to the attribute weight vector of each component.
[0033] The reconstructed multi-channel angular velocity data enters the fusion model processing module 23, which uses the BP model to perform fusion processing on the reconstructed angular velocity data and predict and output a fused angular velocity signal.
[0034] The BP model is obtained by the model training unit of the fusion model training module 25 according to the training set, and the training set is composed of reconstructed angular velocity data and corresponding predicted output signals. The model training unit performs model training based on the training set and verification set including the reconstructed angular velocity data and predicted data, and provides the BP neural network model to the fusion model processing module.
[0035] like Figure 6 As shown, a schematic diagram of the model training unit is given, and the model training unit includes a BP model sub-unit. The training set and the verification set generated according to the reconstructed signal are input into the BP model sub-unit. During the model training process, the model training parameters are regenerated according to the BP model training results in each training cycle to form a trained BP model. After the effect of the BP model is verified by using a high-precision inertial signal, the trained BP model is updated to the fusion model processing module 24.
[0036] The peak-to-peak value of the weight of the BP model is limited to a closed interval of [-a, a]; ; is the number of input units of the weight matrix, which is the same as the number of gyroscopes, is the number of output units of the weight matrix and .
[0037] Example 1 The gyro array fusion system based on improved VMD and BP is specifically implemented as follows: the raw data acquisition module 21 includes a gyro array composed of K inertial measurement units, and may also include a turntable mode setting unit for setting the turntable mode of the gyro array; the temperature control unit of the temperature control module includes a control subunit, a temperature sensor, a heating subunit and a cooling subunit; the preprocessing module includes a high-pass filter and a temperature compensation unit; the decomposition and reconstruction module includes a decomposition unit based on improved VMD decomposition and a reconstruction unit based on information entropy; the model training unit includes a Xavier subunit and a BP model subunit for initialization, and the fusion model processing module includes a trained BP neural network model.
[0038] In the specific implementation, the temperature sensor is an integrated digital temperature sensor (model LM75), which is responsible for monitoring the temperature of the array fusion system. The inertial measurement unit is responsible for collecting relevant data; the preprocessing module is implemented using an STM32 signal processor, which is responsible for executing the corresponding control and algorithm, and the obtained results are transmitted to the host computer through the RS422 serial port; the host computer collects the corresponding angular velocity data and processes the output angular velocity data. The decomposition and reconstruction module, the fusion model processing module, and the fusion model training module are located in the host computer.
[0039] In specific implementation, it is assumed that the turntable mode set by the turntable mode setting unit is a sine mode; the turntable mode setting unit is located in the raw data acquisition module. The inertial measurement unit in the raw data acquisition module collects the sinusoidal changes. The dynamic angular velocity signal is collected by the original data acquisition module, and then passes through the preprocessing module and the decomposition and reconstruction module to output the reconstructed signal as the training set and the verification set, and the corresponding high-precision inertial signal is used as the learning group. Through training, a BP fusion model that meets the requirements is obtained. The reconstructed signal is a sinusoidal signal. The reconstructed signal is an angular velocity signal; the training set and the validation set are both partial signals of the reconstructed signal in this part of the time period; the high-precision inertial signal is the true value of the angular velocity, which is used to measure the quality of the BP network obtained by training.
[0040] In the process of training the BP neural network model, the first step is to collect the sinusoidal data at room temperature. Dynamic raw angular velocity data that changes the rotation speed in a certain way; represents the angular velocity, Represents the acquisition time; Taking the Y axis as an example, the 4 Y axis original signals are represented as (Y1, Y2, Y3, Y4). These four sets of dynamic original angular velocity data are decomposed and reconstructed by the improved VMD algorithm to obtain reconstructed angular velocity signals. These reconstructed angular velocity signals are used as training sets and input into the input layer of the BP neural network model; The 4 Y axis reconstructed signals are represented as (Y1', Y2', Y3', Y4'); These four reconstructed signals are simultaneously input into a trained BP neural network with an input dimension of 4 and an output dimension of 1; Then they are processed through a three-layer fully connected network to obtain the fusion result of the training model and save the model.
[0041] Signal decomposition helps to distinguish different types of gyro signals and improve signal processing effects. Signal decomposition methods include wavelet transform, singular value decomposition, independent component analysis, and local mean decomposition. Wavelet transform can effectively extract specific frequency components, singular value decomposition is used to suppress noise, independent component analysis realizes blind source separation, and improved VMD achieves noise reduction through adaptive local mean decomposition, especially in processing nonlinear signals. The comprehensive application of these decompositions helps to extract key information and suppress noise, thereby improving the performance and accuracy of the gyroscope. Compared with the original VMD method, the improved VMD method can more accurately determine the decomposition coefficient and penalty factor, which is simpler and more accurate in capturing signals and local features. The improved VMD uses KL divergence to determine the penalty factor α and the number of decomposition layers k, which can prevent over-decomposition and screen out useful information and noise in the decomposed IMF according to the information entropy weight. Then, the machine learning model is used for fusion processing to improve the accuracy of signal data.
[0042] Example 2 The gyro array fusion method based on improved VMD and BP includes the following steps: S201, obtaining raw angular velocity data of the gyroscope array; The gyroscope array is located in the raw data acquisition module, and the raw angular velocity data of the gyroscope array includes The angular velocity data of the three axes X, Y, and Z generated by the gyroscope; Greater than or equal to 2 and less than or equal to 200.
[0043] In a specific implementation, the gyroscope array includes 4 gyroscopes, and may also include 6, 8, 10, 12, 14 or 16 gyroscopes; the raw angular velocity data generated by each gyroscope in the gyroscope array is divided into three groups of data, corresponding to the angular velocity data of the X, Y and Z axes respectively.
[0044] S202: Based on the improved VMD, the original angular velocity data acquired in S201 is filtered and denoised by the improved VMD, and temperature correction may also be performed to obtain corrected angular velocity data.
[0045] S203, decomposing the corrected angular velocity data, calculating information entropy and attribute weights, and reconstructing the signal according to the attribute weights; S204, establishing a BP fusion model; S205 , input the reconstructed signal outputted from S203 into the BP fusion model established in S204 to obtain angular velocity fusion data.
[0046] The decomposition and reconstruction of step 203 includes the following sub-steps: S2031, performing I-modified VMD decomposition on the corrected angular velocity data to obtain IMF components; S2032, averaging the IMF components of each order to obtain an improved VMD decomposition result of the corresponding order; The IMF components of each order are M in total, the first M-1 orders are IMF components, and the last order is a residual value; M is greater than or equal to 5 and less than or equal to 15; specifically in this embodiment, M=6; S2033, calculating attribute weights for the IMF components obtained by the improved VMD decomposition and a group of residual values according to their information entropy; the information entropy is calculated using the following formula: ; in, Is the attribute value The number of values of ; Is the attribute value The probability of The value of is equal to the number of sample points actually collected from the original angular velocity data; specifically in this embodiment, =72000; Corresponding to The order IMF component, The value range of is 1 to M; the M-th order IMF component corresponds to the residual value; the attribute weight is calculated by the following formula: ;in, It is The first attribute is The weight of the PF component of order, It is information about an attribute.
[0047] S2034. Add the calculated attribute weight to the improved VMD decomposition result of the corresponding order in step S2033 to obtain a reconstructed signal.
[0048] In step S204, a BP fusion model is established; the BP fusion model is a BP neural network. The peak-to-peak value of the weight of the BP network is limited to a closed interval of [-a, a], where , is the number of input units of the weight matrix, which is the same as the number of gyroscopes. is the number of output units of the weight matrix and this ; The value range is greater than or equal to 2 and less than or equal to 200.
[0049] The input of the BP fusion model is the reconstructed signal, and the output is the predicted output angular velocity fusion data, that is, three groups of angular velocity fusion data of the X, Y, and Z axes are fused.
[0050] Example 3 The array gyro acquisition board of the inertial measurement unit (IMU), where the IMU model is MPU6050 and the sampling frequency of the inertial measurement unit is 50Hz. The data of the four gyroscopes were acquired through data acquisition, including a total of 12-axis angular velocity data, and divided into three groups, namely (X1, X2, X3, X4), (Y1, Y2, Y3, Y4), and (Z1, Z2, Z3, Z4). Then, based on the improved VMD, the original angular velocity data is filtered and denoised by the improved VMD, and then the information entropy and attribute weights are calculated, and the reconstructed signal is obtained according to the attribute weights, which specifically includes the following steps: S301, performing improved VMD decomposition on the corrected angular velocity data to obtain IMF components; S302, averaging the IMF components of each order to obtain an improved VMD decomposition result of the corresponding order; The IMF components of each order are M in total, the first M-1 orders are IMF components, and the last order is a residual value; in specific implementation, M is 7. The components of each order are 7 in total, the first 6 orders are IMF components, and the last order is a residual value.
[0051] S303, calculating attribute weights based on the information entropy of the IMF components obtained by the improved VMD decomposition and a group of residual values; information entropy Use the following formula to calculate: ; in, is the number of possible values for the attribute. The value is 72000. Is the attribute value The probability of the attribute value Corresponding to The order IMF component, The value range of is 1 to M; the Mth-order IMF component corresponds to the residual value, and the value of M is 7. Calculated by the following formula: ;in, It is The first attribute is The weights of the order IMF components, It is information about an attribute.
[0052] S304: The calculated attribute weights The reconstructed signal is obtained by weighting the improved VMD decomposition result of the corresponding order in step S303. The value of M is 7, and the attribute weight vector of the 7th order IMF component is obtained. The calculated attribute weight is weighted to the improved VMD decomposition result corresponding to the 7th order (i.e., the 6th order IMF component and the last 1st order residual value), and finally the reconstructed signal is obtained.
[0053] S305, input the reconstructed signal into the BP fusion model to obtain a predicted fusion signal. Finally, the reconstructed signal is input into the BP fusion model to obtain angular velocity fusion data; the angular velocity fusion data is three sets of angular velocity fusion data that fuse the three axes of X, Y, and Z.
[0054] The peak-to-peak value of the weight of the BP model network is limited to a closed interval of [-a, a], , is the number of input units of the weight matrix, which is the same as the number of gyroscopes. is the number of output units of the weight matrix, ; This example The value is 4. The value is 1.
[0055] BP neural network is a typical multi-layer feedforward type (usually three layers) artificial neural network. Structurally, it consists of input, hidden and output layers, and each layer has nodes. The nodes of adjacent layers are connected by weights, but the nodes in each layer are independent of each other. In BP neural network, the selection of initial weights is crucial to the training efficiency and stability of the network. Xavier initialization aims to provide a suitable initial value for the weights of the neural network to promote the stability and convergence of the training and avoid the gradient vanishing or gradient exploding problems of deep neural networks; the Xavier initialization takes into account the number of weights and the network structure to ensure that the initialized weights do not cause the gradient to be too small or too large; Xavier initialization is to randomly select weight values from a uniform distribution so that they are in the range of [-a, a], where a is a scaling factor related to the number of input and output units. The scaling factor a is calculated based on the size of the weight matrix.
[0056] In each unit of the BP network, at time step t, the input layer , hidden layer , output layer Three layers are used to regulate and manage the information flow of each unit in the BP network: Input Layer : The input layer is responsible for receiving external stimuli or data. The number of nodes in the input layer usually corresponds to the dimension of the input data. In addition, the nodes in the input layer and the nodes in the hidden layer transmit information through weighted connections. These weights will be adjusted during the neural network training process to optimize the network performance.
[0057] Hidden Layer : The hidden layer is the core part of the BP neural network, which is responsible for processing input data and performing nonlinear transformation through activation functions.
[0058] For each neuron in the hidden layer, its input is the weighted sum of the outputs of all neurons in the previous layer. The weighted sum of the inputs to the neurons is , then: Where n is the number of neurons in the previous layer, It is Neuron to The weight of a neuron, It is The output of neurons (for the input layer, is the input data), It is The bias term of a neuron.
[0059] The output of the hidden layer neurons is calculated through the activation function, and the Sigmoid function is used here. The output of a neuron for: Output Layer : The output layer is the last layer of the BP neural network. It receives the processing results from the hidden layer and further processes them through the activation function to finally generate the output of the network. The output layer usually uses a specific activation function to process the output of the hidden layer. The activation function used here is the sigmoid function. Therefore, the calculation formula of the output layer of the BP neural network can usually be expressed as: in, The output layer The output of a neuron, is the activation function, The output layer The input of a neuron is calculated as: in, To connect the hidden layer neurons and the output layer The weight of a neuron, For the hidden layer The output of a neuron, The output layer The bias term of a neuron.
[0060] During the training process of the BP neural network, the output of the output layer is compared with the expected output, and the weights and biases in the network are adjusted through the back propagation algorithm to reduce the prediction error. The back propagation algorithm is based on the chain rule and gradient descent, and updates these parameters by calculating the gradient of the output layer error with respect to the weights and biases.
[0061] For updating the output layer weights and biases, the following formula is usually used: in, is the learning rate, is the cost function (such as mean square error), and are the partial derivatives of the cost function with respect to weight and bias, respectively.
[0062] The process of training the X-BP neural network model is as follows: The dynamic raw angular velocity data of the rotation speed is changed in a certain way, and they are decomposed into training set, validation set and test set. Then the weights and biases of the neural network are initialized by Xavier. The training set data is passed from the input layer to the output layer through the hidden layer. The actual output of the process network is output after calculation by the output layer. At the same time, the error between the actual output and the expected output of the sample data is collected, and the error information is back-propagated to the input layer based on the obtained error information. At the same time, the weights and thresholds between the layers are adjusted; the validation set is used to detect whether the model is overfitting, so as to select the best model. The test set will be used to finally evaluate the performance of the selected model.
[0063] The 4 sets of collected original angular velocity data are decomposed and reconstructed by the improved VMD algorithm. The four X-axis signals are represented as (X1', X2', X3', X4') respectively. These four signals are simultaneously input into a BP neural network with an input dimension of 4 and an output dimension of 1. A set of static data of a high-precision quartz gyroscope is used as the learning target. After the neural network predicts and outputs a set of angular velocity data, finally, these four signals are processed and fused by the neural network to generate an X-axis angular velocity signal, which is the X-axis output value after the array gyroscope is fused. The same processing method is also applicable to the Y-axis and Z-axis data, and finally three-way X, Y, and Z-axis angular velocity fusion data are generated.
[0064] like Figure 7 As shown in the figure, a schematic diagram of the comparison of the angular velocity fusion signal obtained by the information entropy reconstruction method and the classification Kalman filter reconstruction method is given. It can be seen that the information entropy reconstruction method outputs a reconstructed signal that is closer to the true value than the reconstruction method in previous studies. The experimental environment of the collected signal in the figure is static and at room temperature, and the true value is 0° / s.
[0065] like Figure 8 As shown in the figure, a comparison of angular velocity signals in different temperature ranges and different stages of data fusion processing is given. It can be seen that compared with the original angular velocity, the decomposed and reconstructed signal is better than the original signal, and the fused angular velocity data finally predicted and output by the BP network is further better than the decomposed and reconstructed signal.
[0066] By comparing the raw angular velocity data of a certain axis directly collected with the angular velocity data output after noise reduction and fusion by the improved VMD combined with the BP neural network, it can be intuitively seen that the noise of the fused data processed by the algorithm is smaller and closer to 0. In order to evaluate the improvement effect on the gyroscope performance, the ALLAN variance of the fused signal is calculated to obtain the zero bias instability and angle random walk of the original data and the fused data. The performance of the unprocessed original angular velocity signal is: zero bias instability: 0.67887° / h, angle random walk: 14.2189 , bias stability: 14.2189° / h; bias instability after data fusion: 0.019882° / h, angle random walk: 2.2835 , bias stability: 2.2835° / h; bias instability is reduced to 3% of the original, angle random walk is reduced to 16% of the original, and bias instability is reduced to 16% of the original. Compared with the prior art, the original output signal of the MEMS gyroscope array based on the improved VMD decomposition combined with the BP neural network is fused, and the improved VMD decomposition can reduce noise and the memory function of the BP neural network model can improve the accuracy of the fused output signal.
[0067] Although the present invention has been described in conjunction with specific features and embodiments thereof, it is apparent that various modifications and combinations may be made thereto without departing from the spirit and scope of the present invention. Accordingly, this specification and the accompanying drawings are merely exemplary illustrations of the present invention as defined by the appended claims and are deemed to cover any and all modifications, variations, combinations or equivalents within the scope of the present invention. Obviously, those skilled in the art may make various modifications and variations to the present invention without departing from the spirit and scope of the present invention. Thus, the present invention is intended to include such modifications and variations if they fall within the scope of the claims of the present invention and their equivalents.
[0068] The present invention proposes a gyro array fusion system and method based on improved VMD and BP. The system includes: a temperature control module, a raw data acquisition module, a preprocessing module, a decomposition and reconstruction module and a fusion model processing module; the raw data acquisition module acquires raw angular velocity data within a preset temperature range; the preprocessing module corrects each of the raw angular velocity data to obtain corrected angular velocity data; the decomposition and reconstruction module decomposes the corrected angular velocity data, reconstructs the attribute weights determined according to the information entropy of the decomposed data, and obtains reconstructed angular velocity data; the fusion model processing module, according to the reconstructed angular velocity data, relies on the BP model for training and testing, and obtains a predicted output fusion angular velocity signal. The application of the present invention can meet the complex high-precision navigation and stability control requirements based on a low-precision gyroscope.
Claims
1. A gyro array fusion system based on improved VMD and BP, characterized in that: include: Temperature control module, raw data acquisition module, preprocessing module, decomposition and reconstruction module and fusion model processing module; The raw data acquisition module is controlled by the temperature control module. The collected data needs to be pre-processed by the pre-processing module and then sent to the decomposition and reconstruction module. After the data is decomposed and reconstructed, it is sent to the fusion model processing module. The raw data acquisition module includes a plurality of inertial measurement units for acquiring raw angular velocity data within a preset temperature range; The temperature control module includes one or more temperature control units, which include: a control subunit, a heating subunit, a cooling subunit, and a temperature sensor, which controls the working temperature of the raw data acquisition module within a preset temperature range, and provides a temperature feedback signal for the preprocessing module. Specifically, the temperature sensor collects the ambient temperature, and the control subunit gives a control signal based on the set temperature to control the heating subunit for heating, and / or controls the cooling subunit for cooling, thereby achieving temperature control, so that the raw data acquisition module works within the preset temperature range.
2. A gyro array fusion system based on improved VMD and BP according to claim 1, characterized in that: The preprocessing module includes a temperature compensation unit. The temperature control unit sends a temperature feedback signal of the temperature sensor to the temperature compensation unit. The temperature compensation unit then compensates the angular velocity signal output by the high-pass filter. The temperature compensation is specifically achieved by sending the temperature feedback signal to a temperature error neural network, and then subtracting the angular velocity error output by the temperature error neural network from the original angular velocity. The temperature error neural network is a BP model, which is used to predict the output angular velocity error based on the temperature feedback signal.
3. The gyro array fusion system based on improved VMD and BP according to claim 1, characterized in that: The decomposition and reconstruction module comprises decomposing the corrected angular velocity and reconstructing it according to the attribute weights determined by the information entropy of the decomposed data to obtain reconstructed angular velocity data; The fusion model processing module performs training and testing based on the reconstructed angular velocity data and relies on the BP model to obtain a predicted output fusion angular velocity signal.
4. The gyro array fusion system based on improved VMD and BP according to claim 1, characterized in that: The raw angular velocity data is specifically obtained by acquiring multiple channels of raw angular velocity data within a preset temperature range through a raw data acquisition module according to a set mode and filtering out low-frequency noise through a high-pass filter.
5. The gyro array fusion system based on improved VMD and BP according to claim 1, characterized in that: The preset temperature range is from -40 to +85 degrees Celsius.
6. The gyro array fusion system based on improved VMD and BP according to claim 1, characterized in that: The number of the inertial measurement units is greater than or equal to 2 and less than or equal to 200.
7. The gyro array fusion system based on improved VMD and BP according to claim 1, characterized in that: The training and testing are performed to obtain a predicted output fused angular velocity signal, specifically: after model training is performed according to a training set and a validation set constructed using the reconstructed angular velocity data, testing is performed based on the trained model to output a fused predicted value.
8. The gyro array fusion system based on improved VMD and BP according to claim 1, characterized in that: The training set and the validation set are specifically obtained through the following process: setting the turntable mode of the inertial measurement unit and collecting data through the inertial measurement unit, and then preprocessing, decomposing and reconstructing the data, and the obtained reconstructed signal is decomposed into the training set and the validation set.
9. A gyro array fusion method based on improved VMD and BP, characterized in that: include: Acquire multiple channels of raw angular velocity data within a preset temperature range; Performing preprocessing including filtering on each channel of the original angular velocity data to obtain corrected angular velocity data; Decomposing the corrected angular velocity data using improved VMD, and reconstructing the decomposed data according to information entropy and attribute weights to obtain reconstructed angular velocity data; The reconstructed angular velocity data is fused based on a neural network model and a fused angular velocity signal is predicted and output; the neural network model is a BP fusion model.