RLG temperature drift compensation method and system based on lightweight convolutional neural network

Abstract

The invention discloses an annular laser gyroscope temperature drift compensation method and system based on a lightweight convolutional neural network. The method comprises the following steps: acquiring angular velocities and temperature measurement values of a ring laser gyroscope RLG at different temperatures as training data, and preprocessing the training data; constructing a multi-channel time sequence input feature based on the preprocessed temperature measurement value; taking the temperature drift residual error as a training label; and designing and training a lightweight convolutional neural network model. And deploying the trained model to an RLG processing unit. During operation, the system acquires the angular velocity and the temperature measurement value of the RLG in real time, constructs a real-time multi-channel input feature based on the temperature value, predicts a real-time temperature drift error by using a deployed model, subtracts the prediction error from the real-time angular velocity measurement value, and outputs a compensated angular velocity value. According to the method, the RLG temperature excursion can be effectively and efficiently compensated by using the lightweight CNN and specific multi-channel temperature characteristics, the output precision of the RLG temperature excursion in a variable temperature environment is improved, and the model is easy to deploy.

Description

Technical Field

[0001] The present invention relates to the technical field of ring laser gyroscopes, and particularly relates to a temperature drift compensation method and system for a ring laser gyroscope (RLG) based on a lightweight convolutional neural network (Convolutional Neural Networks, CNN), which can be used for temperature drift compensation of ring laser gyroscopes. Background Art

[0002] A ring laser gyroscope (RLG) is an inertial device for measuring the angular velocity of a vehicle, and is commonly used in the inertial navigation systems of aircraft, ships, and unmanned underwater vehicles. The angular velocity measurement accuracy of the RLG is affected by temperature. When the ambient temperature fluctuates greatly, the angular velocity output of the RLG will generate a temperature drift (usually referred to as temperature drift for short) error. To solve the temperature drift problem of the RLG in a complex temperature environment, traditional methods use methods such as polynomial fitting, linear regression, and traditional BP neural networks for temperature drift compensation. However, traditional polynomial fitting or linear regression cannot effectively handle high-order non-linear errors under dynamic temperature changes, and high-precision temperature drift compensation cannot be achieved under complex temperature changes; the traditional BP neural network has a large number of parameters due to its fully connected structure, making it difficult to achieve low-latency real-time compensation on an embedded platform (such as an FPGA), and it is easy to fall into a local optimum. In addition, the traditional temperature drift method has insufficient handling of the hysteresis effect, and fails to effectively model the timing delay between temperature mutation and gyro response, resulting in a significant increase in the compensation residual as the temperature change rate increases. Summary of the Invention

[0003] The present invention proposes a temperature drift compensation method and system for a ring laser gyroscope based on a lightweight convolutional neural network. Through multi-channel time series feature fusion and physical constraint training strategies, high-precision and low-latency embedded deployment is achieved, which can significantly improve the navigation accuracy of the RLG in a complex temperature environment.

[0004] To solve the above technical problems, the technical solution proposed by the present invention is as follows:

[0005] An RLG temperature drift compensation method based on a lightweight convolutional neural network, comprising:

[0006] S1 Collect the angular velocity measurement values of the ring laser gyroscope RLG in different temperature environments and the temperature measurement values output by the internal temperature sensor of the RLG at the corresponding moments as training data, and preprocess the training data;

[0007] S2 Based on the preprocessed temperature measurement values, construct multi-channel time series input features; based on the angular velocity measurement values and the mean value of the angular velocity measurement values in the constant temperature stage, calculate the temperature drift residual as the training label; S3 designs and trains a lightweight convolutional neural network model. The lightweight convolutional neural network model takes multi-channel time series input features as input and uses the temperature drift residual as the output for training to learn the mapping relationship between the input features and the temperature drift residual. The second layer of the lightweight convolutional neural network model is a dilated convolutional layer, the third layer is a depthwise separable convolutional layer, and the sixth layer is a global average pooling layer. S4 deploys the trained lightweight convolutional neural network model to the internal processing unit of the RLG. When the RLG is running, it obtains the angular velocity measurement value and the temperature measurement value in real time. Based on the temperature measurement value obtained in real time, it constructs real-time multi-channel time series input features. It inputs the real-time multi-channel time series input features into the deployed lightweight convolutional neural network model to obtain the predicted real-time temperature drift error. It subtracts the predicted real-time temperature drift error from the angular velocity measurement value obtained in real time to obtain the compensated angular velocity value.

[0008] Furthermore, the preprocessing in step S1 is as follows: The angular velocity measurement value and the temperature measurement value are filtered using a moving average filter, and the filter window size is 10 time steps.

[0009] Furthermore, the steps of constructing multi-channel time series input features in step S2 include: Based on the temperature measurement values at times k and k-1 , , the following four channels are constructed: Channel 1 is the output value of the temperature sensor ;

[0010] Channel 2 is the temperature change rate : ;

[0011] Channel 3 is the square of the output value of the temperature sensor ;

[0012] Channel 4 is the product of the output value of the temperature sensor and the temperature change rate .

[0013] Furthermore, the training label in step S2 is the residual of the RLG measurement value : , where is the angular velocity measurement value at the current time k; is the average value of the angular velocity measurement values of the RLG during the constant temperature stage of the high and low temperature test chamber.

[0014] Further, the lightweight convolutional neural network model in step S3 includes the following layers connected in sequence: an input layer that receives multi-channel time series input features containing a predetermined number of time steps; a dilated convolutional layer configured with a first number of convolutional kernels, a convolutional kernel of a predetermined size, and a predetermined dilation rate; a depthwise separable convolutional layer including a depthwise convolution step and a pointwise convolution step; a pooling layer for reducing the data dimension and extracting significant features; a third convolutional layer for further processing the features; a global average pooling layer for compressing the time dimension features and reducing the number of parameters; a fully connected layer for feature mapping; and an output layer that outputs the predicted real-time temperature drift error.

[0015] Further, the dilated convolutional layer uses 32 convolutional kernels of the first number, the convolutional kernel size is 5 time steps, and the dilation rate is 2.

[0016] Further, in the depthwise separable convolutional layer, the depthwise convolution step uses the same number of convolutional kernels as the number of input channels and a predetermined convolutional kernel size, and the pointwise convolution step uses 1x1 convolutional kernels and a second number of convolutional kernels to change the number of channels.

[0017] Further, the loss function used when training the lightweight convolutional neural network model includes a mean square error term and an L1 regularization term for the model weights.

[0018] An RLG temperature drift compensation system includes: a ring laser gyroscope (RLG) configured with an angular velocity sensor and a temperature sensor for outputting an angular velocity measurement value and an internal temperature measurement value; a processing unit; and a memory connected to the processing unit, wherein instructions are stored in the memory. When the instructions are executed by the processing unit, it receives in real time the angular velocity measurement value and the temperature measurement value output by the RLG; constructs multi-channel time series input features based on the temperature measurement value, and the multi-channel time series input features include at least one channel calculated based on the temperature measurement value; processes the multi-channel time series input features using a lightweight convolutional neural network model pre-trained and stored in the memory, and the model is trained to learn the mapping relationship between the input features and the temperature drift residual, so as to output the predicted real-time temperature drift error; and subtracts the predicted real-time temperature drift error from the real-time angular velocity measurement value to output a compensated angular velocity value.

[0019] Further, the lightweight convolutional neural network model stored in the memory includes a dilated convolutional layer and a depthwise separable convolutional layer.

[0020] Compared with the prior art, the beneficial effects of the present invention are as follows:

[0021] 1. Based on the lightweight convolutional neural network, the present invention can automatically learn the complex non-linear relationship between temperature and sensor error through multi-layer convolution, and has a higher compensation accuracy compared with traditional methods.

[0022] 2. The convolutional kernels of the lightweight convolutional neural network can automatically extract local correlation features from the original data, avoiding the cumbersome and subjective nature of manual feature engineering;

[0023] 3. The lightweight convolutional neural network model designed in the present invention is customized based on the MobileNet-type lightweight convolutional neural network. By adopting the design method of combining the second dilated convolutional layer with the third depthwise separable convolutional layer, in the scenario of ring laser gyroscope temperature drift compensation, it can effectively capture the complex non-linear dependence relationship between temperature changes and gyroscope drift across different time scales, while significantly reducing the number of model parameters and computational complexity.

[0024] 4. Through the model compression technology of the present invention, that is, by setting a global average pooling layer, the convolutional neural network can be deployed to resource-constrained embedded devices to meet the real-time requirements. After testing, after being deployed on the FPGA, the real-time compensation speed of this lightweight model is 1.8 ms / frame, meeting the real-time output requirement of 200 Hz, and there is a significant improvement compared with the computational speed of the traditional BP neural network (>10 ms / frame). BRIEF DESCRIPTION OF THE DRAWINGS

[0025] In order to more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the following will briefly introduce the drawings required for use in the description of the embodiments or the prior art. Obviously, the drawings in the following description are some embodiments of the present invention. For those of ordinary skill in the art, without creative efforts, other drawings can also be obtained based on these drawings.

[0026] Figure 1 is the implementation flowchart of the present invention;

[0027] Figure 2 is the temperature change curve of the high and low temperature test chamber;

[0028] Figure 3 is the comparison diagram of the temperature drift compensation effect. DETAILED DESCRIPTION OF THE EMBODIMENTS

[0029] For the convenience of understanding the present invention, the following will describe the present invention more comprehensively and meticulously in combination with the accompanying drawings of the specification and the preferred embodiments. However, the protection scope of the present invention is not limited to the following specific embodiments.

[0030] Unless otherwise defined, all the technical terms used hereinafter have the same meaning as commonly understood by those skilled in the art. The technical terms used herein are only for the purpose of describing specific embodiments and are not intended to limit the protection scope of the present invention.

[0031] The following will further describe the present invention in detail with reference to the accompanying drawings:

[0032] SeeFigure 1 and Figure 2 , this invention discloses an RLG temperature drift compensation method based on a lightweight convolutional neural network, including:

[0033] S1: Acquisition and preprocessing of training data for the ring laser gyro data model;

[0034] The training data includes the angular velocity measurement values of the RLG under different temperature environments, the temperature measurement values output by the internal temperature sensor of the RLG at the corresponding moments, and the temperature field data constructed from the above data; the specific steps are as follows:

[0035] S1.1 Set the temperature control program of the high and low temperature test chamber to collect the output of the ring laser gyro and the output of the temperature sensor under a complex temperature change environment;

[0036] Open the high and low temperature test chamber and start the temperature change program:

[0037] The entire data acquisition process takes 17.5 hours. Among them, from the 0th to the 2nd hour, the temperature of the high and low temperature test chamber is stable at 20°C; from the 2nd to the 3rd hour, the temperature of the test chamber is reduced to -40°C at a temperature change rate of -1°C / min; from the 3rd to the 5th hour, the temperature of the test chamber is stable at -40°C; from the 5th hour to the 6.67th hour, the temperature of the test chamber is increased to 60°C at a temperature change rate of 1°C / min; from the 6.67th to the 8.67th hour, the temperature of the test chamber is stable at 60°C; from the 8.67th to the 10.34th hour, the temperature of the test chamber is reduced to -40°C at a temperature change rate of -1°C / min; from the 10.34th to the 12.34th hour, the temperature of the test chamber is stable at -40°C; from the 12.34th to the 14th hour, the temperature of the test chamber is increased to 60°C at a temperature change rate of 1°C / min; from the 14th to the 16th hour, the temperature of the test chamber is stable at 60°C; from the 16th to the 16.67th hour, the temperature of the test chamber is reduced to 20°C at a temperature change rate of -1°C / min; from the 16.67th to the 17.5th hour, the temperature of the test chamber is stable at 20°C;

[0038] S1.2 Acquisition of the original output of the ring laser gyro;

[0039] While starting the temperature change program, power on the ring laser gyro and record the real-time output angular velocity measurement value of the gyro and the output value of the temperature sensor , where the subscript k represents the current moment k, and the sampling frequency is 200Hz;

[0040] S1.3 Data preprocessing, using moving average filtering (window = 10) to suppress high-frequency noise;

[0041] S1.4 Construction of multi-channel time series input and labels;

[0042] The multi-channel time series includes 4 channels:

[0043] Channel 1 is the output value of the temperature sensor ;

[0044] Channel 2 is the temperature change rate : ;

[0045] Channel 3 is the square of the output value of the temperature sensor ;

[0046] Channel 4 is the product of the output value of the temperature sensor and the temperature change rate ; ;

[0047] The label is the residual of the RLG measurement value : , where is the mean value of the angular velocity measurement values in the 1st - 2nd hour interval when the RLG stabilizes during the constant temperature stage of the high - low temperature test chamber;

[0048] The training data uses the residual as the labeled data to learn the mapping relationship between the temperature field and the residual for the convolutional neural network model;

[0049] S2: Design a convolutional neural network model;

[0050] S2.1 The network structure of the convolutional neural network model is as follows:

[0051] The network structure of the convolutional neural network model is divided into eight layers, which are used to process the input multi - channel time series in sequence, gradually extract the key features related to the temperature drift error, and finally output the predicted temperature drift compensation amount, specifically as follows:

[0052] Input layer

[0053] The input layer is used to receive the pre - processed multi - channel time series. Each input sample of the input layer is data of 100 consecutive time steps, and each time step includes the features of 4 channels constructed in S2. The input sample

[0054] is as follows:

[0055]

[0056] The training data of 17.5 hours totals 12,600,000 (17.5×24×3600×200 = 12,600,000) time steps and can be separated into 126,000 input samples, .

[0057] The input layer does not perform any calculations and passes 126,000 input samples to the first convolutional layer. The data input shape is: (126,000, 100, 4), representing a three-dimensional tensor.

[0058] The first convolutional layer, which is also the dilated convolutional layer

[0059] The first convolutional layer is used to expand the receptive field and capture long-range dependencies in the time series without significantly increasing the computational burden. The input data of the first convolutional layer is the output data of the input layer.

[0060] The first convolutional layer uses 32 convolutional kernels with a kernel size of 5 time steps and a dilation rate = 2. Dilated convolution expands the receptive field by interval sampling. The formula for the effective kernel size is: effective kernel size = kernel size + (kernel size - 1) × (dilation rate - 1) = 9, that is, each convolutional kernel actually covers 9 time steps. In the mode without padding, the formula for the output number of time steps is: output time steps = input time steps - effective kernel size + 1. Substituting the values, the output time steps can be obtained as 92. By using 32 convolutional kernels, each convolutional kernel generates an independent feature map, automatically expanding from the original 4 temperature field features to 32 new features, which are used to automatically capture and extract different patterns in the temperature data. The first convolutional layer converts the input data of (126,000, 100, 4) into an output of shape (126,000, 92, 32), where 92 is the number of time steps and 32 is the number of convolutional kernels (number of channels).

[0061] Depthwise separable convolutional layer

[0062] The depthwise separable convolutional layer is used to decompose the standard convolution into two steps: depthwise convolution and pointwise convolution, significantly reducing the amount of computation and the number of parameters while maintaining the feature extraction ability. The following are the specific steps:

[0063] Step 1: Depthwise convolution

[0064] The parameter settings are as follows:

[0065] Kernel size: 3 time steps; Number of convolutional kernels: 32; Stride: 1;

[0066] Output calculation: Number of time steps = 92 (input time steps) - 3 (kernel size) + 1 = 90; Number of channels: 32;

[0067] Output shape: (126,000, 90, 32)

[0068] Step 2: Pointwise convolution

[0069] Kernel size: 1 time step, Number of convolutional kernels: 64; Stride: 1;

[0070] Output calculation: Number of time steps = 90, number of channels = 64;

[0071] Output shape: (12600, 90, 64)

[0072] The number of parameters of the depthwise convolution is 32 (number of channels) × 3 (kernel size) = 96 parameters,

[0073] The number of parameters of the pointwise convolution is 32 (input channels) × 64 (output channels) × 1 (kernel size) = 2048 parameters.

[0074] Total number of parameters: 96 (depthwise convolution) + 2048 (pointwise convolution) = 2144 parameters;

[0075] Comparison with standard convolution: If directly using standard convolution (number of filters = 64, kernel size = 3), the number of parameters is: 32 (input channels) × 64 (output channels) × 3 (kernel size) = 6144 parameters, and the number of parameters is reduced by about 65%, significantly improving the computational efficiency.

[0076] Max pooling layer

[0077] The max pooling layer is used to downsample the data of each channel, retaining the maximum value every two time steps to filter out high-frequency noise and highlight the overall trend of temperature changes. The max pooling layer extracts local maxima through a sliding window, compressing the input shape from (126000, 90, 64) to (126000, 45, 64), while reducing the data volume and retaining key features. This operation significantly improves the model's robustness to noise and provides more abstract high-order temporal features for subsequent layers.

[0078] Third convolutional layer

[0079] The third convolutional layer uses 64 kernels, each kernel covering 3 time steps, to perform cross-channel feature combination on the pooled data. The input data shape of the third convolutional layer is (126000, 45, 64), and with 64 kernels and each kernel processing 3 time steps, the time step dimension of the output shape is reduced from 45 to 43, gradually abstracting temporal features. The output data shape is (126000, 43, 64);

[0080] Global average pooling layer

[0081] The global average pooling layer compresses the input shape from (126000, 43, 64) to (126000, 64) through simple mean calculation, completely eliminating parameters while retaining global channel information, significantly reducing the model complexity. This design makes the CNN model more lightweight, adapts to the real-time requirements of embedded devices, and reduces the risk of overfitting at the same time.

[0082] Global connection layer

[0083] The global connection layer maps the 64-dimensional feature vector to 32 dimensions and uses the swish non-linear activation function to enhance the model's fitting ability for complex temperature-error relationships.

[0084] Output layer

[0085] The output layer converts the 32-dimensional feature into a single numerical value, which is the predicted value of the model , with the unit of degrees per hour (° / h).

[0086] S2.2 Design the loss function;

[0087] Corresponding to the input sample The label is , and the loss function is:

[0088]

[0089] denotes and the mean square error between denotes the L1 regularization of

[0090] S3: Model deployment, using a convolutional neural network model for real-time compensation;

[0091] S3.1 Deploy the convolutional neural network model trained in S2 to the FPGA of the RLG;

[0092] In the engineering application of the RLG, the angular velocity measurement value and the corresponding temperature measurement value will be output in real time at time j (any time in the actual application process, represented by j);

[0093] S3.3 Calculate the temperature field data based on , including , and , and the calculation formula can refer to S1.4;

[0094] S3.2 Input the temperature field data into the convolutional neural network model deployed on the FPGA to obtain the temperature drift error predicted by the model at time j;

[0095] S3.3 Perform real-time compensation within the FPGA;

[0096] The real-time compensation formula at the j-th moment is as follows:

[0097] .

[0098] Thus, the angular velocity value after temperature drift compensation at the j-th moment is obtained , compared with the angular velocity measurement value directly output by the RLG , the temperature drift caused by temperature change is compensated, and its measurement accuracy is significantly improved.

[0099] Embodiment:

[0100] To illustrate the technical solution disclosed in the present invention in detail, the following further elaborates with specific embodiments. The specific embodiments described herein are only used to explain the present application and are not used to limit the present application.

[0101] Figure 1 The implementation process of the present invention is divided into the following steps:

[0102] S1: Acquisition and preprocessing of training data for the ring laser gyro data model;

[0103] S2: Input feature design and construction of a multi-channel time series;

[0104] S3: Design of the convolutional neural network model

[0105] S4: Real-time compensation

[0106] S1 to S3 are the model construction part, and S4 is the model deployment part. After obtaining the CNN model in the model construction part, the model is deployed and real-time compensation is performed.

[0107] A test experiment was carried out in this embodiment:

[0108] The hardware configuration is a type 90 ring laser gyro. Based on S1, a set of data was collected and preprocessed;

[0109] The data collected in S1 was constructed into the multi-channel time series described in S2. Further, it was input into the convolutional neural network model designed in S3 to obtain a convolutional neural network model specifically designed for ring laser gyro temperature drift compensation;

[0110] The model obtained in S3 was deployed to the FPGA inside the ring laser gyro;

[0111] The ring laser gyro deployed with the convolutional neural network model was put into the high and low temperature test chamber again for temperature compensation effect testing, Figure 3It is a comparison chart of temperature compensation effects before and after warm compensation. RAW refers to the output pulse frequency data of the ring laser gyroscope in its original state without any temperature compensation treatment. LR refers to the gyroscope output pulse frequency obtained after temperature compensation using a linear regression model. NN refers to the gyroscope output pulse frequency obtained after temperature compensation using a traditional neural network. CNN refers to the gyroscope output pulse frequency obtained after temperature compensation using a lightweight convolutional neural network. The output curve of the lightweight convolutional neural network is smoother, directly proving that the lightweight convolutional neural network model has superior performance in the temperature drift compensation task.

[0112] The compensation accuracy compared with traditional methods is shown in Table 1:

[0113] Table 1 Comparison of temperature drift compensation by different methods Linear regression Traditional neural network The method of the present invention Zero bias instability (° / h) 0.00324 0.00292 0.00182

[0114] As can be seen from Table 1, the zero-bias instability of the method proposed in the present invention is the smallest, indicating that the temperature compensation accuracy of this method is the highest, verifying the effectiveness and superiority of this method, and reflecting the beneficial effects of the present invention.

Claims

1. A RLG temperature drift compensation method based on lightweight convolutional neural network, characterized in that: The following steps are involved: S1 collects the angular velocity measurement values ​​of the ring laser gyro RLG in different temperature environments and the temperature measurement values ​​output by the internal temperature sensor of the RLG at the corresponding time as training data, and pre-processes the training data; S2 constructs a multi-channel time series input feature based on the pre-processed temperature measurement value, and the multi-channel time series input feature includes at least one channel calculated based on the temperature measurement value; based on the angular velocity measurement value and the average of the angular velocity measurement value in the constant temperature stage, the temperature drift residual is calculated as a training label; S3 designs and trains a lightweight convolutional neural network model, and the lightweight convolutional neural network model uses the multi-channel time series input feature as input and the temperature drift residual as output for training, and learns the mapping relationship between the input feature and the temperature drift residual; the second layer of the lightweight convolutional neural network model is a hole convolution layer, the third layer is a depth-separable convolution layer, and the sixth layer is a global average pooling layer; S4 deploys the trained lightweight convolutional neural network model to the internal processing unit of the RLG; when the RLG is running, its angular velocity measurement value and temperature measurement value are obtained in real time; based on the real-time obtained temperature measurement value, the real-time multi-channel time series input feature is constructed; The real-time multi-channel time series input feature is input into the deployed lightweight convolutional neural network model to obtain a predicted real-time temperature drift error; the predicted real-time temperature drift error is subtracted from the real-time angular velocity measurement value to obtain a compensated angular velocity value.

2. The RLG temperature drift compensation method based on a lightweight convolutional neural network according to claim 1, characterized in that: The preprocessing in step S1 is: using a sliding average filter to filter the angular velocity measurement value and the temperature measurement value, and the filter window size is 10 time steps.

3. The RLG temperature drift compensation method based on a lightweight convolutional neural network according to claim 1, characterized in that: The step of constructing the multi-channel time series input feature in step S2 includes: based on the temperature measurement values ​​at time k and k-1 , , construct the following four channels: Channel 1 is the temperature sensor output value ; Channel 2 is the temperature change rate : ; Channel 3 is the square of the temperature sensor output value ; Channel 4 is the temperature sensor output value Rate of change with temperature The product of .

4. The RLG temperature drift compensation method based on a lightweight convolutional neural network according to claim 1, characterized in that: The training label in step S2 is the residual of the RLG measurement value : ,in is the angular velocity measurement value at the current moment k; It is the average of the angular velocity measurements of RLG in the constant temperature stage of the high and low temperature test chamber.

5. The RLG temperature drift compensation method based on lightweight convolutional neural network according to claim 1, characterized in that: The lightweight convolutional neural network model in step S3 includes the following layers connected in sequence: an input layer, receiving the multi-channel time series input features including a predetermined number of time steps; a dilated convolution layer, configured with a first number of convolution kernels, a convolution kernel of a predetermined size, and a predetermined dilation rate; a depthwise separable convolution layer, including a depthwise convolution step and a pointwise convolution step; a pooling layer, for reducing data dimensions and extracting significant features; and a third convolution layer, for further processing features; The global average pooling layer is used to compress the time dimension features and reduce the number of parameters; the fully connected layer is used for feature mapping; and the output layer outputs the predicted real-time temperature drift error.

6. The RLG temperature drift compensation method based on lightweight convolutional neural network according to claim 5, characterized in that: The atrous convolution layer uses a first number of 32 convolution kernels, a convolution kernel size of 5 time steps, and a atrous rate of 2.

7. A RLG temperature drift compensation method based on a lightweight convolutional neural network according to claim 5 or 6, characterized in that: The depthwise separable convolution layer, whose depthwise convolution step uses the same number of convolution kernels as the number of input channels and a predetermined convolution kernel size, and whose pointwise convolution step uses a 1x1 convolution kernel and a second number of convolution kernels to change the number of channels.

8. The RLG temperature drift compensation method based on lightweight convolutional neural network according to claim 1, characterized in that: The loss function used when training the lightweight convolutional neural network model includes a mean square error term and an L1 regularization term for the model weights.

9. A RLG temperature drift compensation system, characterized in that: include: A ring laser gyro RLG is configured with an angular velocity sensor and a temperature sensor for outputting an angular velocity measurement value and an internal temperature measurement value; Processing unit; A memory connected to the processing unit, wherein the memory stores instructions; wherein when the instructions are executed by the processing unit, the system implements the method according to any one of claims 1 to 8, specifically for: receiving the angular velocity measurement value and the temperature measurement value output by the RLG in real time; constructing a multi-channel time series input feature based on the temperature measurement value, wherein the multi-channel time series input feature includes at least one channel calculated based on the temperature measurement value; processing the multi-channel time series input feature using a lightweight convolutional neural network model pre-trained and stored in the memory, wherein the model training is used to learn the mapping relationship between the input feature and the temperature drift residual, thereby outputting a predicted real-time temperature drift error; and subtracting the predicted real-time temperature drift error from the real-time angular velocity measurement value to output a compensated angular velocity value.

10. The RLG temperature drift compensation system according to claim 9, characterized in that: The lightweight convolutional neural network model stored in the memory includes a hole convolution layer and a depth-separable convolution layer.

Images