An optimization method for improving the quick start performance of a quartz accelerometer

Abstract

The application relates to an optimization method for improving the quick starting performance of a quartz accelerometer, which comprises the following steps: 1, connecting the accelerometer table head and the servo circuit through a connecting line; 2, arranging multiple temperature sensors inside and outside the table body for temperature gradient field monitoring; 3, establishing a temperature-error mapping model by adopting an LSTM algorithm; 4, training and verifying the LSTM model; and 5, compensating the starting data of the accelerometer based on the mapping output of the verified LSTM model. The application optimizes the structure of the accelerometer, realizes the separation of the table body and the servo circuit without affecting the output precision of the accelerometer, avoids the influence of the heat dissipation of the servo circuit on the temperature field of the table body, improves the effective temperature compensation of the accelerometer through the arrangement of the multiple-position temperature sensors and the application of the LSTM compensation algorithm, and the 5min starting performance precision of the accelerometer is improved by 20% through testing.

Description

Technical Field

[0001] This invention belongs to the field of inertial navigation technology, and in particular relates to a method for improving the rapid startup performance of accelerometers through structural optimization and intelligent temperature compensation algorithms. Background Technology

[0002] Quartz accelerometers (hereinafter referred to as accelerometers), as a high-precision inertial sensor, are widely used in aerospace, weapon guidance, geological exploration, and other fields. In recent years, with the increasing demand for rapid guidance, the requirements for rapid start-up of accelerometers have also been continuously increasing. Currently, the main factors affecting start-up performance are as follows: First, the heat generated by the integrated servo circuit and the heat generated by the torque coil energized under high g conditions lead to changes in the internal temperature field of the accelerometer body. Second, due to the influence of the heat conduction rate, the external body and the internal core of the accelerometer cannot reach thermal equilibrium quickly during the rapid start-up time, resulting in a temperature gradient. This temperature gradient changes nonlinearly with changes in the external temperature environment and the internal thermal field generated by the internal torque coil. Existing temperature compensation methods mostly use a single temperature sensor on the outside of the accelerometer body, which cannot accurately capture the dynamic temperature change characteristics inside the accelerometer body in a short period of time. Therefore, it is impossible to construct an effective model of accelerometer output and temperature under dynamic conditions for accurate temperature compensation.

[0003] Existing compensation methods for accelerometer startup performance include polynomial fitting compensation, and compensation algorithms based on CNN and CMAC neural networks. Polynomial fitting can only describe static or low-order nonlinear relationships and cannot construct temperature compensation models under multiple temperature sensors. CNNs, due to their limited receptive field (only time-seconds), have difficulty modeling long-term temperature variation dependencies. CMAC is suitable for static mappings, such as temperature compensation under constant temperature, but cannot adapt to dynamic temperature compensation. None of the above compensation methods can meet the data compensation requirements under dynamic temperature variation environments.

[0004] LSTM (Long Short-Term Memory) is a special type of recurrent neural network model that is particularly good at processing sequential data. Through gating mechanisms, it can meet the requirements of dynamic temporality and multi-source fusion, and can realize temperature compensation in dynamic temperature environments with multiple temperature sensors.

[0005] Based on the above research, a systematic optimization scheme from structure to algorithm needs to be designed to improve the startup performance accuracy of the accelerometer. Summary of the Invention

[0006] This invention addresses the shortcomings of existing technologies by proposing an optimized method to improve the rapid start-up performance of quartz accelerometers.

[0007] An optimization method for improving the fast start-up performance of a quartz accelerometer, characterized by comprising the following steps: Step 1: Disconnect the accelerometer head from the servo circuit using a connecting cable; Step 2: Arrange multiple temperature sensors inside and outside the body of the instrument for monitoring the temperature gradient field; Step 3: Establish a temperature-error mapping model using the LSTM algorithm; Step 4: Train and validate the LSTM model; Step 5: Based on the mapping output of the validated LSTM model, compensate for the accelerometer's startup data.

[0008] Furthermore, in step 1, the connecting cable should be an FPC cable or a coaxial shielded cable, with a length not exceeding 50mm and a signal shielding rate ≥60dB; the connecting cable should be fixed to the insulating frame and routed in a serpentine manner.

[0009] Furthermore, in step 2, a first thermistor and a second thermistor are installed on the outer side of the upper excitation ring and the outer side of the lower excitation ring of the accelerometer head, respectively, to monitor the temperature on the outside of the accelerometer body; a third thermistor 8 is installed on the inner wall of the upper excitation ring to monitor the temperature inside the accelerometer body. The three thermistors are attached to their respective positions with thermally conductive silicone, and the temperature measurement accuracy of the thermistors is within ±0.1℃.

[0010] Furthermore, step 3 includes: Step 3.1: Acquire the raw output a_raw of the accelerometer and the data from the three temperature sensors (T1, T2, T3), ensuring that the timestamps are aligned and the time error is less than 1ms; Step 3.2: Preprocess the raw accelerometer output a_raw and the data from the three temperature sensors (T1, T2, T3) through three steps: normalization, sliding window, and outlier filtering. Normalization maps the temperature data to the [1,1] interval using the Min-Max normalization method. The sliding window divides the acquired data into continuous segments with a length of 30 seconds and a step size of 1 second. Outlier filtering uses the 3σ criterion, discarding data points exceeding 3σ as jump points. Step 3.3: Design the LSTM model. The LSTM model adopts a two-layer stacked structure of 64 units + 32 units, and the output layer is a fully connected layer (Dense). The first LSTM layer has 64 units and is used to extract the temporal features of short-term temperature fluctuations. The return_sequences=true is set in the first LSTM layer to output the hidden state of each time step, retaining complete temporal information for use by the next layer. The second LSTM layer has 32 units and return_sequences=false is set in the second LSTM layer to output only the hidden state of the last time step, capturing global temporal dependencies. The fully connected layer maps the 32-dimensional features output by the LSTM to a one-dimensional accelerometer temperature compensation value ΔE. The activation function of the fully connected layer is linear.

[0011] Furthermore, in step 3.3, an attention weighting mechanism is added to the LSTM model design, and the inputs from multiple temperature sensors are dynamically weighted using the Softmax function.

[0012] Furthermore, step 4, the model training process includes data partitioning, selection of loss function and optimizer, training strategy and tuning, and finally, validation of the training effect through validation set calculation. Specifically, the model training data is divided into a 70% training set, a 15% validation set, and a 15% test set. The training model uses the MSE loss function and the Adam optimizer, with an initial learning rate of 0.001. The model training strategy employs learning rate decay and early stopping. The learning rate decay is set to 0.5 times the original value every 20 epochs, and the early stopping patience value is set to 10. Model tuning is achieved by adjusting the number of LSTM units. Validation of the training effect includes validation set RMSE calculation, comparison of loss curves between the training and validation sets, and test set analysis. The formula for calculating the validation set RMSE is:

[0013] in: N: The number of samples in the validation set; The model's predicted value for the i-th sample; : The true value of the i-th sample.

[0014] If the validation data does not meet the standards, return to step S3 to readjust the model parameters, including the number of LSTM layers and units, learning rate, and dropout rate. Note that only 1 to 2 parameters should be adjusted each time. If the data meets the standards, then deploy and apply the model.

[0015] Furthermore, in step 5, the LSTM output is mapped to a temperature compensation amount Δa, and this compensation amount is superimposed onto the original accelerometer output. Then: a_compensated=a_raw+Δa Where a_compensated is the compensated accelerometer value.

[0016] The above-mentioned objective of this invention is achieved through the following technical solution: The advantages and positive effects of this invention are as follows: 1. This invention optimizes the structure of the accelerometer, achieving separation of the meter body and servo circuit without affecting the output accuracy of the accelerometer. This avoids the impact of the servo circuit's heat dissipation on the meter body's temperature field. By arranging multiple temperature sensors and applying the LSTM compensation algorithm, the effective temperature compensation of the accelerometer is improved. According to the test, the accelerometer's 5-minute startup performance accuracy is improved by 20%.

[0017] 2. This invention captures short-term fluctuations and long-term dependencies in temperature-acceleration using a dual-layer LSTM. Simultaneously, to enhance dynamic adaptability, an attention mechanism is added, automatically assigning weights to multiple temperature sensors to improve the accuracy of spatial temperature distribution modeling. Attached Figure Description

[0018] Figure 1 This is a schematic diagram of the split structure of the quartz accelerometer of the present invention; Figure 2 This is a schematic diagram showing the distribution of the thermistor inside and outside the instrument body in this invention; Figure 3 This is a flowchart of the LSTM compensation algorithm of the present invention; In the diagram: 1. Pad; 2. Accelerometer head; 3. Transition plate; 4. Connecting line; 5. Servo circuit; 6. Lower excitation loop; 7. Upper excitation loop; 8. Third thermistor; 9. First thermistor; 10. Second thermistor. Detailed Implementation

[0019] The structure of the present invention will be further described below with reference to the accompanying drawings and embodiments. It should be noted that these embodiments are descriptive and not limiting.

[0020] For an optimization method to improve the fast start-up performance of quartz accelerometers, please refer to [link / reference]. Figures 1-3 Its inventive point is that it includes the following steps: Step 1: Disconnect the accelerometer head 2 from the servo circuit 5. (See attached image) Figure 1This diagram shows the structure of the accelerometer head 2 and the servo circuit 5, which are connected separately. In the diagram, the accelerometer head 2 is connected to the meter head via pad 1 on the transition plate 3. The two ends of the connecting line 4 are soldered to the transition plate 3 and the servo circuit 5, respectively. The connecting line 4 uses an FPC cable or a coaxial shielded cable to achieve the separate connection between the accelerometer head 2 and the servo circuit 5. The use of an FPC cable or a coaxial shielded cable reduces the impact of stray capacitance caused by changes in lead position on the accelerometer. The length of the connecting line 4 cannot exceed 50mm to reduce signal screening. In practical applications, since the connecting line 4 should be fixed to a metal frame, a serpentine routing method is used to account for the effects of thermal expansion and contraction.

[0021] Step 2: Install a first thermistor 9 and a second thermistor 10 on the outer sides of the upper excitation ring 7 and lower excitation ring 6 of the accelerometer head, respectively, to monitor the temperature on the outer side of the accelerometer body; install a third thermistor 8 on the inner wall of the upper excitation ring to monitor the temperature inside the accelerometer body. The three thermistors are attached to their respective positions using thermally conductive silicone. The temperature measurement accuracy of the thermistors is within ±0.1℃. Thus, the internal and external temperature gradient fields of the accelerometer body are monitored through three temperature points. (Appendix) Figure 2 The diagram shows the distribution of thermistors inside and outside the accelerometer body.

[0022] Step 3: Establish a temperature-error mapping model using the LSTM algorithm to achieve effective and accurate temperature compensation for the accelerometer output, including the following steps: Step 3.1: Data Acquisition: Acquire the raw output a_raw from the accelerometer and data from the three temperature sensors (T1, T2, T3), ensuring timestamp alignment with a time error of less than 1ms. In this embodiment, a parallel acquisition module is used for data acquisition. To ensure the accuracy of subsequent model training and real-time compensation, the data acquisition environment should cover actual usage conditions. Therefore, in this embodiment, the data acquisition method involves placing the accelerometer in a temperature chamber and cycling it across the entire temperature range at different temperature change rates. The temperature range is -55℃ to 85℃, the temperature change rate range is 0.5℃ / min to 5℃ / min, and the gradient of the temperature change rate is 0.5℃ / min. The continuous output values ​​of the accelerometer at different temperature change rates and the corresponding temperature data from the three internal and external temperature sensors are acquired in parallel.

[0023] Step 3.2: Data Preprocessing: The raw accelerometer output a_raw and the data from the three temperature sensors (T1, T2, T3) are preprocessed through three steps: normalization, sliding window, and outlier filtering. Specifically: Data preprocessing involves "cleaning" and "organizing" the collected data, removing invalid and bad data, and is a crucial step in data analysis or machine learning. This embodiment performs data preprocessing in three steps: normalization, sliding window, and outlier filtering. Normalization maps temperature data to the [1,1] interval using the Min-Max normalization method. The sliding window segments the collected data into continuous fragments to help the LSTM model capture the dynamic patterns of temperature changes. In this embodiment, the sliding window length is 30 seconds, with a step size of 1 second. Outlier filtering uses statistical rules to remove unqualified data, preventing erroneous data from misleading model training. In this embodiment, the data filtering uses the 3σ criterion, discarding data points exceeding 3σ as jump points.

[0024] Step 3.3: LSTM Model Design. The LSTM model adopts a two-layer stacked structure of 64 units + 32 units, with a fully connected (Dense) output layer. The advantage of this stacked structure is that it allows for deeper learning of complex nonlinear relationships. The first LSTM layer has 64 units, used to extract temporal features of short-term temperature fluctuations. `return_sequences=true` is set in the first LSTM layer to output the hidden state at each time step, preserving complete temporal information for use by the next layer. The second LSTM layer has 32 units, designed to reduce overfitting risk through feature compression. `return_sequences=false` is set in the second LSTM layer to output only the hidden state of the last time step, capturing global temporal dependencies. The fully connected layer maps the 32-dimensional features output by the LSTM to a one-dimensional accelerometer temperature compensation value ΔE. The activation function of the fully connected layer is linear. An attention weighting mechanism can be added to the model design, dynamically weighting the inputs from multiple temperature sensors using a Softmax function, allowing the model to dynamically adjust the importance of each temperature sensor.

[0025] Step 4: LSTM Model Training and Validation. Model training includes data partitioning, selection of loss function and optimizer, training strategy and tuning, and finally, validation of the training effect through validation set calculation. Specifically: The model training data was divided into a 70% training set, a 15% validation set, and a 15% test set. The training model used the MSE loss function and the Adam optimizer, with an initial learning rate of 0.001. The training strategy employed learning rate decay and early stopping. The learning rate decay was set to decrease to 0.5 times the original value every 20 epochs, and the early stopping patience value was set to 10, meaning training would terminate if there was no improvement after 10 consecutive epochs. Model tuning was achieved by adjusting the number of LSTM units. The training effect was verified through validation set RMSE calculation, comparison of loss curves between the training and validation sets, and test set analysis. The validation set RMSE calculation formula is:

[0026] in: N: The number of samples in the validation set; The model's predicted value for the i-th sample; : The true value of the i-th sample.

[0027] If the validation data does not meet the standards, return to step S3 to readjust the model parameters, including the number of LSTM layers and units, learning rate, and dropout rate. Note that only 1 to 2 parameters should be adjusted each time. If the data meets the standards, then deploy and apply the model.

[0028] Step 5: Implement compensation. Based on the validated LSTM model, compensate the accelerometer's startup data. Specifically: Map the LSTM output to a temperature compensation amount Δa, and then superimpose the compensation amount onto the original accelerometer output. a_compensated=a_raw+Δa Where a_compensated is the compensated accelerometer value.

[0029] Comparing the 5-minute startup performance data of the three accelerometers before and after structural modification and data compensation, the test environment temperature was 40℃±2℃ and the acceleration input value was 1g, as shown in Table 1. The startup performance was improved by 20% after implementation.

[0030]

[0031] Although embodiments and drawings of the present invention have been disclosed for illustrative purposes, those skilled in the art will understand that various substitutions, variations and modifications are possible without departing from the spirit and scope of the present invention and the appended claims. Therefore, the scope of the present invention is not limited to the contents disclosed in the embodiments and drawings.

Claims

1. An optimized method for improving the rapid start-up performance of a quartz accelerometer, characterized in that, Includes the following steps: Step 1: Disconnect the accelerometer head from the servo circuit using a connecting cable; Step 2: Arrange multiple temperature sensors inside and outside the body of the instrument for monitoring the temperature gradient field; Step 3: Establish a temperature-error mapping model using the LSTM algorithm; Step 4: Train and validate the LSTM model; Step 5: Based on the mapping output of the validated LSTM model, compensate for the accelerometer's startup data; Step 3 includes: Step 3.1: Collect the raw output a_raw of the accelerometer and the data from the three temperature sensors, namely T1, T2, and T3. Ensure that the timestamps are aligned and that the time error is less than 1ms. Step 3.2: The raw output a_raw of the accelerometer and the data from the three temperature sensors are preprocessed through three steps: normalization, sliding window, and outlier filtering. The sliding window is constructed by cutting the acquired data into continuous segments with a length of 30 seconds and a step size of 1 second. The outlier filtering adopts the 3σ criterion, and data points exceeding 3σ are discarded as jump points. Step 3.3: Design the LSTM model. The LSTM model adopts a two-layer stacked structure of 64 units + 32 units, and the output layer is a fully connected layer (Dense). The first LSTM layer has 64 units and is used to extract the temporal features of short-term temperature fluctuations. The return_sequences=true is set in the first LSTM layer to output the hidden state of each time step, retaining complete temporal information for use by the next layer. The second LSTM layer has 32 units and return_sequences=false is set in the second LSTM layer to output only the hidden state of the last time step, capturing global temporal dependencies. The fully connected layer maps the 32-dimensional features output by the LSTM to a one-dimensional accelerometer temperature compensation value ΔE. The activation function of the fully connected layer is linear.

2. The optimization method for improving the rapid start-up performance of a quartz accelerometer according to claim 1, characterized in that: In step 1, the connecting cable is an FPC cable or a coaxial shielded cable, the length of the connecting cable does not exceed 50mm, and the signal shielding rate is ≥60dB; The connecting wires are fixed to the insulating frame and are routed in a serpentine pattern.

3. The optimized method for improving the rapid start-up performance of a quartz accelerometer according to claim 1, characterized in that: In step 2, a first thermistor and a second thermistor are installed on the outer sides of the upper and lower excitation rings of the accelerometer head, respectively, to monitor the temperature on the outside of the accelerometer body; a third thermistor is installed on the inner wall of the upper excitation ring to monitor the temperature inside the accelerometer body. The three thermistors are attached to their respective positions with thermally conductive silicone, and the temperature measurement accuracy of the thermistors is within ±0.1℃.

4. The optimization method for improving the rapid start-up performance of a quartz accelerometer according to claim 1, characterized in that: In step 3.3, an attention weighting module is added to the LSTM model design to dynamically weight the inputs from multiple temperature sensors using the Softmax function.

5. The optimization method for improving the rapid start-up performance of a quartz accelerometer according to claim 1, characterized in that: Step 4 involves model training, including data partitioning, selection of loss function and optimizer, training strategy and tuning, and finally, validation of the training effect through validation set calculation. The model training data is divided into a 70% training set, a 15% validation set, and a 15% test set. The training model uses the MSE loss function and the Adam optimizer, with an initial learning rate of 0.

001. The model training strategy employs learning rate decay and early stopping. The learning rate decay is set to 0.5 times its original value every 20 epochs, and the early stopping patience value is set to 10. Model tuning is achieved by adjusting the number of LSTM units. Validation of the training effect includes validation set RMSE calculation, comparison of loss curves between the training and validation sets, and test set analysis. The formula for calculating the validation set RMSE is: in: N: The number of samples in the validation set; The model's predicted value for the i-th sample; : The true value of the i-th sample; If the validation data does not meet the standards, return to step S3 to readjust the model parameters, including the number of LSTM layers and units, learning rate, and dropout rate. Adjust 1 to 2 parameters each time. If the data meets the standards, deploy and apply the system.

6. The optimized method for improving the rapid start-up performance of a quartz accelerometer according to claim 1, characterized in that: In step 5, the LSTM output is mapped to a temperature compensation amount Δa, and this compensation amount is superimposed onto the original accelerometer output. Then: a_compensated=a_raw+Δa Where a_compensated is the compensated accelerometer value.

Images