An active deployment type end fitting device optimization control method based on a deep learning model

Abstract

The application discloses an active deployment type end device optimization control method based on a deep learning model. First, high-precision sensors are installed on each slider of the active deployment type end device, and key data in a test process, including slider displacement and speed, shearing force, friction force and the like, are collected in real time through a high-speed data acquisition card; subsequently, the collected data are preprocessed; a CNN-LSTM hybrid model is constructed, and the processed data are used for training; after the training is completed, the model is further verified and optimized; finally, the optimized CNN-LSTM model is applied to the slider active control system, the displacement and speed of the slider are predicted and adjusted in real time, the high matching of the slider and the sample in a large deformation process is ensured, and thus the friction effect is effectively eliminated. Through the deep learning model, high-precision control of the slider is realized, the accuracy of test data and the adaptability of the system are greatly improved, and excellent control effect is shown under complex experimental conditions.

Description

Technical Field

[0001] This invention relates to the field of deep learning technology. Background Technology

[0002] Triaxial testing in geotechnical engineering is one of the core methods for studying the mechanical properties of soil and rock, and its test results directly affect the reliability and safety of geotechnical engineering designs. In triaxial testing, the end friction effect of the specimen is one of the key issues affecting the accuracy of the test results. Traditional loading methods often generate significant sliding friction at the specimen ends, leading to uneven stress and strain distribution and affecting the reliability of the test results.

[0003] To address this issue, CN118883231A proposes an active deployment control end device for triaxial compression tests in geotechnical engineering, such as... Figure 1 , Figure 2 , Figure 3 , Figure 4 As shown, the device structure includes: a circular support base for supporting other components and equipped with corresponding sliding parts; a guide rail system comprising multiple sets of guide rail blocks and several precision balls, the guide rail blocks being able to move radially on the circular support base; a slider system comprising multiple rings of sliders arranged radially around the center of the base, the sliders cooperating with the guide rail blocks, and a pressure-resistant arc-shaped rubber diaphragm between the sliders; and a slider active control system including a micro servo motor, carbon fiber force transmission rod, sensors, and a control unit. This device converts the sliding friction between the sample and the loading head into rolling friction between the slider and the balls, and combines this with servo control to achieve active control of the slider, thereby better conforming to the motion law of the sample during large deformation. However, the servo control system has certain limitations in practical applications, especially when facing complex and variable loading conditions, making it difficult to respond to the dynamic changes of the sample in real time and accurately, which may lead to deviations in the stress-strain state of the sample.

[0004] With the rapid development of deep learning technology, it has demonstrated outstanding performance in fields such as pattern recognition and predictive control, and has achieved remarkable results in numerous engineering practices. Deep learning models can capture the potential patterns in complex systems by learning from massive amounts of historical data, thereby making more accurate predictions and responses in real-time control processes.

[0005] Therefore, exploring an optimized control method that applies a deep learning model to an active deployment control end device to achieve full-process loading of the unit mechanical test under the conditions of maintaining the unit stress state and uniform deformation is of great significance for studying the mechanical characteristics of geotechnical materials throughout the entire process. Summary of the Invention

[0006] To address the problem of inaccurate test results caused by end-friction effects in triaxial compression tests of geotechnical engineering, this invention provides an active unfolding end-cap device optimization control method based on a deep learning model. Specifically, it is an active unfolding end-cap device optimization control method based on a hybrid model of convolutional neural network (CNN) and long short-term memory network (LSTM). By introducing deep learning technology, especially combining the advantages of CNN and LSTM, this invention overcomes the limitation of traditional servo control in slow response under complex, nonlinear, and variable loading conditions. It achieves real-time high-precision control of the slider during the test, and more comprehensively eliminates end-friction effects, thereby significantly improving the accuracy and reliability of the test results.

[0007] To achieve the above-mentioned technical objectives, the present invention is implemented through the following technical solution:

[0008] An optimized control method for an actively deployable end-cap device based on a deep learning model, the actively deployable end-cap device comprising: a circular support base for supporting other components and equipped with corresponding sliding parts; a guide rail system comprising multiple sets of guide rail blocks and several precision ball bearings, the guide rail blocks being movable radially on the circular support base; a slider system comprising multiple rings of sliders arranged radially around the center of the base, the sliders cooperating with the guide rail blocks, and a pressure-resistant arc-shaped rubber diaphragm between the sliders; and an active slider control system comprising a micro servo motor, a force transmission rod, a sensor, and a control unit; characterized in that the implementation steps of the control method include:

[0009] Step 1: Sensor deployment and data acquisition;

[0010] Step 2: Data preprocessing and dataset creation;

[0011] Step 3: Construction of the CNN-LSTM hybrid model;

[0012] Step 4: Training the CNN-LSTM hybrid model;

[0013] Step 5: Validation and optimization of the CNN-LSTM hybrid model;

[0014] Step 6: Model Deployment and Real-time Control: The validated and optimized CNN-LSTM hybrid model is deployed in the active control system of the slider of the active unfolding end device to achieve real-time high-precision control of the slider movement.

[0015] Advantages and beneficial effects of the present invention:

[0016] This invention introduces a CNN-LSTM hybrid model to achieve precise real-time control of the slider motion of an actively deployable end-cap device, effectively solving the problem of inaccurate test results caused by end-cap friction effects in triaxial compression tests of geotechnical engineering. Specific effects are as follows:

[0017] (1) The control precision of the slider is significantly improved.

[0018] This invention utilizes a deep learning model to achieve real-time, high-precision control of the slider's motion, significantly reducing the impact of friction on experimental data. Using this method, the standard deviation of slider displacement control decreased from ±5 μm to ±2 μm, and the speed control error decreased from ±0.2 mm / s to ±0.05 mm / s. This high-precision control ensures the synchronization of the slider and the specimen throughout the large deformation process, significantly reducing the influence of end friction on the experimental results. The uniformity of stress-strain distribution in the specimen during the large deformation process is improved by approximately 15% to 20%.

[0019] (2) Enhanced system adaptability and robustness

[0020] The CNN-LSTM hybrid model possesses powerful learning capabilities, enabling it to adapt to different types of soil and rock materials and varying loading conditions. Under different experimental conditions, the CNN-LSTM hybrid model continuously learns and updates, ensuring the excellent control performance of the slider active control system under various loading conditions. Practical applications show that the system response time is reduced by approximately 30%, and the control accuracy is improved by approximately 25%.

[0021] (3) Improved experimental efficiency

[0022] Because the active slider control system can precisely control the slider in real time, it reduces experimental errors caused by friction, thereby reducing the need for repeated experiments and improving experimental efficiency. Specifically, the time for a single experiment is shortened by about 20%, the experimental cycle is greatly reduced, and researchers can obtain high-precision experimental data more quickly, optimizing the entire experimental process.

[0023] (4) Expansion of the scope of application

[0024] The method of this invention is not only applicable to traditional geotechnical material testing, but also adaptable to different types and shapes of specimens, as well as different loading conditions. This makes the method widely applicable in various complex engineering tests, expanding the scope of geotechnical engineering testing.

[0025] (5) Safety and stability are guaranteed

[0026] The slider active control system incorporates anomaly detection and fault protection mechanisms to ensure the safety of the experimental process and the stability of system operation. When the CNN-LSTM mixture model prediction value exceeds the safe range or an anomaly occurs in the system, the system can promptly issue an alarm and stop the slider movement to prevent sample damage or equipment failure. Long-term operation tests show that the system operates without failure for 48 hours continuously, meeting the needs of practical engineering applications.

[0027] (6) Promote the development of geotechnical engineering testing technology

[0028] This invention combines deep learning technology with geotechnical engineering testing, pioneering a testing method based on intelligent control. Through precise slider motion control, it enables a more in-depth study of the mechanical behavior of geotechnical materials under complex stress states, providing more reliable experimental data and scientific basis for the design, analysis, and construction of geotechnical engineering projects, and promoting the development of geotechnical engineering testing technology. Attached Figure Description

[0029] Figure 1 This is one of the structural schematic diagrams of an actively deployable end-cap device (structural exploded view).

[0030] Figure 2 This is the second structural schematic diagram of the active unfolding end device (when the end slider is not unfolded).

[0031] Figure 3 This is the third structural schematic diagram of the active unfolding end device (when the end slider is extended).

[0032] Figure 4 The fourth schematic diagram of the active deployment end device (schematic diagram of the active control system for the slider).

[0033] Figure 5 This is a schematic diagram of the method flow of the present invention.

[0034] Figure 6 This is a schematic diagram of the sensor arrangement in the method of the present invention.

[0035] In the picture:

[0036] 1—Circular support base plate; 2—Permeable stone; 3—Sliding block stop bar;

[0037] Rectangular grooves: 4—long rectangular groove, 5—medium rectangular groove, 6—short rectangular groove;

[0038] Guide rails: 7—long guide rail, 8—medium guide rail, 9—short guide rail;

[0039] Guide rail blocks: 10—first guide rail block, 11—second guide rail block, 12—third guide rail block, 13—fourth guide rail block, 14—fifth guide rail block, 15—sixth guide rail block, 16—seventh guide rail block;

[0040] 17—Arch-shaped groove, 18—Rectangular hole, 19—Tongue and tenon;

[0041] Circle sliders: 21—first slider, 22—second slider, 23—third slider, 24—fourth slider, 25—fifth slider, 26—sixth slider, 27—seventh slider;

[0042] 28—Tenon, 29—Fan-shaped body, 30—Column;

[0043] Miniature servo motors: 31—First miniature servo motor, 32—Second miniature servo motor, 33—Third miniature servo motor, 34—Fourth miniature servo motor, 35—Fifth miniature servo motor, 36—Sixth miniature servo motor, 37—Seventh miniature servo motor;

[0044] 38—Carbon fiber force transmission rod, 39—Hinged support, 40—Limiting ring, 41—Pressure-resistant arc-shaped rubber diaphragm, 42—Drainage hole, 43—High-precision miniature sensor;

[0045] 431—Displacement sensor, 432—Velocity sensor, 433—Pressure sensor, 434—Shear force sensor, 435—Friction force sensor. Detailed Implementation

[0046] CN118883231A proposes an active deployment control end device for triaxial compression tests in geotechnical engineering. The device includes a circular bearing base, a guide rail system, a slider system, and an active slider control system. The active slider control system includes a micro servo motor, a carbon fiber force transmission rod, sensors, and a control unit. This device converts the sliding friction between the sample and the loading head into rolling friction between the slider and the ball bearings, and combines this with servo control to achieve active control of the slider.

[0047] This invention introduces a deep learning model into the active control system of the slider in the aforementioned active deployment end device. Based on servo control, the deep learning model enables high-precision real-time control of the slider's motion, better addressing the complex changes in the specimen throughout the large deformation process, further optimizing the specimen's stress-strain state, and ensuring the scientific validity and reliability of triaxial test results in geotechnical engineering. This invention will provide a more reliable scientific basis for basic research and engineering design in the field of geotechnical engineering.

[0048] The technical solutions provided in this application will be further described below with reference to specific embodiments and accompanying drawings. The advantages and features of this application will become clearer from the following description.

[0049] This invention proposes an optimized control method for an active unfolding end device based on a deep learning model. By combining a convolutional neural network (CNN) and a long short-term memory network (LSTM), it achieves real-time high-precision control of the slider to ensure the uniformity of stress and strain of the specimen throughout the large deformation process.

[0050] like Figure 5 As shown, the specific implementation steps of the method of the present invention include:

[0051] Step 1: Sensor Deployment and Data Acquisition

[0052] On each slider of the active deployment end device (see CN118883231A for the specific structure of the device), a variety of high-precision sensors are arranged, including displacement sensors, velocity sensors, pressure sensors, shear force sensors and friction force sensors, to collect key multidimensional data in real time during the test.

[0053] Specifically, such as Figure 6 As shown, the sensors are arranged as follows:

[0054] The displacement sensor 431 is installed on the curved side of the slider, facing the direction of the slider's movement, to measure the displacement change of the slider in real time with an accuracy of ±0.1 μm.

[0055] The speed sensor 432 (photoelectric encoder) is installed on the corresponding guide block of the slider to obtain the real-time speed information of the slider. The encoder resolution is 1000 P / R. Combined with the counter, high-precision speed measurement can be achieved.

[0056] The pressure sensor 433 is embedded in the end face of the slider that contacts the sample, and measures the normal pressure applied by the slider to the sample. The range is 0-10 MPa and the accuracy is ±0.5% FS.

[0057] Shear force sensor 434 and friction force sensor 435 are embedded in the end face of the slider in contact with the sample, and are used to measure the shear force and friction force between the slider and the sample, respectively. The sensor range is 0-5 kN and the accuracy is ±0.2% FS.

[0058] Each slider is equipped with a displacement sensor, a velocity sensor, a pressure sensor, a shear force sensor, and a friction force sensor. As an example, there are 126 sliders in total, and correspondingly, 126 displacement sensors, 126 velocity sensors, 126 pressure sensors, 126 shear force sensors, and 126 friction force sensors are installed.

[0059] During the data acquisition process, a high-speed data acquisition card was used to synchronously acquire data from the aforementioned sensors at a sampling rate of 2000 Hz. This data included multi-dimensional information during the slider's movement, such as displacement, velocity, pressure, shear force, and friction, ensuring comprehensive and accurate information about the experimental process to improve the model's generalization ability.

[0060] Furthermore, the data acquisition card is equipped with an anti-interference design to ensure the stability and consistency of data in complex experimental environments.

[0061] Step 2: Data preprocessing and dataset creation

[0062] Step 2.1 Preprocess the collected raw sensor data to improve data quality and the effectiveness of subsequent model training. Data preprocessing includes the following steps:

[0063] First, data cleaning: To address potential noise and outliers in sensor data, signal processing methods such as median filtering and Kalman filtering are employed for data cleaning. Median filtering effectively removes spike noise using a sliding window technique, maintaining data smoothness, while Kalman filtering is used to smooth time-series data, reducing the impact of random noise and ensuring data continuity and accuracy.

[0064] Next, normalization is performed: the Min-Max normalization method is used to scale feature data of different dimensions and magnitudes to the range [0,1], eliminating scale differences between data and preventing a single feature from having an excessive impact on the loss function during model training. Normalization not only speeds up the convergence of the model but also improves the stability of training.

[0065] Subsequently, feature extraction is performed: Principal Component Analysis (PCA) and mutual information methods are used to extract key features that significantly influence the slider's motion. For example, PCA projects high-dimensional data into a low-dimensional space by reducing dimensionality while retaining the main information of the data; the mutual information method selects features with high mutual information values, such as displacement change rate, velocity fluctuation, and peak shear force, by quantifying the correlation between each feature and the target variable. Feature extraction not only reduces the dimensionality of the data and lowers the complexity of the model, but also improves the training efficiency and prediction accuracy of the model.

[0066] In addition, data augmentation techniques were employed to improve the model's generalization ability and prevent overfitting. By performing operations such as translation, scaling, and adding noise to the original data, more training samples were generated, enriching the diversity of the dataset and enhancing the model's robustness to input variations.

[0067] Step 2.2 Build a dataset based on the preprocessed data, and divide the dataset into training, validation, and test sets according to a set ratio. The specific ratio can be as follows:

[0068] Training set: accounting for 70% of the total data, used for model training.

[0069] Validation set: accounting for 15% of the total data, used for hyperparameter tuning of the model and implementation of early stopping strategies.

[0070] Test set: 15% of the total data, used for the final evaluation of model performance.

[0071] Organize the data into a format suitable for model input. For example, structure the time series data into a three-dimensional array, where the dimensions include the number of samples, the time step, and the number of features. Ensure that the data format meets the input requirements of the CNN-LSTM hybrid model.

[0072] Step 3: Construction of the CNN-LSTM hybrid model

[0073] The CNN-LSTM hybrid model includes a CNN part and an LSTM part, wherein the CNN part is used to extract the spatial features of the slider motion, and the LSTM part is used to capture the advantages of time series features.

[0074] The CNN component comprises multiple layers of one-dimensional convolutional layers, pooling layers, and batch normalization layers, wherein:

[0075] The first convolutional layer uses 64 one-dimensional convolutional kernels, each 3×3 in size, to extract low-level features of the slider motion. This layer captures local features and performs feature mapping through convolution operations, with an output shape of (L-3+1, 64), where L is the input length. The activation function used is ReLU (Rectified Linear Function) to introduce non-linearity.

[0076] The first max pooling layer follows the first convolutional layer, using a pooling window size of 2 to reduce feature dimensionality, computational complexity, and prevent overfitting. After pooling, the output shape is ((L-3+1) / 2, 64).

[0077] The second convolutional layer uses 128 one-dimensional convolutional kernels, each with a size of 3×3, to further extract deeper features. This layer enhances the model's spatial understanding of slider motion, and the output shape is (((L-3+1) / 2-3+1), 128), with ReLU activation function also used.

[0078] The second max pooling layer follows the second convolutional layer, further reducing the feature dimension, and the output shape becomes (((L-3+1) / 2-3+1) / 2, 128).

[0079] Batch normalization layers are added after each convolutional layer to stabilize the network training process, prevent internal covariate shift, and improve the training speed and stability of the model.

[0080] The features extracted by the CNN part are then fed into the LSTM part.

[0081] The LSTM section comprises two layers of LSTM cells, wherein:

[0082] The first LSTM layer contains 256 hidden units, effectively processing the temporal information in the input sequence data. This layer is configured to return sequences (return_sequences=1), ensuring that the output of each time step is passed to the next layer, thus preserving the complete time series information. This is crucial for subsequent LSTM layers, as it enables them to learn the short-term and long-term dependencies of the slider's motion.

[0083] The second LSTM layer contains 128 hidden units, further deepening the understanding of time series data. This layer does not return the sequence (return_sequences=0), but only outputs the state of the last time step to meet the input requirements of the fully connected layer.

[0084] Following the second LSTM unit is a fully connected layer. The output of this fully connected layer consists of two neurons, which correspond to the optimal displacement and velocity control commands of the slider, respectively. The activation function of these two neurons is a linear function to achieve direct output of continuous control parameters.

[0085] The two-layer LSTM structure enables the CNN-LSTM hybrid model to capture complex motion patterns and changing trends more effectively, thereby improving prediction accuracy.

[0086] The LSTM unit internally employs forget gates, input gates, and output gates to dynamically adjust the flow of information at each time step, ensuring that important information is retained while unimportant information is forgotten. This mechanism is particularly effective for capturing long-term dependencies, enabling the model to perform excellently when handling the nonlinear and time-varying characteristics of slider motion.

[0087] Step 4: Training the CNN-LSTM hybrid model

[0088] The CNN-LSTM hybrid model takes sensor data as input and outputs optimal displacement and velocity control commands for the slider. The trained CNN-LSTM hybrid model can effectively extract the spatial and temporal features of the slider's motion, enabling accurate prediction and control of the slider's displacement and velocity.

[0089] The CNN-LSTM hybrid model is trained using the training set data, employing the Adaptive Moment Estimation (AdamW) optimization algorithm combined with a weight decay strategy to prevent overfitting and improve the generalization ability of the CNN-LSTM hybrid model. The loss function uses a weighted combination of mean squared error (MSE) and mean absolute error (MAE), adjusting the ratio of MSE to MAE to balance error sensitivity and robustness.

[0090] During training, a cosine annealing learning rate scheduling strategy is employed to gradually decrease the learning rate, promoting model convergence. Simultaneously, an early stopping strategy is implemented: training is halted if the loss on the validation set fails to decrease within 15 consecutive epochs to prevent overfitting. Furthermore, K-fold cross-validation (e.g., K=5) is used to evaluate the model's performance across different dataset partitions, ensuring model stability and generalization ability.

[0091] Step 5: Validation and Optimization of the CNN-LSTM Hybrid Model

[0092] After the CNN-LSTM hybrid model is trained, it undergoes rigorous verification and optimization to ensure its stability and accuracy in practical applications.

[0093] First, the CNN-LSTM hybrid model is evaluated using a test set, and metrics such as MSE, MAE, and coefficient of determination (R²) are calculated for the CNN-LSTM hybrid model on the test set.

[0094] Then, residual analysis was performed. If the prediction error of the CNN-LSTM hybrid model is found to be normally distributed with a mean close to zero, it indicates that the CNN-LSTM hybrid model has high reliability in predicting the slider motion. The experimental results show that the CNN-LSTM hybrid model exhibits high prediction accuracy and good generalization ability on the test set.

[0095] If the performance on the test set is unsatisfactory, optimization can be achieved through the following methods: adjusting the hyperparameters of the CNN-LSTM hybrid model, such as increasing or decreasing the number of LSTM units, adjusting the size and number of convolutional kernels, and changing the learning rate; increasing the amount of training data to provide more samples for the CNN-LSTM hybrid model to learn from; or introducing regularization methods, such as Dropout layers, to further prevent overfitting. Through iterative optimization, a high-performance CNN-LSTM hybrid model can eventually be obtained.

[0096] Step 6: Model Deployment and Real-Time Control

[0097] The validated and optimized CNN-LSTM hybrid model will be deployed in the active control system of the slider of the active unfolding end device to achieve real-time high-precision control of the slider's motion.

[0098] Step 6.1 Model Deployment: Convert the trained CNN-LSTM model into a format suitable for real-time control (such as TensorFlow Lite) and embed it into the control unit of the active control system of the slider of the active unfolding end device. Specifically, this includes:

[0099] Model conversion: The trained CNN-LSTM hybrid model is converted into a format suitable for embedded systems and optimized to reduce model size and improve inference speed. For example, TensorFlow Lite can be used to convert the trained CNN-LSTM model into a lightweight model suitable for the real-time inference needs of embedded devices.

[0100] Embedded Platform Selection: Choose a suitable embedded platform (such as Raspberry Pi, NVIDIA Jetson, or a dedicated industrial controller) as the hardware foundation for model deployment. Install the necessary software libraries and dependencies, such as the TensorFlow Lite runtime, on the selected embedded platform to ensure that the converted model can run efficiently on the device.

[0101] Software integration: Develop embedded software to enable real-time inference of the model and generation of control commands.

[0102] Step 6.2 Real-time control, the specific implementation steps are as follows:

[0103] First, the control unit of the slider active control system receives various sensor data in real time. After data synchronization (ensuring that the acquisition and transmission of sensor data are consistent with the time step of model inference, avoiding data delay or loss) and data preprocessing, the consistency of the data input to the CNN-LSTM hybrid model is ensured.

[0104] Secondly, the preprocessed data is input into the CNN-LSTM hybrid model for real-time inference to generate the optimal displacement and velocity control commands for the slider.

[0105] Then, the control unit sends the optimal displacement and velocity control commands output by the CNN-LSTM hybrid model to the servo motor. The servo motor adjusts its output torque to drive the slider to perform corresponding movements, thereby adjusting the slider's movement path and speed to ensure that the slider's movement is highly synchronized with the sample's deformation process.

[0106] Finally, the information on the slider's motion is collected in real time by sensors and transmitted to the control unit of the slider active control system to form a closed-loop control. The CNN-LSTM hybrid model continuously adjusts the control commands to ensure the timeliness and accuracy of the response.

[0107] Furthermore, to ensure the safety and stability of the active control system for the actively deployable end-cap device's slider, an anomaly detection and fault protection mechanism has been implemented. Specific measures include:

[0108] Anomaly detection: Real-time monitoring of CNN-LSTM hybrid model predictions and actual control commands to detect anomalies or situations exceeding preset ranges.

[0109] Fault protection: When an abnormal situation is detected, the active control system of the active unfolding end device automatically issues an alarm and stops all slider movement to prevent sample damage or equipment failure.

[0110] The above description is merely a description of preferred embodiments of this application and is not intended to limit the scope of this application in any way. Any changes or modifications made by those skilled in the art based on the above-disclosed technical content should be considered as equivalent and valid embodiments and fall within the scope of protection of the technical solution of this application.

Claims

1. An optimization control method for an actively deployable end-cap device based on a deep learning model, wherein the actively deployable end-cap device comprises: A circular support base is used to support other components and devices and is equipped with corresponding sliding parts; The guide rail system comprises multiple sets of guide rail blocks and several precision ball bearings, the guide rail blocks being able to move radially on a circular support base; the slider system comprises multiple rings of sliders arranged radially around the center of the base, the sliders cooperating with the guide rail blocks, and a pressure-resistant arc-shaped rubber diaphragm between the sliders; the slider active control system comprises a micro servo motor, a force transmission rod, a sensor, and a control unit; characterized in that the implementation steps of the control method include: Step 1: Sensor deployment and data acquisition; Step 2: Data preprocessing and dataset creation; Step 3: Construction of the CNN-LSTM hybrid model; Step 4: Training the CNN-LSTM hybrid model; Step 5: Validation and optimization of the CNN-LSTM hybrid model; Step 6: Model Deployment and Real-Time Control: The validated and optimized CNN-LSTM hybrid model is deployed in the active control system of the slider of the active unfolding end device to achieve real-time, high-precision control of the slider's motion; The CNN-LSTM hybrid model includes a CNN part and an LSTM part, wherein the CNN part is used to extract the spatial features of the slider motion, and the LSTM part is used to capture the advantages of time series features. The CNN component comprises multiple layers of one-dimensional convolutional layers, pooling layers, and batch normalization layers, wherein: The first convolutional layer uses 64 one-dimensional convolutional kernels, each of which is 3×3 in size. It aims to extract low-level features of the slider motion. This layer captures local features and performs feature mapping through convolution operations. The output shape is (L-3+1, 64), where L is the input length. The activation function is ReLU, which is used to introduce non-linearity. The first max pooling layer follows the first convolutional layer and uses a pooling window size of 2 to reduce feature dimensionality, computational complexity, and prevent overfitting. After pooling, the output shape is ((L-3+1) / 2, 64). The second convolutional layer uses 128 one-dimensional convolutional kernels, each of which is 3×3 in size, to further extract deeper features. This layer enhances the model's spatial understanding of slider motion, and the output shape is (((L-3+1) / 2-3+1), 128), with ReLU as the activation function. The second max pooling layer follows the second convolutional layer, further reducing the feature dimension, and the output shape becomes (((L-3+1) / 2-3+1) / 2, 128); Batch normalization layers are added after each convolutional layer to stabilize the network training process, prevent internal covariate shift, and improve the training speed and stability of the model. The features extracted by the CNN are then fed into the LSTM. The LSTM section comprises two layers of LSTM cells, wherein: The first LSTM unit contains 256 hidden units, which can effectively process the time information in the input sequence data; this layer is set to return the sequence, so that the output of each time step is passed to the next layer, thus preserving the complete time series information. The second LSTM unit contains 128 hidden units, further deepening the understanding of time series data; this layer does not return the sequence, but only outputs the state of the last time step to meet the input requirements of the fully connected layer; Following the second LSTM unit is a fully connected layer. The output of the fully connected layer consists of two neurons, which correspond to the optimal displacement and velocity control commands of the slider, respectively. The activation function of these two neurons is a linear function to achieve direct output of continuous control parameters.

2. The method for optimizing control of an active unfolding end-device device based on a deep learning model as described in claim 1, characterized in that, Step 1: On each slider of the actively deployable end-cap device, various sensors are arranged, including displacement sensors, velocity sensors, pressure sensors, shear force sensors, and friction force sensors. The various types of sensors are arranged as follows: The displacement sensor is installed on the side of the slider, facing the direction of the slider's movement, to measure the slider's displacement changes in real time; The speed sensor is installed on the guide block corresponding to the slider to obtain the real-time speed information of the slider; A pressure sensor is embedded in the end face of the slider that contacts the sample to measure the normal pressure applied by the slider to the sample. Shear force sensor and friction force sensor are embedded in the end face of the slider that contacts the sample, and are used to measure the shear force and friction force between the slider and the sample, respectively. During the experiment, the required data is collected synchronously via a high-speed data acquisition card as the slider moves.

3. The method for optimizing control of an active unfolding end-device device based on a deep learning model as described in claim 1, characterized in that, Step 2: The collected raw sensor data undergoes preprocessing, including data cleaning, normalization, and feature extraction. A dataset is built based on the preprocessed data, and the dataset is divided into a training set, a validation set, and a test set.

4. The method for optimizing control of an active unfolding end-device device based on a deep learning model as described in claim 1, characterized in that, Step 4: The training set data is used to train the CNN-LSTM hybrid model, which employs the Adam optimization algorithm of adaptive moment estimation combined with a weight decay strategy to prevent overfitting and improve the generalization ability of the CNN-LSTM hybrid model; the loss function is a weighted combination of mean squared error (MSE) and mean absolute error (MAE). During training, a cosine annealing learning rate scheduling strategy is adopted to gradually reduce the learning rate during training, promoting model convergence. At the same time, an early stopping strategy is set to stop training when the loss on the validation set does not decrease within 15 consecutive epochs to prevent overfitting. In addition, K-fold cross-validation is used to evaluate the model's performance under different dataset partitions, ensuring the model's stability and generalization ability.

5. The optimized control method for an active unfolding end-device device based on a deep learning model as described in claim 1, characterized in that, Step 5: First, the CNN-LSTM hybrid model is evaluated using a test set, and the MSE, MAE, and coefficient of determination R² of the CNN-LSTM hybrid model on the test set are calculated. Then, residual analysis is performed. If the prediction error of the CNN-LSTM hybrid model is found to be normally distributed with a mean close to zero, it indicates that the CNN-LSTM hybrid model has high reliability in predicting the slider motion. Conversely, optimization can be achieved through the following methods: adjusting the hyperparameters of the CNN-LSTM hybrid model, increasing the amount of training data, introducing regularization methods, and iteratively optimizing to ultimately obtain a high-performance CNN-LSTM hybrid model.

6. The method for optimizing control of an active unfolding end-device device based on a deep learning model as described in claim 1, characterized in that, Step 6: The trained CNN-LSTM model is converted into a format suitable for real-time control and embedded into the control unit of the active control system of the slider of the active unfolding end device. It receives sensor data in real time, generates and executes the optimal displacement and velocity control commands to achieve a high degree of synchronization between slider movement and sample deformation. The real-time control process is as follows: First, the control unit of the slider active control system receives data from various sensors in real time. After data synchronization and data preprocessing, the consistency of the data input to the CNN-LSTM hybrid model is ensured. Secondly, the preprocessed data is input into the CNN-LSTM hybrid model for real-time inference to generate the optimal displacement and velocity control commands for the slider. Then, the control unit sends the optimal displacement and velocity control commands output by the CNN-LSTM hybrid model to the servo motor. The servo motor adjusts its output torque to drive the slider to perform corresponding movements, thereby adjusting the slider's movement path and speed to ensure that the slider's movement is highly synchronized with the sample's deformation process. Finally, the information on the slider's motion is collected in real time by sensors and transmitted to the control unit of the slider active control system to form a closed-loop control. The CNN-LSTM hybrid model continuously adjusts the control commands to ensure the timeliness and accuracy of the response.

Images