A real-time prediction method for lunar soil drilling temperature based on spatiotemporal feature fusion

Abstract

The present invention relates to the fields of computer science, artificial intelligence, lunar exploration and drilling technology, and particularly to a method for real-time prediction of lunar soil drilling temperature by integrating spatiotemporal features. The method comprises the following steps: collecting drilling state parameters and drilling temperature data, then performing data preprocessing, incorporating a proposed model optimization method of early stopping-Bayesian hyperparameter optimization into the training process to effectively adjust the model's hyperparameters, and finally inputting a test set into the trained model to achieve intelligent real-time prediction of drilling temperature. Furthermore, actual application effect tests were conducted, demonstrating good robustness and generalization capabilities. The method utilizes deep learning technology, combined with a Bi-directional Long Short-Term Memory (Bi-LSTM) network and a self-attention mechanism, to automatically focus on features that are critical to drilling temperature prediction, thereby accurately predicting lunar soil drilling temperature in real time.

Description

Technical Field

[0001] The present invention relates to the fields of computer science, artificial intelligence, lunar exploration and drilling technology, and in particular to a method for real-time prediction of lunar soil drilling temperature by integrating spatiotemporal features. Background Art

[0002] As a close neighbor of the Earth, the Moon has important scientific and resource value. In lunar exploration missions, drilling is an important means to obtain deep information about the lunar soil. However, the lunar surface environment is complex, and the temperature generated during drilling is difficult to accurately predict, which affects drilling efficiency and safety. Therefore, studying a reliable real-time prediction method for lunar soil drilling temperature is crucial to improving the efficiency of exploration missions and ensuring equipment safety. The current drilling temperature prediction methods exist. So far, many scholars have carried out research on lunar soil drilling temperature, which can be roughly divided into three categories: experimental methods, theoretical methods, and numerical simulation methods. Although the above methods provide insights into the research on lunar soil drilling temperature prediction, these methods cannot be directly applied to the real-time prediction of lunar soil drilling temperature. The experimental method is more inclined to reveal the law of temperature change during the drilling process and cannot provide specific temperature predictions. The theoretical method makes a lot of assumptions based on experience and speculation, which may be quite different from the actual drilling process. The numerical simulation method is prone to calculation errors in certain drilling scenarios, and accurate modeling is difficult and the computational cost is too high. In addition, previous studies mainly focused on the spatial relationship between variables, rarely considering the time dependence within the variables, and there are certain limitations on the real-time requirements. In order to solve these limitations, the present invention aims to consider both the spatial relationship between variables and the time dependence within the variables, and proposes a real-time prediction method for lunar soil drilling temperature with high precision based on the fusion of spatiotemporal features. Summary of the Invention

[0003] In order to overcome the problems described in the above-mentioned prior art that cannot provide specific temperature predictions, the theoretical method makes a large number of assumptions based on experience and speculation, which may differ greatly from the actual drilling process, the numerical simulation method is prone to calculation errors in certain drilling scenarios, the difficulty of accurate modeling is high, and the computational cost is too high, the present invention provides a real-time prediction method for lunar soil drilling temperature by fusing spatiotemporal features.

[0004] In order to solve the above technical problems, the technical solutions of the present invention are as follows:

[0005] A real-time prediction method for lunar soil drilling temperature based on spatiotemporal feature fusion includes:

[0006] S1: Acquire lunar soil drilling sampling data, wherein the lunar soil drilling sampling data includes real-time drilling state parameters and drilling temperature data at multiple positions of the drilling tool;

[0007] S2: Preprocessing the lunar soil drilling sampling data, wherein the preprocessing includes packaging the drilling state parameters at an earlier time and the drilling temperature data at a later time;

[0008] S3: Build a model that introduces a self-attention mechanism and a Bi-LSTM network;

[0009] S4: training the model using the lunar soil drilling sampling data pre-processed in step S2;

[0010] S5: Optimize the model trained in step S4 to obtain an optimized model;

[0011] S6: Input the previously acquired drilling state parameters into the optimized model to predict the drilling temperature data of the drilling tool in real time.

[0012] Furthermore, in step S1, the lunar soil drilling sampling data is obtained by using a plurality of sensors to collect seven drilling state parameters and eight drilling temperature data based on a simulated lunar soil environment system and a drilling test bench. The seven drilling state parameters include drilling thrust, rotary torque, drilling depth, feed speed, rotary speed, feed motor current, and rotary motor current.

[0013] The 8 types of drilling temperature data include 4 drill rod temperatures, drill bit rake temperature, drill bit inner surface temperature, drill bit end temperature and core barrel temperature.

[0014] Furthermore, in step S2, the lunar soil drilling sampling data is preprocessed, wherein:

[0015] S2.1: Filter out the values ​​outside the preset range from the sampled data and treat them as abnormal data;

[0016] S2.2: Process abnormal data;

[0017] S2.3: interpolate the drilling temperature data after abnormal data processing;

[0018] S2.4: Align the drilling state parameters and drilling temperature data;

[0019] S2.5: normalize the aligned drilling state parameters and drilling temperature data;

[0020] S2.6: Finally, the 7 drilling status parameter data of the first 20 seconds after processing in steps S2.1 to S2.5 and the temperature data of 8 positions on the drilling tool in the next 1 second are packaged and then divided into training set, validation set and test set according to the ratio of 6:2:2.

[0021] Furthermore, in step S2.2, the abnormal data is processed, specifically:

[0022]

[0023] Among them, Q t represents the replacement value of the outlier, Q t-1 represents the value of the abnormal value at the moment before, Q t+1 Represents the value one moment after the outlier.

[0024] Furthermore, in step S2.3, the drilling temperature data after abnormal data processing is interpolated. The interpolation processing is specifically as follows:

[0025]

[0026] Among them, t i+n -t i is the time difference between data point i+n and data point i, T i is the temperature at data point i, T i+1 is the temperature at data point i+1, T i+n -T i is the temperature difference between the two data points.

[0027] Furthermore, in step S2.4, the drilling state parameters and the drilling temperature data are aligned, specifically:

[0028]

[0029] Among them, F t Represents the replacement value of the drilling status parameter in that second, N represents N drilling status parameters stored in that second, Represents the nth drilling status parameter within that second.

[0030] Furthermore, in step S3, in the constructed model, the model includes: an input layer, a Bi-LSTM layer, a BN layer, an Attention layer, a Dropout layer, an FC layer, and an output layer, wherein:

[0031] The input of the input layer is the preprocessed drilling state parameters. The Bi-LSTM layer, BN layer, Attention layer, Dropout layer, FC layer and output layer are connected in sequence after the input layer. The output layer outputs the predicted value of the drilling temperature data of the drill tool in the next 1 second.

[0032] Furthermore, in step S5, the model trained in step S4 is optimized, and the model optimization method is specifically as follows:

[0033] The model is optimized using the early stopping-Bayesian hyperparameter optimization method to obtain a hyperparameter combination, and the optimized model is obtained based on the hyperparameter combination.

[0034] Furthermore, the model is optimized using the early stopping-Bayesian hyperparameter optimization method, specifically:

[0035] S5.1: Define hyperparameter space;

[0036] S5.2: Initialize the Bayesian optimizer;

[0037] S5.3: The Bayesian optimizer obtained in step S5.2 recommends the next hyperparameter combination and continues training the model;

[0038] S5.4: Use early stopping during each model training process to prevent model overfitting;

[0039] S5.5: Repeat steps S5.3 and S5.4 until the optimizer converges and the optimal hyperparameter combination is found or the preset upper limit of hyperparameter searches is reached.

[0040] S5.6: Output the current hyperparameter combination.

[0041] Furthermore, in step S6, the model is used to predict the lunar soil drilling temperature in real time, specifically:

[0042] The seven drilling status parameter data of the previous 20 seconds are collected in real time, and the drilling status parameter data are input into the model to predict the temperature of eight locations on the drilling tool in real time in the next 1 second.

[0043] Compared with the prior art, the beneficial effects of the technical solution of the present invention are:

[0044] The present invention adopts a model that combines the self-attention mechanism and the bidirectional long short-term memory network. It can predict the temperature of eight locations on the drilling tool in the next second in real time based on the seven drilling status parameters in the previous 20 seconds. In order to further optimize the model performance, the present invention also proposes an early stopping-Bayesian hyperparameter optimization model optimization method to effectively adjust the model's hyperparameters, considering both the spatial relationship between variables and the time dependency within the variables, thereby improving the prediction performance. BRIEF DESCRIPTION OF THE DRAWINGS

[0045] Figure 1 This is a flow chart of a method for real-time prediction of lunar soil drilling temperature by integrating spatiotemporal features according to the present invention;

[0046] Figure 2 This is a model framework diagram in an embodiment of the present invention;

[0047] Figure 3 This is a graph showing the loss results of the training process of the model in an embodiment of the present invention;

[0048] Figure 4This is an optimization history chart of ESBO (early stopping Bayesian optimization) in an embodiment of the present invention;

[0049] Figure 5 This is a diagram showing the prediction effect of the model in the test set according to an embodiment of the present invention;

[0050] Figure 6 1 is a comparison chart of prediction results of five typical prediction models in an embodiment of the present invention;

[0051] Figure 7 This is a diagram showing the predicted results of the model in the simulated lunar environment drilling test in an embodiment of the present invention. DETAILED DESCRIPTION

[0052] The accompanying drawings are for illustrative purposes only and are not to be construed as limiting this patent;

[0053] In order to better illustrate this embodiment, some parts in the drawings may be omitted, enlarged, or reduced, and do not represent the actual product size;

[0054] It is understandable to those skilled in the art that some well-known structures and their descriptions may be omitted in the drawings.

[0055] The technical solution of the present invention is further described below with reference to the accompanying drawings and embodiments.

[0056] Example 1

[0057] A real-time prediction method for lunar soil drilling temperature based on the fusion of spatiotemporal features, such as Figure 1 Shown, including:

[0058] S1: Acquire lunar soil drilling sampling data, wherein the lunar soil drilling sampling data includes real-time drilling state parameters and drilling temperature data at multiple positions of the drilling tool;

[0059] S2: Preprocessing the lunar soil drilling sampling data, wherein the preprocessing includes packaging the drilling state parameters at an earlier time and the drilling temperature data at a later time;

[0060] S3: Build a model that introduces a self-attention mechanism and a Bi-LSTM network;

[0061] S4: training the model using the lunar soil drilling sampling data pre-processed in step S2;

[0062] S5: Optimize the model trained in step S4 to obtain an optimized model;

[0063] S6: Input the previously acquired drilling state parameters into the optimized model to predict the drilling temperature data of the drilling tool in real time.

[0064] The present invention provides a real-time prediction method for lunar soil drilling temperature by fusing spatiotemporal features, aiming to solve the problem of reliable real-time prediction of drilling tool drilling temperature in the complex environment of the lunar surface. The model incorporates a self-attention mechanism and a Bi-LSTM network to effectively adjust the model's hyperparameters. The test set is input into the optimized model to achieve end-to-end intelligent real-time prediction of drilling temperature. From the actual application effect, the model demonstrates good robustness and generalization ability. The present invention considers both the spatial relationship between variables and the temporal dependency within the variables.

[0065] Example 2

[0066] This embodiment, based on the first embodiment, further discloses the following contents:

[0067] In step S2, the lunar soil drilling sampling data is preprocessed, wherein:

[0068] S2.1: Filter out the values ​​outside the preset range from the sampled data and treat them as abnormal data;

[0069] S2.2: Process abnormal data;

[0070] S2.3: interpolate the drilling temperature data after abnormal data processing;

[0071] S2.4: Align the drilling state parameters and drilling temperature data;

[0072] S2.5: normalize the aligned drilling state parameters and drilling temperature data;

[0073] S2.6: Finally, the 7 drilling status parameter data of the first 20 seconds after processing in steps S2.1 to S2.5 and the temperature data of 8 positions on the drilling tool in the next 1 second are packaged and then divided into training set, validation set and test set according to the ratio of 6:2:2.

[0074] In step S1, the lunar soil drilling sampling data is obtained by using a variety of sensors to collect seven drilling state parameters and eight drilling temperature data based on a simulated lunar soil environment system and a drilling test bench. The seven drilling state parameters include drilling thrust, rotary torque, drilling depth, feed speed, rotary speed, feed motor current, and rotary motor current.

[0075] The 8 types of drilling temperature data include 4 drill rod temperatures, drill bit rake temperature, drill bit inner surface temperature, drill bit end temperature and core barrel temperature.

[0076] During the specific implementation process, data collected from simulated lunar soil drilling and sampling tests conducted in a vacuum chamber simulating the lunar surface environment were used. The experimental system in the present invention is mainly a thermal characteristics test bench for robotic lunar drilling tests. A robotic drill is installed on the top of the thermal vacuum environment simulation device to stretch a 2m-level drill to penetrate in the simulated lunar soil environment. The drill consists of a drill bit for crushing the lunar soil, a hollow drill rod with a spiral thread for powder removal, and a core tube for obtaining lunar soil samples. Eight very thin thermocouples (0.5 mm in diameter) are fixed at the drill bit, spiral drill rod, and core tube of the drill to monitor temperature changes in the lunar soil environment during penetration. The custom-designed drill rod integrates three vacuum feedthroughs and a slip ring to transmit temperature signals from the internal vacuum environment to the external room environment. Based on the simulated lunar soil environment system and the drilling test bench, drilling and sampling tests were carried out under vacuum conditions. During the test, multiple sensors were used to collect drilling state parameters and drilling temperature data. 22 drilling sampling experiments were carried out in the simulated lunar soil, with a penetration depth of 2 meters.

[0077] Abnormal data processing: Due to factors such as sensor failure and environmental interference, the collected data may contain abnormal data or missing data. These problems will have a negative impact on the prediction effect of the established model. Therefore, these abnormal data need to be processed. First, according to the theoretical range of the parameters, the values ​​that are obviously out of range are screened out and regarded as outliers. These outliers will be excluded from the data set and treated as missing values.

[0078] Step S2.2 abnormal data processing, specifically:

[0079]

[0080] Among them, Q t represents the replacement value of the outlier, Q t-1 represents the value of the abnormal value at the moment before, Q t+1 Represents the value one moment after the outlier.

[0081] Temperature interpolation processing, although the thin thermocouple is tightly fixed to obtain the most accurate possible temperature change of the drill tool, the measurement accuracy of the sensor is 1 degree Celsius, which means that all measured data are integers, and the temperature remains at the same T∈N value until the next different value T+m or Tn can be reached. In addition, due to the temperature sampling frequency, when the temperature changes sharply in some extreme penetration conditions, not all integral values ​​are recorded. Due to the above reasons, the actual temperature curve collected changes in the form of steps, especially when the temperature change is small. This step-like data is not suitable for data-based modeling, so linear interpolation processing is performed on the temperature data. For a specific continuous sampling sequence, the temperature may remain at T for a long time. i The value (from time ti to t i+n-1 ) is that the temperature fluctuation is not severe; then at time t i+n , then it becomes T i+n , there is good reason to believe that this temperature increase is not sudden, but rather occurs over time t i+n -t i The accumulation of i and t i+n The temperature data of the two endpoints is removed from t i+1 to t i+n-1 The temperature data between the two values ​​is used, and the linear interpolation method is used to create the temperature data that is removed in the middle. i+1 Taking the temperature interpolation as an example, in step S2.3, the drilling temperature data after abnormal data processing is interpolated. The temperature interpolation process is specifically as follows:

[0082]

[0083] Among them, t i+n -t i is the time difference between data point i+n and data point i, T i is the temperature at data point i, T i+1 is the temperature at data point i+1, T i+n -T i is the temperature difference between the two data points.

[0084] In step S2.4, data alignment is performed. In actual operation, the acquisition frequencies of different parameters are different. In the specific implementation process, drilling force data is usually collected at a very high frequency because the robot's drilling is controlled based on the feedback of the drilling force. Temperature is sampled at a lower frequency because the generated heat takes time to transfer and dissipate. In order to study the relationship between drilling state parameters and drilling temperature, a better strategy is to make the data frequency of all parameters equal, which is also more suitable for model training. Therefore, for data with high acquisition frequency, such as drilling force (including drilling pressure and rotary torque), the following processing is required (taking drilling pressure data as an example below), specifically:

[0085]

[0086] Among them, F t Represents the drilling pressure data of that second, and N represents N drilling pressure data stored in that second. Represents the nth drilling pressure data within that second.

[0087] Data normalization: When modeling, seven drilling state parameters are involved, which have different dimensions and units. This difference may affect the results of data analysis and increase the difficulty of model training. To solve this problem, the data needs to be normalized to eliminate the dimensional influence between different drilling state parameters. The min-max normalization method is adopted. The present invention performs a linear transformation on each parameter and maps its value to a specific range. The calculation formula is as follows:

[0088]

[0089] Among them, x represents a drilling state parameter in the original data, x normalized represents the normalized drilling state parameter, x min and x max Respectively represent the minimum and maximum values ​​of the parameter in the data set. Through this formula, the parameter value can be mapped to the range between 0 and 1, achieving a unified dimension so that they can be compared and analyzed on the same scale. This will help improve the accuracy and reliability of data analysis and reduce the difficulty of model training.

[0090] Data packaging and processing: Before training the model, the data needs to be packaged and processed. In practical applications, since there is no temperature sensor on the sampling drill, only a small number of sensors can collect some drilling state parameter data. Therefore, this study aims to develop a model that aims to predict the temperature of 8 locations on the drill in the next second in real time based on the 7 drilling state parameter data of the previous 20 seconds. Therefore, the data set must be structured first to determine the input and output shapes of the model. The 7 drilling state parameters at the i-th second are represented as x i ={x i (1) ,x i (2) ,x i (3) ,x i (4) ,x i (5) ,x i (6) ,x i (7)}, where x i (1) ,x i (2) ,x i (3) ,x i (4) ,x i (5) ,x i(6) ,x i (7) They represent the drilling thrust, rotary torque, drilling depth, feed speed, rotary speed, feed motor current and rotary motor current at the i-th second respectively. The temperature of the eight positions on the drilling tool at the i-th second is expressed as y i ={y i (1) ,y i (2) ,y i (3) ,y i (4) ,y i (5) ,y i (6) ,y i (7) ,y i (8)}, where y i (1) ,y i (2) ,y i (3) ,y i (4) ,y i (5) ,y i (6) ,y i (7) ,y i (8) They represent the drill rod temperature 1, drill rod temperature 2, drill rod temperature 3, drill rod temperature 4, drill bit front cutter temperature, drill bit inner surface temperature, drill bit end temperature and core tube temperature at the i-th second respectively. The single sample input of the data set is X n and its corresponding expected output Y n The matrix expression after packaging is as follows. According to this packaging rule, the entire data set is traversed and all samples are input into X n and its corresponding expected output Y n Pack them one by one:

[0091]

[0092]

[0093] In addition, in order to improve the training efficiency and prediction effect of the model, the entire data set is divided into several batches for training. The data set is packaged in the form of Batch, and each Batch contains 64 samples. After constructing the data set, the data set is divided into three parts, namely training set, validation set and test set, with a division ratio of 6:2:2. The training set consists of X train and Y trainIt consists of 11372 samples for model training, and the validation set consists of X valid and Y valid It consists of 3690 samples, which are used for model optimization. The test set consists of X test and Y test It consists of 3690 samples for model testing.

[0094] Example 3

[0095] This embodiment, based on Embodiments 1 and 2, further discloses the following contents:

[0096] In step S3, in the constructed model, the model includes: an input layer, a Bi-LSTM layer, a BN layer, an Attention layer, a Dropout layer, an FC layer, and an output layer, wherein:

[0097] The input of the input layer is the preprocessed drilling state parameters. The Bi-LSTM layer, BN layer, Attention layer, Dropout layer, FC layer and output layer are connected in sequence after the input layer. The output layer outputs the predicted value of the drilling temperature data.

[0098] Model design: Designing a reasonable model is the key to successfully predicting drilling temperature. This paper designs a drilling temperature model that integrates spatiotemporal features. The model considers the spatial relationship between drilling temperature and drilling state parameters as well as the time dependence between parameters. The framework of the model is as follows: Figure 2 As shown in Figure 1, the overall model consists of seven parts: input layer, Bi-LSTM layer, BN layer, Attention layer, Dropout layer, fully connected layer, and output layer. The output layer contains eight drilling temperatures for the next second. The model input is seven drilling status parameter data for the previous 20 seconds. Because the generation and dissipation of drill tool heat is affected by multiple historical and future inputs, a Bi-LSTM recurrent neural network is selected as the basic model. It has stronger feature capture capabilities. During the feedforward process of the neural network, the distribution of input data may change. To improve model accuracy, a batch normalization (BN) layer is added after the Bi-LSTM layer. In addition, an attention mechanism (Attention layer) is introduced into the drilling temperature prediction model to enhance the influence of key features of the input data and better capture the internal correlation of drill tool temperature changes, thereby improving the model's prediction effect. To avoid overfitting and improve the model's generalization ability, a Dropout layer is also introduced. Finally, the extracted features are aggregated into a fully connected (FC) layer for prediction output. The output layer outputs the predicted value of the drilling tool's drilling temperature data for the next second.

[0099] Model training: After the model framework is built, the model is trained using the prepared training set. The training process includes multiple steps such as forward propagation, back propagation, and weight update. Through these steps, the model continuously adjusts the weights and gradually learns the characteristics of the input samples. In each training cycle, the model is optimized by monitoring the performance of the model. The combination of these steps enables the model to automatically learn the characteristics of the data from the training set and perform better in subsequent task processing. Through the training process, the model can gradually improve its ability to predict drilling temperature, so that the model can more accurately predict future temperature changes. The model uses the best hyperparameter combination searched by the ESBO method for training. The training round is set to 1000. The loss results of the training process are as follows: Figure 3 As shown in the figure, after 564 epochs, the termination condition of the early stopping algorithm was reached and further training was stopped. It can be observed that in the first 50 epochs, the training loss and validation loss decreased rapidly, and then gradually stabilized with small fluctuations. The training loss and validation loss showed a similar downward trend, but the training loss decreased slightly faster than the validation loss. After 500 epochs, the stable values ​​of the normalized mean square error (MSE) converged to about 0.0004 and 0.0009, respectively. Due to the use of the BN layer, Dropout layer and early stopping strategy, the difference between the training loss and the validation loss was small, which indicated that there was no overfitting problem and verified the reliability of the model. These results show that the designed model can effectively learn the characteristics of the data during training and achieve good performance.

[0100] Example 4

[0101] This embodiment further discloses the following contents based on Embodiments 1, 2, and 3:

[0102] In step S5, the model is optimized to obtain an optimized model. The model optimization method is specifically as follows:

[0103] The model is optimized using the early stopping-Bayesian hyperparameter optimization method to obtain a hyperparameter combination, and the optimized model is obtained based on the hyperparameter combination.

[0104] The model is optimized using the early stopping-Bayesian hyperparameter optimization method, specifically:

[0105] S5.1: Define hyperparameter space;

[0106] S5.2: Initialize the Bayesian optimizer;

[0107] S5.3: The Bayesian optimizer obtained in step S5.2 recommends the next hyperparameter combination and continues training the model;

[0108] S5.4: Use early stopping during each model training process to prevent model overfitting;

[0109] S5.5: Repeat steps S5.3 and S5.4 until the optimizer converges and the optimal hyperparameter combination is found or the preset upper limit of hyperparameter searches is reached.

[0110] S5.6: Output the current hyperparameter combination.

[0111] The number of hyperparameter searches set in the ESBO of the present invention is 100, the number of early stopping is set to 30, and the optimization history of ESBO is as follows: Figure 4 As shown, the line is the optimal value currently achieved (the highest R 2 ), point is the current value. It can be seen that the optimization effect shows an upward trend with the increase of search times. When the search reaches 62 times, the optimal point is found. Its R 2 The value is 0.972, where R 2 Used to evaluate the accuracy of the model, MAPE is used to evaluate the average value of the model:

[0112]

[0113] Where n is the total number of samples in the test set, y is the actual output, is the predicted output, is the average value of multiple actual outputs y.

[0114] Example 5

[0115] This embodiment, based on Embodiments 1, 2, 3, and 4, further discloses the following contents:

[0116] In step S6, the model is used to predict the lunar soil drilling temperature in real time, specifically: 7 types of drilling status parameter data in the previous 20 seconds are collected in real time, the drilling status parameter data are input into the model, and the temperature of 8 positions on the drilling tool in the next 1 second is predicted in real time.

[0117] Model prediction results: The final model was trained using the best hyperparameter combination obtained through ESBO. Finally, the trained model was tested on the test set to predict the drilling temperature of 3690 samples in the test set (the temperature at 8 different locations on the drill bit). The prediction results are as follows: Figure 5 As shown in the figure, the line represents the line where the predicted value and the true value are completely fitted (predicted value = true value). The points in the figure are composed of the true value and predicted value of each test set sample. The more the points converge to the fitting line, the better the prediction effect. It can be seen from the figure that the point sets composed of the predicted values ​​and true values ​​of the 8 drilling temperatures are obviously clustered near the fitting line, indicating that the predicted values ​​and the true values ​​are basically consistent. The R2 The values ​​are 0.956, 0.941, 0.945, 0.935, 0.970, 0.977, 0.982, and 0.947, all above 0.9; the MAPE values ​​are 5.22%, 5.24%, 5.06%, 5.11%, 4.98%, 4.91%, 4.86%, and 5.04%, all below 6%. After statistics, the average R 2 The value is 0.957, and the MAPE value is 5.05%. The experimental results show that the developed model can accurately predict the eight drilling temperatures in the next 1 second during the lunar soil drilling process. The experimental results successfully verify the feasibility and effectiveness of the method. In addition, the model proposed in this invention only takes 0.0021 seconds to predict a single sample, which can meet the requirements of real-time prediction.

[0118] Example 6

[0119] This embodiment, based on Embodiments 1, 2, 3, 4, and 5, further discloses the following contents:

[0120] Comparison of different models: In order to further verify the effectiveness of the method proposed in this invention and highlight its advantages, comparative experiments were conducted and compared with five different models, including Bi-LSTM and LSTM in recurrent neural networks, feedforward neural network model (FNN) in artificial neural network methods, random forest (RF) in ensemble learning, and support vector machine (SVM) in statistical machine learning models. Among these models, FNN, RF and SVM do not have memory function and therefore cannot be directly used for the prediction task of the project of this invention. In order to use them for experimental comparison, it is necessary to transform the input shape of the data set sample while maintaining the expected output Y of the sample. n No change, the original single sample input X n The dimension is changed from 5×20 to 1×100, and the corresponding matrix expression is as follows:

[0121]

[0122] in Represent the drilling thrust, rotary torque, drilling depth, feed speed, rotary speed, feed motor current and rotary motor current at the i-th second respectively. After reconstructing the corresponding data set, the five models of Bi-LSTM, LSTM, FNN, RF and SVM were used to predict the 8 drilling temperatures of 3690 samples in the test set, and the ESBO algorithm was used to optimize the models. The specific prediction comparison results are shown in the figure. Figure 6 As shown in the comparison results, it can be seen that the developed model has a good 2The MAPE value is significantly better than that of the other five typical models, and the MAPE value is also significantly lower than that of the other five typical models, which shows that the method proposed in the present invention has better prediction accuracy. The comparative experiment successfully verifies the effectiveness of the method proposed in the present invention and highlights the advantages of the method proposed in the present invention. In addition, according to the comparison results of the prediction performance, the prediction performance is arranged from good to bad: ATT-Bi-LSTM>Bi-LSTM>LSTM>ANN>RF>SVM. It can be seen that the performance of the three models of recurrent neural network (ATT-Bi-LSTM, Bi-LSTM and LSTM) is significantly better than other models, which once again reflects the importance of fusion extraction of spatiotemporal features for drilling temperature prediction.

[0123] In the specific implementation process, in order to verify the effectiveness and reliability of the method proposed in the present invention, the developed model was used to conduct actual application effect tests in a vacuum chamber simulating the lunar environment using the same drilling test bench as the source data. The eight drilling temperatures during the entire drilling process were predicted in real time. During the drilling process, seven drilling state parameters were collected at a frequency of once per second, and the eight drilling temperatures in the next 1 second were predicted in real time. The drilling test took a total of 1107 seconds and the drilling depth was 2 meters, which is similar to the drilling and sampling requirements of China's Chang'e 5 mission. The prediction effect of drilling temperature is as follows: Figure 7 As shown, the predicted drilling temperature R 2 The values ​​are all above 0.86, and the MAPE values ​​are all below 0.055. The drill bit temperature (temperature 5 to temperature 7) has the best prediction effect. 2 The value can reach above 0.9. These experimental results show that the model proposed in the present invention can provide accurate real-time prediction for the drilling test in the vacuum chamber simulating the lunar environment, and successfully verify the effectiveness and reliability of the method proposed in the present invention.

[0124] The same or similar reference numerals correspond to the same or similar components;

[0125] The terms used in the drawings to describe positional relationships are for illustrative purposes only and should not be construed as limiting this patent;

[0126] Obviously, the above embodiments of the present invention are merely examples for the purpose of clearly illustrating the present invention, and are not intended to limit the embodiments of the present invention. Those skilled in the art will appreciate that other variations or modifications can be made based on the above description. It is not necessary and impossible to enumerate all embodiments here. Any modifications, equivalent substitutions, and improvements made within the spirit and principles of the present invention shall be included within the scope of protection of the claims of the present invention.

Claims

1. A real-time prediction method for lunar soil drilling temperature based on spatiotemporal feature fusion, characterized in that: include: S1: Acquire lunar soil drilling sampling data, wherein the lunar soil drilling sampling data includes real-time drilling state parameters and drilling temperature data at multiple positions of the drilling tool; S2: Preprocessing the lunar soil drilling sampling data, wherein the preprocessing includes packaging the drilling state parameters at an earlier time and the drilling temperature data at a later time; S3: Build a model that introduces a self-attention mechanism and a Bi-LSTM network; S4: training the model using the lunar soil drilling sampling data pre-processed in step S2; S5: Optimize the model trained in step S4 to obtain an optimized model; S6: Inputting the previously acquired drilling state parameters into the optimized model to predict the drilling temperature data of the drilling tool in real time; Among them, S1 obtains lunar soil drilling sampling data, including the following steps: based on the simulated lunar soil environment system and drilling test bench, uses multiple sensors to collect 7 drilling state parameters and 8 drilling temperature data. The 7 drilling state parameters include drilling thrust, rotary torque, drilling depth, feed speed, rotary speed, feed motor current and rotary motor current; 8 types of drilling temperature data including 4 drill rod temperatures, drill bit rake temperature, drill bit inner surface temperature, drill bit end temperature and core barrel temperature; In S2, the lunar soil drilling sampling data is preprocessed, including the following steps: S2.1: Filter out the values ​​outside the preset range from the sampled data and treat them as abnormal data; S2.2: Process abnormal data; S2.3: interpolate the drilling temperature data after abnormal data processing; S2.4: Align the drilling state parameters and drilling temperature data, specifically: Among them, F t Represents the replacement value of the drilling status parameter at the tth second, N represents N drilling status parameters stored in the tth second, Represents the nth drilling status parameter within the tth second; S2.5: normalize the aligned drilling state parameters and drilling temperature data; S2.6: Finally, the 7 drilling status parameter data of the first 20 seconds after processing in steps S2.1 to S2.5 and the temperature data of 8 positions on the drilling tool in the next 1 second are packaged and then divided into training set, validation set and test set according to the ratio of 6:2:

2.

2. The method for real-time prediction of lunar soil drilling temperature based on spatiotemporal feature fusion according to claim 1 is characterized in that: In step S2.2, the abnormal data is processed as follows: Among them, Q t represents the replacement value of the outlier, Q t-1 represents the value of the abnormal value at the moment before, Q t+1 Represents the value one moment after the outlier.

3. The method for real-time prediction of lunar soil drilling temperature based on spatiotemporal feature fusion according to claim 1 is characterized in that: In step S2.3, interpolation processing is performed on the drilling temperature data after abnormal data processing. The interpolation processing is specifically as follows: Among them, t i+n -t i is the time difference between data point i+n and data point i, T i is the temperature at data point i, T i+1 is the temperature at data point i+1, T i+n -T i is the temperature difference between the two data points.

4. The method for real-time prediction of lunar soil drilling temperature based on spatiotemporal feature fusion according to claim 1 is characterized in that: In step S3, in the constructed model, the model includes: an input layer, a Bi-LSTM layer, a BN layer, an Attention layer, a Dropout layer, an FC layer, and an output layer, wherein: The input of the input layer is the preprocessed drilling state parameters. The Bi-LSTM layer, BN layer, Attention layer, Dropout layer, FC layer and output layer are connected in sequence after the input layer. The output layer outputs the predicted value of the drilling temperature data of the drill tool in the next 1 second.

5. The method for real-time prediction of lunar soil drilling temperature based on spatiotemporal feature fusion according to claim 1 is characterized in that: In step S5, the model trained in step S4 is optimized, and the optimization method of the model is specifically as follows: The model is optimized using the early stopping-Bayesian hyperparameter optimization method to obtain a hyperparameter combination, and the optimized model is obtained based on the hyperparameter combination.

6. The method for real-time prediction of lunar soil drilling temperature based on spatiotemporal feature fusion according to claim 5 is characterized in that: The model is optimized using the early stopping-Bayesian hyperparameter optimization method, specifically: S5.1: Define hyperparameter space; S5.2: Initialize the Bayesian optimizer; S5.3: The Bayesian optimizer obtained in step S5.2 recommends the next hyperparameter combination and continues training the model; S5.4: Use early stopping during each model training process to prevent model overfitting; S5.5: Repeat steps S5.3 and S5.4 until the optimizer converges and the optimal hyperparameter combination is found or the preset upper limit of hyperparameter searches is reached. S5.6: Output the current hyperparameter combination.

7. The method for real-time prediction of lunar soil drilling temperature based on spatiotemporal feature fusion according to claim 1 is characterized in that: In step S6, the optimized model is used to predict the drilling temperature data of the drilling tool in real time, specifically: The seven drilling status parameter data of the previous 20 seconds are collected in real time, and the drilling status parameter data are input into the model to predict the temperature of eight locations on the drilling tool in real time in the next 1 second.

Images