Slope parameter inversion method and device based on active learning
By training an ANN using an active learning method, and combining it with a non-dominated genetic algorithm and high-fidelity numerical simulation, the problem of large computational sample requirements in slope parameter inversion is solved, achieving fast and efficient slope parameter inversion and improving the timeliness and accuracy of slope stability analysis.
Patent Information
- Application Number
- CN202511651458.8
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2025-11-12
- Publication Date
- 2026-02-27
AI Technical Summary
Existing slope parameter inversion methods require a large number of calculation samples, resulting in excessively high simulation time costs and making it difficult to meet the timeliness and accuracy requirements of slope stability analysis.
An active learning-based approach is adopted, which trains an ANN by acquiring data from monitoring equipment, and combines a non-dominated genetic algorithm and high-fidelity numerical simulation to gradually optimize the parameter vector set, thereby achieving fast and efficient slope parameter inversion.
It achieves rapid, efficient and accurate slope parameter inversion, reduces training costs, and improves the efficiency and accuracy of slope parameter inversion.
Smart Images
Figure CN121580787A_ABST
Abstract
Description
Technical Field
[0001] This disclosure relates to the field of machine learning technology, and in particular to a method and apparatus for slope parameter inversion based on active learning. Background Technology
[0002] The accuracy of slope stability analysis depends on the accuracy of slope parameters, which refer to some geological parameters related to the slope.
[0003] Collecting geological parameters through field experiments is costly and time-consuming. Related technologies employ data-driven parameter inversion methods to invert slope parameters; however, under current conditions, most parameter inversions require a large number of computational samples, resulting in excessively high simulation time costs and making it difficult to meet the timeliness and accuracy requirements of slope stability analysis. Summary of the Invention
[0004] This disclosure provides a method and apparatus for slope parameter inversion based on active learning, which can achieve fast, efficient and accurate slope parameter inversion. The technical solution includes at least the following: Firstly, a slope parameter inversion method based on active learning is provided, comprising: acquiring first monitoring data collected by each monitoring device in a first monitoring group, wherein the first monitoring data is the actual slope displacement; in the i-th training round, training a first ANN using multiple first parameter vectors from a parameter vector set to obtain the predicted slope displacement corresponding to each first parameter vector, wherein the first ANN corresponds to the first monitoring group; using the parameter vector set as the initial population and the first monitoring data as the optimization target, performing multi-objective optimization using a non-dominated genetic algorithm to obtain a non-dominated solution set in the i-th training round, wherein the non-dominated solution set includes multiple second parameter vectors; performing numerical simulation on the multiple second parameter vectors using a high-fidelity numerical simulation method to obtain the reference slope displacement corresponding to each second parameter vector; storing the multiple second parameter vectors and the reference slope displacement corresponding to each second parameter vector in the parameter vector set to update the parameter vector set, wherein the updated parameter vector set is used to train the first ANN in the (i+1)-th training round; and selecting an optimal second parameter vector from the last training round of the first ANN as the parameter vector corresponding to the first monitoring data.
[0005] Optionally, in the initial training round, multiple first parameter vectors in the parameter vector set are randomly sampled from the parameter space.
[0006] Optionally, the training epoch that reaches the preset stopping iteration target is the last training epoch of the first ANN. The stopping iteration target includes: the number of parameter vectors in the parameter vector set reaches a number threshold, and the average absolute error between the reference slope displacement corresponding to the second parameter vector and the first monitoring data is less than an error threshold.
[0007] Optionally, selecting an optimal second parameter vector from the last training round of the first ANN as the parameter vector corresponding to the first monitoring data includes: calculating the average absolute error between the reference slope displacement and the first monitoring data corresponding to the plurality of second parameter vectors in the last training round of the first ANN, and selecting the second parameter vector with the smallest average absolute error as the optimal second parameter vector.
[0008] Optionally, after the first ANN is trained, if a new monitoring device is added to the first monitoring group, the trained first ANN is updated quickly using transfer learning. The transfer learning includes freezing the network weights of the first ANN except for the last two layers, and only fine-tuning the network weights of the last two layers of the first ANN.
[0009] Optionally, for any training round, the k-centers algorithm or the maximum-minimum distance algorithm is used to perform redundancy removal on the non-dominated solution set.
[0010] Secondly, a slope parameter inversion device based on active learning is also provided, comprising: an acquisition module for acquiring first monitoring data collected by each monitoring device in a first monitoring group, wherein the first monitoring data is the actual slope displacement; a training module for training a first ANN using multiple first parameter vectors from the parameter vector set in the i-th training round to obtain the predicted slope displacement corresponding to each first parameter vector, wherein the first ANN corresponds to the first monitoring group; and an optimization module for using the parameter vector set as the initial population, using the first monitoring data as the optimization target, and employing a non-dominated genetic algorithm for multi-objective optimization to obtain a set of non-dominated solutions in the i-th training round. The non-dominated solution set includes multiple second parameter vectors; a numerical simulation module is used to perform numerical simulation on the multiple second parameter vectors using a high-fidelity numerical simulation method to obtain the reference slope displacement corresponding to each second parameter vector; an update module is used to store the multiple second parameter vectors and the reference slope displacement corresponding to each second parameter vector into the parameter vector set to update the parameter vector set, and the updated parameter vector set is used to train the first ANN in the (i+1)th training round; a parameter inversion module is used to select an optimal second parameter vector from the last training round of the first ANN as the parameter vector corresponding to the first monitoring data.
[0011] Optionally, the training module is further configured to randomly sample multiple first parameter vectors from the parameter vector set in the parameter space during the initial training round.
[0012] Optionally, the training epoch that reaches the preset stopping iteration target is the last training epoch of the first ANN. In the parameter inversion module, the stopping iteration target includes: the number of parameter vectors in the parameter vector set reaches a number threshold, and the average absolute error between the reference slope displacement corresponding to the second parameter vector and the first monitoring data is less than an error threshold.
[0013] Optionally, the parameter inversion module is further configured to calculate the average absolute error between the reference slope displacement and the first monitoring data corresponding to the plurality of second parameter vectors in the last training round of the first ANN, and to take the second parameter vector with the smallest average absolute error as the optimal second parameter vector.
[0014] Optionally, the device further includes a transfer learning module, which is used to quickly update the trained first ANN if a new monitoring device is added to the first monitoring group after the first ANN has been trained. The transfer learning includes freezing the network weights of the first ANN except for the last two layers and only fine-tuning the network weights of the last two layers of the first ANN.
[0015] Optionally, the optimization module is further configured to perform redundancy removal processing on the non-dominated solution set using the k-centers algorithm or the maximum-minimum distance algorithm for any training round.
[0016] Thirdly, a computer device is also provided, comprising: a memory and a processor, wherein the memory stores at least one computer program, the at least one computer program being loaded and executed by the processor to perform the slope parameter inversion method based on active learning described in the above embodiments.
[0017] Fourthly, a computer-readable storage medium is also provided, wherein at least one computer program is stored in the computer-readable storage medium, the at least one computer program being loaded and executed by a processor to perform the slope parameter inversion method based on active learning described in the above embodiments.
[0018] Fifthly, a computer program product is provided, including a computer program / instructions that, when executed by a processor, implement the method described in the first aspect.
[0019] The beneficial effects of the technical solutions provided in this disclosure include at least the following: In this embodiment, the first monitoring data collected by each monitoring device in the first monitoring group is obtained, and the first monitoring data is the actual slope displacement. In the i-th training round, a first ANN is trained using multiple first parameter vectors from the parameter vector set to obtain the predicted slope displacement corresponding to each first parameter vector. The first ANN corresponds to the first monitoring group. Using the parameter vector set as the initial population and the first monitoring data as the optimization target, a non-dominated genetic algorithm is used for multi-objective optimization to obtain the non-dominated solution set in the i-th training round. The non-dominated solution set includes multiple second parameter vectors. A high-fidelity numerical simulation method is used to numerically simulate the multiple second parameter vectors to obtain the reference slope displacement corresponding to each second parameter vector. The multiple second parameter vectors and the reference slope displacement corresponding to each second parameter vector are stored in the parameter vector set to update the parameter vector set. The updated parameter vector set is used to train the first ANN in the (i+1)-th training round. An optimal second parameter vector is selected from the last training round of the first ANN as the parameter vector corresponding to the first monitoring data. Since the first monitoring data is used as the optimization target, the multiple second parameter vectors in the non-dominated solution set obtained by multi-objective optimization using the non-dominated genetic algorithm during the iteration process will gradually approach the target parameter vector (the correct inversion result of the first monitoring data). These second parameter vectors are added back to the parameter vector set, so that the expansion of the parameter vector set is in the correct direction (the direction of the target parameter vector). This is equivalent to adding effective and accurate data to the training set of the ANN, which can speed up the training efficiency of the ANN, thereby reducing the training cost of the ANN and enabling fast and efficient training of the ANN, thus achieving fast, efficient and accurate slope parameter inversion. Attached Figure Description
[0020] To more clearly illustrate the technical solutions in the embodiments of this disclosure, the accompanying drawings used in the description of the embodiments will be briefly introduced below. Obviously, the accompanying drawings described below are only some embodiments of this disclosure. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.
[0021] Figure 1 A flowchart illustrating an exemplary embodiment of the slope parameter inversion method based on active learning provided in this disclosure is shown. Figure 2 This is a schematic diagram illustrating the distribution of the second parameter vector in the parameter space as iteratively updated; Figure 3 A flowchart illustrating another exemplary embodiment of the present disclosure provides a slope parameter inversion method based on active learning; Figure 4A schematic diagram of the structure of a slope parameter inversion device based on active learning provided in an exemplary embodiment of this disclosure is shown. Figure 5 This is a schematic diagram of the structure of a computer device provided in an embodiment of this disclosure. Detailed Implementation
[0022] Unless otherwise defined, the technical or scientific terms used herein shall have the ordinary meaning as understood by one of ordinary skill in the art to which this disclosure pertains. The terms “first,” “second,” “third,” and similar terms used in this patent application specification and claims do not indicate any order, quantity, or importance, but are merely used to distinguish different components. Similarly, the terms “an” or “a” and similar terms do not indicate a quantity limitation, but rather indicate the presence of at least one. The terms “comprising” or “including” and similar terms mean that the elements or objects preceding “comprising” or “including” encompass the elements or objects listed following “comprising” or “including” and their equivalents, but do not exclude other elements or objects.
[0023] To make the objectives, technical solutions, and advantages of this disclosure clearer, the embodiments of this disclosure will be described in further detail below with reference to the accompanying drawings.
[0024] Figure 1 A flowchart illustrating an exemplary embodiment of the present disclosure provides a slope parameter inversion method based on active learning, which can be executed by a computer device. See also Figure 1 The method includes: In step 101, multiple monitoring devices are installed on the slope and divided into multiple monitoring groups.
[0025] For example, monitoring groups can be divided according to the installation batch of the monitoring equipment or according to the displacement change characteristics at the monitoring points. These displacement change characteristics could be, for example, average rate, peak rate, or trend slope.
[0026] Here, when dividing the monitoring groups according to the installation batch of the monitoring equipment, monitoring equipment installed in the same batch is grouped into the same monitoring group.
[0027] When dividing monitoring groups according to displacement change characteristics, the displacement change characteristics of each monitoring point are first calculated based on the monitoring data collected by each monitoring device. Then, these displacement change characteristics are divided into multiple intervals, and monitoring devices belonging to the same interval are in the same monitoring group.
[0028] This disclosure describes the process of inverting geological parameters for each monitoring group for the first time, that is, each monitoring group in this disclosure performs subsequent steps 102 to 107 for the first time.
[0029] In step 102, the first monitoring data collected by each monitoring device in the first monitoring group is obtained.
[0030] The first monitoring data is the actual slope displacement.
[0031] Here, the first monitoring group includes multiple monitoring devices deployed on a real slope that requires geological parameter inversion. For example, the slope is a rockfill dam.
[0032] These monitoring devices are used to collect actual slope displacements (i.e., the initial monitoring data) of real slopes. Examples of such monitoring devices include, but are not limited to, surface displacement gauges, multi-point displacement gauges, and interferometric synthetic aperture radar (InSAR).
[0033] Optionally, the acquired first monitoring data also needs to be preprocessed. Preprocessing of the first monitoring data includes, but is not limited to, denoising, outlier removal, and time alignment. The types of denoising include, but are not limited to, moving average filtering, wavelet denoising, and empirical mode decomposition.
[0034] In step 103, in the i-th training round, the first ANN is trained using multiple first parameter vectors from the parameter vector set to obtain the predicted slope displacement corresponding to each first parameter vector.
[0035] Different monitoring groups correspond to different ANNs; for example, the first ANN corresponds to the first monitoring group. The parameter vector contains various geological parameters of the actual slope that need to be inverted. The dimension of the parameter vector depends on the number and types of geological parameters that need to be inverted; that is, each dimension of the parameter vector represents a geological parameter that needs to be inverted.
[0036] In this embodiment of the disclosure, the input to the ANN is a parameter vector, and the output of the ANN is the predicted slope displacement corresponding to the parameter vector. During the training process of the ANN, it gradually learns the nonlinear relationship between the parameter vector and the slope displacement.
[0037] An Artificial Neural Network (ANN) is a computer model that can automatically learn and extract inherent patterns from complex, non-linear data, and use it for prediction, classification, recognition, or generation of new data. There are many types of ANNs, such as Feedforward Neural Networks (FNN), Convolutional Neural Networks (CNN), and Generative Adversarial Networks (GAN). This disclosure does not limit the type of the first ANN.
[0038] During the training process of an ANN, the ANN can learn the nonlinear relationship between the parameter vector and the slope displacement. In other words, the trained ANN can accurately predict the slope displacement based on the input parameter vector.
[0039] Steps 103 to 106 describe the training process of the first ANN in the i-th training epoch. It can be seen that during training, the parameter vectors in the parameter vector set are continuously expanded and increased. The parameter vector set used in the (i+1)-th training epoch is obtained after expansion in the i-th training epoch. This process continues sequentially: the parameter vector set used in the i-th training epoch is obtained after expansion in the (i-1)-th training epoch… the parameter vector set used in the second training epoch is obtained after expansion in the first training epoch; and the parameter vector set used in the first training epoch (the initial training epoch) is obtained through random sampling.
[0040] Optionally, in the initial training rounds, multiple first parameter vectors are randomly sampled from the parameter space to obtain the parameter vector set. Here, the parameter space is a multi-dimensional space, where each dimension of the parameter space corresponds to a dimension of the parameter vector. The upper and lower limits of each dimension in the parameter space are preset based on experience, and this embodiment does not limit this.
[0041] For example, a Latin hypercube sampling method is used to randomly sample from the parameter space. When training the first ANN in the initial training rounds, in addition to obtaining multiple first parameter vectors in the parameter vector set, it is also necessary to obtain the reference slope displacement corresponding to each first parameter vector in the parameter vector set. The slope displacement corresponding to each first vector in the parameter vector set can be obtained using a high-fidelity numerical simulation method. For example, high-fidelity numerical simulation methods include the finite element method, the finite difference method, and the material point method.
[0042] Here, the predicted slope displacement is the slope displacement output by the ANN. Before the ANN is fully trained, the predicted slope displacement output by the ANN may not be accurate. Therefore, high-fidelity numerical simulation methods are needed to obtain the accurate slope displacement (i.e., reference slope displacement) corresponding to each parameter vector. During the training of the ANN, the error between the predicted slope displacement output by the ANN and the reference slope displacement is minimized, thereby enabling the ANN to learn the nonlinear relationship between the parameter vector and the slope displacement.
[0043] As can be seen from subsequent steps 105 to 106, in each iteration, the newly added parameter vectors in the parameter vector set all have corresponding reference slope displacements. Therefore, no matter how the parameter vector set is updated, each parameter vector in the parameter vector set contains the corresponding reference slope displacement.
[0044] When training an ANN using a set of parameter vectors, multiple first parameter vectors from the set of parameter vectors are input into the first ANN to obtain the predicted slope displacement corresponding to each first parameter vector output by the first ANN. Then, in this round, the error between the predicted slope displacement corresponding to each first parameter vector output by the first ANN and the reference slope displacement corresponding to each first parameter vector is minimized, thereby enabling the first ANN to learn the nonlinear relationship between the parameter vector and the slope displacement.
[0045] In step 104, the parameter vector set is used as the initial population, the first monitoring data is used as the optimization target, and a non-dominated genetic algorithm is used to perform multi-objective optimization to obtain the non-dominated solution set in the i-th training round.
[0046] The set of non-dominated solutions includes multiple second parameter vectors.
[0047] For example, the non-dominated genetic algorithm is the NSGA-III (Non-dominated Sorting Genetic Algorithm III) algorithm, and the non-dominated solution set is also the Pareto solution set.
[0048] Since solutions in the non-dominated solution set may be duplicated, redundancy removal is necessary. For example, for any training epoch, the k-centers algorithm or the maximum-minimum distance algorithm can be used to remove redundancy from the non-dominated solution set.
[0049] In this embodiment of the present disclosure, when the parameter vector set is used as the initial population, the initial population not only includes multiple first parameter vectors in the parameter vector set, but also includes the reference slope displacement corresponding to each first parameter vector.
[0050] When using a non-dominated genetic algorithm for multi-objective optimization, the multi-objective here includes the first monitoring data. This means that the first monitoring data is one of the objectives of the non-dominated genetic algorithm, and the other optimization objectives can be set as needed.
[0051] If the parameter vector derived from the first monitoring data is called the target parameter vector, the goal of slope parameter inversion is to obtain an accurate target parameter vector. In this embodiment, through multi-objective optimization, a parameter vector (i.e., the second parameter vector) that is relatively close to the target parameter vector can be gradually found.
[0052] If the goal is simply to train an ANN to learn the nonlinear relationship between the parameter vector and the slope displacement, it often requires generating a large amount of data randomly and training the ANN multiple times with that data. This results in high training costs for the ANN, and the ANN trained in this way can only generate slope displacement based on the parameter vector, but cannot perform slope parameter inversion.
[0053] In this embodiment of the disclosure, a parameter vector set is used as the initial population, and multi-objective optimization is performed with the first monitoring data as the target. This allows the search for a second parameter vector in the parameter space that is close to the target parameter vector. Through continuous iteration and updating in training rounds, the second parameter vector found in each training round becomes closer to the target parameter vector, thereby effectively realizing slope parameter inversion.
[0054] Furthermore, through steps 105 to 106, the second parameter vector found each time and the reference slope displacement corresponding to each second parameter vector can be stored in the parameter vector set, thereby expanding the parameter vector set (equivalent to the training set of ANN).
[0055] In related technologies, when a large amount of randomly generated data is used to train an ANN, much of this randomly generated data is actually useless, which increases the training cost and reduces the training efficiency of the ANN. In the embodiments of this disclosure, the expansion of the parameter vector set is carried out in the correct direction (the direction of the target parameter vector), which is equivalent to adding effective and accurate data to the training set of the ANN. Therefore, the training efficiency of the ANN can be accelerated, thereby reducing the training cost of the ANN and achieving fast, efficient and accurate training of the ANN.
[0056] In step 105, a high-fidelity numerical simulation method is used to numerically simulate multiple second parameter vectors to obtain the reference slope displacement corresponding to each second parameter vector.
[0057] In step 106, multiple second parameter vectors and the reference slope displacement corresponding to each second parameter vector are stored in the parameter vector set to update the parameter vector set.
[0058] The updated parameter vector set is used to train the first ANN in the (i+1)th training epoch.
[0059] After storing multiple second parameter vectors and the corresponding reference slope displacements of each second parameter vector into a parameter vector set, these multiple second parameter vectors are actually transformed into the first parameter vector in the parameter vector set.
[0060] For example, if there are a first parameter vectors in the parameter vector set in the i-th round, and b second parameter vectors are found in the i-th round, then after storing these b second parameter vectors in the parameter vector set, there will be a+b first parameter vectors in the parameter vector set in the (i+1)-th round.
[0061] Here, when training the first ANN using the updated parameter vector set in the (i+1)th training epoch, since all the original first parameter vectors (a first parameter vectors) in the initial state of the parameter vector set in the i-th training epoch have already participated in the training of the first ANN, there is no need to repeat the training of the first ANN. Therefore, it is only necessary to fine-tune the first ANN in the (i+1)th training epoch using the newly added part (b second parameter vectors) in the parameter vector set. There are many related technologies regarding the implementation of fine-tuning, which will not be detailed here.
[0062] During the iterative training process described above, the parameter vector set is continuously expanded, which is equivalent to active learning. Figure 2 This is a schematic diagram illustrating the distribution of the second parameter vector in the parameter space as iteratively updated. For example... Figure 2 As shown, the circular dots represent multiple first parameter vectors randomly sampled during the initial training rounds in the parameter space, i.e., the initial samples. The triangular dots represent the second parameter vectors subsequently added to the parameter vector set. The dashed part is the Pareto front. It can be seen that as active learning progresses, the newly expanded samples (second parameter vectors) gradually concentrate on the Pareto front.
[0063] By continuously executing steps 102 to 106, the first ANN can be trained iteratively multiple times until a preset stopping iteration target is reached, at which point training will stop. The training epoch at which the preset stopping iteration target is reached is the last training epoch of the first ANN.
[0064] Optionally, the stopping iteration objectives include: the number of parameter vectors in the parameter vector set reaches a number threshold, and the average absolute error between the reference slope displacement corresponding to the second parameter vector and the first monitoring data is less than an error threshold.
[0065] The expansion of the parameter vector set in each iteration is finite (depending on the number of second parameter vectors in the non-dominated solution set after redundancy removal). Therefore, the number of parameter vectors in the parameter vector set can reflect the iteration round. If the number of parameter vectors in the parameter vector set reaches the threshold, it means that enough iterations have been performed, and the iteration can be terminated at this time.
[0066] If, in a certain training round, the average absolute error between the reference slope displacement corresponding to a certain second parameter vector and the first monitoring data is less than the error threshold, it means that the second parameter vector can be approximately equal to the target parameter vector, which is equivalent to completing the slope geological parameter inversion of the first monitoring data, and therefore the iteration can be terminated.
[0067] In step 107, an optimal second parameter vector is selected from the last training round of the first ANN as the parameter vector corresponding to the first monitoring data.
[0068] Once the preset stopping iteration target is reached, it indicates that the first ANN has been trained. At this point, iteration can be stopped. When the first ANN is trained, the multiple second parameter vectors in the last training round of the first ANN will be close enough to the target parameter vector. Therefore, at this point, an optimal second parameter vector can be selected from the multiple second parameter vectors in the last training round of the first ANN as the parameter vector corresponding to the first monitoring data. In other words, there is an optimal second parameter vector in the last training round of the first ANN, and this second parameter vector is the target parameter vector.
[0069] Optionally, the following method can be used to select an optimal second parameter vector from multiple second parameter vectors in the last training round of the first ANN: calculate the average absolute error between the reference slope displacement and the first monitoring data corresponding to multiple second parameter vectors in the last training round of the first ANN, and obtain the average absolute error corresponding to each second parameter vector; take the second parameter vector with the smallest average absolute error as the optimal second parameter vector, and thus obtain the target parameter vector.
[0070] This target parameter vector can be used to upload to the digital twin model of the slope, enabling stability analysis and safety assessment of the slope where the first monitoring group is located.
[0071] Furthermore, the first ANN at this point has fully learned the mapping relationship between the geological parameters (parameter vectors) and slope displacement in the first monitoring group. Therefore, the first ANN can also play a role in the stability analysis and safety assessment of the slope where the first monitoring group is located (e.g., coupling with the digital twin model).
[0072] In this embodiment, the first monitoring data collected by each monitoring device in the first monitoring group is obtained, and the first monitoring data is the actual slope displacement. In the i-th training round, a first ANN is trained using multiple first parameter vectors from the parameter vector set to obtain the predicted slope displacement corresponding to each first parameter vector. The first ANN corresponds to the first monitoring group. Using the parameter vector set as the initial population and the first monitoring data as the optimization target, a non-dominated genetic algorithm is used for multi-objective optimization to obtain the non-dominated solution set in the i-th training round. The non-dominated solution set includes multiple second parameter vectors. A high-fidelity numerical simulation method is used to numerically simulate the multiple second parameter vectors to obtain the reference slope displacement corresponding to each second parameter vector. The multiple second parameter vectors and the reference slope displacement corresponding to each second parameter vector are stored in the parameter vector set to update the parameter vector set. The updated parameter vector set is used to train the first ANN in the (i+1)-th training round. An optimal second parameter vector is selected from the last training round of the first ANN as the parameter vector corresponding to the first monitoring data. Since the first monitoring data is used as the optimization target, the multiple second parameter vectors in the non-dominated solution set obtained by multi-objective optimization using the non-dominated genetic algorithm during the iteration process will gradually approach the target parameter vector (the correct inversion result of the first monitoring data). These second parameter vectors are added back to the parameter vector set, so that the expansion of the parameter vector set is in the correct direction (the direction of the target parameter vector). This is equivalent to adding effective and accurate data to the training set of the ANN, which can speed up the training efficiency of the ANN, thereby reducing the training cost of the ANN and enabling fast and efficient training of the ANN, thus achieving fast, efficient and accurate slope parameter inversion.
[0073] For large-scale slopes, the monitoring equipment is not deployed all at once, but in batches, with new monitoring equipment being installed on the slope.
[0074] In step 101, when the monitoring groups are divided according to the installation batch of the monitoring equipment, it is only necessary to treat the subsequently deployed monitoring equipment as the new monitoring group and perform steps 102 to 107 on the new monitoring group. Details are omitted here.
[0075] In the case where monitoring groups are divided according to displacement change characteristics in step 101, no new monitoring groups will be created for subsequently deployed monitoring equipment. Newly deployed monitoring equipment will be assigned to existing ANN monitoring groups based on displacement characteristics. In this situation, for existing ANN monitoring groups with newly added monitoring equipment, the ANN still needs to be migrated and updated. The following section will combine... Figure 3The steps for migrating and updating the first ANN after the first ANN corresponding to the first monitoring group has been trained and monitoring devices have been added to the first monitoring group are explained.
[0076] Figure 3 A flowchart illustrating an exemplary embodiment of the present disclosure provides a slope parameter inversion method based on active learning, which can be executed by a computer device. See also Figure 3 The method includes: In step 301, the second monitoring data collected by the newly added monitoring device in the first monitoring group is obtained.
[0077] In step 302, the first ANN corresponding to the first monitoring group is obtained.
[0078] The first ANN is obtained using the methods described in steps 101 to 107 above. In other words, the first ANN here is an ANN trained on the basis of the original first monitoring group using the first monitoring data.
[0079] In step 303, the first ANN is transferred to the second monitoring data.
[0080] Transfer learning involves freezing the network weights of all layers of the first ANN except for the last two layers, and only fine-tuning the network weights of the last two layers of the first ANN.
[0081] Here, the implementation process of step 303 is basically the same as that of steps 103 to 107. When implementing step 303, it is only necessary to replace the first monitoring data in steps 103 to 107 with the second monitoring data, replace the first ANN in steps 103 to 107 with the first ANN that has already been trained in step 302, and freeze the network weights of the first ANN except for the last two layers when training the first ANN, and only fine-tune the network weights of the last two layers of the first ANN. Detailed descriptions are omitted here.
[0082] By transferring learning to the newly added monitoring equipment, the training efficiency of ANN can be further improved, thereby increasing the efficiency of slope parameter inversion.
[0083] The following are device embodiments of this application. For details not described in detail in the device embodiments, please refer to the above method embodiments.
[0084] Figure 4 A schematic diagram of an active learning-based slope parameter inversion device provided in an exemplary embodiment of this disclosure is shown. See also Figure 4The slope parameter inversion device 400 based on active learning includes: an acquisition module 401, a training module 402, an optimization module 403, a numerical simulation module 404, an update module 405, and a parameter inversion module 406.
[0085] The acquisition module 401 is used to acquire the first monitoring data collected by each monitoring device in the first monitoring group. The first monitoring data is the actual slope displacement. Training module 402 is used to train the first ANN in the i-th training round using multiple first parameter vectors in the parameter vector set, and obtain the predicted slope displacement corresponding to each first parameter vector. The first ANN corresponds to the first monitoring group. The optimization module 403 is used to use the parameter vector set as the initial population, the first monitoring data as the optimization target, and a non-dominated genetic algorithm to perform multi-objective optimization, so as to obtain the non-dominated solution set in the i-th training round. The non-dominated solution set includes multiple second parameter vectors. The numerical simulation module 404 is used to perform numerical simulation on multiple second parameter vectors using a high-fidelity numerical simulation method to obtain the reference slope displacement corresponding to each second parameter vector. The update module 405 is used to store multiple second parameter vectors and the reference slope displacement corresponding to each second parameter vector into the parameter vector set to update the parameter vector set. The updated parameter vector set is used to train the first ANN in the (i+1)th training round. The parameter inversion module 406 is used to select an optimal second parameter vector from the last training round of the first ANN as the parameter vector corresponding to the first monitoring data.
[0086] Optionally, the training module 402 is also used to randomly sample multiple first parameter vectors from the parameter vector set in the parameter space during the initial training rounds.
[0087] Optionally, the training round at which the preset stopping iteration target is reached is the last training round of the first ANN. In the parameter inversion module 406, the stopping iteration target includes: the number of parameter vectors in the parameter vector set reaches the number threshold, and the average absolute error between the reference slope displacement corresponding to the second parameter vector and the first monitoring data is less than the error threshold.
[0088] Optionally, the parameter inversion module 406 is also used to calculate the average absolute error between the reference slope displacement and the first monitoring data corresponding to multiple second parameter vectors in the last training round of the first ANN, and to take the second parameter vector with the smallest average absolute error as the optimal second parameter vector.
[0089] Optionally, the device further includes a transfer learning module 407, which is used to quickly update the trained first ANN if a new monitoring device is added in the first monitoring group after the first ANN has been trained. The transfer learning includes freezing the network weights of the first ANN except for the last two layers and only fine-tuning the network weights of the last two layers of the first ANN.
[0090] Optionally, the optimization module 403 is also used to perform redundancy removal on the non-dominated solution set for any training round using the k-centers algorithm or the maximum-minimum distance algorithm.
[0091] It should be noted that the slope parameter inversion device based on active learning provided in the above embodiments is only illustrated by the division of the functional modules described above. In practical applications, the functions can be assigned to different functional modules as needed, that is, the internal structure of the device can be divided into different functional modules to complete all or part of the functions described above. Furthermore, the slope parameter inversion device based on active learning provided in the above embodiments and the slope parameter inversion method embodiments based on active learning belong to the same concept, and their specific implementation process is detailed in the method embodiments, which will not be repeated here.
[0092] The module division in this embodiment is illustrative and represents only one logical functional division. In actual implementation, other division methods are possible. Furthermore, the functional modules in the various embodiments of this disclosure can be integrated into a single processor, exist as separate physical entities, or be integrated into a single module. The integrated modules described above can be implemented in hardware or as software functional modules.
[0093] If the integrated module is implemented as a software functional module and sold or used as an independent product, it can be stored in a computer-readable storage medium. Based on this understanding, the technical solution of this disclosure, in essence, or the part that contributes to the prior art, or all or part of the technical solution, can be embodied in the form of a software product. This computer software product is stored in a storage medium and includes several instructions to cause a terminal device (which may be a personal computer, mobile phone, or communication device, etc.) or processor to execute all or part of the steps of the methods of the various embodiments of this disclosure. The aforementioned storage medium includes various media capable of storing program code, such as USB flash drives, portable hard drives, read-only memory (ROM), random access memory (RAM), magnetic disks, or optical disks.
[0094] Figure 5This is a schematic diagram of the structure of a computer device provided in an embodiment of this disclosure. For example... Figure 5 As shown, the computer device 500 includes a processor 501 and a memory 502.
[0095] Processor 501 may include one or more processing cores, such as a quad-core processor, an octa-core processor, etc. Processor 501 may be implemented using at least one hardware form selected from DSP (Digital Signal Processing), FPGA (Field-Programmable Gate Array), and PLA (Programmable Logic Array). Processor 501 may also include a main processor and a coprocessor. The main processor, also known as a CPU (Central Processing Unit), is used to process data in the wake-up state; the coprocessor is a low-power processor used to process data in the standby state. In some embodiments, processor 501 may integrate a GPU (Graphics Processing Unit), which is responsible for rendering and drawing the content to be displayed on the screen. In some embodiments, processor 501 may also include an AI (Artificial Intelligence) processor, which is used to handle computational operations related to machine learning.
[0096] The memory 502 may include one or more computer-readable storage media, which may be non-transitory. The memory 502 may also include high-speed random access memory and non-volatile memory, such as one or more disk storage devices or flash memory devices. In some embodiments, the non-transitory computer-readable storage media in the memory 502 is used to store at least one instruction, which is executed by the processor 501 to implement the active learning-based slope parameter inversion method provided in this disclosure embodiment.
[0097] Those skilled in the art will understand that Figure 5 The structure shown does not constitute a limitation on the computer device 500, and may include more or fewer components than shown, or combine certain components, or use different component arrangements.
[0098] This disclosure also provides a non-transitory computer-readable storage medium, wherein when the instructions in the storage medium are executed by the processor of a computer device, the computer device is able to execute the active learning-based slope parameter inversion method provided in this disclosure.
[0099] This disclosure also provides a computer program product, including a computer program / instruction, which, when executed by a processor, implements the active learning-based slope parameter inversion method provided in this disclosure.
[0100] The above description is merely an optional embodiment of this disclosure and is not intended to limit this disclosure. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of this disclosure should be included within the protection scope of this disclosure.
Claims
1. A slope parameter inversion method based on active learning, characterized in that, The method includes: Acquire the first monitoring data collected by each monitoring device in the first monitoring group, where the first monitoring data is the actual slope displacement; In the i-th training round, the first ANN is trained using multiple first parameter vectors from the parameter vector set to obtain the predicted slope displacement corresponding to each first parameter vector. The first ANN corresponds to the first monitoring group. Using the parameter vector set as the initial population and the first monitoring data as the optimization target, a non-dominated genetic algorithm is used to perform multi-objective optimization to obtain the non-dominated solution set in the i-th training round. The non-dominated solution set includes multiple second parameter vectors. A high-fidelity numerical simulation method is used to numerically simulate the multiple second parameter vectors to obtain the reference slope displacement corresponding to each second parameter vector. The plurality of second parameter vectors and the reference slope displacement corresponding to each second parameter vector are stored in the parameter vector set to update the parameter vector set. The updated parameter vector set is used to train the first ANN in the (i+1)th training round. Select an optimal second parameter vector from the last training round of the first ANN as the parameter vector corresponding to the first monitoring data.
2. The method according to claim 1, characterized in that, In the initial training round, multiple first parameter vectors are randomly sampled from the parameter space to obtain the parameter vector set.
3. The method according to claim 1, characterized in that, The training epoch that reaches the preset stopping iteration target is the last training epoch of the first ANN. The stopping iteration target includes: the number of parameter vectors in the parameter vector set reaches a number threshold, and the average absolute error between the reference slope displacement corresponding to the second parameter vector and the first monitoring data is less than an error threshold.
4. The method according to claim 3, characterized in that, The step of selecting an optimal second parameter vector from the last training round of the first ANN as the parameter vector corresponding to the first monitoring data includes: calculating the average absolute error between the reference slope displacement and the first monitoring data corresponding to the plurality of second parameter vectors in the last training round of the first ANN, and selecting the second parameter vector with the smallest average absolute error as the optimal second parameter vector.
5. The method according to claim 1, characterized in that, After the first ANN is trained, if a new monitoring device is added to the first monitoring group, the trained first ANN is updated quickly using transfer learning. The transfer learning includes freezing the network weights of the first ANN except for the last two layers, and only fine-tuning the network weights of the last two layers of the first ANN.
6. The method according to claim 1, characterized in that, The method further includes: for any training round, using the k-centers algorithm or the maximum-minimum distance algorithm to perform redundancy removal processing on the non-dominated solution set.
7. A slope parameter inversion device based on active learning, characterized in that, The device includes: The acquisition module is used to acquire the first monitoring data collected by each monitoring device in the first monitoring group, wherein the first monitoring data is the actual slope displacement; The training module is used to train the first ANN using multiple first parameter vectors in the parameter vector set in the i-th training round, so as to obtain the predicted slope displacement corresponding to each first parameter vector. The first ANN corresponds to the first monitoring group. The optimization module is used to use the parameter vector set as the initial population, the first monitoring data as the optimization target, and a non-dominated genetic algorithm to perform multi-objective optimization to obtain the non-dominated solution set in the i-th training round. The non-dominated solution set includes multiple second parameter vectors. The numerical simulation module is used to perform numerical simulation on the plurality of second parameter vectors using a high-fidelity numerical simulation method to obtain the reference slope displacement corresponding to each second parameter vector. An update module is used to store the plurality of second parameter vectors and the reference slope displacement corresponding to each second parameter vector into the parameter vector set to update the parameter vector set. The updated parameter vector set is used to train the first ANN in the (i+1)th training round. The parameter inversion module is used to select an optimal second parameter vector from the last training round of the first ANN as the parameter vector corresponding to the first monitoring data.
8. A computer device, characterized in that, The computer device includes a memory and a processor, wherein the memory stores at least one computer program, which is loaded and executed by the processor to implement the method according to any one of claims 1 to 6.
9. A computer-readable storage medium, characterized in that, The computer-readable storage medium stores at least one computer program, which is loaded and executed by a processor to implement the method according to any one of claims 1 to 6.
10. A computer program product comprising a computer program / instructions, characterized in that, When the computer program / instructions are executed by the processor, they implement the method described in any one of claims 1 to 6.