A deep learning-based worst-case scenario recognition method and system

Abstract

The application provides a deep learning-based most unfavorable working condition identification method and system, relates to the technical field of engineering structure test, and comprises the following steps: performing an initial test to obtain data, training a prediction model; generating multiple groups of to-be-tested working conditions by using a conditional generative adversarial network, inputting each group of to-be-tested working conditions into the prediction model, and outputting a predicted comprehensive risk index and corresponding mean and variance; constructing working condition interaction terms, and selecting target interaction terms through a random forest model; calculating the expected improvement of each group of to-be-tested working conditions based on the mean and variance of the predicted comprehensive risk index of the to-be-tested working conditions and the target interaction terms, selecting the to-be-tested working conditions for impact test, updating the conditional generative adversarial network and the prediction model based on the test results, and iterating until convergence to obtain the most unfavorable working condition. According to the scheme, high-risk working conditions are generated by using the conditional generative adversarial network, the working conditions are screened in combination with the prediction model, the most unfavorable working condition is quickly approached, the number of tests is significantly reduced, and the research and development cost and period are greatly reduced.

Description

Technical Field

[0001] This invention relates to the field of engineering structure testing technology, specifically a method and system for identifying the most unfavorable working conditions based on deep learning. Background Technology

[0002] The response of shield tunnel segments to unconventional impacts (such as train derailment and falling object impacts) involves multi-body contact, soil-structure coupling, and high-frequency transients. Experimentally, variable-angle impact devices, soil chambers, and multi-channel data acquisition are commonly used to achieve impact tests at different angles and locations. To save costs, the overall risk is often extrapolated from a small number of representative angle / location tests. Data collection typically involves individual channels such as steel reinforcement strain, concrete surface strain / cracks, or earth pressure, with the results relying on peak values ​​or selected characteristics as proxies for "severity."

[0003] In train-segment-soil impact tests, the working parameters (impact angle θ, impact position p, velocity v, mass m, projectile attitude ψ, whether it is located at a joint, etc.) are highly dimensional and have many combinations. Traditional methods rely on manual experience or equal-interval scanning. In order to find the most unfavorable working condition, multiple tests need to be repeated. Each test uses disposable projectiles and segments. This is not only costly and inefficient, but also cannot guarantee coverage of the most unfavorable working condition.

[0004] Therefore, how to automatically, quickly, and verifiably find the most unfavorable combination of working conditions for the segments, reinforcing bars, and the soil behind them under a limited experimental budget is a technical problem that urgently needs to be solved in this field. Summary of the Invention

[0005] The purpose of this invention is to provide a method and system for identifying the most unfavorable working conditions based on deep learning, so as to improve the above-mentioned problems. To achieve the above objective, the technical solution adopted by this invention is as follows:

[0006] Firstly, this application provides a deep learning-based method for identifying the most unfavorable working conditions, including:

[0007] Initial impact data is collected, which is obtained after an impact test is conducted using preset initial conditions.

[0008] Multiple sets of test conditions are generated using a conditional generative adversarial network. Each set of test conditions is input into a preset prediction model, which outputs a comprehensive risk index and the corresponding mean and variance. The prediction model is trained based on the initial test conditions and initial impact data.

[0009] Construct working condition interaction items and filter them using a random forest model to obtain target interaction items;

[0010] Based on the mean and variance of the comprehensive risk index of the test conditions and the target interaction term, the expected improvement amount of each test condition is calculated.

[0011] The test conditions are selected based on the expected improvement amount for impact testing. The conditions are updated based on the test results to generate an adversarial network and a prediction model. The process is iterated until convergence to obtain the test condition with the largest comprehensive risk index.

[0012] Secondly, this application also provides a deep learning-based worst-case scenario identification system, comprising:

[0013] The first module is used to collect initial impact data, which is obtained after an impact test is conducted using preset initial conditions.

[0014] The second module is used to generate multiple sets of test conditions using a conditional generative adversarial network, input each set of test conditions into a preset prediction model, and output a comprehensive prediction risk index and the corresponding mean and variance; the prediction model is trained based on the initial test conditions and initial impact data.

[0015] The third module is used to construct working condition interaction items and filter them using a random forest model to obtain target interaction items;

[0016] The fourth module is used to predict the mean and variance of the comprehensive risk index and the target interaction term based on the test conditions, and to calculate the expected improvement amount for each group of test conditions.

[0017] The fifth module is used to select the test conditions for impact testing based on the expected improvement amount, update the conditions based on the test results to generate an adversarial network and a prediction model, and iterate until convergence to obtain the test condition with the largest comprehensive risk index.

[0018] Beneficial effects:

[0019] This method trains a conditional generative adversarial network using existing experimental data to generate potential high-risk operating conditions. By combining a random forest model with a prediction model, it comprehensively utilizes prediction uncertainty (variance) and parameter interaction effects to enhance the ability to identify high-risk operating conditions. Through an active learning mechanism, iterative optimization improves the comprehensiveness and accuracy of the operating condition search, thereby quickly approximating the most unfavorable operating condition. This method significantly reduces the number of physical impact tests, substantially lowering R&D costs and timelines.

[0020] Other features and advantages of the invention will be set forth in the description which follows, and will be apparent in part from the description, or may be learned by practicing embodiments of the invention. Attached Figure Description

[0021] To more clearly illustrate the technical solutions of the embodiments of the present invention, the accompanying drawings used in the embodiments will be briefly introduced below. It should be understood that the following drawings only show some embodiments of the present invention and should not be regarded as a limitation on the scope. For those skilled in the art, other related drawings can be obtained based on these drawings without creative effort.

[0022] Figure 1 This is a flowchart of the deep learning-based worst-case identification method in the embodiment;

[0023] Figure 2 This is a schematic diagram of the experimental device structure in the embodiment;

[0024] Figure 3 This is a schematic diagram of the installation of the tunnel lining segments and the soil box in the embodiment;

[0025] Figure 4 This is a schematic diagram of the soil box structure in the embodiment;

[0026] Figure 5 This is a schematic diagram of the structure of a deep learning-based worst-case identification device for an example.

[0027] Symbol Explanation: 1-Impact side rectangular flange; 2-Semi-circular flange; 3-Triangular steel plate; 4-Square rectangular tube; 5-Rectangular flange; 6-First bolt hole; 7-Second bolt hole; 8-Third bolt hole; 9-Soil box; 10-Tube segment; 800-Deep learning-based worst-case identification device; 801-Processor; 802-Memory; 803-Multimedia component; 804-I / O interface; 805-Communication component. Detailed Implementation

[0028] To make the objectives, technical solutions, and advantages of the embodiments of the present invention clearer, the technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some, not all, of the embodiments of the present invention. The components of the embodiments of the present invention described and shown in the accompanying drawings can generally be arranged and designed in various different configurations. Therefore, the following detailed description of the embodiments of the present invention provided in the accompanying drawings is not intended to limit the scope of the claimed invention, but merely to illustrate selected embodiments of the invention. All other embodiments obtained by those skilled in the art based on the embodiments of the present invention without inventive effort are within the scope of protection of the present invention.

[0029] It should be noted that similar labels and letters in the following figures indicate similar items. Therefore, once an item is defined in one figure, it does not need to be further defined and explained in subsequent figures.

[0030] Example 1:

[0031] See Figure 1 This embodiment provides a method for identifying the most unfavorable working conditions based on deep learning, including steps S100, S200, S300, S400 and S500;

[0032] S100. Collect initial impact data, which is obtained after conducting an impact test using preset initial conditions;

[0033] The impact test was conducted on the tunnel segments with variable angles using a simulated train projectile, with a soil box placed on the back side of the tunnel segments.

[0034] The test monitoring system includes a steel bar strain measurement unit (resistance strain gauge), an earth pressure measurement unit (earth pressure cell), an acceleration measurement unit (IEPE or charge-type accelerometer), a force measurement unit (impact sensor), a displacement / opening measurement unit (LVDT or visual measurement), and high-speed imaging (high-speed camera).

[0035] The impact test is conducted under preset initial conditions. The conditions parameters may include impact velocity, projectile mass, impact angle, impact position, projectile attitude, etc.

[0036] The following data was collected after each impact:

[0037] Reinforcing bar strain ( ): Used to calculate peak strain and equivalent strain (durable strength index) of steel bars; used to calculate comprehensive risk index R;

[0038] Earth pressure (p): used to calculate peak pressure, pressure impulse, and residual pressure, and ultimately to calculate the comprehensive risk index R;

[0039] Acceleration (a): used for secondary impact / bounce identification and energy estimation (velocity integral). The final secondary impact / bounce identification is used for anomaly labeling, and the energy estimation is used to calculate the comprehensive risk index R.

[0040] Impact force (F): used for triggering, energy / work estimation, and ultimately for calculating the comprehensive risk index R;

[0041] Displacement / Opening (Linear Variable Differential Transformer / Vision): Used for joint opening displacement and shear slip, ultimately used to calculate the comprehensive risk index R;

[0042] High-speed imaging: used for crack segmentation and length / width / area, displacement field statistics, and ultimately used to calculate the comprehensive risk index R.

[0043] After acquiring the data, the data processing and preprocessing workflow includes:

[0044] Timing unification: All channels are aligned to t0, and a unified window is captured. (Example: t) b =2ms,t a =50ms); t0 is the time reference point, t b This is the starting offset of the time window (relative to t0), i.e., the left boundary of the window; t a This is the end offset of the time window (relative to t0), i.e., the right boundary of the window;

[0045] Filtering and robustness: Anti-aliasing low-pass filter (20–40 kHz, fourth order) + DC removal; peak values ​​are filtered using P99 or Hampel and then the extreme values ​​are taken to suppress occasional spikes; impulse / work are integrated using trapezoidal / Simpson integrals;

[0046] Anomaly detection: If a secondary peak (force / acceleration) appears after contact is released, mark it as a secondary impact, and the main scoring window will only take the first impact window;

[0047] Image processing: Crack segmentation to obtain length / maximum width / area; DIC (Digital Image Correlation) displacement / strain field extreme values ​​or P95 quantiles.

[0048] As an optional implementation, this step first determines the selection range of each operating condition parameter based on experience or pre-simulation results, ensuring that the most unfavorable operating condition falls within the range; then, the initial operating condition is determined within the range using the following method:

[0049] Obtain the parameter range for each operating condition parameter, and divide each parameter range into multiple value layers at equal intervals; the total number of layers is the same for each operating condition parameter.

[0050] For each operating condition parameter, the sampling order of each layer is randomly generated to obtain a random arrangement; for example, the random arrangement is [3, 1, 5, 2, 4], which means that the first layer is sampled in the 3rd test, the second layer is sampled in the 1st test, and so on.

[0051] By randomly permuting and combining all the operating parameters, an integer matrix is ​​obtained;

[0052] According to the test sequence number, the value layer of each working condition parameter is found in the integer matrix. For example, for the first test, the value is taken in the third layer of parameter A, the fifth layer of parameter B, and the second layer of parameter C. Random sampling is performed within the value layer to obtain the parameter values ​​of each working condition parameter in each test. The obtained working conditions can effectively cover the entire multidimensional parameter space with fewer sample points.

[0053] S200. Multiple sets of test conditions are generated using a conditional generative adversarial network. Each set of test conditions is input into a preset prediction model, which outputs a comprehensive prediction risk index and its corresponding mean and variance. The prediction model is trained based on the initial test conditions and initial impact data. Specifically, as follows:

[0054] Based on the initial impact data, the comprehensive risk index for each working condition is calculated; the comprehensive risk index is calculated as follows:

[0055] ;

[0056] ;

[0057] ;

[0058] Where R is the comprehensive risk index. For the first A physical quantity , The lower / upper limit is set for a physical quantity; n is the total number of physical quantities; For the first The weights of each physical quantity; as an example, the physical quantities used to calculate R include: peak strain of the steel reinforcement; equivalent strain of the steel reinforcement; peak soil pressure; impulse of soil pressure; residual stress in the soil; peak acceleration; peak impact force; crack length (obtained from segment image recognition); maximum crack width (obtained from segment image recognition);

[0059] Based on the comprehensive risk index of each working condition, the sampling weight information of each working condition parameter is calculated, and a standardized dataset is generated:

[0060] Obtain the type of each operating condition parameter, including continuous and discrete types;

[0061] For each continuous operating condition parameter, operating point is generated with the parameter value as the x-axis and the corresponding comprehensive risk index as the y-axis.

[0062] The weighted curve is obtained by fitting the operating points of all tested operating conditions.

[0063] The horizontal axis of the curve is discretized into several sampling intervals. The average comprehensive risk index value of each sampling interval is calculated, and the entire horizontal axis is normalized to obtain the weight of each sampling interval.

[0064] Based on the preset initial operating conditions and the corresponding comprehensive risk index, a weighted curve is constructed for each continuous operating condition parameter. Specifically, taking the speed parameter as an example, the speed values ​​of a set of tested operating conditions are used as the horizontal axis, and the comprehensive risk index corresponding to the operating condition is used as the vertical axis to generate speed operating condition points. All tested speed operating condition points are connected by a smooth curve (fitted), and the range of the horizontal axis of the resulting curve is the value range of the speed parameter. The height of the curve reflects the comprehensive risk index value that the parameter point may correspond to, and this value reflects the weight of the parameter point. The horizontal axis of the curve is discretized into several sampling intervals, the average comprehensive risk index value of each sampling interval is calculated, and the entire horizontal axis is normalized to obtain the weight of each sampling interval.

[0065] For each discrete working condition parameter, its value is finite. We only need to calculate the average value of the comprehensive risk index corresponding to each value and perform normalization to obtain the weight of each value and construct the parameter set for each discrete working condition.

[0066] Obtain the sampling intervals or parameter sets of all operating condition parameters to form a standardized dataset, which is the set of all available operating condition parameters; the parameter weights in this set will also be continuously updated as the number of experiments increases.

[0067] Based on weights, a multinomial sampling strategy is adopted to sample from the sampling interval or parameter set of each working condition parameter to generate an enhanced dataset; among which, random sampling is performed in the sampling interval, for example, taking the median of the interval and adding random perturbation (not exceeding the interval range).

[0068] Based on the calculated weights, M samples are randomly selected with replacement from the sampling interval or parameter set; the higher the weight, the greater the probability of being selected, thus resulting in a new dataset rich in high-risk samples.

[0069] Gaussian noise is randomly added to the continuous parameters of the augmented dataset, allowing the generator to learn to explore the neighborhood of the original high-risk conditions and generate new conditions with minor variations.

[0070] The conditional generative adversarial network is trained adversarially using an augmented dataset with added noise until the generator and discriminator are iteratively optimized to convergence;

[0071] The generator G is input to random noise and a conditional vector c (the vector of high-risk operating condition parameters in the augmented dataset). The discriminator D is input to the fake operating condition parameter vector generated by the generator; it needs to determine whether this vector comes from real high-risk data or is fabricated by the generator. Through continuous adversarial optimization, the operating condition parameter vector generated by the generator is eventually made to approximate the real high-risk operating conditions.

[0072] Multiple random noise vectors are input into the trained generator to obtain multiple sets of test condition parameter vectors output by the generator. The test condition parameter vectors are then denormalized and decoded to obtain multiple sets of test conditions. Here, it is necessary to screen to ensure that the test conditions generated in each iteration are untested.

[0073] Input each set of test conditions into the preset prediction model, and output the comprehensive risk index and the corresponding mean and variance;

[0074] A deep neural network, specifically a fully connected network, is used to train a prediction model by inputting the experimental operating parameters and the R-values ​​calculated from the raw impact data. The prediction model then regresses data through two regression heads. (mean of R) and σ 2 (Variance / uncertainty of R); The prediction model can predict the corresponding R value based solely on the operating parameters, and its prediction accuracy gradually improves with the increase of the number of tests; thus, it quickly approximates the most unfavorable operating conditions, greatly reducing the actual number of tests required and lowering the test costs.

[0075] S300. Construct working condition interaction items and filter them using a random forest model to obtain target interaction items;

[0076] Combinations of different operating conditions may produce synergistic or inhibitory effects. This method needs to explore whether paired combinations will "amplify" or "weaken" the risk. Therefore, it is not only necessary to consider the individual effects of parameters such as angle θ, velocity v, mass m, and position d on the results, but also to consider the combined effects of operating conditions such as θ×d (angle + position), θ×v (angle + velocity), and d×v (position + velocity), as detailed below:

[0077] S310. Obtain the type of each working condition parameter in each group of working conditions, wherein the type includes continuous type and discrete type;

[0078] S320. For interaction combinations consisting of continuous parameters, the interaction terms are obtained by calculating the parameter product or constructing a nonlinear polynomial.

[0079] S330. For an interaction combination consisting of continuous and discrete parameters, perform one-hot encoding on the discrete parameters, and multiply the encodings of the continuous and discrete parameters to obtain the interaction term.

[0080] S340. For interactive combinations consisting of discrete parameters, perform one-hot encoding on the discrete parameters, calculate the outer product of all discrete parameter encodings, and obtain the interactive term.

[0081] For example, combinations could be formed such as: parameter A × parameter B; parameter A × parameter C; parameter A × parameter B × parameter C;

[0082] S350. Train a random forest model using the working parameters, interaction terms and corresponding comprehensive risk indicators of each working condition, and output the importance of each interaction term based on the trained random forest model.

[0083] Random Forest (RF) integrates the prediction results of multiple decision trees, which can objectively quantify the contribution of each feature (including original operating parameters and candidate interaction terms) to the target variable (comprehensive risk index R).

[0084] After training, each feature (original parameter / candidate interaction term) is randomly permuted to break the association between the feature and the target variable. The prediction error after the permutation is then calculated. The greater the increase in error, the greater the influence of the feature on the result. This allows us to obtain the importance of each original working condition parameter and interaction term.

[0085] S360. Filter out the target interaction items based on their importance;

[0086] In many cases, the interaction effect may be overestimated because a certain original parameter in the interaction item is more important. Therefore, this step first needs to compare the importance of the interaction item with the importance of the individual parameter in the interaction item, and only retain the interaction items whose importance is higher than the corresponding original parameter to ensure that the interaction effect is significant. Then, the retained interaction items are filtered based on the preset importance threshold, and the parts with lower importance are removed to obtain the target interaction items.

[0087] S400: Based on the mean and variance of the comprehensive risk index and the target interaction term of the test conditions, calculate the expected improvement amount for each group of test conditions.

[0088] Based on the mean, variance, interaction term, and the current maximum composite risk index, the cumulative distribution function of the standard normal distribution is constructed as follows: ;

[0089] The cumulative distribution function of the standard normal distribution. The weighted sum of all target interactions. This is currently the largest comprehensive risk indicator; Preset the weight coefficients for interaction items; The standard deviation of R for the regression of the prediction model; This is the mean of R-squared for the regression of the prediction model;

[0090] Based on the mean, variance, interaction term, and the current maximum composite risk index, the probability density function of the standard normal distribution is constructed as follows: ;

[0091] The cumulative distribution function of the standard normal distribution; The weighted sum of all target interactions. This is currently the largest comprehensive risk indicator; Preset the weight coefficients for interaction items; The standard deviation of R for the regression of the prediction model; This is the mean of R-squared for the regression of the prediction model;

[0092] The expected improvement formula is constructed based on the cumulative distribution function and the probability density function, and the expected improvement amount for each set of test conditions is calculated based on the expected improvement formula.

[0093] The desired improved formula is:

[0094] ;

[0095] EI To the desired improvement amount, < When this condition is not improved, it indicates that there is no potential for improvement. The weighted sum of all target interactions. This is currently the largest comprehensive risk indicator; Preset the weight coefficients for interaction items; The standard deviation of R for the regression of the prediction model; This is the mean of R-squared for the regression of the prediction model; The cumulative distribution function of the standard normal distribution; The cumulative distribution function of the standard normal distribution;

[0096] S500: Select the test condition based on the expected improvement amount for impact test, update the conditions based on the test results to generate an adversarial network and prediction model, and iterate until convergence to obtain the test condition with the largest comprehensive risk index.

[0097] As an example, the convergence condition can be set as follows: the improvement of R by ΔR in three consecutive trials is less than 5%, and the average prediction uncertainty Δ <5%.

[0098] This embodiment also provides an experimental apparatus for implementing the above method, see [link to apparatus]. Figure 2-4 ,include:

[0099] Square rectangular tube 4, which is a hollow steel tube, is mainly concentrated on the outside of the device and serves as the load-bearing structure of the device.

[0100] The semi-circular flange 2 serves as a connecting and positioning component of the structure. The semi-circular flange 2 can perform impact tests at different angles based on the position of its bolt holes. There are two semi-circular flanges 2, one upper and one lower, connected by an impact-side rectangular flange 1. The impact-side rectangular flange 1 can serve as an installation positioning plate for the pipe segments and soil boxes, enabling precise installation of the pipe segments and soil boxes.

[0101] A triangular steel plate 3 has a first bolt hole 6 along the arc surface of the semi-circular flange, which can be connected to the semi-circular flange 2 by bolts.

[0102] The test apparatus also includes:

[0103] The rectangular flange 5, located on the back side of the impact, can improve the overall stability of the device structure.

[0104] Bolt holes are pre-drilled at preset intervals on the triangular steel plate 3 and the semi-circular flange 2, allowing for different angles to be changed according to test requirements.

[0105] The triangular steel plate 3 and the square rectangular tube 4 are connected by welding.

[0106] The semi-circular flange 2 and the impact-side rectangular flange 1 are connected by welding, which allows the semi-circular flange 2 to move along the circumference together with the impact-side rectangular flange 1.

[0107] The impact-side rectangular flange 1 is provided with a second bolt hole 7 and a third bolt hole 8, which can be used to connect to the soil box and the pipe segment respectively.

[0108] The square rectangular tubes 4 are connected to each other by welding, and the side areas are reinforced by connecting diagonally, while the top area is reinforced by connecting the midpoints of triangles.

[0109] The rectangular flange 5 on the impact back side has bolt holes, which enables the device to be assembled and positioned on the test site.

[0110] The test operation method of the above-mentioned device includes the following steps:

[0111] The test device is fixed to the test site using the bolt holes on the rectangular flange 5 on the impact back side;

[0112] Rotate the semi-circular flange 2 to the angle required for the test, and connect the semi-circular flange 2 and the triangular steel plate 3 with bolts;

[0113] The pipe segment 10 and the soil box 9 are hoisted to the corresponding positions of the impact-side rectangular flange and connected by bolts.

[0114] Rotating the semi-circular flange 2 includes the following steps:

[0115] According to the test requirements, determine the angle of the impact test, adjust the semi-circular flange 2 to the corresponding position, ensure that the first bolt holes 6 can be aligned with each other, and tighten all the aligned bolt holes with bolts. The operation is the same for the upper and lower semi-circular flanges.

[0116] The following steps are included when hoisting the segment 10 and the soil box 9:

[0117] First, the segment 10 is lifted as a whole by a crane. The position of the segment is continuously adjusted until it is aligned with the third bolt hole 8. The upper and lower fixing bolts are tightened. Then, the soil box is fixed through the second bolt holes 7 on both the left and right sides.

[0118] Example 2:

[0119] A deep learning-based worst-case scenario identification system includes:

[0120] The first module is used to collect initial impact data, which is obtained after an impact test is conducted using preset initial conditions.

[0121] The second module is used to generate multiple sets of test conditions using a conditional generative adversarial network, input each set of test conditions into a preset prediction model, and output a comprehensive prediction risk index and the corresponding mean and variance; the prediction model is trained based on the initial test conditions and initial impact data.

[0122] The third module is used to construct working condition interaction items and filter them using a random forest model to obtain target interaction items;

[0123] The fourth module is used to predict the mean and variance of the comprehensive risk index and the target interaction term based on the test conditions, and to calculate the expected improvement amount for each group of test conditions.

[0124] The fifth module is used to select the test conditions for impact testing based on the expected improvement amount, update the conditions based on the test results to generate an adversarial network and a prediction model, and iterate until convergence to obtain the test condition with the largest comprehensive risk index.

[0125] As an optional implementation, the system further includes:

[0126] The sixth module is used to obtain the parameter range of each working condition parameter and divide each parameter range into multiple value layers at equal intervals.

[0127] The seventh module is used to randomly generate the sampling order of each layer for each working condition parameter, resulting in a random permutation. The random permutations of all working condition parameters are combined to obtain an integer matrix.

[0128] The eighth module is used to find the value layer of each working condition parameter in the integer matrix according to the test sequence number, and to perform random sampling within the value layer to obtain the parameter values ​​of each working condition parameter for each test.

[0129] As an optional implementation, the second module includes:

[0130] The first unit is used to calculate the comprehensive risk index for each set of working conditions based on the initial impact data.

[0131] The second unit is used to calculate the sampling weight information of each working condition parameter based on the comprehensive risk index of each working condition, and generate a standardized dataset.

[0132] The third unit is used to sample from the standardized dataset based on a weighted multinomial sampling strategy to generate an enhanced dataset;

[0133] The fourth unit is used to randomly add Gaussian noise to the continuous parameters of the augmented dataset;

[0134] The fifth unit is used to perform adversarial training on the conditional generative adversarial network using an augmented dataset with added noise, until the generator and discriminator are iteratively optimized to convergence;

[0135] The sixth unit is used to input multiple random noise vectors into the trained generator to obtain multiple sets of test condition parameter vectors output by the generator. The test condition parameter vectors are then denormalized and decoded to obtain multiple sets of test conditions.

[0136] Example 3:

[0137] Corresponding to the above method embodiments, this embodiment also provides a deep learning-based worst-case condition identification device. The deep learning-based worst-case condition identification device described below and the deep learning-based worst-case condition identification method described above can be referred to each other.

[0138] Figure 5 This is a block diagram illustrating a deep learning-based worst-case scenario identification device 800 according to an exemplary embodiment. Figure 5As shown, the deep learning-based worst-case scenario identification device 800 includes a processor 801 and a memory 802. The deep learning-based worst-case scenario identification device 800 may also include one or more of the following: a multimedia component 803, an input / output (I / O) interface 804, and a communication component 805. The processor 801 controls the overall operation of the deep learning-based worst-case scenario identification device 800 to complete all or part of the steps in the deep learning-based worst-case scenario identification method described above. The memory 802 stores various types of data to support the operation of the deep learning-based worst-case scenario identification device 800. This data may include, for example, commands for any application or method operating on the deep learning-based worst-case scenario identification device 800, and application-related data such as contact data, sent and received messages, images, audio, video, etc. The memory 802 can be implemented by any type of volatile or non-volatile storage device or a combination thereof, such as static random access memory (SRAM), electrically erasable programmable read-only memory (EEPROM), erasable programmable read-only memory (EPROM), programmable read-only memory (PROM), read-only memory (ROM), magnetic storage, flash memory, magnetic disk or optical disk.

[0139] Multimedia component 803 may include a screen and an audio component. The screen may be, for example, a touchscreen, and the audio component is used to output and / or input audio signals. For example, the audio component may include a microphone for receiving external audio signals.

[0140] The received audio signal can be further stored in memory 802 or transmitted via communication component 805. The audio component also includes at least one speaker for outputting audio signals. I / O interface 804 provides an interface between processor 801 and other interface modules, such as a keyboard, mouse, buttons, etc. These buttons can be virtual or physical. Communication component 805 is used for wired or wireless communication between the deep learning-based worst-case scenario identification device 800 and other devices. Wireless communication includes, for example, Wi-Fi, Bluetooth, Near Field Communication (NFC), 2G, 3G, or 4G, or a combination thereof; therefore, the corresponding communication component 805 may include a Wi-Fi module, a Bluetooth module, or an NFC module.

[0141] Example 4:

[0142] Corresponding to the above embodiment of the deep learning-based worst-case identification method, this embodiment also provides a readable storage medium. The readable storage medium described below can be referred to in correspondence with the deep learning-based worst-case identification method described above.

[0143] A readable storage medium storing a computer program, which, when executed by a processor, implements the steps of the above-described embodiment of the deep learning-based worst-case identification method.

[0144] Specifically, the readable storage medium can be a USB flash drive, external hard drive, read-only memory (ROM), random access memory (RAM), magnetic disk, or optical disk, or any other readable storage medium capable of storing program code.

[0145] It should be noted that, in this document, relational terms such as "first" and "second" are used merely to distinguish one entity or operation from another, and do not necessarily require or imply any such actual relationship or order between these entities or operations. Furthermore, the terms "comprising," "including," or any other variations thereof are intended to cover non-exclusive inclusion, such that a process, method, article, or apparatus that comprises a list of elements includes not only those elements but also other elements not expressly listed, or elements inherent to such a process, method, article, or apparatus. Without further limitations, an element defined by the phrase "comprising one..." does not exclude the presence of other identical elements in the process, method, article, or apparatus that includes said element.

[0146] The above description is merely a specific embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any variations or substitutions that can be easily conceived by those skilled in the art within the technical scope disclosed in the present invention should be included within the scope of protection of the present invention. Therefore, the scope of protection of the present invention should be determined by the scope of the claims.

Claims

1. A method for identifying the most unfavorable working condition based on deep learning, characterized in that, include: Initial impact data is collected, which is obtained after an impact test is conducted using preset initial conditions. Multiple sets of test conditions are generated using a conditional generative adversarial network. Each set of test conditions is input into a preset prediction model, and the predicted comprehensive risk index and the corresponding mean and variance are output. The prediction model is trained based on the initial working conditions and initial impact data; The working condition interaction terms are constructed and filtered using a random forest model to obtain the target interaction terms, including: Obtain the type of each working condition parameter in each group of working conditions, including continuous and discrete types; For interaction combinations consisting of continuous parameters, the interaction terms can be obtained by calculating the parameter product or constructing a nonlinear polynomial. For an interaction combination consisting of continuous and discrete parameters, the discrete parameters are one-hot encoded, and the codes of the continuous and discrete parameters are multiplied together to obtain the interaction term. For an interaction combination consisting entirely of discrete parameters, one-hot encoding is performed on the discrete parameters, and the outer product of all discrete parameter encodings is calculated to obtain the interaction term. The random forest model is trained using the working parameters, interaction terms and corresponding comprehensive risk indicators of each working condition. The importance of each interaction term is output based on the trained random forest model. Target interaction items are obtained by filtering based on their importance; Based on the mean and variance of the comprehensive risk index predicted for the test conditions, and the objective interaction term, the expected improvement for each set of test conditions is calculated, including: Construct the cumulative distribution function of the standard normal distribution based on the mean, variance, interaction term, and the current maximum comprehensive risk index; Construct the probability density function of the standard normal distribution based on the mean, variance, interaction term, and the current maximum comprehensive risk index; The expected improvement formula is constructed based on the cumulative distribution function and the probability density function, and the expected improvement amount for each set of test conditions is calculated based on the expected improvement formula. The desired improved formula is: ； EI The desired improvement amount; I The weighted sum of all target interactions. This is currently the largest comprehensive risk indicator; Preset the weight coefficients for interaction items; The standard deviation of R for the regression of the prediction model; This is the mean of the R-value for the regression of the predictive model; The cumulative distribution function of the standard normal distribution; The cumulative distribution function of the standard normal distribution; The test conditions are selected based on the expected improvement amount for impact testing. The conditions are updated based on the test results to generate an adversarial network and a prediction model. The process is iterated until convergence to obtain the test condition with the largest comprehensive risk index.

2. The method for identifying the most unfavorable working condition based on deep learning according to claim 1, characterized in that, The method includes: Obtain the parameter range for each operating condition parameter, and divide each parameter range into multiple value layers at equal intervals; For each operating condition parameter, the sampling order of each layer is randomly generated to obtain a random arrangement. The random arrangements of all operating condition parameters are combined to obtain an integer matrix. According to the test sequence number, the value layer of each working condition parameter is found in the integer matrix, and random sampling is performed within the value layer to obtain the parameter value of each working condition parameter for each test.

3. The method for identifying the most unfavorable working condition based on deep learning according to claim 1, characterized in that, Multiple test cases are generated using a conditional generative adversarial network, including: Based on the initial impact data, calculate the comprehensive risk index for each working condition. Based on the comprehensive risk index of each working condition, the sampling weight information of each working condition parameter is calculated, and a standardized dataset is generated. Based on weights, a multinomial sampling strategy is used to sample from the standardized dataset to generate an enhanced dataset; Randomly add Gaussian noise to the continuous parameters of the augmented dataset; The conditional generative adversarial network is trained adversarially using an augmented dataset with added noise until the generator and discriminator are iteratively optimized to convergence; Multiple random noise vectors are input into the trained generator to obtain multiple sets of test condition parameter vectors output by the generator. The test condition parameter vectors are then denormalized and decoded to obtain multiple sets of test conditions.

4. The method for identifying the most unfavorable working condition based on deep learning according to claim 1, characterized in that, Based on the comprehensive risk index of each working condition, the sampling weight information of each working condition parameter is calculated, and a standardized dataset is generated, including: Obtain the type of each operating condition parameter, including continuous and discrete types; For each continuous operating condition parameter, operating point is generated with the parameter value as the x-axis and the corresponding comprehensive risk index as the y-axis. The weighted curve is obtained by fitting the operating points of all tested operating conditions. The horizontal axis of the curve is discretized into several sampling intervals. The average comprehensive risk index value of each sampling interval is calculated, and the entire horizontal axis is normalized to obtain the weight of each sampling interval. For each discrete working condition parameter, calculate the average value of the comprehensive risk index corresponding to each value and normalize it to obtain the weight of each value, and construct the parameter set for each discrete working condition. Obtain the sampling intervals or parameter sets of all operating condition parameters to form a standardized dataset.

5. The method for identifying the most unfavorable working condition based on deep learning according to claim 1, characterized in that, The test conditions are selected based on the expected improvement amount, including: The test conditions are sorted from largest to smallest according to the expected improvement amount. When the difference between the expected improvement amount of the first test condition and the subsequent test conditions is greater than or equal to the preset threshold, the first test condition is used for the test. If the difference between the expected improvement of the first test condition and the subsequent test conditions is less than the preset threshold, the predicted comprehensive risk index corresponding to the test condition is compared. If the predicted comprehensive risk indexes are not equal, the test condition with the largest predicted comprehensive risk index is selected. If the predicted comprehensive risk indicators are equal, the convergence stage is determined by the difference between the historical maximum expected improvement. In the early stage of convergence, the test condition with a larger predicted standard deviation is selected, and in the later stage of convergence, the test condition with a smaller standard deviation is selected.

6. A deep learning-based system for identifying the most unfavorable working conditions, characterized in that, include: The first module is used to collect initial impact data, which is obtained after an impact test is conducted using preset initial conditions. The second module is used to generate multiple sets of test conditions using a conditional generative adversarial network. Each set of test conditions is input into a preset prediction model, and the prediction comprehensive risk index and the corresponding mean and variance are output. The prediction model is trained based on the initial working conditions and initial impact data; The third module is used to construct working condition interaction items and filter them using a random forest model to obtain target interaction items; it includes: Obtain the type of each working condition parameter in each group of working conditions, including continuous and discrete types; For interaction combinations consisting of continuous parameters, the interaction terms can be obtained by calculating the parameter product or constructing a nonlinear polynomial. For an interaction combination consisting of continuous and discrete parameters, the discrete parameters are one-hot encoded, and the codes of the continuous and discrete parameters are multiplied together to obtain the interaction term. For an interaction combination consisting entirely of discrete parameters, one-hot encoding is performed on the discrete parameters, and the outer product of all discrete parameter encodings is calculated to obtain the interaction term. The random forest model is trained using the working parameters, interaction terms and corresponding comprehensive risk indicators of each working condition. The importance of each interaction term is output based on the trained random forest model. Target interaction items are obtained by filtering based on their importance; The fourth module is used to predict the mean and variance of the comprehensive risk index based on the test conditions, the target interaction term, and to calculate the expected improvement for each set of test conditions, including: Construct the cumulative distribution function of the standard normal distribution based on the mean, variance, interaction term, and the current maximum comprehensive risk index; Construct the probability density function of the standard normal distribution based on the mean, variance, interaction term, and the current maximum comprehensive risk index; The expected improvement formula is constructed based on the cumulative distribution function and the probability density function, and the expected improvement amount for each set of test conditions is calculated based on the expected improvement formula. The desired improved formula is: ； EI The desired improvement amount; I The weighted sum of all target interactions. This is currently the largest comprehensive risk indicator; Preset the weight coefficients for interaction items; The standard deviation of R for the regression of the prediction model; This is the mean of the R-value for the regression of the predictive model; The cumulative distribution function of the standard normal distribution; The cumulative distribution function of the standard normal distribution; The fifth module is used to select the test conditions for impact testing based on the expected improvement amount, update the conditions based on the test results to generate an adversarial network and a prediction model, and iterate until convergence to obtain the test condition with the largest comprehensive risk index.

7. The deep learning-based worst-case scenario identification system according to claim 6, characterized in that, The system also includes: The sixth module is used to obtain the parameter range of each working condition parameter and divide each parameter range into multiple value layers at equal intervals. The seventh module is used to randomly generate the sampling order of each layer for each working condition parameter, resulting in a random permutation. The random permutations of all working condition parameters are combined to obtain an integer matrix. The eighth module is used to find the value layer of each working condition parameter in the integer matrix according to the test sequence number, and to perform random sampling within the value layer to obtain the parameter values ​​of each working condition parameter for each test.

8. The deep learning-based worst-case condition identification system according to claim 6, characterized in that, The second module includes: The first unit is used to calculate the comprehensive risk index for each set of working conditions based on the initial impact data. The second unit is used to calculate the sampling weight information of each working condition parameter based on the comprehensive risk index of each working condition, and generate a standardized dataset. The third unit is used to sample from the standardized dataset based on a weighted multinomial sampling strategy to generate an enhanced dataset; The fourth unit is used to randomly add Gaussian noise to the continuous parameters of the augmented dataset; The fifth unit is used to perform adversarial training on the conditional generative adversarial network using an augmented dataset with added noise, until the generator and discriminator are iteratively optimized to convergence; The sixth unit is used to input multiple random noise vectors into the trained generator to obtain multiple sets of test condition parameter vectors output by the generator. The test condition parameter vectors are then denormalized and decoded to obtain multiple sets of test conditions.

Images