Giant estuary ship navigation capability prediction model test method fused with machine learning

Abstract

The invention relates to the technical field of ship navigation model tests, and discloses a giant estuary ship navigation capability prediction model test method fused with machine learning, which comprises the following steps: initializing and defining a scene vector, training an agent model, generating a population, and setting archives and parameters; executing dynamic fitness evaluation to obtain a fitness score; generating filial generations based on score selection and genetic operators and ensuring physical effectiveness; selecting a scene with the highest score to obtain a predicted risk; when the error is greater than a threshold value, storing the candidate scene; if yes, outputting the file; and if not, replacing the population and returning evaluation. According to the method, a feedback closed loop is constructed through failure judgment and updating of a failure mode file, after a failure scene verified by a high-fidelity simulator is stored in the file, the areas are avoided when the novel fitness is calculated through dynamic fitness evaluation, the search direction is automatically adjusted, and automatic and self-adaptive exploration of a high-risk scene is achieved.

Description

Technical Field

[0001] This invention relates to the field of ship navigation model testing technology, specifically to a test method for predicting the navigation capacity of ships in giant estuaries by incorporating machine learning. Background Technology

[0002] Giant estuaries are critical nodes in global shipping, and their complex navigation environments pose significant challenges to vessel safety. The industry increasingly relies on advanced machine learning predictive models (MUTs) to assess navigational capacity and risks in real time. The reliability of these models directly impacts the safety of both vessels and ports. Therefore, comprehensive and rigorous testing and validation are essential before deploying any model.

[0003] The industry standard for validating these predictive models relies on high-fidelity simulators (HFS). High-fidelity simulations (such as those based on computational fluid dynamics) can provide risk results that can be considered as real-world scenarios, but HFS is extremely computationally expensive. To find model flaws within an acceptable cost, existing technologies have begun to employ automated strategies such as evolutionary or genetic algorithms to attempt to search for high-risk test scenarios in a vast parameter space before submitting them to the HFS for validation.

[0004] However, existing evolutionary algorithms, when generating test scenarios, use genetic operators (such as crossover and mutation) that are essentially random and not coupled with the physical constraints between scenario parameters. This process generates a large number of physically invalid parameter combinations. Submitting these invalid scenarios to high-cost high-fidelity simulators for verification leads to a serious waste of computational resources. Existing search methods generally rely on static fitness functions, which are usually based on only a single physical risk indicator. This makes the search prone to focusing on local optima. Furthermore, existing testing processes often lack a mechanism to feed authoritative verification results back to the search algorithm. Failure data output by high-fidelity simulators is only used as the final test report and is not used to dynamically adjust subsequent search strategies. Therefore, this invention provides a test method for predicting the navigation capacity of ships in giant estuaries by integrating machine learning, in order to address the shortcomings of existing technologies. Summary of the Invention

[0005] To address the shortcomings of existing technologies, this invention provides an experimental method for predicting the navigation capacity of ships in giant estuaries by integrating machine learning. This method solves the problems of low testing efficiency, easy getting stuck in local optima, and inability to adaptively explore unknown areas caused by the lack of physical constraints in the generation of test scenarios, the reliance of search strategies on static fitness functions, and the lack of feedback loops in the testing process.

[0006] To achieve the above objectives, the present invention provides the following technical solution: a test method for a prediction model of navigation capacity of giant estuaries integrating machine learning, comprising the following steps: S1. System initialization includes defining scene gene vectors describing the general aviation scenario, training constrained agent models offline, generating an initial scene population, initializing an empty failure mode file, setting the maximum number of generations and the number of target failure modes, and initializing the current generation. S2. Using the initial scene population as the current scene population, start the iterative loop and perform dynamic fitness evaluation on the scene gene vectors in the current scene population to obtain fitness scores. S3. Based on the fitness score, perform selection operations, perform crossover and mutation operations through adaptive genetic operators to generate offspring population, and call the constraint proxy model to ensure that the generated offspring scene gene vector conforms to physical laws; S4. Select the candidate scene with the highest fitness score from the current scene population, and call the high-fidelity simulator and the test model as the test object to perform simulation or prediction respectively, so as to obtain the risk result and prediction risk result of the candidate scene. S5. Calculate the prediction error between the risk result and the predicted risk result. When the prediction error is greater than the preset failure threshold, store the corresponding candidate scenario in the failure mode file. S6. If the current generation number is greater than the maximum generation number or the number of scenarios in the failure mode file reaches the target number of failure modes, the iteration is terminated and the failure mode file is output as the exploration result. If the condition is not met, the current scenario population is replaced by the offspring population, and the dynamic fitness evaluation is returned.

[0007] Preferably, in step S1, the constraint proxy model further includes: The discrimination function is a classifier, which receives the complete scene gene vector as input and outputs a Boolean value to determine whether the parameter combination represented by the scene gene vector conforms to the preset physical law. The constraint interval function, which is a regression model or boundary prediction model, is used to determine and give the allowed physical effective range of values ​​for unknown parameters based on known parameter values ​​when the parameter value of a certain dimension in the scene gene vector is unknown.

[0008] Preferably, in step S2, the fitness score is calculated based on the following criteria: The uncertainty of the test model's prediction of the scene gene vector as the test object; The prediction uncertainty is quantified by enabling Dropout or employing a model ensemble approach during prediction, and by calculating the statistical variance of multiple prediction results.

[0009] Preferably, in step S2, the calculation of the fitness score further includes: The novelty fitness of the scene gene vector and all existing scenes in the failure mode archive; The novelty fitness is used to drive evolutionary search to avoid discovered failure regions; The novelty fitness is calculated by obtaining the distance metric between the scene gene vector and the failure scenes stored in the failure mode archive, and taking the minimum value among them.

[0010] Preferably, when the failure mode file has not yet been stored in any failure scenarios, the novelty fitness is a preset constant maximum value.

[0011] Preferably, in step S3, the step of the adaptive genetic operator performing the crossover operation further includes: The discrimination function of the constraint proxy model is invoked to ensure that the offspring scene gene vector conforms to physical laws. The invocation process includes: After generating the offspring vector, the discrimination function is called to perform discrimination; When the child vector is determined to be inconsistent with physical laws, the child vector is discarded and the crossover operation is re-executed until the generated child vector is determined to be consistent with physical laws by the discrimination function.

[0012] Preferably, in step S3, the step of the adaptive genetic operator performing mutation operation further includes: The constraint interval function of the constraint proxy model is invoked to ensure that the offspring scene gene vector conforms to physical laws. The invocation process includes: Before performing mutation operations on specific genes in the scene gene vector, the constraint interval function is invoked to obtain the physical valid value range of the specific gene. After generating a new gene value, the new gene value is restricted to the physical valid value range. When the new gene value exceeds the physical valid value range, the new gene value is set as the boundary value of the physical valid value range.

[0013] Preferably, in step S4, the step of selecting the candidate scene with the highest fitness score further includes: The current scene population is sorted in descending order based on the fitness score; The scene gene vector with the highest score is selected from the ranking results as the candidate scene.

[0014] Preferably, in step S5, the step of calculating the prediction error further includes: quantifying the prediction error by calculating the absolute value of the difference between the risk result and the predicted risk result.

[0015] It also provides a test system for predicting the navigation capacity of giant estuaries using machine learning, including: The scenario definition and interface module is used to define parameters affecting navigation capability as scenario gene vectors and call the tested model as the test object to obtain the predicted risk results of candidate scenarios. A high-fidelity simulator module is used to receive the candidate scenarios and perform simulations to obtain the risk outcomes of the candidate scenarios; The constrained proxy model module provides a discrimination function and a constraint interval function. The discrimination function is used to determine whether the parameter combination in the input scene gene vector conforms to the preset physical law. The constraint interval function is used to obtain the physical valid value interval of the undetermined parameter based on the existing parameters when a certain parameter in the scene gene vector is undetermined. The failure mode file module is used to store the corresponding candidate scenario when the prediction error between the risk result and the predicted risk result is greater than a preset failure threshold, thus forming a confirmed failure mode. The dynamic fitness assessment module is used to calculate the fitness score of the scene gene vector in the current population; The evolutionary scene generator module is used to determine whether the termination condition is met. When the termination condition is not met, the current scene population is replaced with the offspring population, and the dynamic fitness evaluation module is triggered to return to calculate the fitness score.

[0016] This invention provides an experimental method for predicting the navigation capacity of ships in large estuaries using a machine learning-integrated model. It offers the following advantages: 1. This invention constructs a feedback loop through failure determination and failure mode archive updates. After the failure scenarios verified by the high-fidelity simulator are stored in the archive, the dynamic fitness evaluation will avoid these areas when calculating the novelty fitness in subsequent iterations. This enables the entire test system to automatically adjust the search direction, realize automated and adaptive exploration of high-risk scenarios, and ensure that the potential defects of the tested model are discovered to the maximum extent within a limited test cycle.

[0017] 2. This invention is based on the coupling of the prediction uncertainty of the model under test and the distance metric between the scenario and the failure mode archive. This dynamic fitness guides the evolutionary search to prioritize regions where the prediction of the model under test is uncertain and novel, thereby overcoming the defect of traditional testing methods that are prone to getting trapped in local optima and improving the ability to explore unknown and diverse failure modes.

[0018] 3. By setting a constraint proxy model, the present invention calls the discrimination function or constraint interval function in real time when the evolutionary scene generator module performs crossover and mutation operations, ensuring that the newly generated offspring scene gene vector always conforms to the preset physical laws, avoiding searching in an invalid parameter space, thereby improving the efficiency of evolutionary search. Attached Figure Description

[0019] Figure 1 This is a system architecture diagram of the present invention; Figure 2 This is a flowchart of the method of the present invention; Figure 3 This is a schematic diagram of the constraint proxy model of the present invention.

[0020] The modules are: 10. Scene definition and interface module; 20. High-fidelity simulator module; 30. Constraint proxy model module; 40. Failure mode archive module; 50. Dynamic fitness evaluation module; and 60. Evolutionary scene generator module. Detailed Implementation

[0021] The technical solutions in 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 embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0022] See attached document Figure 1 , Figure 1 This is a system architecture diagram according to an embodiment of the present invention. The present invention provides an experimental system for predicting the navigation capacity of vessels in large estuaries by incorporating machine learning, comprising: The scenario definition and interface module 10 is used to define multiple parameters affecting navigation capability as scenario gene vectors with a unified data structure, and to provide a standardized interface for data interaction with the model under test (MUT). This scenario definition and interface module 10 receives the scenario gene vectors, transmits them to the model under test, and receives the predicted risk value and prediction uncertainty measure returned by the model under test.

[0023] The high-fidelity simulator module 20 integrates a validated, high-precision physical environment simulation model. This module receives one or more scene gene vectors as input and performs costly simulation calculations based on these vectors, outputting a realistic risk result for the ground.

[0024] The constraint surrogate model module 30 is a lightweight model obtained through offline training. The constraint surrogate model module 30 includes: a discrimination function, which is used to quickly determine whether the parameter combination in the input scene gene vector conforms to the preset physical law; and a constraint interval function, which is used to return the physical valid value interval of the undetermined parameter based on other existing parameters when a certain parameter in the scene gene vector is undetermined.

[0025] The failure mode file module 40 stores scene gene vectors that, as verified by the high-fidelity simulator module 20, cause the prediction error of the tested model to exceed a preset failure threshold. The stored set of scene gene vectors constitutes the confirmed failure modes.

[0026] The dynamic fitness evaluation module 50 is used to calculate the fitness score of the scene gene vector in the current population. The fitness score is calculated based on: the potential impact of the scene, the prediction uncertainty measure output by the tested model, and the distance measure between the scene gene vector and all scene gene vectors stored in the failure mode archive module 40. The dynamic fitness evaluation module 50 provides the calculated fitness score to the evolutionary scene generator module 60 as the basis for its selection operation.

[0027] The evolutionary scene generator module 60 internally implements a search algorithm based on a biological model, such as a genetic algorithm. This module performs selection, crossover, and mutation operations on the population composed of scene gene vectors to generate new offspring populations. The execution of the crossover and mutation operations is controlled by the discrimination results or constraint intervals output by the constraint proxy model module 30 to ensure that the generated offspring scene gene vectors conform to physical laws.

[0028] See attached document Figure 2 , Figure 2 This is a flowchart of a method according to an embodiment of the present invention. The present invention provides a test method for a prediction model of navigation capacity of vessels in large estuaries that integrates machine learning, comprising the following steps: S1, system initialization includes: defining scene gene vectors describing navigation scenarios; offline training of constraint surrogate models (CSM); generating an initial scene population and initializing an empty failure mode profile (CFA).

[0029] S2, for each scenario gene vector in the current scenario population, perform dynamic fitness evaluation. The fitness score is calculated based on: the potential impact of the general aviation scenario, the prediction uncertainty of the model under test (MUT) for the scenario, and the novelty measure of the general aviation scenario and all existing scenarios in the failure mode archive (CFA).

[0030] S3, perform a selection operation based on the fitness score to determine the parent individuals, and apply an adaptive genetic operator to generate the offspring population; when performing crossover and mutation operations, the adaptive genetic operator actively calls the discrimination function or constraint interval function of the constrained surrogate model (CSM) to ensure that the generated offspring scene gene vector conforms to physical laws.

[0031] S4. Periodically select several candidate scenarios with the highest fitness scores from the current scenario population, and call the high-fidelity simulator (HFS) and the model under test (MUT) to perform simulations respectively to obtain the authoritative risk results and predicted risk results of the candidate scenarios.

[0032] S5, calculate the prediction error between the authoritative risk result and the predicted risk result; when the prediction error is greater than the preset failure threshold, store the corresponding candidate scenario in the failure mode file (CFA); determine whether the termination condition is met, if not, return to execute dynamic fitness evaluation.

[0033] S6. When the termination condition is met (e.g., the maximum number of iterations is reached or the number of scenarios in the failure mode file reaches a predetermined target), the iteration is terminated and the failure mode file (CFA) is output as the exploration result.

[0034] In the initialization phase of step S1, this initialization phase includes defining a scene gene vector describing a navigation scenario in a giant estuary. The scene's gene vector Constructed as 3D real vector Each dimension of the vector (in These parameters correspond to key parameters that affect navigation capability. In a preferred embodiment, these parameters can be divided into: environmental parameters (e.g., tide level, average current velocity, current direction, wind speed, wind direction, visibility, etc. in a specified area), ship parameters (e.g., the ship's draft, speed, ship type, maneuverability indicators, etc.), and traffic parameters (e.g., whether there are encountering ships, the relative position and course of the encountering ships, the type of encounter situation, etc.).

[0035] The initialization phase also includes offline training of the Constrained Agent Model (CSM). This training process specifically includes: training the model offline using a high-fidelity simulator (HFS) in a complete... Sampling is performed within the parameter space (e.g., using Latin hypercube sampling or orthogonal experimental design) to generate a large amount of physical scene sample data; based on the sample data, the constraint surrogate model is trained. The constraint surrogate model (CSM) is a lightweight machine learning model, such as Gaussian process regression (GPR), support vector machine (SVM), or small artificial neural network (ANN).

[0036] The training process enables the Constrained Surrogate Model (CSM) to possess two key functions: Discriminant function This discriminative function is trained as a classifier to receive complete scene gene vectors. As input, and quickly output a Boolean value (true / false) to determine the vector. Whether the combination of parameters represents conforms to physical laws (e.g., whether a specific tidal level and a specific flow velocity can coexist).

[0037] Constraint interval function This constrained interval function is trained as a regression model or boundary prediction model to be used in scene vectors. The Vigene In the case of pending determination, depending on other The value of the dimension parameter is returned for the first dimension. Vigene Physical effective value range .

[0038] The initialization phase further includes setting global parameters for system operation, such as the maximum number of generations. Population size Authoritative verification period Number of candidate scenarios for a single verification and failure determination threshold .

[0039] Simultaneously, the Failure Mode and Effects File (CFA) used to store confirmed failure modes is initialized. Initially, this file is set to an empty set, i.e. .

[0040] This initialization phase includes generating an initial scene population. The generation process specifically involves: random generation. Each scene gene vector, and utilizes a trained constrained surrogate model (CSM). The discrimination function, for the Each vector is individually evaluated for physical validity. The system retains all vectors evaluated as "compliant with physical laws" and repeats this process until a result is obtained. A valid scene gene vector, which The vectors together constitute the initial population. .

[0041] During the dynamic fitness assessment phase of step S2, the system evaluates the current population. ( For each scene gene vector in the current generation) Calculate its dynamic fitness This fitness is the core basis for guiding the direction of evolution. The dynamic fitness... From basic fitness and novelty fitness Multiplying them together, the formula is as follows: ; in, For dynamic fitness; For novelty fitness; Basic fitness, basic fitness Used to measure this scenario This inherently carries the potential to place the model under test (MUT) into a state of high risk or high uncertainty. In a preferred embodiment, Further breakdown into: ; in: The potential impact It is a scene Rapid assessment of physical hazard levels can be achieved using a Model Under Test (MUT) or a Constraint Surrogate Model (CSM) based on scene gene vectors. A rapid estimate of one or more risk indicators, such as scalar values ​​calculated based on ship kinetic energy, the urgency of the time to minimum encounter (TTCPA), or the congestion of navigable waters.

[0042] The prediction uncertainty It is the Model Under Test (MUT) for the scene Give a prediction The confidence level is measured at that time. In a preferred embodiment, if the Model Under Test (MUT) is built based on a Bayesian neural network, then The prediction variance can be obtained directly. In another implementation, if the model under test (MUT) is a standard neural network, the statistical variance of multiple predictions can be calculated by enabling Dropout during prediction or using an ensemble method, and this variance can be used as the basis for the prediction. The quantized value.

[0043] novelty fitness Used to measure scenarios With failure mode profile The differences among all "known" failure modes are used to drive evolutionary search to avoid discovered failure regions and instead explore new, unknown regions. The novelty fitness is calculated as follows: [The text abruptly ends here, likely due to an incomplete sentence or missing information.] and Each failure scenario stored in Measure the distance between them and take the minimum value: ; in, Indicates the calculation of the current scene With stored failure scenarios The distance between them; Indicates in All of the files In the middle, find The minimum value of the calculation result; For failure mode files; This indicates the scenario currently being evaluated; This indicates a stored failure scenario.

[0044] The system calculates the dynamic fitness of the population for all scenarios. Then, the fitness score set is passed to the next step as the basis for performing the selection operation.

[0045] See attached document Figure 3 In step S3, the adaptive evolutionary generation phase, the system operates on the current population based on the calculated dynamic fitness to generate the next generation population. .

[0046] This phase begins with a selection operation, where individuals are chosen from the current population to form a mating pool based on their dynamic fitness scores. In a preferred embodiment, this selection operation employs a tournament selection strategy, where several individuals are randomly selected for fitness comparison, and the best individual (the one with the highest dynamic fitness) is placed into the mating pool. This process is repeated until the mating pool reaches a preset size.

[0047] An offspring population is generated by applying adaptive genetic operators (including constrained crossover and constrained mutation) to individuals in the mating pool. The core of this process is that the execution of the genetic operators is guided and constrained in real time by a constrained surrogate model (CSM), ensuring that the newly generated offspring individuals always conform to physical laws.

[0048] The execution process of the constraint crossover operator is as follows: The system selects two parent scene gene vectors from the mating pool. and The system applies a crossover operator (e.g., simulated binary crossover SBX or arithmetic crossover) to generate one or more child vectors. For each generated child vector, the system immediately invokes the discriminative function of the Constrained Surrogate Model (CSM). If the discriminant function of the constraint surrogate model returns "false" (0), it indicates that the parameter combination of the child vector (e.g., combining...) The "full tide" characteristics and If the "maximum ebb current velocity" characteristic violates the laws of physics, then the offspring vector is discarded. The system will re-execute the crossover operation (e.g., select a different parent or a different crossover point) until the generated offspring vector satisfies the discrimination function of the constraint surrogate model and returns "true" (1), and only then will the valid offspring be accepted into the offspring population.

[0049] The execution process of the constrained mutation operator is as follows: The system selects offspring vectors (or individuals in the mating pool) for mutation, and determines the first individual to be mutated. One gene Before performing the mutation operation, the system first calls the constraint interval function of the Constraint Surrogate Model (CSM). This constraint interval function will exclude the child vectors from the constraint interval function. Other than The dimensional gene value is used as input, and the result is returned accordingly. Vigene The effective range of values ​​under the current physical constraints .

[0050] right New gene values ​​are generated by applying mutation operators (such as Gaussian mutation or polynomial mutation). The new gene value must be limited to... The returned valid value range is within this range. If a new gene value exceeds this range, it is set as the boundary value of the range (i.e., ...). or This operation ensures the mutation of the scene vector. It remains physically valid and requires no additional verification.

[0051] Through the aforementioned constrained crossover and constrained mutation operations, the system generates a complete offspring population in which all individuals conform to physical laws. And pass it on to the next step.

[0052] During the periodic authoritative verification phase of step S4, the preset authoritative verification cycle is followed. To trigger.

[0053] The system first checks the current generation. Whether the triggering condition is met. In a preferred embodiment, the triggering condition is: That is, the current generation is The number of iterations must be an integer multiple of the given number. If the current generation does not meet this trigger condition, the process will proceed directly to the termination condition check.

[0054] When the triggering condition When the conditions are met, the system performs candidate scenario screening. This screening process involves the system calculating the dynamic fitness of each scenario's gene vector from the current population. Sort in descending order.

[0055] The system selects from the sorting results highest score Each scenario's gene vector constitutes a high-risk candidate set. .in This is the preset number of candidate scenarios for a single verification. The high-risk candidate set... Each candidate scenario (in Submit them to the High Fidelity Simulator (HFS) module and the Model Under Test (MUT) module respectively.

[0056] After receiving candidate scenarios, the High Fidelity Simulator (HFS) module performs high-precision, computationally expensive authoritative simulations and outputs the authoritative risk result for the candidate scenario. .

[0057] After receiving a candidate scenario, the Model Under Test (MUT) module executes its own fast prediction logic and outputs the predicted risk result for that candidate scenario. .

[0058] In one implementation, for The authoritative verification and prediction process for each candidate scenario can be executed in parallel to shorten the time required for the verification phase. The system, after acquiring all... After comparing the authoritative risk results and predicted risk results for each candidate scenario, these results are used to perform failure determination and file updates.

[0059] During the failure determination and file update stage of step S5, the acquired data is processed. The data on authoritative risk outcomes and predicted risk outcomes are processed.

[0060] The failure determination and archive update phase addresses each candidate scenario. (in Calculate its prediction error In a preferred embodiment, the prediction error Through the loss function To quantify, absolute error is used: ; in, Candidate scenarios The prediction error; Authoritative risk results output by the High Fidelity Simulator (HFS); The predicted risk output by the Model Under Test (MUT).

[0061] Perform a failure determination, comparing the calculated prediction error with a preset failure determination threshold. Compare, if the conditions are met If the candidate scenario is determined to be a "confirmed failure mode", the system will summarize the gene vectors of all scenarios determined to be "confirmed failure modes" in the high-risk candidate set to form a failure scenario set.

[0062] Perform an update operation on the Failure Mode and Effects File (CFA), incorporating the set of failure scenarios into the current Failure Mode and Effects File to generate a file for the next generation. New failure mode profile used: ; in, For failure mode files; Indicates a new failure mode profile; This represents a set of failure scenarios.

[0063] This feedback loop forces the Evolutionary Scenario Generator (ESG) to automatically avoid already explored areas and redirect its search resources to other potential, undiscovered failure mode regions in the parameter space. Termination conditions are determined after the failure mode profile is updated to a new one.

[0064] In the termination and output phase of step S6, a determination is made as to whether the iteration process should be terminated. This termination determination specifically includes: Check whether the current generation number is greater than the preset maximum generation number; or check whether the total number of failure scenario gene vectors stored in the updated failure mode file has reached or exceeded the preset target number. In another embodiment, the termination condition may also include determining whether the fitness of the population has converged (e.g., the highest fitness value has not significantly improved for several consecutive generations).

[0065] If any of the above termination conditions are met, the evolutionary iteration loop is terminated. The system then performs a result output operation, outputting the current final failure mode file as the complete result of this failure exploration. The failure mode file contains a set of scene gene vectors that, during the iteration process, were verified by a high-fidelity simulator (HFS) and caused the prediction error of the model under test (MUT) to exceed the failure threshold.

[0066] If the system determines that all termination conditions have not been met, the current generation will be changed. Increment by 1 and update the current failure mode file. Passed on to the next generation, returning to the dynamic fitness assessment phase to perform a new generation of dynamic fitness assessment.

[0067] Although embodiments of the invention have been shown and described, it will be understood by those skilled in the art that various changes, modifications, substitutions and alterations can be made to these embodiments without departing from the principles and spirit of the invention, the scope of which is defined by the appended claims and their equivalents.

Claims

1. An experimental method for predicting the navigation capacity of ships in giant estuaries using machine learning, characterized in that: Includes the following steps: S1. System initialization includes defining scene gene vectors describing the general aviation scenario, training constrained agent models offline, generating an initial scene population, initializing an empty failure mode file, setting the maximum number of generations and the number of target failure modes, and initializing the current generation. S2. Using the initial scene population as the current scene population, start the iterative loop and perform dynamic fitness evaluation on the scene gene vectors in the current scene population to obtain fitness scores. S3. Based on the fitness score, perform selection operations, perform crossover and mutation operations through adaptive genetic operators to generate offspring population, and call the constraint proxy model to ensure that the generated offspring scene gene vector conforms to physical laws; S4. Select the candidate scene with the highest fitness score from the current scene population, and call the high-fidelity simulator and the test model as the test object to perform simulation or prediction respectively, so as to obtain the risk result and prediction risk result of the candidate scene. S5. Calculate the prediction error between the risk result and the predicted risk result. When the prediction error is greater than the preset failure threshold, store the corresponding candidate scenario in the failure mode file. S6. If the current generation number is greater than the maximum generation number or the number of scenarios in the failure mode file reaches the target number of failure modes, the iteration is terminated and the failure mode file is output as the exploration result. If the condition is not met, the current scenario population is replaced by the offspring population, and the dynamic fitness evaluation is returned.

2. The experimental method for predicting the navigation capacity of giant estuaries using a machine learning-integrated model as described in claim 1, characterized in that, In step S1, the constraint proxy model further includes: The discrimination function is a classifier, which receives the complete scene gene vector as input and outputs a Boolean value to determine whether the parameter combination represented by the scene gene vector conforms to the preset physical law. The constraint interval function, which is a regression model or boundary prediction model, is used to determine and give the allowed physical effective range of values ​​for unknown parameters based on known parameter values ​​when the parameter value of a certain dimension in the scene gene vector is unknown.

3. The experimental method for predicting the navigation capacity of giant estuaries using a machine learning-integrated model as described in claim 1, characterized in that, In step S2, the fitness score is calculated based on the following criteria: The uncertainty of the test model's prediction of the scene gene vector as the test object; The prediction uncertainty is quantified by enabling Dropout or employing a model ensemble approach during prediction, and by calculating the statistical variance of multiple prediction results.

4. The experimental method for predicting the navigation capacity of giant estuaries using a machine learning-integrated model as described in claim 1, characterized in that, In step S2, the calculation of the fitness score also includes: The novelty fitness of the scene gene vector and all existing scenes in the failure mode archive; The novelty fitness is used to drive evolutionary search to avoid discovered failure regions; The novelty fitness is calculated by obtaining the distance metric between the scene gene vector and the failure scenes stored in the failure mode archive, and taking the minimum value among them.

5. The experimental method for predicting the navigation capacity of giant estuaries using a machine learning-integrated model as described in claim 4, characterized in that, When the failure mode file has not yet been stored in any failure scenarios, the novelty fitness is a preset constant maximum value.

6. The experimental method for predicting the navigation capacity of giant estuaries using a machine learning-integrated model as described in claim 1, characterized in that, In step S3, the step of the adaptive genetic operator performing the crossover operation further includes: The discrimination function of the constraint proxy model is invoked to ensure that the offspring scene gene vector conforms to physical laws. The invocation process includes: After generating the offspring vector, the discrimination function is called to perform discrimination; When the child vector is determined to be inconsistent with physical laws, the child vector is discarded and the crossover operation is re-executed until the generated child vector is determined to be consistent with physical laws by the discrimination function.

7. The experimental method for predicting the navigation capacity of giant estuaries using a machine learning-integrated model as described in claim 1, characterized in that, In step S3, the step of the adaptive genetic operator performing mutation operation further includes: The constraint interval function of the constraint proxy model is invoked to ensure that the offspring scene gene vector conforms to physical laws. The invocation process includes: Before performing mutation operations on specific genes in the scene gene vector, the constraint interval function is invoked to obtain the physical valid value range of the specific gene. After generating a new gene value, the new gene value is restricted to the physical valid value range. When the new gene value exceeds the physical valid value range, the new gene value is set as the boundary value of the physical valid value range.

8. The experimental method for predicting the navigation capacity of giant estuaries using a machine learning-integrated model as described in claim 1, characterized in that, In step S4, the step of selecting the candidate scene with the highest fitness score further includes: The current scene population is sorted in descending order based on the fitness score; The scene gene vector with the highest score is selected from the ranking results as the candidate scene.

9. The experimental method for predicting the navigation capacity of giant estuaries using a machine learning-integrated model as described in claim 1, characterized in that, In step S5, the calculation step of the prediction error further includes: quantifying the prediction error by calculating the absolute value of the difference between the risk result and the predicted risk result.

10. A test system for predicting the navigation capacity of vessels in giant estuaries by incorporating machine learning, applied to the test method for predicting the navigation capacity of vessels in giant estuaries by incorporating machine learning as described in any one of claims 1-9, characterized in that, include: The scenario definition and interface module is used to define parameters affecting navigation capability as scenario gene vectors and call the tested model as the test object to obtain the predicted risk results of candidate scenarios. A high-fidelity simulator module is used to receive the candidate scenarios and perform simulations to obtain the risk outcomes of the candidate scenarios; The constrained proxy model module provides a discrimination function and a constraint interval function. The discrimination function is used to determine whether the parameter combination in the input scene gene vector conforms to the preset physical law. The constraint interval function is used to obtain the physical valid value interval of the undetermined parameter based on the existing parameters when a certain parameter in the scene gene vector is undetermined. The failure mode file module is used to store the corresponding candidate scenario when the prediction error between the risk result and the predicted risk result is greater than a preset failure threshold, thus forming a confirmed failure mode. The dynamic fitness assessment module is used to calculate the fitness score of the scene gene vector in the current population; The evolutionary scene generator module is used to determine whether the termination condition is met. When the termination condition is not met, the current scene population is replaced with the offspring population, and the dynamic fitness evaluation module is triggered to return to calculate the fitness score.

Images