Method and system for testing launching heat flow field of offshore rocket

Abstract

The invention discloses an offshore rocket launching heat flow field test method and system, and the method comprises the following steps: S1, collecting multi-source heat flow data on an offshore rocket launching platform, and carrying out the synchronous processing, and generating a heat flow test data set; s2, constructing a sparse auto-encoder model; s3, optimizing model structure parameters by utilizing a king butterfly optimization algorithm to obtain an optimal parameter combination; s4, processing the multi-source heat flow data by applying the optimal structure parameters, and outputting a two-dimensional heat flow distribution diagram; and S5, identifying the heat flux area and the potential thermal anomaly, and outputting thermal shock early warning information. According to the method, the sparse auto-encoder and the empire butterfly optimization algorithm are fused, so that high-precision modeling and thermal shock early warning of the maritime rocket launching heat flow field are realized.

Description

Technical Field

[0001] The present invention relates to the technical fields of thermophysical measurement and intelligent space systems, and particularly relates to a method and system for testing the launch heat flux field of a sea rocket. Background Technique

[0002] With the wide application of sea launch platforms in modern space launch missions, how to efficiently and reliably evaluate the heat flux field characteristics during rocket launch in a complex marine environment has become a key issue for enhancing platform safety, mission stability, and structural adaptability. During rocket launch, due to the superposition effects of multiple factors such as engine flame jet action, high-temperature gas reflux, tail flame impact, and platform structure reflection, complex, intense, and highly unstable heat flux responses will be generated in different regions of the platform, even leading to local structural overheating, abnormal thermal stress, and even thermal failure. Therefore, establishing a set of accurate, real-time, and intelligent heat flux field testing and heat response identification methods has become a key research direction in the current fields of rocket launch thermal protection and structural health management.

[0003] Existing heat flux field testing methods mainly include three categories: one is multi-point heat flux measurement based on physical sensors, such as thermocouples, infrared thermometers, heat flux meters, etc.; the second is visualization heat distribution acquisition technology based on thermal imagers or infrared imaging systems; the third is an indirect estimation method combined with simulation modeling, such as finite element heat conduction analysis and CFD (Computational Fluid Dynamics) simulation. However, these traditional methods have several significant problems in the actual sea rocket launch scenario. First, the installation density of physical sensors is limited, and it is impossible to restore the planar heat flux field with high spatial resolution, and the sensors are easily affected by the marine environment, vibration disturbances, and high-temperature damage. Second, although the infrared imaging system can provide a two-dimensional heat map, it relies on manual analysis or traditional filtering methods in data processing, making it difficult to automatically identify heat anomaly regions, and it is significantly affected by smoke and water vapor. Third, although numerical simulation methods have auxiliary significance in the design stage, they are extremely sensitive to boundary conditions and initial conditions, making it difficult to adapt to real-time changing data stream inputs and cannot be used for dynamic monitoring and heat warning at the launch site.

[0004] In recent years, significant progress has been made in structural feature extraction, anomaly recognition, and image modeling using artificial intelligence and deep learning methods. Especially in the field of image reconstruction and sparse representation, autoencoder-based neural networks have demonstrated superior performance. Among them, the sparse autoencoder can more effectively capture the essential structural features in the input data by imposing a sparse constraint on neuron activation, and is particularly suitable for learning the low-dimensional latent representation of heatmap-like data. However, in the existing literature, sparse autoencoders are mostly applied to medical imaging, industrial inspection, and image compression, and their application in rocket heat flux distribution modeling is not yet mature. In particular, there is a lack of an adaptation and optimization mechanism for structural design parameters (such as encoder depth, sparse intensity, path structure, etc.) and task characteristics (such as heat flux concentration, heat spot boundary diffusivity).

[0005] In addition, the parameter configuration of the deep model structure greatly affects its performance, and existing methods mostly rely on manual setting or empirical tuning, lacking the ability of automated and globally optimal structure search. In recent years, a class of swarm intelligence optimization algorithms, such as particle swarm optimization, grey wolf optimization, firefly algorithm, etc., have gradually been used for neural network structure parameter optimization. However, these algorithms are prone to falling into local optima in complex high-dimensional search spaces, and the search strategies lack the hierarchy of biological behavior modeling. In contrast, the monarch butterfly optimization algorithm is a newly emerging bio-inspired optimization algorithm that simulates the migration and survival behavior of monarch butterflies between North America and Central America. It achieves global search and local refinement through the cooperation of migration operations and adjustment operations, and has strong convergence and the ability to jump out of local extrema, making it suitable for the search and optimization of neural network structures with multiple parameters and multiple modules.

[0006] However, there is currently a lack of a technical solution that combines sparse autoencoders with swarm intelligence optimization algorithms for rocket launch heat flux image modeling and heat anomaly region recognition. Existing research either focuses on model construction while ignoring parameter adaptation, or only targets heatmap visualization without intelligent structural analysis. Especially in the environment of a sea-launch platform, the heat flux perturbation is intense, the data is complex and multi-source, and the heat anomaly region is not easy to pre-calibrate in advance, making it difficult for traditional methods to meet comprehensive requirements such as accuracy, stability, and real-time performance. At the same time, in the existing technology, there is a lack of a subsequent physical-meaning-based analysis process for the heat flux maps output by the model. For example, there is a lack of a structural understanding of the heat flux distribution and intelligent early warning means for the heat shock level.

[0007] Therefore, how to provide a test method and system for the launch heat flux field of a sea rocket is an urgent problem that needs to be solved by those skilled in the art. Summary of the Invention

[0008] An object of the present invention is to propose a method and system for testing the launch heat flow field of a sea rocket. The present invention integrates key links such as multi-source heat flow data acquisition, deep feature extraction, swarm intelligence optimization, and heat anomaly recognition and early warning, constructs a sparse autoencoder model with an adaptive structure configuration ability, and uses the monarch butterfly optimization algorithm to globally optimize and search its structural parameters to ensure that the model structure is highly adapted to the heat flow distribution characteristics. On this basis, the system can efficiently restore the two-dimensional heat flow image during the launch process, identify the distribution of heat flux on the platform structure, accurately locate potential heat anomaly areas, and output heat shock early warning information, forming a closed-loop control process for thermal state monitoring and risk warning. This method has the advantages of intelligent adjustable structure, accurate image generation, and timely warning response, significantly improving the safety monitoring and thermal protection capabilities of the sea rocket launch platform in a complex thermal environment.

[0009] A method for testing the launch heat flow field of a sea rocket according to an embodiment of the present invention includes the following steps: S1. Collect multi-source heat flow data on the sea rocket launch platform, and synchronize the multi-source heat flow data according to a unified time stamp to generate a heat flow test data set; S2. Construct a sparse autoencoder model, and use the heat flow test data set as the input. The sparse autoencoder model includes an encoder module, a heat flow attention focusing module, a sparse constraint module, and a decoder module; S3. Use the monarch butterfly optimization algorithm to optimize the structural parameters of the sparse autoencoder model, encode each population individual into a set of structural parameter combinations, and obtain optimized structural parameter combinations; S4. Apply the optimized structural parameters to the sparse autoencoder model, use the optimized sparse autoencoder model to process the multi-source heat flow data, and output two-dimensional heat flow distribution map data; S5. Based on the two-dimensional heat flow distribution map data, identify the heat flux area in the sea rocket launch platform structure, analyze the distribution of heat flux at different spatial positions, determine the location and intensity of potential heat anomaly areas, and output corresponding heat shock early warning information.

[0010] Optionally, the multi-source heat flow data specifically includes an infrared image sequence, a thermocouple measurement point heat flux value, environmental wind speed and temperature data, rocket attitude parameters, and platform structure state information.

[0011] Optionally, S2 specifically includes: S21. Set the heat flow test data set as the input matrix , where is the set of real numbers, represents the number of samples, Indicates the thermal flux feature dimension of each sample, including infrared image sequence, thermocouple measurement point heat flux value, ambient wind speed and temperature data, rocket attitude parameters, and platform structure status information; S22. Construct the encoder module of the sparse autoencoder model, which adopts a multi-path parallel structure, including a full-image perception path, a local hot area path, and an edge detail path, to process the whole image, local hot area, and edge contour information respectively. The full-image perception path outputs a low-frequency feature map, the local hot area path outputs a medium-frequency feature map, and the edge detail path outputs a high-frequency feature map. The stacking depth of the convolutional layers of each path is controlled by the multi-scale path fusion depth factor ; S23. Set the kernel size, stride, and normalization parameters of each convolutional path, and splice and fuse the low-frequency feature map, medium-frequency feature map, and high-frequency feature map to form a latent feature representation ; S24. Construct a thermal flux attention focusing module, introduce channel attention and spatial attention mechanisms to the latent feature representation . Set the spatial attention diffusion radius as the attention diffusion radius control coefficient , and generate a weighted feature representation according to the attention response ; S25. Construct a sparse constraint module, set the sparse target value , and introduce a regional sparse adjustment factor. According to the response characteristics of different input feature regions, dynamically adjust the activation and inhibition intensity of each hidden neuron according to the regional sparse distribution coefficient to form an adjustable regional sparse control strategy: S26. Introduce an inter-layer sparse adjustment unit between the encoder module, the thermal flux attention focusing module, and the sparse constraint module to monitor the activation output of each hidden layer according to the regional feature importance, and dynamically adjust the sparse intensity layer by layer according to the regional sparse distribution coefficient; S27. Construct a decoder module, input the weighted feature representation into the decoding path, and adopt a combination structure of transposed convolution and upsampling to restore the output dimension and generate output data ; S28. Define the set of structural parameters of the sparse autoencoder model, including the number of layers of the encoder and decoder, the number of neurons in each layer, the sparse target value , the regional sparse distribution coefficient , the multi-scale path fusion depth factor , the attention diffusion radius control coefficient ; S29. Complete the structural initialization of the sparse autoencoder model, and load the current parameter configuration for pre-training.

[0012] Optionally, the specific content of S3 includes: S31. Initialize the population size of the monarch butterfly optimization algorithm and divide it into a northern migration sub-population and a southern regulation sub-population; S32. Represent each monarch butterfly individual using a structure-aware coding method, and encode the monarch butterfly individual as : ; Among them, represents the structure diagram of the sparse autoencoder model, defining the module connection relationship, enabling status, and path layout, represents the corresponding combination of structure parameters, including the number of layers of the encoder and decoder, the number of neurons in each layer, the global sparsity target value , the regional sparse distribution coefficient , the multi-scale path fusion depth factor , the attention diffusion radius control coefficient ; S33. Initialize and pre-train the structure of the sparse autoencoder model corresponding to each monarch butterfly individual, and use the heat flow test data set to evaluate the reconstruction ability and the high-heat area response ability; S34. Define the fitness function of the monarch butterfly individual , combined with the reconstruction error of the heat flow image, the reconstruction error of the high-heat area and the structural complexity ; Among them, is the weighting coefficient; S35. Calculate the structural covariance entropy diversity index of the current generation population: ; Among them, represents the value of the th individual in the th structural parameter dimension, is the population average of the dimension, is the population size, is the structural parameter dimension, is a small positive number, is the logarithmic function; S36. Adopt a non-linear migration probability adjustment mechanism based on the diversity index to calculate the migration probability of the th generation: ; Among them, is the steepness adjustment factor, is the balance threshold value, is an exponential function; generate a random number for each individual , if it satisfies , then mark the individual as a migration candidate individual; S37. Perform multi-center attraction update on the migration candidate individuals. Let the set of guiding centers be , where is the combination of structural parameters of the j-th guiding center individual, , is the number of migration guiding centers. Update the combination of structural parameters based on the fitness entropy weight as : ; Among them, is a random scaling factor, is the guiding weight of the j-th guiding center, is the combination of structural parameters of the j-th guiding center individual, is the guiding weight of the k-th guiding center; S38. Perform a regulation operation on the individuals in the sub-population that have not participated in migration. Dynamically control the perturbation amplitude according to the distance from the current optimal individual, and update the structural parameter vector as : ; Among them, is the reference perturbation amplitude, is the maximum individual distance in the current generation population, is a stability constant, is the current combination of structural parameters of the i-th individual, is the complete coding of the individual with the optimal fitness in the current generation; S39. Evaluate the fitness values of all newly generated individuals, and select the top q% of the optimal individuals to form an elite set for retention; S310. Determine whether the iteration termination condition is satisfied, including reaching the maximum number of iterations or the fitness change of the optimal individual being lower than the set threshold in several consecutive generations; S311. If the termination condition is not satisfied, then combine the elite individuals and the remaining updated individuals to form the next generation population, and return to step S33; S312. If the termination condition is satisfied, then output the individual with the optimal fitness , where is the structure diagram coding corresponding to the individual with the optimal fitness, is the optimal combination of structural parameters of the sparse autoencoder model.

[0013] Optionally, the S4 specifically includes: S41. The obtained optimal combination of structural parameters Applied to the sparse autoencoder model, decompose the optimal structural parameter combination into , where represents the optimal number of encoder layers, represents the optimal number of decoder layers, represents the optimal number of neurons per layer, is the optimal sparsity target value, is the regional sparse distribution coefficient, is the multi-scale path fusion depth factor, is the attention diffusion radius control coefficient; S42. Configure the parallel convolution path structure of the multi-scale encoder module of the sparse autoencoder according to the optimal number of encoder layers and the multi-scale path fusion depth factor, set the number of convolution layers and receptive field of each path, and set the diffusion window size of the spatial attention convolution according to the attention diffusion radius control coefficient; S43. Configure the density of all neurons based on the optimal number of neurons per layer, and set the sparse activation constraint weights for different heat flux regions according to the regional sparse distribution coefficient; S44. Construct the sparse autoencoder model structure based on the optimal structural parameter combination , expressed as a function mapping , where represents the input heat flux test data set, is the latent representation feature tensor; S45. Input the newly collected multi-source heat flux data during the sea rocket launch into the optimized sparse autoencoder model, perform the encoding-focusing-decoding process, and obtain the two-dimensional heat flux feature map ; S46. Normalize and filter the reconstructed output of the optimized sparse autoencoder model, and output the two-dimensional heat flux distribution map data.

[0014] Optionally, the specific steps of S5 include: S51. Obtain the output two-dimensional heat flux distribution map data, which contains the calorific value distribution information at different spatial positions; S52. Utilize the calorific value gradient change feature to extract the heat response mutation region from the two-dimensional heat flux distribution map data, and combine the launch platform structure layout information to identify the heat flux regions in the sea rocket launch platform structure, covering the rocket flame impact area, the ground reflection area, and the wake flow around the flame area; S53. Conduct spatial modeling analysis on the identified heat flux regions, evaluate the coverage range, relative position relationship, and calorific value density distribution of different heat flux regions on the platform structure, and determine the distribution characteristics and propagation trends of each heat flux region in the horizontal and vertical directions; S54. Based on the spatial modeling results and calorific value characteristics, conduct a focused analysis on the phenomenon of abnormal temperature concentration in the heat flux region to determine the location of potential thermal anomaly regions in the platform structure and the level of thermal response intensity; S55. Compare and verify the determined thermal anomaly regions with the two-dimensional heat flow distribution map data, evaluate the consistency performance of the thermal anomaly regions in the actual heat flow response, and form a thermal shock determination result; S56. Generate corresponding thermal shock warning information according to the thermal shock determination result, clarify the warning level, influence range, and structural risk regions, and assist in the pre-launch structural thermal safety assessment and risk prompts during the mission execution process.

[0015] A launch heat flow field test system for a sea rocket according to an embodiment of the present invention includes the following modules: A data acquisition module for collecting multi-source heat flow data of a sea rocket launch platform, completing time synchronization and preprocessing, and constructing a heat flow data set; A model construction module for constructing a sparse autoencoder model; A feature extraction module for inputting the heat flow data set into the sparse autoencoder model to extract the potential features of the two-dimensional heat flow distribution; A structure optimization module for globally optimizing the structural parameters of the sparse autoencoder model; A heat flow generation module for processing multi-source heat flow data based on the optimized sparse autoencoder model and outputting two-dimensional heat flow distribution map data; A heat flux identification module for analyzing the two-dimensional heat flow distribution map data to identify the heat flux regions and potential thermal anomalies in the platform structure; A warning module for outputting thermal shock warning information and marking the abnormal regions and risk levels.

[0016] The beneficial effects of the present invention are: The present invention breaks through the limitations of the existing offshore rocket launch heat flux field testing technology in aspects such as inaccurate feature extraction, fixed model structure, low response efficiency, and lack of intelligent early warning by introducing a deep modeling strategy that combines a sparse autoencoder and a monarch butterfly optimization algorithm. The sparse autoencoder model constructed using a structure-aware encoding mechanism can extract potential information with spatial distribution characteristics from multi-source heat flux data to form a more discriminative two-dimensional heat flux map; the structure parameters of the model are globally optimized through the monarch butterfly optimization algorithm, enabling core configurations such as the encoding depth, sparsity, and neuron density to be adaptively adjusted according to the characteristics of the heat flux field, enhancing the model's expressive ability and generalization performance. The present invention not only realizes the high-precision restoration of heat flux field images, but also further combines the map data to model and analyze the heat flux regions in the structure of the offshore rocket launch platform, can effectively identify the location and intensity of potential heat anomaly regions, and automatically output heat shock early warning information, thus realizing the full-process closed-loop monitoring and control from data collection, model construction, image generation to risk warning. Compared with existing methods, the present invention has significant advantages such as intelligent structure configuration, high accuracy of image generation, strong automation of heat risk identification, and applicability to dynamic and complex launch environments. Description of the Drawings

[0017] The drawings are used to provide a further understanding of the present invention and constitute a part of the specification. They are used together with the embodiments of the present invention to explain the present invention and do not constitute a limitation to the present invention. In the drawings: Figure 1 is a flowchart of a method and system for testing the launch heat flux field of an offshore rocket proposed by the present invention; Figure 2 is a flowchart for optimizing the structure parameters of the monarch butterfly optimization algorithm of a method and system for testing the launch heat flux field of an offshore rocket proposed by the present invention. Detailed Embodiments

[0018] Now, the present invention will be further described in detail with reference to the drawings. These drawings are all simplified schematic diagrams, only illustrating the basic structure of the present invention in a schematic manner, so they only show the components related to the present invention.

[0019] Refer to Figure 1 , a method for testing the launch heat flux field of an offshore rocket, including the following steps: S1. Collect multi-source heat flux data on the offshore rocket launch platform and synchronize the multi-source heat flux data according to a unified time stamp to generate a heat flux test data set; S2. Construct a sparse autoencoder model, using the heat flux test data set as input. The sparse autoencoder model includes an encoder module, a heat flux attention focusing module, a sparse constraint module, and a decoder module; S3. Optimize the structural parameters of the sparse autoencoder model using the monarch butterfly optimization algorithm, encode each population individual as a set of structural parameter combinations, and obtain the optimized structural parameter combinations; S4. Apply the optimized structural parameters to the sparse autoencoder model, use the optimized sparse autoencoder model to process the multi-source heat flux data, and output the two-dimensional heat flux distribution map data; S5. Based on the two-dimensional heat flux distribution map data, identify the heat flux regions in the structure of the offshore rocket launch platform, analyze the distribution of heat flux at different spatial positions, determine the location and intensity of potential heat anomaly regions, and output the corresponding heat shock warning information.

[0020] In this invention, by constructing a sparse autoencoder model integrated with the monarch butterfly optimization algorithm, intelligent modeling and risk identification of the heat flux data of the offshore rocket launch platform are realized. Compared with the defects of traditional heat flux testing methods, such as low density of physical sensor layout, large response delay, and manual analysis dependence for image processing, this invention can automatically extract heat flux features based on multi-source data through the optimized model structure, effectively generate a high-precision two-dimensional heat flux distribution map. Using the output results of the model to further identify the heat flux regions in the platform structure and analyze the aggregation law of heat values at different spatial positions, it can accurately judge the location and intensity of potential heat anomaly regions, output heat shock warning information, and realize dynamic monitoring and intelligent warning of the heat response state. The overall solution has the advantages of strong structure adaptability, high image expression accuracy, and high degree of automation in risk identification, is suitable for platform heat safety guarantee in complex launch environments, and has good engineering practicability and deployment value.

[0021] In this embodiment, the multi-source heat flux data specifically includes infrared image sequences, thermocouple measurement point heat flux values, ambient wind speed and temperature data, rocket attitude parameters, and platform structure state information.

[0022] In this embodiment, the S2 specifically includes: S21. Set the heat flux test data set as the input matrix , where is the set of real numbers, represents the number of samples, represents the heat flux feature dimension of each sample, including infrared image sequences, thermocouple measurement point heat flux values, ambient wind speed and temperature data, rocket attitude parameters, and platform structure state information; S22. Construct the encoder module of the sparse autoencoder model, which adopts a multi-path parallel structure, including a full-map perception path, a local hot-spot path, and an edge detail path, to process the information of the whole map, local hot spots, and edge contours respectively. The full-map perception path outputs a low-frequency feature map, the local hot-spot path outputs a medium-frequency feature map, and the edge detail path outputs a high-frequency feature map. The stacking depth of the convolutional layers in each path is controlled by the multi-scale path fusion depth factor ; S23. Set the kernel size, stride, and normalization parameters of each convolutional path, and splice and fuse the low-frequency feature map, medium-frequency feature map, and high-frequency feature map to form a latent feature representation ; S24. Construct a heat flow attention focusing module, introduce channel attention and spatial attention mechanisms into the latent feature representation , set the spatial attention diffusion radius as the attention diffusion radius control coefficient , and generate a weighted feature representation according to the attention response ; S25. Construct a sparse constraint module, set the sparse target value , introduce a regional sparse adjustment factor, and according to the response characteristics of different input feature regions, dynamically adjust the activation and inhibition intensity of each hidden neuron according to the regional sparse distribution coefficient to form an adjustable regional sparse control strategy: S26. Introduce an inter-layer sparse adjustment unit between the encoder module, the heat flow attention focusing module, and the sparse constraint module to monitor the activation output of each hidden layer according to the regional feature importance, and dynamically adjust the sparse intensity layer by layer according to the regional sparse distribution coefficient; S27. Construct a decoder module, input the weighted feature representation into the decoding path, adopt a combination structure of transposed convolution and upsampling to restore the output dimension, and generate output data ; S28. Define the set of structural parameters of the sparse autoencoder model, including the number of layers of the encoder and decoder, the number of neurons in each layer, the sparse target value , the regional sparse distribution coefficient , the multi-scale path fusion depth factor , the attention diffusion radius control coefficient ; S29. Complete the structural initialization of the sparse autoencoder model, and load the current parameter configuration for pre-training.

[0023] Through the systematic design and module-level improvement of the sparse autoencoder model structure, the present invention realizes the efficient multi-scale extraction and regional sparse-aware expression of heat flow image features. By adopting a multi-path parallel encoder structure to process the whole image, local heat regions, and edge detail information respectively, and flexibly controlling the convolution depth with a multi-scale path fusion depth factor, the model has good hierarchical perception ability. By introducing channel attention and spatial attention mechanisms, the model can adaptively focus on key heat regions and effectively express hot spot regions through the control of the attention diffusion radius. For the heat feature responses in different regions, the present invention sets a regional sparse adjustment factor, combined with an inter-layer sparse adjustment mechanism, to realize the dynamic control of the activation intensity of hidden neurons, improving the model's feature retention ability for high-heat regions and compression ability for low-correlation regions. The sparse autoencoder constructed by this method not only has adjustable structural parameters and strong perception ability but also takes into account the reconstruction accuracy of global and local heat features, laying an efficient expression foundation for subsequent heat flow map generation and heat anomaly recognition, and having the significant advantages of stable modeling performance, strong optimizability, and high adaptability.

[0024] In this embodiment, the specific steps of S3 are as follows: S31. Initialize the population size of the monarch butterfly optimization algorithm and divide it into a northern migration sub-population and a southern adjustment sub-population; S32. Represent each monarch butterfly individual using a structure-aware coding method, and encode the monarch butterfly individual as : ; Among them, represents the structure diagram of the sparse autoencoder model, defining the module connection relationship, enabling status, and path layout, represents the corresponding combination of structural parameters, including the number of layers of the encoder and decoder, the number of neurons in each layer, the global sparse target value , the regional sparse distribution coefficient , the multi-scale path fusion depth factor , the attention diffusion radius control coefficient ; S33. Initialize the structure and pre-train the sparse autoencoder model corresponding to each monarch butterfly individual, and use the heat flow test data set to evaluate the reconstruction ability and high-heat region response ability; S34. Define the fitness function of the monarch butterfly individual, combined with the reconstruction error of the heat flow image, the reconstruction error of the high-heat region, and the structural complexity : Among them, is the weighting coefficient; Fitness function Its practical significance lies in comprehensively considering the overall reconstruction error of the heat flux image, the local reconstruction error of the high-heat region, and the model structure complexity to comprehensively measure the expression ability and resource efficiency of different model structures in the heat flux image modeling task. The three terms in the formula respectively reflect the performance of the model in overall reconstruction accuracy, key heat feature retention ability, and computational resource consumption; through adjustment by the weighting coefficient, the optimization process can focus on accuracy or complexity control according to application requirements, thus achieving a dynamic balance between performance and efficiency. Specifically, the smaller the reconstruction error, the more accurate the model is in restoring the heat flux image; the smaller the error in the high-heat region, the stronger the model's ability to capture important heat spot information; and the lower the structure complexity, the more concise the model and the higher the inference efficiency. As the optimization objective, this fitness function enables the monarch butterfly algorithm to effectively guide the search process to converge to the optimal structure combination with high accuracy, fast response, and reasonable resources.

[0025] S35. Calculate the structural covariance entropy diversity index of the current generation population : ; Among them, represents the value of the th individual in the th structural parameter dimension, is the population average of the dimension, is the population size, is the structural parameter dimension, is a small positive number, is the logarithmic function; The formula for the structural covariance entropy index used to measure the diversity of the current population. Its practical significance lies in calculating the variance of each structural parameter dimension in the entire population and introducing the form of the entropy function for measurement, thereby reflecting the breadth and dispersion of the distribution of the model structure parameters. This index not only considers the degree of change of each parameter dimension but also strengthens the response to the high-dispersion state through the logarithmic term, enabling the dynamic evaluation of whether the population tends to converge or fall into a local optimum during the optimization process. Specifically, when the structural parameters of each individual in the population are highly similar, the covariance entropy value is low, indicating that the search space shrinks and the individual diversity decreases, and it may be necessary to increase the exploration intensity; conversely, when the entropy value is high, it indicates that the population structure has rich diversity, which helps to maintain the global nature of the search and jump out of local extrema. As the basis for the migration probability adjustment mechanism, this index enables the algorithm to adaptively adjust the search strategy, achieve stable convergence from early wide-area exploration to late-stage, effectively improve the efficiency and robustness of the optimization, and is a key means to ensure the global search ability and optimization quality.

[0026] S36. Adopt a non-linear migration probability adjustment mechanism based on the diversity index to calculate the Migration probability of the generation :[[]] ; Among them, is the steepness adjustment factor, is the balance threshold, is the exponential function; generate a random number for each individual , if it satisfies , then mark the individual as a migration candidate individual; The non-linear migration probability adjustment function used to dynamically control the migration probability, its practical significance lies in, according to the structural covariance entropy diversity index of the current population, calculate the migration probability through the Sigmoid function form, so as to realize the adaptive adjustment of the population behavior state. This mechanism can flexibly decide whether an individual participates in the migration operation according to the change of population diversity, so as to effectively balance the strategy selection between global search and local development. When is relatively high, it means that the distribution of population structure parameters is wide and the individual differences are large, and the migration probability is suppressed, encouraging to retain the current exploration direction; when is relatively low, it indicates that the population tends to converge and the degree of structural homogenization is high, and the migration probability increases, promoting individuals to perform jump updates to multiple centers, so as to jump out of the local extreme value trap. This formula controls the response sensitivity and turning point of the change of migration probability through the steepness factor and the threshold parameter , ensuring to dynamically adjust the optimization behavior at different evolutionary stages, and is one of the important mechanisms to realize the adaptive search strategy of the swarm intelligence algorithm.

[0027] S37. Perform multi-center attraction update on the migration candidate individuals. Let the set of guiding centers be , among which is the combination of structural parameters of the j-th guiding center individual, , is the number of migration guiding centers, and update the combination of structural parameters based on the fitness entropy weight as :[[]] ; Among them, is the random scaling factor, is the guiding weight of the j-th guiding center, is the combination of structural parameters of the j-th guiding center individual, is the guiding weight of the k-th guiding center; Update the combination of structural parameters The formula reflects the multi-center attraction update mechanism introduced in the optimization process of the present invention. Its core idea is that instead of allowing individuals to evolve only towards a single optimal solution, multiple structural configurations with better performance are used as guiding centers to form a more directional and diverse parameter update path. In each optimization iteration, the algorithm constructs weights through entropy according to the performance scores of multiple guiding centers, and these weights reflect the importance of each center in the current optimization stage. Subsequently, candidate individuals will adjust their parameters under the guidance of these centers, and different guiding centers will have different degrees of influence on the individuals, so that the individual update direction integrates multiple advantageous features and avoids falling into local optima. In addition, by introducing a random adjustment factor within a certain range, each parameter update has appropriate perturbation, enhancing the exploration ability of the search. This mechanism effectively improves the flexibility and stability of the optimization path, makes the structure search more extensive and in-depth, can find a better configuration scheme in the sparse autoencoder structure space, and ultimately improves the overall effect of the heat flow image modeling and the adaptability of the model structure.

[0028] S38. Perform an adjustment operation on the individuals in the sub-population that have not participated in migration, dynamically control the perturbation amplitude according to the distance from the current optimal individual, and update the structure parameter vector to : ; where is the reference perturbation amplitude, is the maximum individual distance in the current generation population, is the stability constant, is the current structural parameter combination of the i-th individual, is the complete coding of the individual with the optimal fitness in the current generation; Update the structure parameter vector The formula is a local perturbation update strategy implemented by the present invention for individuals not participating in the migration operation during the optimization process, aiming to enhance the local search ability and maintain the diversity of the population structure. The practical significance of this formula lies in: by measuring the difference between the current individual and the global optimal individual, adaptively adjusting the perturbation amplitude of its structural parameters according to the size of the difference, so as to achieve a dynamic adjustment mechanism of "taking big steps to explore when the distance is far, and making fine adjustments with small steps when the distance is close". When an individual differs greatly from the optimal structure configuration, the system will assign it a strong perturbation amplitude to guide it to quickly jump out of the current state and search for new structural possibilities; while when an individual is close to the optimal structure, a small perturbation amplitude is used for fine-tuning to improve the accuracy and reduce the structural fluctuation. This mechanism realizes the controllability and adaptability of the perturbation process by introducing the relative distance from the optimal solution as a regulation factor and combining the basic perturbation intensity with the stability parameter. Overall, this perturbation strategy based on the difference in the position state of individuals helps the algorithm improve the local convergence quality while maintaining the search activity, and is a key step to improve the optimization accuracy and stability of the final sparse autoencoder structural parameters.

[0029] S39. Evaluate the fitness values of all newly generated individuals, and select the top q% of the optimal individuals to form an elite set for retention; S310. Determine whether the iteration termination condition is satisfied, including reaching the maximum number of iterations or the fitness change of the optimal individual being lower than the set threshold in several consecutive generations; S311. If the termination condition is not satisfied, then merge the elite individuals and the remaining updated individuals to form the next generation population, and return to step S33; S312. If the termination condition is satisfied, then output the individual with the optimal fitness , where is the structure diagram encoding corresponding to the individual with the optimal fitness, is the optimal structure parameter combination of the sparse autoencoder model.

[0030] In the present invention, the monarch butterfly optimization algorithm is introduced to adaptively search for and globally optimize the structural parameters of the sparse autoencoder model, effectively overcoming the defects of traditional deep models that rely on manual configuration of structural parameters and are prone to falling into local optima. By constructing a structure-aware individual coding method, key parameters such as the number of encoder layers, the number of neurons, the sparse control factor, the depth of the multi-scale path, and the attention diffusion range are jointly optimized to ensure a high degree of matching between the model structure and the characteristics of the heat flux data. The covariance entropy index is used to evaluate the population diversity, and combined with the non-linear migration probability adjustment mechanism and the multi-center fitness entropy weight guidance strategy, the algorithm has the ability of dynamic search adjustment and local fine exploration. The perturbation amplitude between individuals is dynamically adjusted according to the structural distance from the optimal individual, improving the search efficiency and stability. The optimized structural parameters not only improve the reconstruction accuracy of the model and the perception ability of the high-temperature region, but also enhance the adaptability of the model to the complex heat flux data structure, realizing the optimal configuration of the structure for heat flux map construction, and having the beneficial effects of strong search intelligence, high convergence efficiency, and good result stability.

[0031] In this embodiment, step S4 specifically includes: S41. Apply the obtained optimal structural parameter combination to the sparse autoencoder model, and decompose the optimal structural parameter combination into , where represents the optimal number of encoder layers, represents the optimal number of decoder layers, represents the optimal number of neurons per layer, is the optimal sparsity target value, is the regional sparse distribution coefficient, is the multi-scale path fusion depth factor, is the attention diffusion radius control coefficient; S42. Configure the parallel convolution path structure of the multi-scale encoder module of the sparse autoencoder according to the optimal number of encoder layers and the multi-scale path fusion depth factor, set the number of convolution layers and the receptive field of each path, and set the diffusion window size of the spatial attention convolution according to the attention diffusion radius control coefficient; S43. Configure the density of all neurons based on the optimal number of neurons per layer, and set the sparse activation constraint weights for different heat flux regions according to the regional sparse distribution coefficient; S44. Construct a sparse autoencoder model structure based on the optimal structural parameter combination , expressed as a function mapping , where represents the input heat flux test data set, is the latent representation feature tensor; S45. Input the multi-source heat flux data during the offshore rocket launch process newly collected into the optimized sparse autoencoder model, and perform the encoding-focusing-decoding process to obtain a two-dimensional heat flux feature map. ; S46. Normalize and filter the reconstruction output of the optimized sparse autoencoder model, and output the two-dimensional heat flux distribution map data.

[0032] In the present invention, the optimal structural parameter combination obtained by the monarch butterfly optimization algorithm is accurately applied to the sparse autoencoder model to construct an optimal network structure that highly matches the heat flux characteristics. This method not only ensures the effective configuration of structural factors such as the number of encoder layers, neuron density, and multi-scale path depth, but also realizes targeted control in terms of the attention mechanism and sparse strategy, enabling the model to have stronger expression and compression capabilities when facing heat flux regions with different scales and different response characteristics. Through the collaborative setting of the optimal attention diffusion radius and regional sparse distribution coefficient, the focusing effect of the model on high heat density regions and the constraint effect on low-correlation regions are enhanced. The model structure is constructed in the form of function mapping, which is convenient for system integration and deployment. In practical applications, the model can accurately restore the two-dimensional heat flux distribution during the offshore rocket launch process, with high-quality reconstructed images and clear detail expression, and can normalize the output results of the model to improve the consistency and interpretability of the data. The overall solution has the advantages of high structural matching degree, strong output accuracy, and good reconstruction stability, providing high-quality heat map data support for the subsequent heat flux identification and warning module.

[0033] In this embodiment, the specific steps of S5 are as follows: S51. Obtain the output two-dimensional heat flux distribution map data, which contains the calorific value distribution information at different spatial positions. S52. Utilize the calorific value gradient change characteristics to extract the heat response mutation regions from the two-dimensional heat flux distribution map data, and combine the structural layout information of the launch platform to identify the heat flux regions in the offshore rocket launch platform structure, including the rocket flame impact area, ground reflection area, and wake flow area. S53. Conduct spatial modeling analysis on the identified heat flux regions, evaluate the coverage range, relative position relationship, and calorific value density distribution of different heat flux regions on the platform structure, and determine the distribution characteristics and propagation trends of each heat flux region in the horizontal and vertical directions. S54. Based on the spatial modeling results and calorific value characteristics, conduct focused analysis on the phenomenon of abnormal concentration of temperature in the heat flux regions to determine the location and heat response intensity level of potential heat anomaly regions in the platform structure. S55. Compare and verify the determined heat anomaly regions with the two-dimensional heat flux distribution map data, evaluate the consistency performance of the heat anomaly regions in the actual heat flux response, and form a heat shock determination result. S56. Generate corresponding thermal shock warning information according to the thermal shock determination result, clarify the warning level, influence range and structural risk area, and assist in the structural thermal safety assessment before launch and risk prompts during the mission execution process.

[0034] The present invention constructs a complete thermal flux region recognition and thermal shock warning analysis process based on the two-dimensional heat flux distribution map data output by the optimized model, significantly improving the response perception ability and safety guarantee level of the offshore rocket launch platform in a complex thermal environment. By using the characteristics of sudden change in heat value gradient and combining the platform structure layout information, this method can accurately identify typical thermal flux regions such as the rocket flame impact area, ground reflection area and wake flow around the flame, and quantify the thermal coverage range, relative position relationship and propagation trend of each region based on the spatial modeling method. On this basis, the system further focuses on the region with abnormal temperature concentration, comprehensively evaluates its position, intensity and spatial continuity, and scientifically determines the risk level of thermal shock. Through the comparison and verification of the abnormal region and the heat flow map, the accuracy and credibility of the warning are effectively improved. The finally output thermal shock warning information has the characteristics of hierarchical prompt, risk positioning and strong platform adaptability, and can provide strong support for the structural thermal safety assessment before launch and risk prompts during the launch process. The overall method has the beneficial effects of high recognition accuracy, strong response timeliness and good engineering applicability.

[0035] Reference Figure 2 , a thermal flow field test system for an offshore rocket, includes the following modules: A data acquisition module, which is used to collect multi-source heat flow data of the offshore rocket launch platform, complete time synchronization and preprocessing, and construct a heat flow data set; A model construction module, which is used to construct a sparse autoencoder model; A feature extraction module, which is used to input the heat flow data set into the sparse autoencoder model to extract the latent features of the two-dimensional heat flow distribution; A structure optimization module, which is used to globally optimize the structural parameters of the sparse autoencoder model; A heat flow generation module, which is used to process multi-source heat flow data based on the optimized sparse autoencoder model and output two-dimensional heat flow distribution map data; A thermal flux recognition module, which is used to analyze the two-dimensional heat flow distribution map data and identify the thermal flux regions and potential thermal anomalies in the platform structure; A warning module, which is used to output thermal shock warning information and mark the abnormal region and risk level.

[0036] Example 1: To verify the feasibility of the present invention in implementation, the present invention is applied to the launch mission of a certain liquid carrier rocket on a certain offshore mobile platform as the application scenario. The platform is equipped with various types of heat flux sensors and infrared image acquisition devices, which can record in real time the relevant heat flux data such as the flame jet, the reflection of the tail flame, and the thermal response of the platform structure during the rocket launch process. The duration of the launch mission is 163 seconds, involving multiple working stages, including ignition, takeoff, passing through the platform section, and the stage of pulling the tail flame away, which are key periods with significant changes in the heat load distribution.

[0037] In this embodiment, the system collected the heat flux change data of 124 measurement points through the thermocouple sensors pre - arranged in areas such as the platform deck, the reflection wall, and the support arm. At the same time, an infrared image sequence with a resolution of 512×512 was obtained through a dual - band infrared camera. In addition, the platform environmental data (temperature, humidity, wind speed) and the rocket attitude parameters (pitch angle, nozzle direction vector, etc.) were also synchronously collected as the multi - source input features of the model.

[0038] First, the data acquisition module was used to perform unified timestamp alignment and normalization processing on the above - mentioned multi - source data to generate a heat flux test data set for training. In the model construction stage, the sparse auto - encoder model adopted a three - path parallel encoder structure to process the full - map information, the local high - temperature area, and the edge details respectively, and configured an attention mechanism for spatial focus extraction. In the optimization stage, the monarch butterfly optimization algorithm was introduced to globally search and optimize the structural parameters through 200 rounds of evolution process to obtain the optimal model configuration. Its encoder has 4 layers, and the number of single - layer neurons is 128, 256, 128, and 64 respectively. The optimal sparsity target is 0.18, the regional sparsity factor is 0.26, the multi - scale fusion depth factor is set to 3, and the attention diffusion control coefficient is 4.5.

[0039] Subsequently, the optimized sparse auto - encoder model was used to process the newly collected heat flux data during the launch process. The system can generate a two - dimensional heat flux distribution map for each frame within 2.3 seconds, and the error between the image output result and the measured value of the sensor has decreased significantly. In the heat anomaly analysis stage, the model automatically identified the high - heat - flux areas on the platform structure and successfully located three areas with sudden increases in heat values, namely the flame - jet impact area (about 8.6 meters from the platform center), the tail - flame flow - around area (the left - rear edge of the platform), and the strong ground - reflection tropical area (the bottom of the platform reflector), and gave early warning prompt levels of Ⅲ, Ⅱ, and Ⅳ respectively. The system output to the mission control terminal in a graphic and text manner.

[0040] Compared with the traditional method based on infrared imaging + manual interpretation, this method has obvious advantages in terms of response time, modeling accuracy, and thermal anomaly recognition accuracy. Under actual measurement conditions, the average reconstruction error of the model for infrared images is 3.26%, and the local reconstruction error in the high-temperature area is 2.01%, which is about 31% higher than that of the unoptimized sparse autoencoder. At the same time, the positioning accuracy of the thermal anomaly area has increased from 72% of the original manual annotation method to 91%, and the average advance time of the system's automatically generated warning prompt is 6.8 seconds, significantly improving the ability to predict the structural thermal safety during rocket launch.

[0041] Table 1 Performance comparison of different methods in the heat flux test task

[0042] From the perspective of the overall accuracy of heat flux image modeling, the method of the present invention performs optimally in terms of both average reconstruction error and local reconstruction error in the high-temperature area. Specifically, the average error in the overall thermal image reconstruction of the present invention is only 3.26%, which is about 1.45 percentage points lower than that of the traditional sparse autoencoder and more than 2 percentage points lower than that of the convolutional neural network method, fully indicating that the optimized structural parameter configuration is more suitable for the distribution characteristics of heat flux images, especially in maintaining details in key thermal areas such as high-temperature flame jets and tail flame reflections, with stronger expression ability. The error in the high-temperature area is only 2.01%, which is of great significance in the actual rocket launch safety analysis.

[0043] In terms of response speed, the single-frame processing time of the method of the present invention is only 2.3 seconds, saving about 2.5 seconds and 1.2 seconds respectively compared with the convolutional neural network and the traditional unoptimized method, with obvious real-time advantages. The traditional manual image analysis method takes longer, usually exceeding 20 seconds, and obviously it is difficult to meet the requirements of real-time monitoring and early warning of the task.

[0044] From the perspective of thermal anomaly recognition and early warning capabilities, the recognition accuracy of the present invention is as high as 91.2%, far superior to 72.3% of the traditional manual analysis and better than the 80% - 85% range of other deep learning models. The system can also achieve an average early warning time of 6.8 seconds, effectively helping operators make intervention judgments before the heat load increases, while the traditional manual method lacks the ability of early perception.

[0045] In terms of false alarm control, the regional false alarm rate of the present invention is controlled at 4.6%, which is significantly lower than 9.2% and 14.7% of the traditional sparse autoencoder and manual analysis, indicating that the model has better regional thermal feature discrimination ability, and the warning information is more reliable, suitable for the variable and complex thermal environment during the launch mission.

[0046] Based on the above performance indicators, it can be seen that the method of the present invention is significantly superior to the existing methods in terms of modeling accuracy, response speed, anomaly recognition rate, false alarm control, etc., reflecting the synergistic advantages of structural optimization and model customization.

[0047] The above is only a preferred specific embodiment of the present invention, but the protection scope of the present invention is not limited thereto. Any person skilled in the art within the technical scope disclosed by the present invention, according to the technical solution and inventive concept of the present invention, makes equivalent substitutions or changes, and should be covered by the protection scope of the present invention.

Claims

1. A test method for the launch heat flow field of a sea rocket, characterized in that, It includes the following steps: S1. Collect multi-source heat flux data on the offshore rocket launch platform, synchronize the multi-source heat flux data according to a unified timestamp, and generate a heat flux test data set; S2. Construct a sparse autoencoder model, use the heat flux test data set as input, and the sparse autoencoder model includes an encoder module, a heat flux attention focusing module, a sparse constraint module, and a decoder module; S3. Use the monarch butterfly optimization algorithm to optimize the structural parameters of the sparse autoencoder model, encode each population individual into a set of structural parameter combinations, and obtain the optimized structural parameter combinations; S4. Apply the optimized structural parameters to the sparse autoencoder model, use the optimized sparse autoencoder model to process the multi-source heat flux data, and output two-dimensional heat flux distribution map data; S5. Based on the two-dimensional heat flux distribution map data, identify the heat flux regions in the offshore rocket launch platform structure, analyze the distribution of heat flux at different spatial positions, determine the location and intensity of potential heat anomaly regions, and output corresponding heat shock warning information.

2. The method for testing the launch heat flow field of an offshore rocket according to claim 1, wherein The multi-source heat flux data specifically includes infrared image sequences, thermocouple measurement point heat flux values, environmental wind speed and temperature data, rocket attitude parameters, and platform structure state information.

3. A method for testing the launch heat flow field of an offshore rocket according to claim 1, characterized in that The S2 specifically includes: S21. Set the heat flux test data set as the input matrix , where is a set of real numbers, represents the number of samples, represents the heat flux feature dimension of each sample, including the infrared image sequence, the heat flux value of the thermocouple measurement point, the environmental wind speed and temperature data, the rocket attitude parameters and the platform structure state information; S22. Construct the encoder module of the sparse autoencoder model, which adopts a multi-path parallel structure, including a full-graph perception path, a local hot-spot path, and an edge detail path, to process the information of the whole graph, local hot-spots, and edge contours respectively. The full-graph perception path outputs a low-frequency feature map, the local hot-spot path outputs a medium-frequency feature map, and the edge detail path outputs a high-frequency feature map. The stacking depth of the convolutional layers of each path is controlled by the multi-scale path fusion depth factor Control; S23. Set the kernel size, stride, and normalization parameters for each convolutional path, and splice and fuse the low-frequency feature map, intermediate-frequency feature map, and high-frequency feature map to form a potential feature representation ; S24. Construct a heat flow attention focusing module for potential feature representation Introduce the channel attention and spatial attention mechanisms, and set the spatial attention diffusion radius as the attention diffusion radius control coefficient and generate a weighted feature representation according to the attention response ; S25. Construct a sparse constraint module and set a sparse target value , and introduce a regional sparse adjustment factor. According to the response characteristics of different input feature regions, the activation and inhibition strengths of each hidden neuron are dynamically adjusted according to the regional sparse distribution coefficient to form an adjustable regional sparse control strategy: S26. Introduce an inter-layer sparse adjustment unit between the encoder module, the heat flux attention focusing module, and the sparse constraint module to monitor the activation outputs of each hidden layer according to the importance of regional features, and realize dynamic adjustment of the sparse intensity layer by layer according to the regional sparse distribution coefficient; S27. Construct a decoder module to input the weighted feature representation into the decoding path, and use a combination structure of transposed convolution and upsampling to restore the output dimension and generate output data ; S28. Define the set of structural parameters of the sparse autoencoder model, including the number of layers of the encoder and decoder, the number of neurons in each layer, the sparse target value , the regional sparse distribution coefficient , the multi-scale path fusion depth factor , the attention diffusion radius control coefficient ; S29. Complete the structural initialization of the sparse autoencoder model, and load the current parameter configuration for pre-training.

4. The method for testing the launch heat flow field of an offshore rocket according to claim 3, characterized in that The S3 specifically includes: S31. Initialize the population size of the monarch butterfly optimization algorithm, and divide it into a northern migration sub-population and a southern adjustment sub-population; S32. Represent each monarch butterfly individual using a structure-aware encoding method, and encode the monarch butterfly individual as : ； Among them, represents the structural diagram of the sparse autoencoder model, defining the module connection relationship, enabling state, and path layout, represents the corresponding combination of structural parameters, including the number of layers of the encoder and decoder, the number of neurons in each layer, the global sparse target value , the regional sparse distribution coefficient , the multi-scale path fusion depth factor , the attention diffusion radius control coefficient ; S33. Initialize and pre-train the structural sparse autoencoder model corresponding to each monarch butterfly individual, and use the heat flux test data set to evaluate the reconstruction ability and the high-heat region response ability; S34. Define the fitness function of the monarch butterfly individuals , combined with the reconstruction error of the heat flux image , the reconstruction error of the high-temperature region and the structural complexity : ； Among them, is the weighting coefficient; S35. Calculate the structural covariance entropy diversity index of the current generation population : ； Among them, represents the value of the -th individual in the -th structural parameter dimension, is the population average of the dimension, is the population size, is the structural parameter dimension, is a small positive number, is the logarithmic function; S36. Adopt a non-linear migration probability adjustment mechanism based on diversity metrics to calculate the migration probability of the generation : ； Among them, is the steepness adjustment factor, is the balance threshold, is the exponential function; generate a random number for each individual. If is satisfied, then mark the individual as a migration candidate individual; S37. Perform multi-center attraction update on the migration candidate individuals. Let the set of guiding centers be , where is the combination of structural parameters of the j-th guiding center individual, , is the number of migration guiding centers. Update the combination of structural parameters based on the fitness entropy weight as : ； Among them, is a random scaling factor, is the guiding weight of the j-th guiding center, is the combination of structural parameters of the j-th guiding center individual, is the guiding weight of the k-th guiding center; S38. Perform a regulation operation on the individuals in the sub-population that did not participate in migration, dynamically control the perturbation amplitude according to the distance from the current optimal individual, and update the structure parameter vector to :[[]]END]] ； Among them, is the reference perturbation amplitude, is the maximum individual distance in the current generation population, is the stability constant, is the current structural parameter combination of the i-th individual, is the complete coding of the individual with the optimal fitness in the current generation; S39. Evaluate the fitness values of all newly generated individuals, and select the top q% of the optimal individuals to form an elite set for retention; S310. Determine whether the iteration termination condition is met, including reaching the maximum number of iterations or the fitness change of the optimal individual being lower than the set threshold in several consecutive generations; S311. If the termination condition is not met, merge the elite individuals and the remaining updated individuals to form the next generation population, and return to step S33; S312. If the termination condition is satisfied, output the individual with the optimal fitness , where is the structure diagram encoding corresponding to the individual with the optimal fitness, is the optimal structure parameter combination of the sparse autoencoder model.

5. A method for testing the launch heat flow field of a sea-based rocket according to claim 4, characterized in that The S4 specifically includes: S41. Apply the obtained optimal structural parameter combination to the sparse autoencoder model, and decompose the optimal structural parameter combination into , where represents the optimal number of encoder layers, represents the optimal number of decoder layers, represents the optimal number of neurons in each layer, is the optimal sparsity target value, is the regional sparse distribution coefficient, is the multi-scale path fusion depth factor, is the attention diffusion radius control coefficient; S42. Configure the parallel convolution path structure of the multi-scale encoder module of the sparse autoencoder according to the optimal encoder layer number and the multi-scale path fusion depth factor, set the number of convolution layers and the receptive field of each path, and set the diffusion window size of the spatial attention convolution according to the attention diffusion radius control coefficient; S43. Configure the density of all neurons based on the optimal number of neurons in each layer, and set the sparse activation constraint weights for different heat flux regions according to the regional sparse distribution coefficient; S44. Construct a sparse autoencoder model structure based on the optimal combination of structural parameters , expressed as a function mapping , where represents the input heat flux test data set is the latent representation feature tensor; S45. Input the multi-source heat flux data during the offshore rocket launch newly collected into the optimized sparse autoencoder model, perform the encoding-focusing-decoding process, and obtain the two-dimensional heat flux feature map ; S46. Normalize and filter the reconstructed output of the optimized sparse autoencoder model, and output two-dimensional heat flux distribution map data.

6. The method for testing the launch heat flow field of an offshore rocket according to claim 5, characterized in that The S5 specifically includes: S51. Obtain the data of the output two-dimensional heat flux distribution map, where the two-dimensional heat flux distribution map data contains the heat value distribution information at different spatial positions; S52. Utilize the characteristics of heat value gradient change to extract the heat response mutation regions from the two-dimensional heat flux distribution map data, and combine with the structural layout information of the launch platform to identify the heat flux regions in the structure of the offshore rocket launch platform, covering the rocket flame impact area, the ground reflection area, and the wake flow around the flame area; S53. Conduct spatial modeling and analysis on the identified heat flux regions to evaluate the coverage range, relative position relationship, and heat value density distribution of different heat flux regions on the platform structure, and determine the distribution characteristics and propagation trends of each heat flux region in the horizontal and vertical directions; S54. Based on the spatial modeling results and heat value characteristics, conduct focused analysis on the phenomenon of abnormal concentration of temperature in the heat flux regions to determine the location and heat response intensity level of potential heat anomaly regions in the platform structure; S55. Compare and verify the determined heat anomaly regions with the two-dimensional heat flux distribution map data to evaluate the consistency performance of the heat anomaly regions in the actual heat flux response, and form a heat shock determination result; S56. According to the heat shock determination result, generate corresponding heat shock warning information, clarify the warning level, influence range, and structural risk regions, and assist in the pre-launch structural thermal safety assessment and risk prompt during the mission execution process.

7. A test system for the launch heat flux field of a sea rocket, and a test method for the launch heat flux field of a sea rocket according to any one of claims 1 to 6, characterized in that, It includes the following modules: The data acquisition module is used to collect multi-source heat flux data of the offshore rocket launch platform, complete time synchronization and preprocessing, and construct a heat flux data set; The model construction module is used to construct a sparse autoencoder model; The feature extraction module is used to input the heat flux data set into the sparse autoencoder model to extract the potential features of the two-dimensional heat flux distribution; The structure optimization module is used to globally optimize the structural parameters of the sparse autoencoder model; The heat flux generation module is used to process the multi-source heat flux data based on the optimized sparse autoencoder model and output the two-dimensional heat flux distribution map data; The heat flux identification module is used to analyze the two-dimensional heat flux distribution map data to identify the heat flux regions and potential heat anomalies in the platform structure; The warning module is used to output heat shock warning information and mark the abnormal regions and risk levels.

Images