Rocket platform sea condition self-adaptive lifting control method based on self-supervised learning

Abstract

The invention discloses a rocket platform sea condition adaptive lifting control method based on self-supervised learning. The method comprises the following steps: S1, generating a preprocessed sea condition disturbance data set; s2, outputting a multi-scale disturbance characteristic representation set; s3, obtaining a dynamic prior variation disturbance encoder; s4, reasoning the current preprocessed sea condition disturbance data set by using a dynamic prior variational disturbance encoder, outputting a real-time sea condition disturbance feature vector, and setting a control performance evaluation function; s5, guiding the secretary bird population to initialize the position by adopting the real-time sea condition disturbance feature vector, and carrying out iterative calculation through a dynamic guiding type secretary bird optimization algorithm to obtain an optimal lifting control parameter group; and S6, the optimal lifting control parameter set is input into a rocket platform lifting execution system, platform lifting control is completed, and platform lifting response data are generated. According to the method, the problems that a traditional disturbance feature extraction method is insufficient in non-structural strong noise signal recognition capability and excessively static in disturbance prior hypothesis are solved, and disturbance sensing precision and system robustness are greatly improved.

Description

Technical Field

[0001] The present invention relates to the technical field of rocket platforms, and in particular to a sea condition adaptive lifting and lowering control method for a rocket platform based on self-supervised learning. Background Art

[0002] As the application scenarios of rocket platforms expand into the complex environments of near-shore launches and offshore recovery, the dynamic disturbance factors faced by the platforms during mission execution are becoming increasingly complex. During the ascent and descent phases, their operational stability and control accuracy have a decisive impact on the safety and success rate of the overall flight mission. Currently, in actual deployment, rocket platforms mainly rely on traditional controller design schemes, such as fixed-gain PID controllers, fuzzy controllers, or feedback control systems with preset rule bases. Although existing methods have a certain degree of controllability under static or weak disturbance conditions, they often exhibit response lag, unstable regulation, and insufficient control robustness under conditions of severe sea disturbances, multi-scale dynamic interference, and strong nonlinear responses.

[0003] Existing rocket platform control systems generally face two key difficulties: First, sea state disturbance signals are characterized by high dimensionality, multi-sources, and strong noise interference. Most of the disturbance perception modules that current control systems rely on adopt simple filtering, threshold judgment, or empirical feature extraction methods based on the original sensor signals. It is difficult to extract high-quality disturbance features from complex signals, resulting in control input lag or even failure; second, the control parameters are multidimensional and complex and show strong nonlinear dependencies. Traditional control strategies based on manual experience or linear parameter adjustment cannot be adapted in real time in time-varying disturbance scenarios, and are prone to falling into local optimal or out-of-control states, and cannot meet the rapid response requirements of rocket platform lifting systems in highly dynamic environments.

[0004] In addition, although some studies in recent years have begun to try to introduce intelligent optimization algorithms such as particle swarms, genetic algorithms or reinforcement learning strategies to assist in the optimization of rocket platform control systems, the relevant methods have problems such as slow algorithm convergence, strong initialization sensitivity, and inability to effectively integrate high-dimensional disturbance perception data. Most of them lack the ability to represent and model the disturbance data itself, making it difficult to form a dynamic adjustment mechanism for control parameters for disturbance structures.

[0005] Therefore, it is urgent to build an integrated method of perception and control of multi-scale sea state disturbances for rocket platforms to solve the key problems of inaccurate disturbance feature extraction, untimely update of control parameters, and unstable optimization strategies, so as to adapt to the stable lifting and lowering control needs of future rocket platforms in highly dynamic and complex environments. Summary of the Invention

[0006] One purpose of the present invention is to propose a sea-condition adaptive lifting and lowering control method for a rocket platform based on self-supervised learning. The present invention effectively solves the problems of insufficient recognition ability of non-structured strong noise signals and overly static prior assumptions about disturbances in traditional disturbance feature extraction methods, thereby greatly improving the disturbance perception accuracy and system robustness.

[0007] According to an embodiment of the present invention, a method for sea-state adaptive lifting and lowering control of a rocket platform based on self-supervised learning includes the following steps:

[0008] S1. Synchronously deploy an acquisition system on the rocket platform to continuously acquire and preprocess the sea state disturbance dataset to generate a preprocessed sea state disturbance dataset;

[0009] S2. Construct a multi-scale self-supervised variational contrastive learning model and perform self-supervised training on the pre-processed sea state disturbance dataset as input, outputting a set of multi-scale disturbance feature representations.

[0010] S3. Construct a dynamic prior distribution model of the sea state disturbance of the rocket platform based on the multi-scale disturbance feature representation set. This dynamic prior distribution model is embedded in the multi-scale self-supervised variational contrastive learning model to complete the model update, thus obtaining a dynamic prior variational disturbance encoder.

[0011] S4. Use a dynamic prior variational disturbance encoder to infer the current preprocessed sea state disturbance dataset, output a real-time sea state disturbance feature vector, and construct a multidimensional search space for the lifting control parameters based on the real-time sea state disturbance feature vector and the multi-scale disturbance feature representation set, and set a control performance evaluation function;

[0012] S5. Using the real-time sea state disturbance characteristic vector to guide the initial position of the secretary bird population, the dynamic guided secretary bird optimization algorithm is started to perform a global search in the multidimensional search space of the lift control parameters, and the optimal lift control parameter set is obtained through iterative calculation of the dynamic guided secretary bird optimization algorithm;

[0013] S6. Input the optimal lifting control parameter group into the rocket platform lifting execution system and complete the platform lifting control to generate platform lifting response data.

[0014] Optionally, the S1 includes the following steps:

[0015] S11. Accelerometers, gyroscopes, and sea state monitors are placed at key structural locations on the rocket platform. Sampling intervals and total sampling durations are set to continuously collect disturbance signals from the rocket platform in a dynamic sea state environment. This generates an original sea state disturbance dataset. Each disturbance sampling data item d in the original sea state disturbance dataset is i Including timestamp t i , three-dimensional acceleration vector a i , three-dimensional angular velocity vector ωi And the instantaneous effective wave height h i , the total number of perturbation data is N raw =T total / Δt;

[0016] S12. Preprocess the original sea state disturbance data set using a time window length of T smooth The sliding filtering method is used to smooth the three-dimensional acceleration vector, three-dimensional angular velocity vector and instantaneous effective wave height value, eliminating the high-frequency disturbance components caused by sensor sampling jitter. Based on the adaptive threshold mechanism, the signal outliers are detected and eliminated, and atypical disturbance samples affected by external abnormal interference are excluded to obtain the intermediate processed data set after filtering and denoising.

[0017] S13. Perform disturbance feature enhancement on the intermediate processing data set, calculate the disturbance change rate feature between continuous disturbance data according to the time step, and extract the three-dimensional acceleration change rate Δa i , three-dimensional angular velocity change rate Δω i and the effective wave height change rate Δh i , respectively represent the acceleration fluctuation trend, attitude disturbance fluctuation trend and sea state disturbance fluctuation trend of the rocket platform at the current moment relative to the previous moment. The disturbance change rate feature is spliced with the three-dimensional acceleration vector, three-dimensional angular velocity vector and instantaneous effective wave height value at the corresponding moment to form an enhanced disturbance data sample set, which is used as the final preprocessed sea state disturbance data set D pre .

[0018] Optionally, S2 includes the following steps:

[0019] S21. Construct a multi-scale self-supervised variational contrastive learning model to preprocess the sea state disturbance dataset D pre As input, according to the multi-frequency oscillation characteristics contained in the disturbance sample, a scale set S is set. Each scale s in the scale set S is k The response window corresponding to the different frequency disturbance characteristics of the rocket platform lifting system is obtained through the scale filter Acting on the channel splicing feature vector to construct the scale perturbation feature signal

[0020]

[0021] S22. Calculate the energy of the scale disturbance characteristic signal to measure its potential impact on the rocket platform's lifting stability. Define the scale energy as:

[0022]

[0023] in, Denotes the perturbation sample di In scale s k The total energy under the scale is the unit sum of squares of the combined signal, which represents the driving amplitude of the disturbance signal on the dynamic response capability of the rocket platform at this scale;

[0024] S23. Normalize the scale energy to obtain the scale energy weight of the perturbation sample

[0025]

[0026] in, Indicates the perturbation sample at scale s k The scale energy weight under , reflects the importance of the scale disturbance to the rocket platform lifting decision;

[0027] S24. Perturb the characteristic signal at each scale Set up and parallel input global perturbation encoder With local perturbation encoder Extract the slow-changing trend disturbance and instantaneous disturbance during the lifting and lowering of the rocket platform respectively:

[0028]

[0029] in, Indicates scale s k The mean vector of the perturbed feature distribution, Scale s k The standard deviation vector of the perturbed feature distribution;

[0030] S25. Generate standard normal noise independently for each scale based on the reparameterization strategy and Combined with the corresponding mean vector and standard deviation vector, the long-term perturbation latent vector is obtained and the instantaneous perturbation latent vector

[0031]

[0032] S26. Calculate the long-term perturbation latent vector and the instantaneous perturbation latent vector The cosine similarity between them, combined with the latent variable dimension d z Perform temperature scaling and then apply the softmax function to obtain the cross-scale attention coefficient The cross-scale attention coefficient reflects the mutual driving relationship of disturbances at different scales and is used to construct a coordinated ascending and descending regulatory signal pathway;

[0033] S27. According to scale energy weight and cross-scale attention coefficient For instantaneous perturbation latent vector Perform weighted summation to generate the perturbation fusion vector f i The disturbance fusion vector integrates the comprehensive impact of the sea state disturbances of various scales on the lift state of the rocket platform at the current moment:

[0034]

[0035] S28. Fusion perturbation vector f i Input disturbance stability prediction head P θ , output the predicted value of future wave height change slope The actual wave height change trend s of the rocket platform under the current disturbance i and predicted value The mean square error between them defines the stability prediction loss L p , used to measure the accuracy of predicting the future stability of the rocket platform:

[0036]

[0037] Among them, N pre represents the total number of perturbation samples used to train the perturbation stability prediction head;

[0038] S29. By jointly optimizing the multi-scale variational reconstruction loss L v , self-supervised contrast loss L c and stability prediction loss L p The adjustment coefficients λ1 and λ2 are used to control the relative weights of each loss in the sea disturbance perception task, and the parameters of the multi-scale self-supervised variational contrastive learning model are updated:

[0039] L total =L v +λ1L c +λ2L p ;

[0040] Among them, λ1 and λ2 are the adjustment coefficients of contrastive learning and prediction loss, respectively, which control the relative weights of feature learning and stability judgment in the disturbance perception task;

[0041] S210. After completing the joint training, output the multi-scale perturbation feature representation set

[0042] Optionally, S3 includes the following steps:

[0043] S31. Multi-scale perturbation feature representation set F multi , according to the disturbance sample timestamp t i Construct the disturbance feature time series trajectory and fuse each disturbance vector f i Its corresponding timestamp ti The disturbance characteristic time series state trajectory set T is formed to characterize the evolution characteristics of the sea state disturbance of the rocket platform in a continuous time period;

[0044] S32. Set a fixed-length sliding time window T on the disturbance feature time series state trajectory set T prior , the disturbance feature time series state trajectory set is divided into several local sub-window sequences of disturbance using a sliding time window, and each local sub-window sequence of disturbance contains the time period [t j ,t j +T prior ) in the fused perturbation feature vectors, recorded as the perturbation feature sub-window set W j , used to extract the local disturbance distribution pattern of the rocket platform at different stages of sea state change;

[0045] S33. For each disturbance feature sub-window set W j The fused disturbance feature vector in is counted and the mean vector of all disturbance features in the disturbance feature sub-window set is calculated. and the covariance matrix They are used to describe the central trend of disturbance and the diffusion range of disturbance characteristics of the rocket platform in the current time period respectively;

[0046] S34. Constructing a priori prediction network P φ , the perturbation sample timestamp t i and the corresponding perturbation fusion vector f i Input, output estimated prior distribution parameters To estimate the potential response structure of the sea state disturbance of the rocket platform at the current moment, and define the prior estimation error loss function with KL divergence:

[0047]

[0048] in, are the posterior distribution parameters calculated by the multi-scale encoder at the current moment, is the estimated prior distribution parameter of the prior prediction network output, and the KL divergence measures the structural deviation between the current state of the platform and the predicted disturbance trend;

[0049] S35. Estimated prior distribution parameters of the prior prediction network output It is embedded in the multi-scale self-supervised variational contrastive learning model, replacing the standard normal distribution as a new variational prior distribution, completing the dynamic structural update of the disturbance perception model during the rocket platform lifting and lowering process, and forming a dynamic prior variational perturbation encoder.

[0050] Optionally, the S4 includes the following steps:

[0051] S41. Based on the dynamic prior variational disturbance encoder, receive the latest enhanced disturbance data sample d in the current preprocessed sea state disturbance dataset t , and its channel splicing feature x t As input, infer the posterior distribution parameters of the disturbance at the current moment At the same time, the current prior distribution parameters are generated by the prior prediction network Based on the dynamic prior distribution, a reparameterization strategy is used to generate the current moment's real-time perturbation potential feature vector:

[0052]

[0053] Among them, z t is the real-time disturbance potential feature vector of the platform in the multi-scale disturbance space at the current moment, ∈ t is a standard normal noise vector, which represents the abstract representation of the current sea state disturbance during the ascent and descent of the rocket platform;

[0054] S42. Perturb the potential feature vector z in real time t and the multi-scale perturbation feature representation set F multi Perform scale alignment and obtain attention weights through the cosine similarity-softmax mechanism adjusted by the temperature coefficient τ Then use the attention weight to perturb the fusion vector f i Weighted summation to obtain the disturbance driven fusion vector

[0055] S43. Fusion vector driven by disturbance Construct a multi-dimensional search space P for lifting control parameters. The multi-dimensional search space for lifting control parameters is composed of the control parameter vector p = (p1, p2, ..., p d ), each control parameter p in the control parameter vector i are constrained to the corresponding minimum value With the maximum value between;

[0056] S44. Define the control performance evaluation function J(p, z t ), the control performance evaluation function is composed of the height fluctuation residual ΔH res (p,z t ), attitude stability error Δθ stab (p,z t ) and control response delay time T delay (p,z t ) is composed of the sum of three weighted squares, with weight coefficients λ1, λ2, and λ3 respectively, which are used to comprehensively measure the lift control accuracy and response efficiency of the rocket platform under the current potential disturbance state.

[0057] Optionally, the S5 includes the following steps:

[0058] S51. Based on the real-time perturbation potential feature vector z t , calculate the current disturbance intensity index I t , where the disturbance intensity index I t By real-time perturbation of the latent feature vector z t The second norm of is used to measure the overall energy amplitude of the disturbance experienced by the rocket platform under the current sea conditions;

[0059] S52. Construct a secretary bird population consisting of M individuals in the multidimensional search space P of the lifting control parameters, and use real-time perturbation to adjust the potential feature vector z t To inspire information, through the weight matrix W (m) With the bias vector b (m) After the linear transformation of , the perturbation mapping coefficient Γ is obtained by Sigmoid function compression. (m) , then according to the lower limit vector p min With the upper limit vector p max The initial control parameter vector of each secretary bird is generated by the interval proportional mapping method Complete the disturbance-driven initialization of the secretary bird population;

[0060] S53. Based on the disturbance intensity index I t Set the adaptive step size adjustment factor α t , adaptive step size adjustment factor α t By the basic step size α base Together with the step-size attenuation coefficient κ, when the disturbance intensity index I t When it increases, the adaptive step size adjustment factor α t Decrease in an inverse relationship;

[0061] S54. During the optimization iteration of the secretary bird, for each secretary bird individual in the kth generation, the current control parameter vector is updated by the jump-tracking strategy. Plus the adaptive step size adjustment factor α t The direction increment mixed with random weights r1 and r2 is realized to achieve the optimal control parameter vector With random individual vector and get closer at the same time, thus obtaining the next generation control parameter vector

[0062]

[0063] in, represents the control performance evaluation function J(p,z t ), is a random individual vector in the current population, r1 and r2 are random weights independently and uniformly sampled in the interval [0,1], which are used to simulate the hunting focus behavior and global reconnaissance behavior of secretary birds;

[0064] S55. Repeat the secretary bird population update until the set termination condition is reached. The termination condition includes the maximum number of iterations or the control performance evaluation function J(p, z t )The improvement is lower than the threshold, and the final output makes the control performance evaluation function J(p,z t ) to obtain the minimum control parameter vector p * , control parameter vector p * As the current optimal lifting control parameter group of the rocket platform lifting execution system.

[0065] Optionally, the S6 includes the following steps:

[0066] S61. The optimal lifting control parameter vector p * Input the rocket platform lifting execution system to drive the platform to perform the current lifting task. The lifting execution system adjusts the key control signals of lifting rate, attitude angle stability factor, and disturbance compensation gain according to the strategy values corresponding to each control dimension in the optimal control parameter vector;

[0067] S62. During the lifting process, a sensor system fixed to the rocket platform structure is used to monitor the lifting dynamic process in real time. The platform's altitude change curve, attitude change trajectory, attitude offset rate, and lifting time delay platform response variables are recorded within the lifting response cycle to generate the platform lifting response dataset R. exec ;

[0068] S63. Platform lift response dataset R exec Each response data r i Includes the following parameter fields: The platform's current lifting height H i , platform attitude angle change rate Δθ i , control command response time Disturbance compensation error System control status label L i , system control status label L i The expression labels the current platform operation status.

[0069] Optionally, the tag identifier is defined as follows:

[0070] Normal operating state: disturbance compensation error Response time Attitude angle change rate Δθ i ≤3° / s;

[0071] Disturbance compensation error Response time Or attitude angle change rate 3° / s<Δθ i ≤5° / s;

[0072] Control instability: disturbance compensation error Response time Or attitude angle change rate Δθ i >5° / s.

[0073] The beneficial effects of the present invention are:

[0074] (1) The present invention proposes a multi-scale variational contrastive learning architecture in the disturbance perception link. By introducing scale energy distribution and self-attention cross-scale collaborative mechanism, the model can extract the temporal variation law of acceleration, angular velocity and effective wave height multimodal information from different frequency disturbance response windows, and integrate the variational sampling mechanism to obtain a stable representation. On this basis, a dynamic disturbance prior distribution network constructed by time series is embedded to replace the traditional fixed standard normal distribution. By predictively modeling the evolution trend of the disturbance distribution, the sensitivity and stability of the model to the temporal characteristics of sea state disturbance are enhanced. This effectively solves the problems of the traditional disturbance feature extraction method's insufficient ability to recognize non-structured strong noise signals and the overly static disturbance prior assumption, greatly improving the disturbance perception accuracy and system robustness.

[0075] (2) The present invention constructs a secretary bird population initialization strategy guided by the disturbance potential eigenvector, and dynamically generates the initialization control parameter position distribution in combination with the disturbance intensity, so that the secretary bird population starts from the highly correlated area of the disturbance state, skipping the inefficient exploration process caused by a large number of random initializations in the traditional algorithm. At the same time, the step size adjustment coefficient is driven by the disturbance intensity index to construct a two-way closed-loop mechanism of disturbance perception-optimization search, which dynamically converges to a high-precision local area when the sea conditions change drastically and maintains the global search capability under stable disturbance conditions. The improved strategy is significantly better than the traditional secretary bird and particle swarm fixed step size optimization algorithms. In the high-dimensional parameter space and nonlinear control function scenarios, the control parameter optimization speed is improved.

[0076] (3) The present invention establishes a direct feedback channel between the disturbance perception model and the platform lifting execution system. The attitude change rate, wave height response, and control delay indicators generated in real time during the lifting execution process constitute the platform lifting response data set, and the operating state label is defined according to the disturbance compensation error and attitude stability. It is not only used to evaluate the effect of the current control strategy, but also to reversely act on the dynamic prior prediction network and the secretary bird optimizer to update the control strategy and the disturbance prior model online, thereby realizing intelligent adaptive control under disturbance drive. This linkage closed-loop system is significantly superior to the traditional "perception-control-execution" one-way architecture and has extremely high dynamic adaptability and engineering practicality. BRIEF DESCRIPTION OF THE DRAWINGS

[0077] The accompanying drawings are used to provide a further understanding of the present invention and constitute a part of the specification. Together with the embodiments of the present invention, they are used to explain the present invention and do not constitute a limitation of the present invention. In the accompanying drawings:

[0078] Figure 1 This is a flow chart of a sea condition adaptive lifting and lowering control method for a rocket platform based on self-supervised learning proposed by the present invention. DETAILED DESCRIPTION

[0079] The present invention will now be described in further detail with reference to the accompanying drawings, which are simplified schematic diagrams that illustrate the basic structure of the present invention in a schematic manner.

[0080] refer to Figure 1 A method for sea-state adaptive lifting control of a rocket platform based on self-supervised learning includes the following steps:

[0081] S1. Synchronously deploy an acquisition system on the rocket platform to continuously acquire and preprocess the sea state disturbance dataset to generate a preprocessed sea state disturbance dataset;

[0082] S2. Construct a multi-scale self-supervised variational contrastive learning model and perform self-supervised training on the pre-processed sea state disturbance dataset as input, outputting a set of multi-scale disturbance feature representations.

[0083] S3. Construct a dynamic prior distribution model of the sea state disturbance of the rocket platform based on the multi-scale disturbance feature representation set. This dynamic prior distribution model is embedded in the multi-scale self-supervised variational contrastive learning model to complete the model update, thus obtaining a dynamic prior variational disturbance encoder.

[0084] S4. Use a dynamic prior variational disturbance encoder to infer the current preprocessed sea state disturbance dataset, output a real-time sea state disturbance feature vector, and construct a multidimensional search space for the lifting control parameters based on the real-time sea state disturbance feature vector and the multi-scale disturbance feature representation set, and set a control performance evaluation function;

[0085] S5. Using the real-time sea state disturbance characteristic vector to guide the initial position of the secretary bird population, the dynamic guided secretary bird optimization algorithm is started to perform a global search in the multidimensional search space of the lift control parameters, and the optimal lift control parameter set is obtained through iterative calculation of the dynamic guided secretary bird optimization algorithm;

[0086] S6. Input the optimal lifting control parameter group into the rocket platform lifting execution system and complete the platform lifting control to generate platform lifting response data.

[0087] In this embodiment, S1 includes the following steps:

[0088] S11. Accelerometers, gyroscopes, and sea state monitors are placed at key structural locations on the rocket platform. Sampling intervals and total sampling durations are set to continuously collect disturbance signals from the rocket platform in a dynamic sea state environment. This generates an original sea state disturbance dataset. Each disturbance sampling data item d in the original sea state disturbance dataset is i Including timestamp t i , three-dimensional acceleration vector a i , three-dimensional angular velocity vector ω i And the instantaneous effective wave height h i , the total number of perturbation data is N raw =T total / Δt;

[0089] S12. Preprocess the original sea state disturbance data set using a time window length of T smooth The sliding filtering method is used to smooth the three-dimensional acceleration vector, three-dimensional angular velocity vector and instantaneous effective wave height value, eliminating the high-frequency disturbance components caused by sensor sampling jitter. Based on the adaptive threshold mechanism, the signal outliers are detected and eliminated, and atypical disturbance samples affected by external abnormal interference are excluded to obtain the intermediate processed data set after filtering and denoising.

[0090] S13. Perform disturbance feature enhancement on the intermediate processing data set, calculate the disturbance change rate feature between continuous disturbance data according to the time step, and extract the three-dimensional acceleration change rate Δa i , three-dimensional angular velocity change rate Δω i and the effective wave height change rate Δh i , respectively represent the acceleration fluctuation trend, attitude disturbance fluctuation trend and sea state disturbance fluctuation trend of the rocket platform at the current moment relative to the previous moment. The disturbance change rate feature is spliced with the three-dimensional acceleration vector, three-dimensional angular velocity vector and instantaneous effective wave height value at the corresponding moment to form an enhanced disturbance data sample set, which is used as the final preprocessed sea state disturbance data set D pre .

[0091] In this embodiment, S2 includes the following steps:

[0092] S21. Construct a multi-scale self-supervised variational contrastive learning model to preprocess the sea state disturbance dataset D pre As input, according to the multi-frequency oscillation characteristics contained in the disturbance sample, a scale set S is set. Each scale s in the scale set S is k The response window corresponding to the different frequency disturbance characteristics of the rocket platform lifting system is obtained through the scale filter Acting on the channel splicing feature vector to construct the scale perturbation feature signal

[0093]

[0094] S22. Calculate the energy of the scale disturbance characteristic signal to measure its potential impact on the rocket platform's lifting stability. Define the scale energy as:

[0095]

[0096] in, Denotes the perturbation sample d i In scale s k The total energy under the scale is the unit sum of squares of the combined signal, which represents the driving amplitude of the disturbance signal on the dynamic response capability of the rocket platform at this scale;

[0097] S23. Normalize the scale energy to obtain the scale energy weight of the perturbation sample

[0098]

[0099] in, Indicates the perturbation sample at scale s k The scale energy weight under , reflects the importance of the scale disturbance to the rocket platform lifting decision;

[0100] S24. Perturb the characteristic signal at each scale Set up and parallel input global perturbation encoder With local perturbation encoder Extract the slow-changing trend disturbance and instantaneous disturbance during the lifting and lowering of the rocket platform respectively:

[0101]

[0102] in, Indicates scale s k The mean vector of the perturbed feature distribution, Scale s k The standard deviation vector of the perturbed feature distribution;

[0103] S25. Generate standard normal noise independently for each scale based on the reparameterization strategy and Combined with the corresponding mean vector and standard deviation vector, the long-term perturbation latent vector is obtained and the instantaneous perturbation latent vector

[0104]

[0105]

[0106] S26. Calculate the long-term perturbation latent vector and the instantaneous perturbation latent vector The cosine similarity between them, combined with the latent variable dimension d z Perform temperature scaling and then apply the softmax function to obtain the cross-scale attention coefficient The cross-scale attention coefficient reflects the mutual driving relationship of disturbances at different scales and is used to construct a coordinated ascending and descending regulatory signal pathway;

[0107] S27. According to scale energy weight and cross-scale attention coefficient For instantaneous perturbation latent vector Perform weighted summation to generate the perturbation fusion vector f i The disturbance fusion vector integrates the comprehensive impact of the sea state disturbances of various scales on the lift state of the rocket platform at the current moment:

[0108]

[0109] S28. Fusion perturbation vector f i Input disturbance stability prediction head P θ , output the predicted value of future wave height change slope The actual wave height change trend s of the rocket platform under the current disturbance i and predicted value The mean square error between them defines the stability prediction loss L p , used to measure the accuracy of predicting the future stability of the rocket platform:

[0110]

[0111] Among them, N pre represents the total number of perturbation samples used to train the perturbation stability prediction head;

[0112] S29. By jointly optimizing the multi-scale variational reconstruction loss L v , self-supervised contrast loss L c and stability prediction loss Lp The adjustment coefficients λ1 and λ2 are used to control the relative weights of each loss in the sea disturbance perception task, and the parameters of the multi-scale self-supervised variational contrastive learning model are updated:

[0113] L total =L v +λ1L c +λ2L p ;

[0114] Among them, λ1 and λ2 are the adjustment coefficients of contrastive learning and prediction loss, respectively, which control the relative weights of feature learning and stability judgment in the disturbance perception task;

[0115] Multi-scale variational reconstruction loss L v It is used to constrain the generation capacity of the perturbation latent variables, so that the model can accurately reconstruct the perturbation characteristics at all scales and maintain the distribution consistency and structural fidelity between the perturbation potential representation and the observed data. Its construction logic is as follows:

[0116] For each scale s k , respectively construct the global perturbation encoder With local perturbation encoder Extract the long-term mean vector of the disturbance sample at the current scale Standard deviation vector and transient disturbances

[0117] Using the reparameterization strategy, the perturbation latent variables are generated based on the above mean and standard deviation and

[0118] Input the hidden variables into the perturbation reconstructor Generate reconstructed values of scale perturbation signals

[0119] Comparing true-scale perturbation signals and reconstruction value The difference between the two scales is defined as the reconstruction error loss of each scale: the observation error term (MSE loss) KL divergence term: measures the deviation between the posterior and the prior (or standard normal), and the losses of all scales are accumulated to form the overall variational reconstruction loss Used to maintain the integrity and multi-scale fidelity of perturbation information during the encoding-decoding process.

[0120] The multi-scale variational reconstruction loss term ensures that the model can effectively reconstruct disturbance characteristics at different scales (low-frequency wave trends and high-frequency jitter), maintain the rocket platform's multi-scale disturbance recognition and recovery capabilities under complex sea conditions, and effectively support the precise perception of subsequent lifting and control tasks.

[0121] Self-supervised contrast loss is used to improve the discriminability and temporal structure consistency of perturbation features in the latent space. By using the strategy of pulling positive samples closer and pushing negative samples further away, similarity between perturbations is used to drive expression learning. The construction logic is as follows:

[0122] The perturbation sample d i Generate two positive sample representations under different data enhancement methods (scale perturbation, feature perturbation) From the same disturbance source;

[0123] Extract negative samples from other perturbation samples in the same batch (different time or different perturbation scenarios)

[0124] The contrastive learning goal is to make the positive samples To keep the proximity in the feature space and the distinction between negative samples, the contrast loss is constructed using the temperature-scaled InfoNCE loss:

[0125]

[0126] Where sim(·) represents the cosine similarity function and τ is the temperature adjustment coefficient.

[0127] In the rocket platform lifting and lowering mission, different sea state disturbances may have similar trends but large differences in details. Through self-supervised comparative loss, the model is equipped with the ability to identify "fine-grained differences" in disturbances, thereby improving the model's response sensitivity to sudden disturbance changes and critical moments of platform attitude transitions, thereby supporting high-robustness decision-making in lifting and lowering control.

[0128] S210. After completing the joint training, output the multi-scale perturbation feature representation set

[0129] In this embodiment, S3 includes the following steps:

[0130] S31. Multi-scale perturbation feature representation set F multi , according to the disturbance sample timestamp t i Construct the disturbance feature time series trajectory and fuse each disturbance vector f i Its corresponding timestamp t i The disturbance characteristic time series state trajectory set T is formed to characterize the evolution characteristics of the sea state disturbance of the rocket platform in a continuous time period;

[0131] S32. Set a fixed-length sliding time window T on the disturbance feature time series state trajectory set T prior , the disturbance feature time series state trajectory set is divided into several local sub-window sequences of disturbance using a sliding time window, and each local sub-window sequence of disturbance contains the time period [t j ,tj +T prior ) in the fused perturbation feature vectors, recorded as the perturbation feature sub-window set W j , used to extract the local disturbance distribution pattern of the rocket platform at different stages of sea state change;

[0132] S33. For each disturbance feature sub-window set W j The fused disturbance feature vector in is counted and the mean vector of all disturbance features in the disturbance feature sub-window set is calculated. and the covariance matrix They are used to describe the central trend of disturbance and the diffusion range of disturbance characteristics of the rocket platform in the current time period respectively;

[0133] S34. Constructing a priori prediction network P φ , the perturbation sample timestamp t i and the corresponding perturbation fusion vector f i Input, output estimated prior distribution parameters To estimate the potential response structure of the sea state disturbance of the rocket platform at the current moment, and define the prior estimation error loss function with KL divergence:

[0134]

[0135] in, are the posterior distribution parameters calculated by the multi-scale encoder at the current moment, is the estimated prior distribution parameter of the prior prediction network output, and the KL divergence measures the structural deviation between the current state of the platform and the predicted disturbance trend;

[0136] S35. Estimated prior distribution parameters of the prior prediction network output It is embedded in the multi-scale self-supervised variational contrastive learning model, replacing the standard normal distribution as a new variational prior distribution, completing the dynamic structural update of the disturbance perception model during the rocket platform lifting and lowering process, and forming a dynamic prior variational perturbation encoder.

[0137] In this embodiment, S4 includes the following steps:

[0138] S41. Based on the dynamic prior variational disturbance encoder, receive the latest enhanced disturbance data sample d in the current preprocessed sea state disturbance dataset t , and its channel splicing feature x t As input, infer the posterior distribution parameters of the disturbance at the current moment At the same time, the current prior distribution parameters are generated by the prior prediction network Based on the dynamic prior distribution, a reparameterization strategy is used to generate the current moment's real-time perturbation potential feature vector:

[0139]

[0140] Among them, z t is the real-time disturbance potential feature vector of the platform in the multi-scale disturbance space at the current moment, ∈ t is a standard normal noise vector, which represents the abstract representation of the current sea state disturbance during the ascent and descent of the rocket platform;

[0141] S42. Perturb the potential feature vector z in real time t and the multi-scale perturbation feature representation set F multi Perform scale alignment and obtain attention weights through the cosine similarity-softmax mechanism adjusted by the temperature coefficient τ Then use the attention weight to perturb the fusion vector f i Weighted summation to obtain the disturbance-driven fusion vector

[0142] S43. Fusion vector driven by disturbance Construct a multi-dimensional search space P for lifting control parameters. The multi-dimensional search space for lifting control parameters is composed of the control parameter vector p = (p1, p2, ..., p d ), each control parameter p in the control parameter vector i are constrained to the corresponding minimum value With the maximum value between;

[0143] S44. Define the control performance evaluation function J(p, z t ), the control performance evaluation function is composed of the height fluctuation residual ΔH res (p,z t ), attitude stability error Δθ stab (p,z t ) and control response delay time T delay (p,z t ) is composed of the sum of three weighted squares, with weight coefficients λ1, λ2, and λ3 respectively, which are used to comprehensively measure the lift control accuracy and response efficiency of the rocket platform under the current potential disturbance state.

[0144] In this embodiment, the height fluctuation residual ΔH res (p,z t ) is used to measure the potential state z in the current disturbance t The degree of deviation of the rocket platform's lifting height from the desired trajectory under the conditions of the control parameter p is constructed as follows:

[0145] During the lifting control process, the platform will execute a lifting trajectory according to the control parameter p. The actual height sequence during this period is The unit is meter (m);

[0146] Assume that the expected height trajectory corresponding to this period is Can be generated by the target pose planner;

[0147] The square root of the average residual between the actual trajectory and the expected trajectory during the entire response period is taken as the height fluctuation residual: the smaller the residual, the more stable the platform height change and the higher the control accuracy; the unit is still meter, which is consistent with the output height dimension of the lifting system.

[0148] Attitude stability error Δθ stab (p,z t ) is used to describe the dynamic change of the attitude angle of the rocket platform during the lifting process, reflecting the degree of jitter of the attitude control during the lifting process. The construction method is as follows:

[0149] During the control execution, the change of the platform attitude angle over time is recorded to obtain the attitude angle change rate sequence: The unit is degree / second (° / s); it can be obtained from the platform attitude inertial measurement unit;

[0150] Calculate the variance or maximum rate of change of the attitude change rate within the response period as the stability indicator of the platform under the parameter conditions; the smaller the variance / maximum value, the smaller the platform jitter and the smoother the lifting and lowering; finally normalize it to a single stability error indicator Δθ stab (p,z t ), unit is degree / second.

[0151] Control response delay time T delay (p,z t ) is used to measure the delay between the platform receiving a control command and generating an effective action, reflecting the system's agile response capability under disturbances. It is constructed as follows:

[0152] For each control cycle, record the control command issuance time t cmd The time t when the platform detects the first response action res ;

[0153] The control response delay time is defined as: T delay =t res -t cmd The platform's initial response can be determined by the mutation trend collected by the speed sensor or height sensor. The shorter the delay, the faster the system response. The unit is seconds (s).

[0154] In this embodiment, S5 includes the following steps:

[0155] S51. Based on the real-time perturbation potential feature vector z t , calculate the current disturbance intensity index It , where the disturbance intensity index I t By real-time perturbation of the latent feature vector z t The second norm of is used to measure the overall energy amplitude of the disturbance experienced by the rocket platform under the current sea conditions;

[0156] S52. Construct a secretary bird population consisting of M individuals in the multidimensional search space P of the lifting control parameters, and use real-time perturbation to adjust the potential feature vector z t To inspire information, through the weight matrix W (m) With the bias vector b (m) After the linear transformation of , the perturbation mapping coefficient Γ is obtained by Sigmoid function compression. (m) , then according to the lower limit vector p min With the upper limit vector p max The initial control parameter vector of each secretary bird is generated by the interval proportional mapping method Complete the disturbance-driven initialization of the secretary bird population;

[0157] S53. Based on the disturbance intensity index I t Set the adaptive step size adjustment factor α t , adaptive step size adjustment factor α t By the basic step size α base Together with the step-size attenuation coefficient κ, when the disturbance intensity index I t When it increases, the adaptive step size adjustment factor α t Decrease in an inverse relationship;

[0158] S54. During the optimization iteration of the secretary bird, for each secretary bird individual in the kth generation, the current control parameter vector is updated by the jump-tracking strategy. Plus the adaptive step size adjustment factor α t The direction increment mixed with random weights r1 and r2 is realized to achieve the optimal control parameter vector With random individual vector and get closer at the same time, thus obtaining the next generation control parameter vector

[0159]

[0160] in, represents the control performance evaluation function J(p,z t ), is a random individual vector in the current population, r1 and r2 are random weights independently and uniformly sampled in the interval [0,1], which are used to simulate the hunting focus behavior and global reconnaissance behavior of secretary birds;

[0161] S55. Repeat the secretary bird population update until the set termination condition is reached. The termination condition includes the maximum number of iterations or the control performance evaluation function J(p, z t )The improvement is lower than the threshold, and the final output makes the control performance evaluation function J(p,z t ) to obtain the minimum control parameter vector p * , control parameter vector p * As the current optimal lifting control parameter group of the rocket platform lifting execution system.

[0162] In this embodiment, S6 includes the following steps:

[0163] S61. The optimal lifting control parameter vector p * Input the rocket platform lifting execution system to drive the platform to perform the current lifting task. The lifting execution system adjusts the key control signals of lifting rate, attitude angle stability factor, and disturbance compensation gain according to the strategy values corresponding to each control dimension in the optimal control parameter vector;

[0164] S62. During the lifting process, a sensor system fixed to the rocket platform structure is used to monitor the lifting dynamic process in real time. The platform's altitude change curve, attitude change trajectory, attitude offset rate, and lifting time delay platform response variables are recorded within the lifting response cycle to generate the platform lifting response dataset R. exec ;

[0165] S63. Platform lift response dataset R exec Each response data r i Includes the following parameter fields: The platform's current lifting height H i , platform attitude angle change rate Δθ i , control command response time Disturbance compensation error System control status label L i , the meaning of each parameter is as follows:

[0166] Lifting height H i : The actual height of the platform at the i-th sampling moment after executing the control command, in meters (m);

[0167] Attitude angle change rate Δθ i : The change in attitude angle of the platform per unit time during the lifting process, reflecting the stability of the platform, in degrees per second (° / s);

[0168] Control command response time The time delay from when the control command is issued to when the platform completes execution, in seconds (s);

[0169] Disturbance compensation error The deviation between the platform's expected attitude state and its actual state. The unit is consistent with the attitude angle and indicates the accuracy of disturbance compensation.

[0170] System control status label L i : Label the current platform operation status.

[0171] In this embodiment, the tag identifier is defined as follows:

[0172] Normal operating state: disturbance compensation error Response time Attitude angle change rate Δθ i ≤3° / s;

[0173] Disturbance compensation error Response time Or attitude angle change rate 3° / s<Δθ i ≤5° / s;

[0174] Control instability: disturbance compensation error Response time Or attitude angle change rate Δθ i >5° / s.

[0175] Example 1:

[0176] The complete perception-optimization-control process for a certain type of rocket platform encountering complex, multi-scale sea disturbances before launch. During a pre-launch inspection of a rocket platform at sea, the system planned to perform a standard lift simulation to test the platform's responsiveness. The platform was equipped with this invention. During the lift preparation phase, the platform was in a high-frequency disturbance band, and the system continuously recorded multiple disturbance surges.

[0177] At T-60 seconds after the platform is ready for lift, the sensor system on the platform starts, and the three types of sensors deployed at the bottom and four corners of the platform begin sampling data at a frequency of 50 Hz:

[0178] Accelerometer returns data: timestamp t 2218 = 60.3 seconds, three-dimensional acceleration vector a 2218 =(0.42,-0.18,1.54)m / s 2 ; Gyroscope return data: angular velocity vector ω 2218 =(0.05,0.11,-0.07)rad / s; Wave height monitored by sea state instrument: significant wave height h 2218 =2.37m.

[0179] The data constitutes a perturbation sample d 2218 , combined with historical samples to form the original sea state disturbance data set Then it is filtered by sliding window (window size T smooth =0.6s), and using the acceleration rate Δa i , angular velocity change rate Δω i , wave height change rate Δh i Construct perturbation enhancement samples and generate preprocessed dataset D pre .

[0180] The system will preprocess the dataset D pre Input to the multi-scale self-supervised variational contrastive learning model, and set three scales S = {s1 = 0.5s, s2 = 1.0s, s3 = 2.0s} in the model. At each scale:

[0181] Concatenate feature vector x i =a i ||ω i ||h i ||Δa i ||Δω i ||Δh i Apply a scale filter Extracting scale perturbation features Calculate its scale energy For example, the high-frequency scale s3 obtains the maximum energy of 7.94; the scale attention weight is obtained To reflect the importance of this scale disturbance to lift control; extract the long / short disturbance latent vector Calculate the cross-scale attention coefficient and generate the perturbation feature vector f after fusion i The platform outputs the current disturbance feature representation at the current time t=60.3s

[0182] The system starts the dynamic disturbance modeling module at t=60.3s:

[0183] Sliding window T prior = 5s, covering the past 250 samples; extract the disturbance vector set W within the window j ={f 1968 ,...,f 2218}, calculate the prior mean and covariance Prior prediction network P φ Receive f 2218 ,t 2218 Output estimated prior distribution This distribution is used as the basis for the next step of variational inference to complete the update of the dynamic prior variational perturbation encoder.

[0184] The system will 2218 and Combined with reparameterization to generate perturbation potential vector z2218 , the second norm ‖z 2218 ‖2=1.98. 2218 As a benchmark:

[0185] Construct the control parameter vector p = (p1: lifting rate, p2: stability coefficient, p3: disturbance compensation gain) ∈ P; define the evaluation function:

[0186]

[0187] The control goal is to minimize the platform lifting offset and delay time.

[0188] According to the disturbance intensity I 2218 =1.98, the system adaptively sets the step size adjustment factor α 2218 = 0.014, initialize the secretary bird population M = 20: the position of each secretary bird is determined by the perturbation vector z 2218 After projection and control parameter upper and lower limits p min ,p max Interval mapping generation; perform 20 rounds of secretary bird jump-tracking strategy iterations, updating the control vector in each round The optimal parameter vector p is obtained in the 16th round * =(rise and fall rate = 1.07, stability coefficient = 0.83, gain = 0.74).

[0189] The system takes the optimal parameter vector p * Control the lifting platform action, the platform response is as follows: actual lifting height: 2.95m; attitude change rate Δθ=1.28° / s; control delay time T delay =0.92s; compensation error∈ comp =0.34°.

[0190] According to the label rules, the system assesses this round of control as "normal operation status". Response data r 2218 ={H,Δθ,T delay ,∈ comp , label} is added to the response dataset R exec , used for subsequent model optimization and control strategy adjustment.

[0191] This embodiment gradually demonstrates how to start from sensor sampling, go through multi-scale disturbance characterization, dynamic prior update, control space construction, secretary bird optimal search, and finally control distribution and execution feedback, reflecting the feasibility and effectiveness advantages of the present invention in actual engineering deployment.

[0192] The present invention proposes a multi-scale variational contrastive learning architecture for disturbance perception. By introducing scale energy distribution and a self-attention cross-scale collaborative mechanism, the model can extract the temporal variation patterns of multimodal information of acceleration, angular velocity, and significant wave height from different frequency disturbance response windows, and integrates a variational sampling mechanism to obtain a stable representation. On this basis, a dynamic disturbance prior distribution network constructed by embedding time series replaces the traditional fixed standard normal distribution. By predictively modeling the evolution trend of the disturbance distribution, the model enhances its sensitivity and stability to the temporal characteristics of sea state disturbances. This effectively solves the problems of traditional disturbance feature extraction methods' insufficient ability to recognize unstructured strong noise signals and their overly static prior assumptions about disturbances, significantly improving disturbance perception accuracy and system robustness.

[0193] The present invention constructs a secretary bird population initialization strategy guided by the potential eigenvector of disturbance, and dynamically generates the initialization control parameter position distribution in combination with the disturbance intensity, so that the secretary bird population starts from the highly correlated area of the disturbance state, skipping the inefficient exploration process caused by a large number of random initializations in the traditional algorithm. At the same time, the step size adjustment coefficient is driven by the disturbance intensity index to construct a two-way closed-loop mechanism of disturbance perception-optimization search, which dynamically converges to a high-precision local area when the sea conditions change drastically, and maintains global search capabilities under stable disturbance conditions. The improved strategy is significantly better than the traditional secretary bird and particle swarm fixed step size optimization algorithms, and the control parameter optimization speed is improved in high-dimensional parameter space and nonlinear control function scenarios.

[0194] The present invention establishes a direct feedback channel between the disturbance perception model and the platform lifting execution system. The attitude change rate, wave height response, and control delay indicators generated in real time during the lifting execution process constitute the platform lifting response data set, and the operating status label is defined according to the disturbance compensation error and attitude stability. It is not only used to evaluate the effectiveness of the current control strategy, but also to reversely apply structural feedback information to the dynamic prior prediction network and secretary bird optimizer to update the control strategy and disturbance prior model online, realizing intelligent adaptive regulation under disturbance drive. This linked closed-loop system is significantly superior to the traditional "perception-control-execution" one-way architecture and has extremely high dynamic adaptability and engineering practicality.

[0195] The above description is only a preferred specific embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any technician familiar with the technical field, within the technical scope disclosed by the present invention, who makes equivalent replacements or changes based on the technical solution and inventive concept of the present invention, should be covered by the scope of protection of the present invention.

Claims

1. A sea-condition adaptive lifting control method for a rocket platform based on self-supervised learning, characterized in that: The steps include: S1. Synchronously deploy an acquisition system on the rocket platform to continuously acquire and preprocess the sea state disturbance dataset to generate a preprocessed sea state disturbance dataset; S2. Construct a multi-scale self-supervised variational contrastive learning model and perform self-supervised training on the pre-processed sea state disturbance dataset as input, outputting a set of multi-scale disturbance feature representations. S3. Construct a dynamic prior distribution model of the sea state disturbance of the rocket platform based on the multi-scale disturbance feature representation set. This dynamic prior distribution model is embedded in the multi-scale self-supervised variational contrastive learning model to complete the model update, thus obtaining a dynamic prior variational disturbance encoder. S4. Use a dynamic prior variational disturbance encoder to infer the current preprocessed sea state disturbance dataset, output a real-time sea state disturbance feature vector, and construct a multidimensional search space for the lifting control parameters based on the real-time sea state disturbance feature vector and the multi-scale disturbance feature representation set, and set a control performance evaluation function; S5. Using the real-time sea state disturbance characteristic vector to guide the initial position of the secretary bird population, the dynamic guided secretary bird optimization algorithm is started to perform a global search in the multidimensional search space of the lift control parameters, and the optimal lift control parameter set is obtained through iterative calculation of the dynamic guided secretary bird optimization algorithm; S6. Input the optimal lifting control parameter group into the rocket platform lifting execution system and complete the platform lifting control to generate platform lifting response data.

2. The method for sea-condition adaptive lifting control of a rocket platform based on self-supervised learning according to claim 1, characterized in that: Said S1 comprises the following steps: S11. Accelerometers, gyroscopes, and sea state monitors are placed at key structural locations on the rocket platform. Sampling intervals and total sampling durations are set to continuously collect disturbance signals from the rocket platform in a dynamic sea state environment. This generates an original sea state disturbance dataset. Each disturbance sampling data item d in the original sea state disturbance dataset is i Including timestamp t i , three-dimensional acceleration vector a i , three-dimensional angular velocity vector ω i And the instantaneous effective wave height h i , the total number of perturbation data is N raw =T total / Δt; S12. Preprocess the original sea state disturbance data set using a time window length of T smooth The sliding filtering method is used to smooth the three-dimensional acceleration vector, three-dimensional angular velocity vector and instantaneous effective wave height value, eliminating the high-frequency disturbance components caused by sensor sampling jitter. Based on the adaptive threshold mechanism, the signal outliers are detected and eliminated, and atypical disturbance samples affected by external abnormal interference are excluded to obtain the intermediate processed data set after filtering and denoising. S13. Perform disturbance feature enhancement on the intermediate processing data set, calculate the disturbance change rate feature between continuous disturbance data according to the time step, and extract the three-dimensional acceleration change rate Δa i , three-dimensional angular velocity change rate Δω i and the effective wave height change rate Δh i , respectively represent the acceleration fluctuation trend, attitude disturbance fluctuation trend and sea state disturbance fluctuation trend of the rocket platform at the current moment relative to the previous moment. The disturbance change rate feature is spliced with the three-dimensional acceleration vector, three-dimensional angular velocity vector and instantaneous effective wave height value at the corresponding moment to form an enhanced disturbance data sample set, which is used as the final preprocessed sea state disturbance data set D pre .

3. The method for sea-condition adaptive lifting control of a rocket platform based on self-supervised learning according to claim 2, characterized in that: The S2 comprises the following steps: S21. Construct a multi-scale self-supervised variational contrastive learning model to preprocess the sea state disturbance dataset D pre As input, according to the multi-frequency oscillation characteristics contained in the disturbance sample, a scale set S is set. Each scale s in the scale set S is k The response window corresponding to the different frequency disturbance characteristics of the rocket platform lifting system is obtained through the scale filter Acting on the channel splicing feature vector to construct the scale perturbation feature signal S22. Calculate the energy of the scale disturbance characteristic signal to measure its potential impact on the rocket platform's lifting stability. Define the scale energy as: in, Denotes the perturbation sample d i In scale s k The total energy under this scale represents the driving amplitude of the disturbance signal on the dynamic response capability of the rocket platform at this scale; S23. Normalize the scale energy to obtain the scale energy weight of the perturbation sample S24. Perturb the characteristic signal at each scale Set up and parallel input global perturbation encoder With local perturbation encoder Extract the slow-changing trend disturbance and instantaneous disturbance during the lifting and lowering of the rocket platform respectively: in, Indicates scale s k The mean vector of the perturbed feature distribution, Scale s k The standard deviation vector of the perturbed feature distribution; S25. Generate standard normal noise independently for each scale based on the reparameterization strategy and Combined with the corresponding mean vector and standard deviation vector, the long-term perturbation latent vector is obtained and the instantaneous perturbation latent vector S26. Calculate the long-term perturbation latent vector and the instantaneous perturbation latent vector The cosine similarity between them, combined with the latent variable dimension d z Perform temperature scaling and then apply the softmax function to obtain the cross-scale attention coefficient The cross-scale attention coefficient reflects the mutual driving relationship of disturbances at different scales and is used to construct a coordinated ascending and descending regulatory signal pathway; S27. According to scale energy weight and cross-scale attention coefficient For instantaneous perturbation latent vector Perform weighted summation to generate the perturbation fusion vector f i ,The disturbance fusion vector integrates the comprehensive impact of the sea disturbance of various scales on the lifting and lowering states of the rocket platform at the current moment; S28. Fusion perturbation vector f i Input disturbance stability prediction head P θ , output the predicted value of future wave height change slope The actual wave height change trend s of the rocket platform under the current disturbance i and predicted value The mean square error between them defines the stability prediction loss L p , used to measure the accuracy of predicting the future stability trend of the rocket platform; S29. By jointly optimizing the multi-scale variational reconstruction loss L v , self-supervised contrast loss L c and stability prediction loss L p The adjustment coefficients λ1 and λ2 are used to control the relative weights of each loss in the sea disturbance perception task, and the parameters of the multi-scale self-supervised variational contrastive learning model are updated: L total =L v +λ1L c +λ2L p ; Among them, λ1 and λ2 are the adjustment coefficients of contrastive learning and prediction loss, respectively, which control the relative weights of feature learning and stability judgment in the disturbance perception task; S210. After completing the joint training, output the multi-scale perturbation feature representation set 4. The method for sea-condition adaptive lifting control of a rocket platform based on self-supervised learning according to claim 3, characterized in that: The S3 includes the following steps: S31. Multi-scale perturbation feature representation set F multi , according to the disturbance sample timestamp t i Construct the disturbance feature time series trajectory and fuse each disturbance vector f i Its corresponding timestamp t i The disturbance characteristic time series state trajectory set T is formed to characterize the evolution characteristics of the sea state disturbance of the rocket platform in a continuous time period; S32. Set a fixed-length sliding time window T on the disturbance feature time series state trajectory set T prior , the disturbance feature time series state trajectory set is divided into several local sub-window sequences of disturbance using a sliding time window, and each local sub-window sequence of disturbance contains the time period [t j ,t j +T prior ) in the fused perturbation feature vectors, recorded as the perturbation feature sub-window set W j , used to extract the local disturbance distribution pattern of the rocket platform at different stages of sea state change; S33. For each disturbance feature sub-window set W j The fused disturbance feature vector in is counted and the mean vector of all disturbance features in the disturbance feature sub-window set is calculated. and the covariance matrix They are used to describe the central trend of disturbance and the diffusion range of disturbance characteristics of the rocket platform in the current time period respectively; S34. Constructing a priori prediction network P φ , the perturbation sample timestamp t i and the corresponding perturbation fusion vector f i Input, output estimated prior distribution parameters To estimate the potential response structure of the sea state disturbance of the rocket platform at the current moment, and define the prior estimation error loss function with KL divergence: in, are the posterior distribution parameters calculated by the multi-scale encoder at the current moment, is the estimated prior distribution parameter of the prior prediction network output, and the KL divergence measures the structural deviation between the current state of the platform and the predicted disturbance trend; S35. Estimated prior distribution parameters of the prior prediction network output It is embedded in the multi-scale self-supervised variational contrastive learning model, replacing the standard normal distribution as a new variational prior distribution, completing the dynamic structural update of the disturbance perception model during the rocket platform lifting and lowering process, and forming a dynamic prior variational perturbation encoder.

5. The method for sea-condition adaptive lifting control of a rocket platform based on self-supervised learning according to claim 4, characterized in that: The S4 comprises the following steps: S41. Based on the dynamic prior variational disturbance encoder, receive the latest enhanced disturbance data sample d in the current preprocessed sea state disturbance dataset t , and its channel splicing feature x t As input, infer the posterior distribution parameters of the disturbance at the current moment At the same time, the current prior distribution parameters are generated by the prior prediction network Based on the dynamic prior distribution, a reparameterization strategy is used to generate the current moment's real-time perturbation potential feature vector: Among them, z t is the real-time disturbance potential feature vector of the platform in the multi-scale disturbance space at the current moment, ∈ t is a standard normal noise vector, which represents the abstract representation of the current sea state disturbance during the ascent and descent of the rocket platform; S42. Perturb the potential feature vector z in real time t and the multi-scale perturbation feature representation set F multi Perform scale alignment and obtain attention weights through the cosine similarity-softmax mechanism adjusted by the temperature coefficient τ Then use the attention weight to perturb the fusion vector f i Weighted summation to obtain the disturbance driven fusion vector S43. Fusion vector driven by disturbance Construct a multi-dimensional search space P for lifting control parameters. The multi-dimensional search space for lifting control parameters is composed of the control parameter vector p = (p1, p2, ..., p d ), each control parameter p in the control parameter vector i are constrained to the corresponding minimum value With the maximum value between; S44. Define the control performance evaluation function J(p, z t ), the control performance evaluation function is composed of the height fluctuation residual ΔH res (p,z t ), attitude stability error Δθ stab (p,z t ) and control response delay time T delay (p,z t ) is composed of the sum of three weighted squares, with weight coefficients λ1, λ2, and λ3 respectively, which are used to comprehensively measure the lift control accuracy and response efficiency of the rocket platform under the current potential disturbance state.

6. The method for sea-state adaptive lifting control of a rocket platform based on self-supervised learning according to claim 5, characterized in that: The S5 comprises the following steps: S51. Based on the real-time perturbation potential feature vector z t , calculate the current disturbance intensity index I t , where the disturbance intensity index I t By real-time perturbation of the latent feature vector z t The second norm of is used to measure the overall energy amplitude of the disturbance experienced by the rocket platform under the current sea conditions; S52. Construct a secretary bird population consisting of M individuals in the multidimensional search space P of the lifting control parameters, and use real-time perturbation to adjust the potential feature vector z t To inspire information, through the weight matrix W (m) With the bias vector b (m) After the linear transformation of , the perturbation mapping coefficient Γ is obtained by Sigmoid function compression. (m) , then according to the lower limit vector p min With the upper limit vector p max The initial control parameter vector of each secretary bird is generated by the interval proportional mapping method Complete the disturbance-driven initialization of the secretary bird population; S53. Based on the disturbance intensity index I t Set the adaptive step size adjustment factor α t , adaptive step size adjustment factor α t By the basic step size α base Together with the step-size attenuation coefficient κ, when the disturbance intensity index I t When it increases, the adaptive step size adjustment factor α t Decrease in an inverse relationship; S54. During the optimization iteration of the secretary bird, for each secretary bird individual in the kth generation, the current control parameter vector is updated by the jump-tracking strategy. Plus the adaptive step size adjustment factor α t The direction increment mixed with random weights r1 and r2 is realized to achieve the optimal control parameter vector With random individual vector and get closer at the same time, thus obtaining the next generation control parameter vector in, represents the control performance evaluation function J(p,z t ), is a random individual vector in the current population, r1 and r2 are random weights independently and uniformly sampled in the interval [0,1], which are used to simulate the hunting focus behavior and global reconnaissance behavior of secretary birds; S55. Repeat the secretary bird population update until the set termination condition is reached. The termination condition includes the maximum number of iterations or the control performance evaluation function J(p, z t )The improvement is lower than the threshold, and the final output makes the control performance evaluation function J(p,z t ) to obtain the minimum control parameter vector p * , control parameter vector p * As the current optimal lifting control parameter group of the rocket platform lifting execution system.

7. The method for sea-state adaptive lifting control of a rocket platform based on self-supervised learning according to claim 1, characterized in that: The S6 comprises the following steps: S61. The optimal lifting control parameter vector p * Input the rocket platform lifting execution system to drive the platform to perform the current lifting task. The lifting execution system adjusts the key control signals of lifting rate, attitude angle stability factor, and disturbance compensation gain according to the strategy values corresponding to each control dimension in the optimal control parameter vector; S62. During the lifting process, a sensor system fixed to the rocket platform structure is used to monitor the lifting dynamic process in real time. The platform's altitude change curve, attitude change trajectory, attitude offset rate, and lifting time delay platform response variables are recorded within the lifting response cycle to generate the platform lifting response dataset R. exec ; S63. Platform lift response dataset R exec Each response data r i Includes the following parameter fields: The platform's current lifting height H i , platform attitude angle change rate Δθ i , control command response time Disturbance compensation error System control status label L i , system control status label L i The expression labels the current platform operation status.

8. The method for sea-state adaptive lifting control of a rocket platform based on self-supervised learning according to claim 1, characterized in that: The tag identifier is defined as follows: Normal operating state: disturbance compensation error Response time Attitude angle change rate Δθ i ≤3° / s; Disturbance compensation error Response time Or attitude angle change rate 3° / s<Δθ i ≤5° / s; Control instability: disturbance compensation error Response time Or attitude angle change rate Δθ i >5° / s.

Images