Method for simulation and verification of complex system launch control logic

Abstract

The application relates to the field of computer simulation and automatic control technology, and discloses a complex system launch control logic simulation and verification method.The method comprises the following steps: fusing multi-source flight states, environment perception and task information to construct a multi-dimensional state space representation; outputting candidate control rules through a logic generation model fused by a graph neural network and an attention mechanism; carrying out double verification by combining formal semantic checking and safety reachability analysis; after verification, smoothly deploying new logic in an isolated sandbox, and online fine-tuning the model based on execution feedback.The system comprises five modules, namely, multi-source data fusion, logic generation, formal verification, logic deployment and monitoring and updating.The application realizes dynamic evolution and safety closed-loop optimization of launch control logic, and significantly improves the task continuity and control reliability of a complex system under non-expected working conditions.

Description

Technical Field

[0001] This invention belongs to the field of computer simulation and automatic control technology, specifically relating to a method for simulating and verifying the launch control logic of complex systems. Background Technology

[0002] With the increasing demands for mission reliability and environmental adaptability from complex systems such as aerospace, missile defense, and hypersonic vehicles, launch control logic, as a core component ensuring system safety and precise mission execution, is increasingly facing critical technological bottlenecks in terms of its intelligence and dynamic response capabilities. Traditional launch control systems generally employ a fixed, static logic architecture, with control strategies fully determined and programmed into dedicated controllers during ground testing. While this approach offers high stability under normal operating conditions, it essentially relies on prior exhaustive enumeration of mission scenarios and failure modes, making it difficult to cope with sudden extreme failures (such as sensor failure or propulsion system anomalies) or dynamic mission changes (such as target redirection or temporary route adjustments) during flight. Once the actual flight conditions exceed preset boundaries, the system lacks the ability to perceive, evaluate, and reconstruct the control logic online, which can easily lead to mission interruption or even catastrophic consequences.

[0003] Dynamic reconfiguration technology for launch control logic aims to overcome the limitations of traditional static architectures. By transforming control rules from hard-coded forms into runtime-modifiable software modules, it enables real-time responses to changes in flight status. The core of this approach lies in empowering the flight management computer with the ability to autonomously adjust, combine, and verify the control logic while ensuring system safety and determinism. In recent years, some research has attempted to introduce finite state machines or rule engines to enhance flexibility; however, these methods suffer from coarse reconfiguration granularity, reliance on manually pre-built rule bases, and lack of deep coupling with the aircraft's real-time model state and historical data, making it difficult to support multi-dimensional collaborative decision-making under complex fault conditions.

[0004] In existing technologies, control logic updates typically rely on ground station command injection or preset redundant path switching, which suffers from inherent drawbacks such as high communication latency, limited coverage, and inability to handle unknown faults. Meanwhile, completely entrusting the online generation of control logic to a general-purpose AI agent faces risks such as poor interpretability, verification difficulties, and uncontrollable security. Furthermore, existing simulation verification systems are mostly conducted offline, making it difficult to simulate the complex scenarios of model mismatch, data noise, and concurrent faults encountered in real flight. This results in the online reconstructed logic being executed without sufficient and reliable verification, creating significant safety hazards. Therefore, achieving online dynamic reconstruction of control logic driven by both high-fidelity models and real-time flight data during flight, while simultaneously completing lightweight, high-confidence embedded simulation verification, has become a pressing technical challenge for improving the resilience and autonomy of complex systems. Summary of the Invention

[0005] This invention provides a method for simulating and verifying launch control logic for complex systems, aiming to solve the technical problems that fixed launch control logic cannot cope with extreme faults or mission changes beyond the preset range during flight, and lacks online self-adjustment and reconfiguration capabilities. The method constructs an adaptive launch control logic system integrating dynamic modeling, real-time reasoning, logic reconfiguration, and closed-loop verification. By performing multi-source sensing and fusion analysis of system state, environmental disturbances, and mission objectives during on-orbit operation, it drives the online evolution and safety verification of the control logic, thereby ensuring the mission continuity and control reliability of complex systems under unexpected operating conditions.

[0006] This invention provides a method for simulating and verifying the launch control logic of complex systems, comprising:

[0007] Acquire the current flight state sequence, real-time environmental perception data, and current mission planning information of complex systems;

[0008] Based on the current flight state sequence, real-time environmental perception data, and current mission planning information, a multi-dimensional state space representation is constructed.

[0009] The multi-dimensional state space representation is input into a pre-trained logic generative neural network model to generate a set of candidate emission control logic rules;

[0010] Formal semantic consistency verification and security boundary constraint verification are performed on the candidate launch control logic rule set;

[0011] After verification, the candidate launch control logic rule set is loaded into the flight control execution unit, and the logic version identifier is updated synchronously.

[0012] During subsequent control cycles, the execution effect of the candidate launch control logic rule set is continuously monitored, and the logic generator neural network model is fine-tuned online based on the feedback signals.

[0013] Preferably, the acquisition of the current flight state sequence, real-time environmental perception data, and current mission planning information of the complex system specifically includes:

[0014] Real-time three-dimensional position coordinates, velocity, and acceleration data of complex systems are obtained through a global positioning system module.

[0015] The attitude angles (roll angle, pitch angle, yaw angle), angular velocity, and angular acceleration data of complex systems are obtained through the inertial measurement unit module.

[0016] The atmospheric data sensor module acquires air pressure, temperature, humidity, wind speed and direction data for the airspace where the complex system is located.

[0017] Acquire spatial distribution, relative position, and relative velocity data of obstacles around a complex system using airborne vision or lidar sensors;

[0018] Obtain the target waypoint, flight mode, and no-fly zone information for the current flight mission through the mission planning interface.

[0019] Preferably, the construction of the multi-dimensional state space representation specifically includes:

[0020] The current flight status sequence, real-time environmental perception data, and current mission planning information are aligned according to a unified time base to form a synchronized time-series data stream;

[0021] Normalize the data of each dimension in the synchronized time-series data stream to eliminate dimensional differences;

[0022] A sliding window mechanism is used to extract continuous data segments of a preset window size to form a state observation window;

[0023] Principal component analysis is performed on the data within the state observation window to reduce dimensionality, retaining principal components with a cumulative variance contribution rate of not less than 95%, and generating a compressed state feature vector.

[0024] The compressed state feature vector is concatenated with the task target code to form a complete multi-dimensional state space representation.

[0025] Preferably, the logic generative neural network model is a sequence-to-rule generation model based on a fusion architecture of graph neural network and attention mechanism. Its input layer receives the multi-dimensional state space representation, the intermediate layer extracts the high-order dependencies between state elements through a multi-head self-attention mechanism, and uses graph convolutional layers to model the topological constraints between control variables. The output layer generates a structured emission control logic rule set, which is expressed in the form of predicate logic and includes a triplet of preconditions, action instructions and execution priorities.

[0026] Preferably, the formal semantic consistency check specifically includes:

[0027] The candidate launch control logic rule set is converted into a first-order predicate logic formula;

[0028] Based on a predefined axiom library of system behavior, an automatic theorem prover is used to verify whether the first-order predicate logic formula is compatible with the axiom library;

[0029] If a logical conflict exists, the conflicting rules are marked and returned to the logical generative neural network model for correction and iteration until a conflict-free rule set is generated.

[0030] Preferably, the security boundary constraint verification specifically includes:

[0031] Establish a multi-dimensional set of safety constraints that includes flight envelope limits, structural load thresholds, energy margin lower limits, and communication link reliability.

[0032] Map the candidate launch control logic rule set to a discrete event system model;

[0033] Perform reachability analysis on the discrete event system model to determine whether all possible state transition paths satisfy the multidimensional set of security constraints.

[0034] If a state transition path that violates any security constraint exists, the verification is deemed to have failed, and the rule reconstruction process is triggered.

[0035] Preferably, loading the candidate launch control logic rule set into the flight control execution unit specifically includes:

[0036] Allocate a logic execution sandbox in the isolated memory area of ​​the flight control computer;

[0037] The verified candidate launch control logic rule set is compiled into intermediate representation code and loaded into the logic execution sandbox;

[0038] Initiate the logic switching arbitrator to compare the execution context compatibility of the old and new logic rule sets;

[0039] After confirming that there are no contextual conflicts, smoothly switch control to the new set of logical rules and record the switch timestamp and version identifier.

[0040] Preferably, the execution effect of the continuous monitoring of the candidate launch control logic rule set specifically includes:

[0041] Collect the actual control output, system response deviation, and task achievement indicators during the execution of the acquisition logic.

[0042] Calculate the mean square error between the actual control output and the expected output;

[0043] When the mean square error exceeds the preset tolerance threshold for three consecutive control cycles, the logic execution is deemed abnormal.

[0044] At the same time, if the task completion rate drops by more than 10%, a logic performance degradation alert will be triggered.

[0045] Preferably, the online fine-tuning of the logic generative neural network model based on feedback signals specifically includes:

[0046] The results of performance monitoring are encapsulated into reinforcement learning reward signals;

[0047] A Markov decision process is constructed using the multi-dimensional state space representation as the state input and the candidate launch control logic rule set as the action output.

[0048] The parameter weights of the logical generative neural network model are updated using a proximal strategy optimization algorithm.

[0049] After every 100 control cycles, the fine-tuned model parameters are written to non-volatile memory to form a new baseline version.

[0050] This invention provides a simulation and verification system for launch control logic of complex systems, comprising:

[0051] The multi-source data fusion module is used to acquire the current flight state sequence, real-time environmental perception data and current mission planning information of complex systems, and to construct a multi-dimensional state space representation.

[0052] The logic generation module has a built-in logic generation neural network model, which is used to receive the multi-dimensional state space representation and output a set of candidate emission control logic rules.

[0053] The formal verification module is used to perform formal semantic consistency verification and security boundary constraint verification on the candidate launch control logic rule set;

[0054] The logic deployment module is used to load the candidate launch control logic rule set into the flight control execution unit after verification, and update the logic version identifier;

[0055] The execution monitoring and model update module is used to continuously monitor the execution effect of the candidate launch control logic rule set and to fine-tune the logic generator neural network model online based on feedback signals.

[0056] Preferably, the multi-source data fusion module includes a data synchronization submodule, a normalization submodule, a sliding window submodule, a principal component analysis submodule, and a state splicing submodule. Each submodule is connected in sequence to complete the entire process from raw sensor data to multi-dimensional state space representation.

[0057] Preferably, the logic generation neural network model in the logic generation module is composed of an input embedding layer, a multi-head self-attention layer, a graph convolutional constraint layer, and a rule decoding layer stacked in sequence, and the adjacency matrix of the graph convolutional constraint layer is predefined by the physical coupling relationship between the system control variables.

[0058] Preferably, the formal verification module includes a logic transformation submodule, an axiom matching submodule, a theorem proving submodule, a discrete modeling submodule, and a reachability analysis submodule. The theorem proving submodule is implemented using an SMT solver, and the reachability analysis submodule is based on symbolic model detection technology.

[0059] Preferably, the logic deployment module includes a sandbox allocation unit, an intermediate code compilation unit, a context compatibility check unit, and a smooth switching execution unit. The context compatibility check unit performs verification by comparing the consistency of the domain and value range of shared state variables in the new and old logic rule sets.

[0060] Preferably, the execution monitoring and model update module includes a performance index acquisition unit, an error calculation unit, an anomaly determination unit, and a parameter fine-tuning unit. The parameter fine-tuning unit adopts a gradient pruning strategy to prevent parameter oscillations during the online learning process.

[0061] Compared with the prior art, the beneficial effects of the present invention are as follows:

[0062] This invention overcomes the adaptability bottleneck of traditional fixed logic in the face of extreme failures or sudden task changes by constructing an end-to-end online generation and verification mechanism for launch control logic. The combination of multi-source state perception and principal component analysis enables effective compression and representation of high-dimensional heterogeneous data, providing accurate input for logic generation. The logic generation neural network model integrates graph neural networks and attention mechanisms, enabling explicit modeling of physical constraints and state dependencies between control variables, generating structured rules that conform to the system's dynamic characteristics. The formal verification module introduces first-order predicate logic and discrete event system reachability analysis, ensuring the reliability of the new logic from both semantic consistency and security dimensions, avoiding the introduction of potential risks. Logic deployment employs sandbox isolation and smooth switching mechanisms, ensuring the continuity and stability of the control transfer process. Execution monitoring and online fine-tuning closed-loop enable the system to continuously learn and optimize performance, continuously improving logic adaptability over long-term operation. The overall solution achieves a fundamental transformation of launch control logic from static preset to dynamic self-evolution, significantly improving the survivability and task completion rate of complex systems under unexpected operating conditions. Attached Figure Description

[0063] Figure 1 This is a schematic diagram of the overall technical solution architecture of the present invention;

[0064] Figure 2 This is a schematic diagram illustrating the core principle framework of the logical generative neural network model that integrates graph neural networks and attention mechanisms in this invention.

[0065] Figure 3 This is a logical flowchart of the construction of multi-source data fusion and multi-dimensional state space representation in this invention;

[0066] Figure 4 This is a flowchart illustrating the formal semantic consistency verification and security boundary constraint verification of the candidate launch control logic rule set in this invention.

[0067] Figure 5This is a logical flow diagram of the sandbox loading, context compatibility checking and smooth switching execution of the candidate launch control logic rule set in this invention;

[0068] Figure 6 This is a schematic diagram of the closed-loop feedback interaction relationship and data flow between the logic execution effect monitoring and the online fine-tuning of the logic generation neural network model in this invention. Detailed Implementation

[0069] Please refer to the attached document. Figure 1 To be continued Figure 6 This invention provides a method for simulating and verifying launch control logic for complex systems, aiming to solve the technical problems that fixed launch control logic cannot cope with extreme faults or mission changes beyond the preset range during flight, and lacks online self-adjustment and reconfiguration capabilities. The method constructs an adaptive launch control logic system integrating dynamic modeling, real-time reasoning, logic reconfiguration, and closed-loop verification. Through multi-source perception and fusion analysis of system state, environmental disturbances, and mission objectives during on-orbit operation, it drives the online evolution and safety verification of the control logic, thereby ensuring the mission continuity and control reliability of complex systems under unexpected operating conditions.

[0070] Preferably, the acquisition of the current flight state sequence, real-time environmental perception data, and current mission planning information of the complex system specifically includes: acquiring the real-time three-dimensional position coordinates, velocity, and acceleration data of the complex system through the Global Positioning System module; acquiring the attitude angles (roll angle, pitch angle, yaw angle), angular velocity, and angular acceleration data of the complex system through the Inertial Measurement Unit module; acquiring the air pressure, temperature, humidity, and wind speed and direction data of the airspace where the complex system is located through the atmospheric data sensor module; acquiring the spatial distribution, relative position, and relative velocity data of obstacles around the complex system through airborne vision or lidar sensors; and acquiring the target waypoint, flight mode, and no-fly zone information of the current flight mission through the mission planning interface.

[0071] All the aforementioned sensor modules synchronously acquire data using a unified time reference, with a sampling frequency of no less than 100 Hz to ensure the timeliness and completeness of state observations. Data output from the Global Positioning System (GPS) module is processed using Kalman filtering to eliminate multipath effects and satellite signal noise; attitude angle data output from the Inertial Measurement Unit (INS) module is updated using a quaternion integration algorithm to avoid Euler angle singularity issues; raw air pressure and temperature signals acquired by the atmospheric data sensor module undergo zero-point drift compensation and nonlinear correction; point cloud data output from airborne vision or lidar sensors is preprocessed using voxel grid downsampling and dynamic object culling algorithms to reduce computational load and improve obstacle recognition accuracy; flight mode information received by the mission planning interface includes discrete commands such as hovering, cruise, evasion, and emergency return; no-fly zone information is expressed as a set of polygon vertices in the form of a geofence.

[0072] After completing the initial data acquisition, step S101 is executed: based on the current flight state sequence, real-time environmental perception data, and current mission planning information, a multi-dimensional state space representation is constructed. This step specifically includes: aligning the current flight state sequence, real-time environmental perception data, and current mission planning information according to a unified time reference to form a synchronized time-series data stream; normalizing the data in each dimension of the synchronized time-series data stream to eliminate dimensional differences; using a sliding window mechanism to extract continuous data segments of a preset window size to form a state observation window; performing principal component analysis to reduce the dimensionality of the data within the state observation window, retaining principal components with a cumulative variance contribution rate of not less than 95%, and generating a compressed state feature vector; concatenating the compressed state feature vector with the mission target code to form a complete multi-dimensional state space representation.

[0073] The time alignment process uses hardware timestamps to mark each sensor data frame and linear interpolation to unify all data to the master control clock cycle. Normalization employs a min-max scaling method, mapping each dimension of data to the 0-1 interval; the transformation formula is as follows:

[0074] ;

[0075] These are the normalized data values. The original data, and These represent the minimum and maximum values ​​of the data in this dimension within a historical sliding window, respectively. The window length is set to the most recent 1000 sampling points. The sliding window size is fixed at 200 time steps, corresponding to a two-second observation duration. During principal component analysis dimensionality reduction, the covariance matrix is ​​calculated from all samples within the sliding window, and the top eigenvalues ​​are selected after eigenvalue decomposition. There are principal components such that the cumulative variance contribution rate satisfies:

[0076] ;

[0077] For the first 1 eigenvalue, The original feature dimension is 128. The mission target encoding adopts a one-hot encoding method, which maps the target waypoint index, flight mode category and no-fly zone activation status into fixed-length binary vectors. Finally, these vectors are concatenated with the compressed state feature vectors by channel dimension to form a multi-dimensional state space representation vector with dimension 128.

[0078] Then, step S102 is executed: the multi-dimensional state space representation is input into a pre-trained logic generative neural network model to generate a candidate emission control logic rule set. The logic generative neural network model is a sequence-to-rule generation model based on a fusion architecture of graph neural networks and attention mechanisms. Its input layer receives the multi-dimensional state space representation, the intermediate layer extracts high-order dependencies between state elements through a multi-head self-attention mechanism, and uses graph convolutional layers to model the topological constraints between control variables. The output layer generates a structured emission control logic rule set, which is expressed in predicate logic form and includes a triplet of preconditions, action instructions, and execution priorities.

[0079] The model's input embedding layer first projects the 128-dimensional state vector into a 512-dimensional latent space; the multi-head self-attention layer is configured with 8 attention heads, each with a dimension of 64, and calculates the correlation weights between state elements through query, key, and value matrices. The attention score is calculated using the following formula:

[0080] ;

[0081] These are query, key, and value matrices, respectively. The key vector dimension is used. The adjacency matrix of the graph convolution constraint layer is predefined by the physical coupling relationships between system control variables. For example, there is a strong coupling between the thrust vector and the control surface deflection angle, while the coupling between the energy management module and the navigation module is weaker. The graph convolution operation is performed according to the following formula:

[0082] ;

[0083] For the first The node feature matrix of the layer (each row corresponds to the feature of a node). Given an adjacency matrix with self-loops, Its degree matrix, For the first Layer node features For learnable weight matrix, It is the ReLU activation function. This is the original adjacency matrix. It is an identity matrix. The rule decoding layer adopts a gated loop unit structure, generating predicate logic rules symbol by symbol. The output format strictly follows the syntax of "IF<condition>THEN<action>PRIORITY<value>". The condition part consists of state variable comparisons, and the action part consists of atomic operations of control instructions. The priority value is an integer from 1 to 10.

[0084] After generating the candidate rule set, step S103 is executed: formal semantic consistency verification and security boundary constraint verification are performed on the candidate emission control logic rule set. The formal semantic consistency verification specifically includes: converting the candidate emission control logic rule set into a first-order predicate logic formula; verifying whether the first-order predicate logic formula is compatible with the axiom library based on a predefined system behavior axiom library using an automatic theorem prover; if logical conflicts exist, the conflicting rules are marked and returned to the logic generative neural network model for correction iteration until a conflict-free rule set is generated.

[0085] The system behavior axiom library is stored in read-only memory and includes fundamental constraints such as conservation laws (e.g., conservation of momentum and energy), mutual exclusion of control laws (e.g., prohibiting the simultaneous activation of reverse thrust and forward thrust), and state reachability (e.g., prohibiting direct entry from hovering to supersonic cruise). The automatic theorem prover uses an SMT solver, supporting linear real arithmetic and uninterpreted function theory. The logic transformation process maps comparison operators (e.g., greater than, less than, or equal to) in predicate rules to SMT logical expressions, and action instructions to state transition functions. If the solver returns "unsatisfiable," a semantic conflict is determined to exist.

[0086] The safety boundary constraint verification specifically includes: establishing a multi-dimensional safety constraint set that includes flight envelope limits, structural load thresholds, energy margin lower limits, and communication link reliability; mapping the candidate launch control logic rule set to a discrete event system model; performing reachability analysis on the discrete event system model to determine whether all possible state transition paths satisfy the multi-dimensional safety constraint set; if there is a state transition path that violates any safety constraint, the verification is deemed to have failed, and a rule reconstruction process is triggered.

[0087] Flight envelope limits are defined using a two-dimensional lookup table of speed and altitude. Structural load thresholds are represented by the upper limit of the overload factor g. The lower limit of energy margin is set at 20% of the total capacity. Communication link reliability is constrained by both the signal-to-noise ratio threshold and the upper limit of packet loss rate. The discrete event system model treats control logic rules as event triggers, system states as event contexts, and state transitions driven by rule actions. Reachability analysis employs symbolic model detection technology, traversing all possible state paths through Boolean satisfaction solutions to verify that all path endpoints are within safe regions. If an unsafe path is detected, the violation rule number is recorded, and local regeneration is initiated.

[0088] After successful verification, step S104 is executed: the candidate launch control logic rule set is loaded into the flight control execution unit, and the logic version identifier is updated synchronously. This step specifically includes: allocating a logic execution sandbox in the isolated memory area of ​​the flight control computer; compiling the verified candidate launch control logic rule set into intermediate representation code and loading it into the logic execution sandbox; starting the logic switching arbitrator to compare the execution context compatibility of the old and new logic rule sets; and after confirming that there are no context conflicts, smoothly switching control to the new logic rule set and recording the switching timestamp and version identifier.

[0089] The logic execution sandbox employs a memory protection unit to achieve address space isolation, preventing new logic code from disrupting the core flight control process. The intermediate representation code uses a register-based low-level instruction set, supporting conditional jumps, arithmetic operations, and I / O operations. A context compatibility check unit compares the domain and value range of shared state variables in the old and new logic rule sets for consistency. For example, if the old logic uses a "speed" variable with a domain of 0 to 300 meters per second, while the new logic extends it to 0 to 500 meters per second, it is determined to be incompatible. The smooth switching execution unit uses a double-buffering mechanism, atomically switching the instruction pointer at the start of the next control cycle to ensure no instruction tearing. The version identifier uses an incrementing integer, initially set to 1, and incremented by 1 after each successful switch.

[0090] Then, step S105 is executed: During subsequent control cycles, the execution effect of the candidate launch control logic rule set is continuously monitored, and the logic generation neural network model is fine-tuned online based on feedback signals. Continuous monitoring specifically includes: collecting the actual control output, system response deviation, and task achievement index during logic execution; calculating the mean square error between the actual control output and the expected output; determining logic execution anomaly when the mean square error exceeds a preset tolerance threshold for three consecutive control cycles; and triggering a logic performance degradation alarm if the task achievement index decreases by more than 10%.

[0091] Actual control outputs include control surface deflection angle, engine thrust command, and energy allocation ratio; system response deviation is calculated by aligning the actual flight trajectory with the planned trajectory using dynamic time normalization to determine the Euclidean distance; mission achievement indicators are weighted composites of mission completion progress, resource consumption efficiency, and safety margin. The mean square error tolerance threshold is dynamically adjusted according to the current flight phase; for example, the threshold is set to 0.5 during takeoff and 0.2 during cruise.

[0092] The online fine-tuning based on feedback signals specifically includes: encapsulating the execution effect monitoring results into reinforcement learning reward signals; constructing a Markov decision process using the multi-dimensional state space representation as state input and the candidate emission control logic rule set as action output; updating the parameter weights of the logic generative neural network model using a proximal policy optimization algorithm; and writing the fine-tuned model parameters into a non-volatile memory after every 100 control cycles to form a new baseline version.

[0093] The reward signal is designed as a composite function: task achievement contributes a positive reward, control deviation contributes a negative reward, and a significant penalty is imposed when the safety margin falls below a threshold. The loss function of the near-end policy optimization algorithm includes a policy gradient term and a pruning term, and its update formula is:

[0094] ;

[0095] This is the pruning advantage loss (one of the core loss terms of the PPO algorithm). Time step The expected value calculation (i.e., averaging the loss over all time steps). The probability ratio between the old and new strategies. For the estimation of the advantage function, The clipping parameter is set to 0.2. Gradient clipping is enabled during parameter fine-tuning, with a threshold set to 1.0 to prevent parameter oscillation. This is the pruning function. The non-volatile memory uses ferroelectric memory with a write endurance of no less than 1 billion cycles, ensuring the feasibility of long-term online learning.

[0096] The steps S101 to S105 described above constitute a complete adaptive closed loop for launch control logic. The entire process is scheduled on the flight control computer using a real-time operating system, with each control cycle taking no more than 20 milliseconds, meeting the timeliness requirements of highly dynamic flight missions.

[0097] As another aspect of the present invention, a simulation and verification system for launch control logic of a complex system is provided, comprising: a multi-source data fusion module for acquiring the current flight state sequence, real-time environmental perception data, and current mission planning information of the complex system, and constructing a multi-dimensional state space representation; a logic generation module, which has a built-in logic generation neural network model for receiving the multi-dimensional state space representation and outputting a candidate launch control logic rule set; a formal verification module for performing formal semantic consistency verification and security boundary constraint verification on the candidate launch control logic rule set; a logic deployment module for loading the candidate launch control logic rule set into the flight control execution unit after verification and updating the logic version identifier; and an execution monitoring and model update module for continuously monitoring the execution effect of the candidate launch control logic rule set and fine-tuning the logic generation neural network model online based on feedback signals.

[0098] The multi-source data fusion module comprises a data synchronization submodule, a normalization submodule, a sliding window submodule, a principal component analysis submodule, and a state stitching submodule. These submodules are connected sequentially to complete the entire process from raw sensor data to a multi-dimensional state space representation. The data synchronization submodule receives timestamp data streams from various sensors and outputs aligned synchronization data frames. The normalization submodule maintains a queue of historical extreme values ​​for each dimension of data and updates scaling parameters in real time. The sliding window submodule uses a circular buffer structure to efficiently manage time-series data. The principal component analysis submodule periodically updates the covariance matrix and eigenvectors in a background thread. The state stitching submodule merges the dimensionality-reduced features and task encoding into a fixed-length vector.

[0099] The logic generation module's logic generation neural network model consists of an input embedding layer, a multi-head self-attention layer, a graph convolutional constraint layer, and a rule decoding layer stacked sequentially. The adjacency matrix of the graph convolutional constraint layer is predefined by the physical coupling relationship between the system's control variables. This model is deployed on a dedicated neural network accelerator, supports fixed-point quantization inference, and has a computational precision of 8-bit integers.

[0100] The formal verification module includes a logic transformation submodule, an axiom matching submodule, a theorem proving submodule, a discrete modeling submodule, and a reachability analysis submodule. The theorem proving submodule is implemented using an SMT solver, and the reachability analysis submodule is based on symbolic model checking technology. All submodules run on independent secure core processors, physically isolated from the main control core, ensuring that the verification process is not interfered with by the main system.

[0101] The logic deployment module includes a sandbox allocation unit, an intermediate code compilation unit, a context compatibility check unit, and a smooth switching execution unit. The context compatibility check unit verifies the consistency of the domain and value range of shared state variables in the old and new logic rule sets. The sandbox allocation unit works in conjunction with the memory management unit to ensure that the address space of the new logic code does not overlap with critical system areas.

[0102] The execution monitoring and model update module includes a performance metric acquisition unit, an error calculation unit, an anomaly detection unit, and a parameter fine-tuning unit. The parameter fine-tuning unit employs a gradient pruning strategy to prevent parameter oscillations during online learning. This module shares the model parameter storage area with the logic generation module, and achieves atomicity of parameter updates through a double buffering mechanism.

[0103] This embodiment fully discloses the technical solution of the present invention, covering the entire lifecycle management from state awareness, logic generation, formal verification, secure deployment to execution feedback, realizing online adaptive evolution of launch control logic, and effectively solving the problem of insufficient adaptability of traditional solidified logic under extreme conditions.

Claims

1. A method for simulation and verification of launch control logic in complex systems, characterized in that, include: Acquire the current flight state sequence, real-time environmental perception data, and current mission planning information of complex systems; Based on the current flight state sequence, real-time environmental perception data, and current mission planning information, a multi-dimensional state space representation is constructed. The multi-dimensional state space representation is input into a pre-trained logic generative neural network model to generate a set of candidate emission control logic rules; Formal semantic consistency verification and security boundary constraint verification are performed on the candidate launch control logic rule set; After verification, the candidate launch control logic rule set is loaded into the flight control execution unit, and the logic version identifier is updated synchronously. During subsequent control cycles, the execution effect of the candidate launch control logic rule set is continuously monitored, and the logic generator neural network model is fine-tuned online based on the feedback signals.

2. The method for simulating and verifying the launch control logic of a complex system according to claim 1, characterized in that, Acquire the current flight state sequence, real-time environmental perception data, and current mission planning information of complex systems, including: Real-time three-dimensional position coordinates, velocity, and acceleration data of complex systems are obtained through a global positioning system module. The attitude angle, angular velocity, and angular acceleration data of complex systems are obtained through the inertial measurement unit module. The atmospheric data sensor module acquires air pressure, temperature, humidity, wind speed and direction data for the airspace where the complex system is located. Acquire spatial distribution, relative position, and relative velocity data of obstacles around a complex system using airborne vision or lidar sensors; Obtain the target waypoint, flight mode, and no-fly zone information for the current flight mission through the mission planning interface.

3. The method for simulating and verifying the launch control logic of a complex system according to claim 2, characterized in that, Based on the current flight state sequence, real-time environmental perception data, and current mission planning information, a multi-dimensional state space representation is constructed, including: The current flight status sequence, real-time environmental perception data, and current mission planning information are aligned according to a unified time base to form a synchronized time-series data stream; Normalize the data of each dimension in the synchronized time-series data stream to eliminate dimensional differences; A sliding window mechanism is used to extract continuous data segments of a preset window size to form a state observation window; Principal component analysis is performed on the data within the state observation window to reduce dimensionality, retaining principal components with a cumulative variance contribution rate of not less than 95%, and generating a compressed state feature vector. The compressed state feature vector is concatenated with the task target code to form a complete multi-dimensional state space representation.

4. The method for simulating and verifying the launch control logic of a complex system according to claim 3, characterized in that, The multi-dimensional state space representation is input into a pre-trained logic generative neural network model to generate a set of candidate emission control logic rules, including: The multi-dimensional state space representation is input into a logical generative neural network model consisting of an input embedding layer, a multi-head self-attention layer, a graph convolutional constraint layer, and a rule decoding layer stacked sequentially. The higher-order dependencies between state elements are extracted through the multi-head self-attention layer; The graph convolution constraint layer models the topological constraints between control variables, and the adjacency matrix of the graph convolution constraint layer is predefined by the physical coupling relationship between the system control variables. The rule decoding layer outputs a set of structured emission control logic rules expressed in predicate logic form, which includes a triplet of preconditions, action instructions, and execution priorities.

5. The method for simulating and verifying the launch control logic of a complex system according to claim 4, characterized in that, Perform formal semantic consistency verification on the candidate launch control logic rule set, including: The candidate launch control logic rule set is converted into a first-order predicate logic formula; Based on a predefined axiom library of system behavior, an automatic theorem prover is used to verify whether the first-order predicate logic formula is compatible with the axiom library; If a logical conflict exists, the conflicting rules are marked and returned to the logical generative neural network model for correction and iteration until a conflict-free rule set is generated.

6. The method for simulating and verifying the launch control logic of a complex system according to claim 5, characterized in that, Perform security boundary constraint verification on the candidate launch control logic rule set, including: Establish a multi-dimensional set of safety constraints that includes flight envelope limits, structural load thresholds, energy margin lower limits, and communication link reliability. Map the candidate launch control logic rule set to a discrete event system model; Perform reachability analysis on the discrete event system model to determine whether all possible state transition paths satisfy the multidimensional set of security constraints. If a state transition path that violates any security constraint exists, the verification is deemed to have failed, and the rule reconstruction process is triggered.

7. The method for simulating and verifying the launch control logic of a complex system according to claim 6, characterized in that, After verification, the candidate launch control logic rule set is loaded into the flight control execution unit, and the logic version identifier is updated synchronously, including: Allocate a logic execution sandbox in the isolated memory area of ​​the flight control computer; The verified candidate launch control logic rule set is compiled into intermediate representation code and loaded into the logic execution sandbox; Initiate the logic switching arbitrator to compare the execution context compatibility of the old and new logic rule sets; After confirming that there are no contextual conflicts, smoothly switch control to the new set of logical rules and record the switch timestamp and version identifier.

8. The method for simulating and verifying the launch control logic of a complex system according to claim 7, characterized in that, Continuously monitor the execution effect of the candidate launch control logic rule set, including: Collect the actual control output, system response deviation, and task achievement indicators during the execution of the acquisition logic. Calculate the mean square error between the actual control output and the expected output; When the mean square error exceeds the preset tolerance threshold for three consecutive control cycles, the logic execution is deemed abnormal. At the same time, if the task completion rate drops by more than 10%, a logic performance degradation alert will be triggered.

9. The method for simulating and verifying the launch control logic of a complex system according to claim 8, characterized in that, Online fine-tuning of the logic generative neural network model based on feedback signals includes: The results of performance monitoring are encapsulated into reinforcement learning reward signals; A Markov decision process is constructed using the multi-dimensional state space representation as the state input and the candidate launch control logic rule set as the action output. The parameter weights of the logical generative neural network model are updated using a proximal strategy optimization algorithm. After every 100 control cycles, the fine-tuned model parameters are written to non-volatile memory to form a new baseline version.

10. The method for simulating and verifying the launch control logic of a complex system according to claim 9, characterized in that, The normalization process employs a minimum-maximum scaling method, mapping each dimension of data to the interval between 0 and 1. The transformation formula is as follows: ； These are the normalized data values. The original data, and These are the minimum and maximum values ​​of the data in this dimension within the historical sliding window, respectively, with the window length set to the most recent 1000 sampling points.

Images