Aircraft cabin temperature self-adaptive control system based on predictive control

Abstract

The invention discloses an aircraft cabin temperature adaptive control system based on predictive control, and relates to the technical field of temperature adaptive control, and the system comprises a state cleaning module which filters out the noise of a sensor through manifold boundary projection, and outputs a pure bottom layer state vector; the model wake-up module evaluates a high-dimensional prediction model based on information geometric features and performs local directional updating to realize self-adaptive compensation of cabin aging, passenger distribution change and non-uniform heat flow; the prediction module is combined with a port Hamiltonian energy topological equation, a quadratic programming solver is used for calculating an optimal control sequence under absolute passivity constraint, and hot and cold airflow hedging internal friction is avoided; and the execution module dynamically adjusts an error envelope by using the track information entropy, and intercepts or issues a valve control instruction. According to the system, high and low frequency state estimation, sensitivity-driven parameter updating and physical constraint optimization are fused, high-precision, quick response and self-adaptive control over the temperature in the cabin are achieved, and meanwhile physical rationality and the service life of an execution mechanism are guaranteed.

Description

Technical Field

[0001] This invention relates to the field of temperature adaptive control technology, and in particular to an adaptive control system for cabin temperature of an aircraft based on predictive control. Background Technology

[0002] As the requirements for cabin comfort and safety in aerospace vehicles continue to increase, cabin temperature control systems need to simultaneously meet the demands of high-precision regulation, rapid response, and adaptability to changes in the external environment. Existing cabin temperature control methods mainly rely on traditional PID or linear predictive control, which typically assume that the cabin's thermodynamic parameters are known and fixed, neglecting the impact of cabin material aging, changes in occupant distribution, and fluctuations in cabin micro-pressure differentials on heat transfer.

[0003] Furthermore, traditional methods have limited ability to handle cabin temperature sensor noise and unmodeled heat sources (such as solar radiation, equipment heat dissipation, and crew activity heat), which can easily lead to control deviations or response delays. While existing predictive control methods can calculate future temperature trajectories in advance, they lack effective mechanisms for adaptive updates to model parameters, cannot quantify the effective information of observational data on model parameters in real time, and struggle to incorporate physical constraints to ensure that the model output remains within a reliable temperature state space. Especially in complex flight paths and multi-temperature compartments, the non-dissipative characteristics of air convection coupling and heat flow crosstalk are not fully considered, resulting in decreased control accuracy and impacting valve mechanical lifespan. Summary of the Invention

[0004] In view of the aforementioned existing problems, the present invention is proposed.

[0005] Therefore, this invention provides an adaptive control system for cabin temperature based on predictive control to solve the problem of how to achieve high-precision predictive control of cabin temperature while adaptively compensating for cabin aging, changes in occupant distribution, and non-uniform heat flow disturbances.

[0006] To solve the above-mentioned technical problems, the present invention provides the following technical solution: This invention provides an adaptive control system for cabin temperature based on predictive control, which includes a state cleaning module that collects temperature data of multiple temperature zones in the cabin, performs implicit thermal state cleaning based on manifold boundary projection, and outputs a clean bottom-level state vector with noise filtered out. The model wake-up module receives the pure underlying state vector, extracts information geometric features, performs feature-oriented wake-up of the high-dimensional prediction model based on information geometric evaluation, and realizes the update of the high-dimensional prediction model; The prediction module elevates the physical quantities containing the pure underlying state vector to the updated high-dimensional prediction model, forces constraints on the underlying rules to a port Hamiltonian energy topology that conforms to the passive characteristics, and solves for the optimal control sequence and the synchronously generated future temperature trajectory with the goal of minimizing the total energy dissipation rate. The execution module parses meteorological data of the aircraft's forward trajectory to calculate the trajectory information entropy, and constructs a tracking error invariant set whose error radius expands dynamically inversely with the trajectory information entropy. When the future predicted temperature trajectory is completely within the invariant set, the control command is intercepted and the physical actuator is locked; otherwise, the optimal control sequence is converted into a valve action command and issued.

[0007] As a preferred embodiment of the predictive control-based adaptive control system for cabin temperature of an aircraft, the implicit thermal state cleaning based on manifold boundary projection includes parallel operation of a high-frequency state estimation algorithm and a low-frequency physical optimization algorithm. Extended Kalman filter is used as the high-frequency state estimation algorithm. With cabin temperature sensor data as the observation input, the high-frequency real-time estimated state including the implicit thermal load of the cabin is calculated through recursive filtering. The synchronous call to the moving horizon estimation is used as a low-frequency physics optimization algorithm. Based on the temperature data within the historical sliding time window, it solves the nonlinear programming problem constrained by the heat conduction differential equation and outputs the envelope surface formed by the solution set in the multidimensional space as an absolutely reliable physical manifold surface. In each control cycle, the minimum Euclidean geometric distance from the high-frequency real-time estimated state space coordinates to the boundary of the absolutely reliable physical manifold is calculated; when it is determined that the minimum Euclidean geometric distance is greater than the tolerance of the inherent white noise absolute value calibrated by the hardware sensor, the orthogonal projection mechanism is triggered to forcibly pull the deviated state back into the boundary of the manifold.

[0008] As a preferred embodiment of the predictive control-based adaptive control system for cabin temperature of an aircraft described in this invention, the orthogonal projection mechanism involves calling a quadratic programming solver to perform orthogonal projection optimization.

[0009] As a preferred embodiment of the predictive control-based adaptive control system for cabin temperature of an aircraft described in this invention, the pure bottom-level state vector is up-dimensional through an observer dictionary basis function mapping to obtain the high-dimensional true state vector of the current control cycle. Extract the predicted state vector of the high-dimensional prediction model for the current control period from the previous control period; calculate the prediction error vector between the predicted state vector and the high-dimensional true state vector. Solve for the partial derivatives of the prediction error vector with respect to the parameters of the matrix to be updated within the high-dimensional prediction model to generate a sensitivity vector characterizing the sensitivity of the model parameters. Within a set sliding time window, the tensor cross product of the sensitivity vector and its transpose vector at each sampling time is calculated, and all tensor cross products within the sliding time window are summed to construct an empirical Fisher information matrix driven by measurable data. Calculate the sum of all elements on the main diagonal of the empirical Fisher information matrix, and obtain the trace of the empirical Fisher information matrix as the information geometric feature; The matrix parameters to be updated include the thermodynamic damping coefficient in the dissipative semidefinite matrix and the air convection coupling coefficient in the interconnected skew-symmetric matrix.

[0010] As a preferred embodiment of the predictive control-based adaptive control system for cabin temperature of an aircraft described in this invention, wherein: a steady-state dead zone threshold characterizing the absolute value of the inherent noise floor of the cabin airflow is preset. The information geometric features are compared with the steady-state dead zone threshold; when the information geometric features are determined to be greater than the steady-state dead zone threshold, the local update mechanism of the high-dimensional prediction model is triggered; otherwise, the model parameters are completely frozen. When updating the high-dimensional prediction model, the model parameters are only locally updated along the dominant feature direction indicated by the information geometric features; otherwise, the parameters of the high-dimensional prediction model are frozen. The local update mechanism includes calling the singular value decomposition algorithm to decompose the empirical Fisher information matrix, extracting the left singular column vector corresponding to the maximum value in the singular value diagonal matrix, as the dominant feature direction containing the highest signal-to-noise ratio; The initial parameter update step size generated by the underlying recursive least squares algorithm is calculated. The initial parameter update step size is then forcibly projected onto the dominant feature direction for local accumulation through a vector dot product mechanism. At the same time, the update step sizes in other spatial directions orthogonal to the dominant feature direction are forcibly overwritten to zero, thereby completing the directional local evolution of the parameters of the matrix to be updated.

[0011] As a preferred embodiment of the predictive control-based adaptive control system for cabin temperature of an aircraft described in this invention, the discrete state transition equation of the high-dimensional prediction model is forcibly reorganized into a combination of the difference between the interconnected skew-symmetric matrix and the dissipative positive semi-definite matrix. The algebraic property deadlock constraint of the interconnected skew-symmetric matrix is ​​set so that the transpose of the matrix is ​​equal to its own negative value, in order to characterize the non-dissipative heat flow exchange network between adjacent multi-temperature zones inside the cabin without any increase or decrease in total energy; the algebraic property deadlock constraint of the dissipative semi-definite matrix is ​​set so that all eigenvalues ​​are non-negative, in order to characterize the irreversible heat dissipation of the cabin skin to the external extremely cold atmospheric environment. By constructing the port Hamiltonian energy topology equation using the high-dimensional Hamiltonian energy gradient vector, and calling the quadratic programming solver under the absolute passivity constraint of the port Hamiltonian energy topology equation, the optimal control sequence that minimizes the high-dimensional total energy dissipation rate and avoids the internal friction caused by the clash between hot and cold airflows is calculated.

[0012] As a preferred embodiment of the predictive control-based adaptive control system for cabin temperature of an aircraft, the calculation of the trajectory information entropy by analyzing the meteorological data of the aircraft's forward trajectory includes reading the three-dimensional spatial coordinate sequence of waypoints within the planned future flight line of sight and the corresponding atmospheric meteorological environment temperature prediction sequence in real time through the airborne aviation data bus. The probability distribution density features of the atmospheric meteorological environment temperature prediction sequence are extracted and substituted into the Shannon information entropy measurement model to perform logarithmic summation of probability uncertainty. Output a dimensionless entropy scalar of the trajectory information.

[0013] As a preferred embodiment of the predictive control-based adaptive control system for cabin temperature of an aircraft described in this invention, wherein: the tracking error invariant set whose error radius expands inversely with the entropy of the trajectory information is constructed by using the benchmark target temperature set in the cabin as the center line to construct a virtual tracking error envelope that fluctuates up and down. The radial width of the virtual tracking error envelope is set to be equal to the preset physical mechanical valve sensitivity constant divided by the trajectory information entropy, and the error radius of the basic steady state is superimposed, so that the allowable error range of the envelope and the trajectory information entropy form an inversely proportional dynamic expansion relationship. Extract the future projected temperature trajectory and compare the absolute difference between each discrete point and the baseline target temperature. When it is determined that the absolute difference within the entire projected line of sight is less than the currently calculated radial width of the envelope, the temperature fluctuation is determined to be within the legal containment range. The valve action control increment in the optimal control sequence is forcibly overwritten to zero, and the action signal is intercepted to protect the mechanical life of the pneumatic valve. Otherwise, the instruction is executed according to the optimal control sequence.

[0014] The beneficial effects of this invention are as follows: By introducing a state cleaning module, information geometric feature evaluation, and adaptive updating of the high-dimensional prediction model, high precision, rapid response, and adaptive capability of cabin temperature control are achieved. An absolutely reliable cabin state is generated using a high-low frequency fusion method, and combined with orthogonal projection and a quadratic programming correction mechanism, sensor noise is effectively filtered out and unmodeled heat sources are handled. Information geometric features are calculated using sensitivity vectors and empirical Fisher information matrices to achieve localized, directional updates of the high-dimensional prediction model parameters, enabling online compensation for thermodynamic drift caused by cabin material aging and changes in occupant distribution. Simultaneously, the model parameters are mapped to physical matrices (thermodynamic damping coefficient and air convection coupling coefficient) and port Hamiltonian energy constraints are enforced to minimize total cabin energy dissipation and maintain a non-dissipative coupled network, optimizing the control sequence. The dynamic error envelope assisted by track information entropy further ensures that temperature fluctuations remain within a safe range, extends valve mechanical life, and improves cabin comfort and safety. The control process combines physical rationality and adaptive capability. Attached Figure Description

[0015] To more clearly illustrate the technical solutions of the embodiments of the present invention, the drawings used in the following description of the embodiments will be briefly introduced. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0016] Figure 1 This is a block diagram of the adaptive temperature control system for the cabin of an aircraft based on predictive control. Detailed Implementation

[0017] To make the above-mentioned objects, features and advantages of the present invention more apparent and understandable, the specific embodiments of the present invention will be described in detail below with reference to the accompanying drawings.

[0018] Many specific details are set forth in the following description in order to provide a full understanding of the invention. However, the invention may also be practiced in other ways different from those described herein, and those skilled in the art can make similar extensions without departing from the spirit of the invention. Therefore, the invention is not limited to the specific embodiments disclosed below.

[0019] Secondly, the term "one embodiment" or "embodiment" as used herein refers to a specific feature, structure, or characteristic that may be included in at least one implementation of the present invention. The phrase "in one embodiment" appearing in different places in this specification does not necessarily refer to the same embodiment, nor is it a single or selective embodiment that is mutually exclusive with other embodiments.

[0020] Reference Figure 1This is one embodiment of the present invention, which provides an adaptive control system for cabin temperature of an aircraft based on predictive control, including a state cleaning module, a model wake-up module, a prediction module and an execution module.

[0021] The state cleaning module collects temperature data from multiple temperature zones in the cabin, performs implicit thermal state cleaning based on manifold boundary projection, and outputs a clean, noise-filtered bottom-level state vector.

[0022] In this invention, cabin multi-temperature zone data refers to the real-time temperature information collected by temperature sensors deployed in different spatial areas within the aircraft cabin. Because the temperature distribution within the aircraft cabin varies spatially—for example, temperatures near windows, vents, or cabin walls may differ—the cabin is divided into several temperature zones, with temperature data collected independently for each zone, forming a multi-dimensional temperature vector. This multi-temperature zone data reflects the true temperature distribution characteristics within the cabin, providing accurate input for high-dimensional prediction models. It also supports adaptive control and local parameter updates based on multiple temperature zones, enabling fine-tuning and control of the temperature in each area of ​​the cabin.

[0023] The implicit thermal state cleaning based on manifold boundary projection includes parallel execution of a high-frequency state estimation algorithm and a low-frequency physical optimization algorithm: The Extended Kalman Filter (EKF) is used as the high-frequency state estimation algorithm, with cabin temperature sensor data as the observation input. Recursive filtering is employed to calculate the high-frequency real-time estimated state, including implicit cabin thermal loads. The EKF is a recursive estimation algorithm for nonlinear systems. It linearizes the system state prediction model, transforming the nonlinear state equations into a form that can be handled by linear methods. In each control cycle, the EKF uses multi-temperature zone temperature data collected by the cabin temperature sensor as the observation input, combined with the state estimate from the previous control cycle, to recursively calculate the optimal estimate of the current cabin state. The recursive filtering calculation refers to achieving state estimation through a two-step "prediction-update" loop: first, the current state and its covariance matrix are predicted using the system dynamic model; then, the predicted state is updated based on the observed data, correcting deviations to obtain a more accurate state estimate. Through this process, a high-frequency real-time estimated state, including implicit cabin thermal loads, can be calculated. This state reflects measurable temperature data and indirectly characterizes cabin heat changes caused by unmodeled heat sources such as crew activity, solar radiation, and equipment heat dissipation. The role of high-frequency real-time state estimation is that it can capture rapid dynamic changes in cabin temperature on a millisecond timescale, providing accurate basic input for subsequent information geometric feature extraction, high-dimensional prediction model updates and control decisions, ensuring that the adaptive control system can respond to cabin thermal disturbances in a timely manner and maintain cabin temperature stability and comfort.

[0024] It's also important to understand that, firstly, Extended Kalman Filtering (EKF) is a mature nonlinear recursive filtering method. Even with nonlinear system state equations, it can achieve optimal estimation of the system state through local linearization prediction and observation updates, ensuring the stability and convergence of the state estimation. Secondly, in engineering implementation, using multi-temperature zone temperature data from the cabin as observation input, combined with state information from the previous control cycle, EKF can recursively calculate the current state within each control cycle. This eliminates the need for global optimization, resulting in low computational complexity and suitability for real-time operation with limited onboard computing power, thus achieving millisecond-level high-frequency estimation. Thirdly, the high-frequency real-time estimated state not only includes observable temperature information but also indirectly reflects implicit thermal loads within the cabin through state prediction and observation fusion. These implicit loads, such as unmodeled heat sources like crew activity, solar radiation, and equipment heat dissipation, provide accurate instantaneous state information, offering reliable input for subsequent high-dimensional prediction model updates and adaptive control decisions.

[0025] The moving horizon estimation algorithm is synchronously invoked as a low-frequency physical optimization algorithm. Based on temperature data within a historical sliding time window, it solves a nonlinear programming problem constrained by a thermal conduction differential equation. The output solution set forms an envelope in multidimensional space, which serves as the absolutely reliable physical manifold. Moving horizon estimation is an optimization state estimation method based on historical observation data. Its core idea is to use only the most recent temperature data within each control cycle, combined with the system dynamic model and physical constraints, to solve for the optimal state sequence. Specifically, it uses multi-temperature zone temperature data within a historical sliding time window to solve a nonlinear programming problem constrained by a thermal conduction differential equation, ensuring that the predicted state satisfies the thermodynamic laws within the cabin. The multiple feasible solutions obtained form a solution set, distributed in a multidimensional temperature space. The outer boundaries of these solutions form an envelope, i.e., the absolutely reliable physical manifold. This envelope represents the range of possible temperature states within the cabin under physical constraints. When the cabin state output by the high-frequency extended Kalman filter deviates from this manifold, an orthogonal projection mechanism can be triggered to correct the deviation back to the envelope. In this way, low-frequency MHE provides a reliable state boundary based on physical constraints, enabling high-frequency state estimation to respond quickly to changes in cabin temperature while maintaining physical rationality, thus providing a robust foundation for adaptive predictive control of cabin temperature.

[0026] It's worth noting that, unlike high-frequency filtering algorithms (such as extended Kalman filtering) that rely solely on a single current measurement for iteration, the moving horizon estimation used in this system, as a low-frequency physics optimization algorithm, possesses global extrapolation capabilities that include memory features. Specifically, the controller extracts a multi-dimensional temperature observation sequence from the cabin's historical sliding time window (e.g., the past N consecutive sampling periods) from memory and treats it as a whole optimization object to construct a nonlinear programming problem.

[0027] In this optimization process, the system forcibly introduces partial differential equations of heat conduction (such as Fourier's law of heat conduction and the fluid convection heat transfer equation) that characterize the essential thermodynamic laws of the cabin into the underlying code as global equality constraints for optimization. This mechanism is equivalent to putting "physical shackles" on the purely mathematical optimization solver, ensuring that any state trajectory it fits and derives must strictly obey the objective thermodynamic axioms that "heat cannot be transferred instantaneously and temperature cannot change abruptly." After solving the above nonlinear programming problem with strict thermodynamic partial differential equation constraints, due to the basic thermodynamic white noise and bounded measurement interference from sensors in the cabin environment, the optimization solver does not output a single, absolutely precise, isolated coordinate point in the multidimensional state space, but rather a set of permissible states that both conform to the laws of thermodynamics and satisfy the current noise interference probability boundary, i.e., the "solution set."

[0028] The system extracts the outermost boundary shell formed by the solution set in the multidimensional geometric topological space. In terms of physical judgment logic: all multidimensional state coordinates inside the envelope of this manifold are regarded as "thermodynamically valid real physical states"; conversely, any isolated coordinate point outside the envelope of this manifold is unconditionally judged by the system as a false signal that violates objective physical laws (such as sensor value spikes or hardware glitches caused by sudden electromagnetic pulses).

[0029] In each control cycle, the minimum Euclidean geometric distance from the high-frequency real-time estimated state space coordinates to the boundary of the absolutely reliable physical manifold is calculated; when it is determined that the minimum Euclidean geometric distance is greater than the tolerance of the inherent white noise absolute value calibrated by the hardware sensor, the orthogonal projection mechanism is triggered to forcibly pull the deviated state back into the boundary of the manifold.

[0030] Existing methods typically rely on a single high-frequency filtering algorithm (fast but prone to drift and divergence with time integration) or a single statistical algorithm (extremely accurate but extremely time-consuming and unsuitable for airborne operation). This solution employs a "slow-but-fast approach with a safety net": allowing the low-frequency, precise algorithm to construct an insurmountable "physical baseline (manifold)," within which the high-frequency estimation algorithm operates freely; if it deviates from this baseline, orthogonal projection is immediately used to force it back. This completely solves the industry-wide problem of inaccurate and real-time observation of hidden heat sources (such as passengers and sunlight) in complex aviation environments.

[0031] The orthogonal projection mechanism involves calling a quadratic programming solver to perform orthogonal projection optimization. in, This represents the pure bottom-level state vector output after projection onto the manifold boundary; A mathematical optimization operator representing the independent variable that achieves the global minimum; Represents any legal state point on an absolutely reliable physical manifold surface; Represents an absolutely reliable physical manifold; Indicates the high-frequency real-time estimation status; The Euclidean square norm represents the difference between two vectors.

[0032] The model wake-up module receives the pure underlying state vector, extracts information geometric features, performs feature-oriented wake-up of the high-dimensional prediction model based on information geometric evaluation, and realizes the update of the high-dimensional prediction model.

[0033] The pure bottom-level state vector is up-dimensioned through an observer dictionary basis function mapping to obtain the high-dimensional true state vector for the current control cycle. The observer dictionary basis function is forced to include a high-power term representing the physical intensity of cabin thermal radiation, and a product term of intake flow rate and temperature representing the fluid convection heat transfer mechanism. The pure bottom-level state vector is substituted into the observer dictionary basis function, and a high-dimensional true state vector is generated through nonlinear mapping. Essentially, this is a rigorous and continuous algebraic mapping process. First, in any discrete control cycle, the state cleaning module outputs the cleaned low-dimensional pure bottom-level state vector at the current moment in real time, whose mathematical expression is: Indicates the first The system acquires a low-dimensional pure underlying state vector in one discrete control cycle; This represents the current discrete sampling time step scalar. Indicates the first Periodicity, the actual base temperature value of the first temperature zone of the cabin (such as the front cabin) after manifold boundary cleaning; Indicates the first Periodicity refers to the actual base temperature value of the second temperature zone of the cabin (such as the aft cabin) after manifold boundary cleaning; This represents the transpose operator of a matrix or vector. To encompass the complete feature set of the cabin's nonlinear thermodynamic system, the system pre-constructs an observer dictionary basis function in the controller memory. This dictionary basis function not only includes the basic state variables themselves but also has nonlinear higher-order terms and coupling terms characterizing objective thermodynamic laws hard-coded. Its typical mathematical construction is as follows: The observer dictionary basis function mapping operator is essentially a multidimensional column vector function; It represents any combination of low-dimensional state variables as input; The fourth term, representing absolute temperature, is used to accurately characterize the nonlinear thermal radiation effect that conforms to the Stefan-Boltzmann Law. This indicates the intake air mass flow rate of the cabin air conditioning system. This term represents the product of mass flow rate and temperature, used to characterize the nonlinear convective heat transfer effect under fluid intervention.

[0034] Based on the function construction incorporating prior physical knowledge, the system can perform nonlinear algebraic mapping operations, thus achieving the core "dimensionality increase" action. The controller extracts low-dimensional vectors with definite numerical values. Then, it is directly substituted into the aforementioned preset dictionary base function for real-time algebra calculation, and its mathematical expression is: Indicates the first The period, a high-dimensional column vector obtained through mapping formula calculation, is the "high-dimensional true state vector of the current control period" as described in the claims of this invention. This represents the complete high-dimensional feature array expanded by substituting the specific values ​​of the low-dimensional pure underlying state vector of the current period into the dictionary basis function.

[0035] Extract the predicted state vector of the high-dimensional prediction model for the current control period from the previous control period; calculate the prediction error vector between the predicted state vector and the high-dimensional true state vector.

[0036] It should be noted that, in the embodiments of this invention, the "high-dimensional prediction model" described in the claims plays a core role as a "digital twin inference engine" in the entire adaptive predictive control system. This model is not a black-box artificial intelligence neural network lacking physical mechanisms, but rather a white-box linear discrete state-space transition equation constructed in the controller's memory through rigorous algebraic topology. After nonlinear up-dimensional mapping using the observer dictionary, the extremely complex nonlinear thermodynamic dynamics of the cabin (including thermal radiation, fluid coupling, etc.) are rigorously and equivalently collapsed into purely linear evolutionary laws in this high-dimensional topological space. The core mathematical foundation of this model is a first-order linear difference equation: This indicates that the system's future value, calculated by the model, is... A high-dimensional abstract state prediction vector for each discrete sampling time; This represents the current control step sequence scalar on the discrete-time axis, and the corresponding... This indicates the next deduction action step that is immediately following on the timeline; The high-dimensional state transition characteristic matrix is ​​a set of constant coefficients containing the inherent thermodynamic properties of the system, used to characterize the spontaneous evolution and dissipation of high-dimensional thermodynamic energy inside the cabin without external intervention. Indicates the first At each discrete sampling time, the high-dimensional true state vector (or the inferred state vector representing the previous step in multi-step iteration) is input into the model. This represents the high-dimensional control input weight matrix, used to characterize the specific physical driving efficiency and influence weight of the cabin air conditioning physical actuators on the high-dimensional state of the system; Indicates the first At each discrete sampling moment, the system assumes or actually issues a low-dimensional control input command vector (such as the target opening degree or intake air velocity of a pneumatic valve).

[0037] In the original low-dimensional physical space, the transitions of the cabin system state are controlled by extremely complex nonlinear physical functions, such as the fourth-order temperature term representing thermal radiation and the product coupling term representing convective heat transfer. However, this system performs topological dimensionality reduction and dimensionality transpose operations on the nonlinear features at the underlying level through the basis functions of the observer dictionary. Specifically, the system no longer treats the aforementioned higher-order terms and coupling terms as nonlinear derivative functions of the original basic temperature variables, but instead forces them to be mathematically redefined as completely independent new first-order linear coordinate axes in the high-dimensional topological space. Through this variable substitution and dimensional expansion, the original nonlinear power function operation is cleverly and equivalently reduced to a first-order linear combination in the high-dimensional space. Furthermore, based on the Koopman operator theory, when the high-dimensional feature space constructed by the mapping is sufficiently complete, the mutual transition probabilities and dynamic coupling relationships between these newly generated independent high-dimensional variables will strictly converge to constants. During the system's offline initialization or online targeted update phases, the underlying Extended Dynamic Mode Decomposition (EDMD) algorithm has accurately calculated all the aforementioned mapping constant coefficients through extreme value optimization and hard-coded them into the high-dimensional state transition matrix A. Therefore, the physical essence of this high-dimensional prediction model, which exhibits an absolutely linear relationship mathematically, is that the system performs an equivalent algebraic transformation of "trading high-dimensional space for linear time," permanently folding and solidifying the extremely large spatial computational complexity of nonlinear thermodynamic laws into the static memory occupation of a high-dimensional linear matrix. During the real-time derivation phase on the onboard computer, this linear algebraic model, while perfectly adhering to the complex objective physical laws of the cabin, completely eliminates the real-time calculus differentiation process of nonlinear functions, thereby realizing nonlinear dynamic ultra-fast derivation based on pure linear matrix multiplication and addition operations.

[0038] The core mechanism behind this model's ultra-real-time predictive computation lies in the synergy between offline folding of complex nonlinear laws and online ultra-fast algebraic multiplication and addition. During system initialization, the extremely complex partial differential thermodynamic evolution laws are fully calculated using the Extended Dynamic Mode Decomposition (EDMD) data-driven algorithm and "folded" into the elements of the aforementioned constant matrices A and B in a hard-coded form. Therefore, within the real-time control cycle of the onboard computer, when the quadratic programming solver (described later) needs to evaluate a set of hypothetical future control action sequences, the underlying central processing unit (CPU) no longer needs to perform any calculus or Jacobi partial derivative iterations. The controller only needs to extract the current high-dimensional state vector and strictly follow the aforementioned linear equations to execute the most basic matrix multiplication and addition instructions. By multiplying the current state by matrix A and the control instructions by matrix B and summing the results, the high-dimensional state at the next moment can be calculated analytically in one step. The solver, through forward loop calls to this multiplication and addition operation, can complete the coherent deduction of multiple future control steps under extremely demanding onboard computing power conditions.

[0039] Furthermore, the partial derivatives of the prediction error vector with respect to the parameters of the matrix to be updated within the high-dimensional prediction model are calculated to generate a sensitivity vector characterizing the sensitivity of the model parameters. Since the high-dimensional prediction model possesses linear parameterization characteristics, this sensitivity vector is numerically equivalent to the high-dimensional true state vector of the previous control cycle.

[0040] It's important to note that during closed-loop system operation, when the actual physical evolution of the cabin deviates from the high-dimensional prediction model (i.e., prediction errors occur), these errors are often caused by the combined effects of multiple environmental abrupt changes or physical and mechanical aging. To accurately implement "responsibility tracing" within a high-dimensional parameter space of hundreds or thousands of parameters—that is, to identify which specific physical attribute (such as a decrease in the thermal insulation performance of a certain skin layer or a reduction in the convection efficiency of a certain ventilation duct) caused the current prediction distortion—the system must introduce sensitivity vectors for quantitative evaluation. The physical essence of partial derivatives is to explore the absolute ratio of the dramatic change in results (error amplification) caused by a small cause (parameter drift). Because the model uses an algebraic topological structure of high-dimensional linear equations, a mathematically valuable phenomenon occurs when solving for partial derivatives of the prediction error: the latest high-dimensional true state vector acquired at the current moment, and the independent control command vector that was solidified and issued at the previous moment, both degenerate into absolute objective constants in the mathematical perspective of differentiating the parameters to be updated, and their derivative values ​​automatically resolve to zero. Therefore, the extremely complex calculation of prediction error differentiation is elegantly and rigorously simplified to differentiating only the state transition feature matrix itself, supplemented by historical real states as constant weighting coefficients. This highly specialized solution method mathematically avoids the computational disaster of solving complex Jacobian matrices in high-dimensional space; physically, it acts like a high-precision "algebraic microscope," successfully stripping away external data interference caused by real-time control actions and transient environmental fluctuations, and purely extracting the endogenous structural drift characteristics caused by the decay of the cabin's inherent physical architecture. It is precisely by relying on this pure sensitivity guidance that the system, in subsequent parameter update stages, avoids blind trial and error, but rather, like precision guidance, performs localized, directional evolution only along the physical features that cause the most severe errors. The matrix parameters to be updated include the thermodynamic damping coefficient in the dissipative semidefinite matrix, which maps to the physical thermal insulation performance of the aircraft cabin skin. The controller tracks and compensates online for the aging of thermal insulation materials and the drift of heat leakage rate to the outside atmosphere caused by the increase of aircraft service time by adaptively updating the thermodynamic damping coefficient. The matrix parameters to be updated also include the air convection coupling coefficient in the interconnected skew-symmetric matrix, which maps to the physical airflow connectivity between adjacent temperature zones inside the cabin. The controller tracks and compensates online for the sudden changes in non-dissipative heat flow rate caused by changes in cabin occupant distribution or micro-pressure difference fluctuations in the cabin by adaptively updating the air convection coupling coefficient.

[0041] Within a set sliding time window, the tensor cross product of the sensitivity vector and its transpose vector at each sampling time is calculated. All tensor cross products within the sliding time window are then summed to construct an empirical Fisher information matrix driven by measurable data. The sum of all elements along the main diagonal of the empirical Fisher information matrix is ​​calculated, and the trace of the empirical Fisher information matrix is ​​obtained as the information geometric feature.

[0042] Specifically, in a macroscopic physical sense, this empirical Fisher information matrix is ​​equivalent to an "information topology map" that depicts the degree to which cabin operation data stimulates various underlying physical parameters in real time. Each element on the main diagonal of the matrix precisely quantifies the sensitivity (i.e., local information intensity) of the cabin observation data in the corresponding parameter dimension. To aggregate the highly dispersed local sensitivity information in high-dimensional space into a macroscopic scalar that allows the underlying processor to perform absolute threshold decisions, the system performs a global summation extraction operation on the main diagonal elements of this feature matrix (i.e., solving for the trace of the matrix). This scalar sum intuitively quantifies the "total effective physical information" contained in the current observation data.

[0043] In the adaptive triggering logic, the system relies on the aforementioned information scalar to intelligently distinguish between environmental noise and actual physical degradation: if the total information content is below a preset threshold, the system determines that the current error mainly originates from high-frequency random noise, thus forcibly suppressing blind model updates; conversely, once the total information content exceeds the threshold, the system immediately confirms that the cabin physical environment has undergone substantial degradation. Under this certainty, the system will formally activate the parameter evolution mechanism and strictly follow the "maximum local information intensity" direction guided by the empirical Fisher information matrix, accurately locking and finding the core physical parameters that have become distorted in the complex high-dimensional parameter space, and then initiating highly targeted local parameter repair. This parameter search mechanism based on information geometry completely avoids the waste of computational power and the risk of model divergence caused by blind global parameter updates in traditional adaptive control.

[0044] A preset steady-state dead zone threshold is established to characterize the absolute value of the inherent noise floor of the cabin airflow. The information geometric features are compared with the steady-state dead zone threshold; if the information geometric features are determined to be greater than the steady-state dead zone threshold, the local update mechanism of the high-dimensional prediction model is triggered; otherwise, the model parameters are completely frozen.

[0045] When updating the high-dimensional prediction model, the model parameters are only locally updated along the dominant feature direction indicated by the information geometric features; otherwise, the parameters of the high-dimensional prediction model are frozen.

[0046] It's important to note that in complex adaptive control engineering, high-frequency environmental background noise and low-frequency physical component degradation are often deeply coupled in the observed data. The fundamental design purpose of setting a steady-state dead zone threshold is to construct an extremely strict physical barrier of "information signal-to-noise ratio" for the prediction model. If the control system provides real-time feedback for any tiny numerical prediction residual, it is highly likely that the model parameters will fall into disordered algebraic oscillations caused by high-frequency random sampling noise from sensors. Once this high-frequency jitter at the parameter level is output, it will inevitably transmit high-frequency noise to downstream physical actuators, directly inducing severe surge and accelerated fatigue wear of mechanical components such as pneumatic valves. The introduction of this dead zone threshold forces the system to strip away invalid random environmental disturbances at the logical level, allowing the model to break its silent state only when it is certain that it has captured an effective excitation sufficient to cross the noise floor threshold, caused by substantial physical material decay or structural distortion. This asymmetric triggering mechanism fundamentally prevents the ineffective idling of onboard computing power and achieves a perfect balance between the absolute numerical stability of the model and its self-diagnostic sensitivity.

[0047] On the other hand, the mandatory requirement for the model to perform local directional updates only along the dominant feature direction and freeze the remaining dimensions after wake-up aims to completely avoid the "curse of dimensionality" and "algebraic overfitting" traps in the ultra-high-dimensional parameter optimization space. In an extremely large multidimensional state transition matrix, if a traditional global indiscriminate optimization strategy is used to forcibly smooth out the current prediction error, it will inevitably destroy the existing algebraic manifold of other physical parameters that are still in a healthy steady state in the model. This will cause the model to lose its global physical generalization ability in order to accommodate local transient errors, and may even lead to the complete divergence of the inference algorithm. Based on the update restriction of the dominant feature direction, a "surgical" targeted repair rule with strong physical boundary constraints is essentially established: the system is forced to perform fixed-point compensation only on the core parameter coordinate axis that has been confirmed by objective data to have the most serious deviation, while the remaining massive healthy parameters are absolutely physically locked. This evolutionary strategy, which combines local reconstruction with global freezing, safeguards the global physical consistency of the high-dimensional prediction model as a digital twin to the greatest extent, ensuring that the system still possesses unbreakable long-term robustness under extremely harsh time-varying conditions.

[0048] The local update mechanism includes calling the singular value decomposition algorithm to decompose the empirical Fisher information matrix, extracting the left singular column vector corresponding to the largest value in the singular value diagonal matrix, as the dominant feature direction containing the highest signal-to-noise ratio.

[0049] The initial parameter update step size generated by the underlying recursive least squares algorithm is calculated. The initial parameter update step size is then forcibly projected onto the dominant feature direction for local accumulation through a vector dot product mechanism. At the same time, the update step sizes in other spatial directions orthogonal to the dominant feature direction are forcibly overwritten to zero, thereby completing the directional local evolution of the parameters of the matrix to be updated.

[0050] It should be noted that when the system determines that the trace of the empirical Fisher information matrix has exceeded the steady-state dead zone threshold and is certain that a model update needs to be triggered, the controller first calls the singular value decomposition (SVD) algorithm in high-speed memory to perform dimensionality reduction and decoupling decomposition of the orthogonal and diagonal matrices on the currently generated empirical Fisher information matrix. Its core mathematical expression is: This represents the empirical Fisher information matrix constructed within the current discrete control cycle; This represents the left singular orthogonal matrix generated by the decomposition, where each column vector represents a mutually orthogonal independent feature basis direction in the parameter space of the high-dimensional model. It represents a singular value diagonal matrix, in which the elements (i.e., singular values) on the main diagonal are arranged in descending order, accurately representing the amount of absolute observational data information contained in each corresponding feature basis direction; This represents the transpose of a right singular orthogonal matrix.

[0051] After obtaining the above algebraic decomposition results, the system directly addresses and locks the diagonal matrix. The maximum singular value in the matrix. This maximum value, in information geometry, absolutely maps the evolution dimension with the highest signal-to-noise ratio (SNR) and the most significant physical degradation characteristics in the current cabin thermodynamic evolution data. The controller then operates on the left singular matrix. In this process, the single column vector corresponding to the maximum singular value is precisely extracted and defined and cached as the "dominant feature direction vector" (denoted as ). ).

[0052] After establishing an absolutely reliable physical degradation direction, the system initiates a parameter step size optimization and spatial orthogonal filtering mechanism. First, the underlying layer calls the classic Recursive Least Squares (RLS) algorithm to calculate a multi-dimensional preliminary parameter update step size vector (denoted as ) based on the currently acquired prediction error vector. The initial step size is an omnidirectional vector diverging in the full parameter space, inevitably deeply coupled with the degradation gradient of the real physical environment and high-frequency sensor random noise. To transform this "blind and noisy" initial step size into a "precise and pure" directional repair instruction, the controller forcibly introduces a vector dot product and spatial projection mechanism, the core algebraic mapping formula of which is: This represents the absolutely safe and noise-free directional local update step size vector that the system finally determines after spatial orthogonal filtering; This represents the initial, unfiltered multi-directional parameter update step size vector generated by the underlying RLS algorithm based on the prediction error; The dot product operator represents the inner product of vectors, used to extract the length of the coincident projection of two vectors in spatial geometry; This represents the dominant feature direction vector of unit length extracted by singular value decomposition.

[0053] From this algebraic mapping formula, we can see that the inner product operation... As an extremely demanding scalar extractor, it extracts and retains only the effective physical update magnitude scalar that is perfectly parallel to the dominant feature direction from the chaotic multidimensional initial step size. The system then reassigns this purified magnitude scalar... The geometric direction is determined, thus generating the final update vector. Under extremely rigorous algebraic topological constraints, this projection operation forcibly overwrites and annihilates all other spatial multidimensional components in the initial step size that are orthogonal to the dominant feature direction (i.e., deviating from the core physical degradation direction). This extreme algebraic dimensionality reduction mechanism completely blocks the path of environmental noise penetrating into the parameters of the healthy model, allowing the high-dimensional prediction model to perform local adaptive compensation only for the single physical dimension that has been rigorously proven to have undergone substantial changes. This mechanism perfectly freezes and defends the existing physical manifold of the remaining healthy dimensions in the model, fundamentally eliminating the fatal "catastrophic forgetting" and model divergence collapse phenomena caused by omnidirectional blind updates in traditional adaptive control.

[0054] It's important to note that in ultra-high-dimensional physical systems like aircraft cabins, the dimensionality of the states observable in real-time by sensors is often far lower than the dimensionality of the hidden feature parameters within the predictive model. Attempting to use observational errors of a finite dimension to infer global parameter updates is a classic example of an "underdetermined problem" on an algebraic level. If the underlying optimization algorithm is allowed to diverge freely across the entire parameter space, the solver, in order to forcibly smooth out immediate transient errors as quickly as possible, will inevitably adopt a "robbing Peter to pay Paul" mathematical compromise strategy, arbitrarily altering previously perfectly healthy physical parameters. This rampant, blind updating will instantly destroy the correct thermodynamic algebraic manifold accumulated in the early stages of the model, causing the model to fit the error at the current moment but permanently lose its ability to generalize to future variable conditions (i.e., triggering catastrophic amnesia).

[0055] The mechanism of forcibly implementing spatial projection and overwriting orthogonal directions to zero constructs an absolutely robust "noise-resistant algebraic firewall." Physical engineering experience shows that secondary, non-dominant parametric spatial characteristic directions often do not contain substantial component decay information, but are instead filled with high-frequency sensor measurement noise and random perturbations from cabin airflow. Forcibly concentrating the update step size to a single dominant characteristic direction and implementing complete algebraic locking in the remaining orthogonal directions essentially employs information geometry's dimensionality reduction technique to physically separate real mechanical aging signals from illusory random environmental clutter in spatial topology.

[0056] It's important to understand that traditional adaptive algorithms suffer from a fatal flaw: "covariance explosion." When the aircraft is cruising smoothly at a constant temperature, the algorithm will frantically learn minute noise from the sensors as if it were a physical law, leading to model collapse. This embodiment introduces "Information Value Assessment (FIM)" from the field of communications. It goes beyond simply adding a dead zone; it utilizes Singular Value Decomposition (SVD) to achieve "surgical targeted learning": updating model parameters only on the dimensions with the strongest signals and the most realistic physical stimuli, while freezing the dimensions overwhelmed by noise. This mathematically endows the algorithm with extremely strong anti-divergence immunity.

[0057] The prediction module elevates the physical quantities containing the pure underlying state vector to the updated high-dimensional prediction model, forces constraints on the underlying rules to a port Hamiltonian energy topology that conforms to the passive characteristics, and solves for the optimal control sequence and the synchronously generated future temperature trajectory with the goal of minimizing the total energy dissipation rate.

[0058] Specifically, after the system completes the internal parameter repair of the high-dimensional state transition matrix through a directed local evolution mechanism, the high-dimensional prediction model has the ability to extrapolate the latest physical degradation state of the cabin. At this point, the system extracts the clean bottom-level state vector (i.e., the current real physical environment parameters after noise cleansing) at the latest discrete sampling moment and directly substitutes it into the observer dictionary basis function to perform a nonlinear algebraic mapping, thereby obtaining the high-dimensional real state vector at the current moment. Subsequently, the system forcibly injects this highly timely high-dimensional real state vector as an initial condition into the linear state-space equation that has just completed parameter updates. This step, in terms of algebraic topology, ensures that when the prediction module performs future-oriented multi-step closed-loop optimization extrapolation in the next stage, its extrapolation starting point is absolutely anchored to the current real physical coordinates of the cabin, and the extrapolation path absolutely obeys the latest thermodynamic evolution mechanism of the cabin, thereby completely eliminating the risk of prediction divergence caused by the superposition of model parameter lag and initial state distortion.

[0059] In a specific embodiment of the present invention, after each call to the high-dimensional prediction model to perform forward inference, it generates and outputs only two types of core data results in the controller memory to support the global closed-loop control of the system. First, the model directly outputs a sequence of future high-dimensional inference states on a mathematical level. This sequence consists of multiple consecutive high-dimensional abstract state prediction vectors derived from the model's future inference, implicitly containing the system's future high-order thermodynamic kinetic and potential energy evolution characteristics. The system directly injects this high-dimensional state sequence into the port Hamiltonian energy topology equation to calculate the total energy dissipation rate cost of the system within the entire prediction line of sight with extreme precision. This provides an absolute physical boundary judgment basis for the quadratic programming solver in the prediction module to evaluate the quality of control actions and find optimal extreme values. Secondly, the system synchronously calls the dimensionality reduction extraction operator to strip away the high-order terms and nonlinear coupling terms in the aforementioned high-dimensional state sequence, extracting only the basic temperature physical dimension, thereby generating a set of future temperature trajectory sequences with intuitive physical meaning. This temperature trajectory sequence is then completely transmitted to the execution module for comparison with the spatial absolute value of the elastic tracking error envelope dynamically controlled by the trajectory information entropy at each discrete time point, thus serving as the sole feedforward data source for the system to determine whether to intercept the current control command and trigger the pneumatic valve mechanical anti-surge protection action.

[0060] It should be noted that the "mandatory constraints on the underlying rules" essentially mean that the system rejects any black-box algebraic matrix that is purely generated by data fitting but violates the laws of macroscopic thermodynamics. When using data-driven algorithms to solve the state transition feature matrix of a high-dimensional prediction model, the system forcibly injects the topological algebraic structure constraints of the Port-Hamiltonian system into the underlying extremum optimization function. Specifically, the originally freely divergent high-dimensional state transition matrix is ​​forcibly deadlocked and decomposed into a combination of two sub-matrices with absolute physical meaning: one is an interconnected skew-symmetric matrix (whose transpose is equal to its negative matrix, i.e., ...). The other is a dissipative positive semi-definite matrix (whose eigenvalues ​​are always greater than or equal to zero, i.e., ...). Under this strict algebraic structural constraint, interconnected skew-symmetric matrices... It is limited to characterizing only the lossless routing and exchange of high-dimensional thermodynamic energy within the cabin between different subspaces or physical regions (such as the internal conversion of radiative and convective energy), and the generation of energy out of thin air is strictly prohibited; while the dissipative semidefinite matrix It is strictly defined as the irreversible thermodynamic dissipation loss of system energy to the external cold source environment or due to mechanical friction (i.e., the entropy increase process). This rule constraint, which forcibly "hard-codes" the first law of thermodynamics (energy conservation) and the second law of thermodynamics (direction of energy dissipation) of the objective physical world into the matrix structure in the form of algebraic topology, ensures that the model is fundamentally deprived of the algebraic degrees of freedom to generate "fake energy" or "non-physical oscillations" in mathematical deduction.

[0061] The aforementioned "passivity" is a highly valuable cybernetics attribute that inevitably manifests in the macroscopic dynamic evolution of the system under the aforementioned underlying mandatory constraints. In control theory, "passivity" means that the total internal energy stored in a physical system during any given time period can never exceed the sum of the total work done by external control mechanisms (such as the hot or cold air injected by air conditioning valves) and the initial energy stored. Mapped to the cabin thermodynamic high-dimensional prediction model of this invention, this characteristic indicates that after stripping away the forced cooling or heating input from the external air conditioning system, the cabin system itself is an absolute energy consumer. Any transient thermodynamic disturbance or high-dimensional energy fluctuation within it will irreversibly decay naturally along the direction of energy gradient descent under the dominance of the dissipation matrix R, and will eventually converge strictly to a minimum steady state in equilibrium with the external environment. For airborne closed-loop control systems, endowing the prediction model with this rigorous passivity characteristic has decisive engineering significance. In terms of Lyapunov stability at the most fundamental level, it provides an absolutely convergent convex space for the quadratic programming solver: no matter what extreme external airflow disturbances or unpredictable sensor value jumps the cabin faces, the future state projection trajectory calculated by the solver based on this passive model will always point to the safe physical funnel of energy dissipation and state convergence, thereby fundamentally eliminating the global collapse or high-frequency mechanical surge caused by the algebraic divergence of the model in the control commands.

[0062] Furthermore, the discrete state transition equations of the high-dimensional prediction model are forcibly reorganized into a combination of the difference between the interconnected skew-symmetric matrix and the dissipative positive semi-definite matrix. The algebraic property deadlock constraint of the interconnected skew-symmetric matrix is ​​set so that the transpose of the matrix is ​​equal to its own negative value, to characterize the non-dissipative heat flow exchange network between adjacent multi-temperature zones inside the cabin without any increase or decrease in total energy; the algebraic property deadlock constraint of the dissipative positive semi-definite matrix is ​​set so that all eigenvalues ​​are non-negative, to characterize the irreversible heat dissipation from the cabin skin to the external extremely cold atmospheric environment.

[0063] By constructing the port Hamiltonian energy topology equation using the high-dimensional Hamiltonian energy gradient vector, and calling the quadratic programming solver under the absolute passivity constraint of the port Hamiltonian energy topology equation, the optimal control sequence that minimizes the high-dimensional total energy dissipation rate and avoids the internal friction caused by the clash between hot and cold airflows is calculated.

[0064] It should be noted that, in the embodiments of this invention, to ensure that the high-dimensional prediction model possesses strict mathematical convexity and extremely fast solution capability, the system's underlying layer configures the high-dimensional Hamiltonian energy function of the cabin as a quadratic form of the state vector. Therefore, the high-dimensional Hamiltonian energy gradient vector, in terms of mathematical operations, is directly obtained by multiplying the positive definite heat capacity diagonal matrix pre-stored by the controller with the current high-dimensional true state vector through linear matrix multiplication. This process eliminates the need for real-time complex calculus and partial derivative calculations, thereby ensuring the rigor of the physical mechanism while completely avoiding additional computational consumption.

[0065] Specifically, in control engineering and thermodynamics, the total energy of a system (Hamiltonian energy) The Hamiltonian energy function is never an abstract concept, but rather something that can be calculated from state variables (temperature). In high-dimensional linear spaces, to ensure perfect mathematical convexity and extremely fast solution capabilities, the high-dimensional Hamiltonian energy function is defined with extreme restraint as a standard quadratic form: It is the high-dimensional real state vector that we have obtained at the current moment (such as the current temperature). It is a positive definite diagonal matrix pre-written into the onboard computer; it represents the "high-dimensional thermal energy storage bottle" (that is, the inherent physical constants such as the mass of air in the container and the specific heat capacity of air); these constants are determined when the aircraft leaves the factory and do not require adaptive updates.

[0066] "Energy gradient vector" The gradient is the gradient of the state variable. Find the partial derivatives: According to the basic rules of matrix calculus, the quadratic form right The derivative of is equal to The extremely complex calculus was reduced to pure algebraic multiplication: It's just a "temperature value," but thermodynamics can't be based solely on temperature. Multiplying the temperature by the heat capacity (…) This transforms it into a true "thermal potential Laukut". The result of this multiplication is the real driving force (gradient) that propels the flow of heat.

[0067] After rapidly acquiring the energy gradient vector, the system compares it with the interconnected skew-symmetric matrix that is forced into deadlock. and dissipative positive semidefinite matrix Deep algebraic coupling is used to construct the system's core derivation engine—the port Hamiltonian energy topological equation. Its core mathematical expression in the discrete-time domain is: and These are updated parameters. That is the result.

[0068] The result of the Hamiltonian energy topological equation at the port is defined as the predicted state vector of the system in the next discrete period.

[0069] Furthermore, the Quadratic Programming (QP) solver is a high-performance numerical optimization algorithm integrated into the underlying firmware of the flight control computer. The solver's core mission is to find a set of execution commands with the highest global energy efficiency ratio within each discrete control cycle, considering the thermodynamic evolution of the multi-temperature cabin and satisfying strict physical conservation laws.

[0070] The solver at every time step The core computational task is to solve an extremum optimization problem with a quadratic objective function and linear constraints. The standard mathematical expression for the objective cost function is defined as follows: Represents the total energy dissipation and control cost to be minimized as a scalar; and These represent the system's predicted line-of-sight and control line-of-sight, respectively, and are used to define the time depth of the algorithm's "preview" of the future. This represents the future Hamiltonian derived from the port Hamiltonian equation. The high-dimensional state prediction vector for the step; The preset ideal reference target state vector for the cabin; This is a diagonal matrix of state deviation weights, the magnitude of which determines the system's weight in terms of "temperature control accuracy". For the future number to be solved The virtual control input vector of the step (such as the preset opening degree of the pneumatic valve in each temperature zone); The difference between the two is the control input from the previous moment, and represents the range of motion of the actuator. The incremental weight matrix is ​​used to control the intensity of the valve's movement, thereby achieving the engineering goal of suppressing mechanical fatigue wear and reducing transient energy consumption.

[0071] When performing the above optimization calculations, the QP solver does not search in an unrestricted algebraic space, but is subject to hard constraints of equality formed by the port Hamiltonian energy topological equations.

[0072] At the same time, the solver is also subject to inequalities imposed by the physical limits of the actuator: in, and These correspond to the physical boundaries of the pneumatic valve being fully closed (0%) and fully open (100%), respectively.

[0073] Furthermore, the QP solver employs interior-point or effective-set methods, iterating within a geometric space that satisfies absolute passivity. This is due to the dissipation matrix... The positive semi-definite property of the eigenvalues ​​(i.e., all their eigenvalues ​​are always non-negative) defines a unidirectional downward convergent "physical funnel" on the algebraic surface for the optimization process, forcing all output control command sequences to guide the system energy towards the global minimum. This constraint mechanism plays a crucial "energy arbitration" role in physics: by comparing the total energy dissipation rate generated by valve linkages in different temperature zones in real time, the solver can accurately identify and actively eliminate ineffective control combinations that would cause mutual conflict and internal friction between the hot and cold airflows in the front and rear cabins. For example, when it is detected that a certain control sequence would cause two adjacent temperature zones to perform drastic cooling and drastic heating respectively, resulting in energy cancellation in the middle of the cabin, this action would trigger... The energy term increases dramatically, and the QP solver will automatically reject the scheme, thus forcing the system to find the optimal cooperative action with the best energy efficiency ratio.

[0074] Finally, the optimal control sequence is locked by the QP solver. This is the only optimal solution after balancing all the aforementioned physical constraints and accuracy requirements. The system only extracts the first element of the sequence at the output. This is then converted into specific actuator pulse signals and sent out. Simultaneously, the system substitutes this optimal sequence back into the topological equations to perform a complete forward simulation, thereby synchronously generating the future temperature trajectory. The generation process of this temperature trajectory can be expressed as follows: ,in This is the set of future projected states driven by the optimal instruction. The preset linear dimensionality reduction extraction operator (i.e., temperature extraction matrix) is used to accurately extract the temperature evolution curve with intuitive physical meaning from the high-dimensional abstract space, thereby providing feedforward safety verification support for the elastic envelope comparison in the execution module, ensuring that the physical valve will never produce unexpected surge fluctuations.

[0075] It's worth noting that directly solving nonlinear predictive control leads to onboard computing timeouts, while purely data-driven dimensionality reduction / upgrading models lack understanding of physics (easily calculating the erratic canceling action of extreme cooling and heating in adjacent temperature zones). This improvement involves physically restraining the mathematical black box through deadlock. The oblique symmetry (internal interchange does not consume energy) and The semi-positive definiteness of the system (which can only dissipate heat and cannot absorb heat) forces the first and second laws of thermodynamics to be written as mathematical axioms that the algorithm must follow. This not only reduces nonlinear deadlock to extremely fast linear quadratic programming (QP), but also ensures that the output control law will never violate common sense, achieving ultimate computational power liberation and physical security.

[0076] The execution module parses meteorological data of the aircraft's forward trajectory to calculate the trajectory information entropy, and constructs a tracking error invariant set whose error radius expands dynamically inversely with the trajectory information entropy. When the future predicted temperature trajectory is completely within the invariant set, the control command is intercepted and the physical actuator is locked; otherwise, the optimal control sequence is converted into a valve action command and issued.

[0077] The controller reads in real time the three-dimensional spatial coordinate sequence of waypoints within the planned future flight line of sight from the flight management system via the onboard aviation data bus, as well as the corresponding atmospheric ambient temperature prediction sequence. It extracts the probability distribution density function p(T) from the atmospheric ambient temperature prediction sequence. Substituting the probability distribution density function p(T) into the Shannon information entropy model formula, it performs a time summation operation of probability uncertainty, calculates and outputs the dimensionless trajectory information entropy value. Specifically, the mathematical representation is as follows: Where T represents the dynamic atmospheric temperature variable ahead, and log represents the logarithmic operator; this is the entropy index of the flight path. The numerical characteristics are as follows: it tends to be at a minimum when the aircraft is cruising at a constant altitude and speed, and increases when the aircraft is in convective weather or undergoes violent maneuvers.

[0078] Furthermore, a virtual tracking error envelope with fluctuating vertically is constructed using the cabin's set baseline temperature as the centerline. The radial width of the virtual tracking error envelope is set to be equal to a preset physical mechanical valve sensitivity constant divided by the trajectory information entropy, and the error radius of the basic steady state is superimposed, so that the allowable error range of the envelope and the trajectory information entropy form an inversely proportional dynamic expansion relationship.

[0079] Track information entropy (denoted as This is used to quantify the degree of disturbance in the current operating condition in real time. This is achieved by using the radial width of the virtual tracking error envelope (denoted as...). The configuration is such that the dynamic expansion relationship is inversely proportional to the entropy of the track information, i.e., it follows the formula: in, This is the preset physical and mechanical valve sensitivity constant. The basic steady-state error radius.

[0080] Extract the future projected temperature trajectory and compare the absolute difference between each discrete point and the baseline target temperature. When it is determined that the absolute difference within the entire projected line of sight is less than the currently calculated radial width of the envelope, the temperature fluctuation is determined to be within the legal containment range. The valve action control increment in the optimal control sequence is forcibly overwritten to zero, and the action signal is intercepted to protect the mechanical life of the pneumatic valve. Otherwise, the instruction is executed according to the optimal control sequence.

[0081] It's worth noting that a serious side effect of using high-precision algorithms is "oversensitivity." Even minute temperature disturbances can cause mechanical valves to vibrate at high frequencies and quickly fail. Directly modifying the objective function to penalize valve action would break the algorithm's convexity, making it unsolvable. This solution, however, utilizes an inverse proportional function. When the aircraft's cruise environment is extremely stable (track entropy is minimal), the algorithm's tolerance radius automatically and rapidly expands. Because the "pipe" thickens, minute temperature fluctuations are completely contained, and the system directly determines them as "legal, no action required." Thus, it ingeniously relaxes mathematical tolerance to indirectly achieve ultimate protection for the physical mechanical lifespan.

[0082] In summary, this invention achieves high-precision, adaptive, and physically constrained predictive control of cabin temperature by: a state cleaning module performing implicit thermal state cleaning on multi-temperature zone temperature data of the cabin through manifold boundary projection, outputting a clean, noise-filtered bottom-level state vector; a model wake-up module extracting information geometric features and performing sensitivity-based local directional adaptive updates on the high-dimensional prediction model, enabling online compensation for cabin aging, changes in crew distribution, and non-uniform thermal flow disturbances; a prediction module inputting the high-dimensional state into the port Hamiltonian energy topology equation, using a quadratic programming solver to calculate the optimal control sequence under absolute passivity constraints, while avoiding internal friction from opposing hot and cold airflows; and an execution module combining trajectory information entropy to construct a dynamically expanding tracking error envelope, making real-time judgments on future temperature trajectories and intercepting or issuing valve action commands. This invention, through high- and low-frequency state fusion, high-dimensional model parameter sensitivity updates, information geometric feature evaluation, and port Hamiltonian optimization control, achieves rapid response, physical rationality, and adaptive capability of the cabin temperature control system, effectively improving cabin comfort and control accuracy, while extending the service life of valves and actuators.

[0083] It should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention and are not intended to limit it. Although the present invention has been described in detail with reference to preferred embodiments, those skilled in the art should understand that modifications or equivalent substitutions can be made to the technical solutions of the present invention without departing from the spirit and scope of the technical solutions of the present invention, and all such modifications or substitutions should be covered within the scope of the claims of the present invention.

Claims

1. An adaptive control system for cabin temperature in an aircraft based on predictive control, characterized in that: This includes a state cleaning module that collects temperature data from multiple temperature zones in the cabin, performs implicit thermal state cleaning based on manifold boundary projection, and outputs a clean, noise-filtered bottom-level state vector. The model wake-up module receives the pure underlying state vector, extracts information geometric features, performs feature-oriented wake-up of the high-dimensional prediction model based on information geometric evaluation, and realizes the update of the high-dimensional prediction model; The prediction module elevates the physical quantities containing the pure underlying state vector to the updated high-dimensional prediction model, forces constraints on the underlying rules to a port Hamiltonian energy topology that conforms to the passive characteristics, and solves for the optimal control sequence and the synchronously generated future temperature trajectory with the goal of minimizing the total energy dissipation rate. The execution module parses meteorological data of the aircraft's forward trajectory to calculate the trajectory information entropy, and constructs a tracking error invariant set whose error radius expands dynamically inversely with the trajectory information entropy. When the future predicted temperature trajectory is completely within the invariant set, the control command is intercepted and the physical actuator is locked; otherwise, the optimal control sequence is converted into a valve action command and issued.

2. The adaptive control system for cabin temperature based on predictive control as described in claim 1, characterized in that: The implicit thermal state cleaning based on manifold boundary projection includes parallel operation of a high-frequency state estimation algorithm and a low-frequency physical optimization algorithm. Extended Kalman filter is used as the high-frequency state estimation algorithm. With cabin temperature sensor data as the observation input, the high-frequency real-time estimated state including the implicit thermal load of the cabin is calculated through recursive filtering. The synchronous call to the moving horizon estimation is used as a low-frequency physics optimization algorithm. Based on the temperature data within the historical sliding time window, it solves the nonlinear programming problem constrained by the heat conduction differential equation and outputs the envelope surface formed by the solution set in the multidimensional space as an absolutely reliable physical manifold surface. In each control cycle, the minimum Euclidean geometric distance from the high-frequency real-time estimated state space coordinates to the boundary of the absolutely reliable physical manifold is calculated; when it is determined that the minimum Euclidean geometric distance is greater than the tolerance of the inherent white noise absolute value calibrated by the hardware sensor, the orthogonal projection mechanism is triggered to forcibly pull the deviated state back into the boundary of the manifold.

3. The adaptive control system for cabin temperature based on predictive control as described in claim 2, characterized in that: The orthogonal projection mechanism involves calling a quadratic programming solver to perform orthogonal projection optimization.

4. The adaptive control system for cabin temperature based on predictive control as described in claim 3, characterized in that: The pure underlying state vector is increased in dimension by mapping the observer dictionary basis function to obtain the high-dimensional true state vector of the current control cycle. Extract the predicted state vector of the high-dimensional prediction model for the current control period from the previous control period; calculate the prediction error vector between the predicted state vector and the high-dimensional true state vector. Solve for the partial derivatives of the prediction error vector with respect to the parameters of the matrix to be updated within the high-dimensional prediction model to generate a sensitivity vector characterizing the sensitivity of the model parameters. Within a set sliding time window, the tensor cross product of the sensitivity vector and its transpose vector at each sampling time is calculated, and all tensor cross products within the sliding time window are summed to construct an empirical Fisher information matrix driven by measurable data. The sum of all elements on the main diagonal of the empirical Fisher information matrix is ​​calculated, and the trace of the empirical Fisher information matrix is ​​obtained as the geometric feature of the information.

5. The adaptive control system for cabin temperature based on predictive control as described in claim 4, characterized in that: The matrix parameters to be updated include the thermodynamic damping coefficient in the dissipative semidefinite matrix and the air convection coupling coefficient in the interconnected skew-symmetric matrix.

6. The adaptive control system for cabin temperature based on predictive control as described in claim 5, characterized in that: Preset steady-state dead zone threshold that characterizes the absolute value of the inherent noise floor of cabin airflow; The information geometric features are compared with the steady-state dead zone threshold; when the information geometric features are determined to be greater than the steady-state dead zone threshold, the local update mechanism of the high-dimensional prediction model is triggered; otherwise, the model parameters are completely frozen. When updating the high-dimensional prediction model, the model parameters are only locally updated along the dominant feature direction indicated by the information geometric features; otherwise, the parameters of the high-dimensional prediction model are frozen.

7. The adaptive control system for cabin temperature based on predictive control as described in claim 6, characterized in that: The local update mechanism includes calling the singular value decomposition algorithm to decompose the empirical Fisher information matrix, extracting the left singular column vector corresponding to the maximum value in the singular value diagonal matrix, as the dominant feature direction containing the highest signal-to-noise ratio; The initial parameter update step size generated by the underlying recursive least squares algorithm is calculated. The initial parameter update step size is then forcibly projected onto the dominant feature direction for local accumulation through a vector dot product mechanism. At the same time, the update step sizes in other spatial directions orthogonal to the dominant feature direction are forcibly overwritten to zero, thereby completing the directional local evolution of the parameters of the matrix to be updated.

8. The adaptive control system for cabin temperature based on predictive control as described in claim 7, characterized in that: The discrete state transition equation of the high-dimensional prediction model is forcibly reorganized into a combination of the difference between the interconnected skew-symmetric matrix and the dissipative positive semi-definite matrix; The algebraic property deadlock constraint of the interconnected skew-symmetric matrix is ​​set so that the transpose of the matrix is ​​equal to its own negative value, in order to characterize the non-dissipative heat flow exchange network between adjacent multi-temperature zones inside the cabin without any increase or decrease in total energy; the algebraic property deadlock constraint of the dissipative semi-definite matrix is ​​set so that all eigenvalues ​​are non-negative, in order to characterize the irreversible heat dissipation of the cabin skin to the external extremely cold atmospheric environment. By constructing the port Hamiltonian energy topology equation using the high-dimensional Hamiltonian energy gradient vector, and calling the quadratic programming solver under the absolute passivity constraint of the port Hamiltonian energy topology equation, the optimal control sequence that minimizes the high-dimensional total energy dissipation rate and avoids the internal friction caused by the clash between hot and cold airflows is calculated.

9. The adaptive control system for cabin temperature based on predictive control as described in claim 8, characterized in that: The calculation of trajectory information entropy by analyzing the meteorological data of the aircraft's forward trajectory includes reading the three-dimensional spatial coordinate sequence of waypoints within the planned future flight line of sight and the corresponding atmospheric meteorological environment temperature prediction sequence in real time through the airborne aviation data bus. The probability distribution density features of the atmospheric meteorological environment temperature prediction sequence are extracted and substituted into the Shannon information entropy measurement model to perform logarithmic summation of probability uncertainty. Output a dimensionless entropy scalar of the trajectory information.

10. The adaptive control system for cabin temperature based on predictive control as described in claim 9, characterized in that: The invariant tracking error set, whose error radius expands inversely to the entropy of the track information, is constructed by using the reference target temperature set in the cabin as the center line to construct a virtual tracking error envelope that fluctuates up and down. The radial width of the virtual tracking error envelope is set to be equal to the preset physical mechanical valve sensitivity constant divided by the trajectory information entropy, and the error radius of the basic steady state is superimposed, so that the allowable error range of the envelope and the trajectory information entropy form an inversely proportional dynamic expansion relationship. Extract the predicted future temperature trajectory and compare the absolute difference between each discrete point and the baseline target temperature. When it is determined that the absolute difference within the entire simulation range is less than the currently calculated radial width of the envelope, the temperature fluctuation is determined to be within the legal containment range. The valve action control increment in the optimal control sequence is forcibly overwritten to zero, and the action signal is intercepted to protect the mechanical life of the pneumatic valve. Otherwise, the instruction is executed according to the optimal control sequence.

Images