Helicopter power distribution dynamic priority scheduling method based on physical information neural network

Abstract

This invention discloses a dynamic priority scheduling method for helicopter power distribution based on a physical information neural network. The method includes predefined load classifications, flight phases, and electrical safety boundaries. Fixed power supply loads maintain the highest priority throughout the entire flight cycle, while the rest are adjustable loads. An input feature vector is obtained by acquiring the current flight phase and real-time power supply power, which is input into the physical information neural network to output an initial priority score. Constraints are applied to the fixed power supply loads to maintain their scores at the maximum value. A loss function training network is constructed, incorporating total current safety, adjustable load score differentiation, fixed power supply, and scenario-based constraints for the flight phase. During real-time operation, the priority of adjustable loads is adjusted in conjunction with the network output and power supply power, with gradual disconnection based on scores from low to high. Power is restored in reverse order after restoration. This invention can dynamically adapt to the support needs of different flight phases, alleviating the problems of poor scenario adaptability and lack of physical rule constraints in existing technologies.

Description

Technical Field

[0001] This invention belongs to the field of helicopter electrical system technology, and particularly relates to a dynamic priority scheduling method for helicopter power distribution based on physical information neural networks. Background Technology

[0002] The helicopter electrical system is a critical subsystem for ensuring flight safety, mission execution, and crew comfort; its power distribution management level directly affects the reliability and survivability of the entire aircraft. With the increasing electrification of helicopters, power supply networks are becoming more complex, and the types and number of loads are significantly increasing, encompassing multiple categories such as flight control, navigation and communication, mission equipment, and environmental control. During actual flight, power supply capacity is constantly changing due to factors such as engine power, generator status, and failure modes. Therefore, how to scientifically and rationally allocate limited power resources under constrained power conditions, prioritizing power supply to critical loads while simultaneously ensuring mission completion and system stability, has become one of the core challenges in helicopter power distribution system design.

[0003] Currently, load priority scheduling technology in helicopter power distribution systems typically employs fixed priority scheduling. This technology pre-classifies helicopter loads into several fixed levels according to their importance, such as Level 1 core loads, Level 2 auxiliary loads, and Level 3 other loads, and fixes the load disconnection sequence. When the system detects insufficient total power supply or total current exceeding a preset threshold, it strictly disconnects loads level by level in the fixed sequence of "Level 3 → Level 2" until the total current falls back to the allowable range. In addition, some improved schemes employ threshold-based dynamic scheduling technology, introducing power threshold segmentation logic on top of fixed priorities. For example, when the power supply is below 50A, all Level 3 loads are disconnected; when it is below 40A, some Level 2 loads are disconnected. A few studies have attempted to predict load priorities using traditional machine learning methods such as BP neural networks and support vector machines, but their optimization objective is usually only set as "total current not exceeding the limit."

[0004] Existing technologies have the following significant drawbacks in practical applications: First, fixed-priority scheduling technology cannot dynamically adjust load priorities according to different flight phases such as helicopter takeoff, cruise, hovering, and landing, leading to insufficient power supply to critical auxiliary loads in certain flight phases and poor scenario adaptability. Second, threshold-based scheduling technology's power threshold is only linked to the total power supply and does not dynamically adapt to the characteristics of the flight phase, easily causing voltage fluctuations in the power distribution system and frequent load starts and stops, resulting in poor system stability. Third, traditional machine learning scheduling technology only uses "total current not exceeding the limit" as a single optimization objective, lacking hard constraints on the physical rules of aviation power distribution, causing scheduling results to deviate from actual engineering needs. In addition, existing technologies generally suffer from the problem of convergence in priority scores for non-core loads (such as level 2 and level 3 loads), making it impossible to determine the specific disconnection sequence when power is insufficient, resulting in poor executability of scheduling decisions and difficulty in simultaneously considering flight safety, mission completion, and crew comfort. Summary of the Invention

[0005] Purpose of the invention: The purpose of this invention is to provide a dynamic priority scheduling method for helicopter power distribution based on physical information neural networks. This method aims to solve the problems of poor adaptability of existing helicopter power distribution scheduling technologies, lack of physical rule constraints, easy occurrence of system fluctuations, and convergence of non-core load priorities. It can dynamically adapt to flight phases, integrate multi-dimensional physical rules, and achieve adaptive power linkage and effective priority differentiation.

[0006] Technical solution: The helicopter power distribution dynamic priority scheduling method based on physical information neural network of the present invention includes the following steps:

[0007] S1. Predefine the load classification information, flight phase information and electrical safety boundary information of the helicopter power distribution system. Among them, at least one type of load is set as a fixed power supply load that maintains the highest priority throughout the entire flight cycle, and the remaining loads are set as adjustable loads.

[0008] S2. Obtain the current flight phase information and real-time power supply information of the helicopter, and construct the flight phase information and the real-time power supply information into an input feature vector, input it into the physical information neural network, and output the initial priority score corresponding to each load.

[0009] S3. Apply physical constraints to the initial priority score to keep the priority score of the fixed power-saving load at a preset maximum value, and retain the priority score of the adjustable load.

[0010] S4. Construct a loss function that includes total current safety constraints, adjustable load priority fractionalization constraints, fixed power supply load constraints, and flight phase scenario-based constraints, and perform supervised training on the physical information neural network until the preset convergence condition is met.

[0011] S5. During real-time operation, the priority score of the adjustable load is adjusted in conjunction with the priority score output by the trained physical information neural network and the real-time power supply. The adjustable load is then gradually cut off in order of increasing priority score. When the real-time power supply recovers, the previously cut-off adjustable load is gradually restored in the reverse order.

[0012] This invention distinguishes between fixed power supply loads and adjustable loads by predefining load and flight phase information (S1), solving the problems of poor scenario adaptability and lack of rigid constraints on core loads in existing methods. It uses a physical information neural network to fuse real-time flight phase and power supply (S2) and applies physical constraints to keep fixed power supply loads with the highest priority (S3), making up for the lack of physical rule constraints in traditional methods. At the same time, it constructs a loss function that includes current safety, priority differentiation, and flight phase scenario-based processing for training (S4), which alleviates the convergence of priorities of non-core loads. In real-time operation, it adjusts, gradually cuts off and restores power supply based on priority scores and power supply (S5), effectively suppressing system fluctuations and realizing adaptive power management and fine-grained scheduling for helicopter dynamic operating conditions.

[0013] Preferably, the load classification information includes fixed power-protected loads and adjustable loads, wherein the fixed power-protected loads do not participate in the progressive cutoff control throughout the entire flight cycle; the flight phase information includes seven flight phases: takeoff, climb, cruise, maneuver, hover, descent, and landing, and defines corresponding scenario-based physical constraint rules for each phase, which specify the non-core loads that need to be prioritized for protection in the current phase and their target priorities; the electrical safety boundary information includes the total current safety margin of the helicopter power supply system, and sets the total load current to always be lower than the difference between the real-time power supply current and the current safety margin.

[0014] The aforementioned preferred scheme, by clearly distinguishing between fixed power-protected loads and adjustable loads, ensures that critical loads do not participate in cutoff control throughout the entire flight cycle, thus improving the stability of core power supply. For the seven flight phases—takeoff, climb, cruise, maneuvering, hovering, descent, and landing—scenarios-based physical constraint rules are defined, providing differentiated guidance for non-core loads requiring priority protection and their target priorities at different phases, enhancing the adaptability of scheduling strategies to phase changes. Simultaneously, the introduction of a total current safety margin constraint limits the total load current to within the difference between the real-time power supply and the safety margin, mitigating system overload risks and providing clearer electrical safety boundaries for priority scheduling.

[0015] Preferably, the physical information neural network adopts a fully connected feedforward neural network structure, with 8 neurons in the input layer, corresponding to the 7-dimensional one-hot encoding and 1-dimensional real-time power supply value during the flight phase; the first hidden layer has 256 neurons and uses the ReLU activation function, the second hidden layer has 128 neurons and uses the ReLU activation function, the number of neurons in the output layer is consistent with the number of loads and uses the Sigmoid activation function, and the output value is limited to between 0 and 1.

[0016] The above-mentioned preferred scheme adopts a fully connected feedforward neural network structure. The input layer integrates the one-hot encoding of the flight phase with the real-time power supply, which can effectively represent the coupling relationship between phase switching and power change. The multi-layer hidden layer, combined with the ReLU activation function, helps to extract nonlinear features in power grid scheduling and enhances the model's ability to express complex power supply scenarios. The output layer uses the Sigmoid activation function to limit the priority score within a fixed range, providing a unified numerical basis for subsequent physical constraint application and score comparison, which is convenient for connection with the engineering boundary of the helicopter power distribution system.

[0017] Preferably, the physical constraint processing includes masking the output results corresponding to the fixed power supply load, so that the priority score of the fixed power supply load is constant at 1, and the priority scores of the other adjustable loads are kept at the original output values ​​of the physical information neural network.

[0018] The above-mentioned preferred scheme masks the output results of fixed power-protected loads, keeping their priority scores constant at the upper bound value, so that they do not participate in the cut-off order in subsequent scheduling, which helps to strengthen the full-cycle power supply guarantee of critical loads. At the same time, it retains the original network output values ​​of adjustable loads, so that the priority scores can still reflect the network's response to real-time operating conditions while meeting rigid constraints, alleviating the conflict between physical rules and neural network learning, and improving the interpretability and engineering feasibility of scheduling decisions.

[0019] Preferably, the loss function satisfies: Loss = loss1 + loss2 + loss3 + phase_loss; where loss1 is the total current safety constraint loss, satisfying... I_total is the sum of the products of each load priority score and its corresponding rated current, where the rated current of each load is a predefined fixed parameter. I_power is the real-time power supply, B is the safety margin, and α is the total current safety constraint weight. loss2 is the priority score differentiation constraint loss, which satisfies... Where S_adjustable is the priority score of the adjustable load, the outer mean is calculated for all adjustable load scores within the batch, and the inner mean is calculated for the adjustable load scores within a single sample; β is the differentiation constraint weight; loss3 is the fixed power supply load mandatory power supply constraint loss, satisfying... Where S_core is the priority score of the fixed power supply load, γ is the power supply constraint weight; phase_loss is the flight phase scenario-based constraint loss, which is used to apply target priority constraints to the preset core auxiliary loads under different flight phases.

[0020] The aforementioned preferred scheme constructs a composite loss function comprising four sub-losses: total current safety constraint, priority score differentiation constraint, fixed power supply load constraint, and flight phase scenario-based constraint. This function guides the network to learn priority outputs that conform to helicopter power distribution engineering rules from multiple dimensions. Specifically, the total current safety constraint helps limit the total load current within the range of power supply capacity and safety margin, mitigating overload risks; the priority score differentiation constraint promotes score differences among adjustable loads, reducing the ambiguity in disconnection order caused by priority convergence; the fixed power supply load constraint forces its score to approach the upper bound, ensuring the stability of power supply to critical loads; and the flight phase scenario-based constraint assigns corresponding priority biases to key non-core loads in each phase, improving the adaptability of the scheduling strategy to dynamic flight conditions.

[0021] Preferably, the scenario-based constraint losses during the flight phases include: applying priority constraint losses to hydraulic pumps and fuel pumps during the maneuvering phase; applying priority constraint losses to the air conditioning system and cockpit lighting during the cruise phase; applying priority constraint losses to warning lights and backup equipment during the takeoff phase; and applying priority constraint losses to their respective preset core auxiliary loads during the climb, hovering, descent, and landing phases. All constraint losses are calculated using the mean square error method, with a phase weight of 10 for the maneuvering, cruise, and takeoff phases, and a phase weight of 8 for the climb, hovering, descent, and landing phases.

[0022] The aforementioned preferred scheme applies priority constraint losses to core auxiliary loads such as hydraulic pumps, fuel pumps, air conditioning systems, cockpit lighting, warning lights, and backup equipment for different flight phases, including maneuvering, cruise, takeoff, climb, hovering, descent, and landing. This helps to prioritize non-core loads that require special protection in each phase, enhancing the adaptability of the scheduling strategy to changes in phase conditions. The constraints for each phase are in the form of mean square error, and with differentiated phase weight settings, the constraint strength can be adjusted according to the safety sensitivity of different flight phases. This alleviates the scenario adaptation bias caused by uniform weights and improves the scheduling rationality of the physical information neural network during phase transitions.

[0023] Preferably, the supervised training includes: generating training samples based on random sampling, wherein the training samples include at least flight phase independent encoding, real-time power supply and predefined load rated current parameters; iteratively optimizing the physical information neural network using the Adam optimizer; performing forward propagation, loss calculation, backpropagation and parameter update in each iteration; and determining that training is complete when the total loss value, the fixed power supply load output score and the total current constraint meet the preset convergence conditions.

[0024] The aforementioned preferred scheme generates training samples covering flight phase coding, power supply, and load rated current through random sampling, which helps improve the network's generalization ability to different operating condition combinations. Iterative optimization using the Adam optimizer, combined with a complete process of forward propagation, loss calculation, backpropagation, and parameter updates, guides the network to gradually meet the preset convergence conditions. Training is considered complete only when the total loss value, the fixed power supply load output score, and the total current constraint all reach the expected range simultaneously. This helps achieve a balance across multiple engineering constraint dimensions, mitigating the risk of scheduling decision bias caused by convergence of a single metric while other constraints are not met.

[0025] Preferably, during the supervised training process, the training dataset is constructed as follows: training samples are randomly generated again in each iteration, instead of using a fixed pre-stored dataset; each sample in a single batch contains randomly sampled flight phase independent codes, real-time power supply values ​​randomly sampled within a preset power supply range, and predefined load rated current parameters, so that the training samples seen in each iteration are different, in order to cover all scheduling scenarios in the entire flight cycle and the full power range.

[0026] In the above-mentioned preferred scheme, training samples are randomly regenerated in each iteration during supervised training, rather than relying on a fixed pre-stored dataset. This allows the network to encounter continuously changing combinations of operating conditions throughout the entire flight cycle and power range. This helps alleviate the model's tendency to overfit to limited samples and enhances its generalization ability to unseen scheduling scenarios. Simultaneously, by dynamically covering the intersection of different flight phases and power supply conditions, the physical information neural network can learn more robust priority mapping relationships, providing a more stable decision-making basis for subsequent real-time coordinated adjustments.

[0027] Preferably, the step of adjusting the priority score of the adjustable load in a coordinated manner includes: multiplying the initial priority score output by the physical information neural network by the ratio of the real-time power supply and the rated maximum power to obtain the adjusted priority score; when the total current calculated based on the adjusted priority score exceeds the threshold corresponding to the electrical safety boundary information, only the adjustable loads in the on state are cut off, and only the adjustable load with the lowest priority score is cut off each time, and the total current is recalculated after each cut-off.

[0028] The above-mentioned preferred scheme achieves adaptive linkage adjustment of the priority score to the current power supply capacity by multiplying the initial priority score output by the ratio of the real-time power supply to the rated maximum power. This helps to lower the priority ranking benchmark of adjustable loads as a whole when power is tight. When the total current exceeds the electrical safety boundary, only the adjustable loads in the connected state are disconnected. Each time, only the load with the lowest priority score is disconnected and the total current is recalculated round by round. This gradual single-path disconnection strategy helps to alleviate the risk of sudden changes in system power and power supply fluctuations caused by simultaneous disconnection of multiple paths, and enhances the smoothness and controllability of the power distribution scheduling process.

[0029] Preferably, the restoration of the disconnected adjustable loads includes: when the real-time power supply recovers and the current total current is lower than the threshold corresponding to the electrical safety boundary information, restoring the disconnected adjustable loads step by step in order of priority score from high to low, recalculating the total current after each load is restored, until the current total current is less than the difference between the real-time power supply and the preset safety margin; during the restoration process, the fixed power supply loads are never involved in the restoration order determination, and only the adjustable loads are operated on.

[0030] The preferred scheme described above, when power recovers and the total current is below the safety boundary, gradually restores the disconnected adjustable loads in descending order of priority score. The total current is recalculated after each load is restored, until the difference between the power supply and safety margin is met. This gradual single-load restoration strategy helps mitigate the power surge and secondary overload risks caused by simultaneous access of multiple loads. Simultaneously, fixed power-protected loads are never involved in the restoration sequence determination; only adjustable loads are operated on, further maintaining the power supply stability of critical loads and ensuring a smooth restoration process within the system's capacity limits.

[0031] Beneficial Effects: Compared with existing technologies, this invention has the following significant advantages: 1. By predefining seven flight phases and electrical safety boundaries, this invention constructs flight phases and real-time power supply as input features, and builds a loss function containing total current safety constraints and phase-specific constraints. This enables the model to dynamically adapt to the protection requirements of different flight phases, embedding physical rules into scheduling decisions, thus alleviating the problems of poor scenario adaptability and lack of physical rule constraints in existing technologies; 2. Setting a priority score differentiation constraint term in the loss function promotes a differentiated distribution of priority scores for adjustable loads, enabling non-core loads to have clearer priority ranking and cutoff order in different scenarios. This helps to alleviate the convergence of non-core load priorities and scheduling decisions in existing technologies. 3. Implement a progressive disconnection process from low to high priority scores, disconnecting only one lowest priority load at a time and gradually restoring it in reverse order as power recovers; simultaneously, link the neural network output with the real-time power supply to help suppress voltage fluctuations caused by priority jumps and batch switching, reducing the risk of frequent load starts and stops; 4. Set at least one type of load as a fixed power-saving load that maintains the highest priority throughout the entire flight cycle, apply a mask to the model output to keep its priority score constant at the preset maximum value, and set a high-weight power-saving constraint term in the loss function, combined with the total current safety boundary constraint, to form a dual safety mechanism, so that the core safety load does not participate in the disconnection control, reducing the risk of unexpected power loss of the core load. Attached Figure Description

[0032] Figure 1 This is a schematic diagram of the method flow of the present invention;

[0033] Figure 2 This is a schematic diagram of the PINN model principle of the present invention;

[0034] Figure 3 This is a schematic diagram illustrating the working principle of the system of the present invention. Detailed Implementation

[0035] The technical solution of the present invention will be further described below with reference to the accompanying drawings.

[0036] This embodiment discloses a helicopter automatic power distribution system and dynamic priority allocation method based on Physical Information Neural Network (PINN). The core is to achieve dynamic priority allocation of airborne loads under different flight phases and power supply levels by integrating physical rules of the flight phase with a PINN model of the neural network. This solves the problem of load power supply and disconnection decisions in helicopter power distribution systems when power dynamically changes. (Refer to...) Figure 1The overall system flow is as follows: load / flight phase parameter definition → PINN model construction → multi-constraint loss function design → model training → flight phase simulation verification and visualization verification. The following section uses a 12-channel load as an example to illustrate the specific implementation of this invention.

[0037] Step S1: System parameter definition and load / flight phase division

[0038] 1. Load parameter definition

[0039] The system defines the types and quantities of helicopter loads, as well as their numbers, names, and rated currents (in A), as shown in Table 1. Level 1 loads are core safety loads, maintaining the highest priority throughout the entire flight cycle and not participating in cutoff control. Level 2 / 3 loads are dynamically adjustable loads, with priority scores output by the PINN model and adaptively allocated.

[0040] Table 1 Load Parameter Definition Table

[0041] 0 Inertial navigation 5.0 Level 1 (Fixed Power Supply) Forced priority score is 1 1 Satellite navigation 6.0 Level 1 (Fixed Power Supply) Forced priority score is 1 2 Shortwave radio 7.0 Level 1 (Fixed Power Supply) Forced priority score is 1 3 Warning lights 4.0 Level 2 / 3 (Dynamic) No phase priority constraints 4 hydraulic pump 5.0 Level 2 / 3 (Dynamic) No phase priority constraints 5 fuel pump 3.0 Level 2 / 3 (Dynamic) No phase priority constraints 6 backup equipment 6.0 Level 2 / 3 (Dynamic) No phase priority constraints 7 Cockpit Lighting 3.0 Level 2 / 3 (Dynamic) No phase priority constraints 8 Cabin lighting 2.0 Level 2 / 3 (Dynamic) No phase priority constraints 9 air conditioning system 4.0 Level 2 / 3 (Dynamic) No phase priority constraints 10 Entertainment System 5.0 Level 2 / 3 (Dynamic) No phase priority constraints 11 Heating system 3.0 Level 2 / 3 (Dynamic) No phase priority constraints

[0042] The rated current of the above loads is quantified using the matrix [5.0,6.0,7.0,4.0,5.0,3.0,6.0,3.0,2.0,4.0,5.0,3.0]. Level 1 loads (numbered 0-2) are core safety loads with a mandatory priority score of 1; Level 2 / 3 loads (numbered 3-11) are dynamically adjustable loads, with priority scores output by the PINN model and adaptively allocated.

[0043] 2. Flight phase division and scenario-based rules

[0044] The helicopter flight phases are divided into 7 categories, as shown in Table 2. The scenario-based physical constraints for each phase specify the non-core loads that need to be prioritized and their target priorities for the current phase. The specific constraints are reflected in the loss function.

[0045] Table 2 Flight Phase Division and Scenario-Based Constraints

[0046] 0 take off Except for the Level 1 core security load, all other loads are dynamically adjusted, with no dedicated priority loads. 1 Climb Except for the Level 1 core security load, all other loads are dynamically adjusted, with no dedicated priority loads. 2 cruise Except for the Level 1 core security load, all other loads are dynamically adjusted, with no dedicated priority loads. 3 maneuver Except for the Level 1 core security load, all other loads are dynamically adjusted, with no dedicated priority loads. 4 Hover Except for the Level 1 core security load, all other loads are dynamically adjusted, with no dedicated priority loads. 5 decline Except for the Level 1 core security load, all other loads are dynamically adjusted, with no dedicated priority loads. 6 landing Except for the Level 1 core security load, all other loads are dynamically adjusted, with no dedicated priority loads.

[0047] The flight phase is independently encoded and converted into a 7-dimensional vector input to the PINN model. For example, the takeoff phase is encoded as [1,0,0,0,0,0,0,0] and the cruise phase is encoded as [0,0,1,0,0,0,0].

[0048] Step S2: PINN Model Construction and Physical Hard Constraint Embedding

[0049] 1. Model Structure Design

[0050] The fully connected neural network structure constructed in this system is as follows: the input layer has 8 neurons, corresponding to 7-dimensional one-hot encoding of the flight phase and 1-dimensional real-time power supply value; the first hidden layer has 256 neurons, using the ReLU activation function; the second hidden layer has 128 neurons, using the ReLU activation function; the output layer has 12 neurons, corresponding to the priority scores of 12 load paths, and the Sigmoid activation function is used to limit the output value to between 0 and 1.

[0051] like Figure 2 As shown, the core principle of the PINN model is as follows: First, define the PINN model category, initialize a 7-layer fully connected network (input layer + two hidden layers + output layer, of which there are 2 hidden layers, plus input and output, a total of 4 layers. The 7 layers here are for illustration, and the actual structure is a fully connected structure), and perform forward propagation to obtain the original output. During the training process, each iteration performs gradient zeroing, calculates four types of losses (total current safety constraint loss, priority score differentiation constraint loss, level 1 load forced power supply constraint loss, flight phase scenario constraint loss), and then backpropagates the total loss, updates the network parameters through the optimizer, and returns the loss value for monitoring convergence. In the model evaluation and real-time inference phase, the system traverses 300 time steps, and performs the following operations at each time step: (1) Input the current flight phase and real-time power supply, and the model outputs the priority score of each load; (2) Dynamically adjust the non-core load score according to the real-time power supply; (3) If the total current exceeds the power threshold, cut off the non-level 1 load with the lowest score in order of priority score from low to high.

[0052] 2. Physical hard constraint embedding

[0053] Reference Figure 2 The "hard safety constraint" part sets insurmountable hard safety constraints at the model output: for the outputs corresponding to Level 1 core safety loads (numbered 0-2), the priority score is forcibly set to 1, that is, the original output is overwritten with 1 through a mask; for the outputs corresponding to non-core adjustable loads (numbered 3-11), the original output values ​​of the neural network are retained. This mechanism ensures from the model structure level that the absolute power supply requirement of Level 1 core loads will not be breached by any scenario.

[0054] Step S3: Design of Multi-Constraint Loss Function

[0055] To enable the PINN model to learn the physical rules of helicopter power distribution and the differentiated requirements of flight phases, a loss function consisting of four parts is designed: Total Loss = loss1 + loss2 + loss3 + phase_loss. The physical meaning of each loss term is as follows:

[0056] 1. Total current safety constraint loss (loss1)

[0057] Constraint objective: The actual total current of all loads (the sum of the products of priority score and rated load current) shall not exceed the difference between the real-time power supply and the safety margin, thus avoiding overload. The formula is:

[0058]

[0059] In this embodiment, the weighting coefficient α = 10, and the safety margin B = 2A. I_total is the total load current, and I_power is the real-time power supply.

[0060] 2. Priority score differentiation constraint loss (loss2)

[0061] Constraint Objective: To avoid convergence of priority scores for dynamic loads, ensure the model can distinguish the priorities of different loads, and guarantee a clear disconnection sequence. The formula is:

[0062]

[0063] In this implementation, the weighting coefficient β = 5. By maximizing the variance of the non-core load scores, the scores are forced to exhibit significant differences.

[0064] 3. Level 1 load forced power supply constraint loss (loss3)

[0065] Constraint Objective: As a reinforcement of the hard constraints on the model structure, ensure that the priority score of the Level 1 core security load is strictly 1. The formula is:

[0066]

[0067] In this implementation, the weight coefficient γ=100, which is the highest value among all loss terms, ensuring that this constraint is the highest priority optimization objective of the model.

[0068] 4. Phase loss based on scenario-specific constraints during the flight phase.

[0069] This loss term is used to force the model to learn the scenario-based priority requirements of different flight phases. The flight phase to which each sample in a single batch belongs is identified through independent encoding. When the number of samples in the corresponding phase is greater than 0, the mean square error between the score of the key load to be guaranteed in that phase and the target value of 0.9 is calculated, multiplied by the corresponding constraint weight, and then added to the phase_loss. Specific constraints are shown in Table 3.

[0070] Table 3. Weighting Table of Differentiated Constraints in Flight Phases

[0071] Maneuver Phase (Phase 3) hydraulic pump, fuel pump 10 Cruise Phase (Phase 2) Air conditioning system, cockpit lighting 10 Takeoff phase (Phase 0) Alarm lights, backup equipment 10 Climbing, hovering, descent, and landing phases Each has its own preset core auxiliary load 8

[0072] Step S4: Supervised Training Process of PINN Model

[0073] 1. Training environment and parameter settings: Tensor computing platform based on CUDA or CPU; Adam optimizer is used, and the learning rate is set to 1e-3; the sample size per batch is 256; the total number of training rounds is 3000 rounds, and the loss value is output every 500 rounds to monitor the training convergence.

[0074] 2. Training Data Generation: Training samples are randomly generated in each iteration instead of using a fixed pre-stored dataset. Specifically, integers from 0 to 6 are randomly generated as flight phases and converted into 7-dimensional one-hot encodings; real-time power supply values ​​are randomly generated within the range of 18A to 55A, covering the typical power variation range of helicopters.

[0075] 3. Training steps:

[0076] Initialize the PINN model and load it onto the computing device;

[0077] Initialize the Adam optimizer and set the learning rate to 1e-3;

[0078] Iterate through 3000 training rounds, performing the following operations in each round:

[0079] Generate batch-based flight phase one-hot codes and power supply tensors;

[0080] Perform forward propagation and calculate the original priority score output by the model;

[0081] Calculate the total loss as: Loss = loss1 + loss2 + loss3 + phase_loss;

[0082] The gradient is calculated through backpropagation, and the model parameters are updated using the Adam optimizer.

[0083] Output the current loss value every 500 rounds to monitor training convergence.

[0084] like Figure 3 As shown in the system's working principle diagram, after the overall process begins, the computing equipment is first initialized, then the PINN model is instantiated and the Adam optimizer is initialized. Next, the training loop begins, executing 3000 training rounds. In each round, an independent encoding for one of the seven flight stages is randomly generated, the loss is calculated, and the model is updated. During training, all seven flight stages need to be traversed to ensure the model learns the differentiated constraints of each stage. Once all flight stages have been traversed, training ends, and the converged PINN model is output for subsequent real-time scheduling.

[0085] 4. Convergence Verification Criteria: The model is considered to have successfully converged when it meets all of the following conditions: the total loss value is stable below 5 with no significant fluctuations; the priority score of level 1 load is always 1, with a constraint satisfaction rate of 100%; the total current constraint satisfaction rate is ≥99.9%; the priority scores of the core auxiliary loads in each flight phase are all ≥0.85, with a scenario constraint satisfaction rate of 100%. After training convergence, the model parameters are saved for subsequent real-time scheduling and simulation verification.

[0086] Step S5: Flight Phase Power Distribution Simulation Verification and Real-Time Scheduling Execution

[0087] 1. Simulation hardware environment: CPU is AMD-8945HX, graphics card is RTX5060, operating system is WIN11 25H2.

[0088] 2. Simulation parameter settings: Total duration 300 seconds, 300 time steps generated at 1 second / step; power supply simulates the dynamic attenuation scenario of helicopter power, linearly decreasing from 55A to 18A; load cut-off logic follows the progressive cut-off and recovery rules described below.

[0089] 3. Real-time scheduling and simulation results

[0090] Real-time parameter acquisition and model inference: With a fixed time step of 1 second, the current flight phase (such as takeoff phase, cruise phase) and real-time power supply are acquired in real time to construct an 8-dimensional input feature vector, which is then input into the trained PINN model. The model forward inference outputs the initial dynamic priority scores of each load.

[0091] Smooth adjustment of priority for power linkage: For non-core adjustable loads, the priority is adjusted according to the formula: adjusted score = initial score × (real-time power supply / rated maximum power) to achieve a smooth and gradual change in priority with the power supply.

[0092] Progressive load cut-off control: Calculates total current in real time. When the total current exceeds the safety threshold (real-time power supply - 2A), the cut-off logic is executed: the cut-off range is limited to non-core adjustable loads; only for non-core loads currently connected, the load with the lowest priority score is cut off at a time, in ascending order of priority score; after each load is cut off, the total current is recalculated, and if it still exceeds the limit, the above steps are repeated until the total current meets the safety threshold requirement.

[0093] Recovery logic: When the real-time power supply recovers, the disconnected non-core loads are gradually restored in order of priority score from high to low. The total current is recalculated after each load is restored until the total current is close to but does not exceed the safety threshold.

[0094] Priority score dynamic adjustment results: The priority scores of each load show significant differences.

Claims

1. A physical information neural network-based dynamic priority scheduling method for helicopter power distribution, characterized in that, Includes the following steps: S1. Predefine the load classification information, flight phase information and electrical safety boundary information of the helicopter power distribution system. Among them, at least one type of load is set as a fixed power supply load that maintains the highest priority throughout the entire flight cycle, and the remaining loads are set as adjustable loads. S2. Obtain the current flight phase information and real-time power supply information of the helicopter, and construct the flight phase information and the real-time power supply information into an input feature vector, input it into the physical information neural network, and output the initial priority score corresponding to each load. S3. Apply physical constraints to the initial priority score to keep the priority score of the fixed power-saving load at a preset maximum value, and retain the priority score of the adjustable load. S4. Construct a loss function that includes total current safety constraints, adjustable load priority fractionalization constraints, fixed power supply load constraints, and flight phase scenario-based constraints, and perform supervised training on the physical information neural network until the preset convergence condition is met. S5. During real-time operation, the priority score of the adjustable load is adjusted in conjunction with the priority score output by the trained physical information neural network and the real-time power supply. The adjustable load is then gradually cut off in order of increasing priority score. When the real-time power supply recovers, the previously cut-off adjustable load is gradually restored in the reverse order.

2. The method according to claim 1, characterized in that, The load classification information includes fixed power-protected loads and adjustable loads, wherein the fixed power-protected loads do not participate in the gradual cutoff control throughout the entire flight cycle; the flight phase information includes seven flight phases: takeoff, climb, cruise, maneuver, hover, descent, and landing, and defines corresponding scenario-based physical constraint rules for each phase. The scenario-based physical constraint rules specify the non-core loads that need to be prioritized for protection in the current phase and their target priorities; the electrical safety boundary information includes the total current safety margin of the helicopter power supply system, and sets the total load current to always be lower than the difference between the real-time supply current and the current safety margin.

3. The method according to claim 1, characterized in that, The physical information neural network adopts a fully connected feedforward neural network structure. The input layer has 8 neurons, corresponding to the 7-dimensional one-hot encoding and 1-dimensional real-time power supply value during the flight phase. The first hidden layer has 256 neurons and uses the ReLU activation function, the second hidden layer has 128 neurons and uses the ReLU activation function, and the number of neurons in the output layer is the same as the number of loads and uses the Sigmoid activation function. The output value is limited to between 0 and 1.

4. The method according to claim 1, characterized in that, The physical constraint processing includes masking the output results corresponding to the fixed power supply load, so that the priority score of the fixed power supply load is constant at 1, and the priority scores of the other adjustable loads are kept at the original output values ​​of the physical information neural network.

5. The method according to claim 1, characterized in that, The loss function satisfies: Loss = loss1 + loss2 + loss3 + phase_loss; where loss1 is the total current safety constraint loss, satisfying... I_total is the sum of the products of each load priority score and its corresponding rated current, where the rated current of each load is a predefined fixed parameter. I_power is the real-time power supply, B is the safety margin, and α is the total current safety constraint weight. loss2 is the priority score differentiation constraint loss, which satisfies... Where S_adjustable is the priority score of the adjustable load, the outer mean is calculated for all adjustable load scores within the batch, and the inner mean is calculated for the adjustable load scores within a single sample; β is the differentiation constraint weight; loss3 is the fixed power supply load mandatory power supply constraint loss, satisfying... Where S_core is the priority score of the fixed power supply load, γ is the power supply constraint weight; phase_loss is the flight phase scenario-based constraint loss, which is used to apply target priority constraints to the preset core auxiliary loads under different flight phases.

6. The method according to claim 5, characterized in that, The scenario-based constraint losses during the flight phases include: priority constraint losses applied to hydraulic pumps and fuel pumps during the maneuvering phase; priority constraint losses applied to the air conditioning system and cockpit lighting during the cruise phase; priority constraint losses applied to warning lights and backup equipment during the takeoff phase; and priority constraint losses applied to their respective preset core auxiliary loads during the climb, hovering, descent, and landing phases. All constraint losses are calculated using the mean square error method, with a phase weight of 10 for the maneuvering, cruise, and takeoff phases, and a phase weight of 8 for the climb, hovering, descent, and landing phases.

7. The method according to claim 1, characterized in that, The supervised training includes: generating training samples based on random sampling, the training samples including at least flight phase independent encoding, real-time power supply and predefined load rated current parameters; using the Adam optimizer to iteratively optimize the physical information neural network; performing forward propagation, loss calculation, backpropagation and parameter update in each iteration; and determining that training is complete when the total loss value, the fixed power supply load output score and the total current constraint meet the preset convergence conditions.

8. The method according to claim 7, characterized in that, During the supervised training process, the training dataset is constructed as follows: training samples are randomly generated again in each iteration, instead of using a fixed pre-stored dataset; each sample in a single batch contains randomly sampled flight phase independent codes, real-time power supply values ​​randomly sampled within a preset power supply range, and predefined load rated current parameters, so that the training samples seen in each iteration are different, in order to cover all scheduling scenarios in the entire flight cycle and the full power range.

9. The method according to claim 1, characterized in that, The linkage adjustment of the priority score of the adjustable load includes: multiplying the initial priority score output by the physical information neural network by the ratio of the real-time power supply and the rated maximum power to obtain the adjusted priority score; when the total current calculated based on the adjusted priority score exceeds the threshold corresponding to the electrical safety boundary information, only the adjustable load in the on state is cut off, and only the adjustable load with the lowest priority score is cut off each time, and the total current is recalculated after each cut-off.

10. The method according to claim 1, characterized in that, The restoration of the disconnected adjustable loads includes: when the real-time power supply recovers and the current total current is lower than the threshold corresponding to the electrical safety boundary information, the disconnected adjustable loads are restored step by step in order of priority score from high to low. The total current is recalculated after each load is restored until the current total current is less than the difference between the real-time power supply and the preset safety margin. During the restoration process, the fixed power supply loads are not involved in the restoration order determination, and only the adjustable loads are operated on.

Images