Fuel cell adaptive control method and system based on multi-dimensional environment perception

Abstract

The application discloses a fuel cell adaptive control method and system based on multi-dimensional environment perception, relates to the technical field of fuel cells, and collects multi-dimensional state parameters in a preset historical time window to construct a state sequence, carries out differential calculation on the state sequence to extract a first-order change rate and a second-order change rate, arranges original data in the state sequence, the first-order change rate and the second-order change rate according to time steps, maps into a two-dimensional feature matrix, inputs the two-dimensional feature matrix into a pre-trained deep learning network, extracts a dynamic trend feature through the deep learning network, outputs a dynamic compensation amount, superimposes the reference collaborative control instruction set and the dynamic compensation amount, obtains a final collaborative control instruction set, and through dynamic compensation, further improves the dynamic response precision and environmental adaptability of the system on the basis of original feedforward prediction control based on a multi-dimensional performance mapping table.

Description

Technical Field

[0001] This invention relates to the field of fuel cell technology, and in particular to an adaptive control method and system for fuel cells based on multi-dimensional environmental perception. Background Technology

[0002] Proton exchange membrane fuel cells (PEMFCs) have become an ideal power source for long-endurance unmanned aerial vehicles (UAVs) due to their advantages such as high energy density, low noise, and zero emissions. However, UAVs often experience drastic changes in altitude, ambient temperature, and humidity when performing tasks such as reconnaissance, mapping, and inspection. In high-altitude environments, such as plateaus or high-altitude flights, the air becomes thinner, atmospheric pressure decreases, and oxygen partial pressure drops significantly, directly affecting the electrochemical reaction rate and output power of the fuel cell stack.

[0003] To ensure sufficient oxygen supply to the fuel cell stack for the chemical reaction, current technologies typically increase the cathode air intake by increasing the fan speed. However, this simple control strategy suffers from one or more technical problems. First, parasitic losses increase dramatically. In low-pressure environments, the fan's power consumption increases exponentially to achieve the same pressurization effect as at sea level. This power consumption, i.e., parasitic losses, severely erodes the net output power of the fuel cell system, thereby shortening the effective endurance of the drone. Under extreme altitude conditions, the fan power consumption may even exceed a significant proportion of the stack's own output, leading to a sharp decline in system efficiency.

[0004] Second, water and thermal management deteriorate. Excessively high fan speeds blow large amounts of dry, cold air into the fuel cell stack, easily causing the proton exchange membrane to lose water and dry out, increasing membrane resistance and reducing stack performance and lifespan. At the same time, the low-temperature environment makes inlet air temperature difficult to manage. Without effective inlet air heating or flow control, uneven temperature distribution inside the stack further exacerbates the challenges of water and thermal management.

[0005] Third, the risk of oxygen starvation. During a sudden surge in load current, if the turbine response is insufficient, or if the turbine reaches its maximum speed at extreme altitudes but still cannot meet the oxygen demand, the fuel cell stack will instantly fall into an "oxygen starvation" state. This state may cause irreversible damage to the cathode catalyst layer, such as carbon corrosion and platinum dissolution, seriously affecting the reliability and safety of the system.

[0006] While some studies have proposed coordinated control of the blower and cathode back pressure valve to more efficiently regulate cathode pressure and flow, these methods are mostly reactive control, adjusting based on feedback from internal sensors. For example, the blower speed is increased only after a pressure sensor detects a pressure drop. This reactive approach lacks anticipation of the rapidly changing operating environment of drones, and the lag in adjustment can easily lead to momentary oxygen shortages. Furthermore, these methods still prioritize "meeting load requirements" above all else, lacking the ability to proactively manage the load in extreme environments and failing to fundamentally prevent the fuel cell stack from entering hazardous operating conditions. Summary of the Invention

[0007] This invention aims to address the technical problems of excessively high parasitic losses, poor environmental adaptability, and risks of oxygen starvation and membrane dehydration when existing fuel cell control technologies are applied to UAVs in high-altitude and low-temperature coupled environments. It also provides a fuel cell adaptive control method and system based on multi-dimensional environmental perception.

[0008] To achieve the above objectives, the present invention provides the following technical solution:

[0009] An adaptive control method for fuel cells based on multi-dimensional environmental perception includes the following steps:

[0010] The system acquires multi-dimensional state parameters, which include at least: the current altitude acquired by the positioning module, the ambient temperature acquired by the sensor, and the target load current requirement received from the power management unit.

[0011] Using the altitude, ambient temperature, and target load current requirement as input indices, the system queries a multi-dimensional performance mapping table pre-set in the controller to obtain a benchmark collaborative control instruction set corresponding to the current input index. The benchmark instruction set includes at least: benchmark fan speed, benchmark cathode back pressure valve opening, and benchmark maximum allowable output current threshold.

[0012] Multidimensional state parameters within a preset historical time window are collected to construct a state sequence, and the state sequence is differentially calculated to extract the first-order rate of change and the second-order rate of change.

[0013] The original data, the first-order rate of change, and the second-order rate of change in the state sequence are arranged according to the time step and mapped into a two-dimensional feature matrix.

[0014] The two-dimensional feature matrix is ​​input into a pre-trained deep learning network, which extracts dynamic trend features and outputs dynamic compensation amount.

[0015] The reference coordinated control instruction set is superimposed with the dynamic compensation amount to obtain the final coordinated control instruction set, which includes at least: the final fan speed, the final cathode back pressure valve opening, and the final maximum allowable output current threshold.

[0016] The operation of the wind turbine is controlled according to the final wind turbine speed, the operation of the back pressure valve is controlled according to the final cathode back pressure valve opening, and the final maximum allowable output current threshold is sent to the power converter or power management unit to limit the actual output current of the fuel cell stack from exceeding the threshold.

[0017] Preferably, the multidimensional state parameters also include system internal parameters, which include at least one or more of the following: temperature, voltage, and current of the fuel cell stack; the state sequence is composed of the altitude, ambient temperature, target load current requirement, and the system internal parameters.

[0018] Preferably, the process of constructing the two-dimensional feature matrix includes:

[0019] The original values ​​of each parameter in the state sequence at each time step are taken as zero-order data;

[0020] The difference between each parameter at adjacent time steps is calculated as the first-order rate of change;

[0021] The difference between the first-order rate of change at adjacent time steps is calculated as the second-order rate of change;

[0022] The zero-order data, the first-order rate of change, and the second-order rate of change are stacked in the vertical direction according to parameter type, and expanded in the horizontal direction according to time steps to generate the two-dimensional feature matrix.

[0023] Preferably, the deep learning network is a convolutional neural network, comprising an input layer, at least one convolutional layer, a pooling layer, and a fully connected layer connected in sequence.

[0024] The convolutional layer is used to extract local dynamic texture features from the two-dimensional feature matrix;

[0025] The fully connected layer is used to map the extracted features to the dynamic compensation values, which include: fan speed compensation value, back pressure valve opening compensation value, and maximum allowable output current threshold compensation value.

[0026] Preferably, the training objective function of the convolutional neural network includes a load tracking error term and a system parasitic loss term. The loss value is obtained by weighted summation of the load tracking error term and the system parasitic loss term, so as to simultaneously optimize the current tracking accuracy and system energy consumption during the training process.

[0027] Preferably, the multidimensional performance mapping table is obtained by calibration using experimental data or a high-fidelity system simulation model. The calibration range covers the entire altitude range, ambient temperature range, and load current range of the expected flight of the UAV. The goal of constructing the mapping table is to minimize the total parasitic loss of the system while ensuring that the oxygen stoichiometry is within a safe range.

[0028] Preferably, the fan is any one of the following: a blower, an air compressor, or an axial fan.

[0029] An adaptive control system for fuel cells based on multi-dimensional environmental perception includes:

[0030] The status acquisition module is configured to acquire multi-dimensional status parameters in real time. The multi-dimensional status parameters include at least: the current altitude acquired by the positioning module, the ambient temperature acquired by the temperature sensor, and the target load current requirement received from the power management unit.

[0031] The main channel control module stores a multi-dimensional performance mapping table internally and is connected to the status acquisition module. It is used to receive the multi-dimensional status parameters and query the multi-dimensional performance mapping table using the altitude, ambient temperature and target load current requirements as input indexes. It outputs a reference collaborative control instruction set, which includes at least: reference fan speed, reference cathode back pressure valve opening, and reference maximum allowable output current threshold.

[0032] The feature construction module, connected to the state acquisition module, is used to collect multi-dimensional state parameters within a preset historical time window to construct a state sequence, perform difference calculation on the state sequence to extract the first-order rate of change and the second-order rate of change, and arrange the original data, the first-order rate of change and the second-order rate of change in the state sequence according to the time step and map them into a two-dimensional feature matrix.

[0033] The prediction compensation module internally stores a pre-trained deep learning network and is connected to the feature construction module. It is used to receive the two-dimensional feature matrix, extract dynamic trend features through the deep learning network, and output dynamic compensation amount.

[0034] The instruction synthesis module is connected to the main channel control module and the prediction compensation module respectively. It is used to superimpose the reference collaborative control instruction set with the dynamic compensation amount to obtain the final collaborative control instruction set. The final instruction set includes at least: the final fan speed, the final cathode back pressure valve opening, and the final maximum allowable output current threshold.

[0035] The multi-channel execution module is connected to the instruction synthesis module, the wind turbine driver, the back pressure valve actuator, and the power converter, respectively. It is used to receive the final coordinated control instruction set, send a first control signal to the wind turbine driver according to the final wind turbine speed, send a second control signal to the back pressure valve actuator according to the final cathode back pressure valve opening, and send the final maximum allowable output current threshold to the power converter to limit the actual output current of the fuel cell stack from exceeding the threshold.

[0036] Preferably, the multidimensional state parameters also include system internal parameters, which at least include: the temperature, voltage, and current of the fuel cell stack; the state sequence is composed of the altitude, ambient temperature, target load current requirement, and the system internal parameters.

[0037] Preferably, the deep learning network is a convolutional neural network, comprising an input layer, at least one convolutional layer, a pooling layer, and a fully connected layer connected in sequence; the fully connected layer outputs a fan speed compensation value, a back pressure valve opening compensation value, and a maximum allowable output current threshold compensation value.

[0038] Compared with the prior art, the beneficial effects of the present invention are as follows: Compared with the prior art, the present invention, by introducing a deep learning compensation channel, further improves the dynamic response accuracy and environmental adaptability of the system on the basis of the original feedforward predictive control based on a multi-dimensional performance mapping table, bringing the following significant beneficial effects:

[0039] First, it significantly reduces parasitic losses and improves endurance. The main channel avoids excessive fan operation by coordinating and optimizing the fan speed and back pressure valve opening, minimizing parasitic losses while meeting oxygen supply requirements. The compensation channel predicts and compensates for dynamic deviations through a deep learning model, ensuring that the system always operates at a better operating point, thereby further improving the net output power of the fuel cell system and the effective flight time of the UAV.

[0040] Second, it enhances the adaptability to the entire environment and the smoothness of dynamic response. Based on predictive mapping table control, the system can anticipate upcoming altitude changes; at the same time, by constructing a two-dimensional feature matrix containing second-order rates of change and using a convolutional neural network to extract dynamic texture features, it can keenly capture trends before load abrupt changes and drastic environmental changes, and output compensation amounts in advance, making the adjustment of fan speed, back pressure valve opening and current threshold smoother and more precise, effectively suppressing overshoot and oscillation.

[0041] Third, it improves system reliability and lifespan. The main channel actively limits the output by setting a maximum permissible output current threshold, fundamentally eliminating oxygen starvation. The addition of the compensation channel further optimizes the gas supply matching during the dynamic process, avoiding membrane dryness and local overheating, and extending the stack's lifespan. Even if the deep learning model fails due to extreme interference, the system can still maintain safe operation by relying on the mapping table control of the main channel, demonstrating extremely high engineering robustness.

[0042] Fourth, it is simple to implement and highly scalable. The main channel uses lookup table control, which has low requirements for computing resources; the lightweight convolutional neural network of the compensation channel can run in real time on mainstream embedded controllers. The decoupled architecture of "main channel + compensation channel" adopted in this invention facilitates subsequent algorithm upgrades and model iterations. Attached Figure Description

[0043] Figure 1 This is a diagram of the adaptive control system architecture for a fuel cell according to an embodiment of the present invention.

[0044] Figure 2 This is a flowchart of the adaptive control method for fuel cells according to an embodiment of the present invention.

[0045] The markings for each component in the attached diagram are explained below:

[0046] 10: Positioning Module

[0047] 20: Temperature sensor

[0048] 30: Power Management Unit

[0049] 40: Fan drive

[0050] 50: Back pressure valve actuator

[0051] 60: Power Converter

[0052] 70: Fuel cell stack

[0053] 80: Fan

[0054] 90: Cathode back pressure valve

[0055] 100: Status Acquisition Module

[0056] 200: Main Channel Control Module

[0057] 300: Multi-channel execution module

[0058] 400: Feature Building Module

[0059] 500: Predictive Compensation Module

[0060] 600: Instruction Synthesis Module Detailed Implementation

[0061] To make the objectives, technical solutions, and advantages of this application clearer, the following detailed description is provided in conjunction with the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are merely illustrative and not intended to limit the scope of this application.

[0062] Example 1: Adaptive Control Method

[0063] This embodiment provides a fuel cell adaptive control method based on multi-dimensional environmental perception. This method is applied to a fuel cell system including a fuel cell stack 70, a wind turbine 80, a cathode back pressure valve 90, and a controller. The following is in conjunction with... Figures 1 to 2 This method will be explained in detail.

[0064] This method employs a dual-channel control architecture consisting of a "main channel + compensation channel". The main channel implements feedforward predictive control based on a multi-dimensional performance mapping table to ensure the safety of the system under steady-state and extreme conditions. The compensation channel is based on a deep learning network, which analyzes the dynamic trends of historical state sequences and outputs compensation amounts to correct the main channel commands, thereby improving the accuracy of dynamic response.

[0065] include:

[0066] Obtaining multidimensional state parameters:

[0067] Multidimensional state parameters are acquired in real time by the state acquisition module 100. These multidimensional state parameters include at least three types of parameters:

[0068] External environmental parameters: The current altitude is obtained through the positioning module 10, and the ambient temperature is obtained through the temperature sensor 20. The positioning module 10 can be a Global Positioning System (GPS) module or a BeiDou navigation module, which sends positioning data to the controller via a serial communication interface (such as UART) using the NMEA-0183 protocol. The temperature sensor 20 can be installed on the outside of the UAV in a shaded area, and uses a thermistor or semiconductor temperature sensor, which sends the ambient temperature value to the controller via an analog-to-digital converter interface (such as I2C or SPI).

[0069] Internal system parameters: temperature, voltage, current, etc. of the fuel cell stack 70, for more precise control. The stack temperature can be obtained by thermocouples or thin-film temperature sensors embedded inside the stack; the stack voltage and current can be directly measured by voltage sensors and Hall current sensors, and transmitted to the controller via analog input interface or CAN bus.

[0070] Load demand parameters: Target load current demand received from the power management unit 30. The power management unit 30 is part of the UAV flight control system and is responsible for overall power distribution. It calculates the required electrical power based on flight control commands and the current flight phase, and sends the current demand to the fuel cell controller via CAN bus or PWM duty cycle signal.

[0071] Query the multidimensional performance mapping table to obtain the benchmark command:

[0072] The main channel control module 200 uses the current altitude, ambient temperature, and target load current requirement as input indices to query a pre-calibrated and stored multi-dimensional performance mapping table to obtain the corresponding benchmark collaborative control instruction set. This instruction set includes at least: benchmark wind turbine speed. Reference cathode back pressure valve opening and the reference maximum allowable output current threshold .

[0073] The multidimensional performance mapping table is calibrated through extensive experiments or high-fidelity simulations covering the entire flight envelope at various operating conditions. The calibration process is as follows: First, steady-state tests of the fuel cell stack are conducted at operating condition grid points covering an altitude range of 0 to 6000 meters (500-meter step), an ambient temperature range of -40°C to 50°C (10°C step), and a load current range of 0A to 200A (10A step). At each operating condition point, key indicators such as net system output power, fan parasitic losses, and oxygen stoichiometry are recorded by adjusting the fan speed to 80 rpm and the cathode back pressure valve opening to 90 rpm. Then, data processing (such as linear interpolation or cubic spline interpolation) generates the optimal combination of control commands that achieves the preset performance target across all operating condition grid points. The preset performance target is to minimize the total parasitic losses of the system while ensuring the oxygen stoichiometry remains within a safe range (e.g., 1.8 to 2.5). Finally, these command combinations are stored in Flash or EEPROM in the form of a three-dimensional lookup table, forming the multidimensional performance mapping table 210.

[0074] Construct a historical state sequence and extract the second-order rate of change:

[0075] To capture the dynamic trends of the system, the feature construction module 400 collects multidimensional state parameters from the most recent N time steps (e.g., N=10) to construct a state sequence. The state vector at each time step Includes: altitude Ambient temperature Target load current And optional internal states (such as stack voltage) Current ,temperature ).

[0076] Perform a difference operation on each parameter in the state sequence:

[0077] First-order rate of change (velocity): Calculates the difference between each parameter at adjacent time steps, i.e. This reflects the changing trend of the parameters.

[0078] Second-order rate of change (acceleration): Calculates the difference between the first-order rate of change at adjacent time steps, i.e. Curvature or acceleration that reflects the changing trend.

[0079] The physical significance of introducing the second-order rate of change is that fuel cell systems have thermal inertia and electrochemical inertia. For example, the acceleration of current error can predict an impending overshoot or oscillation. This "dynamic texture" is difficult to represent with a single numerical value.

[0080] Mapped to a two-dimensional feature matrix:

[0081] Arrange the original data (zeroth order), first-order rate of change, and second-order rate of change according to time steps to construct a two-dimensional feature matrix. Assuming M state parameters are selected (e.g., 5), and the time window length N=10, a matrix of size [size missing] is constructed. The matrix (which can be viewed as a single-channel grayscale image). Specific mapping method:

[0082] Rows 1 to M: Zero-order raw data (normalized and mapped to the [0,1] interval).

[0083] Rows M+1 to 2M: First-order rate of change.

[0084] Rows 2M+1 to 3M: Second-order rate of change.

[0085] All data is normalized before mapping to facilitate neural network training.

[0086] Deep learning network predicts dynamic compensation amount:

[0087] The generated two-dimensional feature matrix is ​​input into a pre-trained deep learning network 510. In this embodiment, the deep learning network uses a lightweight convolutional neural network (CNN), including:

[0088] Input layer: receiving The two-dimensional feature matrix.

[0089] Convolutional layers: using multiple A convolutional kernel with a stride of 1 and a ReLU activation function is used to extract local dynamic texture features. For example, the convolutional kernel can identify a pattern where "the current error increases sharply and the second derivative is positive".

[0090] Pooling layer: Employs max pooling (e.g.) This reduces feature dimensionality while retaining salient features.

[0091] Fully connected layer: Flattens the pooled feature map and maps it to three output nodes, each corresponding to a wind turbine speed compensation value. Back pressure valve opening compensation value Maximum allowable output current threshold compensation value .

[0092] This network can run in real time in mainstream embedded controllers (such as STM32H7 and Infineon TC3xx), with an inference time typically less than 100μs, meeting the current loop control cycle requirements.

[0093] Instruction synthesis:

[0094] The instruction synthesis module 600 superimposes the reference instruction from the main channel with the dynamic compensation amount from the compensation channel to obtain the final coordinated control instruction set:

[0095] Final fan speed = Base fan speed + Fan speed compensation value

[0096] Final cathode back pressure valve opening = reference cathode back pressure valve opening + back pressure valve opening compensation value

[0097] Final maximum allowable output current threshold = Reference maximum allowable output current threshold + Maximum allowable output current threshold compensation value

[0098] If the compensation amount causes the threshold to exceed the physically permissible range, then amplitude limiting will be performed.

[0099] Execute coordinated control:

[0100] The multi-channel execution module 300 sends control signals (such as PWM or CAN commands) to the wind turbine driver 40 according to the final wind turbine speed, sends control signals (such as pulses or analog signals) to the back pressure valve actuator 50 according to the final back pressure valve opening, and sends the final maximum allowable output current threshold to the power converter 60 (or power management unit 30) via the CAN bus to limit the actual output current of the fuel cell stack from exceeding the threshold.

[0101] Through the above steps, the system maintains the robustness of the main channel based on the physical model, and significantly improves the dynamic response accuracy by leveraging the deep learning compensation channel.

[0102] Example 2: Training Method of Convolutional Neural Networks

[0103] This embodiment details an offline training method for deep learning networks.

[0104] 1. Data Acquisition and Preprocessing:

[0105] Full-condition dynamic testing was conducted on a fuel cell test bench, covering different scenarios including altitude simulation (via a pressure chamber), temperature changes, and load step changes. Multi-dimensional state parameters (altitude, temperature, load current, stack voltage, current, etc.) and corresponding optimal control commands were simultaneously acquired. The optimal control commands were calculated offline using high-precision optimization algorithms (such as particle swarm optimization or model predictive control) to ensure the system achieves maximum efficiency while meeting oxygen supply constraints.

[0106] 2. Tag Generation:

[0107] For each sampling time, the difference between the "optimal command" and the "main channel baseline command" is calculated and used as the label for the training data. That is:

[0108] Fan speed compensation value label = optimal fan speed - reference fan speed

[0109] Back pressure valve opening compensation value label = optimal back pressure valve opening - reference back pressure valve opening

[0110] Maximum allowable output current threshold compensation value label = Optimal maximum allowable output current threshold - Reference maximum allowable output current threshold

[0111] 3. Feature Construction:

[0112] A corresponding two-dimensional feature matrix is ​​generated for each time step as the input sample.

[0113] 4. Loss function design:

[0114] To simultaneously ensure output current tracking accuracy and system efficiency, the loss function is designed to consist of two weighted components: the first is the load tracking error term, which is the squared deviation between the actual output current and the target load current requirement; the second is the system parasitic loss term, which is the power consumption of auxiliary components such as the fan. The total loss value is obtained by weighting and summing these two components, allowing the model to simultaneously learn and optimize tracking accuracy and energy consumption during training. The weighting coefficients can be adjusted according to actual needs; for example, a larger weight can be assigned to the first component when prioritizing tracking accuracy.

[0115] 5. Training and Deployment:

[0116] The convolutional neural network is trained using the backpropagation algorithm. After training, the model parameters are fixed, converted into C code, or a library file that can run on the target MCU is generated using a neural network compiler, and then deployed to the prediction compensation module 500. During training, an early stopping strategy can be used to prevent overfitting, and the model performance is evaluated by dividing the training set, validation set, and test set.

[0117] Example 3: Adaptive Control System

[0118] This embodiment provides a fuel cell adaptive control system based on multi-dimensional environmental perception, the structure of which is as follows: Figure 1 As shown, this corresponds to the method in Embodiment 1. The system includes:

[0119] Status acquisition module 100: integrates multiple communication interfaces (UART, I2C, SPI, CAN, ADC), connects to positioning module 10, temperature sensor 20, power management unit 30 and internal sensors of fuel cell stack, and is responsible for real-time acquisition of multi-dimensional status parameters.

[0120] Main channel control module 200: Internally stores a multi-dimensional performance mapping table, receives data from the status acquisition module 100, performs table lookup operations, and outputs a baseline collaborative control instruction set.

[0121] Feature construction module 400: includes a circular buffer for caching state data from the most recent N time steps; a built-in difference calculation unit for calculating first-order and second-order rates of change; and maps the data into a two-dimensional feature matrix through a data recombination unit.

[0122] Prediction compensation module 500: includes a pre-trained deep learning network (convolutional neural network in this embodiment), receives a two-dimensional feature matrix, and outputs dynamic compensation amount through forward inference.

[0123] Instruction synthesis module 600: Performs addition operations to superimpose the base instruction and the compensation amount, and performs amplitude limiting processing.

[0124] Multi-channel execution module 300: includes a multi-channel PWM generator, digital-to-analog converter, CAN controller, etc., and generates physical control signals according to the final instruction to drive the fan driver 40, back pressure valve actuator 50 and power converter 60.

[0125] The modules interact with each other via an internal bus or direct memory access to ensure real-time performance.

[0126] Example 4: Exception Handling and Degradation Strategies

[0127] This embodiment provides an exception handling mechanism to enhance the engineering robustness of the system.

[0128] When the prediction compensation module 500 detects abnormal input data (such as sensor failure or data exceeding the normal range) or the model output value exceeds the preset threshold, the system automatically enters the "compensation failure mode": the command synthesis module 600 sets the dynamic compensation amount to zero and uses only the reference command of the main channel for control. At the same time, the system records the fault log and reports it to the flight control unit through the communication interface.

[0129] When the status acquisition module 100 detects that the positioning module 10 has lost its signal, it can estimate the current altitude using inertial navigation data or atmospheric pressure sensor. If the estimation fails, it will revert to the preset conservative control mode (such as control according to the highest altitude condition) to ensure basic safety.

[0130] In addition, a watchdog timer can be set to monitor the operating status of the main control chip. If the program crashes or enters an infinite loop, the system will be automatically reset.

[0131] Example 5: A complete application example combining drone flight scenarios

[0132] The following is combined with Figure 1 and Figure 2 Taking a complete drone flight mission as an example, the specific implementation process of the present invention will be explained in detail.

[0133] A fuel cell drone equipped with the control system of this invention is parked at a take-off and landing site at an altitude of 500 meters, ready to carry out a reconnaissance mission over a mountain peak at an altitude of 4,000 meters.

[0134] Takeoff phase (altitude 500 meters, ambient temperature 20℃):

[0135] The status acquisition module 100 obtains GPS altitude of 500 meters and ambient temperature of 20°C, while the power management unit 30 requests an 80A takeoff current. The main channel control module 200 queries the multi-dimensional performance mapping table to obtain the baseline commands: baseline fan speed of 15000 rpm, baseline back pressure valve opening of 20%, and baseline maximum allowable output current threshold of 120A. The feature construction module 400 collects historical data from the last 10 moments to construct a two-dimensional feature matrix, and the prediction compensation module 500 outputs the compensation amount (assuming the current operating conditions are stable and the compensation amount is close to zero). The command synthesis module 600 outputs the final command unchanged. The multi-channel execution module 300 executes the control, and the UAV takes off smoothly.

[0136] Ascent Phase (Rapid Increase in Altitude):

[0137] As the drone ascends, its altitude continuously changes, and the status acquisition module 100 updates the data in real time. The main channel continuously adjusts the baseline commands based on the current altitude, temperature, and load current. Simultaneously, the feature construction module 400 continuously builds a feature matrix, and the prediction compensation module 500 detects a significant rate of altitude change and outputs appropriate compensation (e.g., a slight increase in fan speed compensation, such as +200 rpm, to anticipate a drop in oxygen partial pressure), making control smoother. The entire system operates stably.

[0138] Hovering reconnaissance phase (altitude 4000 meters, ambient temperature -10℃, load requirement 50A):

[0139] The main channel query mapping table yields: a reference turbine speed of 25,000 rpm, a reference back pressure valve opening of 60%, and a reference maximum permissible output current threshold of 70A. If the operator suddenly issues a high-maneuver command, the flight controller requests 90A of current. While the main channel's reference maximum permissible current threshold is 70A, the compensation channel, by analyzing historical sequences (such as the acceleration of current errors), predicts the impending power conflict and outputs compensation (e.g., further reducing the threshold or increasing the turbine speed in advance). The final synthesized maximum permissible output current threshold may still be around 70A, but the turbine speed has been increased in advance to prepare for a possible short-term overload. The multi-channel execution module 300 forces the power converter to limit the current to 70A and feeds back the current-limiting information to the flight controller. The flight controller then replans the maneuver power, thereby protecting the fuel cell stack from oxygen starvation damage.

[0140] As can be seen from the above embodiments, the present invention, by integrating deep learning compensation, significantly improves dynamic response capability and environmental adaptability while retaining the stability of the original mapping table control, and perfectly solves the core pain points of UAV operation in plateau and high-altitude environments.

[0141] The embodiments described above are merely illustrative of several implementation methods of this application, and while the descriptions are relatively specific and detailed, they should not be construed as limiting the scope of the invention patent. It should be noted that those skilled in the art can make various modifications and improvements without departing from the concept of this application, and these all fall within the protection scope of this application. Therefore, the protection scope of this patent application should be determined by the appended claims.

Claims

1. A fuel cell adaptive control method based on multi-dimensional environmental perception, characterized in that, Includes the following steps: The system acquires multi-dimensional state parameters, which include at least: the current altitude acquired by the positioning module, the ambient temperature acquired by the sensor, and the target load current requirement received from the power management unit. Query the multidimensional performance mapping table preset in the controller to obtain the corresponding baseline collaborative control instruction set; Multidimensional state parameters within a preset historical time window are collected to construct a state sequence, and the state sequence is differentially calculated to extract the first-order rate of change and the second-order rate of change. The original data, the first-order rate of change, and the second-order rate of change in the state sequence are arranged according to the time step and mapped into a two-dimensional feature matrix. The two-dimensional feature matrix is ​​input into a pre-trained deep learning network, which extracts dynamic trend features and outputs dynamic compensation amount. The reference coordinated control instruction set is superimposed with the dynamic compensation amount to obtain the final coordinated control instruction set, which includes at least: the final fan speed, the final cathode back pressure valve opening, and the final maximum allowable output current threshold. The operation of the wind turbine is controlled according to the final wind turbine speed, the operation of the back pressure valve is controlled according to the final cathode back pressure valve opening, and the final maximum allowable output current threshold is sent to the power converter or power management unit to limit the actual output current of the fuel cell stack from exceeding the threshold.

2. The method according to claim 1, characterized in that, The multidimensional state parameters also include system internal parameters, which include at least one or more of the following: temperature, voltage, and current of the fuel cell stack; the state sequence is composed of the altitude, ambient temperature, target load current requirement, and the system internal parameters.

3. The method according to claim 1 or 2, characterized in that, The process of constructing the two-dimensional feature matrix includes: The original values ​​of each parameter in the state sequence at each time step are taken as zero-order data; The difference between each parameter at adjacent time steps is calculated as the first-order rate of change; The difference between the first-order rate of change at adjacent time steps is calculated as the second-order rate of change; The zero-order data, the first-order rate of change, and the second-order rate of change are stacked in the vertical direction according to parameter type, and expanded in the horizontal direction according to time steps to generate the two-dimensional feature matrix.

4. The method according to claim 1 or 2, characterized in that, The deep learning network is a convolutional neural network, comprising an input layer, at least one convolutional layer, a pooling layer, and a fully connected layer connected in sequence. The convolutional layer is used to extract local dynamic texture features from the two-dimensional feature matrix; The fully connected layer is used to map the extracted features to the dynamic compensation values, which include: fan speed compensation value, back pressure valve opening compensation value, and maximum allowable output current threshold compensation value.

5. The method according to claim 4, characterized in that, The training objective function of the convolutional neural network includes a load tracking error term and a system parasitic loss term. The loss value is obtained by weighted summation of the load tracking error term and the system parasitic loss term, so as to optimize the current tracking accuracy and system energy consumption simultaneously during the training process.

6. The method according to claim 1, characterized in that, The multidimensional performance mapping table is obtained by calibration using experimental data or high-fidelity system simulation models. The calibration range covers the entire altitude range, ambient temperature range, and load current range of the expected flight of the UAV. The goal of constructing the mapping table is to minimize the total parasitic loss of the system while ensuring that the oxygen stoichiometry is within a safe range.

7. The method according to claim 1 or 6, characterized in that, The fan can be any one of the following: a blower, an air compressor, or an axial fan.

8. A fuel cell adaptive control system based on multi-dimensional environmental perception, characterized in that, include: The status acquisition module is configured to acquire multi-dimensional status parameters in real time. The multi-dimensional status parameters include at least: the current altitude acquired by the positioning module, the ambient temperature acquired by the temperature sensor, and the target load current requirement received from the power management unit. The main channel control module internally stores a multi-dimensional performance mapping table and is connected to the status acquisition module. It is used to receive the multi-dimensional status parameters, and use the altitude, ambient temperature and target load current requirements as input indexes to query the multi-dimensional performance mapping table and output a set of benchmark collaborative control instructions. The feature construction module, connected to the state acquisition module, is used to collect multi-dimensional state parameters within a preset historical time window to construct a state sequence, perform difference calculation on the state sequence to extract the first-order rate of change and the second-order rate of change, and arrange the original data, the first-order rate of change and the second-order rate of change in the state sequence according to the time step and map them into a two-dimensional feature matrix. The prediction compensation module internally stores a pre-trained deep learning network; The instruction synthesis module is connected to the main channel control module and the prediction compensation module respectively, and is used to superimpose the benchmark collaborative control instruction set with the dynamic compensation amount to obtain the final collaborative control instruction set. The multi-channel execution module is connected to the instruction synthesis module, the fan driver, the back pressure valve actuator, and the power converter, respectively. It is used to receive the final coordinated control instruction set, send a first control signal to the fan driver according to the final fan speed, and send a second control signal to the back pressure valve actuator according to the final cathode back pressure valve opening.

9. The system according to claim 8, characterized in that, The multidimensional state parameters also include system internal parameters, which include at least the temperature, voltage, and current of the fuel cell stack; the state sequence is composed of the altitude, ambient temperature, target load current requirement, and the system internal parameters.

10. The system according to claim 8 or 9, characterized in that, The deep learning network is a convolutional neural network, which includes an input layer, at least one convolutional layer, a pooling layer and a fully connected layer connected in sequence; the fully connected layer outputs the fan speed compensation value, the back pressure valve opening compensation value and the maximum allowable output current threshold compensation value.

Images