A fuel cell valve self-adaptive calibration method based on bayesian optimization

Abstract

The application discloses a fuel cell valve self-adaptive calibration method based on Bayesian optimization, and comprises the following steps: selecting a fuel cell valve action segment, and presetting a fuel cell operation safety boundary; selecting a matched kernel distance calculation path according to an opening degree change direction; performing time offset correction on a covariance calculation sequence of a state transition kernel; performing gate regulation according to a calibration working condition; recursively selecting a candidate valve action, and constraining the safety boundary; rewriting a collection function to select a next round of valve calibration action; and correcting the state transition kernel by using an execution feedback, and updating a calibration atlas. The application relates to the technical field of fuel cell system control, realizes safe and efficient calibration of a fuel cell valve, and improves working condition adaptability and control precision.

Description

Technical Field

[0001] This invention relates to the field of fuel cell system control technology, and in particular to an adaptive calibration method for fuel cell valves based on Bayesian optimization. Background Technology

[0002] The operation of a fuel cell system relies on actuators such as hydrogen inlet valves, air regulating valves, back pressure valves, and hydrogen exhaust valves to complete gas supply, exhaust, and pressure regulation. Current valve calibration is mostly based on bench tests to create calibration tables, which are then used by the controller to call fixed valve opening, duty cycle, and action time parameters according to the load range and pressure target. Some control schemes introduce PID control, model predictive control, or adaptive correction methods to reduce pressure and flow deviations and improve the stability of fuel cell gas supply.

[0003] Fuel cell valves exhibit issues such as inconsistent opening and closing responses, valve lag, and abrupt response changes caused by operating condition switching during actual operation. Fixed calibration tables struggle to adapt to valve aging, variations in pipeline resistance, hydrogen and water discharge disturbances, and rapid load changes. Single-path state prediction models tend to mix opening and closing responses in the same calculation process, leading to calibration results that deviate from the actual valve response.

[0004] Existing Bayesian optimization methods can be used for finding optimal parameters with few samples, but when directly applied to fuel cell valve calibration, candidate parameters tend to approach safety boundaries such as pressure, flow rate, and voltage fluctuations. Current methods lack modifications to the state transition kernel for valve action direction, action lag, and operating condition switching, and also lack acquisition functions that combine future recursive results with safety boundary constraints, thus limiting calibration efficiency, safety, and online correction capabilities.

[0005] Therefore, how to provide an adaptive calibration method for fuel cell valves based on Bayesian optimization is a problem that urgently needs to be solved by those skilled in the art. Summary of the Invention

[0006] One objective of this invention is to propose an adaptive calibration method for fuel cell valves based on Bayesian optimization. This invention utilizes a Gaussian process state-space model, state transition kernel modification, and safety boundary fitting acquisition function to refine the calibration process under valve opening and closing direction, action lag, and operating condition switching. It has the advantages of high calibration safety, fewer test runs, and strong adaptability to operating conditions.

[0007] An adaptive calibration method for fuel cell valves based on Bayesian optimization according to an embodiment of the present invention includes the following steps: S1. Select the valve action segment of the fuel cell according to the calibration conditions and preset the safe operating boundary of the fuel cell; S2. Set up open kernel distance calculation paths and closed kernel distance calculation paths in the state transition kernel of the Gaussian process state space model, and select matching kernel distance calculation paths according to the opening change direction of the fuel cell valve action segment. S3. Perform time offset correction on the covariance calculation order of the state transition kernel according to the valve action lag in the fuel cell valve action segment, so that the valve response before and after the lag participates in the same state transition kernel calculation. S4. Set up a working condition switching gating in the state transition kernel, and adjust the gating of the matching kernel distance calculation path and the covariance calculation order corrected by time offset according to the calibrated working condition. S5. The state transition kernel, after being adjusted by the operating condition switching gating, is used to recursively calculate the future calibration cycle of the candidate valve action input formed by the fuel cell valve action segment, and the result of the future calibration cycle recursion is limited to the preset fuel cell operating safety boundary. S6. Rewrite the Bayesian optimization acquisition function into an acquisition function with safety boundary fitting constraints, and select the next round of valve calibration action according to the rule of being close to the preset fuel cell operating safety boundary without touching the preset fuel cell operating safety boundary. S7. Execute the next round of valve calibration actions to obtain execution feedback, synchronously correct the state transition kernel based on the execution feedback, and update the fuel cell valve calibration chart according to the calibration conditions.

[0008] Optionally, the step of selecting the fuel cell valve action segment according to the calibration conditions in S1 includes: The historical calibration process of fuel cell valves is segmented according to the calibration conditions. The continuous segment between the command triggering and the valve response entering the allowable range under the corresponding calibration conditions is taken as the fuel cell valve action segment. Read the allowable pressure range, allowable flow range, and allowable voltage fluctuation range under the calibrated operating conditions corresponding to the valve action segment of the fuel cell, and use the allowable pressure range, allowable flow range, and allowable voltage fluctuation range as the boundary composition of the preset fuel cell operating safety boundary.

[0009] Optionally, the step of setting the open-ended kernel distance calculation path and the closed-ended kernel distance calculation path in S2 includes: Before performing kernel distance calculation in the state transition kernel, the open kernel distance calculation path and the closed kernel distance calculation path are divided; Based on the direction of the opening change of the fuel cell valve action segment, a matching core distance calculation path is selected from the opening core distance calculation path and the closing core distance calculation path; The fuel cell valve action segment with increased opening degree is sent to the open core distance calculation path, and the fuel cell valve action segment with decreased opening degree is sent to the closed core distance calculation path. Within the selected nuclear distance calculation path, only fuel cell valve action segments with the same opening change direction are called to participate in the nuclear distance calculation; The unselected kernel distance calculation path is excluded from the current state transition kernel calculation; The kernel distance calculation results of the selected kernel distance calculation path are sent to the time offset corrected covariance calculation sequence, so that the state transition kernel performs state recursion of candidate valve actions based on the kernel distance calculation results.

[0010] Optionally, the time offset correction step in S3 includes: Using the command trigger sampling position in the fuel cell valve action segment as the starting point for covariance calculation, the valve response sampling positions are adjusted from being paired at the same sampling time to being paired at a delayed sampling time according to the valve action lag. Within the state transition kernel, the accumulation of covariance for sampling positions that do not produce a valve response is cancelled, and the first sampling position that produces a valve response is included in the covariance calculation order. Continue to access subsequent valve response sampling positions along the sampling time sequence, so that the instruction trigger sampling position and the delayed valve response sampling position participate in the kernel calculation in the same state recursion process; The time-offset corrected covariance calculation order receives the kernel distance calculation results output by the open kernel distance calculation path and the closed kernel distance calculation path, thus completing the time offset correction of the covariance calculation order of the state transition kernel.

[0011] Optionally, the gating adjustment step in S4 includes: Before the time-off-corrected covariance calculation sequence enters the state transition kernel calculation, the operating condition switching gating determines the starting point and ending point of the covariance calculation according to the current calibrated operating condition. When the future calibration cycle corresponding to the candidate valve action remains within the same calibration condition, the covariance calculation order corresponding to the current calibration condition is maintained, and the kernel distance calculation results output by the selected kernel distance calculation path are continuously connected. When the future calibration cycle corresponding to the candidate valve action crosses the calibration condition switching position, the covariance calculation corresponding to the calibration condition before the switching is terminated by taking the calibration condition switching position as the covariance breakpoint. Restart the covariance calculation from the first valve response sampling position after the switch, and select the matching kernel distance calculation path from the opening kernel distance calculation path and the closing kernel distance calculation path according to the opening change direction corresponding to the candidate valve action after the switch. The covariance calculation results corresponding to the calibration conditions before the switchover and the covariance calculation results corresponding to the calibration conditions after the switchover are entered into the state transition kernel in the order of the calibration conditions, so as to complete the gating adjustment of the covariance calculation order after time offset correction.

[0012] Optionally, the step of recursively calculating the future calibration period in S5 includes: The candidate valve action is input into the state transition kernel after the working condition switching gating adjustment, and the selected kernel distance calculation path calls the covariance calculation order corrected by time offset; The state recursion is executed step by step according to the sampling order of the future calibration cycle. After each state recursion is completed, the state transition kernel outputs the recursion uncertainty of the corresponding sampling position and makes a judgment on the recursion result of the future calibration cycle with the preset fuel cell operation safety boundary. If the future calibration cycle recursion result does not enter the inner contact range of the preset fuel cell operating safety boundary, the recursion direction of the current candidate valve action will be maintained. When the future calibration cycle recursion result enters the inner contact range of the preset fuel cell operating safety boundary but does not reach the preset fuel cell operating safety boundary, the adjustment range of the current candidate valve action in the next state recursion will be reduced. When the future calibration cycle recursion results reach the preset fuel cell operating safety boundary, the subsequent recursion of the current candidate valve action will be stopped.

[0013] Optionally, the step of selecting the next round of valve calibration action in S6 includes: The future calibration cycle recursion result corresponding to the candidate valve action is used as the calculation object of the Bayesian optimization acquisition function; Before the acquisition function is calculated, the actions of candidate valves that reach the preset fuel cell operating safety boundary are blocked; For candidate valve actions that do not reach the preset fuel cell operating safety boundary, the collected function values ​​are updated in the order of decreasing distance between the future calibration cycle recursive results and the preset fuel cell operating safety boundary. When the future calibration cycle recursion result will reach the preset fuel cell operation safety boundary in the next sampling step, the acquisition function value will be rolled back to the acquisition function value of the previous sampling step. For candidate valve actions with the same acquired function value, they are reordered in order of increasing recursive uncertainty. The candidate valve action ranked first is determined as the valve calibration action for the next round.

[0014] Optionally, in step S7, the state transition kernel is synchronously corrected based on execution feedback, and the fuel cell valve calibration chart is updated according to the calibration conditions. Specifically, this includes: Align the execution feedback of the next round of valve calibration actions with the corresponding future calibration cycle recursion results according to the sampling order; When the direction of the opening change in the execution feedback is inconsistent with the selected kernel distance calculation path, the corresponding sampling position is migrated to the kernel distance calculation path that matches the opening kernel distance calculation path and the closing kernel distance calculation path. When the sampling position of the valve response in the execution feedback is offset from the corresponding sampling position of the valve response in the future calibration cycle recursion result, the covariance calculation order after time offset correction is reset according to the sampling position offset amount; When the sampling sequence corresponding to the execution feedback crosses the calibration condition switching position, the covariance breakpoint is re-determined, and the covariance calculation is restarted from the first valve response sampling position after the switch. The corrected nuclear distance calculation path, the time-off corrected covariance calculation order, and the operating condition switching gating are written into the fuel cell valve calibration chart under the corresponding calibration conditions.

[0015] The beneficial effects of this invention are: (1) The present invention sends the valve action segments of the fuel cell into the opening kernel distance calculation path and the closing kernel distance calculation path according to the opening change direction, so as to avoid the opening action and the closing action being mixed in the same state transition kernel calculation process, thereby improving the accuracy of valve response prediction and calibration parameter selection.

[0016] (2) The present invention corrects the time offset of the covariance calculation order according to the valve action lag, and through the working condition switching gating process, the cross-working condition recursion process is processed so that the actual response after the valve action can participate in the same core calculation, thereby reducing the calibration deviation caused by load change, hydrogen discharge and drainage, and back pressure change.

[0017] (3) The present invention rewrites the Bayesian optimization acquisition function into an acquisition function with safety boundary fitting constraints, selects the next round of valve calibration action based on the future calibration cycle recursion results, completes the optimization under the condition of being close to the safety boundary and not touching the safety boundary, reduces the number of invalid tests, and improves calibration safety and working condition adaptability. Attached Figure Description

[0018] The accompanying drawings are provided to further illustrate the invention and form part of the specification. They are used in conjunction with embodiments of the invention to explain the invention and do not constitute a limitation thereof. In the drawings: Figure 1 This is a flowchart of an adaptive calibration method for fuel cell valves based on Bayesian optimization proposed in this invention; Figure 2 This is a schematic diagram of the state transition core modification and operating condition switching gating structure proposed in this invention; Figure 3 This is a schematic diagram of the safe boundary fitting Bayesian optimization and feedback update structure proposed in this invention. Detailed Implementation

[0019] The present invention will now be described in further detail with reference to the accompanying drawings. These drawings are simplified schematic diagrams, illustrating only the basic structure of the invention, and therefore only show the components relevant to the invention.

[0020] refer to Figures 1-3 An adaptive calibration method for fuel cell valves based on Bayesian optimization includes the following steps: S1. Select the valve action segment of the fuel cell according to the calibration conditions and preset the safe operating boundary of the fuel cell; S2. Set up open kernel distance calculation paths and closed kernel distance calculation paths in the state transition kernel of the Gaussian process state space model, and select matching kernel distance calculation paths according to the opening change direction of the fuel cell valve action segment. S3. Perform time offset correction on the covariance calculation order of the state transition kernel according to the valve action lag in the fuel cell valve action segment, so that the valve response before and after the lag participates in the same state transition kernel calculation. S4. Set up a working condition switching gating in the state transition kernel, and adjust the gating of the matching kernel distance calculation path and the covariance calculation order corrected by time offset according to the calibrated working condition. S5. The state transition kernel, after being adjusted by the operating condition switching gating, is used to recursively calculate the future calibration cycle of the candidate valve action input formed by the fuel cell valve action segment, and the result of the future calibration cycle recursion is limited to the preset fuel cell operating safety boundary. S6. Rewrite the Bayesian optimization acquisition function into an acquisition function with safety boundary fitting constraints, and select the next round of valve calibration action according to the rule of being close to the preset fuel cell operating safety boundary without touching the preset fuel cell operating safety boundary. S7. Execute the next round of valve calibration actions to obtain execution feedback, synchronously correct the state transition kernel based on the execution feedback, and update the fuel cell valve calibration chart according to the calibration conditions.

[0021] In this embodiment, the step of selecting the fuel cell valve action segment according to the calibration condition in S1 includes: The historical calibration process of fuel cell valves is segmented according to the calibration conditions. The continuous segment between the command triggering and the valve response entering the allowable range under the corresponding calibration conditions is taken as the fuel cell valve action segment. In this invention, the calibration conditions are divided according to the operating state of the fuel cell. Each calibration condition corresponds to a set of valve control targets and safety constraints. The calibration conditions can be distinguished according to the load change state, gas supply regulation state, back pressure regulation state, and hydrogen discharge and drainage state, which are used to limit the segment boundaries of the historical calibration process of fuel cell valves and the reading range of the preset fuel cell operating safety boundary. The historical calibration process of fuel cell valves is a record of valve actions generated during the completion of bench calibration, online test calibration, or operation playback of fuel cells. The historical calibration process of fuel cell valves is arranged in the order of sampling time. Each sampling time corresponds to the valve control command and the fuel cell operation response, which serves as the data source for extracting fuel cell valve action segments. When segmenting the historical calibration process of fuel cell valves, the start and end positions of the calibration conditions are first identified according to the calibration conditions, and then the valve command trigger positions are identified within the same calibration conditions. If there is a switch in the operating conditions, a protection shutdown, or a sampling interruption between two adjacent valve command trigger positions, the current segment is stopped and the next segment identification process is restarted. The same valve action is determined according to the continuous control behavior of the same target valve after a single command trigger; when the same target valve exhibits a continuous process of opening adjustment, holding, and retraction after a single command trigger, it is classified as the same valve action; when different target valves act synchronously within the same sampling window, the segment is assigned according to the control relationship between the master control valve and the follow-up valve. The command trigger is determined by the sampling position when the valve control command changes from the holding state to the adjustment state; when the valve control command changes any of the following between adjacent sampling times: opening degree change, duty cycle change, opening command, or closing command, the sampling position at the beginning of the change is taken as the command trigger position. The valve response entering the allowable range under the corresponding calibration condition is determined according to the regression state of the fuel cell operating response; after the valve is activated, if the pressure response, flow response, and voltage fluctuation enter the allowable range under the corresponding calibration condition and do not leave the allowable range within several consecutive sampling periods, the first sampling position that enters the allowable range is taken as the end point of the fuel cell valve action segment. Continuous segments are determined based on the continuity of sampling timing and the integrity of actions; fuel cell valve action segments start from the command trigger position and end when the valve response enters the allowable range under the corresponding calibration condition; intermediate sampling points maintain temporal continuity, and segments do not cross new calibration conditions, protection shutdown states, or new independent valve actions. Read the allowable pressure range, allowable flow range, and allowable voltage fluctuation range under the calibrated operating conditions corresponding to the valve action segment of the fuel cell, and use the allowable pressure range, allowable flow range, and allowable voltage fluctuation range as the boundary composition of the preset fuel cell operating safety boundary.

[0022] In this invention, the allowable pressure range, allowable flow rate range, and allowable voltage fluctuation range are read according to the fuel cell safety operation constraints under the corresponding calibrated operating conditions. The allowable pressure range is used to define the boundaries of the anode-side pressure, cathode-side pressure, and pressure difference; the allowable flow rate range is used to define the boundaries of the hydrogen supply and air supply; and the allowable voltage fluctuation range is used to define the boundaries of the stack output disturbance caused by valve operation.

[0023] The preset safety boundary for fuel cell operation consists of the allowable pressure range, the allowable flow rate range, and the allowable voltage fluctuation range. During subsequent state recursion, the recursive results corresponding to candidate valve actions are compared with the aforementioned boundary components; the safety boundary fitting constraints in the Bayesian optimization acquisition function use the preset fuel cell operation safety boundary as the constraint object.

[0024] In this embodiment, the step of setting the open kernel distance calculation path and the closed kernel distance calculation path in S2 includes: Before performing kernel distance calculation in the state transition kernel, the open kernel distance calculation path and the closed kernel distance calculation path are divided; Based on the direction of the opening change of the fuel cell valve action segment, a matching core distance calculation path is selected from the opening core distance calculation path and the closing core distance calculation path; The fuel cell valve action segment with increased opening degree is sent to the open core distance calculation path, and the fuel cell valve action segment with decreased opening degree is sent to the closed core distance calculation path. Within the selected nuclear distance calculation path, only fuel cell valve action segments with the same opening change direction are called to participate in the nuclear distance calculation; The unselected kernel distance calculation path is excluded from the current state transition kernel calculation; The kernel distance calculation results of the selected kernel distance calculation path are sent to the time offset corrected covariance calculation sequence, so that the state transition kernel performs state recursion of candidate valve actions based on the kernel distance calculation results.

[0025] In a specific embodiment of the present invention, the state transition kernel, as the kernel calculation structure for state recursion in the Gaussian process state-space model, is used to determine the influence weight of candidate valve actions on future state recursion based on the similarity between valve action segments in the fuel cell. Instead of using a single path to calculate the kernel distance between all valve action segments, the state transition kernel sets up an opening-direction kernel distance calculation path and a closing-direction kernel distance calculation path at the kernel calculation entry point, allowing fuel cell valve action segments with different opening direction changes to enter different calculation paths.

[0026] The Gaussian process state-space model receives the current recursive state and candidate valve actions at each sampling location. It then uses a state transition kernel to calculate the state mean and prediction variance for the next sampling location. The state mean serves as the recursive state for the next sampling location, and the prediction variance serves as the recursive uncertainty for that location. This process is repeated in future calibration cycles according to the sampling order until the end of the future calibration cycle is reached or the preset fuel cell operating safety boundary is met.

[0027] Before calculating the nuclear distance, the opening change of the same valve at adjacent sampling times is read in the fuel cell valve action segment. If the valve opening at the later sampling time is greater than the valve opening at the previous sampling time, the corresponding fuel cell valve action segment is determined to be an opening increase; if the valve opening at the later sampling time is less than the valve opening at the previous sampling time, the corresponding fuel cell valve action segment is determined to be an opening decrease. Fuel cell valve action segments with increasing openings enter the opening-direction nuclear distance calculation path, and fuel cell valve action segments with decreasing openings enter the closing-direction nuclear distance calculation path.

[0028] In the open-ended core distance calculation path, only the fuel cell valve action segment with increasing opening degree is called to participate in the core distance calculation; in the closed-ended core distance calculation path, only the fuel cell valve action segment with decreasing opening degree is called to participate in the core distance calculation. When a candidate valve action enters the state transition kernel, the selected core distance calculation path is first determined according to the opening degree change direction corresponding to the candidate valve action, and then the core distance calculation is completed within the selected core distance calculation path.

[0029] Unselected kernel distance calculation paths are not included in the current state transition kernel calculation. When a candidate valve action with an increasing opening degree enters the opening-direction kernel distance calculation path, the closing-direction kernel distance calculation path is stopped being called; when a candidate valve action with a decreasing opening degree enters the closing-direction kernel distance calculation path, the opening-direction kernel distance calculation path is stopped being called. Through the above path exclusion method, the state transition kernel only compares fuel cell valve action segments under the same opening degree change direction in a single kernel calculation.

[0030] The time-off-corrected covariance calculation sequence is used to receive the kernel distance calculation results output by the selected kernel distance calculation path. After the kernel distance calculation results enter the covariance calculation sequence, they are aligned with the time offset position corresponding to the valve action lag, ensuring that the response change after the valve action participates in the same covariance calculation. After completing the covariance calculation, the state transition kernel sends the calculation results into the subsequent state recursion process of the candidate valve action to predict the recursive state of the candidate valve action in future calibration periods.

[0031] In this embodiment, the step of time offset correction in S3 includes: Using the command trigger sampling position in the fuel cell valve action segment as the starting point for covariance calculation, the valve response sampling positions are adjusted from being paired at the same sampling time to being paired at a delayed sampling time according to the valve action lag. Within the state transition kernel, the accumulation of covariance for sampling positions that do not produce a valve response is cancelled, and the first sampling position that produces a valve response is included in the covariance calculation order. Continue to access subsequent valve response sampling positions along the sampling time sequence, so that the instruction trigger sampling position and the delayed valve response sampling position participate in the kernel calculation in the same state recursion process; The time-offset corrected covariance calculation order receives the kernel distance calculation results output by the open kernel distance calculation path and the closed kernel distance calculation path, thus completing the time offset correction of the covariance calculation order of the state transition kernel.

[0032] In a specific embodiment of the present invention, the command trigger sampling position is determined according to the sampling position where the valve control command changes within the fuel cell valve action segment. When the valve control command changes from the holding state to the opening adjustment state, the opening adjustment start position is recorded as the command trigger sampling position; when the valve control command changes from the open state to the closed state, the closing command start position is recorded as the command trigger sampling position.

[0033] The valve response sampling position is determined according to the sampling position where the operating response begins to deviate from the reference state before the action within the fuel cell valve action segment. When the operating response deviates from the reference state before the action in the same direction of change within a continuous sampling period, the starting position of the continuous deviation is recorded as the first sampling position that generates a valve response, and the sampling positions after the first sampling position that generates a valve response are successively used as the sampling positions of subsequent valve responses.

[0034] The valve action lag is determined by the sampling interval between the command-triggered sampling position and the first sampling position that generates a valve response. Before performing covariance calculation, the state transition kernel first reads the valve action lag, and then adjusts the pairing relationship between the command-triggered sampling position and the valve response sampling position according to the valve action lag.

[0035] Before time offset correction, the state transition kernel calculates the covariance between the command trigger sampling position and the valve response sampling position at the same sampling time using a pairing method based on the same sampling time. When performing time offset correction, the pairing based on the same sampling time is changed to pairing based on the delayed sampling time, so that a covariance calculation relationship is established between the command trigger sampling position and the first sampling position that generates a valve response.

[0036] Sampling positions that do not produce a valve response are not included in the covariance accumulation. The state transition kernel stops covariance accumulation between the instruction-triggered sampling position and the first sampling position that produces a valve response. It then starts the covariance calculation sequence from the first sampling position that produces a valve response and continues to the subsequent valve response sampling positions along the sampling time sequence.

[0037] The time-offset corrected covariance calculation order is established according to the order of instruction trigger sampling position, the first sampling position that generates a valve response, and the sampling positions of subsequent valve responses. After the kernel distance calculation results output by the open-ended kernel distance calculation path or the closed-ended kernel distance calculation path enter the time-offset corrected covariance calculation order, the state transition kernel calculates the state recursion relationship corresponding to the candidate valve action according to the lagging valve response sampling position.

[0038] In this embodiment, the gating adjustment step in S4 includes: Before the time-off-corrected covariance calculation sequence enters the state transition kernel calculation, the operating condition switching gating determines the starting point and ending point of the covariance calculation according to the current calibrated operating condition. When the future calibration cycle corresponding to the candidate valve action remains within the same calibration condition, the covariance calculation order corresponding to the current calibration condition is maintained, and the kernel distance calculation results output by the selected kernel distance calculation path are continuously connected. When the future calibration cycle corresponding to the candidate valve action crosses the calibration condition switching position, the covariance calculation corresponding to the calibration condition before the switching is terminated by taking the calibration condition switching position as the covariance breakpoint. Restart the covariance calculation from the first valve response sampling position after the switch, and select the matching kernel distance calculation path from the opening kernel distance calculation path and the closing kernel distance calculation path according to the opening change direction corresponding to the candidate valve action after the switch. The covariance calculation results corresponding to the calibration conditions before the switchover and the covariance calculation results corresponding to the calibration conditions after the switchover are entered into the state transition kernel in the order of the calibration conditions, so as to complete the gating adjustment of the covariance calculation order after time offset correction.

[0039] In a specific embodiment of the present invention, the operating condition switching gating is set before the time-offset corrected covariance calculation sequence enters the state transition kernel calculation. The operating condition switching gating reads the current calibration operating condition, determines the starting point and ending point of the covariance calculation under the current calibration operating condition, and restricts the covariance calculation to the sampling range corresponding to the current calibration operating condition.

[0040] When the future calibration cycle is kept within the same calibration condition, the condition switching gating does not set a covariance breakpoint. The covariance calculation order after time offset correction remains continuous. The core distance calculation results output by the open core distance calculation path or the closed core distance calculation path are continuously connected to the state transfer core according to the sampling time sequence.

[0041] When the calibration cycle is recursively pushed across the calibration condition switching point, the condition switching gating will use the calibration condition switching point as the covariance breakpoint. The sampling positions before the covariance breakpoint will be included in the covariance calculation corresponding to the calibration condition before the switch, and the sampling positions after the covariance breakpoint will be included in the covariance calculation corresponding to the calibration condition after the switch.

[0042] At the covariance breakpoint, the covariance calculation corresponding to the calibration condition before the switch is terminated by the condition switching gating, and the sampling positions before and after the switch are not merged into the same covariance calculation chain. The first valve response sampling position after the switch is used as the new starting point for covariance calculation, and the subsequent valve response sampling positions corresponding to the calibration condition after the switch are sequentially connected to the new covariance calculation sequence.

[0043] When restarting covariance calculation after a switch, the operating condition switching gating selects a matching core distance calculation path from the open and closed core distance calculation paths according to the opening change direction corresponding to the candidate valve action after the switch. The core distance calculation result output by the selected core distance calculation path is then used in the covariance calculation sequence corresponding to the calibrated operating condition after the switch.

[0044] The covariance calculation results for the calibration conditions before and after the switch are entered into the state transition kernel in the order of the calibration conditions. The state transition kernel completes the cross-condition kernel calculation in the order of before and after the switch to avoid the valve response under different calibration conditions being merged into the same covariance calculation process.

[0045] In this embodiment, the step of recursively calculating the future calibration period in S5 includes: The candidate valve action is input into the state transition kernel after the working condition switching gating adjustment, and the selected kernel distance calculation path calls the covariance calculation order corrected by time offset; The state recursion is executed step by step according to the sampling order of the future calibration cycle. After each state recursion is completed, the state transition kernel outputs the recursion uncertainty of the corresponding sampling position and makes a judgment on the recursion result of the future calibration cycle with the preset fuel cell operation safety boundary. If the future calibration cycle recursion result does not enter the inner contact range of the preset fuel cell operating safety boundary, the recursion direction of the current candidate valve action will be maintained. When the future calibration cycle recursion result enters the inner contact range of the preset fuel cell operating safety boundary but does not reach the preset fuel cell operating safety boundary, the adjustment range of the current candidate valve action in the next state recursion will be reduced. When the future calibration cycle recursion results reach the preset fuel cell operating safety boundary, the subsequent recursion of the current candidate valve action will be stopped.

[0046] In a specific embodiment of the present invention, the state transition kernel, after being adjusted by the operating condition switching gating, is used to perform future calibration cycle recursion for candidate valve actions. After the operating condition switching gating completes the extraction of the covariance calculation sequence corresponding to the calibration operating condition, the state transition kernel only retains the portion of the current recursion interval that is allowed to participate in the kernel calculation; the covariance calculation segments not enabled by the current calibration operating condition are not included in this recursion.

[0047] When the selected kernel distance calculation path calls the time-offset corrected covariance calculation sequence, it first reads the kernel distance calculation results output by the open-direction or closed-direction kernel distance calculation path, and then writes the kernel distance calculation results into the time-offset corrected covariance calculation sequence. The state transition kernel completes the state recursion of the candidate valve action at the current sampling position according to the kernel distance calculation results and the covariance calculation sequence.

[0048] The sampling order for future calibration cycles is determined according to the recursive timing after the candidate valve action enters the state transition kernel. Each sampling position corresponds to one state recursion, and the state recursion result of the previous sampling position is used as the input for the next sampling position, so that the candidate valve action forms a continuous recursive process in future calibration cycles.

[0049] The future calibration cycle recursive result is the sequence of operating states obtained by progressively recursively calculating the candidate valve actions within the future calibration cycle using a state transition kernel. This future calibration cycle recursive result is used to determine its fit with the preset fuel cell operating safety boundary and also for subsequent Bayesian optimization of the acquisition function calculation.

[0050] The recursive uncertainty is determined by the prediction variance output by the Gaussian process state-space model at each sampling location. After the state transition kernel completes one state recursion, it synchronously outputs the prediction variance of the corresponding sampling location; the prediction variances of each sampling location within future calibration periods are combined to form the recursive uncertainty according to the sampling order. When ranking candidate valve actions, the Bayesian optimized acquisition function uses the recursive uncertainty to distinguish candidate valve actions with different prediction confidence levels.

[0051] The fit determination is based on the proximity between the future calibration cycle result and the preset fuel cell operating safety boundary. The boundary fit interval is determined after reserving a safety margin inward from the preset fuel cell operating safety boundary, with corresponding safety margins set for the allowable pressure range, allowable flow range, and allowable voltage fluctuation range. If the future calibration cycle result is within the preset fuel cell operating safety boundary and does not enter the inner interval corresponding to any safety margin, it is determined that the fit has not been reached; if the future calibration cycle result enters the inner interval corresponding to any safety margin and does not cross the preset fuel cell operating safety boundary, it is determined that the fit has been reached; if the future calibration cycle result reaches or crosses the preset fuel cell operating safety boundary, it is determined that the preset fuel cell operating safety boundary has been reached.

[0052] When reducing the adjustment range of the current candidate valve action in the next state recursion, the valve adjustment range corresponding to the next sampling position is rolled back towards the hold state, based on the original adjustment range of the current candidate valve action. After rolling back, the next state recursion is executed again, so that the future calibration cycle recursion results continue to be within the preset fuel cell operating safety boundary.

[0053] When the recursive results of future calibration cycles reach the preset safety boundary for fuel cell operation, the state transition kernel stops the state recursion of the current candidate valve action at the remaining sampling positions. After stopping, no further recursive results are generated for the current candidate valve action to prevent the recursive state after reaching the boundary from continuing to participate in subsequent calibration selection.

[0054] In this embodiment, the step of selecting the next round of valve calibration action in S6 includes: The future calibration cycle recursion result corresponding to the candidate valve action is used as the calculation object of the Bayesian optimization acquisition function; Before the acquisition function is calculated, the actions of candidate valves that reach the preset fuel cell operating safety boundary are blocked; For candidate valve actions that do not reach the preset fuel cell operating safety boundary, the collected function values ​​are updated in the order of decreasing distance between the future calibration cycle recursive results and the preset fuel cell operating safety boundary. When the future calibration cycle recursion result will reach the preset fuel cell operation safety boundary in the next sampling step, the acquisition function value will be rolled back to the acquisition function value of the previous sampling step. For candidate valve actions with the same acquired function value, they are reordered in order of increasing recursive uncertainty. The candidate valve action ranked first is determined as the valve calibration action for the next round.

[0055] In a specific embodiment of the present invention, the acquisition function containing safety boundary fitting constraints is rewritten from a regular Bayesian optimized acquisition function. After rewriting, the acquisition function does not directly use the candidate valve action itself as the calculation object, but uses the future calibration cycle recursion result corresponding to the candidate valve action as the calculation object, so that the acquisition function selects the next round of valve calibration action according to the boundary fitting state in the future recursion process.

[0056] Before calculating the acquisition function, the future calibration cycle recursion result obtained in the future calibration cycle recursion step is read first. Candidate valve actions that have reached the preset fuel cell operating safety boundary according to the future calibration cycle recursion result are shielded, and shielded candidate valve actions are not included in the acquisition function value calculation process.

[0057] For candidate valve actions that do not reach the preset fuel cell operating safety boundary, the data acquisition function is updated in the order of decreasing distance between the recursive result of future calibration cycles and the preset fuel cell operating safety boundary. The data acquisition function value corresponding to candidate valve actions with larger distances decreases, while the data acquisition function value corresponding to candidate valve actions that are within the boundary-fitting range but still have not reached the preset fuel cell operating safety boundary increases.

[0058] During the update of the acquisition function value, the boundary distance is calculated sequentially according to each sampling position within the future calibration period. If the recursive result of the future calibration period corresponding to the next sampling step will reach the preset fuel cell operating safety boundary, the acquisition function value corresponding to the next sampling step will be stopped, and the acquisition function value will be rolled back to the acquisition function value of the previous sampling step.

[0059] After the acquisition function value is rolled back, the candidate valve action remains in the boundary-fitting state corresponding to the previous sampling step and participates in the sorting. The acquired function value after rollback will not continue to be updated in the next sampling step to avoid reaching the recursive state of the preset fuel cell operating safety boundary and entering the acquisition function sorting.

[0060] After the acquisition function value rollback is completed, the acquisition function value, recursive uncertainty, and boundary contact state corresponding to the previous sampling step are retained as the sorting criteria for candidate valve actions; the acquisition function value formed when the preset fuel cell operating safety boundary is reached in the next sampling step is not included in the sorting process. The subsequent sorting of candidate valve actions is only based on the rolled-back acquisition function value.

[0061] When two or more candidate valve actions have the same data acquisition function value, they are reordered according to the order of recursive uncertainty from smallest to largest. The recursive uncertainty is output by the Gaussian process state-space model during the recursive process of future calibration cycles and is used to characterize the reliability of the recursive state corresponding to the candidate valve action.

[0062] After sorting, the candidate valve action at the top of the list is determined as the next round of valve calibration action. The next round of valve calibration action then proceeds to the subsequent execution process, and the execution feedback is used to correct the state transition kernel and update the fuel cell valve calibration map.

[0063] In this embodiment, step S7, which involves synchronously correcting the state transition kernel based on execution feedback and updating the fuel cell valve calibration chart according to the calibration conditions, specifically includes: Align the execution feedback of the next round of valve calibration actions with the corresponding future calibration cycle recursion results according to the sampling order; When the direction of the opening change in the execution feedback is inconsistent with the selected kernel distance calculation path, the corresponding sampling position is migrated to the kernel distance calculation path that matches the opening kernel distance calculation path and the closing kernel distance calculation path. When the sampling position of the valve response in the execution feedback is offset from the corresponding sampling position of the valve response in the future calibration cycle recursion result, the covariance calculation order after time offset correction is reset according to the sampling position offset amount; When the sampling sequence corresponding to the execution feedback crosses the calibration condition switching position, the covariance breakpoint is re-determined, and the covariance calculation is restarted from the first valve response sampling position after the switch. The corrected nuclear distance calculation path, the time-off corrected covariance calculation order, and the operating condition switching gating are written into the fuel cell valve calibration chart under the corresponding calibration conditions.

[0064] In a specific embodiment of the present invention, the execution feedback is the actual response record obtained after the next round of valve calibration is completed, which is used to verify the recursive results of future calibration cycles. Before the execution feedback enters the state transition kernel correction process, a correspondence is established with the recursive results of future calibration cycles according to the sampling order, so that the actual valve response and the recursive valve response at the same sampling position are at the same comparison benchmark.

[0065] When aligning according to the sampling order, the execution start point of the next round of valve calibration is used as the alignment start point, and the first recursive sampling position of the future calibration cycle recursive result is used as the prediction start point. The sampling positions in the execution feedback are sequentially matched with the corresponding sampling positions in the future calibration cycle recursive results; when the sampling intervals are inconsistent, the future calibration cycle recursive result with the closest time distance to the execution feedback sampling position is selected as the corresponding result.

[0066] After alignment, the direction of the opening change in the execution feedback is read and compared with the kernel distance calculation path selected in this state recursion. If the direction of the opening change in the execution feedback is an increase in opening and the selected kernel distance calculation path is a closed kernel distance calculation path, the corresponding sampling position is migrated to the open kernel distance calculation path; if the direction of the opening change in the execution feedback is a decrease in opening and the selected kernel distance calculation path is an open kernel distance calculation path, the corresponding sampling position is migrated to the closed kernel distance calculation path.

[0067] After the corresponding sampling position is migrated, the state transition kernel no longer accumulates the kernel distance calculation results of the corresponding sampling position along the original kernel distance calculation path. The migrated corresponding sampling position enters the matching kernel distance calculation path and participates in the state transition kernel calculation again according to the migrated kernel distance calculation path, so that the actual opening change direction in the execution feedback is consistent with the kernel distance calculation path.

[0068] A sampling position offset is considered to have occurred when the valve response sampling position in the execution feedback is not at the same sampling position as the corresponding valve response sampling position in the future calibration cycle recursion result. The sampling position offset is determined according to the difference in sampling sequence number between the valve response sampling position in the execution feedback and the corresponding valve response sampling position in the future calibration cycle recursion result.

[0069] When resetting the time-offset corrected covariance calculation order according to the sampling position offset, the command trigger sampling position is retained as the starting point for covariance calculation, and the valve response sampling position in the execution feedback is used as the new valve response alignment position. The hysteresis sampling time pairing relationship between the command trigger sampling position and the new valve response alignment position is re-established. After the reset, sampling positions that have not generated a valve response are not included in the covariance accumulation, and sampling positions after the new valve response alignment position are included in the covariance calculation order according to the sampling order.

[0070] When the sampling sequence corresponding to the execution feedback crosses the calibration condition switching position, first locate the last sampling position before the switch and the first sampling position after the switch in the execution feedback. Then, determine the position between the last sampling position before the switch and the first sampling position after the switch as the new covariance breakpoint. The sampling positions before the new covariance breakpoint are assigned to the calibration condition before the switch, and the sampling positions after the new covariance breakpoint are assigned to the calibration condition after the switch.

[0071] After re-determining the covariance breakpoint, the state transition kernel terminates the covariance calculation corresponding to the calibration condition before the switch, and restarts the covariance calculation from the first valve response sampling position after the switch. The covariance calculation after the switch reselects the opening kernel distance calculation path or the closing kernel distance calculation path according to the opening change direction corresponding to the execution feedback after the switch, and re-establishes the pairing relationship of the lag sampling time corresponding to the valve action lag.

[0072] The fuel cell valve calibration map stores the corrected kernel distance calculation path, the time-off corrected covariance calculation order, and the condition switching gating according to the calibration conditions. When a candidate valve under the same calibration conditions enters the state transition kernel, it calls the corrected kernel distance calculation path, the time-off corrected covariance calculation order, and the condition switching gating under the corresponding calibration conditions.

[0073] When the fuel cell controller enters any calibration condition, it retrieves the fuel cell valve calibration map according to the current calibration condition, and calls the kernel distance calculation path, the time-off corrected covariance calculation order, and the condition switching gating under the corresponding calibration condition. After the candidate valve action enters the state transition kernel, the state transition kernel performs future calibration cycle recursion according to the corresponding kernel calculation structure in the fuel cell valve calibration map, and then the acquisition function containing safety boundary fitting constraints determines the next round of valve calibration action.

[0074] Example 1: To verify the feasibility of this invention in practice, it was applied to the valve calibration process of a proton exchange membrane fuel cell system with a rated power level. This fuel cell system includes a hydrogen inlet valve, an air regulating valve, a back pressure valve, a hydrogen discharge valve, and a drain valve. The calibration covered operating states such as low-load steady state, rapid load increase, back pressure adjustment, and hydrogen and water discharge operations. Traditional calibration methods mainly rely on fixed calibration tables and manual correction. When faced with valve opening increases, decreases, response lags, and operating condition changes, it is easy to include valve responses in different directions in the same calculation chain, leading to larger pressure tracking deviations and some candidate valve actions approaching the safety boundary.

[0075] In this scenario, the valve action segments of the fuel cell are first captured according to the calibration conditions. Each segment retains the continuous process from the command triggering to the valve response entering the allowable range for the same valve action. The controller reads the allowable pressure range, allowable flow range, and allowable voltage fluctuation range under the corresponding calibration conditions to form the preset safe operating boundary of the fuel cell. The hydrogen inlet valve and back pressure valve exhibit significant action lag during the load increase phase, and the hydrogen discharge valve and drain valve cause short-term pressure disturbances after action. Traditional fixed calibration instruments are unable to separate these response differences.

[0076] After establishing the Gaussian process state-space model, the state transition kernel is transformed into two paths for calculating distances: one for the opening direction and one for the closing direction. Actions that increase valve opening only enter the opening direction kernel distance calculation path, and actions that decrease valve opening only enter the closing direction kernel distance calculation path. Before a candidate valve action enters the state transition kernel, a matching kernel distance calculation path is selected according to the direction of opening change; unselected paths do not participate in the current kernel calculation. With this processing, the model no longer mixes the opening and closing responses in its calculations, and the response differences caused by valve hysteresis are preserved in the state recursion.

[0077] To address valve action lag, the state transition kernel no longer directly pairs valve commands and responses at the same sampling time. Instead, it uses the command trigger sampling position as the starting point for covariance calculation, adjusting the valve response sampling position to the lagging sampling time based on the valve action lag. Sampling positions that do not produce a valve response are not included in the covariance accumulation; the first sampling position to produce a valve response is included in the covariance calculation sequence. This process ensures that the actual pressure, flow, and voltage changes after valve action are included in the same kernel calculation, reducing the interference of unresponsive sampling points on the calibration results.

[0078] When the load changes rapidly or hydrogen discharge / drainage operations switch, the operating condition switching gating truncates and restarts the time-off-corrected covariance calculation sequence. When future calibration cycle recursion crosses the calibration operating condition switching position, the calibration operating condition switching position is used as the covariance breakpoint, the covariance calculation before the switch stops, and the sampling position of the first valve response after the switch is used as the starting point for covariance calculation again. Under the new operating condition, the state transition kernel then reselects the opening kernel distance calculation path or the closing kernel distance calculation path according to the opening change direction corresponding to the candidate valve action. This method avoids placing the responses of the load rise phase, hydrogen discharge phase, and back pressure regulation phase in the same covariance calculation process.

[0079] During the candidate valve action selection phase, candidate valve actions first enter the state transition kernel after being adjusted by the operating condition switching gating. The model then recurses according to the future calibration cycle. After each state recursion, the state transition kernel outputs the recursion result for the future calibration cycle and the recursion uncertainty at the corresponding sampling position. When the recursion result enters the inner-close range of the preset fuel cell operating safety boundary but does not reach the safety boundary, the valve adjustment amplitude in the next recursion is reduced; when the recursion result reaches the safety boundary, the current candidate valve action stops further recursion. The Bayesian optimization acquisition function no longer directly evaluates the valve action itself, but rather evaluates the recursion result for the future calibration cycle. The acquisition function masks candidate valve actions that reach the safety boundary, and improves the ranking position of candidate valve actions that are close to the safety boundary but have not reached it. When the next sampling step is about to reach the safety boundary, the acquisition function value is rolled back to the corresponding value of the previous sampling step.

[0080] After completing the next round of valve calibration, the controller will align the feedback with the recursive results of future calibration cycles according to the sampling order. If the direction of the opening change in the feedback is inconsistent with the selected core distance calculation path, the corresponding sampling position will be migrated to the matching core distance calculation path. If the sampling position of the valve response in the feedback shifts, the covariance calculation order corrected by time offset will be readjusted according to the sampling position shift. If the sampling order corresponding to the feedback crosses the calibration condition switching position, the condition switching gating will re-determine the covariance breakpoint. The corrected core distance calculation path, the covariance calculation order corrected by time offset, and the condition switching gating are written into the fuel cell valve calibration map for subsequent calibration under the same calibration conditions.

[0081] Table 1: Comparison of the Implementation Effects of Adaptive Calibration for Fuel Cell Valves

[0082] The data in the table are the average results of repeated tests on the same fuel cell system under the same load step, gas supply regulation, back pressure regulation, and hydrogen and water discharge conditions.

[0083] Table 1 shows that the present invention achieves significant improvements in pressure control, response hysteresis identification, safety constraints, and calibration efficiency. The maximum deviation of the hydrogen inlet pressure decreased from 5.8 kPa to 2.1 kPa, and the maximum deviation of the air-side pressure decreased from 6.4 kPa to 2.4 kPa, indicating that after the open-side kernel distance calculation path and the closed-side kernel distance calculation path are separated, the state transition kernel can more accurately distinguish the responses of the valve in different action directions. The peak value of the anode-cathode pressure difference fluctuation decreased from 9.2 kPa to 4.1 kPa, reflecting that the operating condition switching gating and covariance breakpoint handling have a suppressive effect on the cross-operating condition response mixing problem.

[0084] The valve response hysteresis identification error decreased from 118ms to 46ms, indicating that the covariance calculation order after time offset correction can more accurately align valve commands and actual responses. The peak value of dynamic loading voltage fluctuation decreased from 4.6V to 2.0V, and the number of sampling steps for hydrogen emission disturbance recovery decreased from 14 steps to 7 steps, indicating that after the candidate valve actions are constrained by the safety boundary during future calibration cycle recursion, the impact of valve actions on stack output disturbances and hydrogen emission disturbances is controlled.

[0085] The number of successful calibration tests decreased from 42 to 18, and the number of times the safety boundary was reached decreased from 7 to 0, indicating that the rewritten Bayesian optimization acquisition function can find more valuable candidate valve actions within the safety boundary and reduce invalid tests. The hydrogen flow tracking error decreased from 4.8% to 1.9%, and the hydrogen utilization rate increased from 96.1% to 98.0%, indicating that after the execution feedback was written back to the state transition kernel and the fuel cell valve calibration map, the calibration results more closely matched the actual valve response.

[0086] The above description is only a preferred embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any equivalent substitutions or modifications made by those skilled in the art within the scope of the technology disclosed in the present invention, based on the technical solution and inventive concept of the present invention, should be covered within the scope of protection of the present invention.

Claims

1. A Bayesian optimization-based adaptive calibration method for fuel cell valves, characterized in that, Includes the following steps: S1. Select the valve action segment of the fuel cell according to the calibration conditions and preset the safe operating boundary of the fuel cell; S2. Set up open kernel distance calculation paths and closed kernel distance calculation paths in the state transition kernel of the Gaussian process state space model, and select matching kernel distance calculation paths according to the opening change direction of the fuel cell valve action segment. S3. Perform time offset correction on the covariance calculation order of the state transition kernel according to the valve action lag in the fuel cell valve action segment, so that the valve response before and after the lag participates in the same state transition kernel calculation. S4. Set up a working condition switching gating in the state transition kernel, and adjust the gating of the matching kernel distance calculation path and the covariance calculation order corrected by time offset according to the calibrated working condition. S5. The state transition kernel, after being adjusted by the operating condition switching gating, is used to recursively calculate the future calibration cycle of the candidate valve action input formed by the fuel cell valve action segment, and the result of the future calibration cycle recursion is limited to the preset fuel cell operating safety boundary. S6. Rewrite the Bayesian optimization acquisition function into an acquisition function with safety boundary fitting constraints, and select the next round of valve calibration action according to the rule of being close to the preset fuel cell operating safety boundary without touching the preset fuel cell operating safety boundary. S7. Execute the next round of valve calibration actions to obtain execution feedback, synchronously correct the state transition kernel based on the execution feedback, and update the fuel cell valve calibration chart according to the calibration conditions.

2. The adaptive calibration method for fuel cell valves based on Bayesian optimization according to claim 1, characterized in that, The step of selecting the fuel cell valve action segment according to the calibration conditions in S1 includes: The historical calibration process of fuel cell valves is segmented according to the calibration conditions. The continuous segment between the command triggering and the valve response entering the allowable range under the corresponding calibration conditions is taken as the fuel cell valve action segment. Read the allowable pressure range, allowable flow range, and allowable voltage fluctuation range under the calibrated operating conditions corresponding to the valve action segment of the fuel cell, and use the allowable pressure range, allowable flow range, and allowable voltage fluctuation range as the boundary composition of the preset fuel cell operating safety boundary.

3. The adaptive calibration method for fuel cell valves based on Bayesian optimization according to claim 1, characterized in that, The steps in S2 for setting the open-direction kernel distance calculation path and the closed-direction kernel distance calculation path include: Before performing kernel distance calculation in the state transition kernel, the open kernel distance calculation path and the closed kernel distance calculation path are divided; Based on the direction of the opening change of the fuel cell valve action segment, a matching core distance calculation path is selected from the opening core distance calculation path and the closing core distance calculation path; The fuel cell valve action segment with increased opening degree is sent to the open core distance calculation path, and the fuel cell valve action segment with decreased opening degree is sent to the closed core distance calculation path. Within the selected nuclear distance calculation path, only fuel cell valve action segments with the same opening change direction are called to participate in the nuclear distance calculation; The unselected kernel distance calculation path is excluded from the current state transition kernel calculation; The kernel distance calculation results of the selected kernel distance calculation path are sent to the time offset corrected covariance calculation sequence, so that the state transition kernel performs state recursion of candidate valve actions based on the kernel distance calculation results.

4. The adaptive calibration method for fuel cell valves based on Bayesian optimization according to claim 1, characterized in that, The time offset correction step in S3 includes: Using the command trigger sampling position in the fuel cell valve action segment as the starting point for covariance calculation, the valve response sampling positions are adjusted from being paired at the same sampling time to being paired at a delayed sampling time according to the valve action lag. Within the state transition kernel, the accumulation of covariance for sampling positions that do not produce a valve response is cancelled, and the first sampling position that produces a valve response is included in the covariance calculation order. Continue to access subsequent valve response sampling positions along the sampling time sequence, so that the instruction trigger sampling position and the delayed valve response sampling position participate in the kernel calculation in the same state recursion process; The time-offset corrected covariance calculation order receives the kernel distance calculation results output by the open kernel distance calculation path and the closed kernel distance calculation path, thus completing the time offset correction of the covariance calculation order of the state transition kernel.

5. The adaptive calibration method for fuel cell valves based on Bayesian optimization according to claim 4, characterized in that, The steps for gating adjustment in S4 include: Before the time-off-corrected covariance calculation sequence enters the state transition kernel calculation, the operating condition switching gating determines the starting point and ending point of the covariance calculation according to the current calibrated operating condition. When the future calibration cycle corresponding to the candidate valve action remains within the same calibration condition, the covariance calculation order corresponding to the current calibration condition is maintained, and the kernel distance calculation results output by the selected kernel distance calculation path are continuously connected. When the future calibration cycle corresponding to the candidate valve action crosses the calibration condition switching position, the covariance calculation corresponding to the calibration condition before the switching is terminated by taking the calibration condition switching position as the covariance breakpoint. Restart the covariance calculation from the first valve response sampling position after the switch, and select the matching kernel distance calculation path from the opening kernel distance calculation path and the closing kernel distance calculation path according to the opening change direction corresponding to the candidate valve action after the switch. The covariance calculation results corresponding to the calibration conditions before the switchover and the covariance calculation results corresponding to the calibration conditions after the switchover are entered into the state transition kernel in the order of the calibration conditions, so as to complete the gating adjustment of the covariance calculation order after time offset correction.

6. The adaptive calibration method for fuel cell valves based on Bayesian optimization according to claim 1, characterized in that, The steps in S5 for recursively calculating the future calibration period include: The candidate valve action is input into the state transition kernel after the working condition switching gating adjustment, and the selected kernel distance calculation path calls the covariance calculation order corrected by time offset; The state recursion is executed step by step according to the sampling order of the future calibration cycle. After each state recursion is completed, the state transition kernel outputs the recursion uncertainty of the corresponding sampling position and makes a judgment on the recursion result of the future calibration cycle with the preset fuel cell operation safety boundary. If the future calibration cycle recursion result does not enter the inner contact range of the preset fuel cell operating safety boundary, the recursion direction of the current candidate valve action will be maintained. When the future calibration cycle recursion result enters the inner contact range of the preset fuel cell operating safety boundary but does not reach the preset fuel cell operating safety boundary, the adjustment range of the current candidate valve action in the next state recursion will be reduced. When the future calibration cycle recursion results reach the preset fuel cell operating safety boundary, the subsequent recursion of the current candidate valve action will be stopped.

7. The adaptive calibration method for fuel cell valves based on Bayesian optimization according to claim 6, characterized in that, The steps for selecting the next round of valve calibration action in S6 include: The future calibration cycle recursion result corresponding to the candidate valve action is used as the calculation object of the Bayesian optimization acquisition function; Before the acquisition function is calculated, the actions of candidate valves that reach the preset fuel cell operating safety boundary are blocked; For candidate valve actions that do not reach the preset fuel cell operating safety boundary, the collected function values ​​are updated in the order of decreasing distance between the future calibration cycle recursive results and the preset fuel cell operating safety boundary. When the future calibration cycle recursion result will reach the preset fuel cell operation safety boundary in the next sampling step, the acquisition function value will be rolled back to the acquisition function value of the previous sampling step. For candidate valve actions with the same acquired function value, they are reordered in order of increasing recursive uncertainty. The candidate valve action ranked first is determined as the valve calibration action for the next round.

8. The adaptive calibration method for fuel cell valves based on Bayesian optimization according to claim 7, characterized in that, In step S7, the state transition kernel is synchronously corrected based on execution feedback, and the fuel cell valve calibration chart is updated according to the calibration conditions. Specifically, this includes: Align the execution feedback of the next round of valve calibration actions with the corresponding future calibration cycle recursion results according to the sampling order; When the direction of the opening change in the execution feedback is inconsistent with the selected kernel distance calculation path, the corresponding sampling position is migrated to the kernel distance calculation path that matches the opening kernel distance calculation path and the closing kernel distance calculation path. When the sampling position of the valve response in the execution feedback is offset from the corresponding sampling position of the valve response in the future calibration cycle recursion result, the covariance calculation order after time offset correction is reset according to the sampling position offset amount; When the sampling sequence corresponding to the execution feedback crosses the calibration condition switching position, the covariance breakpoint is re-determined, and the covariance calculation is restarted from the first valve response sampling position after the switch. The corrected nuclear distance calculation path, the time-off corrected covariance calculation order, and the operating condition switching gating are written into the fuel cell valve calibration chart under the corresponding calibration conditions.

Images