MAP self-updating lean burn air-fuel ratio feedforward control method and system based on iterative learning

Abstract

The invention belongs to the technical field of automobile engine air-fuel ratio control, particularly discloses an MAP self-updating lean-burn air-fuel ratio feedforward control method and system based on iterative learning, and aims to solve the problems of high calibration cost, oxygen sensor feedback delay and long-term precision reduction in lean-burn engine air-fuel ratio control. According to the method, accurate control over the air-fuel ratio is achieved through parameterized feedforward controller design, MAP table dynamic updating and a bilinear interpolation algorithm, the calibration cost is remarkably reduced by 80% or above, the transient working condition control error is stabilized within + / -1.5%, high precision is still kept after long-term operation, and the method is suitable for in-cylinder direct injection, air inlet channel injection and mixed injection engines.

Description

Technical Field

[0001] The present invention belongs to the technical field of automobile engine air-fuel ratio control, and specifically relates to a MAP self-updating lean burn air-fuel ratio feedforward control method and system based on iterative learning, which is applicable to lean burn engines with direct injection, intake port injection or mixed injection. Background Art

[0002] Lean-burn technology, which reduces fuel consumption and nitrogen oxide (NOx) emissions by increasing the air-fuel ratio (typically greater than 18:1), is currently a key technology for achieving energy conservation and emission reduction in the internal combustion engine industry. However, its practical application faces significant challenges in air-fuel ratio control. When approaching the lean-burn limit, even slight deviations in the air-fuel ratio can easily lead to combustion instability, misfire, or detonation. Furthermore, oxygen sensor signal delays exacerbate control errors under transient conditions. Therefore, achieving rapid and precise air-fuel ratio adjustment has become a core challenge in the promotion of lean-burn technology.

[0003] Currently, air-fuel ratio control primarily relies on a composite control strategy combining feedforward and feedback. Feedforward control uses a preset MAP (map spectrum) table to directly look up the injection amount based on engine speed and throttle opening, while feedback control uses oxygen sensor signals to correct steady-state deviations. However, existing technologies have the following limitations:

[0004] Static MAP tables used in feedforward control suffer from drawbacks and high calibration costs: Lean burn requires coverage of a wider operating range (such as switching between lean and stoichiometric conditions), resulting in an exponential increase in MAP table calibration workload. Factors such as engine aging and carbon deposits can cause MAP table parameters to gradually deviate from their true values, and traditional static MAP tables cannot adaptively update. MAP tables store parameters on a discrete grid, requiring interpolation to estimate parameters under transient conditions. However, linear interpolation methods have limited accuracy in nonlinear systems. Oxygen sensor signals exhibit a delay of approximately 100-300ms, leading to hysteresis in feedback control under transient conditions and failing to independently guarantee control accuracy. Existing feedback controllers (such as PID algorithms) rely on empirical parameter tuning, lacking robustness in lean burn scenarios where the air-fuel ratio reference value frequently changes. Traditional ILC is primarily used for repetitive tasks (such as robot trajectory tracking). However, lean burn air-fuel ratio reference values ​​dynamically change with operating conditions, making it a non-repetitive task. Directly applying ILC results in inefficient learning and even divergence. Existing research attempts to combine ILC with model predictive control, but this relies on a precise mathematical model of the engine, making it difficult to scale in practical applications due to model mismatch. Some technologies (such as patent CN113250843A) propose dynamically updating feedforward parameters based on real-time data, but these methods are not integrated with the MAP structure and still rely on manual calibration of initial parameters. Other approaches, such as Shaanxi Diesel Heavy Industry's dual-fuel control technology, use exhaust gas recirculation (EGR) to expand the air-fuel ratio control range, but they do not address the long-term drift of the MAP.

[0005] In summary, existing technologies struggle to achieve high precision, low calibration costs, and long-term stability for lean-burn air-fuel ratio control. A feedforward control method that can dynamically update the MAP, adapt to non-repetitive tasks, and not rely on precise models is urgently needed to overcome the engineering bottleneck of lean-burn technology. Summary of the Invention

[0006] To address the problems of high MAP table calibration cost, long-term accuracy degradation, and poor adaptability to non-repetitive tasks in existing lean burn air-fuel ratio control, the present invention proposes a MAP self-updating feedforward control method based on iterative learning. The method aims to achieve precise and stable control of the air-fuel ratio by dynamically updating MAP table parameters, combining a parameterized feedforward controller and a bilinear interpolation algorithm, while reducing calibration costs and improving system robustness.

[0007] The purpose of the present invention is achieved through the following technical solutions:

[0008] The core of the present invention is to deeply integrate iterative learning control (ILC) with the dynamic update mechanism of the MAP table. The specific technical solutions are as follows:

[0009] Step 1: Parameterized feedforward controller design

[0010] The feedforward controller of the air-fuel ratio system is modeled as a parameterized expression of the inverse model of the system:

[0011]

[0012] Among them, the parameter vector θ = [b1, b2, a0, a1]T is dynamically updated according to the working conditions.

[0013] Model order adaptation: Automatically selects a first-order or second-order model structure based on the current engine operating conditions (such as speed and load) (for example, simplifying to a first-order model F(θ,z)=1+b1z-1a0 at low loads) to balance computational efficiency and control accuracy.

[0014] Step 2: MAP table initialization and iterative learning calibration.

[0015] Gridded MAP table construction: The MAP table is divided into grids (e.g., 20×20) with the speed n and throttle opening α as the two-dimensional coordinate axes, and each grid point stores the feedforward parameters θi,j of the corresponding working condition.

[0016] Model-free initial calibration: An iterative learning algorithm (Formula 5) is used to traverse all operating points and directly estimate the initial parameter θ0 using real-time input and output data without relying on an accurate engine mathematical model.

[0017] Step 3: Bilinear interpolation and transient operating condition parameter optimization.

[0018] Dynamic parameter acquisition: During transient operation, the local grid area in the MAP table is located according to the current operating point (n, α), and the feedforward parameters are calculated using a bilinear interpolation algorithm:

[0019] θ=i,j∑wi,jθi,j

[0020] The weight wi,j is determined by the distance from the current point to the grid vertex, and only four adjacent parameters of the current area need to be updated each time.

[0021] Local update mechanism: When the system detects that the error of the current working condition parameters exceeds the limit, it only iteratively learns and updates the parameters in the grid area to avoid wasting global computing resources.

[0022] Step 4: Iterative learning and MAP table self-update.

[0023] Error-driven update: The air-fuel ratio tracking error e = Sr - SPf - Sv is calculated in real time. If the performance indicator E(e) ≠ 0 (i.e., the expected error is non-zero), the parameter update process is started:

[0024]

[0025] Among them, QIV is the forgetting factor matrix, LIV is the learning gain matrix, and the convergence speed and stability are balanced by adjusting the weights of the two.

[0026] Step 5: MAP table closed-loop update:

[0027] The converged parameter θ is written back to the corresponding grid point in the MAP table to form a "learning-update-feedback" closed loop to continuously optimize the control accuracy.

[0028] As a more optimal technical solution of the present invention, the order of the dynamic model in step 1 is adaptively adjusted according to the working conditions, specifically:

[0029] During the parameter updating process, the first-order or second-order model structure is dynamically selected through an iterative learning algorithm to make the feedforward controller approach the system inverse model.

[0030] As a more optimal technical solution of the present invention, the acquisition of the initial feedforward controller parameter θ0 in step 2 does not need to rely on the initial model of the system, but can be directly generated through iterative learning of real-time measurement data.

[0031] As a more optimal technical solution of the present invention, the implementation method of the bilinear interpolation algorithm in step 3 is:

[0032] When the engine operating condition is in a certain grid area of ​​the MAP table, only the parameters of the four adjacent grid points in the area are interpolated and calculated, and only the parameters of the current coverage area are updated each time.

[0033] As a more optimal technical solution of the present invention, the performance indicator described in step 4 is the expected value E(e)=0 of the error e, which is specifically calculated by the formula e=Sr-SPf-Sv, where S is the sensitivity function and v is the measurement noise.

[0034] As a more optimal technical solution of the present invention, the grid division density of the MAP table is dynamically adjusted according to the control accuracy requirement. The denser the grid, the greater the number of parameters θ.

[0035] As a more optimal technical solution of the present invention, the feedforward controller is suitable for a lean burn engine with direct injection, port injection or mixed injection.

[0036] Yet another object of the present invention is to provide an electronic control unit comprising a memory and a processor, wherein the memory stores a computer program that processes the steps of the method.

[0037] Another object of the present invention is to provide an automobile engine control system, comprising the electronic control unit as claimed in claim 8, and a sensor module for collecting speed, throttle opening, and oxygen sensor signals.

[0038] Beneficial effects

[0039] The core innovation of this invention lies in the organic combination of iterative learning, parametric modeling and dynamic updating of the MAP table, which breaks through the limitations of traditional feedforward control in nonlinear and non-repetitive tasks and provides an efficient solution for the engineering application of lean burn technology.

[0040] The present invention automatically generates the initial MAP table through iterative learning, reducing the manual calibration workload by more than 80%. The dynamic update mechanism of the MAP table can compensate for parameter drift caused by factors such as engine aging and carbon deposits, so that the air-fuel ratio control error is stable within ±1.5% for a long time. The parameter estimation accuracy of the bilinear interpolation algorithm under transient conditions is more than 30% higher than that of traditional linear interpolation. The parameterized feedforward controller combined with the ILC framework can adapt to dynamic changes in the air-fuel ratio reference value (such as switching between lean burn and stoichiometric ratio modes). It is suitable for direct injection, intake port injection and mixed injection systems without changing the existing sensor and ECU hardware architecture. BRIEF DESCRIPTION OF THE DRAWINGS

[0041] Figure 1 is a schematic diagram of an engine air-fuel ratio system to which the present invention is applicable;

[0042] 1: Flow sensor

[0043] 2: Pressure sensor

[0044] 3: Temperature sensor

[0045] 4: Fuel injector

[0046] 5: Spark plug

[0047] 6: Cylinder pressure sensor

[0048] 7: Front oxygen sensor

[0049] 8: Rear oxygen sensor

[0050] Figure 2 This is a control block diagram after the present invention is implemented in the air-fuel ratio system. Figure 2 In the example, P represents the engine air-fuel ratio system object, C fb represents the feedback controller, F(θ, z) represents the parameterized model of the feedforward controller, f is the feedforward adjustment signal, r is the air-fuel ratio reference value, y is the air-fuel ratio output value, and v is the measurement noise;

[0051] Figure 3 Schematic diagram of MAP grid update for the working condition coverage area of ​​the present invention. In the figure, the red solid line is the working condition change curve, and the seven-pointed star points are the grid points for parameter update under the working condition change. DETAILED DESCRIPTION

[0052] The present invention is described in further detail below with reference to the accompanying drawings.

[0053] The present invention first establishes a functional relationship between the speed, throttle opening and air-fuel ratio feedforward controller parameters, and establishes an initial MAP table based on this. During engine operation, based on the initial MAP table, a bilinear interpolation method is used to perform a more accurate search. At the same time, an iterative learning method is used to update the parameters in the MAP table in real time when the control accuracy deviates under operating conditions. This feedforward control strategy uses an iterative learning method to improve control accuracy while using a method of reasonable parameterization of the feedforward controller to improve the flexibility of non-repetitive tasks such as lean burn air-fuel ratio tracking control. The MAP table plus bilinear interpolation algorithm can improve the accuracy of the lookup table, and ultimately achieve fast and accurate feedforward adjustment of the lean burn air-fuel ratio.

[0054] The present invention proposes a feedforward parameter MAP self-updating control strategy based on iterative learning for the air-fuel tracking control problem of lean burn engines. The present invention is applicable to air-fuel ratio systems such as Figure 1 As shown in FIG, the control strategy proposed by the present invention is still applicable after the fuel injection mode is changed from direct injection in the cylinder to port injection or mixed injection. The control block diagram after the control strategy proposed by the present invention acts on the air-fuel ratio system is shown in FIG. Figure 2 As shown, where P represents the engine air-fuel ratio system object, C fb represents the feedback controller, F(θ, z) represents the parameterized model of the feedforward controller, f is the feedforward adjustment signal, r is the air-fuel ratio reference value, y is the air-fuel ratio output value, and v is the measurement noise. The implementation of the control strategy mainly includes the following steps:

[0055] Step 1: Determine the engine air-fuel ratio system model and describe the model as a first-order or two-order model. Determine the feedforward controller structure based on the system model structure, and thus determine the parameter θ to be updated:

[0056] according to Figure 2 , the size of the error e is shown in formula (1). Therefore, when the input f of the feedforward controller to the system satisfies formula (2), the expected error e can be obtained as E(e) = -E(Sv) = 0:

[0057] e=Sr-SPf-Sv

[0058] f=P -1 r=F(θ,z)r

[0059] Then, the design form of the feedforward controller can be obtained as:

[0060] F(θ, z) = P -1 (z)

[0061] Considering the air-fuel ratio system as a first-order or second-order system, F(θ, z) is expressed as:

[0062]

[0063] From this we can get the parameter θ to be updated:

[0064] θ=[b1 b2 a0 a1] T

[0065] During actual operation, the order of the system model will change with the operating conditions. The learning of the parameter θ will make the learned feedforward controller model closer to the system inverse model under different operating conditions.

[0066] Step 2: Traverse the engine operating conditions according to the division of the feedforward parameter MAP table, obtain the feedforward controller parameters through iterative learning when traversing each operating condition, and store the parameters in the MAP table:

[0067] When first obtaining the engine feedforward parameter MAP table or updating the table, first give the initial parameter θ 0 , according to θ 0 Construct F(θ 0 , z) and generates a feedforward signal F(θ 0 , z)r, and then it will run iteratively after each task starts. Set the task index j, and the error e is obtained in the jth task j and output y j The update process of θ in the jth task can be described by the following formula:

[0068]

[0069] Where k is the number of generations updated at the jth task; Q IV，(k) and L IV，(k) They are robust filtering and learning filtering:

[0070]

[0071] Q IV，(k) =L IV，(k) B(θ j )Φ IV (θ j )I 4×4

[0072]

[0073] Where, A and Ψ B is the convolution matrix corresponding to the polynomial basis function; Z is the instrumental variable defined to be related to r and unrelated to v; C = C fb +F(θ j , z).

[0074] The above is a feedforward parameter learning method based on the instrumental variable method. Through the above method, the iterative learning process can be started based only on measurement data when there is no accurate system model.

[0075] The iteration stop condition can be set as:

[0076]

[0077] Where η is selected as a smaller number according to the requirements, and the feedforward controller parameters are recorded in the MAP table when the iterative learning meets the stopping condition.

[0078] Step 3: Calculate the feedforward controller parameters under the corresponding working conditions during engine transient operation and implement the control signal F(θ, z)r into the system according to the reference value:

[0079] During transient operation, it is necessary to look up the MAP table to obtain the feedforward controller parameter θ and further obtain the feedforward control signal based on the parameter. θ can be described as a function of the engine speed n and the throttle opening α as follows:

[0080] θ=f θ (n, α)

[0081] The grid division of the lookup table is defined as formula (11), where a, b∈R and c, d∈R are the limits of the input values, and p1 and p2 are the grid points divided in the interval. Then The two-dimensional lookup table is described using the piecewise bilinear interpolation formula as follows:

[0082]

[0083] By expanding the input interval to (n,α)∈[a,b]×[c,d], formula (12) can be rewritten as the product of two vectors:

[0084] f T (θ i,j ,n,α)=Φ(n,α)·θ (13)

[0085] in:

[0086]

[0087]

[0088] in:

[0089]

[0090] Through the above bilinear interpolation algorithm, we can get the value of * The throttle opening is α * The parameters of the feedforward controller are The feedforward signal input is thus obtained as

[0091] Step 4: During engine operation, the iterative learning algorithm is always running. If a large error occurs, it returns to step 2 and updates the parameters until the error requirement is met. At this time, the feedforward controller parameters θ are updated.

[0092] Step 5: Update the parameters that meet the control performance index into the MAP table of the corresponding working condition. i，j The number of parameters at the grid points in the MAP table is related to the division of the grid. The finer the grid division, the more parameters there are. However, within a specified area, only four parameters can be calculated simultaneously at a time. Only the areas covered by the operating conditions will have their parameters calculated and updated. The bilinear interpolation formula only updates the four parameters within the area where the current operating condition is located during each calculation. When the operating conditions cover the entire area, as the parameters are continuously updated, the estimated values ​​obtained by the bilinear interpolation method will gradually converge to the true values, resulting in an optimal update curve, thus realizing the self-update process of the MAP.

[0093] Example 1: Parameterized feedforward controller design and model order adaptation

[0094] Implementation goal: To verify the order adaptation capability of the feedforward controller under different operating conditions.

[0095] System Identification and Model Selection

[0096] A step response test was conducted on a 1.5L lean-burn gasoline engine, and data on speed n and throttle opening α were collected.

[0097] Under low load conditions (n ​​= 1500 rpm, α = 20%), system identification shows that the dynamic characteristics of the air-fuel ratio are close to the first-order model, and the transfer function is:

[0098]

[0099] The corresponding feedforward controller is designed as:

[0100]

[0101] Under high load conditions (n ​​= 4000 rpm, α = 80%), system identification shows that the second-order model is better:

[0102]

[0103] The feedforward controller is:

[0104]

[0105] Effect verification

[0106] In the WLTC transient cycle, model order adaptation reduces the air-fuel ratio tracking error to ±1.2% (the fixed-order model error is ±2.5%).

[0107] Example 2: MAP table initialization and iterative learning calibration

[0108] Implementation goal: Demonstrate the model-free calibration process and efficiency improvement.

[0109] MAP table grid division

[0110] Define the speed range n∈[800,6000]rpm and throttle opening α∈[0,100]%, and divide it into a 20×20 grid (a total of 400 operating points).

[0111] Instrumental variable method iterative learning

[0112] The initial parameter θ0 is randomly set (e.g., θ0 = [0, 0, 0, 0]T).

[0113] The iterative learning of formula (5) is performed for each operating point, and the forgetting factor QIV=0.9I and the learning gain LIV=0.1I are set.

[0114] Take the operating point (n = 2000rpm, α = 50%) as an example:

[0115] After the first iteration, the error ∥e∥2=4.5%, and after the fifth iteration, it drops to 0.8%0.8%, meeting the termination condition η=1%.

[0116] The final learning parameters are θ = [-0.4, 0.08, 0.75, 0.18] T, which are stored in the corresponding grid of the MAP table.

[0117] 3. Calibration efficiency comparison

[0118] Traditional manual calibration takes 50 hours, while the present invention only takes 10 hours through automatic iterative learning, improving efficiency by 80%.

[0119] Example 3: Application of bilinear interpolation parameters in transient conditions

[0120] Implementation goal: To verify the accuracy advantage of bilinear interpolation under transient conditions.

[0121] Interpolation calculation

[0122] The current operating point is (n*=2500 rpm, α*=60%), which is located in the grid area (n2=2000, n3=3000)×(α3=50%, α4=70%).

[0123] According to formula (12), the interpolation weight is calculated:

[0124] ω 3，3 =0.75×0.5=0.3

[0125] w 2，4 =0.5×0.5×0.25, w 3，4 =0.25×0.5=0.125

[0126] Final parameters:

[0127] θ(n * , α * )=0.25θ 2，3 +0.375θ 3，3 +0.25θ 2，4 +0.125θ 3，4

[0128] Accuracy comparison: The error of bilinear interpolation is ±1.0%, and the error of traditional linear interpolation is ±2.3%.

[0129] Example 4: MAP table closed-loop self-update and long-term stability verification

[0130] Verify the effect of the parameter update mechanism on maintaining long-term control accuracy.

[0131] Aging simulation and parameter updating

[0132] After the simulated engine has run for 1000 hours, the MAP table parameters drift due to carbon deposits (such as θi,j deviation +10%).

[0133] At the operating point (n = 3000 rpm, α = 40%), an error of ∥e∥2 = 3.8% is detected, triggering a parameter update:

[0134] The error dropped to 2.1% after the first update and converged to 0.9% after the third iteration.

[0135] Update the parameters θ3,4 = [-0.42, 0.09, 0.73, 0.17]T and write it back to the MAP table. The update result is as follows Figure 3 shown.

[0136] Long-term stability test: The updated MAP table keeps the control error stable within ±1.5% for a long time. When not updated, the error increases to ±4.2%. Industrial application data

[0137]

[0138] It can be seen from the above embodiments that the present invention is significantly superior to traditional methods in terms of parameterized model design, dynamic update of MAP table, bilinear interpolation algorithm and long-term stability, and provides an efficient and low-cost solution for air-fuel ratio control of lean burn engines.

Claims

1. A MAP self-updating lean burn air-fuel ratio feedforward control method based on iterative learning, characterized in that: The following steps are involved: Step 1: Establish a dynamic model of the engine air-fuel ratio system and parameterize the feedforward controller into an inverse model expression F(θ,z)=B(θ)A(θ), where θ is the parameter vector to be updated, including the coefficients a0, a1, b1, b2 of the first-order or second-order model; Step 2: Divide the engine operating conditions into a two-dimensional MAP table of speed and throttle opening, traverse the operating conditions and obtain the initial feedforward controller parameters θ0 through an iterative learning algorithm, and store the parameters in the MAP table; Step 3: During transient engine operation, the corresponding feedforward controller parameter θ is obtained from the MAP table using a bilinear interpolation algorithm based on the current operating point to generate a feedforward control signal F(θ,z)r. Step 4: Real-time monitoring of the air-fuel ratio output error e. If the performance index does not meet the preset conditions, update the parameter θ through the iterative learning algorithm. The update formula is: Until the error converges; Step 5: Write the updated parameter θ back to the corresponding working condition grid point in the MAP table to complete the self-update of the MAP table.

2. The control method according to claim 1, characterized in that: The order of the dynamic model in step 1 is adaptively adjusted according to the working conditions, specifically: During the parameter updating process, the first-order or second-order model structure is dynamically selected through an iterative learning algorithm to make the feedforward controller approach the system inverse model.

3. The control method according to claim 1, characterized in that: The acquisition of the initial feedforward controller parameter θ0 in step 2 does not rely on the initial system model and can be directly generated through iterative learning of real-time measurement data.

4. The control method according to claim 1, wherein: The implementation method of the bilinear interpolation algorithm in step 3 is: When the engine operating condition is in a certain grid area of ​​the MAP table, only the parameters of the four adjacent grid points in the area are interpolated and calculated, and only the parameters of the current coverage area are updated each time.

5. The control method according to claim 1, characterized in that: The performance indicator in step 4 is the expected value E(e)=0 of the error e, which is specifically calculated by the formula e=Sr-SPf-Sv, where S is the sensitivity function and v is the measurement noise.

6. The control method according to claim 1, characterized in that: The grid division density of the MAP table is dynamically adjusted according to the control accuracy requirement. The denser the grid, the greater the number of parameters θ.

7. The control method according to claim 1, characterized in that: The feedforward controller is suitable for a lean burn engine with direct injection, port injection or mixed injection.

8. An electronic control unit, characterized in that: The method comprises a memory and a processor, wherein the memory stores a computer program, and when the processor executes the program, the steps of the method according to any one of claims 1 to 7 are implemented.

9. An automobile engine control system, characterized in that: The invention comprises the electronic control unit as claimed in claim 8, and a sensor module for collecting rotational speed, throttle opening, and oxygen sensor signals.

Images