Composite interference compensation control method based on online learning and multi-model fusion

Abstract

The invention belongs to the technical field of system control, and particularly relates to a composite interference compensation control method based on online learning and multi-model fusion, which comprises the following steps: S1, establishing a discrete time state space model of a nonlinear dynamic system influenced by composite interference; s2, designing a nonlinear interference observer, and estimating lumped interference suffered by the system on line in real time; s3, decomposing the lumped interference estimated value in real time to obtain a plurality of interference components with different dynamic characteristics; s4, aiming at the dynamic characteristics of each type of interference components obtained by decomposition, designing a corresponding special compensator, and forming a parallel multi-model compensation architecture; s5, dynamically fusing the outputs of the three compensators according to the current interference characteristics through an intelligent arbiter, and generating a composite interference compensation control vector; and S6, the composite interference compensation control vector and the nominal control vector are added to obtain a final total control input vector, and the final total control input vector is applied to a controlled object after being subjected to amplitude limiting processing of an execution mechanism.

Description

TECHNICAL FIELD

[0001] The present application belongs to the technical field of system control, and in particular to a compound disturbance compensation control method based on online learning and multi-model fusion. BACKGROUND

[0002] In complex industrial process control and high dynamic systems (such as chemical reaction processes, biological fermentation devices, aerospace servo mechanisms, etc.), the controlled object is often affected by multiple types of disturbances. These disturbance sources include, but are not limited to, slow parameter drift caused by equipment aging and wear, periodic load fluctuations caused by rotating or reciprocating parts, and random impulse disturbances caused by raw material mutations, external environmental shocks, etc. These disturbances often have different dynamic characteristics (such as frequency band, amplitude, and change rate), and are coupled with each other in actual systems, forming compound disturbances, which seriously affect the control accuracy, stability, and dynamic performance of the system. In order to improve the anti-disturbance ability of the system, various methods have been developed in the control field, such as integral control and adaptive control for constant or slowly varying disturbances, repetitive control and iterative learning control for periodic disturbances, and intelligent feedforward compensation strategies based on data-driven. These methods have shown certain effectiveness in specific disturbance scenarios, and how to effectively integrate the advantages of multiple strategies to build an intelligent control architecture that can perceive, decompose, and cooperatively suppress compound disturbances has become an important research direction in the design of high-precision, robust control systems.

[0003] The anti-disturbance control method in the prior art has the following limitations when faced with multiple-source, time-varying, and strongly coupled compound disturbances in complex industrial processes such as chemical industry and aerospace: First, the control strategy is single and cannot take into account disturbances with different characteristics, such as poor suppression of periodic disturbances by PID and ineffective suppression of random disturbances by repetitive control, resulting in limited overall performance. Second, it lacks adaptive ability, and most methods use fixed parameters or offline trained models, which cannot track the dynamic changes of disturbances online, and the performance significantly decreases when the disturbance characteristics change. Third, the multi-compensation mechanism lacks intelligent coordination, and simple parallel or switching strategies can cause control conflicts or chattering, which can damage system stability. In addition, theoretical methods often lack engineering practicality and do not fully consider physical constraints such as actuator saturation, making it difficult to be directly and safely implemented. Finally, the control accuracy, robustness, and application range of the existing technology are significantly insufficient when dealing with extremely complex compound disturbance scenarios.

[0004] Therefore, the present application provides a compound disturbance compensation control method based on online learning and multi-model fusion. SUMMARY

[0005] In order to compensate for the deficiencies of the prior art and solve at least one of the technical problems proposed in the background art.

[0006] The technical scheme adopted by the present application to solve its technical problems is: the composite interference compensation control method based on online learning and multi-model fusion, comprising the following steps:

[0007] S1, a discrete-time state space model of a nonlinear dynamic system affected by composite interference is established, including a system state equation, a control input decomposition equation and a system output equation;

[0008] S2, a nonlinear disturbance observer is designed to estimate the lumped disturbance suffered by the system in real time;

[0009] S3, the lumped disturbance estimate value is decomposed in real time to obtain a plurality of disturbance components with different dynamic characteristics;

[0010] S4, for the dynamic characteristics of each type of disturbance component obtained by decomposition, a corresponding special compensator is designed to form a parallel multi-model compensation architecture;

[0011] S5, the intelligent arbitrator dynamically fuses the outputs of the three compensators according to the current disturbance characteristics to generate a composite interference compensation control vector;

[0012] S6, the composite interference compensation control vector and the nominal control vector are added to obtain the final total control input vector, which is applied to the controlled object after amplitude limiting processing by the actuator.

[0013] Preferably, the specific implementation steps of S1 are as follows:

[0014] Suppose that the nonlinear dynamic system affected by composite interference has a discrete-time state space model represented as:

[0015] x(k+1)=f(x(k),u(k))+d(k)

[0016] u(k)=u n (k)+u c (k)

[0017] y(k)=m(x(k))

[0018] Preferably, the specific implementation steps of S2 are as follows:

[0019] In order to estimate the lumped disturbance d(k) in real time, the following nonlinear disturbance observer is designed:

[0020]

[0021] z(k+1)=z(k)+τ{-L(x(k))z(k)+L(x(k))[x(k+1)-f(x(k),u(k))-p(x(k))]}

[0022] Preferably, the implementation steps of S3 are as follows:

[0023] By using adaptive signal processing techniques, the interference signal d (k) is decomposed into three interference components: including slow varying component d s (k), periodic component d p (k) and random residual component d r (k).

[0024] Preferably, the method for obtaining the interference components is as follows:

[0025] Slow varying component:

[0026] A first-order low-pass filter LPF is used to extract the slow varying trend:

[0027]

[0028] wherein α ∈ (0, 1) is the filter coefficient, which is adaptively adjusted according to the slow varying degree of the interference, and the adjustment criterion is:

[0029]

[0030] Periodic component:

[0031] The periodic component is extracted from the residual d (k), and online autocorrelation analysis and fast Fourier transform FFT are used to detect the main period T p (k);

[0032] Autocorrelation function calculation:

[0033]

[0034] An adaptive comb filter is designed to extract the periodic component:

[0035]

[0036] Random residual component:

[0037] Preferably, the implementation steps of S4 are as follows:

[0038] For the three interference components, three special compensators are designed: slow varying interference compensator, periodic interference compensator and random residual interference compensator.

[0039] Preferably, the operation logic of the special compensator is as follows:

[0040] Slow varying interference compensator:

[0041]

[0042] Integral gain K I (k) Adaptive adjustment based on error:

[0043] K I (k)=K I0 +λ·tanh(μ|e s (k)|)

[0044] Periodic interference compensator:

[0045] Based on the internal model principle, design a discrete-time repetitive control mechanism:

[0046]

[0047] Random residual disturbance compensator:

[0048] Real-time feedforward compensation is performed using an online sequence extreme learning machine (OS-ELM), whose network structure is a single hidden layer feedforward neural network (SLFN).

[0049] Input vector: composed of random residuals from the past n time steps, x r (k)=[d r (k), d r (k-1), ..., d r (k-n+1)] T ;

[0050] Output layer: β(k)∈R D Online updates, network output is scalar;

[0051] At time k, the current compensation control quantity is calculated using the weight β(k-1) from the previous time step:

[0052]

[0053] The online learning algorithm uses Recursive Least Squares (RLS) to update the network output weights.

[0054] β(k)=β(k-1)+P(k)h(k)[d r (k)-h T (k)β(k-1)]

[0055]

[0056] Preferably, the specific implementation steps of S5 are as follows:

[0057] The outputs of the three dedicated compensators are dynamically weighted and fused using an intelligent arbitrator to generate a total composite disturbance compensation control vector. The calculation formula is as follows:

[0058] u c (k)=α s(k)u s (k)+α p (k)u p (k)+α r (k)u r (k)

[0059] Preferably, the dynamic weighting coefficient α s (k), α p (k), α r (k) Generate online using the following method:

[0060] Feature extraction: Based on interference estimates And its decomposition components, calculate the following characteristic quantities characterizing the interference properties:

[0061] Slowly varying interference energy ratio:

[0062] Periodic interference energy ratio:

[0063] Random residual disturbance energy ratio:

[0064] Total interference energy:

[0065] rate of change of disturbance:

[0066] Arbitration decision: The feature vector The input is fed into a lightweight neural network, which is trained offline on historical data and performs only forward computation online. Its output layer is normalized using the Softmax function to obtain the initial weight vector.

[0067] Output smoothing: To avoid control chattering caused by abrupt weight changes, a first-order inertial smoothing filter is applied to the initial weights of the neural network output.

[0068]

[0069] Preferably, the specific implementation steps of S6 are as follows:

[0070] Integrate the composite disturbance compensation control vector and the nominal control vector into the total control input vector:

[0071] u(k)=u n (k)+u c (k)

[0072] To ensure that the control input remains within the feasible range of the physical actuator, a saturation limit is applied to the total control input:

[0073] u(k)=sat(u(k), u min u max )

[0074] The total control input u(k), after being limited, is output to the actual actuator to complete the closed-loop operation of the current control cycle.

[0075] The beneficial effects of this invention are as follows:

[0076] 1. The composite interference compensation control method based on online learning and multi-model fusion described in this invention uses adaptive signal processing technology to decompose lumped interference into three characteristic components in real time: slow-varying, periodic, and random residuals. A dedicated compensator is designed for each type of component: adaptive integral compensation is used for slow-varying interference, repetitive control is used for periodic interference, and online extreme learning machine feedforward compensation is used for random and unmodeled dynamics. This "differentiated treatment" architecture ensures that all types of interference can be handled by the most suitable compensation mechanism, thereby significantly improving the overall interference suppression accuracy, response speed, and steady-state performance of the system.

[0077] 2. The composite interference compensation control method based on online learning and multi-model fusion described in this invention deeply embeds an online learning mechanism throughout the entire process of interference perception, decomposition, and compensation decision-making: In the interference decomposition stage, the filter coefficients are adjusted online according to the interference change rate; in the period extraction stage, the dominant period is detected in real time through online autocorrelation analysis; in the random interference compensation stage, the network weights are updated in real time using an online sequential extreme learning machine combined with recursive least squares; in the fusion decision-making stage, the weights are dynamically generated based on real-time interference characteristics through a lightweight neural network. This enables the system to continuously adapt to changes in interference characteristics, exhibiting strong robustness and environmental adaptability to unknown, time-varying, and sudden interferences, significantly outperforming traditional control methods with fixed parameters.

[0078] 3. The composite interference compensation control method based on online learning and multi-model fusion described in this invention introduces a lightweight intelligent arbitrator. The system can calculate the optimal fusion weight of each compensator output online based on the multi-dimensional characteristics of the current interference, such as the spectral energy distribution and rate of change. After smoothing and filtering, the composite compensation amount is generated. This mechanism acts as an intelligent command center, ensuring that control resources can be dynamically and coordinated when multiple interferences coexist or their primary and secondary aspects change. This fully leverages the strengths of each compensator while avoiding system oscillations caused by sudden weight changes or conflicts. As a result, the system can maintain stable, smooth, and coordinated high-performance operation even in multi-interference coupling scenarios. Attached Figure Description

[0079] The invention will now be further described with reference to the accompanying drawings.

[0080] Figure 1 This is a schematic diagram of the method flow of the present invention. Detailed Implementation

[0081] To make the technical means, creative features, objectives and effects of this invention easier to understand, the invention will be further described below in conjunction with specific embodiments.

[0082] like Figure 1 As shown, the composite disturbance compensation control method based on online learning and multi-model fusion described in this invention includes the following steps:

[0083] S1. Establish a discrete-time state-space model of a nonlinear dynamic system affected by compound disturbances, including the system state equations, control input decomposition equations, and system output equations.

[0084] S2. Design a nonlinear disturbance observer to estimate the lumped disturbances experienced by the system in real time online;

[0085] S3. Perform real-time decomposition on the lumped interference estimate to obtain multiple interference components with different dynamic characteristics;

[0086] S4. For the dynamic characteristics of each type of interference component obtained by decomposition, design a corresponding dedicated compensator to form a parallel multi-model compensation architecture.

[0087] S5. The intelligent arbitrator dynamically fuses the outputs of the three compensators based on the current interference characteristics to generate a composite interference compensation control vector.

[0088] S6. Add the composite disturbance compensation control vector to the nominal control vector to obtain the final total control input vector, and apply it to the controlled object after being limited by the actuator.

[0089] As one embodiment of the present invention, the specific implementation steps of S1 are as follows:

[0090] Consider a nonlinear dynamic system subjected to combined disturbances. Its discrete-time state-space model can be expressed as:

[0091] x(k+1) = f(x(k), u(k)) + d(k)

[0092] u(k)=u n (k)+u c (k)

[0093] y(k)=m(x(k))

[0094] Where k is the discrete-time index; x(k) is the system state vector; u(k) is the total control input vector; u n (k) is the nominal control vector; u c(k) is the composite disturbance compensation control vector; y(k) is the system output vector; f(·) and m(·) are the known nominal system dynamic and output functions; d(k) is the lumped disturbance vector acting on the system, which contains various types and characteristics of disturbance components, and may change over time and be coupled with each other.

[0095] As one embodiment of the present invention, the specific implementation steps of S2 are as follows:

[0096] To estimate the lumped disturbance d(k) in real time, the following nonlinear disturbance observer (NDOB) is designed:

[0097]

[0098] z(k+1)=z(k)+τ{-L(x(k))z(k)+L(x(k))[x(k+1)-f(x(k),u(k))-p(x(k))]}

[0099] in, τ is the disturbance estimate; z(k) is the internal state of the observer; p(x(k)) is the nonlinear function to be designed, which can usually be simplified to the linear form p(x) = L0x, where L0 is a constant design matrix of appropriate dimension; L(x(k)) is the observer gain matrix, which is designed using the Lyapunov method to ensure convergence; τ is the sampling time.

[0100] This interference observer is responsible for observing the type of interference the system is experiencing in real time and accurately, providing the most critical input signal for all subsequent processing.

[0101] As one embodiment of the present invention, the specific implementation steps of S3 are as follows:

[0102] Using adaptive signal processing techniques, Decomposed into three disturbance components: including the slowly varying component d s (k), periodic component d p (k) and random residual components d r (k);

[0103] The method for obtaining the interference components is as follows:

[0104] Slowly varying components:

[0105] A first-order low-pass filter (LPF) is used to extract slowly varying trends:

[0106]

[0107] Where α∈(0,1) are the filter coefficients, which are adaptively adjusted according to the slow variation of the interference. The adjustment criterion is:

[0108]

[0109] in, The gradient is the disturbance estimate; α0 and γ are design parameters; when the disturbance changes slowly ( When the interference is small, α is decreased to enhance the filtering effect and better extract the slow-changing components. When the interference changes drastically, α is increased to track quickly and avoid lag.

[0110] The significance of extracting slowly varying components lies in separating out interference caused by long-term, slow changes such as equipment aging and environmental temperature drift.

[0111] Periodic components:

[0112] From residuals Periodic components were extracted, and the main period T was detected using online autocorrelation analysis and Fast Fourier Transform (FFT). p (k);

[0113] Autocorrelation function calculation:

[0114]

[0115] Where N is the sliding window length, and the current dominant period T is determined by finding the peak value of R(τ, k). p (k);

[0116] Design an adaptive comb filter to extract the periodic component:

[0117]

[0118] Where M is the number of cycles, usually taken as 1 or 2;

[0119] The significance of extracting periodic components lies in separating out regular and periodic disturbances caused by rotating machinery, reciprocating motion, etc.

[0120] Random residual components:

[0121]

[0122] The significance of extracting the random residual component lies in capturing all remaining disturbances, including random impulses, burst noise, and unmodeled dynamics not accurately described by the first two components. It is a containment item of the "decomposition-compensation" framework, ensuring that the total disturbance is fully processed.

[0123] As one embodiment of the present invention, the specific implementation steps of S4 are as follows:

[0124] Three dedicated compensators are designed for the three interference components: a slowly varying interference compensator, a periodic interference compensator, and a random residual interference compensator.

[0125] The operating logic of the dedicated compensator is as follows:

[0126] Slow-varying interference compensator:

[0127] This compensator is used to counteract slowly varying interference components caused by equipment aging, environmental temperature drift, etc., and its output is:

[0128]

[0129] Among them, u s (k) represents the slow-varying disturbance compensation control component; This is the estimation error for slowly varying disturbances; The observed values ​​of the slowly varying components can be obtained from the slowly varying disturbance observer;

[0130] Integral gain K I (k) Adaptive adjustment based on error:

[0131] K I (k)=K I0 +λ·tanh(μ|e s (k)|)

[0132] Among them, K I0 λ and μ are design parameters that increase the gain when the error is large to speed up the response, and decrease the gain when the error is small to avoid overshoot.

[0133] Periodic interference compensator:

[0134] Based on the internal model principle, design a discrete-time repetitive control mechanism:

[0135]

[0136] Among them, u p (k) represents the periodic disturbance compensation control component; N(k) = round(T) p (k) / T s ) represents the delay step number corresponding to the current period; T s The sampling time is denoted by ; Q(z) is a low-pass filter used to enhance robustness, typically set to . This is due to periodic interference tracking error;

[0137] Random residual disturbance compensator:

[0138] Real-time feedforward compensation is performed using an online sequence extreme learning machine (OS-ELM), whose network structure is a single hidden layer feedforward neural network (SLFN).

[0139] Input vector: composed of random residuals from the past n time steps, x r(k)=[d r (k), d r (k-1), ..., d r (k-n+1)] T ;

[0140] Hidden layer: D nodes in total, activation function g(·) uses Sigmoid or RBF, input weights w j ∈R n With bias b j (j = 1, ..., D) are randomly generated during initialization and remain fixed.

[0141] Output layer: β(k)∈R D Online updates, network output is scalar;

[0142] At time k, the current compensation control quantity is calculated using the weight β(k-1) from the previous time step:

[0143]

[0144] That is, the compensation amount is taken as the negative value of the current random residual prediction value to achieve feedforward cancellation;

[0145] Among them, u r (k) represents the random residual disturbance compensation control component; D is the number of hidden layer nodes; g(·) is the activation function; w j and b j The input weights and biases are randomly initialized and fixed;

[0146] The online learning algorithm uses recursive least squares (RLS) to update the network output weights:

[0147] β(k)=β(k-1)+P(k)h(k)[d r (k)-h T (k)β(k-1)]

[0148]

[0149] in, P(k)∈R is the hidden layer output vector. D×D Let be the covariance matrix.

[0150] As one embodiment of the present invention, the specific implementation steps of S5 are as follows:

[0151] The outputs of the three dedicated compensators are dynamically weighted and fused using an intelligent arbitrator to generate a total composite disturbance compensation control vector. The calculation formula is as follows:

[0152] u c (k)=α s (k)us (k)+α p (k)u p (k)+α r (k)u r (k)

[0153] Wherein, the dynamic weighting coefficient α s (k), α p (k), α r (k)∈[0,1], and satisfy α s (k)+α p (k)+α r (k) = 1.

[0154] The dynamic weighting coefficient α s (k), α p (k), α r (k) Generate online using the following method:

[0155] Feature extraction: Based on interference estimates And its decomposition components, calculate the following characteristic quantities characterizing the interference properties:

[0156] Slowly varying interference energy ratio:

[0157] Periodic interference energy ratio:

[0158] Random residual disturbance energy ratio:

[0159] Total interference energy:

[0160] rate of change of disturbance:

[0161] Arbitration decision: The feature vector (in For the rate of change of disturbance The input (norm of the vector) is fed into a lightweight neural network (e.g., a three-layer perceptron). This network is trained offline on historical data and performs only forward computation online. Its output layer is normalized using the Softmax function to obtain the initial weight vector.

[0162] Output smoothing: To avoid control chattering caused by abrupt weight changes, a first-order inertial smoothing filter is applied to the initial weights of the neural network output.

[0163]

[0164] Where η∈(0,1) is the smoothing coefficient, used to adjust the speed of weight updates; i=s,p,r.

[0165] This step, as the core decision-making unit of the control method, analyzes the real-time dynamic characteristics of the disturbance online, adaptively allocates the contribution weights of each dedicated compensator, coordinates their outputs through a smoothing mechanism, and finally merges them to generate a precise composite disturbance compensation control quantity.

[0166] As one embodiment of the present invention, the specific implementation steps of S6 are as follows:

[0167] Integrate the composite disturbance compensation control vector and the nominal control vector into the total control input vector:

[0168] u(k)=u n (k)+u c (k)

[0169] To ensure that the control input remains within the feasible range of the physical actuator, a saturation limit is applied to the total control input:

[0170] u(k)=sat(u(k), u min u max )

[0171] Where, sat(·) is the component-wise saturation function; u min and u max These are the minimum and maximum control input vectors allowed by the actuator, respectively;

[0172] The total control input u(k), after being limited, is output to the actual actuator to complete the closed-loop operation of the current control cycle.

[0173] The method of this invention features a modular structure and clear hierarchy. From interference observation, decomposition, parallel compensation to intelligent fusion, each step has a clear mathematical description and implementation path. The components used, such as the nonlinear interference observer, digital filter, repetitive controller, and lightweight neural network, all have mature digital implementation foundations, manageable computational burden, and are easy to implement in embedded systems or industrial control platforms. Furthermore, the final control quantity undergoes saturation limiting processing, balancing theoretical advancement with safety constraints in engineering practice, and has promising prospects for practical application.

[0174] This invention is particularly applicable to complex industrial environments with multiple coupled interferences, such as chemical processes, bio-fermentation, precision machining, and aerospace servo systems. Its "targeted and dynamic coordination" design concept effectively addresses complex interference problems caused by equipment aging, load cycle fluctuations, raw material mutations, and external shocks, providing a practical technical path to improve the control precision, operational reliability, and intelligence level of high-end equipment and complex processes.

[0175] The terms "front," "back," "left," "right," "top," and "bottom" all refer to the figures in the accompanying drawings. Figure 1Based on the perspective of the observer, the side of the device facing the observer is defined as the front, the left side of the observer is defined as the left, and so on.

[0176] In the description of this invention, it should be understood that the terms "center", "longitudinal", "lateral", "front", "rear", "left", "right", "vertical", "horizontal", "top", "bottom", "inner", "outer", etc., indicate the orientation or positional relationship based on the orientation or positional relationship shown in the accompanying drawings. They are only for the convenience of describing this invention and simplifying the description, and do not indicate or imply that the device or element referred to must have a specific orientation, or be constructed and operated in a specific orientation. Therefore, they should not be construed as limiting the scope of protection of this invention.

[0177] The foregoing has shown and described the basic principles, main features, and advantages of the present invention. Those skilled in the art should understand that the present invention is not limited to the above embodiments. The embodiments and descriptions in the specification are merely illustrative of the principles of the invention. Various changes and modifications can be made to the invention without departing from its spirit and scope, and all such changes and modifications fall within the scope of the present invention as claimed. The scope of protection of the present invention is defined by the appended claims and their equivalents.

Claims

1. A composite disturbance compensation control method based on online learning and multi-model fusion, characterized in that: Includes the following steps: S1. Establish a discrete-time state-space model of a nonlinear dynamic system affected by compound disturbances, including the system state equations, control input decomposition equations, and system output equations. S2. Design a nonlinear disturbance observer to estimate the lumped disturbances experienced by the system in real time online; S3. Perform real-time decomposition on the lumped interference estimate to obtain multiple interference components with different dynamic characteristics; S4. For the dynamic characteristics of each type of interference component obtained by decomposition, design a corresponding dedicated compensator to form a parallel multi-model compensation architecture. S5. The intelligent arbitrator dynamically fuses the outputs of the three compensators based on the current interference characteristics to generate a composite interference compensation control vector. S6. Add the composite disturbance compensation control vector to the nominal control vector to obtain the final total control input vector, and apply it to the controlled object after being limited by the actuator.

2. The composite disturbance compensation control method based on online learning and multi-model fusion according to claim 1, characterized in that: The specific implementation steps of S1 are as follows: Consider a nonlinear dynamic system subjected to combined disturbances. Its discrete-time state-space model can be expressed as: x(k+1) = f(x(k), u(k)) + d(k) u(k)=u n (k)+u c (k) y(k)=m(x(k)) Where k is the discrete-time index; x(k) is the system state vector; u(k) is the total control input vector; u n (k) is the nominal control vector; u c (k) is the composite disturbance compensation control vector; y(k) is the system output vector; f(·) and m(·) are the known nominal system dynamics and output functions; d(k) is the lumped disturbance vector acting on the system.

3. The composite disturbance compensation control method based on online learning and multi-model fusion according to claim 2, characterized in that: The specific implementation steps of S2 are as follows: To estimate the lumped disturbance d(k) in real time, the following nonlinear disturbance observer is designed: z(k+1)=z(k)+τ{-L(x(k))z(k)+L(x(k))[x(k+1)-f(x(k),u(k))-p(x(k))]} in, τ is the disturbance estimate; z(k) is the internal state of the observer; p(x(k)) is the nonlinear function to be designed; L(x(k)) is the observer gain matrix; τ is the sampling time.

4. The composite disturbance compensation control method based on online learning and multi-model fusion according to claim 3, characterized in that: The specific implementation steps of S3 are as follows: Using adaptive signal processing techniques, Decomposed into three disturbance components: including the slowly varying component d s (k), periodic component d p (k) and random residual components d r (k).

5. The composite disturbance compensation control method based on online learning and multi-model fusion according to claim 4, characterized in that: The method for obtaining the interference components is as follows: Slowly varying components: Slowly varying trends are extracted using a first-order low-pass filter (LPF). Where α∈(0,1) are the filter coefficients, which are adaptively adjusted according to the slow variation of the interference. The adjustment criterion is: in, The gradient is the disturbance estimate; α0 and γ are the design parameters. Periodic components: From residuals Periodic components were extracted, and the main period T was detected using online autocorrelation analysis and Fast Fourier Transform (FFT). p (k); Autocorrelation function calculation: Where N is the sliding window length, and the current dominant period T is determined by finding the peak value of R(τ, k). p (k); Design an adaptive comb filter to extract the periodic component: Where M is the number of cycles; Random residual components:

6. The composite disturbance compensation control method based on online learning and multi-model fusion according to claim 5, characterized in that: The specific implementation steps of S4 are as follows: Three dedicated compensators are designed for the three interference components: a slowly varying interference compensator, a periodic interference compensator, and a random residual interference compensator.

7. The composite disturbance compensation control method based on online learning and multi-model fusion according to claim 6, characterized in that: The operating logic of the dedicated compensator is as follows: Slow-varying interference compensator: Among them, u s (k) represents the slow-varying disturbance compensation control component; This is the estimation error for slowly varying disturbances; These are the observations of the slowly varying components; Integral gain K I (k) Adaptive adjustment based on error: K I (k)=K I0 +λ·tanh(μ|e s (k)|) Among them, K I0 λ and μ are design parameters; Periodic interference compensator: Based on the internal model principle, design a discrete-time repetitive control mechanism: Among them, u p (k) represents the periodic disturbance compensation control component; N(k) = round(T) p (k) / T s ) represents the delay step number corresponding to the current period; T s Where z is the sampling time; Q(z) = 0.25 + 0.5z -1 +0.25z -2 It is a low-pass filter used to enhance robustness; This is due to periodic interference tracking error; Random residual disturbance compensator: Real-time feedforward compensation is performed using an online sequence extreme learning machine (OS-ELM), whose network structure is a single hidden layer feedforward neural network (SLFN). Input vector: composed of random residuals from the past n time steps, x r (k)=[d r (k), d r (k-1), ..., d r (k-n+1)] T ; Output layer: β(k)∈R D Online updates, network output is scalar; At time k, the current compensation control quantity is calculated using the weight β(k-1) from the previous time step: Among them, u r (k) represents the random residual disturbance compensation control component; D is the number of hidden layer nodes; g(·) is the activation function; w j and b j The input weights and biases are randomly initialized and fixed; The online learning algorithm uses Recursive Least Squares (RLS) to update the network output weights. β(k)=β(k-1)+P(k)h(k)[d r (k)-h T (k)β(k-1)] in, P(k)∈R is the hidden layer output vector. D×D Let be the covariance matrix.

8. The composite disturbance compensation control method based on online learning and multi-model fusion according to claim 7, characterized in that: The specific implementation steps of S5 are as follows: The outputs of the three dedicated compensators are dynamically weighted and fused using an intelligent arbitrator to generate the total composite disturbance compensation control vector. The calculation formula is as follows: you c (k)=a s (k)u s (k)+a p (k)u p (k)+a r (k)u r (k) Wherein, the dynamic weighting coefficient α s (k), α p (k), α r (k)∈[0,1], and satisfy α s (k)+α p (k)+α r (k) = 1.

9. The composite disturbance compensation control method based on online learning and multi-model fusion according to claim 8, characterized in that: The dynamic weighting coefficient α s (k), α p (k), α r (k) Generate online using the following method: Feature extraction: Based on interference estimates And its decomposition components, calculate the following characteristic quantities characterizing the interference properties: Slowly varying interference energy ratio: Periodic interference energy ratio: Random residual disturbance energy ratio: Total interference energy: rate of change of disturbance: Arbitration decision: The feature vector [E] s (k), E p (k), E r (k), E total (k), The input is fed into a lightweight neural network, which is trained offline on historical data and performs only forward computation online. Its output layer is normalized using the Softmax function to obtain the initial weight vector. Output smoothing: To avoid control chattering caused by abrupt weight changes, a first-order inertial smoothing filter is applied to the initial weights of the neural network output. Where η∈(0,1) is the smoothing coefficient; i=s,p,r.

10. The composite disturbance compensation control method based on online learning and multi-model fusion according to claim 9, characterized in that: The specific implementation steps of S6 are as follows: Integrate the composite disturbance compensation control vector and the nominal control vector into the total control input vector: u(k)=u n (k)+u c (k) To ensure that the control input remains within the feasible range of the physical actuator, a saturation limit is applied to the total control input: u(k)=sat(u(k),u min ,in max ) Where, sat(·) is the component-wise saturation function; u min and u max These are the minimum and maximum control input vectors allowed by the actuator, respectively; The total control input u(k), after being limited, is output to the actual actuator to complete the closed-loop operation of the current control cycle.

Images