Deep learning optimization method, system and device based on active disturbance rejection control and medium

Abstract

The invention relates to the technical field of deep learning, in particular to a deep learning optimization method, system and device based on active disturbance rejection control and a medium, and the method comprises the steps: obtaining a model parameter of a deep learning model in a current iteration step and a control input value of the deep learning model in a last iteration step in a deep learning model iteration training process, inputting an extended state observer, and outputting a state estimation value and a total disturbance estimation value by the extended state observer; inputting the state estimation value and the total disturbance estimation value into an active-disturbance-rejection controller, and calculating a control input value of the deep learning model in the current iteration step by the active-disturbance-rejection controller; according to the control input value of the current iteration step and the model parameter of the current iteration step, the model parameter of the deep learning model is updated, and the updated model parameter is used as the model parameter of the deep learning model in the next iteration step, so that the robustness of the optimization process to the complex environment and interference is enhanced, and the optimization efficiency is improved. And the training efficiency and performance of the deep learning model are obviously improved.

Description

Technical Field

[0001] This application relates to the field of deep learning technology, and in particular to a deep learning optimization method, system, device and medium based on active disturbance rejection control. Background Technology

[0002] With the rapid development of artificial intelligence technology, deep neural networks automatically extract high-level features from data through multi-layer nonlinear transformations, demonstrating powerful representation learning capabilities on large-scale datasets. Training a deep learning model is essentially a large-scale optimization problem, aiming to find a set of model parameters (such as the weights and biases of the neural network) that minimizes a predefined loss function. The loss function landscape of deep learning is typically extremely complex, exhibiting characteristics such as high nonlinearity, non-convexity, multiple local minima, saddle points, and flat regions. Furthermore, the training process often uses mini-batch data to calculate gradients, introducing significant random noise (gradient noise). These factors pose significant challenges to optimization algorithms, potentially leading to slow convergence during training, stagnation at local optima or saddle points, sensitivity to hyperparameters such as initial parameters and learning rate, and oscillations in the optimization path.

[0003] To improve optimization performance, related techniques generally involve introducing feedback control mechanisms from control theory into the deep learning optimization process. One representative approach is to analogize the parameter update process to a dynamic system control problem and design the optimizer using a PID (Proportional-Integral-Derivative) controller. For example, optimization algorithms based on PID control laws (such as PIDAO) treat the gradient signal as a system error, utilizing the proportional term to respond to the current error, the integral term to eliminate the steady-state error, and the derivative term to predict the trend of change, aiming to improve the dynamic response characteristics of the optimization process.

[0004] However, these methods still face significant drawbacks in practical applications: First, PID parameters (K_p, Ki, K_d) are difficult to tune, lack systematic guidance, and require readjustment for different models and tasks; second, the performance of the PID controller depends on the accuracy of the assumptions about the model of the controlled object (optimization process), while the dynamic characteristics of the optimization process are complex and variable and difficult to model accurately; finally, PID has limited ability to suppress strong nonlinearity and time-varying disturbances (such as gradient noise and drastic changes in the shape of the loss function), and its robustness needs to be improved. Summary of the Invention

[0005] This application provides a deep learning optimization method, system, device, and medium based on active disturbance rejection control, to solve the problems of deep learning optimization methods in related technologies being sensitive to gradient noise and loss function complexity during training, resulting in unstable optimization paths, difficulty in convergence, and poor robustness and adaptability.

[0006] The first aspect of this application provides a deep learning optimization method based on active disturbance rejection control, comprising the following steps: during the iterative training of a deep learning model, obtaining the model parameters of the deep learning model in the current iteration step and the control input value of the previous iteration step; inputting the model parameters of the current iteration step and the control input value of the previous iteration step into an extended state observer, the extended state observer outputting a state estimate and a total disturbance estimate; inputting the state estimate and the total disturbance estimate into an active disturbance rejection controller, the active disturbance rejection controller calculating the control input value of the deep learning model in the current iteration step; updating the model parameters of the deep learning model according to the control input value of the current iteration step and the model parameters of the current iteration step, and using the updated model parameters as the model parameters of the deep learning model in the next iteration step.

[0007] Optionally, before obtaining the model parameters of the deep learning model in the current iteration step and the control input values ​​of the previous iteration step, the process includes: establishing a dynamic system model of the deep learning model optimization process, treating the model parameters of the deep learning model to be optimized as state variables of the dynamic system model, and using the dynamic system model to obtain the model parameters of the current iteration step and the control input values ​​of the previous iteration step.

[0008] Optionally, the calculation formula for the dynamic system model is:

[0009] in, For model parameter vectors, The rate of change of the parameter over time. The acceleration of the parameter change over time. loss function Regarding parameters gradient, For gradient noise, To control the gain coefficient, To control the input signal.

[0010] Optionally, the calculation formula for the extended state observer is:

[0011]

[0012]

[0013]

[0014] in, For observation error, , Rate of change of parameter Real-time estimates, Total disturbance Real-time estimates, Here, is the observer gain, and is a positive constant. for , A pre-defined nonlinear function.

[0015] Optionally, the active disturbance rejection controller is equipped with an active disturbance rejection control law algorithm, the calculation formula of which is:

[0016] in, It is the active disturbance rejection controller's estimate of the total disturbance. It is a disturbance compensation term. It is a nominal control law calculated based on state error feedback.

[0017] Optionally, the model parameters of the deep learning model are updated based on the control input value of the current iteration step and the model parameters of the current iteration step, including: determining the direction and / or magnitude of the parameter update using the control input value of the current iteration step; updating the model parameters of the current iteration step according to the direction and / or magnitude of the parameter update to obtain the updated model parameters.

[0018] Optionally, before inputting the model parameters of the current iteration step and the control input values ​​of the previous iteration into the extended state observer, the method further includes: obtaining the gradient of the loss function of the deep learning model; inputting the gradient of the loss function into the extended state observer, and using the gradient of the loss function to assist the extended state observer in calculation.

[0019] A second aspect of this application provides a deep learning optimization system based on active disturbance rejection control, comprising: an interface module for acquiring model parameters of a deep learning model in the current iteration step and control input values ​​of the previous iteration step; an extended state observer for estimating state estimates and total disturbance estimates of the model parameters based on the model parameters of the current iteration step and the control input values ​​of the previous iteration step; a control law calculation module for calculating the control input values ​​of the deep learning model in the current iteration step based on the state estimates and the total disturbance estimates; and an update module for updating the model parameters of the deep learning model based on the control input values ​​of the current iteration step and the model parameters of the current iteration step, and using the updated model parameters as the model parameters of the deep learning model in the next iteration step.

[0020] A third aspect of this application provides an electronic device, including: a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor executes the program to implement a deep learning optimization method based on active disturbance rejection control as described in the above embodiments.

[0021] A fourth aspect of this application provides a computer-readable storage medium having a computer program stored thereon, which is executed by a processor to implement the deep learning optimization method based on active disturbance rejection control as described in the above embodiments.

[0022] Therefore, this application has the following beneficial effects: This application's embodiments model the parameter optimization process of a deep learning model as a disturbed dynamic system. By extending the state observer, it estimates and actively compensates for the state estimates and total disturbance estimates that affect the optimization process in real time. The disturbance is actively compensated in the active disturbance rejection controller, generating the control input value of the deep learning model in the current iteration step. The model parameters are updated using the control input value, which enhances the robustness of the optimization process to complex environments and disturbances. It effectively solves the problems of slow convergence, poor stability, sensitivity to hyperparameters, and weak anti-interference ability in related technologies under complex loss terrain, and significantly improves the training efficiency and performance of deep learning models.

[0023] Additional aspects and advantages of this application will be set forth in part in the description which follows, and in part will be obvious from the description, or may be learned by practice of this application. Attached Figure Description

[0024] The above and / or additional aspects and advantages of this application will become apparent and readily understood from the following description of the embodiments taken in conjunction with the accompanying drawings, wherein: Figure 1 This is a flowchart of a deep learning optimization method based on active disturbance rejection control provided in an embodiment of this application; Figure 2 This is an example diagram of a deep learning optimization system based on active disturbance rejection control according to an embodiment of this application; Figure 3 This is a detailed structural diagram of an extended state observer provided according to an embodiment of this application; Figure 4 This is a detailed structural diagram of a control law calculation module provided according to an embodiment of this application; Figure 5 This is a schematic diagram of the structure of an electronic device according to an embodiment of this application. Detailed Implementation

[0025] The embodiments of this application are described in detail below. Examples of the embodiments are shown in the accompanying drawings, wherein the same or similar reference numerals denote the same or similar elements or elements having the same or similar functions throughout. The embodiments described below with reference to the accompanying drawings are exemplary and intended to explain this application, and should not be construed as limiting this application.

[0026] The following description, with reference to the accompanying drawings, outlines a deep learning optimization method, system, device, and medium based on active disturbance rejection control (ADRC) according to embodiments of this application. Addressing the problems mentioned in the background section, this application provides a deep learning optimization method based on ADRC. In this method, the parameter optimization process of the deep learning model is modeled as a disturbed dynamic system. An extended state observer is used to estimate and actively compensate for the state estimates and total disturbance estimates affecting the optimization process in real time. The ADRC actively compensates for the disturbance, generating control input values ​​for the deep learning model at the current iteration step. These control input values ​​are then used to update the model parameters, enhancing the robustness of the optimization process to complex environments and disturbances. This effectively solves the problems of slow convergence, poor stability, sensitivity to hyperparameters, and weak anti-interference ability in related technologies under complex loss terrain, significantly improving the training efficiency and performance of the deep learning model.

[0027] Specifically, Figure 1 This is a flowchart illustrating a deep learning optimization method based on active disturbance rejection control provided in an embodiment of this application.

[0028] like Figure 1 As shown, the deep learning optimization method based on active disturbance rejection control includes the following steps: In step S101, during the iterative training of the deep learning model, the model parameters of the deep learning model in the current iteration step and the control input values ​​of the previous iteration step are obtained.

[0029] Deep learning models include, but are not limited to, one or a combination of: Convolutional Neural Network (CNN), Recurrent Neural Network (RNN), Long Short-Term Memory Network (LSTM), Bi-Long Short-Term Memory Network (Bi-LSTM), and Transformer model.

[0030] In one embodiment of this application, before obtaining the model parameters of the deep learning model in the current iteration step and the control input values ​​of the previous iteration step, the method includes: establishing a dynamic system model of the deep learning model optimization process, treating the model parameters of the deep learning model to be optimized as state variables of the dynamic system model, and using the dynamic system model to obtain the model parameters of the current iteration step and the control input values ​​of the previous iteration step.

[0031] The embodiments of this application can optimize the parameters of the deep learning model to be optimized ( ) are considered as state variables in a dynamic system model; a parameter influencing the state variables is defined. Update the total disturbance of all unmodeled dynamics and disturbances ( ).

[0032] It is understood that, in the embodiments of this application, the parameters of the deep learning model to be optimized (e.g., a vector composed of all weights and biases) can be denoted as... It is considered as an iteration of optimization (discrete time) or continuous time The state of an evolving dynamic system. The optimization process can be abstracted into a controlled dynamic system model; for example, a second-order system model can be represented as:

[0033] in, For model parameter vectors, The rate of change (velocity) of the parameter over time. The acceleration of the parameter change over time. loss function Regarding parameters gradient, For gradient noise, The control gain coefficient represents the system's response to the control input. The degree of response is usually a known or estimated constant. The input signal is used to drive parameter updates. The "total disturbance" representing the system is an unknown but bounded function that aggregates all unmodeled dynamics affecting the parameter acceleration (such as coupling effects between parameters and the inherent damping characteristics of the model) and external disturbances (such as gradient noise introduced by mini-batch sampling). ), and the known state and gradient Related but difficult to model precisely nonlinear terms.

[0034] In step S102, the model parameters of the current iteration step and the control input values ​​of the previous iteration are input into the extended state observer, and the extended state observer outputs the state estimate and the total disturbance estimate.

[0035] Understandably, an extended state observer (ESO) is constructed for the aforementioned dynamic system model. The core task of the extended state observer (ESO) is to utilize the system's measurable outputs (typically current parameters). , recorded as ) and known control inputs To estimate the internal state of the system (parameters) in real time. and its derivative and unknown total disturbance estimates .

[0036] For a second-order system, a third-order extended state observer is typically designed, letting and the total disturbance Consider it a new expansion state Assuming ,in It is the unknown rate of change of the disturbance. The state vector of ESO is , respectively corresponding Estimated values. A typical extended state observer uses a nonlinear function to handle observation errors, with the following structure:

[0037]

[0038]

[0039]

[0040] in, For observation error, , Rate of change of parameter Real-time estimates, Total disturbance Real-time estimates, Here, is the observer gain, and is a positive constant. for Its value affects the estimation speed and noise resistance of ESO, and can usually be tuned using bandwidth parameterization methods. It is a selected nonlinear function (or a linear function). For example, the commonly used fal function: when hour, when The nonlinear function can be the fal function, which can generally improve the convergence speed of ESO when the error is large and the tracking accuracy when the error is small.

[0041] In each optimization iteration step , will the current parameter As and using the control input calculated in the previous moment. The state of the ESO is updated through numerical integration (such as the Euler method). .

[0042] In step S103, the state estimate and the total disturbance estimate are input to the active disturbance rejection controller, which calculates the control input value of the deep learning model in the current iteration step.

[0043] Among them, control input Includes a feedback control law part based on state error ( ) and an estimate for compensating the total disturbance ( The perturbation compensation part of the active disturbance rejection controller (ESO) in this application embodiment is the core of the ESO's real-time estimation and compensation of the total perturbation. This makes the optimization process highly resistant to gradient noise, nonlinearity / nonconvexity of the loss function, dynamic changes within the model, and other unmodeled factors, improving the stability and reliability of training. By actively eliminating the influence of perturbations, parameter updates can be made smoother and more direct towards the optimal solution, potentially achieving faster convergence speed, reducing oscillations during the optimization process, and possibly enhancing the ability to escape local optima and saddle points.

[0044] It is understood that the embodiments of this application may be based on the state estimates provided by ESO. Total disturbance estimation Design control input The core advantage of an active disturbance rejection controller (ADRC) lies in its control law structure, which can explicitly compensate for the estimated total disturbance. The ADRC incorporates an ADRC control law algorithm, the calculation formula of which is:

[0045] in, It is the active disturbance rejection controller's estimate of the total disturbance. It is a disturbance compensation term, whose function is to offset the effect of the total disturbance on the system acceleration in real time; It is a nominal control law calculated based on state error feedback, and its goal is to make the compensated system (approximately) It has the desired dynamic performance (e.g., fast and stable convergence to the optimal solution).

[0046] It should be noted that the active disturbance rejection control law algorithm in this embodiment does not depend on the precise mathematical model of the controlled object (optimization process), but only needs to know the approximate order and control gain of the system. It provides coarse information. This makes it well-suited for deep learning optimization scenarios where the model is complex and uncertain; In the embodiments of this application, It can be designed in various forms, for example, a nonlinear proportional-derivative (PD) controller:

[0047]

[0048]

[0049] in, Indicates positional error. Indicates speed error, and This refers to the expected parameter trajectory and its rate of change. In optimization problems, the objective is usually to minimize the loss function, which corresponds to the parameters stabilizing at a certain optimal value. ,Right now In practice, it is possible to set... Set to an implicit objective (driven by gradient information) or set to 0 (if the problem is normalizable). It is usually set to 0.

[0050] It is the controller gain, a positive constant, which determines the response speed and damping of the closed-loop system.

[0051] This is a selected (usually nonlinear) error feedback function, such as the FAL function, to achieve the characteristic of "small error, large gain; large error, small gain," thereby improving response speed and stability. A linear form can also be used. .

[0052] In each optimization iteration step Based on the current ESO output and the set target (or error definition) and controller parameters Calculate the current control input value This approach decouples disturbance estimation (ESO parameters) from control performance tuning (controller parameters). Parameter tuning typically employs systematic bandwidth-based methods, which may be more intuitive and efficient than the trial-and-error process of PID parameter tuning. Furthermore, due to its robustness, the embodiments in this application may exhibit lower sensitivity to hyperparameters (such as estimation accuracy or even the traditional learning rate).

[0053] In step S104, the model parameters of the deep learning model are updated according to the control input value and model parameters of the current iteration step, and the updated model parameters are used as the model parameters of the deep learning model in the next iteration step.

[0054] In one embodiment of this application, updating the model parameters of a deep learning model based on the control input value of the current iteration step and the model parameters of the current iteration step includes: determining the direction and / or magnitude of the parameter update using the control input value of the current iteration step; updating the model parameters of the current iteration step according to the direction and / or magnitude of the parameter update to obtain the updated model parameters.

[0055] Specifically, the control input is calculated by estimating the state of the model parameters (including parameter values ​​and their rates of change) and the impact of the total disturbance in real time using an ESO (Electronic Disturbance Rejection Optimizer). Based on these estimates, an active disturbance rejection controller (ADO) is designed to calculate the control input. The control input consists of two main parts: a feedback control part based on the state error and a disturbance compensation part to compensate for the total disturbance.

[0056] In this embodiment, the control input is calculated using the error between the current state (including the rate of change of model parameters) estimated by ESO and the target state; this part belongs to feedback control. The existence of error determines the update direction; the larger the error, the greater the adjustment of the control input, ensuring that the model parameters are updated towards the optimal solution. The specific update process can be described as: adjusting the control input... As an "acceleration" of parameter change, it is applied to the current model parameters through integration or discretization (such as the Euler method). The above, that is, the update formula can be approximated as: ,in For the rate of change of the parameter, The iteration step size is determined by the direction from... The sign and state error are determined.

[0057] Secondly, the magnitude of the control input values ​​is regulated by the "acceleration" of parameter updates. This acceleration is calculated based on the current control input and the model's dynamic characteristics (such as gradient information and perturbation estimates). The perturbation compensation term, estimated by ESO, is used to adjust these magnitudes, ensuring a stable and robust update process and preventing training oscillations or instability caused by gradient noise or nonlinear effects. The specific update process involves adjusting the control input... With disturbance compensation Combined, the final amplitude adjustment factor is generated, for example, through... ,in The learning rate, with a magnitude of [missing information]. The absolute value and compensation term are adjusted to achieve smooth updates.

[0058] Therefore, when updating the model parameters in the current iteration step based on the direction and / or magnitude of the parameter update, the direction vector (based on the sign of the control input and error feedback) is first calculated, and then an magnitude factor (integrating perturbation compensation and dynamic characteristics) is applied to the parameter vector to achieve iterative parameter updates, making each update smoother and more stable. This method can effectively improve the robustness and convergence speed of the model in complex environments.

[0059] Furthermore, embodiments of this application can utilize calculated control inputs. To update the parameters of the deep learning model The parameter update rules should be consistent with the established dynamic model and control objectives. As one possible implementation method, embodiments of this application can directly utilize... The acceleration of the driving parameters is calculated, and new parameter values ​​are obtained through integration. For example, using discretized form: (Theoretically, if the compensation is accurate) or more practically, considering the delays and errors in ESO estimation and compensation: (Calculation of control acceleration) (Update speed estimate) (Updated location) Among them It is the time step size for discretization, which can usually be set as the learning rate. Another way is to... This can be considered as an adjustment signal for traditional gradient descent (or its variants), for example, using This can be used to modulate the learning rate or gradient direction; the specific update rules can be designed and adjusted according to the actual application.

[0060] The deep learning optimization method based on active disturbance rejection control proposed in this application model models the parameter optimization process of the deep learning model as a disturbed dynamic system. By extending the state observer, the estimated state value and the total disturbance value affecting the optimization process are estimated in real time and actively compensated. The disturbance is actively compensated in the active disturbance rejection controller to generate the control input value of the deep learning model in the current iteration step. The model parameters are updated using the control input value, which enhances the robustness of the optimization process to complex environments and disturbances. It effectively solves the problems of slow convergence, poor stability, sensitivity to hyperparameters and weak anti-interference ability in related technologies under complex loss terrain, and significantly improves the training efficiency and performance of the deep learning model.

[0061] The following will provide a more detailed explanation of the deep learning optimization based on active disturbance rejection control in the embodiments of this application, using specific examples.

[0062] Example 1: Optimization of Image Classification Model for Industrial Visual Quality Inspection Production line environments are subject to physical disturbances such as conveyor belt vibration (causing image blurring) and changes in workshop lighting (causing overexposure or underexposure). These disturbances are reflected in the training data as image noise. Models trained using traditional methods struggle to handle such disturbances, leading to high false negative and false positive rates in practical applications. This embodiment applies the method of the present invention to the visual inspection scenario of minute scratches on the surface of precision electronic components, specifically including the following steps: Step 1: Defining and Observing the Disturbance The noise at the data level caused by the physical environment, as well as the parameter coupling effect of model training itself, are collectively considered as the "total perturbation." An ESO (Earth Scraping Analysis) is constructed to observe the changes in the gradient vector in each iteration. All weights and bias parameters of the CNN model are combined into a parameter vector X. The optimization process is modeled as a dynamic system model, and a control gain is assumed. After preliminary estimation or setting.

[0063] Step 2: Perturbation Estimation When a batch of image data, mixed with blur and uneven lighting, is input into the model, the Extended State Observer (ESO) estimates the gradient bias caused by noise in this batch of data in real time, based on the calculated raw gradient. For example, the ESO might estimate: "The confidence of the current gradient pointing to the optimal direction has decreased by 30% due to image blur." The ESO module implements a third-order nonlinear ESO. In each training iteration... At the beginning, by obtaining the current model parameters ESO reception (as) ) and the control input from the previous moment Internal calculation of observation error Then, discretized ESO equations (e.g., based on the Euler method, with a step size of ) are used. ):

[0064] The state estimate at the current time is calculated. Total disturbance estimation (Note: Here) It means based on The estimate of the next state calculated from the time-information is used to calculate... (Control quantity at time). To simplify the notation, the estimated value calculated in the current iteration in this embodiment can be denoted as... .parameter It needs to be pre-configured, for example, based on bandwidth through simulation or experimentation. To adjust: (Approximate relationship in the linear case).

[0065] Step 3: Disturbance Compensation Before updating the weights, the algorithm corrects the original gradient based on the ESO estimation results. This correction operation is equivalent to "filtering out" the negative impact of factors such as image blurring, generating a pure gradient that is more focused on "learning scratch features".

[0066] The active disturbance rejection controller receives data from the ESO. The optimization objective is set as parameter stability, i.e., desired speed. Expected location It can be set to 0 (if the problem is normalized) or allowed to be implicitly driven by gradient information. A nonlinear PD control law is used for calculation. controller parameters It also needs to be preset, for example, based on the expected closed-loop bandwidth. To adjust: (in This is the damping ratio, an approximate relationship in the case of linear PD. Then, the final control input is calculated:

[0067] Step 4: Optimization and Update Use this compensated, purified gradient to update the weights of the convolutional neural network (CNN). Obtain the control input values. The following update rule can then be used: Calculate the acceleration contribution: Update speed estimate: (Note that the speed estimate used here is ESO) Update parameters: Step length here Equivalent to the learning rate in a traditional optimizer This can be set to a constant or a scheduling strategy can be employed. Alternatively, an update method incorporating gradient information can be designed, for example: (control signal) (Interpreted as the equivalent, perturbation-compensated, and dynamically adjusted "effective gradient" direction and magnitude)

[0068] In summary, the embodiments of this application can compare the loss reduction curve or test accuracy improvement curve of the CNN model trained using the deep learning optimization method based on active disturbance rejection control on a standard image classification dataset (such as CIFAR-10 or ImageNet) with the results using traditional optimizers such as Adam and SGD Momentum. The results demonstrate faster convergence speed, higher final accuracy, or smaller loss oscillations (more stability) during training. Because the trained model has actively adapted to and compensated for various disturbances during training, it maintains high-precision scratch recognition capabilities even in the face of changes in lighting and conveyor belt vibrations after deployment on the production line, significantly reducing false positives and false negatives and improving the level of quality inspection automation.

[0069] Example 2: Optimization of Bi-LSTM Networks for Financial Anti-Fraud Credit card fraud is typically a long sequence of events (such as login from a different location, small-amount probing, information modification, and large-amount fraud). When processing such long sequences, Bi-LSTM networks struggle to effectively correlate large-amount fraud at the end of the sequence with suspicious behavior at the beginning due to the vanishing gradient problem. This embodiment applies the deep learning optimization method based on active disturbance rejection control from the embodiments of this application to the real-time detection scenario of credit card fraud. When optimizing the Bi-LSTM network parameters, the steps are similar to those in Embodiment 1: Step 1: Definition and Observation of Disturbance The inherent "vanishing / exploding gradient" tendency of Bi-LSTM in long sequence backpropagation is defined as the core "internal perturbation". An ESO is constructed specifically to monitor the magnitude change of the gradient as it propagates over time steps.

[0070] Step 2: Perturbation estimation: When a sequence containing dozens of actions is input into the model for training, ESO monitors the entire process of error signal propagation from back to front. Once it detects an exponential decay of gradient magnitude (gradient vanishing), ESO immediately estimates the rate and extent of decay and quantifies it as a "perturbation value".

[0071] Step 3: Disturbance Compensation Before the gradient propagates to the next time step, a compensation mechanism intervenes. Based on the ESO estimate, it dynamically and non-linearly amplifies the gradient that is about to disappear, "restoring" it to its normal magnitude so that it can continue to propagate forward. Conversely, if gradient explosion occurs, it smooths it out.

[0072] Step 4 Optimization and Update: The weights of the Bi-LSTM network are updated using the gradient after gradient stability compensation.

[0073] In summary, due to the strong disturbance rejection capability of the active disturbance rejection controller, the embodiments of this application are expected to better address the optimization challenges posed by gradient noise and long-term dependencies when training models such as Bi-LSTM, thereby improving training stability and final model performance. Through the method of this application, the gradient propagation path of Bi-LSTM is opened, enabling it to truly establish a long-term memory of the entire transaction sequence. The final risk control model can accurately identify complex fraud patterns spanning several hours, greatly improving the early warning capability and accuracy of potential fraudulent activities, while ensuring efficient and stable model training.

[0074] Next, referring to the accompanying drawings, a deep learning optimization system based on active disturbance rejection control is described according to an embodiment of this application.

[0075] Figure 2 This is a block diagram of a deep learning optimization system based on active disturbance rejection control according to an embodiment of this application.

[0076] like Figure 2 As shown, the deep learning optimization system based on active disturbance rejection control includes: an interface module for extending the state observer, a control law calculation module, and an update module.

[0077] The system includes: an interface module for acquiring the model parameters of the deep learning model in the current iteration step and the control input value of the previous iteration step; an extended state observer for estimating the state estimate and total perturbation estimate of the model parameters based on the model parameters of the current iteration step and the control input value of the previous iteration step; a control law calculation module for calculating the control input value of the deep learning model in the current iteration step based on the state estimate and the total perturbation estimate; and an update module for updating the model parameters of the deep learning model based on the control input value and the model parameters of the current iteration step, and using the updated model parameters as the model parameters of the deep learning model in the next iteration step.

[0078] It is understood that, in this embodiment of the application, the interface module is used to obtain the parameters of the current deep learning model. and (optionally) obtain the gradient of the loss function. A detailed structural diagram of the Extended State Observer (ESO) module is shown below. Figure 3 As shown, the ESO algorithm is internally embedded and receives model parameters. and the control input of the previous iteration Output state estimation Total disturbance estimation A detailed structural diagram of the control law calculation module is shown below. Figure 4 As shown, the internal active disturbance rejection control law algorithm is embedded, and the receiver state estimation... and disturbance estimation (Optionally receive target trajectory) (or gradient information), calculate and output the current control input. Parameter update module: Receives control input. and current parameters (and possible) The model parameters are calculated and updated according to the rules in the above embodiments to obtain the parameters for the next iteration. .

[0079] It should be noted that the foregoing explanation of the deep learning optimization method for active disturbance rejection control also applies to the deep learning optimization system for active disturbance rejection control in this embodiment, and will not be repeated here.

[0080] According to the active disturbance rejection control deep learning optimization system proposed in the embodiments of this application, the parameter optimization process of the deep learning model is modeled as a disturbed dynamic system. The system estimates and actively compensates for the state estimates and total disturbance estimates that affect the optimization process in real time by extending the state observer. The system actively compensates for the disturbance in the active disturbance rejection controller and generates the control input value of the deep learning model in the current iteration step. The model parameters are updated using the control input value, which enhances the robustness of the optimization process to complex environments and disturbances. It effectively solves the problems of slow convergence, poor stability, sensitivity to hyperparameters and weak anti-interference ability in related technologies under complex loss terrain, and significantly improves the training efficiency and performance of the deep learning model.

[0081] Figure 5 A schematic diagram of the structure of an electronic device provided in an embodiment of this application. The electronic device may include: The memory 501, the processor 502, and the computer program stored on the memory 501 and capable of running on the processor 502.

[0082] When the processor 502 executes the program, it implements the deep learning optimization method for active disturbance rejection control provided in the above embodiments.

[0083] Furthermore, electronic devices also include: Communication interface 503 is used for communication between memory 501 and processor 502.

[0084] The memory 501 is used to store computer programs that can run on the processor 502.

[0085] The memory 501 may include high-speed RAM (Random Access Memory) memory, and may also include non-volatile memory, such as at least one disk storage.

[0086] If the memory 501, processor 502, and communication interface 503 are implemented independently, then the communication interface 503, memory 501, and processor 502 can be interconnected via a bus to complete communication between them. The bus can be an ISA (Industry Standard Architecture) bus, a PCI (Peripheral Component Interconnect) bus, or an EISA (Extended Industry Standard Architecture) bus, etc. The bus can be divided into address bus, data bus, control bus, etc. For ease of representation, Figure 5 The bus is represented by a single thick line, but this does not mean that there is only one bus or one type of bus.

[0087] Optionally, in a specific implementation, if the memory 501, processor 502, and communication interface 503 are integrated on a single chip, then the memory 501, processor 502, and communication interface 503 can communicate with each other through an internal interface.

[0088] The processor 502 may be a CPU (Central Processing Unit), an ASIC (Application Specific Integrated Circuit), or one or more integrated circuits configured to implement the embodiments of this application.

[0089] This application also provides a computer-readable storage medium storing a computer program thereon, which, when executed by a processor, implements the above-described deep learning optimization method for active disturbance rejection control.

[0090] In the description of this specification, the references to terms such as "one embodiment," "some embodiments," "example," "specific example," or "some examples," etc., refer to specific features, structures, materials, or characteristics described in connection with that embodiment or example, which are included in at least one embodiment or example of this application. In this specification, the illustrative expressions of the above terms do not necessarily refer to the same embodiment or example. Furthermore, the specific features, structures, materials, or characteristics described may be combined in any suitable manner in one or more embodiments or examples. Moreover, without contradiction, those skilled in the art can combine and integrate the different embodiments or examples described in this specification, as well as the features of different embodiments or examples.

[0091] Furthermore, the terms "first" and "second" are used for descriptive purposes only and should not be construed as indicating or implying relative importance or implicitly specifying the number of technical features indicated. Thus, a feature defined as "first" or "second" may explicitly or implicitly include at least one of that feature. In the description of this application, "N" means at least two, such as two, three, etc., unless otherwise explicitly specified.

[0092] Any process or method described in the flowchart or otherwise herein can be understood as representing a module, segment, or portion of code comprising one or N executable instructions for implementing custom logic functions or processes, and the scope of the preferred embodiments of this application includes additional implementations in which functions may be performed not in the order shown or discussed, including substantially simultaneously or in reverse order depending on the functions involved, as should be understood by those skilled in the art to which embodiments of this application pertain.

[0093] It should be understood that various parts of this application can be implemented using hardware, software, firmware, or a combination thereof. In the above embodiments, steps or methods can be implemented using software or firmware stored in memory and executed by a suitable instruction execution system. For example, if implemented in hardware, as in another embodiment, it can be implemented using any of the following techniques known in the art, or a combination thereof: discrete logic circuits having logic gates for implementing logical functions on data signals, application-specific integrated circuits (ASICs) having suitable combinational logic gates, programmable gate arrays (FPGAs), field-programmable gate arrays (FPGAs), etc.

[0094] Those skilled in the art will understand that all or part of the steps of the methods implementing the above embodiments can be implemented by a program instructing related hardware. The program can be stored in a computer-readable storage medium, and when executed, the program includes one or a combination of the steps of the method embodiments.

[0095] Although embodiments of this application have been shown and described above, it is understood that the above embodiments are exemplary and should not be construed as limiting this application. Those skilled in the art can make changes, modifications, substitutions and variations to the above embodiments within the scope of this application.

Claims

1. A deep learning optimization method based on active disturbance rejection control, characterized in that, Includes the following steps: During the iterative training of a deep learning model, the model parameters of the deep learning model in the current iteration step and the control input values ​​of the previous iteration step are obtained. The model parameters of the current iteration step and the control input values ​​of the previous iteration are input into the extended state observer, and the extended state observer outputs the state estimate and the total disturbance estimate. The state estimate and the total disturbance estimate are input into the active disturbance rejection controller, which calculates the control input value of the deep learning model in the current iteration step. Based on the control input value and model parameters of the current iteration step, update the model parameters of the deep learning model, and use the updated model parameters as the model parameters of the deep learning model in the next iteration step.

2. The deep learning optimization method based on active disturbance rejection control according to claim 1, characterized in that, Before obtaining the model parameters of the deep learning model in the current iteration step and the control input values ​​of the previous iteration step, the process includes: A dynamic system model for the optimization process of the deep learning model is established. The model parameters of the deep learning model to be optimized are regarded as the state variables of the dynamic system model. The model parameters of the current iteration step and the control input values ​​of the previous iteration step are obtained using the dynamic system model.

3. The deep learning optimization method based on active disturbance rejection control according to claim 2, characterized in that, The calculation formula for the dynamic system model is as follows: in, For model parameter vectors, The rate of change of the parameter over time. The acceleration of the parameter change over time. loss function Regarding parameters gradient, For gradient noise, To control the gain coefficient, To control the input signal.

4. The deep learning optimization method based on active disturbance rejection control according to claim 1, characterized in that, The calculation formula for the extended state observer is: in, For observation error, , Rate of change of parameter Real-time estimates, For total disturbance Real-time estimates, Here, is the observer gain, and is a positive constant. for , A pre-defined nonlinear function.

5. The deep learning optimization method based on active disturbance rejection control according to claim 1, characterized in that, The active disturbance rejection controller is equipped with an active disturbance rejection control law algorithm, and the calculation formula of the active disturbance rejection control law algorithm is as follows: in, It is the active disturbance rejection controller's estimate of the total disturbance. It is a disturbance compensation term. It is a nominal control law calculated based on state error feedback.

6. The deep learning optimization method based on active disturbance rejection control according to claim 1, characterized in that, The step of updating the model parameters of the deep learning model based on the control input value of the current iteration step and the model parameters of the current iteration step includes: The direction and / or magnitude of the parameter update are determined using the control input value of the current iteration step; The model parameters of the current iteration step are updated according to the direction and / or magnitude of the parameter update to obtain the updated model parameters.

7. The deep learning optimization method based on active disturbance rejection control according to claim 1, characterized in that, Before inputting the model parameters of the current iteration step and the control input values ​​of the previous iteration into the extended state observer, the method further includes: Obtain the gradient of the loss function of the deep learning model; The gradient of the loss function is input into the extended state observer, and the gradient of the loss function is used to assist the extended state observer in calculation.

8. A deep learning optimization system based on active disturbance rejection control, characterized in that, include: The interface module is used to obtain the model parameters of the deep learning model in the current iteration step and the control input values ​​of the previous iteration step; An extended state observer is used to estimate the state estimates of the model parameters and the total disturbance estimate based on the model parameters of the current iteration step and the control input values ​​of the previous iteration. The control law calculation module is used to calculate the control input value of the deep learning model in the current iteration step based on the state estimate and the total disturbance estimate. The update module is used to update the model parameters of the deep learning model based on the control input value of the current iteration step and the model parameters of the current iteration step, and use the updated model parameters as the model parameters of the deep learning model in the next iteration step.

9. An electronic device, characterized in that, include: A memory, a processor, and a computer program stored in the memory and executable on the processor, the processor executing the program to implement the deep learning optimization method based on active disturbance rejection control as described in any one of claims 1-7.

10. A computer-readable storage medium having a computer program or instructions stored thereon, characterized in that, When the computer program or instructions are executed, they implement the deep learning optimization method based on active disturbance rejection control as described in any one of claims 1-7.

Images