A method and system for automatically optimizing process parameters for particle board production

Abstract

The present application relates to the field of wood processing automation, and particularly relates to a method and system for automatically optimizing process parameters for particle board production, comprising: collecting production line operation indicators and quality error vectors, and constructing a nonlinear dynamics model containing time-varying unknown composite disturbance vectors; using a fixed-time disturbance observer to reconstruct the disturbance, calculate the observed estimate and the truncation error bound; combining the continuous differentiable function representing the actuator limit to construct an adaptive control barrier function constraint to generate a safe topological space boundary; at the same time, inputting the state and error vectors into the strategy gradient agent to output the nominal nominal control instruction; then inputting the nominal instruction into the convex optimization solver, using the safe topological space boundary to trim it, and solving the optimal robust process parameter instruction; finally, converting it into a driving electrical signal and issuing an execution. The present application realizes quality optimization of process parameters while ensuring physical safety of the underlying equipment without overloading.

Description

Technical Field

[0001] This invention relates to the field of wood processing automation, and in particular to a method and system for self-optimizing process parameters in particleboard production. Background Technology

[0002] In the continuous flat-press particleboard production process, the wood chips must undergo core processes such as glue application and mixing, and high-temperature and high-pressure curing. To ensure the surface quality and mechanical strength of the finished boards, control algorithms are usually introduced to dynamically adjust process parameters such as glue flow rate and hot-pressing pressure.

[0003] However, traditional particleboard parameter optimization mostly uses unconstrained intelligent algorithms. Due to the lack of underlying safety safeguards, it is easy to output out-of-bounds commands under disturbed working conditions, which can lead to equipment damage. Summary of the Invention

[0004] To overcome the above shortcomings, this invention provides a method and system for self-optimization of particleboard production process parameters. It aims to improve the problem that traditional particleboard parameter optimization methods mostly use unconstrained intelligent algorithms, which, due to the lack of underlying safety defenses, are prone to outputting out-of-bounds commands under disturbed operating conditions, leading to equipment damage.

[0005] In a first aspect, the present invention provides the following technical solution: a method for self-optimizing process parameters in particleboard production, comprising the following steps:

[0006] S1. Collect the operating indicators of the particleboard production line to construct a continuous-time state vector, collect the quality error vector output by the visual inspection station, and construct a nonlinear dynamic model containing time-varying unknown composite disturbance vectors based on the continuous-time state vector.

[0007] S2, input the continuous-time state vector into the fixed-time disturbance observer, reconstruct the time-varying unknown composite disturbance vector, and calculate the disturbance observation estimate and the truncation error limit;

[0008] S3, extract the continuously differentiable function that represents the limit of the actuator, substitute the disturbance observation estimate and the truncation error limit into the continuously differentiable function, construct the adaptive control obstacle function constraint, and generate the safe topology space boundary;

[0009] S4, input the continuous-time state vector and the quality error vector into the policy gradient agent, and output the nominal control command;

[0010] S5, input the nominal control command into the convex optimization solver, use the safe topology space boundary to trim the nominal control command, and solve for the optimal robust process parameter command;

[0011] S6, the optimal robust process parameter instruction is converted into a drive electrical signal and sent to the logic controller for execution.

[0012] By adopting the above technical solution, a safe topological boundary is constructed by reconstructing unknown disturbances and combining execution limits, thereby safely pruning the nominal instructions output by the intelligent agent. This improves the problem that traditional particleboard parameter optimization mostly uses unconstrained intelligent algorithms, which, due to the lack of underlying safety defenses, easily output out-of-bounds instructions under disturbance conditions, leading to equipment damage.

[0013] Optionally, in S1, constructing a nonlinear dynamic model containing a time-varying unknown composite disturbance vector based on the continuous-time state vector includes:

[0014] The temperature parameters of the hot press plate in the hot pressing section, the internal pressure parameters of the adhesive application pipe in the adhesive application section, and the load parameters of the paving strip in the paving section are collected. The temperature parameters of the hot press plate, the internal pressure parameters of the adhesive application pipe, and the load parameters of the paving strip are then spliced ​​together according to the time series to construct the continuous time state vector.

[0015] The displacement increment of the servo cylinder of the hot press and the speed adjustment increment of the glue pump are defined as system control input variables and combined to form the corresponding control command vector.

[0016] The nominal state drift vector and the nominal control gain matrix are extracted. The time-varying unknown composite disturbance vector, which characterizes the jump in wood chip moisture content, actuator wear and parameter drift, is introduced. The nominal state drift vector, the nominal control gain matrix and the time-varying unknown composite disturbance vector are nonlinearly combined in an affine manner to construct the nonlinear dynamic model.

[0017] Optionally, in S2, calculating the boundary between the disturbance observation estimate and the truncation error includes:

[0018] An auxiliary state variable with the same data dimension as the continuous-time state vector is introduced into the fixed-time perturbation observer;

[0019] Calculate the state deviation between the auxiliary state variable and the continuous-time state vector, and set a signed power operation rule for the state deviation;

[0020] The state deviation after being processed by the power operation rule, the state drift component and the control gain component in the nonlinear dynamic model are combined to construct a continuous evolution dynamic equation;

[0021] The continuous evolution dynamic equation is iteratively solved to extract the estimated value of the disturbance observation, and a fixed time constant independent of the initial state of the system is set according to the fixed-time convergence theory.

[0022] When the system running time reaches the fixed time constant, the maximum truncation deviation extreme value between the disturbance observation estimate and the actual disturbance is extracted as the truncation error limit.

[0023] Optionally, in S3, the extraction of the continuously differentiable function characterizing the limit of the actuator includes:

[0024] Obtain the maximum allowable pressure boundary value of the hot press and the maximum allowable flow boundary value of the glue applicator;

[0025] The maximum allowable pressure boundary value and the maximum allowable flow boundary value are transformed into topological space inequalities corresponding to the continuous-time state vector;

[0026] The continuously differentiable function characterizing the limit of the actuator is generated using the aforementioned topological space inequalities;

[0027] The state mapping region where the continuously differentiable function takes values ​​greater than zero and equal to zero is defined as the desired safe set.

[0028] Optionally, in S3, generating a secure topology space boundary includes:

[0029] The partial derivative expansion is generated by calculating the Lie derivative of the continuously differentiable function along the state evolution trajectory of the nonlinear dynamic model.

[0030] An extended class K function is introduced as a compensation term in the partial derivative expansion;

[0031] The disturbance observation estimate and the truncation error limit are embedded into the partial derivative expansion in the form of boundary scaling to generate the adaptive control obstacle function constraint.

[0032] The system state trajectory after being constrained by the adaptive control barrier function is strictly within the desired safety set, forming the safety topology space boundary.

[0033] Optionally, in S4, the output nominal control command includes:

[0034] Define a control Lyapunov function that represents the absolute deviation between the quality error vector and the expected factory quality standard;

[0035] The continuous-time state vector and the quality error vector are input into the state feature extraction network layer of the policy gradient agent to obtain hidden layer features;

[0036] The hidden layer features are input into the policy network nodes, and the policy network node parameters are updated by action exploration using the policy gradient ascent algorithm.

[0037] The network output action that causes the value of the control Lyapunov function to exhibit a decreasing gradient is extracted as the nominal control command.

[0038] Optionally, in S5, pruning the nominal control command using the secure topology space boundary includes:

[0039] Minimizing the quadratic norm deviation between the quadratic programming variable to be optimized and the nominal control command is set as the basic optimization cost function;

[0040] The adaptive control barrier function constraint is set as a safety boundary condition, and the descent gradient of the control Lyapunov function is set as an error convergence condition.

[0041] The safety boundary conditions, the error convergence conditions, and the mechanical extreme values ​​of the underlying actuators are combined to form a set of physical hard constraint conditions.

[0042] A penalized relaxation variable is introduced into the safety boundary conditions and error convergence conditions in the set of physical hard constraints to construct a pruning model with a relaxation factor.

[0043] Optionally, in S5, the instruction to solve for the optimal robust process parameters includes:

[0044] The basic optimization cost function is combined with the pruning model with relaxation factor to transform it into a standard quadratic programming mathematical model equation;

[0045] The standard quadratic programming mathematical model equations are solved online using an interior-point iterative algorithm.

[0046] When the solution process converges, the solution that minimizes the basic optimization cost function and strictly satisfies the set of physical hard constraints is extracted as the optimal robust process parameter instruction.

[0047] Optionally, in S6, converting the optimal robust process parameter instruction into a drive electrical signal and sending it to the logic controller for execution includes:

[0048] Receive the optimal robust process parameter command output by the convex optimization solver;

[0049] The optimal robust process parameter command is subjected to data frame unpacking operation to parse out the voltage control signal for the servo proportional valve of the hot press and the frequency control signal for the glue application frequency pump.

[0050] The voltage control signal and the frequency control signal are synchronously sent to the corresponding logic controller via an industrial control bus network to execute physical drive.

[0051] Secondly, the present invention provides the following technical solution: a self-optimization system for particleboard production process parameters, comprising the following modules:

[0052] The state perception and dynamic modeling module is used to collect the operating indicators of the particleboard production line to construct a continuous-time state vector, collect the quality error vector output by the visual inspection station, and construct a nonlinear dynamic model containing time-varying unknown composite disturbance vectors based on the continuous-time state vector.

[0053] The disturbance observation and reconstruction module is used to input the continuous-time state vector into the fixed-time disturbance observer, reconstruct the time-varying unknown composite disturbance vector, and calculate the disturbance observation estimate and the truncation error limit.

[0054] The safe topology boundary generation module is used to extract a continuously differentiable function that represents the limit of the actuator, substitute the disturbance observation estimate and the truncation error limit into the continuously differentiable function, construct an adaptive control obstacle function constraint, and generate a safe topology space boundary.

[0055] The nominal instruction optimization module is used to input the continuous-time state vector and the quality error vector into the policy gradient agent and output the nominal control instruction.

[0056] The robust instruction trimming and solving module is used to input the nominal control instruction into the convex optimization solver, trim the nominal control instruction using the safe topology space boundary, and solve for the optimal robust process parameter instruction.

[0057] The drive execution closed-loop module is used to convert the optimal robust process parameter instructions into drive electrical signals and send them to the logic controller for execution.

[0058] The present invention has the following beneficial effects:

[0059] 1. In this invention, by reconstructing unknown disturbances and combining them with execution limits to construct a safe topology boundary, the nominal instructions output by the intelligent agent are safely pruned. This improves the problem that traditional particleboard parameter optimization mostly uses unconstrained intelligent algorithms, which, due to the lack of underlying safety defenses, easily output out-of-bounds instructions under disturbance conditions, leading to equipment damage.

[0060] 2. In this invention, by setting power operation rules for state deviations to construct evolutionary dynamic equations, and then extracting truncation error limits based on fixed-time convergence theory, the problem of traditional disturbance observations, which mostly adopt asymptotic convergence models and are affected by the initial state, causing time delays in disturbance reconstruction and resulting in lag in compensation control, is improved.

[0061] 3. In this invention, by calculating the partial derivative expansion of a continuously differentiable function, the disturbance observation estimate and the error limit are embedded in the form of boundary scaling to generate adaptive constraints. This improves the problem that traditional equipment protection mostly uses static thresholds, which cannot adapt to external shocks and thus cause the physical limit to be dynamically penetrated under extreme disturbance conditions.

[0062] 4. In this invention, by combining the safety boundary with the error convergence condition to form a set of physical hard constraints, and then introducing penalty relaxation variables inside the constraints to construct a pruning model, the problem of traditional constraint optimization, which mostly uses absolute hard boundaries, is improved because the feasible region is prone to overlap and disappear under severe perturbation, thus causing the optimization solver to get stuck in a deadlock with no solution. Attached Figure Description

[0063] Figure 1 This is a flowchart of a method for self-optimization of particleboard production process parameters proposed in this invention;

[0064] Figure 2 This is a flowchart illustrating the safety constraints and optimization solution of a self-optimization method for particleboard production process parameters proposed in this invention.

[0065] Figure 3 This is a flowchart illustrating the physical drive signal conversion execution of a self-optimization method for particleboard production process parameters proposed in this invention.

[0066] Figure 4 This is an architecture diagram of a self-optimization system for particleboard production process parameters proposed in this invention. Detailed Implementation

[0067] The technical solutions in the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0068] Example 1:

[0069] In a first embodiment of the present invention, the present invention provides a method for self-optimizing particleboard production process parameters, such as... Figures 1-3 As shown, it includes the following steps:

[0070] S1. Collect the operating indicators of the particleboard production line to construct a continuous-time state vector, collect the quality error vector output by the visual inspection station, and construct a nonlinear dynamic model containing time-varying unknown composite disturbance vectors based on the continuous-time state vector.

[0071] Furthermore, in S1, the nonlinear dynamic model constructed based on the continuous-time state vector, which includes a time-varying unknown composite perturbation vector, includes:

[0072] The temperature parameters of the hot press plate in the hot pressing section, the internal pressure parameters of the adhesive application pipe in the adhesive application section, and the load parameters of the paving strip in the paving section are collected. The temperature parameters of the hot press plate, the internal pressure parameters of the adhesive application pipe, and the load parameters of the paving strip are then spliced ​​together according to the time series to construct a continuous time state vector.

[0073] The displacement increment of the servo cylinder of the hot press and the speed adjustment increment of the glue pump are defined as system control input variables and combined to form the corresponding control command vector.

[0074] The nominal state drift vector and nominal control gain matrix are extracted. A time-varying unknown composite disturbance vector characterizing the change in wood chip moisture content, actuator wear and parameter drift is introduced. The nominal state drift vector, nominal control gain matrix and time-varying unknown composite disturbance vector are nonlinearly combined to construct a nonlinear dynamic model.

[0075] Specifically, the data input / output process is structured as follows: First, the temperature parameters of the hot press plate in the hot pressing section, the internal pressure parameters of the adhesive application pipeline in the adhesive application section, and the load parameters of the paving belt in the paving section are acquired as basic input data. Then, the aforementioned input data is concatenated according to a time series to output a continuous-time state vector as the system state characterization variable. Next, the displacement increment of the servo cylinder of the hot press and the speed adjustment increment of the adhesive application pump are acquired as system control input variables. These system control input variables are combined to output the corresponding control command vector. Finally, the nominal state drift vector and nominal control gain matrix are extracted, a time-varying unknown composite disturbance vector is introduced, and a nonlinear dynamic model is output to complete the mapping from physical space to mathematical space.

[0076] In practical implementation, the nonlinear dynamic model uses nonlinear affine equations to define specific dynamic evolution laws:

[0077] ;

[0078] in It represents the derivative component of the continuous-time state vector with respect to time. This represents the continuous-time state vector constructed by the aforementioned steps. The continuous-time state vector contains the hot press plate temperature parameters, the adhesive application pipe internal pressure parameters, and the paving tape load parameter sequence. This represents the nominal state drift vector of a particleboard production line under ideal, undisturbed conditions. This represents the nominal control gain matrix of a particleboard production line under ideal conditions. This represents a control command vector composed of the displacement increment for the servo cylinder of the hot press and the speed adjustment increment for the glue pump. This represents a time-varying unknown composite disturbance vector characterizing the jump in wood chip moisture content, actuator wear, and parameter drift.

[0079] By collecting temperature parameters of the hot press plate, internal pressure parameters of the adhesive application pipe, and load parameters of the paving belt and splicing them into a continuous-time state vector according to the time sequence, and introducing a time-varying unknown composite disturbance vector characterizing the jump in wood chip moisture content, actuator wear, and parameter drift, a nonlinear dynamic model is constructed by nonlinearly combining the time-varying unknown composite disturbance vector with the nominal state drift vector and the nominal control gain matrix. This improves upon the shortcomings of traditional particleboard automatic control schemes, which rely entirely on ideal nominal state equations and fail to mathematically decouple the time-varying physical disturbance characteristics of the production line. This avoids the problems of severe distortion in control system state estimation and subsequent failure of dynamic optimization and control caused by transient changes in the properties of wood chip raw materials or implicit aging of the underlying servo actuators in actual production. The nonlinear dynamic model provides the necessary and complete dynamic constraint benchmark for subsequent fixed-time disturbance observation and robust command safety boundary generation.

[0080] S2, input the continuous-time state vector into the fixed-time disturbance observer, reconstruct the time-varying unknown composite disturbance vector, and calculate the disturbance observation estimate and the truncation error limit;

[0081] Furthermore, in S2, the limits between the disturbance observation estimate and the truncation error are calculated, including:

[0082] Introduce auxiliary state variables with the same data dimension as the continuous-time state vector in the fixed-time perturbation observer;

[0083] Calculate the state deviation between the auxiliary state variable and the continuous-time state vector, and set a signed power operation rule for the state deviation;

[0084] The state deviation after being processed by the power operation rule, the state drift component in the nonlinear dynamic model, and the control gain component are combined to construct the continuous evolution dynamic equation;

[0085] The continuous evolution dynamic equation is iteratively solved to extract the estimated values ​​of disturbance observations, and a fixed time constant independent of the initial state of the system is set according to the fixed-time convergence theory.

[0086] When the system running time reaches a fixed time constant, the maximum truncation deviation extreme value between the disturbance observation estimate and the actual disturbance is extracted as the truncation error limit.

[0087] Specifically, the detailed S2 step clarifies the processing path for unknown physical disturbances from an unmeasurable state to mathematical quantification. The data input / output flow is as follows: the system receives a continuous-time state vector and the state drift and control gain components from the nonlinear dynamic model as basic inputs. Within the observation module, an auxiliary state variable with the same data dimension as the continuous-time state vector is introduced, and the state deviation between the two is calculated. This state deviation is processed according to a predefined signed exponentiation rule and then mathematically combined with the state drift and control gain components. Finally, the reconstructed disturbance observation estimate and truncation error limits are output downstream of the system.

[0088] The continuous evolution dynamics equation of the fixed-time perturbation observer upon which this step is based is defined as follows:

[0089] ;

[0090] ;

[0091] in This represents the derivative of the auxiliary state variable with respect to time. This represents the auxiliary state variable introduced to achieve disturbance tracking. This represents the continuous-time state vector constructed from the collected data. This indicates the rules for signed exponentiation. and as well as and These represent the preset power exponent parameters. and as well as and This represents the positive definite observer gain scalar coefficient. This represents the state drift component. This indicates the control gain component. This indicates the actual control commands issued. This represents the estimated value of the disturbance observation to be extracted. This represents the time derivative of the estimated perturbation observation.

[0092] When the system's runtime evolves and reaches the fixed time constant derived from the fixed-time convergence theory, the iterative solution process of the evolutionary dynamics equations reaches theoretical convergence. At this node, the maximum residual truncation deviation extreme value between the observed disturbance estimate and the actual disturbance is extracted and established as the truncation error limit.

[0093] In particleboard production, transient fluctuations in wood chip moisture content or equipment wear and tear are difficult to measure directly and in real-time using conventional sensors in industrial settings. By introducing auxiliary state variables into the control system to construct observer dynamic equations, implicit disturbances at the physical level are transformed into explicit mathematical variables at the control level. The evolution equations established using signed power operations ensure that the disturbance reconstruction process converges within a defined fixed time constant, eliminating the interference of initial state differences on the timeliness of disturbance observation. The extracted disturbance observation estimates and truncation error limits provide direct data for constructing a dynamic safety barrier to resist this disturbance. This improves upon the shortcomings of conventional self-optimization algorithms, which blindly rely on nominal models, leading to the failure of safety boundaries under disturbance impacts. It also cuts off the path to large-scale defective product generation caused by delayed disturbance propagation at its source.

[0094] S3, extract the continuously differentiable function that represents the limit of the actuator, substitute the disturbance observation estimate and the truncation error limit into the continuously differentiable function, construct the adaptive control obstacle function constraint, and generate the safety topology space boundary;

[0095] Furthermore, in S3, the continuously differentiable function representing the limit of the actuator includes:

[0096] Obtain the maximum allowable pressure boundary value of the hot press and the maximum allowable flow boundary value of the glue applicator;

[0097] The maximum allowable pressure boundary value and the maximum allowable flow boundary value are transformed into topological space inequalities corresponding to continuous-time state vectors;

[0098] Utilize topological space inequalities to generate continuously differentiable functions characterizing the limits of actuators;

[0099] Define the region of states where a continuously differentiable function takes values ​​greater than zero and equal to zero as the desired safe set.

[0100] In S3, generating the secure topology space boundary includes:

[0101] Lie derivative calculations are performed on the state evolution trajectory of a continuously differentiable function along a nonlinear dynamic model to generate a partial derivative expansion.

[0102] An extended class K function is introduced as a compensation term in the partial derivative expansion;

[0103] The disturbance observation estimate and the truncation error limit are embedded into the partial derivative expansion in the form of boundary scaling to generate adaptive control barrier function constraints.

[0104] The system state trajectory, after being constrained by the adaptive control barrier function, is strictly contained within the desired safe set, forming a safe topological space boundary.

[0105] Specifically, the detailed S3 step establishes a fundamental mathematical protection barrier for the safe operation of the physical execution equipment. The data input / output process is as follows: the system receives the maximum allowable pressure boundary value of the hot press and the maximum allowable flow rate boundary value of the glue applicator as preliminary inputs. These physical boundary values ​​are transformed into topological space inequalities through spatial mapping, outputting a continuously differentiable function to the system. Subsequently, the system extracts the state evolution trajectory of the nonlinear dynamic model generated in the preceding steps, the disturbance observation estimate, and the truncation error limit. It then performs Lie derivative calculations on the continuously differentiable function and substitutes it into compensation terms, finally outputting adaptive control obstacle function constraints and safe topological space boundaries to the downstream solution module.

[0106] When implementing the constraints, the system maps the desired safe set using continuously differentiable functions:

[0107] ;

[0108] in This represents the expected safe set. This represents a continuous-time state vector. express 3D real space. This represents a continuously differentiable function derived from the physical limits of the underlying device. The mapping region with values ​​greater than zero and equal to zero defines the absolute safe range within which the actuator will not experience physical overload.

[0109] To mitigate the impact of unknown external disturbances on the safe zone, the system further constructs an adaptive control barrier function constraint inequality:

[0110] ;

[0111] in Let Lie derivative be the Lie derivative of a continuously differentiable function along the nominal state drift vector. Let represent the Lie derivative of a continuously differentiable function along the nominal control gain matrix. This represents the actual control command vector issued by the system. This represents the transpose of the gradient vector of a continuously differentiable function. This represents the estimated value of the disturbance observation extracted in the preceding steps. This represents the norm operation term of the gradient vector. This indicates the truncation error limit. This represents the extended class K function introduced as a compensation term.

[0112] By expanding partial derivatives and embedding the disturbance observation estimate and the truncation error limit into the constraint inequality in the form of boundary scaling, this approach breaks through the technical limitations of traditional industrial control schemes that rely entirely on static fixed thresholds for safety interception. When the production site encounters severe physical disturbances such as sudden changes in wood chip material, the adaptive control barrier function constraint can proactively perform adaptive contraction of the safety topology space boundary based on the reconstructed disturbance amplitude and error limit. This mechanism establishes strict mathematical forward invariance conditions, ensuring that the system state trajectory is locked within the desired safety set under various complex and severe operating conditions. It cuts off the path for the optimization algorithm to output extreme out-of-bounds instructions, avoids the risk of irreversible mechanical damage to the underlying execution equipment due to receiving incorrect optimization parameters, and provides a rigid physical defense for solving subsequent intelligent nominal instructions.

[0113] S4 inputs the continuous-time state vector and the quality error vector into the policy gradient agent and outputs the nominal control command.

[0114] Furthermore, in S4, the output nominal control commands include:

[0115] The control Lyapunov function is set to represent the absolute deviation between the quality error vector and the expected factory quality standard.

[0116] The continuous-time state vector and the quality error vector are input into the state feature extraction network layer of the policy gradient agent to obtain the hidden layer features.

[0117] The hidden layer features are input into the policy network nodes, and the policy network node parameters are updated by action exploration using the policy gradient ascent algorithm.

[0118] Extract the network output action that causes the Lyapunov function value to exhibit a decreasing gradient as the nominal control command.

[0119] Specifically, the detailed S4 step clarifies the decision-making and computational path from feature extraction to control command generation for production line quality errors. The data input / output process is as follows: the system receives continuous-time state vectors and quality error vectors as initial input data. This initial input data is then fed into the state feature extraction network layer within the policy gradient agent. The hidden layer features are then calculated and output through nonlinear mapping of network weights. These hidden layer features subsequently enter the policy network nodes, where the policy gradient ascent algorithm is used to explore and update the parameters of these nodes. The system then extracts the network output actions that drive the Lyapunov function values ​​towards a descent gradient and outputs the nominal control command.

[0120] In model training and instruction execution, the control Lyapunov function set by the system is specifically defined as follows:

[0121] ;

[0122] in The Lyapunov function represents the control function that characterizes the absolute deviation. This represents the quality error vector collected from the visual inspection station. The transpose matrix represents the quality error vector. This represents a pre-defined positive definite symmetric weight matrix, responsible for allocating the penalty weight for quality errors in different dimensions.

[0123] To ensure that the factory quality standards are met, the nominal control commands output by the policy gradient agent must follow the following descent gradient conditions:

[0124] ;

[0125] in This represents the time derivative of the Lyapunov function. This represents a convergence rate constant that is greater than zero. This condition requires that the actions generated by the policy network must drive the overall error system energy to exhibit a decaying trend.

[0126] The surface quality of particleboard in industrial settings is influenced by multiple dimensional state variables, making it difficult for conventional linear controllers to calculate accurate correction actions amidst a massive number of nonlinear equipment states. This paper introduces a policy gradient agent for feature extraction and action exploration, constructing a mapping network from high-dimensional physical operational characteristics to process adjustment commands. The Lyapunov function descent gradient is set as the physical evaluation criterion for the network's output actions, giving the neural network's action exploration process a clear mathematical convergence direction. This mechanism eliminates the agent's blind divergent trial-and-error behavior in the early stages of optimization, ensuring that the final nominal control commands possess physical determinism in eliminating quality defects. This directly provides the downstream safe trimming steps with a precise, target-oriented foundation of unsolved variables.

[0127] S5. Input the nominal control command into the convex optimization solver, use the safe topology space boundary to trim the nominal control command, and solve for the optimal robust process parameter command.

[0128] Furthermore, in S5, pruning nominal control commands using the safety topology space boundary includes:

[0129] Minimizing the quadratic norm deviation between the quadratic programming variable to be optimized and the nominal control command is set as the basic optimization cost function;

[0130] The adaptive control barrier function constraint is set as a safety boundary condition, and the descent gradient of the control Lyapunov function is set as an error convergence condition.

[0131] The set of physical hard constraints is formed by combining the safety boundary conditions, error convergence conditions, and the mechanical extreme values ​​of the underlying actuators.

[0132] A pruning model with a relaxation factor is constructed by introducing penalized relaxation variables into the safety boundary conditions and error convergence conditions within the set of physical hard constraints.

[0133] In S5, the instructions for solving for the optimal robust process parameters include:

[0134] The basic optimization cost function is combined with the pruning model with relaxation factors to transform it into a standard quadratic programming mathematical model equation.

[0135] Solve the standard quadratic programming mathematical model equations online using an interior-point iterative algorithm;

[0136] When the solution process converges, the solution that minimizes the basic optimization cost function and strictly satisfies the set of physical hard constraints is extracted as the optimal robust process parameter instruction.

[0137] Specifically, the S5 step refinement establishes a quadratic programming solution path for transforming unverified policy actions into safe and executable instructions. The data input / output flow is as follows: it receives the nominal control instructions generated in the preceding steps, adaptive control barrier function constraints, the descent gradient of the control Lyapunov function, and the mechanical extrema of the underlying actuator as pre-input conditions. Minimizing the quadratic norm deviation between the quadratic programming variable to be optimized and the nominal control instructions is set as the basic optimization cost function. A set of physical hard constraints is constructed by combining safety boundary conditions, error convergence conditions, and mechanical extrema, and penalized relaxation variables are introduced within this set. The basic optimization cost function and the set of physical hard constraints with penalized relaxation variables are transformed into a standard quadratic programming mathematical model equation, input into an interior-point iterative algorithm for online calculation, and after determining iterative convergence, the optimal robust process parameter instruction is output to the underlying logic controller.

[0138] When performing online solution calculations, the standard quadratic programming mathematical model equations are defined as follows:

[0139] ;

[0140] The solution process strictly follows the following set of physical hard constraints with penalty relaxation factors:

[0141] ;

[0142] ;

[0143] ;

[0144] in This represents the operation of minimizing the objective cost function. This represents the variable to be optimized in a quadratic programming problem. This represents the nominal control command input to the solver. This represents the quadratic norm operation applied to the elements of a vector. and This represents a penalty relaxation variable introduced to prevent unsolvable deadlock situations under extreme computational conditions. and This represents the penalty weight coefficient assigned to the penalty slack variable. This represents the Lie derivative of the safety boundary function obtained along the nominal state drift vector of the system. This represents the Lie derivative of the safety boundary function obtained along the nominal control gain matrix. This represents the transpose of the gradient vector of the safety boundary function. This represents the estimated disturbance observation value output by the fixed-time disturbance observer. This represents the norm of the gradient vector at the safety boundary. This represents the truncation error limit that characterizes the accuracy of perturbation reconstruction. This represents an extended class K function used to provide a safety buffer margin. This represents the derivative of the Lyapunov function with respect to the time variable. The value of the control Lyapunov function representing the current state. It represents a positive real constant scalar that defines the convergence rate of the quality error. and These represent the lower limit and upper limit of the ultimate execution of the bottom structure of the hot press and the glue applicator, respectively.

[0145] In continuous particleboard processing scenarios, nominal commands that simply pursue the convergence of quality defects often trigger extremely drastic servo compensation actions when encountering sudden changes in wood moisture content. This quadratic programming model uses the safety topology space boundary as a rigid pruning filter, forcibly pulling dangerous parameters that could cause board breakage or hydraulic cylinder damage back into the safe operating range. Even if the external physical environment deteriorates drastically, the introduction of relaxation variables ensures that the optimization command has continuous iterative capabilities, ensuring that each issued process parameter closely approximates the optimal factory quality standard without causing mechanical damage to the underlying equipment.

[0146] S6 converts the optimal robust process parameter instruction into a drive electrical signal and sends it to the logic controller for execution;

[0147] Furthermore, in S6, the process of converting the optimal robust process parameter instructions into drive electrical signals and sending them to the logic controller for execution includes:

[0148] Receive the optimal robust process parameter command output by the convex optimization solver;

[0149] The optimal robust process parameter command is depacked into data frames to extract the voltage control signal for the servo proportional valve of the hot press and the frequency control signal for the glue application frequency pump.

[0150] Voltage control signals and frequency control signals are synchronously sent to the corresponding logic controllers via an industrial control bus network to execute physical drives.

[0151] Specifically, the detailed S6 step clarifies the cross-domain transformation process from abstract mathematical optimization results to electrical drive signals of the underlying physical execution equipment. The data input / output flow in this step is as follows: the system receives the optimal robust process parameter command calculated by the upper-level convex optimization solver as the sole source input data. System-level data frame unpacking and analog-to-digital mapping operations are performed on the source input data to extract and separate multiple independent drive signals, outputting voltage control signals for the servo proportional valve of the hot press and frequency control signals for the glue application variable frequency pump. Finally, relying on the industrial control bus network, the voltage and frequency control signals are synchronously routed and sent to the corresponding logic controllers on the production line.

[0152] The mathematical model involved in the instruction conversion process during data frame unpacking and underlying driver signal mapping calculations is defined as follows:

[0153] ;

[0154] in This represents the voltage control signal generated by the parsing process for the servo proportional valve of the hot press. This indicates the frequency control signal generated by the parsing process for the variable frequency pump. This represents the digital-to-analog conversion gain matrix pre-calibrated based on the underlying device communication protocol and the sensor's measurement range. This represents the optimal robust process parameter command vector, containing multidimensional control components, output by the convex optimization solver. This represents the zero-point electrical offset vector used to match the dead-zone characteristics of the underlying electrical actuator.

[0155] In actual industrial production environments, the pure mathematical variable sequences derived from upper-level reinforcement learning and quadratic programming calculations cannot be directly recognized and responded to by the hydraulic cylinders of hot presses or the variable frequency motors of glue applicators. The system relies on data frame unpacking operations to convert the optimal robust process parameter commands into specific voltage and frequency control quantities, and then uses a high-real-time industrial control bus network to complete the synchronous distribution across the entire line, effectively implementing the robust safety constraints and quality optimization strategies built in the preceding steps. This conversion and distribution mechanism eliminates process control lag defects caused by communication protocol mismatches of underlying devices and signal parsing delays, preventing range distortion of the safety-verified optimal control commands during the physical execution phase, and completing the final physical closed-loop realization of the entire set of adaptive optimization parameters on the particleboard continuous flat pressing production line.

[0156] Example 2:

[0157] In a continuous flat-press particleboard production line, wood chips are sequentially laid and shaped using a laying belt, mixed with glue using a gluing system, and cured under high temperature and pressure in a continuous hot press. Finally, the finished product quality is inspected at the end-of-line visual inspection station. In actual long-cycle continuous operation scenarios, the nominal dynamic model of the production line is easily distorted due to the transient jumps in the moisture content of the wood chips, the time-varying drift of the thermodynamic parameters inside the hot press, and the unknown combined disturbances caused by the mechanical wear of the underlying servo actuators. When the end-of-line visual inspection station reports surface quality defects in the particleboard, traditional process parameter optimization algorithms, in order to eliminate quality deviations, often deviate from the underlying physical boundary constraints and output out-of-bounds process parameter compensation commands to the underlying equipment. Optimization adjustments lacking safe topological boundary constraints and disturbance reconstruction mechanisms can directly penetrate the absolute physical safety operating limits of the hot press and gluing machine, causing equipment mechanical accidents such as hydraulic cylinder breakage or glue pump overload damage, leading to a large-scale continuous production of irreversible defective products. To solve the above problems, this invention provides a self-optimization system for particleboard production process parameters, the structure of which is as follows: Figure 4 As shown. The specific implementation process of this system is as follows:

[0158] The system modules construct a collaborative control closed loop from bottom-level equipment state perception to top-level safety constraint pruning. The state perception and dynamics modeling module, together with the disturbance observation and reconstruction module, transforms the physical operating indicators of the production line and the hidden time-varying disturbances into explicit dynamic state-space variables. The safety topology boundary generation module further transforms the absolute operational extrema of the hot press and the glue applicator into mathematical topology boundaries with disturbance resistance capabilities, constructing a protective barrier for the bottom-level equipment.

[0159] To address the issue that nominal instructions generated by the policy agent in the nominal instruction optimization module only provide quality correction guidance but lack mechanical safety guarantees, the system relies on the convex optimization solver within the robust instruction pruning solution module. Using the aforementioned safety topology space boundary as a hard filter, it prunes the feasible region of nominal instructions that may cause equipment overload. The optimal robust process parameter instructions, stripped of out-of-bounds danger parameters, are ultimately converted into electrical drive signals and issued by the drive execution closed-loop module. This module architecture solves the underlying technical defect of conventional intelligent algorithms, which easily output extreme out-of-bounds instructions and cause irreversible damage to the actuator when dealing with complex disturbance conditions. It achieves hardware-level closed-loop control of zero mechanical safety exceedances and product quality convergence.

[0160] Finally, it should be noted that the above description is only a preferred embodiment of the present invention and is not intended to limit the present invention. Although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art can still modify the technical solutions described in the foregoing embodiments or make equivalent substitutions for some of the technical features. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the protection scope of the present invention.

Claims

1. A method for self-optimizing process parameters in particleboard production, characterized in that, Includes the following steps: S1. Collect the operating indicators of the particleboard production line to construct a continuous-time state vector, collect the quality error vector output by the visual inspection station, and construct a nonlinear dynamic model containing time-varying unknown composite disturbance vectors based on the continuous-time state vector. S2, input the continuous-time state vector into the fixed-time disturbance observer, reconstruct the time-varying unknown composite disturbance vector, and calculate the disturbance observation estimate and the truncation error limit; S3, extract the continuously differentiable function that represents the limit of the actuator, substitute the disturbance observation estimate and the truncation error limit into the continuously differentiable function, construct the adaptive control obstacle function constraint, and generate the safe topology space boundary; S4, input the continuous-time state vector and the quality error vector into the policy gradient agent, and output the nominal control command; S5, input the nominal control command into the convex optimization solver, use the safe topology space boundary to trim the nominal control command, and solve for the optimal robust process parameter command; S6, the optimal robust process parameter instruction is converted into a drive electrical signal and sent to the logic controller for execution.

2. The method for self-optimizing particleboard production process parameters according to claim 1, characterized in that, In S1, constructing a nonlinear dynamic model containing a time-varying unknown composite perturbation vector based on the continuous-time state vector includes: The temperature parameters of the hot press plate in the hot pressing section, the internal pressure parameters of the adhesive application pipe in the adhesive application section, and the load parameters of the paving strip in the paving section are collected. The temperature parameters of the hot press plate, the internal pressure parameters of the adhesive application pipe, and the load parameters of the paving strip are then spliced ​​together according to the time series to construct the continuous time state vector. The displacement increment of the servo cylinder of the hot press and the speed adjustment increment of the glue pump are defined as system control input variables and combined to form the corresponding control command vector. The nominal state drift vector and the nominal control gain matrix are extracted. The time-varying unknown composite disturbance vector, which characterizes the jump in wood chip moisture content, actuator wear and parameter drift, is introduced. The nominal state drift vector, the nominal control gain matrix and the time-varying unknown composite disturbance vector are nonlinearly combined in an affine manner to construct the nonlinear dynamic model.

3. The method for self-optimizing particleboard production process parameters according to claim 1, characterized in that, In S2, the calculation of the boundary between the disturbance observation estimate and the truncation error includes: An auxiliary state variable with the same data dimension as the continuous-time state vector is introduced into the fixed-time perturbation observer; Calculate the state deviation between the auxiliary state variable and the continuous-time state vector, and set a signed power operation rule for the state deviation; The state deviation after being processed by the power operation rule, the state drift component and the control gain component in the nonlinear dynamic model are combined to construct a continuous evolution dynamic equation; The continuous evolution dynamic equation is iteratively solved to extract the estimated value of the disturbance observation, and a fixed time constant independent of the initial state of the system is set according to the fixed-time convergence theory. When the system running time reaches the fixed time constant, the maximum truncation deviation extreme value between the disturbance observation estimate and the actual disturbance is extracted as the truncation error limit.

4. The method for self-optimizing particleboard production process parameters according to claim 1, characterized in that, In S3, the extraction of the continuously differentiable function representing the limit of the actuator includes: Obtain the maximum allowable pressure boundary value of the hot press and the maximum allowable flow boundary value of the glue applicator; The maximum allowable pressure boundary value and the maximum allowable flow boundary value are transformed into topological space inequalities corresponding to the continuous-time state vector; The continuously differentiable function characterizing the limit of the actuator is generated using the aforementioned topological space inequalities; The state mapping region where the continuously differentiable function takes values ​​greater than zero and equal to zero is defined as the desired safe set.

5. The method for self-optimizing process parameters in particleboard production according to claim 4, characterized in that, In S3, generating a secure topology space boundary includes: The partial derivative expansion is generated by calculating the Lie derivative of the continuously differentiable function along the state evolution trajectory of the nonlinear dynamic model. An extended class K function is introduced as a compensation term in the partial derivative expansion; The disturbance observation estimate and the truncation error limit are embedded into the partial derivative expansion in the form of boundary scaling to generate the adaptive control obstacle function constraint. The system state trajectory after being constrained by the adaptive control barrier function is strictly within the desired safety set, forming the safety topology space boundary.

6. The method for self-optimizing process parameters in particleboard production according to claim 1, characterized in that, In S4, the output nominal control command includes: Define a control Lyapunov function that represents the absolute deviation between the quality error vector and the expected factory quality standard; The continuous-time state vector and the quality error vector are input into the state feature extraction network layer of the policy gradient agent to obtain hidden layer features; The hidden layer features are input into the policy network nodes, and the policy network node parameters are updated by action exploration using the policy gradient ascent algorithm. The network output action that causes the value of the control Lyapunov function to exhibit a decreasing gradient is extracted as the nominal control command.

7. The method for self-optimizing process parameters in particleboard production according to claim 1, characterized in that, In S5, pruning the nominal control command using the secure topology space boundary includes: Minimizing the quadratic norm deviation between the quadratic programming variable to be optimized and the nominal control command is set as the basic optimization cost function; The adaptive control barrier function constraint is set as a safety boundary condition, and the descent gradient of the control Lyapunov function is set as an error convergence condition. The safety boundary conditions, the error convergence conditions, and the mechanical extreme values ​​of the underlying actuators are combined to form a set of physical hard constraint conditions. A penalized relaxation variable is introduced into the safety boundary conditions and error convergence conditions in the set of physical hard constraints to construct a pruning model with a relaxation factor.

8. The method for self-optimizing particleboard production process parameters according to claim 7, characterized in that, In S5, the instruction to solve for the optimal robust process parameters includes: The basic optimization cost function is combined with the pruning model with relaxation factor to transform it into a standard quadratic programming mathematical model equation; The standard quadratic programming mathematical model equations are solved online using an interior-point iterative algorithm. When the solution process converges, the solution that minimizes the basic optimization cost function and strictly satisfies the set of physical hard constraints is extracted as the optimal robust process parameter instruction.

9. The method for self-optimizing process parameters in particleboard production according to claim 1, characterized in that, In S6, converting the optimal robust process parameter instruction into a drive electrical signal and sending it to the logic controller for execution includes: Receive the optimal robust process parameter command output by the convex optimization solver; The optimal robust process parameter command is subjected to data frame unpacking operation to parse out the voltage control signal for the servo proportional valve of the hot press and the frequency control signal for the glue application frequency pump. The voltage control signal and the frequency control signal are synchronously sent to the corresponding logic controller via an industrial control bus network to execute physical drive.

10. A self-optimization system for particleboard production process parameters, characterized in that, A method for self-optimizing process parameters in particleboard production according to any one of claims 1-9, comprising the following modules: The state perception and dynamic modeling module is used to collect the operating indicators of the particleboard production line to construct a continuous-time state vector, collect the quality error vector output by the visual inspection station, and construct a nonlinear dynamic model containing time-varying unknown composite disturbance vectors based on the continuous-time state vector. The disturbance observation and reconstruction module is used to input the continuous-time state vector into the fixed-time disturbance observer, reconstruct the time-varying unknown composite disturbance vector, and calculate the disturbance observation estimate and the truncation error limit. The safe topology boundary generation module is used to extract a continuously differentiable function that represents the limit of the actuator, substitute the disturbance observation estimate and the truncation error limit into the continuously differentiable function, construct an adaptive control obstacle function constraint, and generate a safe topology space boundary. The nominal instruction optimization module is used to input the continuous-time state vector and the quality error vector into the policy gradient agent and output the nominal control instruction. The robust instruction trimming and solving module is used to input the nominal control instruction into the convex optimization solver, trim the nominal control instruction using the safe topology space boundary, and solve for the optimal robust process parameter instruction. The drive execution closed-loop module is used to convert the optimal robust process parameter instructions into drive electrical signals and send them to the logic controller for execution.

Images