Digital feedback type intelligent valve positioner torque motor control method

Abstract

This invention relates to the field of valve positioner control technology and discloses a digital feedback intelligent valve positioner torque motor control method, comprising: a microprocessor acquiring position signals and calculating errors to execute proportional-integral-derivative control, and combining a three-dimensional magnetization history state machine to output an asymmetric pulse width modulation signal; during the flywheel freewheeling window period when the pulse width modulation signal is turned off, controlling the lower tube field-effect transistor of the half-bridge drive circuit to conduct to construct a resistive freewheeling loop, performing sampling and separating the back electromotive force sequence; extracting the attenuation rate of the back electromotive force sequence and mapping it to a pneumatic mechanical damping coefficient, and updating the three-dimensional magnetization history state machine; and using a multilayer perceptron model to predict the equivalent delay parameter to dynamically reconstruct the integral gain coefficient. This invention eliminates semiconductor nonlinear voltage drop interference, realizes the quantitative mapping of mechanical physical state to electrical control parameters, prevents system overshoot divergence, and effectively improves the steady-state control accuracy of the valve.

Description

Technical Field

[0001] This invention relates to the field of valve positioner control technology, specifically to a digital feedback intelligent valve positioner torque motor control method. Background Technology

[0002] Existing digital valve positioners typically use conventional diodes for freewheeling when driving torque motors. However, under the hardware constraints of two-wire low-power systems, conventional freewheeling circuits generate semiconductor nonlinear clamping voltage drops. These voltage drops can interfere with or even directly mask the microvolt-level induced back electromotive force (EMF) signal generated by the armature's mechanical oscillations. Coupled with the influence of complex electromagnetic background noise in the environment, the microprocessor struggles to accurately extract and separate the effective back EMF sequence, preventing the system from achieving sensorless acquisition and high signal-to-noise ratio quantitative evaluation of pneumatic mechanical resistance states through electromagnetic feedback signals.

[0003] Due to the lack of effective feedback on the actual microscopic physical state, traditional control methods struggle to adapt to the dynamic changes in the on-site air source environment and mechanical friction. Furthermore, the inherent hysteresis nonlinearity of the torque motor's core, combined with pneumatic damping fluctuations and the hysteresis effect, means that the controller's fixed feedforward command cannot overcome the specific mechanical resistance disturbances in real time. This causes the valve positioner to easily deviate when reaching the target position, resulting in low steady-state control accuracy.

[0004] Furthermore, pneumatic actuators inherently exhibit pure time lag during the charging and discharging process, while conventional control systems often employ proportional-integral-derivative (PID) algorithms with fixed parameters. When a pneumatic system experiences a significant response delay, the fixed integral gain causes the controller to continuously accumulate position errors during the lag period. This excessive accumulation of control output before the actual system action not only fails to accelerate the initial response speed but also induces phase lag after the valve actually actuates. Excessive control action ultimately translates into mechanical oscillations and system divergence, making it difficult to effectively prevent overshoot and reducing the dynamic robustness of the valve control process. Summary of the Invention

[0005] To address the shortcomings of existing technologies, this invention provides a digital feedback intelligent valve positioner torque motor control method. This method solves the problems of existing digital valve positioners being unable to accurately extract weak back electromotive force signals to quantify pneumatic mechanical states due to interference from the nonlinear voltage drop of semiconductors in conventional freewheeling circuits, and the tendency for fixed control parameters to lead to low steady-state control accuracy and system overshoot divergence under the combined influence of torque motor hysteresis nonlinearity and pure time delay of the pneumatic system.

[0006] To achieve the above objectives, the present invention provides a digital feedback intelligent valve positioner torque motor control method, comprising the following steps: The microprocessor acquires the set position signal and the actual position signal to calculate the macroscopic position error, and executes the proportional-integral-derivative control algorithm to solve the basic target current; The microprocessor inputs the base target current, the sign of the rate of change of the base target current, and the pneumatic damping coefficient obtained in the previous control cycle into the three-dimensional magnetization history state machine for calculation, outputs the base duty cycle, and outputs an asymmetric pulse width modulation signal in combination with the base duty cycle. It injects a time difference perturbation into the coil of the torque motor and defines the turn-off edge of the asymmetric pulse width modulation signal as the time phase zero point. During the flywheel freewheeling window period when the asymmetric pulse width modulation signal is in the off state, the microprocessor controls the upper and lower field-effect transistors of the half-bridge drive circuit to be turned on to construct a resistive freewheeling loop. The microprocessor uses the time phase zero point as a reference to control the analog-to-digital converter to sample the resistive freewheeling loop to obtain the original voltage sample value, and separates the back electromotive force sequence from the original voltage sample value. The microprocessor extracts the decay characteristics of the back electromotive force sequence to obtain the decay rate, maps the decay rate to the aeromechanical damping coefficient of the current control cycle, and feeds it back to the three-dimensional magnetization history state machine to update the input parameters as the basic data for the next control cycle.

[0007] Furthermore, the microprocessor inputs multi-source physical parameters into the built-in multilayer perceptron neural network model to predict the aerodynamic delay time and obtain the equivalent delay parameter; based on the equivalent delay parameter, it dynamically reconstructs the integral gain coefficient of the proportional-integral-derivative control algorithm.

[0008] Furthermore, the specific process of calculating the basic target current using the proportional-integral-derivative (PID) control algorithm is as follows: the basic target current is obtained by summing the product of the basic proportional gain coefficient and the macroscopic position error of the current control cycle, the product of the dynamic integral gain coefficient of the current control cycle and the macroscopic position error accumulated by discrete time steps, and the product of the basic derivative gain coefficient and the error change rate of adjacent control cycles. When the basic target current is greater than (or exceeds) the preset coil safety bearing threshold, the microprocessor pauses (or stops) the integral term accumulation action of the current control cycle. When the basic target current does not exceed the preset coil safety bearing threshold, the normal integral term accumulation action is performed. The preset coil safety load threshold is the maximum continuous operating current parameter calibrated at the factory of the torque motor.

[0009] Furthermore, the specific process of outputting the asymmetric pulse width modulation signal based on the basic duty cycle is as follows: the three-dimensional magnetization history state machine includes two discrete two-dimensional static lookup table planes representing the magnetization rising branch and the magnetization falling branch; the microprocessor locks the corresponding discrete two-dimensional static lookup table plane according to the sign of the basic target current change rate, and performs a bilinear interpolation algorithm on the basic target current and the aeromechanical damping coefficient of the previous control cycle to obtain the basic duty cycle; the microprocessor calculates the perturbation time difference based on the slope derivative of the current hysteresis loop operating point, and applies positive and negative perturbation time differences to the basic duty cycle in two consecutive adjacent pulse width modulation cycles to form an asymmetric pulse width modulation signal with alternating conduction time.

[0010] Furthermore, the specific process of constructing the resistive freewheeling circuit and performing sampling to obtain the original voltage sample value is as follows: After the microprocessor outputs a control command to turn off the upper MOSFET of the half-bridge drive circuit, a fixed dead time is inserted between the drive signals before turning on the lower MOSFET, so that the lower MOSFET enters the deep linear conduction region to construct the resistive freewheeling circuit; the microprocessor, using the time phase zero point as a reference, controls the first sampling action of the analog-to-digital converter to be delayed by a time span equal to the dead time, and then begins to perform sampling on the resistive freewheeling circuit to obtain the original voltage sample value.

[0011] Further, the specific process of separating the back electromotive force sequence from the original voltage sample values ​​is as follows: the microprocessor performs time-domain phase-locked coherent cumulative averaging calculation on the original voltage sample values ​​with the same phase index across multiple consecutive perturbation cycles to extract the average coil terminal voltage; using discrete difference equations, the pure inductance transient voltage obtained by the static equivalent inductance and current time difference calculation is subtracted from the average coil terminal voltage, and the resistance voltage drop obtained by multiplying the actual current value of the resistive freewheeling circuit with the DC rated internal resistance and drain-source on-state resistance of the torque motor coil is subtracted, and the pure back electromotive force value is separated and output to form the back electromotive force sequence.

[0012] Furthermore, the specific process of extracting the decay characteristics of the back electromotive force sequence to obtain the decay rate is as follows: The microprocessor uses a local extremum search algorithm combined with hysteresis intervals to scan the back EMF sequence. When the difference between adjacent data points exceeds the preset noise tolerance threshold, and the time interval between two adjacent valid peaks is greater than half a cycle of the expected mechanical resonance period, the corresponding peak feature point is extracted; otherwise, it is determined as an invalid peak and is discarded. The voltage amplitude of each extracted peak feature point is compared with the preset lower limit voltage threshold and clamped. Then, a natural logarithmic transformation is performed, and the slope of the logarithmic amplitude changing with time is calculated using the least squares method. The absolute value of the slope is defined as the decay rate. If the number of captured peak feature points is less than three, the fitting operation is terminated and the decay rate of the previous control cycle is inherited. If the number of captured peak feature points is greater than or equal to three, then a natural logarithmic transformation and slope fitting operation are performed. The preset noise tolerance threshold is the value obtained by multiplying the background noise variance measured in the field environment when the system is in a static state by the set reliability coefficient. The preset minimum voltage threshold is the minimum resolvable voltage amplitude corresponding to the least significant bit of the analog-to-digital converter, set to prevent overflow of the logarithmic function operation.

[0013] Furthermore, the specific process of mapping the attenuation rate to the aeromechanical damping coefficient of the current control cycle is as follows: using the electromechanical equivalent equation, the attenuation rate is multiplied by the square root of the ratio of mechanical equivalent mass to comprehensive equivalent stiffness, and the aeromechanical damping coefficient is obtained by mapping.

[0014] Furthermore, the process of acquiring the multi-source physical parameters is as follows: the microprocessor performs a first-order inertial low-pass digital filter on the acquired aeromechanical damping coefficient; the aeromechanical damping coefficient, macroscopic position error, and valve stem movement speed after the first-order inertial low-pass digital filter are collected as multi-source physical parameters; and the multi-source physical parameters are resampled using a first-order linear interpolation algorithm to align the physical time sections. The extreme value normalization operation is performed on the aligned multi-source physical parameters. The denominator of the normalization process is the difference between the historical maximum and minimum values ​​of the parameters plus an artificially introduced small positive constant. The normalized data are then concatenated to construct a one-dimensional time-series input feature vector.

[0015] Furthermore, the specific process of obtaining the equivalent delay parameter for predicting the aerodynamic delay time is as follows: the microprocessor feeds the one-dimensional time-series input feature vector constructed from multi-source physical parameters into a multilayer perceptron neural network model containing a modified linear unit activation function for forward propagation calculation, and outputs the predicted aerodynamic delay time of the current control cycle. The predicted aerodynamic delay time of the current control cycle and the predicted aerodynamic delay time of the historical cache are weighted using an exponential forgetting factor based on the time span. The predicted aerodynamic delay time of the current control cycle is given the largest weight coefficient, while the weight coefficient of the predicted aerodynamic delay time of the historical cache decays exponentially with the passage of storage time. The equivalent delay parameter is obtained by weighted averaging.

[0016] This invention provides a digital feedback intelligent valve positioner torque motor control method. It has the following advantages: 1. This invention controls the conduction of the lower MOSFET of the half-bridge drive circuit during the flywheel freewheeling window to construct a resistive freewheeling loop. Combined with time-domain phase-locked coherent cumulative averaging calculation, the back EMF sequence is separated from the original voltage sample value. This eliminates the nonlinear clamping voltage drop interference of traditional diode freewheeling, effectively suppresses background electromagnetic noise, and enables the microprocessor to accurately extract the microvolt-level back EMF signal, providing a high signal-to-noise ratio data foundation for subsequent quantitative evaluation of aerodynamic drag state.

[0017] 2. This invention extracts the attenuation rate of the back electromotive force and maps it to the pneumatic mechanical damping coefficient, feeding it back to the three-dimensional magnetization history state machine to update the input parameters. This directly transforms the mechanical physical state into feedforward control parameters, enabling the output asymmetric pulse width modulation signal to adapt to changes in the field air source environment and mechanical friction. This overcomes the inherent hysteresis nonlinearity and pneumatic damping fluctuations of the torque motor, and improves the steady-state control accuracy of the valve positioner.

[0018] 3. This invention utilizes a built-in multilayer perceptron neural network model to perform forward calculations on multi-source physical parameters, predicts aerodynamic delay time in advance, and dynamically reconstructs the integral gain coefficient of the proportional-integral-derivative control algorithm accordingly. This can proactively reduce the cumulative effect of error integral when the pneumatic actuator system faces a long response lag, prevent system overshoot divergence caused by physical delay, and ensure the dynamic robustness of the valve control process. Attached Figure Description

[0019] Figure 1 A flowchart of a digital feedback intelligent valve positioner torque motor control method provided in an embodiment of the present invention; Figure 2 A flowchart of the macroscopic position closed-loop calculation sub-steps provided in this embodiment of the invention; Figure 3 A flowchart of the feedforward compensation calculation and active excitation output sub-steps provided in this embodiment of the invention; Figure 4 A flowchart of the sub-steps for extracting and cleaning extremely weak microscopic electromagnetic signals provided in an embodiment of the present invention; Figure 5A flowchart of the physical feature cross-domain mapping calculation sub-steps provided in this embodiment of the invention; Figure 6 A flowchart of the cross-domain prediction and closed-loop parameter adjustment sub-steps for delay features provided in an embodiment of the present invention; Figure 7 A comparison chart of the large step response control effect of the pneumatic valve positioner provided in the embodiments of the present invention. Detailed Implementation

[0020] 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.

[0021] The control method of this invention operates based on a two-wire power supply system. The hardware system includes a microprocessor, a half-bridge drive circuit, a torque motor, a pneumatic amplifier, a pneumatic actuator, and a position sensor. The half-bridge drive circuit includes an upper-side field-effect transistor and a lower-side field-effect transistor. The torque motor includes a coil and an armature. The microprocessor is equipped with an analog-to-digital converter, a timer, and non-volatile memory. To meet the stringent milliwatt-level power consumption constraints of the two-wire system, the microprocessor has a built-in DSP hardware coprocessor supporting single-cycle multiply-accumulate operations, executing complex neural networks and filtering algorithms using a time-division multiplexing low-power scheduling mechanism. The non-volatile memory stores a three-dimensional magnetization history state machine. The microprocessor is connected to the torque motor coil through the half-bridge drive circuit. The microprocessor's analog-to-digital converter sampling channel is connected across the torque motor coil via a differential filter network.

[0022] See attached document Figure 1 This invention provides a digital feedback intelligent valve positioner torque motor control method, comprising the following steps: S10, the microprocessor acquires the set position signal from the external input and the actual position signal from the position sensor to calculate the macroscopic position error, and the microprocessor executes the proportional-integral-derivative control algorithm to solve the basic target current for maintaining the steady state of the valve; S20, the microprocessor inputs the basic target current, the sign of the basic target current change rate, and the pneumatic damping coefficient obtained in the previous control cycle as addressing indexes into the three-dimensional magnetization history state machine in the non-volatile memory for piezo-magnetic coupling calculation. The three-dimensional magnetization history state machine outputs the basic duty cycle with feedforward compensation. The microprocessor calculates the perturbation time difference based on the slope derivative of the current hysteresis loop operating point. Combined with the basic duty cycle, it outputs an asymmetric pulse width modulation signal through a timer. The asymmetric pulse width modulation signal injects high-frequency time difference perturbation into the coil of the torque motor to excite the armature to generate mechanical oscillation. The microprocessor defines the off-edge of the asymmetric pulse width modulation signal as the time phase zero point. S30, during the flywheel freewheeling window when the asymmetric pulse width modulation signal is off, the microprocessor controls the upper and lower field-effect transistors of the half-bridge drive circuit to turn on to construct a resistive freewheeling loop. The microprocessor controls the analog-to-digital converter to perform high-speed differential sampling to obtain the original voltage sample value by using the time phase zero point as a reference. The microprocessor performs time-domain phase-locked coherent cumulative averaging calculation on the original voltage sample values ​​with the same phase index across multiple consecutive perturbation cycles to separate the back electromotive force sequence from the sampled data. S40, the microprocessor performs exponential fitting on the peak feature points of the back EMF sequence, extracts the decay rate of the back EMF decay envelope curve, and uses the electromechanical equivalent equation to map the decay rate to the aeromechanical damping coefficient of the current control cycle. The microprocessor latches the aeromechanical damping coefficient and feeds it back to the three-dimensional magnetization history state machine to update the input parameters as the basic data for the next control cycle. In S50, the microprocessor inputs multi-source physical parameters into the built-in multilayer perceptron neural network model to perform advanced prediction of aerodynamic delay time. The microprocessor then performs a weighted average of the single prediction result and historical cached data to obtain the equivalent delay parameter. Based on this equivalent delay parameter, the microprocessor dynamically reconstructs the integral gain parameter of the control algorithm using an exponentially decaying regulation law, thereby decoupling the phase lag risk caused by the delay.

[0023] See attached document Figure 2 In the specific implementation of step S10, the microprocessor performs macroscopic position closed-loop calculation, which includes the following sub-steps: S101, during the operation of the valve control system, in order to convert the macroscopic mechanical displacement expectation into specific electrical control commands, the system needs to establish a deviation model between the target and the feedback in real time. In this embodiment, the microprocessor acquires the set position signal from external input and the actual position signal fed back by the position sensor. In industrial field applications, the set position signal is usually issued by the distributed control system in the control room through a two-wire 4-20mA current loop. The microprocessor reads the analog voltage signal connected across the sampling resistor through its internally configured analog-to-digital converter and linearly converts it into a digital form of the set position value. The actual position signal is generated by the position sensor installed on the main valve stem of the pneumatic actuator. As a preferred method, the specific implementation of the position sensor includes, but is not limited to, a Hall effect sensor or a conductive plastic potentiometer. The microprocessor uses the analog-to-digital converter to convert the analog feedback voltage signal output by the position sensor into a digital form of the actual position value. To ensure the causality and timing matching of the control loop, the microprocessor synchronously captures the two analog signals based on a fixed sampling period set by a hardware timer, thereby ensuring the time alignment of multi-source input data within the same control cycle. After data alignment is completed, the microprocessor performs a difference calculation between the set position value and the actual position value to obtain the macroscopic position error. For the analog-to-digital conversion reading circuit of the two-wire 4-20mA loop signal and the hardware connection method of the Hall sensor, those skilled in the art can select and arrange the hardware according to conventional industrial instrument design specifications. The underlying hardware interface circuit and general data acquisition process are well-known technologies in this field and will not be elaborated upon here.

[0024] S102, after acquiring the instantaneous position deviation, considering the significant mechanical inertia and nonlinear friction of industrial valve systems, simple proportional control is insufficient to eliminate steady-state errors. Therefore, the microprocessor executes a discretized position-based proportional-integral-derivative (PID) control algorithm to calculate the fundamental target current for maintaining valve steady-state. The microprocessor substitutes the calculated macroscopic position error into the calculation equations of the control algorithm. The specific formula for calculating the fundamental target current is as follows: ; In the formula, Characterizing the first The basic target current for each control cycle; Characterizing the first Macroscopic position error per control cycle; Characterized by the previous control cycle, i.e., the first Macroscopic position error per control cycle; Characterizes the basic proportional gain coefficient, used to generate instantaneous linear control based on the current macroscopic position error value; The dynamic integral gain coefficient, which characterizes the current control cycle, is based on a pre-calibrated base value during the first run of the system. In subsequent runs, it inherits the integral gain parameter dynamically reconstructed by the system after the prediction delay of the previous control cycle in real time, thus forming a strict closed loop of parameter adaptation. The discrete-time step index represents the cumulative integral operation; Characterizes the fundamental differential gain coefficient, used to calculate the error rate of change between adjacent control cycles, predict the system motion trend, and provide advance damping control; Characterizing the first The macroscopic position error corresponding to each discrete-time step index; This is the summation symbol.

[0025] The initial values ​​of the aforementioned basic proportional gain coefficient, basic integral gain coefficient, and basic differential gain coefficient were pre-calibrated by technicians based on the Ziegler-Nichols critical proportionality rule and stored in non-volatile memory. Furthermore, to ensure the reliability of the control algorithm in actual engineering environments, this embodiment embeds anti-integral saturation logic within the algorithm. In specific implementation, the microprocessor determines whether the currently calculated basic target current exceeds a preset coil safety carrying capacity threshold. This coil safety carrying capacity threshold is determined by the maximum continuous operating current parameter calibrated at the torque motor's factory. When the basic target current exceeds this threshold, the microprocessor freezes the integral term accumulation action of the current control cycle to prevent severe mechanical overshoot of the valve under large step commands.

[0026] In step S103, the microprocessor internally caches the baseline target current calculated by the control algorithm. From a physical perspective, the calculated baseline target current represents the drive current reference required by the torque motor under ideal linear and hysteresis-free physical conditions to overcome the preload of the main valve stem return spring and the hydrodynamic resistance of the pipeline fluid. The reason for internally caching this data instead of directly outputting it to the drive stage is that traditional direct duty cycle drive would apply static deviations unchanged to the coils, making it difficult to overcome the inherent hysteresis effect of the torque motor. Based on these considerations, this baseline target current is stored as a pre-reference variable in the microprocessor's random access memory, serving as a pre-input parameter for inner-loop piezomagnetic coupling addressing in subsequent control flows.

[0027] See attached document Figure 3 In the specific implementation of step S20, the microprocessor performs feedforward compensation calculation and active excitation output. This implementation process includes the following sub-steps: S201, the electromagnetic force and control current of the torque motor are not in an ideal linear proportional relationship. Due to the irreversible magnetic domain reversal mechanism of the core material, the system exhibits hysteresis nonlinearity. Simultaneously, the mechanical damping generated by the pneumatic actuator stage influences the armature's motion. Based on these physical mechanisms, in this embodiment, the microprocessor utilizes a three-dimensional magnetization history state machine pre-installed in non-volatile memory to characterize this complex interaction. At the underlying data structure level, this state machine is divided into two discrete two-dimensional static lookup table planes representing the magnetization ascending and descending branches. The coordinate axes of the two-dimensional planes correspond to the basic target current and the pneumatic mechanical damping coefficient, respectively.

[0028] In step S202, the microprocessor uses the base target current, the sign of the base target current change rate, and the aeromechanical damping coefficient obtained in the previous control cycle as addressing indices, and inputs them into the three-dimensional magnetization history state machine for piezo-magnetic coupling calculation. Specifically, the microprocessor first performs discrete branch discrimination based on the sign of the base target current change rate (1 or -1), absolutely locking the corresponding ascending or descending two-dimensional data plane. Then, within this selected plane, for the two continuous variables of the base target current and the damping coefficient, it locates four adjacent known data nodes and performs a bilinear interpolation algorithm to avoid abrupt changes in physical quantities at grid points. This interpolation calculation model is expressed as: ; In the formula, Characterizes the base duty cycle obtained by table lookup and interpolation during the current control cycle; Characterizes the bilinear interpolation addressing function; Characterizes the basic target current of the current control cycle; Characterizes the difference between the basic target current in the current control cycle and the previous control cycle; A mathematical function representing the extracted symbol, with an output value of 1, 0, or −1, used to determine whether the torque motor is in the magnetized rising or falling branch; It represents the aeromechanical damping coefficient obtained from the feedback of the previous control cycle.

[0029] S203, after obtaining the basic duty cycle, the system needs to construct a specific time difference signal to overcome the local static friction of the armature. The microprocessor calculates the perturbation time difference based on the slope derivative of the current hysteresis loop operating point. In the digital discrete system, the microprocessor uses the central difference method to approximate continuous partial derivatives and combines it with safety limiting logic to prevent computational overflow. The microprocessor solves for the perturbation time difference according to the following formula: ; ; In the formula, The slope derivative of the hysteresis loop at the current operating point; The current differential step size is defined. To prevent the denominator from approaching zero and causing calculation errors, the microprocessor checks whether the step size parameter is lower than a preset minimum threshold before the division operation. If the value is below the threshold, the difference operation will be skipped and the derivative value will be set to a constant. Characterizes the perturbation time difference in the calculated output; The perturbation gain constant is characterized by its range and specific value, which are determined by the safety assessment test of the on-site pipeline vibration environment.

[0030] Under this design logic, a larger slope derivative indicates that the operating point is in a region with a significantly steep hysteresis loop, increasing the probability of the system experiencing position jumps. Multiplying by the perturbation gain constant adaptively generates a larger perturbation time difference to enhance the observed excitation signal. After acquiring the perturbation time difference, the microprocessor compares it with the system's preset maximum safe delay constraint value to limit the amplitude, ensuring that the calculation result does not exceed the upper limit of the underlying timer's range.

[0031] S204, the microprocessor outputs an asymmetric pulse width modulation (PWM) signal based on the base duty cycle. Traditional PWM signals maintain a constant pulse width over multiple consecutive cycles. In this embodiment, the microprocessor applies positive and negative perturbation time differences to the base duty cycle within two consecutive adjacent PWM cycles. While maintaining a constant average macroscopic drive current to ensure overall thrust stability, the asymmetric PWM signal injects high-frequency time difference perturbations into the torque motor coils. The asymmetric conduction time allocation model is as follows: ; ; In the formula, Periodic index characterizing micro-modulation; Characterizing the on-time of a field-effect transistor with odd-numbered periods; The on-time of a field-effect transistor with even-numbered periods; Characterizes the fixed pulse width modulation period duration set by the hardware timer.

[0032] Before outputting the aforementioned conduction time, the microprocessor executes asymmetric safety limiting logic, that is, it determines the extreme value after adding the perturbation. Has the effective period limit set by the hardware been exceeded? ,in To prevent short-circuit dead time, if the limit is exceeded, forced clamping is applied to prevent the underlying timer register from overflowing and causing malfunction. The alternating on-time disrupts the microscopic steady state of the coil current, forcibly exciting the armature to produce mechanical oscillations.

[0033] In the process of outputting the asymmetric pulse width modulation (PWM) signal, the microprocessor defines the off-edge of the PWM signal as the zero point of the time phase. Specifically, the microprocessor configures a compare-match interrupt mechanism for its internal hardware timer. When the timer's count reaches the on-time setting value for the corresponding period, causing the drive pin level to flip from high to low, a momentary interrupt is generated. The microprocessor uses this interrupt signal to reset the sampling timestamp, thereby providing a precise physical observation time reference for the subsequent extraction of the extremely weak back EMF signal.

[0034] See attached document Figure 4 In the specific implementation of step S30, the microprocessor performs the extraction and cleaning of extremely weak microscopic electromagnetic signals. This implementation process includes the following sub-steps: S301, based on the aforementioned asymmetric perturbation excitation injection, the system needs to capture the weak electromagnetic feedback caused by armature mechanical oscillation within an extremely short response time. In traditional drive topologies, freewheeling typically relies on parasitic diodes or external diodes, whose inherent semiconductor nonlinear on-state voltage drop can overwhelm useful signals at the microvolt level. In this embodiment, to overcome this physical limitation, during the flywheel freewheeling window when the asymmetric pulse width modulation signal is off, the microprocessor controls the operation of the upper and lower MOSFETs in the half-bridge drive circuit to construct a resistive freewheeling loop. Specifically, after the microprocessor outputs a control command to turn off the upper MOSFET, it does not immediately turn on the lower MOSFET. The microprocessor, based on its internal timer settings, forcibly inserts a fixed dead time between the drive signals of the two. The specific value of this dead time is determined in conjunction with the gate charge parasitic parameters of the upper and lower MOSFETs to absolutely prevent power supply short-circuit faults caused by simultaneous conduction of the upper and lower MOSFETs in the bridge arm. After the dead time delay, the lower-side MOSFET enters the deep linear conduction region, at which point the circuit exhibits purely linear impedance characteristics. This controlled action eliminates the nonlinear clamping effect of the diode, providing linear physical conditions for observing weak coil terminal voltages.

[0035] After entering the safe freewheeling window, the microprocessor controls the analog-to-digital converter (ADC) to perform high-speed differential sampling based on the time phase zero. To avoid the nonlinear clamping voltage drop and high-frequency oscillation noise caused by the freewheeling through the parasitic diode during the dead time, the microprocessor controls the ADC's first sampling action to be strictly delayed by a time span equal to the dead time, ensuring that effective voltage data is captured only after the lower MOSFET has fully entered the deep linear region.

[0036] S303, due to the compact structure and minimal micro-displacement of the armature in the torque motor, the induced electromotive force generated by its cutting of magnetic field lines is often drowned out by background electromagnetic noise in industrial environments and the quantization noise of the analog-to-digital converter. To address the Gaussian random fluctuation characteristics exhibited by single-sample data, the microprocessor performs a time-domain phase-locked coherent cumulative averaging calculation on the original voltage sample values ​​with the same phase index across multiple consecutive perturbation cycles. This algorithm is based on the principles of statistical signal processing: the mechanical motion excited by active perturbation is strictly locked to the triggering time, exhibiting strong coherence; while the noise floor of various systems exhibits incoherent characteristics with zero mean. Through aligned accumulation in the time domain, the amplitude of the deterministic signal accumulates linearly, while the energy of random noise cancels each other out in the averaging operation. The mathematical equation for this coherent cumulative averaging calculation is specifically expressed as follows: ; In the formula, Characterizes the high signal-to-noise ratio average coil terminal voltage extracted through coherent accumulation; The total number of cycles to be accumulated is defined, and its value takes into account both the real-time requirements of the system's dynamic response and the upper limit of the random access memory capacity in the microprocessor. It is usually selected as a positive integer greater than 10. The index representing the perturbation cycle in which the current accumulation execution is performed; The index of the specific phase sampling node within a single flywheel freewheel window; Characterization in the The perturbation period, the first The original voltage sample values ​​obtained at each phase sampling node.

[0037] S304, the average coil terminal voltage extracted after the above cleaning is not a single-dimensional motion feedback parameter; it still contains electromagnetic inertial components and ohmic losses. According to Kirchhoff's voltage law in dynamic circuit theory, the total terminal voltage is equal to the algebraic sum of the induced back electromotive force, the pure inductive transient voltage, and the voltage drop across the pure resistance of the circuit. Therefore, the microprocessor separates the back electromotive force sequence from the sampled data. In specific implementation, the microprocessor uses discrete difference equations to remove static electrical coupling terms; the decoupling calculation formula is as follows: ; In the formula, The characterization of the final decoupling and separation output of the first The pure back electromotive force values ​​of each phase node; The static equivalent inductance of the torque motor coil is characterized by a parameter that is measured and written into non-volatile memory during the device's factory initialization phase. Characterizes the actual value of the freewheeling current at the current phase node, which is sampled synchronously with the voltage data; Characterizes the actual value of the freewheeling loop current at the previous phase node; Characterizes the fixed sampling period duration set by the analog-to-digital converter; Characterizes the DC rated internal resistance of the torque motor coil; Characterizes the drain-source on-resistance of the lower-side field-effect transistor in the half-bridge drive circuit at the current operating temperature.

[0038] In the above formula, the intermediate current-time difference multiplication term aims to eliminate the interference voltage generated by the release of inductor energy, and the trailing term aims to compensate for the ohmic losses caused by the actual value of the current flowing through the copper wire and semiconductor wafer from the resistive freewheeling circuit. After the stripping operation is completed, the back electromotive force sequence obtained by the microprocessor purely characterizes the micromechanical oscillation speed characteristics of the armature, thereby realizing the cross-physical field extraction of core parameters.

[0039] See attached document Figure 5 In the specific implementation of step S40, the microprocessor performs cross-domain mapping calculations of physical features. This implementation process includes the following sub-steps: S401, the mechanical oscillation process of the armature after excitation exhibits typical underdamped decay motion, and its corresponding back EMF sequence shows an alternating waveform with gradually decreasing amplitude in the time domain. To quantify this decay trend, the microprocessor needs to extract the peak feature points of the back EMF sequence. In this embodiment, the microprocessor uses a local extremum search algorithm combined with hysteresis intervals to scan the back EMF sequence. When the algorithm determines that the data change trend has reversed, it not only requires that the difference between adjacent data points must exceed a preset noise tolerance threshold, but also introduces a time window dimension restriction, that is, the time interval between two adjacent effective peaks must be greater than half a period of the expected mechanical resonance period. This expected mechanical resonance period is pre-calculated and determined by the mechanical equivalent mass and comprehensive equivalent stiffness pre-stored in the non-volatile memory, thereby eliminating false extrema points. As a preferred implementation, the specific value of the above-mentioned noise tolerance threshold is determined by multiplying the background noise variance measured in the field environment when the system is in a static state by a set reliability coefficient, so as to ensure that the extraction logic can adapt to electromagnetic interference of different intensities in the industrial field.

[0040] S402, after acquiring valid peak data, the microprocessor performs exponential fitting on the peak feature points of the back EMF sequence to extract the decay rate of the back EMF decay envelope curve. According to second-order mechanical vibration theory, the amplitude envelope under viscous damping follows a natural exponential decay law. To reduce the computational resource overhead of the underlying hardware, the microprocessor transforms the nonlinear exponential fitting into a linear regression problem. In specific implementation, the microprocessor performs a natural logarithmic transformation on the extracted peak voltage amplitudes. Considering that the voltage amplitude may approach zero at the end of the decay of a weak signal, leading to logarithmic function overflow, the microprocessor compares all voltage amplitudes with a preset lower voltage threshold before performing the natural logarithmic transformation. This preset lower voltage threshold is the minimum resolvable voltage amplitude corresponding to the least significant bit (LSB) of the analog-to-digital converter, set to prevent logarithmic function overflow. If the actual voltage amplitude is lower than this lower voltage threshold, it is forcibly clamped to this lower voltage threshold. Subsequently, the microprocessor uses the least squares method to calculate the slope of the logarithmic amplitude over time. The absolute value of this slope is the decay rate of the back electromotive force decay envelope curve. In this calculation process, to prevent a singular regression matrix due to too few extreme points, the microprocessor performs a mandatory threshold check on the number of valid peaks. If fewer than three peaks are captured, the microprocessor will terminate the fitting operation for the current cycle and directly inherit the historical decay rate data from the previous control cycle; if three or more peaks are captured, the natural logarithmic transformation and fitting operation will proceed normally. The mathematical model for exponential fitting calculation is specifically expressed as follows: ; In the formula, Characterize the decay rate of the calculated and extracted back electromotive force decay envelope curve; Characterizes the total number of effective peaks successfully extracted within the current perturbation period; Sequence index representing effective peaks; Characterizing the first The timestamp corresponding to each valid wave peak is based on the physical timing reference of the zero point of the time phase. Functions representing the natural logarithm operation; Characterizing the first The voltage amplitude corresponding to each effective peak.

[0041] Since the denominator of the above formula is determined only by the timestamp, and the timestamp is a strictly monotonically increasing sequence in physical sampling, the denominator is always greater than zero provided that the number of effective peaks meets the standard, thus naturally avoiding the risk of overflow when divided by zero.

[0042] S403, after extracting the attenuation rate, the microprocessor uses the electromechanical equivalent equation to map the attenuation rate to the pneumatic mechanical damping coefficient currently experienced by the pneumatic amplifier. From a physical causal perspective, the energy dissipation rate of the armature's micro-oscillation is directly controlled by the viscous resistance of the surrounding medium. This resistance in the pneumatic valve positioner is mainly determined by the dynamic stiffness of the compressed air in the back pressure chamber of the pneumatic amplifier. Based on the above principle, the microprocessor establishes a transdomain mapping based on the damping relationship of the equivalent second-order system, and its solution formula is: ; In the formula, Characterize the current number The aeromechanical damping coefficient obtained from each control cycle; Characterizes the mechanical equivalent mass of the torque motor armature and connecting baffle; Characterizes the combined equivalent stiffness of the return spring and the static attraction force of the magnetic field.

[0043] The aforementioned mechanical equivalent mass and comprehensive equivalent stiffness are both inherent physical properties of the torque motor, ensuring the mathematical validity of the square root and division operations. The specific values ​​of these parameters are calibrated by the equipment manufacturer during the factory stage through a combination of finite element simulation and physical bench testing, and are permanently stored in the microprocessor's non-volatile memory. Through this dynamic equivalent equation, the microprocessor successfully transforms the attenuation characteristics of the electromagnetic signal dimension into a physical resistance state assessment value for the macroscopic pneumatic mechanical system.

[0044] S404, to ensure the dynamic state assessment effectively participates in the closed-loop control of the system, the microprocessor latches the pneumatic damping coefficient and feeds it back to the three-dimensional magnetization history state machine to update the input parameters. Considering that fluctuations in air source pressure and the compressibility of gas in industrial settings can easily introduce high-frequency computational noise, directly injecting the extracted transient results into the control feedforward loop could easily cause high-frequency mechanical oscillations in the servo mechanism. Therefore, before latching, the microprocessor performs a first-order inertial low-pass digital filter on the calculated pneumatic damping coefficient. Specifically, the discrete filtering implementation rule is that the filter output of the current cycle equals the product of the original calculated value of the current cycle and the preset filter coefficient, plus the product of the filter output of the previous cycle and the difference between the original and the preset filter coefficient. The value of the preset filter coefficient is strictly limited to a range greater than 0 and less than 1, and its value is determined by researchers through frequency domain analysis based on the typical frequency range of air source pressure fluctuations in the field, thus achieving an optimal balance between system response sensitivity and noise immunity stability. The filtered smoothed damping coefficient is written to a specific address in the microprocessor's random access memory. At the start of the next control cycle, the microprocessor reads the damping coefficient data at that address and uses it as the pre-associative index for multi-dimensional spatial addressing. This closed-loop parameter update mechanism enables the control duty cycle output of the torque motor's inner loop to adaptively follow the evolution of the on-site air source environment and mechanical friction state.

[0045] See attached document Figure 6 In the specific implementation of step S50, the microprocessor performs cross-domain prediction of latency characteristics and adjustment of closed-loop parameters. This implementation process includes the following sub-steps: S501, due to the compressibility of gas and the fluid resistance of long-distance connecting pipelines, the pneumatic actuator exhibits an inherent pure time lag between the issuance of a command and the generation of substantial physical displacement of the main valve stem. To address this lag effect, in this embodiment, the microprocessor uses a built-in lightweight multilayer perceptron neural network model to predict the pneumatic delay time. As a preprocessing step in the forward computation, the microprocessor collects multi-source physical parameters within a continuous sliding time window. From a physical causality analysis, the pneumatic delay time is not only controlled by the absolute position of the valve but is also strongly correlated with the current motion trend and the state of mechanical friction. Therefore, the microprocessor selects the pneumatic mechanical damping coefficient, macroscopic position error, and valve stem velocity obtained by performing first-order discrete difference calculation on the actual position signal as the initial inputs to the model. Considering that the aforementioned multi-source physical parameters originate from sensors or observation algorithms with different sampling rates, the microprocessor uses the baseline interrupt cycle of the control system as a unified time axis and employs a first-order linear interpolation algorithm to perform timestamp resampling on the asynchronous data, ensuring that all feature parameters are strictly aligned to the same physical time section. To eliminate scale imbalance and numerical overflow caused by data of different dimensions in matrix operations, the microprocessor performs extremum normalization with zero-divide-in-zero protection on the aligned original input sequence. The preprocessing formula is as follows: ; In the formula, The normalized eigenvalues ​​are represented by mappings to the dimensionless interval between 0 and 1. Characterizes the original physical parameters generated by the analog-to-digital converter or its internal algorithm; and These represent the historical minimum and maximum values ​​of the physical parameter, which are pre-fixed in non-volatile memory. It represents an artificially introduced minimal positive constant.

[0046] Since physical parameters may remain constant for a long time under steady-state conditions, causing the denominator of the range to approach 0, the insertion of this minimal positive constant ensures the validity of the division operation and the operational safety of the microprocessor. The normalized multi-source parameters are concatenated in time index order to construct a one-dimensional temporal input feature vector of a specific dimension.

[0047] After data preprocessing, the microprocessor feeds the one-dimensional temporal input feature vector into a multilayer perceptron neural network model for computation. In the specific network topology design, this neural network model sequentially includes an input layer, two fully connected hidden layers, and a single-node output layer. The hidden layer neurons map their features to each other using linear matrix multiplication and a nonlinear activation function. To accommodate the computational limitations of industrial microprocessors, a rectified linear unit (RCU) with high computational efficiency is chosen as the activation function. The data flow and forward propagation computation formulas for this model's hierarchical connections are specifically expressed as follows: ; ; In the formula, The feature tensor representing the output of the first hidden layer; The modified linear unit activation function is used to set all negative inputs to zero, thereby introducing the piecewise nonlinear approximation capability required by the system. Characterize the one-dimensional temporal input feature vector constructed in step S501; , and These represent the weight matrices of the first hidden layer, the second hidden layer, and the output layer, respectively. , and The weight matrix and the bias vectors, which represent the corresponding levels, constitute the core memory parameters of the model and are loaded from non-volatile memory into static random access memory during system initialization. The first character representing the final output of the model The predicted aerodynamic delay time for each control cycle, which physically represents the system's advanced quantitative assessment of the impending aerodynamic lag span.

[0048] S503, to ensure that those skilled in the art can fully reproduce the construction and deployment of the above-mentioned prediction model, the offline training steps and key parameters of the multilayer perceptron neural network model are disclosed here. During the calibration training phase before the equipment leaves the factory, technicians apply step test signals containing various frequencies and amplitudes to the torque motor on a physical test bench, while simultaneously using a high-precision laser displacement sensor to synchronously capture the actual response time of the valve stem at microsecond-level resolution. The training label of the sample data is strictly defined as the absolute time difference between the output time of the drive command and the moment when the valve stem displacement changes abruptly. When constructing the dataset, technicians divide the collected total samples into a training set and a validation set in a fixed ratio of eight to two. The validation set data does not participate in the direct update of the forward and backward gradients, but is dedicated to monitoring the generalization ability of the model through cross-validation during training. The training process uses mean squared error as the core loss function and introduces a weight decay regularization term. In the host computer environment, the network parameters are continuously updated iteratively through the backpropagation algorithm combined with an adaptive moment estimation optimizer. If the evaluation loss on the validation set does not decrease for ten consecutive iterations, the system triggers an early stopping mechanism to terminate training, thereby preventing the model from overfitting to a specific calibrated pipeline length. Finally, the matrix parameters with generalization capabilities are distributed and solidified into the microprocessor.

[0049] S504, after obtaining the instantaneous predicted aerodynamic delay time, the system needs to dynamically reconstruct the basic closed-loop control loop to decouple the phase lag risk caused by the delay. To avoid abrupt changes in control parameters due to high-frequency disturbances caused by a single predicted extreme value, the microprocessor does not make a one-sided judgment on the single model output, but instead performs a weighted average of the current predicted value and historical cached data to obtain a smooth equivalent delay parameter. In the implementation of the weighted logic, the microprocessor uses an exponential forgetting factor based on the time span to assign weights to the predicted aerodynamic delay time of the current control cycle and the predicted aerodynamic delay time of the historical cache, assigning the largest weight coefficient to the predicted aerodynamic delay time of the current control cycle. The weight coefficient of the predicted aerodynamic delay time of the historical cache decays exponentially with the storage time, and the sum of all weight coefficients is constantly calibrated to one, ensuring the mathematical equivalence of the weighted mapping dimension. Based on this equivalent delay parameter, the microprocessor uses an exponential decay adjustment law to dynamically reconstruct the integral gain, and its reconstruction model is as follows: ; In the formula, Characterizing the first The actual integral gain coefficient after dynamic updates for each control cycle; Characterizes the basic integral gain coefficient set at the factory delivery stage; A function representing an exponential operation with the natural constant as its base; The delay sensitivity adjustment coefficient is characterized by a range of positive real numbers, which is determined by the researchers based on the root locus stability margin of the closed-loop system. The equivalent delay parameter is represented after multi-dimensional weighted smoothing.

[0050] From a control mechanism perspective, the technical objective of this step is to adaptively reduce the integral accumulation effect of the controller when the system calculates a long predicted aerodynamic delay time. This dynamic reconfiguration mechanism cuts off the integral accumulation of error dead zone caused by physical delay, preventing destructive mechanical overshoot during system recovery and ensuring the dynamic robustness of valve positioning.

[0051] To avoid the risk of drastic bumpless transfer caused by sudden changes in the integral gain coefficient in the positional PID control algorithm, the microprocessor enforces synchronous shift substitution compensation logic for the discrete error accumulator while updating the integral gain. The microprocessor scales and rewrites the accumulator according to the following formula: ; In the formula, Characterizing the first Each control cycle executes the new value of the error integral accumulator after synchronous translation substitution compensation; Characterized by the previous control cycle, i.e., the first The old value of the error integral accumulator retained for each control cycle; The dynamic integral gain coefficient representing the previous control cycle; Characterizing the first The actual integral gain coefficient after dynamic reconstruction of each control cycle; Characterizing the first Instantaneous macroscopic position error for each control cycle.

[0052] This dynamic reconstruction and smooth switching mechanism of the integral term not only cuts off the accumulation of error dead zone integral caused by physical delay, but also prevents destructive operation of the controller caused by sudden parameter jumps, ensuring the ultimate dynamic robustness of the valve positioning system.

[0053] Specific application examples (industrial scenario reproduction): Application scenario settings: The cooling water flow control system of a chemical plant uses a two-wire linear-stroke pneumatic diaphragm regulating valve. The system is powered by a 4-20mA circuit (total power consumption is strictly limited to within 40mW). Due to pipeline aging, the valve experiences significant changes in friction, and the air supply pipeline is quite long (approximately 5 meters), resulting in pneumatic hysteresis.

[0054] The control process is executed as follows: S10 (Macroscopic Position Closed-Loop Calculation): At this time, the control room issues a command to jump the valve opening from 20% to 60% (corresponding to a sudden change in the current signal from 7.2mA to 13.6mA).

[0055] The microprocessor acquires the current actual position (20%) fed back by the Hall sensor and calculates the macroscopic position error. .

[0056] The discrete PID algorithm is used to calculate the basic target current based on the current error. Since the coil safety threshold (25mA) was not exceeded, the anti-integral saturation logic was not triggered, and the value was stored in the internal cache.

[0057] S20 (Feedforward Compensation Solution and Active Excitation Output): The instruction at this point requires opening the valve, which corresponds to the magnetized rising branch. The microprocessor will then use the damping coefficient evaluated in the previous cycle. and Input a three-dimensional magnetization history state machine, and obtain the basic duty cycle by bilinear interpolation and table lookup. .

[0058] Because the operating point is located in a region of steep hysteresis loop, the microprocessor calculates the perturbation time difference. .

[0059] The timer (with a period set to 100μs) begins to output asymmetric PWM alternately: 54μs for odd-numbered periods and 50μs for even-numbered periods. This high-frequency micro-perturbation causes the torque motor armature to generate subtle high-frequency mechanical oscillations, converting static friction into dynamic friction.

[0060] S30 (Extraction and Cleaning of Extremely Weak Microscopic Electromagnetic Signals): During the 46μs / 50μs window period when the PWM is off, the lower MOSFET turns on. The microprocessor starts the ADC (sampling rate 2Msps) after the dead time (200ns).

[0061] By performing coherent phase-locked accumulation in the time domain over 16 perturbation cycles, the 50Hz power frequency interference and ADC quantization noise in the industrial field were eliminated. After stripping the voltage drop of inductors and resistors, the extremely weak back electromotive force sequence with a peak value of only about 45mV was successfully extracted.

[0062] S40 (Physical Feature Cross-Domain Mapping Calculation): The microprocessor locks the first three valid peaks of the back EMF sequence (e.g., 45mV, 36mV, 29mV).

[0063] Perform log-linear fitting to calculate the envelope attenuation rate. By combining the factory-preserved equivalent mass and stiffness parameters, the actual current aeromechanical damping coefficient is mapped. (Increased damping indicates high back pressure in the air chamber or dry valve stem.) This value, after passing through a first-order low-pass filter, is fed back to the state machine to update parameters.

[0064] S50 (Cross-domain prediction of delayed features and adjustment of closed-loop parameters): The microprocessor will [damping] the current state ,error ,speed Input the built-in multilayer perceptron (MLP) neural network.

[0065] The model's forward computation accurately predicted the challenges this major leap would face. The aerodynamic pure time lag is seconds.

[0066] To prevent the PID integrator from accumulating excessively within this 1.2-second period, leading to subsequent overshoot, the microprocessor uses an exponential decay control law to adjust the integral gain. The nonlinearity is instantly reduced from the base value of 0.5 to 0.08. Simultaneously, perturbationless translational substitution compensation is performed, and the integral accumulator is rewritten.

[0067] Result: After 1.2 seconds, the pneumatic thrust overcame the static friction, and the valve began to move. Due to the effective suppression of integral, the valve smoothly reached the 60% position, achieving zero overshoot.

[0068] Experimental verification and effect comparison: To verify the effectiveness of the invention, a comparative test was conducted on a standard test bench (equipped with a 5-meter gas supply pipeline and a set gas supply pressure of 0.4 MPa).

[0069] Benchmark group (traditional method): adopts conventional position PID control + constant duty cycle PWM drive + solid-state relay freewheeling.

[0070] Experimental group (this invention): The back electromotive force is extracted by using neural network delay prediction adaptive PID + asymmetric perturbation PWM + resistive freewheeling of the lower tube.

[0071] Macroscopic large step response test (testing the overshoot prevention capability of decoupling delay) Issue a step instruction from 20% to 60%.

[0072] Traditional method: There is a 1.3-second pneumatic lag. During the lag period, the traditional PID integrator experiences deep saturation (windup), causing the valve to move directly over 60% of the target value, reaching an extreme value of 72.8%, resulting in severe mechanical overshoot. It takes about 8.5 seconds to slowly stabilize after multiple wave-like oscillations.

[0073] This invention: A neural network successfully predicts latency and dynamically reconstructs the integral gain. After the valve moves, it reaches a critical damping state, with the overshoot reduced to less than 0.2%, and the settling time is only 3.2 seconds.

[0074] Microscopic small-step tracking test (testing the ability to overcome dead zones and hysteresis) Continuously issue tiny step commands with a step size of 0.5% (to test the harsh friction dead zone).

[0075] Traditional methods exhibit a significant step / stickiness effect. Due to hysteresis and static friction, the valve does not move when the command changes, and after accumulating two or three commands, it suddenly jumps by 2%, failing to achieve a precise positioning accuracy of 0.5%.

[0076] This invention utilizes feedforward compensation from a three-dimensional magnetization history state machine and time-difference perturbation via asymmetric PWM injection to ensure the valve stem remains in a state of dynamic friction. For every 0.5% command, the valve responds accurately within 0.8 seconds, with a steady-state error ≤0.1%, eliminating viscous jumps.

[0077] Table 1. Summary of Data Comparison Conclusion Analysis: Combining Table 1 and Figure 7 It can be seen that the control method of the present invention has significant advantages over traditional control methods when dealing with complex working conditions with aerodynamic delay.

[0078] Figure 7 This is a comparison chart of the large step response control effects of pneumatic valve positioners. The horizontal axis represents time (s), with a scale range of 0–10 seconds; the vertical axis represents the actual valve displacement (%), with a scale range of 10–80. The chart includes data lines representing three different physical states and related local feature annotations. First, the thin dotted line in the diagram represents the target position. This target position jumps from 20% to 60% in 1 second, simulating a large step control command in an industrial setting.

[0079] Secondly, within a time span of approximately 1 to 2.3 seconds, the inherent pure lag range of pneumatic systems and the corresponding indicator line are marked below the chart. During this period, due to the fluid resistance of the long-distance air pipe and the compressibility of the gas, the actual valve displacement remains at 20% without any substantial action, clearly demonstrating the inherent physical delay characteristics of the pneumatic system.

[0080] Secondly, the thin dashed line in the figure represents the actual displacement curve of the traditional control method. After crossing the pure time delay interval, due to severe integral saturation generated by the traditional PID control during the time delay, the dashed line rises sharply and penetrates the 60% target line. The figure clearly indicates the severe mechanical overshoot generated by this method using the circle symbol (○), and marks the extreme point of this overshoot as the maximum value of 72.8%. Subsequently, the curve undergoes obvious up-and-down wave-like oscillations before slowly approaching the target value.

[0081] Finally, the bold solid line in the figure represents the actual displacement curve of the method in this embodiment of the invention. As can be seen from the comparison, after experiencing the same aerodynamic inherent pure time lag, thanks to the advanced prediction and dynamic reconstruction of the integral gain of the multilayer perceptron neural network in this invention, the bold solid line exhibits a significantly smooth upward trend. As indicated by the partial text in the figure, this invention achieves adaptive parameter reconstruction: smooth and without overshoot, the bold solid line precisely matches 60% of the set target position, no limit exceedances occur throughout the process, and the time to reach steady state is shorter than that of traditional control methods.

[0082] In conclusion, Figure 7 The waveform characteristics and the quantitative data in Table 1 corroborate each other, fully demonstrating that the control method of the present invention can effectively decouple the phase lag risk caused by the delay, cut off the integral accumulation, and ensure the ultimate dynamic robustness and positioning accuracy of the intelligent valve positioner in harsh industrial environments.

Claims

1. A digital feedback intelligent valve positioner torque motor control method, characterized in that, Includes the following steps: The microprocessor acquires the set position signal and the actual position signal to calculate the macroscopic position error, and executes the proportional-integral-derivative control algorithm to solve the basic target current; The microprocessor inputs the base target current, the sign of the rate of change of the base target current, and the pneumatic damping coefficient obtained in the previous control cycle into the three-dimensional magnetization history state machine for calculation, outputs the base duty cycle, and outputs an asymmetric pulse width modulation signal in combination with the base duty cycle. It injects a time difference perturbation into the coil of the torque motor and defines the turn-off edge of the asymmetric pulse width modulation signal as the time phase zero point. During the flywheel freewheeling window period when the asymmetric pulse width modulation signal is in the off state, the microprocessor controls the upper and lower field-effect transistors of the half-bridge drive circuit to be turned on to construct a resistive freewheeling loop. The microprocessor uses the time phase zero point as a reference to control the analog-to-digital converter to sample the resistive freewheeling loop to obtain the original voltage sample value, and separates the back electromotive force sequence from the original voltage sample value. The microprocessor extracts the decay characteristics of the back electromotive force sequence to obtain the decay rate, maps the decay rate to the aeromechanical damping coefficient of the current control cycle, and feeds it back to the three-dimensional magnetization history state machine to update the input parameters as the basic data for the next control cycle.

2. The digital feedback intelligent valve positioner torque motor control method according to claim 1, characterized in that, Also includes: The microprocessor inputs multi-source physical parameters into a built-in multilayer perceptron neural network model to predict aerodynamic delay time and obtain equivalent delay parameters. Based on the equivalent delay parameter, the integral gain coefficient of the proportional-integral-derivative control algorithm is dynamically reconstructed.

3. The digital feedback intelligent valve positioner torque motor control method according to claim 1, characterized in that, The process of using the proportional-integral-derivative (PID) control algorithm to calculate the basic target current is as follows: The basic target current is obtained by summing the product of the basic proportional gain coefficient and the macroscopic position error of the current control cycle, the product of the dynamic integral gain coefficient of the current control cycle and the macroscopic position error accumulated by discrete time steps, and the product of the basic differential gain coefficient and the error change rate of the adjacent control cycle through the proportional-integral-derivative control algorithm. When the basic target current exceeds the preset coil safety bearing threshold, the microprocessor freezes the integral term accumulation operation of the current control cycle; when the basic target current does not exceed the preset coil safety bearing threshold, the normal integral term accumulation operation is performed. The preset coil safety load threshold is the maximum continuous operating current parameter calibrated at the factory of the torque motor.

4. The digital feedback intelligent valve positioner torque motor control method according to claim 1, characterized in that, The process of outputting an asymmetric pulse width modulation signal based on the aforementioned basic duty cycle is as follows: The three-dimensional magnetization history state machine includes two discrete two-dimensional static lookup table planes representing the magnetization rising branch and the magnetization falling branch; The microprocessor locks the corresponding discrete two-dimensional static lookup table plane according to the sign of the basic target current change rate, and performs a bilinear interpolation algorithm on the basic target current and the aeromechanical damping coefficient of the previous control cycle to obtain the basic duty cycle. The microprocessor calculates the perturbation time difference based on the slope derivative of the current hysteresis loop operating point, and applies positive and negative perturbation time differences to the base duty cycle in two consecutive adjacent pulse width modulation cycles, forming an asymmetric pulse width modulation signal with alternating conduction time.

5. The method for controlling the torque motor of a digital feedback intelligent valve positioner according to claim 1, characterized in that, The microprocessor controls the upper and lower MOSFETs of the half-bridge drive circuit to conduct, thereby constructing a resistive freewheeling circuit. The process by which the microprocessor, using the zero-point of the time phase as a reference, controls the analog-to-digital converter to sample the resistive freewheeling circuit and obtain the original voltage sample value is as follows: After the microprocessor outputs a control command to turn off the upper field-effect transistor of the half-bridge drive circuit, a fixed dead time is inserted between the drive signals to turn on the lower field-effect transistor, so that the lower field-effect transistor enters the deep linear conduction region to construct the resistive freewheeling circuit. The microprocessor, using the zero point of the time phase as a reference, controls the first sampling action of the analog-to-digital converter to be delayed by a time span equal to the dead time, and then begins to sample the resistive freewheeling circuit to obtain the original voltage sample value.

6. The digital feedback intelligent valve positioner torque motor control method according to claim 1, characterized in that, The process of separating the back electromotive force sequence from the original voltage sample values ​​is as follows: The microprocessor performs a time-domain phase-locked coherent cumulative averaging calculation on the original voltage sample values ​​with the same phase index across multiple consecutive perturbation cycles to extract the average coil terminal voltage; Using the discrete difference equation, the pure inductance transient voltage obtained by calculating the static equivalent inductance and current time difference is subtracted from the average coil terminal voltage, and the resistance voltage drop obtained by multiplying the actual current value of the resistive freewheeling circuit with the DC rated internal resistance and drain-source conduction internal resistance of the torque motor coil is subtracted. The pure back electromotive force values ​​are then separated to form the back electromotive force sequence.

7. The method for controlling the torque motor of a digital feedback intelligent valve positioner according to claim 1, characterized in that, The process by which the microprocessor extracts the decay characteristics of the back electromotive force sequence to obtain the decay rate is as follows: The microprocessor uses a local extremum search algorithm combined with hysteresis interval to scan the back EMF sequence. When the difference between adjacent data points exceeds the preset noise tolerance threshold and the time interval between two adjacent valid peaks is greater than half a cycle of the expected mechanical resonance period, the corresponding peak feature points are extracted. Otherwise, they are determined to be invalid peaks and are removed. After comparing and clamping the voltage amplitude of each extracted peak feature point with a preset lower voltage threshold, a natural logarithmic transformation is performed, and the slope of the logarithmic amplitude changing with time is calculated using the least squares method. The absolute value of the slope is defined as the attenuation rate. If the number of captured peak feature points is less than three, the fitting operation is terminated and the decay rate of the previous control cycle is inherited. If the number of captured peak feature points is greater than or equal to three, then the natural logarithmic transformation and slope fitting operation are performed. The preset noise tolerance threshold is the value obtained by multiplying the background noise variance measured in the field environment when the system is in a static state by a set reliability coefficient. The preset minimum voltage threshold is the minimum resolvable voltage amplitude corresponding to the least significant bit of the analog-to-digital converter, set to prevent overflow of the logarithmic function operation.

8. The method for controlling the torque motor of a digital feedback intelligent valve positioner according to claim 1, characterized in that, The process of mapping the attenuation rate to the aeromechanical damping coefficient of the current control cycle is as follows: The aeromechanical damping coefficient is obtained by multiplying the attenuation rate by the square root of the ratio of mechanical equivalent mass to comprehensive equivalent stiffness using the electromechanical equivalent equation.

9. The method for controlling the torque motor of a digital feedback intelligent valve positioner according to claim 2, characterized in that, The process of obtaining the multi-source physical parameters is as follows: The microprocessor performs a first-order inertial low-pass digital filter on the acquired aeromechanical damping coefficient. The pneumatic damping coefficient, macroscopic position error, and valve stem speed after the first-order inertial low-pass digital filter are collected as multi-source physical parameters. The multi-source physical parameters are resampled using a first-order linear interpolation algorithm to align the physical time sections; The aligned multi-source physical parameters are subjected to extreme value normalization. The denominator of the normalization process is the difference between the historical maximum and minimum values ​​of the parameters plus an artificially introduced small positive constant. The normalized data are then concatenated to construct a one-dimensional time-series input feature vector.

10. The digital feedback intelligent valve positioner torque motor control method according to claim 9, characterized in that, The microprocessor inputs the multi-source physical parameters into the built-in multilayer perceptron neural network model to perform aerodynamic delay time prediction, and the process of obtaining the equivalent delay parameter is as follows: The microprocessor feeds the one-dimensional time-series input feature vector constructed from the multi-source physical parameters into the multilayer perceptron neural network model containing the modified linear unit activation function for forward propagation calculation, in order to perform the prediction of the aerodynamic delay time and output the predicted aerodynamic delay time of the current control cycle. The predicted aerodynamic delay time of the current control cycle and the predicted aerodynamic delay time of the historical cache are weighted using an exponential forgetting factor based on the time span. The predicted aerodynamic delay time of the current control cycle is given the largest weight coefficient, while the weight coefficient of the predicted aerodynamic delay time of the historical cache decays exponentially with the passage of storage time. The equivalent delay parameter is obtained by weighted averaging.

Images