Adaptive Control Method for Ultrasonic Welding Power Based on Inductive Temperature Rise Feedback

Abstract

This invention relates to an adaptive power control method for ultrasonic welding based on inductor temperature rise feedback, aiming to solve problems such as the difficulty in accurately identifying the thermodynamic parameters of complex multi-layered inductor windings, inaccurate temperature rise trend prediction, and power regulation lag. The core scheme is as follows: During the no-load start-up phase, a stepped power excitation is executed, and temperature time-series signals are simultaneously acquired at multiple points. After signal denoising, feature separation, and nonlinear least-squares fitting, a physically interpretable third-order thermal inertia model of the inductor is constructed. This model drives real-time temperature rise prediction during welding operation, and combined with slope identification, power feedforward pre-adjustment is achieved. Simultaneously, residual extraction and lightweight PI feedback are used to correct power control commands in a closed loop. The feedforward and feedback work together to dynamically allocate output power, achieving precise control of inductor temperature rise. After welding, the accumulated deviation drives online adaptive fine-tuning of model parameters, improving the long-term prediction and control accuracy of the system. This method adapts to varying loads, improving welding energy efficiency and equipment safety.

Description

Technical Field

[0001] This invention relates to the field of ultrasonic welding temperature rise feedback and dynamic power control technology, and in particular to an ultrasonic welding power adaptive control method based on inductive temperature rise feedback. Background Technology

[0002] Current dynamic power control technology in ultrasonic welding primarily relies on a closed-loop feedback adjustment mechanism based on the physical state of the welding load and the inductor temperature rise signal to achieve control over the stability and consistency of the welding process. Existing adaptive power adjustment methods typically include temperature sensor acquisition of inductor surface temperature rise, power allocation strategy optimization using temperature rise rate thresholds, feedback control based on PID (proportional-integral-derivative) parameter tuning, cooling capacity modeling compensation, and compensation strategies such as resonant frequency tracking or voltage overshoot compensation. These solutions are widely used in mainstream ultrasonic welding equipment to improve the energy distribution efficiency and process robustness of the equipment under complex load variations. Some published patents and papers also suggest improving the accuracy and response speed of welding quality control through multi-source data clustering and temperature rise rate discrimination.

[0003] However, existing technologies generally face the following core problems: First, the dynamic response capability of the inductor temperature rise feedback link is insufficient, and the power adjustment action cannot track the changes in welding load in real time, resulting in significant adjustment lag; second, the closed-loop control algorithm has limited adaptability to sudden disturbances and ultrasonic vibration phase, and the accuracy of dynamic distribution of welding energy urgently needs to be improved; third, under the constraints of software and hardware resources, existing temperature rise feedback schemes are unable to overcome the physical limits brought about by thermal inertia and cannot achieve feedforward control that actively manages thermal dynamics. Summary of the Invention

[0004] This application provides an adaptive control method for ultrasonic welding power based on inductive temperature rise feedback, which aims to solve one of the problems or issues of the prior art mentioned in the background section.

[0005] The ultrasonic welding power adaptive control method based on inductive temperature rise feedback provided in this application specifically includes: S1: During the no-load start-up phase of the ultrasonic welding equipment, a stepped power excitation sequence is executed and the time response curves of multiple temperature sensors arranged on the surface of the wire winding, the inner wall of the skeleton, and the interface of the impregnation varnish curing layer are collected simultaneously to obtain the original temperature rise time series dataset containing the thermal conductivity of the surface fiber varnish film, the thermal diffusion of the intermediate wire bundle winding structure, and the overall thermal mass of the winding.

[0006] S2: Based on the original temperature rise time series dataset, construct an inductor third-order thermal inertia model, and use nonlinear least squares method to fit and calculate the equivalent heat capacity and thermal resistance of the surface fiber coating, the thermal diffusion time constant of the intermediate filament winding structure, and the overall winding thermal mass-dominant time constant, generating a parameter set of the inductor third-order thermal inertia model with physical interpretability.

[0007] S3: During the welding operation phase, the inductor surface temperature is collected in real time as the main input signal, and the main input signal is input into the inductor thermal inertia identification module constructed based on the parameter set of the third-order thermal inertia model of the inductor, and the predicted value sequence of temperature rise trend in three key time domain windows of 50 milliseconds, 100 milliseconds and 200 milliseconds is generated.

[0008] S4: Perform a difference calculation between the current measured temperature and the predicted temperature rise trend sequence to extract the transient disturbance residual signal that cannot be covered by the third-order thermal inertia model of inductance, and use the transient disturbance residual signal as the input variable of the PI feedback controller to generate a closed-loop correction component.

[0009] S5: Based on the slope change rate of the predicted temperature rise value sequence, identify the state characteristics of the temperature rise transitioning from a gradual transition to an inflection point or entering a plateau region, trigger the feedforward channel to generate a power gradient pre-adjustment action command, and delay the power gradient pre-adjustment action command to the next zero voltage crossover interval when the vibration phase deviation is detected to exceed the limit.

[0010] S6: Determine whether the absolute value of the transient disturbance residual signal exceeds a set threshold. If it does, perform integral accumulation processing on the closed-loop correction component; otherwise, perform proportional amplification processing to generate the final feedback correction command.

[0011] S7: The power gradient pre-adjustment action command and the feedback correction command are vector-synthesized to generate a composite power adjustment command that includes both feedforward prediction and feedback correction characteristics.

[0012] S8: After each welding, compare the cumulative deviation pattern between the actual temperature rise trajectory and the predicted temperature rise trend sequence. If the deviations show the same direction for three consecutive times, start the parameter update process and only incrementally correct the surface thermal resistance parameter in the equivalent heat capacity and thermal resistance of the surface fiber coating to complete the online fine-tuning of the model.

[0013] The ultrasonic welding power adaptive control method based on inductive temperature rise feedback provided in this application has the following beneficial effects: (1) By constructing a physically interpretable third-order thermal inertia model and combining it with a multi-point temperature response identification mechanism under stepped power excitation, the thermal inertia characteristics of inductive components in the ultrasonic welding process are transformed from a negative factor regarded as hysteresis disturbance in traditional control into a modelable and predictable feedforward control resource for the first time. This significantly improves the system's ability to predict temperature rise trends and its response foresight. Compared with traditional methods that rely on temperature change rate threshold triggering adjustment or are based on fixed PID parameter tuning, this scheme effectively overcomes the control hysteresis problem caused by thermal response delay. It completes the accurate identification of key heat conduction parameters in the no-load start-up stage, enabling the equipment to implement power gradient pre-adjustment 0.8 mechanical vibration cycles before the welding load changes abruptly, greatly reducing the risk of overshoot and significantly improving the stability and robustness of temperature control under dynamic conditions.

[0014] (2) A feedforward-feedback dual-path collaborative control architecture is adopted. The feedforward path actively adjusts the output power based on the model-driven multi-time-domain temperature rise prediction results. The feedback path only activates the integral action when the residual exceeds the set threshold. During other periods, proportional control is maintained to suppress oscillation. This ensures control accuracy while avoiding the cumulative error and high-frequency oscillation defects that are easily caused by traditional full closed-loop control. Furthermore, a welding phase locking mechanism is introduced to strictly constrain all power adjustment actions to be executed within the zero-voltage crossing interval of the ultrasonic vibration cycle. A phase offset fault-tolerant buffer is configured to ensure that the power command switching does not interfere with the transducer resonance state and effectively prevent resonance loss due to power sudden change. This significantly improves energy transfer efficiency and process stability. This design achieves the organic unity of high dynamic response and strong anti-disturbance capability without the need for complex parameter tuning. It is especially suitable for harsh scenarios with frequent load jumps in continuous multi-welding operations.

[0015] (3) The system integrates an online fine-tuning mechanism. By comparing the cumulative deviation between the actual temperature rise trajectory and the model's predicted trajectory, the surface thermal resistance parameter is adaptively corrected within a limited range only under continuous unidirectional offset conditions. The remaining core model parameters remain frozen. This ensures the model's ability to adapt to slow degradation factors such as environmental aging and material batch differences during long-term operation, while avoiding the risk of model drift and runaway caused by frequent parameter updates. The entire technical path completely avoids the complex and noise-sensitive methods in existing technologies that rely on cooling capacity modeling, resonant frequency tracking, voltage overshoot compensation, or multi-source data clustering analysis. This achieves the construction of a lightweight, highly reliable, and low-maintenance intelligent temperature control system. In summary, this solution realizes a fundamental paradigm shift from "passively responding to temperature rise" to "actively managing thermal dynamics." It not only significantly improves welding quality consistency and equipment availability but also provides a complete new thermal management framework with interpretability, scalability, and engineering feasibility for the manufacturing of high-precision electromagnetic components. Attached Figure Description

[0016] Figure 1 This is the main flowchart of the ultrasonic welding power adaptive control method based on inductive temperature rise feedback.

[0017] Figure 2 This is a sub-flowchart of an ultrasonic welding power adaptive control method based on inductive temperature rise feedback.

[0018] Figure 3 This is another sub-flowchart of the ultrasonic welding power adaptive control method based on inductive temperature rise feedback. Detailed Implementation

[0019] Embodiments of the present invention are described in detail below, examples of which are shown in the accompanying drawings, wherein the same or similar reference numerals denote the same or similar elements or elements having the same or similar functions throughout. The embodiments described below with reference to the accompanying drawings are exemplary and are only used to explain the present invention, and should not be construed as limiting the present invention.

[0020] The following disclosure provides many different embodiments or examples for implementing different structures of the invention. To simplify the disclosure, specific examples of components and arrangements are described below. Of course, these are merely examples and are not intended to limit the invention. Furthermore, reference numerals and / or letters may be repeated in different examples; such repetition is for simplification and clarity and does not in itself indicate a relationship between the various embodiments and / or arrangements discussed.

[0021] like Figure 1 As shown, this application provides an adaptive control method for ultrasonic welding power based on inductive temperature rise feedback, specifically including: S1: During the no-load start-up phase of the ultrasonic welding equipment, a stepped power excitation sequence is executed and the time response curves of multiple temperature sensors arranged on the surface of the wire winding, the inner wall of the skeleton, and the interface of the impregnation varnish curing layer are collected simultaneously to obtain the original temperature rise time series dataset containing the thermal conductivity of the surface fiber varnish film, the thermal diffusion of the intermediate wire bundle winding structure, and the overall thermal mass of the winding.

[0022] S2: Based on the original temperature rise time series dataset, construct an inductor third-order thermal inertia model, and use nonlinear least squares method to fit and calculate the equivalent heat capacity and thermal resistance of the surface fiber coating, the thermal diffusion time constant of the intermediate filament winding structure, and the overall winding thermal mass-dominant time constant, generating a parameter set of the inductor third-order thermal inertia model with physical interpretability.

[0023] S3: During the welding operation phase, the inductor surface temperature is collected in real time as the main input signal, and the main input signal is input into the inductor thermal inertia identification module constructed based on the parameter set of the third-order thermal inertia model of the inductor, and the predicted value sequence of temperature rise trend in three key time domain windows of 50 milliseconds, 100 milliseconds and 200 milliseconds is generated.

[0024] S4: Perform a difference calculation between the current measured temperature and the predicted temperature rise trend sequence to extract the transient disturbance residual signal that cannot be covered by the third-order thermal inertia model of inductance, and use the transient disturbance residual signal as the input variable of the PI feedback controller to generate a closed-loop correction component.

[0025] S5: Based on the slope change rate of the predicted temperature rise value sequence, identify the state characteristics of the temperature rise transitioning from a gradual transition to an inflection point or entering a plateau region, trigger the feedforward channel to generate a power gradient pre-adjustment action command, and delay the power gradient pre-adjustment action command to the next zero voltage crossover interval when the vibration phase deviation is detected to exceed the limit.

[0026] S6: Determine whether the absolute value of the transient disturbance residual signal exceeds a set threshold. If it does, perform integral accumulation processing on the closed-loop correction component; otherwise, perform proportional amplification processing to generate the final feedback correction command.

[0027] S7: The power gradient pre-adjustment action command and the feedback correction command are vector-synthesized to generate a composite power adjustment command that includes both feedforward prediction and feedback correction characteristics.

[0028] S8: After each welding, compare the cumulative deviation pattern between the actual temperature rise trajectory and the predicted temperature rise trend sequence. If the deviations show the same direction for three consecutive times, start the parameter update process and only incrementally correct the surface thermal resistance parameter in the equivalent heat capacity and thermal resistance of the surface fiber coating to complete the online fine-tuning of the model.

[0029] Step S1: During the no-load start-up phase of the ultrasonic welding equipment, a stepped power excitation sequence is executed, and the time response curves of multiple temperature sensors arranged on the surface of the wire winding, the inner wall of the skeleton, and the interface of the impregnated varnish curing layer are simultaneously acquired. This yields the original temperature rise time-series dataset, which includes the thermal conductivity of the surface fiber varnish film, the thermal diffusion of the intermediate wire bundle winding structure, and the dominant characteristics of the overall winding thermal mass. Specifically, this includes: S1.1: Initialize the main control unit of the ultrasonic welding equipment, generate a stepped power excitation sequence control command containing five power gradient steps and the duration of each step covering the thermal diffusion time constant, so as to activate the power drive unit to enter the no-load excitation mode and output the corresponding stepped current signal.

[0030] The pre-stored inductor thermal characteristic calibration configuration file in the main control unit of the ultrasonic welding equipment is read. The gain coefficients of the PWM generator register and current sampling module of the power drive unit are initialized. The basic parameters of the stepped power excitation sequence are set, including the initial power threshold, the reference value of the single-step duration, and the total excitation cycle length. Based on the physical dimensions of the multi-strand wire inductor winding and the thermal diffusivity of the impregnated varnish curing layer, the theoretical estimated values ​​of the thermal conductivity time constant of the surface fiber varnish film and the thermal diffusivity time constant of the intermediate wire bundle winding structure are calculated. The larger of these is selected as the minimum coverage reference for the single-step duration, ensuring that the duration of each power step is sufficient to excite a complete thermal conduction transient response. A control command sequence containing five linearly increasing power gradient steps is constructed. The power values ​​of each step are set to 20%, 40%, 60%, 80%, and 100% of the rated power, respectively. A zero-power interval is set between adjacent steps to eliminate the thermal accumulation interference of the previous step, forming a stepped waveform data table with clear thermal excitation boundaries. The generated stepped power excitation sequence control commands are written into the non-volatile memory mapping area of ​​the main control unit and an activation signal is sent to the power drive unit via the internal bus to force a switch to the no-load excitation mode. At this time, the ultrasonic transducer is in an off state or a light load state, and only the inductor winding bears the main Joule heat load. After receiving the commands, the power drive unit outputs the corresponding stepped current signals step by step according to the preset timing sequence. The current amplitude increases linearly with the power step. At the same time, a precise system clock stamp is recorded at each step switching moment as an absolute reference for the subsequent temperature response curve time axis alignment. Through the above configuration and execution process, the abstract thermal inertia identification requirements are transformed into specific, time-controllable stepped current excitation sources, realizing the staged excitation of heat conduction paths at different depths of the inductor winding. This lays a solid signal foundation for the subsequent acquisition of temperature rise response data with high signal-to-noise ratio and clear physical correspondence by multi-point temperature sensors.

[0031] For example, for a 0.1*60mm multi-strand wire inductor winding with an equivalent cross-sectional area of ​​0.6mm in diameter, the thermal conductivity time constant of its surface fiber enamel film is preset to approximately 2 seconds, and the thermal diffusion time constant of the intermediate wire bundle winding structure is preset to approximately 15 seconds. The main control unit initializes the PWM frequency to 20kHz and the dead time to 2 microseconds. In the constructed five-stage power excitation sequence, the first stage power is set to 20% of the rated power of 3000W, i.e., 600W; the second stage is 1200W; the third stage is 1800W; the fourth stage is 2400W; and the fifth stage is 3000W. To ensure sufficient excitation of the thermal diffusion characteristics, the duration of a single stage is set to 20 seconds, covering 1.33 times the 15-second thermal diffusion time constant of the intermediate wire bundle winding structure, and a 5-second zero-power cooling interval is set between adjacent stages. In no-load mode, the power drive unit sequentially outputs stepped current signals with corresponding effective current values ​​of 10A, 14.14A, 17.32A, 20A, and 22.36A. At the 0-millisecond mark at the start of each step, the system records timestamps T1 to T5. This excitation sequence causes the inductor winding surface temperature to gradually rise from the ambient temperature of 25°C to 45°C within 20 seconds, while the inner wall temperature of the frame rises lagging to 38°C, and the impregnating varnish interface temperature rises slowly to 32°C. This generates multi-channel temperature rise data with significant time delay differences, effectively separating the rapidly changing and slowly changing heat components, and verifying the effectiveness of the stepped excitation strategy in stimulating multilayer heat conduction characteristics.

[0032] S1.2: Based on the Joule heating effect generated by the stepped current signal flowing through the multi-strand wire inductor winding, a multi-channel analog voltage signal reflecting the response characteristics of different heat conduction paths is simultaneously acquired using a first type of temperature sensor arranged on the surface of the wire winding, a second type of temperature sensor arranged on the inner wall of the skeleton, and a third type of temperature sensor arranged on the interface of the impregnation varnish curing layer.

[0033] Driven by a stepped power excitation sequence, a Joule thermal field that varies with time is generated inside the multi-strand wire-insulated inductor winding. This thermal field is conducted to the surrounding environment through different physical paths. A first type of temperature sensor, positioned on the surface of the wire-insulated winding, a second type of temperature sensor, positioned on the inner wall of the frame, and a third type of temperature sensor, positioned at the interface of the varnish curing layer, capture the thermal response signals at their respective locations. The first type of temperature sensor is in close contact with the outer fiber varnish film of the wire-insulated winding, directly sensing the rapid temperature rise of the surface layer dominated by the skin effect. Its output analog voltage signal mainly reflects the equivalent heat capacity and thermal resistance characteristics of the surface fiber varnish film. The second type of temperature sensor is embedded in the inner wall of the inductor frame at the contact surface with the winding, capturing the intermediate temperature during the heat transfer process from the winding to the frame. This signal characterizes the thermal diffusion time constant and thermal resistance distribution of the intermediate wire bundle winding structure. The third type of temperature sensor is located at the interface between the varnish curing layer and the external environment or in the deep curing zone, monitoring the slow temperature trend after being buffered by the overall winding thermal mass. Its signal is mainly affected by the time constant dominated by the overall winding thermal mass. Three types of sensors synchronously initiate data acquisition, ensuring that the multi-dimensional temperature-time evolution process from the surface to the core and then to the interface can be completely recorded at each power step switching instant. Because the high-frequency switching action of the ultrasonic welding power supply induces high-frequency electromagnetic noise in the sensor leads, the analog voltage signals of each channel need to be processed by a pre-amplified hardware low-pass filter circuit to suppress common-mode interference and retain the true low-frequency components of thermal dynamics. The weak millivolt-level voltage signals output by the sensors are impedance matched and gain adjusted by a high-precision instrumentation amplifier, linearly mapping them to the standard input range of the main control unit's ADC module, ensuring that the signal dynamic range covers the entire process from zero-power cold state to full-power thermal steady state. Within each constant power range of the stepped current signal, the three types of sensors continuously output analog voltage values ​​that are linearly related to the local temperature, forming three parallel time-continuous analog signal streams. Through hardware triggering mechanisms or software synchronous interrupts, the sampling time of the three analog signals is strictly locked on the same microsecond-level time base, eliminating timing misalignment caused by multi-channel serial scanning, and ensuring that the data collected at the same time can accurately reflect the instantaneous state of the three-dimensional thermal field inside the inductor. Through the above-mentioned multi-point synchronous acquisition and signal conditioning processing method, the stepped current excitation generated in the previous step is transformed into a multi-channel analog voltage signal that reflects the rapid change on the surface, the diffusion in the middle and the slow change in the whole. This enables all-round real-time perception of the complex heat conduction path of the inductor, and provides a raw data foundation with physical layering characteristics for the subsequent construction of a high-fidelity third-order thermal inertia model.

[0034] For example, during the no-load startup phase, the main control unit generates a five-segment stepped power excitation sequence, with each segment representing 20%, 40%, 60%, 80%, and 100% of the rated power, respectively. The duration of each segment is set to 5 seconds, which is sufficient to cover the longest thermal diffusion time constant of the wire-insulated inductor winding. The first type of temperature sensor uses a miniature NTC thermistor with a response time of less than 10ms, attached to the outermost surface of the 0.1mm diameter * 60-strand wire-insulated winding; the second type of temperature sensor uses a thin-film platinum resistance thermometer with a response time of approximately 50ms, embedded in the inner wall groove of the PBT material frame; the third type of temperature sensor uses a K-type thermocouple with a response time of approximately 200ms, embedded 2mm deep in the impregnated varnish cured layer. When the step current reaches 80% of its rated value (e.g., RMS 15A), the insulated wire winding generates significant Joule heating due to the skin effect. The first type of sensor detects a rapid temperature rise from 25°C to 45°C within 2 seconds, and the output analog voltage linearly increases from 1.25V to 2.25V (assuming a sensitivity of 10mV / °C and a magnification of 100x). Simultaneously, the second type of sensor, due to thermal lag, only reaches 35°C after 3.5 seconds, and the output voltage slowly increases from 1.25V to 1.75V. The third type of sensor, buffered by the overall thermal mass, only reaches 30°C at the end of the 5-second step, and the output voltage remains around 1.50V. After filtering out 20kHz switching noise with a pre-amplified RC low-pass filter (cutoff frequency 10Hz), the three signals are simultaneously sampled by a 16-bit high-precision ADC at a sampling rate set to 1kHz. At the 3-second mark of the 80% power step, the system simultaneously reads three voltage values: V1 = 2.00V (corresponding to 45℃), V2 = 1.60V (corresponding to 31℃), and V3 = 1.40V (corresponding to 29℃). These simulated voltage data, with clear physical hierarchical correspondences, fully record the dynamic gradient of heat diffusion from the surface fiber enamel film of the silk-covered wire to the skeleton and impregnated enamel layer. This ensures that subsequent model identification can accurately separate the rapidly changing and slowly changing heat components, significantly improving the signal-to-noise ratio and physical interpretability of the thermal inertia model parameter identification.

[0035] S1.3: Perform high-precision analog-to-digital conversion and timestamp alignment on the multi-channel analog voltage signal to generate a multi-channel digital temperature time series with a unified sampling reference, so as to eliminate the phase lag error caused by the difference in sensor position and form a standardized digital temperature time series.

[0036] The system receives multi-channel analog voltage signals, reflecting the response characteristics of different heat conduction paths, synchronously acquired by Class I, Class II, and Class III temperature sensors, as the raw input data for analog-to-digital conversion and time alignment processing. The analog-to-digital converter (ADC) module inside the main control unit performs parallel high-speed sampling of the three analog voltage signals according to a preset sampling clock frequency, converting continuously changing analog quantities into discrete digital quantized values, ensuring the sampling rate is not lower than the Nyquist frequency to fully preserve the dynamic characteristics of the temperature rise signal. Addressing the signal transmission delay caused by spatial distribution differences at three physical locations—the surface of the multi-strand inductor winding, the inner wall of the frame, and the interface of the impregnated varnish curing layer—the system reads the fixed transmission delay parameters of each sensor channel in the hardware circuit and establishes a time delay compensation mapping table based on physical location. Linear interpolation is used to resample the digitized temperature data, and microsecond-level time axis shift correction is performed on the data sequence of each channel according to the time delay compensation mapping table to eliminate phase lag errors caused by inconsistent sensor wiring lengths and differences in signal conditioning circuit response. A high-precision system global clock is introduced as a unified time reference, and each frame of temperature data after delay correction is timestamped with a precision down to the microsecond level, constructing a multi-channel data frame structure with strict time sequence correspondence. Through the aforementioned high-precision analog-to-digital conversion and timestamp alignment processing, the phase-deviation analog signal present in the previous step is transformed into a multi-channel digital temperature time series with a unified sampling reference and time synchronization, achieving the expected technical effect of eliminating phase lag errors caused by sensor position differences and forming a standardized digital temperature time series.

[0037] S1.4: Perform sliding window denoising filtering based on the digital temperature time series to remove high-frequency noise components introduced by electromagnetic interference and extract a set of pure temperature rise response curves that characterize the rapid thermal conductivity of the surface fiber coating, the medium thermal diffusion characteristics of the intermediate filament winding structure, and the slow thermal quality characteristics of the overall winding.

[0038] S1.5: The pure temperature rise response curve set is segmented and structured according to the excitation step interval to generate an original temperature rise time series dataset containing complete thermal dynamic evolution information, which can be called in subsequent steps to identify and calculate the parameters of the third-order thermal inertia model of inductance.

[0039] Step S2: Based on the original temperature rise time series dataset, a third-order thermal inertia model of the inductor is constructed. The equivalent heat capacity and thermal resistance of the surface fiber coating, the thermal diffusion time constant of the intermediate filament winding structure, and the overall winding thermal mass-dominated time constant are calculated using the nonlinear least squares method to generate a parameter set for the physically interpretable third-order thermal inertia model. Specifically, this includes: S2.1: Perform multi-timescale feature separation processing on the original temperature rise time series dataset to extract the fast-changing component of thermal conductivity of the surface fiber coating, the mid-frequency component of thermal diffusion of the intermediate filament winding structure, and the slow-changing component dominated by the thermal mass of the overall winding, generating a decoupled temperature response sequence containing the thermal conduction response features of the three layers.

[0040] S2.2: Based on the decoupled temperature response sequence, a third-order series thermal resistance and thermal capacity network topology is constructed to define the connection relationship of the surface fiber coating equivalent thermal capacity node, the intermediate filament winding structure thermal diffusion time constant node, and the overall winding thermal mass dominant time constant node, thereby generating an inductor thermal inertia initial architecture model with physical mapping relationship.

[0041] Based on the decoupled temperature response sequence output from step S2.1, which includes the rapidly changing thermal conductivity component of the surface fiber enamel film, the mid-frequency component of thermal diffusion from the intermediate wire bundle winding structure, and the slowly changing component dominated by the overall winding thermal mass, a third-order series thermal resistance-capacitance network topology is constructed to accurately map the physical heat conduction path of the multi-strand wire inductor. The rapidly changing surface component in the decoupled temperature response sequence is mapped as the first-level node of the network. This node is defined as being composed of the equivalent thermal capacity C1 and equivalent thermal resistance R1 of the surface fiber enamel film connected in parallel, used to characterize the dynamic balance process of rapid heat absorption and heat dissipation from the insulating enamel layer on the surface of the wire under Joule heating. The mid-frequency component of thermal diffusion from the intermediate wire bundle winding structure is mapped as the second-level node of the network. This node is defined as being composed of the equivalent thermal capacity C2 of the intermediate layer and the interlayer contact thermal resistance R2 connected in series, used to characterize the thermal diffusion delay effect when heat penetrates from the surface of the wire bundle to the deep interior of the winding and the interfacial thermal resistance characteristics between the multi-strand wires. The dominant slowly varying component of the overall winding thermal mass is mapped to a third-order node in the network. This node is defined as consisting of the equivalent thermal capacity C3 of the winding frame and core, and the thermal resistance R3 to ground. It is used to characterize the overall thermal accumulation of the inductor and its long-term thermal inertia characteristics of slowly releasing heat to the surrounding air or cooling medium. A set of state-space equations describing the dynamic behavior of this third-order series thermal resistance-capacity network is established. Joule heat generated by the input power P(t) is used as the system excitation source, and the temperatures T1(t), T2(t), and T3(t) of each node are used as state variables. The thermal balance differential equations for each node are listed based on the law of conservation of energy. For the first-order node, its temperature change rate depends on the input thermal power minus the heat flow to the second-order node through the thermal resistance R1 and the heat dissipation through surface heat dissipation. For the second-order node, its temperature change rate depends on the heat flow into the first-order node minus the heat flow to the third-order node. For the third-order node, its temperature change rate depends on the heat flow into the second-order node minus the heat loss to the environment. The three differential equations are integrated into a standard state-space form. The system matrix contains the reciprocal relationships between thermal resistance and thermal capacity, and the input matrix correlates the power input with the thermal capacity of the first-level nodes. Through this topology construction, abstract temperature response data is transformed into a circuit simulation model with clear physical meaning, establishing the coupling constraints between thermal parameters at each level. By constructing a third-order series thermal resistance-thermal capacity network topology and establishing the corresponding state-space differential equations, the decoupled temperature response sequence from the previous step is transformed into an initial architecture model of inductor thermal inertia with physical mapping relationships. This achieves a mathematical reconstruction of the inductor's multilayer heat conduction mechanism, providing a definite model framework for subsequent high-precision parameter identification using the nonlinear least squares method.

[0042] S2.3: The nonlinear least squares method is used to perform parameter iterative identification processing on the initial architecture model of the inductor thermal inertia in order to minimize the sum of squared residuals between the decoupled temperature response sequence and the model output trajectory, and generate the equivalent heat capacity value of the surface fiber varnish film, the equivalent thermal resistance value of the surface fiber varnish film, the thermal diffusion time constant value of the intermediate filament winding structure, and the overall winding thermal mass dominant time constant value in the convergent state.

[0043] Based on the initial architecture model of the inductor's thermal inertia constructed in the previous steps and the decoupled temperature response sequence, a nonlinear least squares method is used to optimize the iterative identification of execution parameters to obtain the key physical parameters of the third-order thermal inertia model in the converged state. The rapidly changing temperature components of the decoupled surface fiber coating, the mid-frequency temperature components of the intermediate filament winding structure, and the slowly changing temperature components of the overall winding are mapped to the observed output vectors of each node in the third-order series RC network. The parameter vector to be identified is defined as a four-dimensional spatial variable containing the surface equivalent heat capacity C1, the surface equivalent thermal resistance R1, the intermediate layer thermal diffusion time constant τ2, and the overall winding thermal mass-dominated time constant τ3. An objective function is constructed to minimize the sum of squared residuals between the measured decoupled temperature sequence and the model predicted temperature sequence. This objective function characterizes the degree of fit between the model output trajectory and the actual thermal dynamic process, and its mathematical expression is as follows: Where J(θ) is the cost function value, θ is the parameter vector to be identified [C1, R1, τ2, τ3], N is the total number of sampling points, and T m,i Let T be the measured decoupling temperature value at the i-th sampling time. p,i(θ) represents the model-predicted temperature value calculated based on the current parameter vector θ. Initial guesses for the parameter vector θ are initialized. Reasonable search boundaries are set based on the physical dimensions and material properties of the wire. The search range for the surface equivalent heat capacity C1 is set to 0.5 J / K to 5.0 J / K, the surface equivalent thermal resistance R1 is set to 10 K / W to 100 K / W, the intermediate layer thermal diffusion time constant τ2 is set to 0.1 s to 2.0 s, and the overall winding thermal mass-dominated time constant τ3 is set to 5.0 s to 60.0 s, ensuring the iteration process is within the physically feasible region. The Levenberg-Marquardt algorithm is used to iteratively solve the objective function. In each iteration, the Jacobian matrix is ​​calculated to evaluate the sensitivity of parameter changes to the model output. The elements of the Jacobian matrix are defined as the partial derivatives of the predicted temperature with respect to the parameters to be identified. The calculation process involves linearizing the third-order heat conduction differential equations. The temperature response changes caused by parameter perturbations are approximated using the numerical difference method. The parameter update step size is calculated based on the Jacobian matrix and residual vector. The convergence characteristics of the gradient descent method and the Gauss-Newton method are introduced to balance the damping factor. When the residual decreases, the damping factor is decreased to accelerate convergence; when the residual increases, the damping factor is increased to ensure stability, gradually correcting the value of the parameter vector θ. The rate of change of the objective function J(θ) and the update amplitude of the parameter vector are monitored in real time. When the relative change of the objective function value is less than a preset threshold 1e-6 and the change of the Euclidean norm of the parameter vector is less than 1e-4 in three consecutive iterations, the algorithm is considered to have reached convergence, and the iteration process is stopped. The values ​​of each component in the converged parameter vector are extracted, corresponding to the equivalent heat capacity of the surface fiber coating, the equivalent thermal resistance of the surface fiber coating, the thermal diffusion time constant of the intermediate filament winding structure, and the overall winding thermal mass-dominated time constant, completing the parameter filling for the initial architecture model of the inductor thermal inertia. By using a nonlinear least squares iterative identification process, the decoupled temperature response sequence generated in the previous step is transformed into a parameter set of the third-order thermal inertia model of the inductor with accurate physical dimensions. This achieves a quantitative mapping from experimental data to physical model parameters, significantly improving the model's accuracy in representing the dynamic characteristics of inductor temperature rise and providing a reliable model foundation for subsequent feedforward compensation control.

[0044] S2.4: Update the state matrix of the initial architecture model of the inductor thermal inertia based on the equivalent heat capacity of the surface fiber enamel film, the equivalent thermal resistance of the surface fiber enamel film, the thermal diffusion time constant of the intermediate filament winding structure, and the dominant time constant of the overall winding thermal mass under the convergence state, so as to complete the physical dimension calibration of the model parameters and generate a parameter set of the third-order thermal inertia model of the inductor with physical interpretability.

[0045] The equivalent heat capacity, equivalent thermal resistance, thermal diffusion time constant of the intermediate wire bundle winding structure, and overall winding thermal mass-dominant time constant of the surface fiber enamel film, all output from the preceding step S2.3 in the convergent state, are used as input variables for updating the parameters of the third-order thermal inertia model of the inductor. A state matrix for the initial architecture model of the inductor's thermal inertia, based on a state-space expression, is constructed. This state matrix consists of a system matrix A, an input matrix B, an output matrix C, and a through matrix D, used to describe the dynamic energy transfer relationship between different heat conduction levels of the multi-strand wire inductor winding. The equivalent heat capacity and equivalent thermal resistance of the surface fiber enamel film are mapped to the elements in the first row and first column of state matrix A, respectively, to characterize the self-feedback characteristics of the rapid thermal response of the surface layer and the thermal coupling strength between adjacent layers. The thermal diffusion time constant of the intermediate wire bundle winding structure is converted into an attenuation coefficient on the diagonal of state matrix A, reflecting the influence of the diffusion rate of the mid-frequency thermal component in the wire bundle gap on the system state evolution. The dominant time constant value of the overall winding thermal mass is integrated into the third diagonal element of the state matrix A, establishing a benchmark for the energy accumulation and dissipation of slowly varying heat components in the long time domain. Based on the principle of dimensional consistency, the input matrix B is scaled to convert the Joule heat power density generated by the stepped power excitation into an equivalent temperature change rate input vector, ensuring unit matching between the input signal and the state variables. The output matrix C is configured with row vectors to select only the state variable components representing the surface temperature of the insulated wire winding, simulating the constraints of actual sensor placement on system observability. Eigenvalue decomposition of the state matrix is ​​performed to verify that all real parts of matrix A are negative, ensuring the asymptotic stability of the inductor thermal inertia model without external excitation. Through the above state matrix update and dimensional calibration processes, discrete thermophysical parameters are transformed into a continuous-time state-space model with linear time-invariant properties, generating a physically interpretable parameter set for the third-order thermal inertia model of the inductor. This represents a technological leap from static parameter identification to dynamic system modeling, providing a precise mathematical foundation for subsequent feedforward extrapolation of real-time temperature rise trends.

[0046] S2.5: Based on the parameter set of the physically interpretable third-order inductive thermal inertia model, perform model validity verification processing to compare the goodness-of-fit index between the model's predicted temperature rise curve and the original temperature rise time series dataset, and generate the final parameter set of the third-order inductive thermal inertia model that is confirmed for subsequent temperature rise trend inference.

[0047] like Figure 2 As shown, step S3 involves: real-time acquisition of the inductor surface temperature as the main input signal during the welding operation phase, and inputting the main input signal into the inductor thermal inertia identification module constructed based on the parameter set of the third-order thermal inertia model of the inductor, to deduce and generate a sequence of predicted temperature rise trends within three key time domain windows: 50 milliseconds, 100 milliseconds, and 200 milliseconds. Specifically, this includes: S3.1: Perform high-frequency sampling and digital filtering on the analog voltage signal output by the temperature sensor arranged on the surface of the inductor winding to obtain the real-time raw data of the inductor surface temperature after removing noise interference, and convert the real-time raw data of the inductor surface temperature into a standardized digital temperature measurement value.

[0048] S3.2: Based on the digital temperature measurement value and the parameter set of the third-order thermal inertia model of the inductor generated in the previous step, the inductor thermal inertia identification module embedded in the main control unit is called to perform the initialization mapping operation of the state space equation, so as to construct the current system state vector including the equivalent heat capacity of the surface fiber coating, the thermal diffusion time constant of the intermediate filament winding structure, and the dominant time constant of the overall winding thermal mass.

[0049] Construction process of the inductive thermal inertia identification module: Obtaining the model parameter set: Through nonlinear least squares fitting in step S2, four key parameters of the third-order thermal inertia model of the inductor have been obtained: the equivalent heat capacity and equivalent thermal resistance of the surface fiber coating, the thermal diffusion time constant of the intermediate wire bundle winding structure, and the overall winding thermal mass-dominated time constant. These parameters quantitatively describe the thermal conduction characteristics from the surface to the core inside the inductor.

[0050] A state-space mathematical model is established: Based on the topology of a third-order series RC network, a continuous-time state-space model is constructed. This model uses three state variables to represent the surface temperature, the intermediate wire bundle temperature, and the overall winding core temperature. The system matrix is ​​composed of the reciprocal relationship between heat capacity and thermal resistance, with off-diagonal elements reflecting the interlayer thermal coupling strength. The input matrix represents the excitation effect of welding power on the surface heat capacity; the output matrix selects observable states (usually only the surface temperature). Substituting the above four parameters into the system matrix, input matrix, and output matrix completes the model concretization.

[0051] Discretization: The main control unit operates with a fixed sampling period (e.g., 1 millisecond), requiring the continuous model to be converted into a discrete form. Using the zero-order hold method or the first-order Euler approximation, the discrete state matrix, discrete input matrix, and discrete output matrix are calculated, allowing the state vector to be updated through simple algebraic iterations in each control cycle.

[0052] Initialize the state vector: Before welding begins, three initial state values ​​must be given. The surface temperature is read in real time by the sensor and directly assigned; the intermediate layer and core temperatures cannot be directly measured and are estimated based on the steady-state heat conduction ratio or by assuming the initial temperature equals the ambient temperature (e.g., 25°C). Alternatively, the predicted final value at the end of the previous welding cycle can be used as the initial value. After initialization, the module has complete state information for the current moment.

[0053] Real-time Iterative Identification: In each control cycle, the module performs the following steps: Acquire the current welding power (from feedback from the power drive unit) and the measured surface temperature from the sensor. Substitute the estimated state vector from the previous cycle and the current power into the discrete state equation to calculate the predicted state vector for the next cycle. Compare the predicted surface temperature with the measured surface temperature to obtain the residual, which can be used for subsequent feedback correction. Use the predicted state vector (or the state after residual correction) as the starting point for the next cycle. Since the physical meaning of the model parameters is clear and the welding process has significant thermal inertia, simple open-loop prediction is sufficient. Output the predicted three-dimensional state vector for subsequent multi-step feedforward extrapolation.

[0054] Software Integration and Execution: The discrete state-space iterative algorithm described above is written as a function and embedded into the real-time task of the main control unit. This function is called at a fixed sampling period, with the measured surface temperature and welding power as inputs, and the estimated temperatures of three layers (surface, intermediate layer, and core) as outputs. The entire algorithm is extremely small, requiring only a few multiply-accumulate operations, meeting the lightweight requirement. It utilizes a pre-identified physical model and current observations to "identify" the intermediate layer and core temperatures, which cannot be directly measured, in real time, and predict future temperature rise trends.

[0055] Parameter fine-tuning support: The module has a built-in interface for modifying parameters. When S8 triggers the parameter update process, only the surface equivalent thermal resistance parameter is adjusted, while other parameters remain unchanged. The updated parameters are then re-introduced into the system matrix to achieve online model optimization.

[0056] Through the above steps, an inductive thermal inertia identification module is constructed to convert the measured temperature into a state vector, and then predict the future temperature through iterative state equations, providing key input for feedforward control.

[0057] The system receives the standardized digital temperature measurement value obtained after processing in step S3.1. This value represents the real-time thermal state of the inductor winding surface at the current moment and serves as the initial observation input for the inductor thermal inertia identification module. It calls the inductor third-order thermal inertia model parameter set pre-stored in the main control unit's non-volatile memory. This parameter set includes the equivalent heat capacity of the surface fiber enamel film, the equivalent thermal resistance of the surface fiber enamel film, the thermal diffusion time constant of the intermediate wire bundle winding structure, and the overall winding thermal mass-dominated time constant identified in step S2. Based on the thermodynamic law of conservation of energy and Fourier's law of heat conduction, a continuous-time state-space equation describing the coupling relationship of the three-layer heat conduction paths within the inductor is constructed, transforming the physical thermal resistance-thermal capacity network topology into a mathematical linear time-invariant system model. The system state vector is defined as a three-dimensional column vector, with its elements corresponding to the surface enamel film node temperature, the intermediate wire bundle node temperature, and the overall winding core node temperature, respectively. The surface node temperature is directly related to the sensor's measured value. Based on the characteristics of the series thermal resistance network, the structure of the state matrix A is derived, where the diagonal elements are determined by the heat capacity of each node and the thermal resistance of adjacent nodes, and the off-diagonal elements reflect the thermal coupling strength between nodes. An input matrix B is defined to characterize the excitation effect of ultrasonic welding power on the system state through the Joule heating effect. Since the heat source is mainly generated in the winding conductor and diffuses through multiple layers of medium, only the first component of the input vector is non-zero (corresponding to the reciprocal of the surface heat capacity), and the remaining components are zero. This simplifies the assumption that power injection mainly affects the surface thermal balance, and the deep thermal response is transmitted through the coupling matrix A. An initialization mapping operation of the state space equations is performed, assigning the current digital temperature measurement value to the first component of the state vector. For the intermediate layer temperature and core temperature, which cannot be directly measured, initial estimates are made based on the steady-state heat conduction ratio or the predicted final value of the previous control cycle to construct the complete system state vector for the current moment. Through the above state-space modeling and initialization mapping process, discrete sensor temperature readings are transformed into system state vectors containing multidimensional thermodynamic information, realizing the transformation from single surface temperature observation to internal global thermal state reconstruction, and providing physically consistent initial boundary conditions for subsequent multi-step feedforward inference.

[0058] S3.3: Using the current system state vector as the iterative reference, perform multi-step forward extrapolation calculations on the third-order heat conduction differential equations to solve the discretized temperature rise state components corresponding to the three specific time nodes of the next 50 milliseconds, 100 milliseconds and 200 milliseconds respectively.

[0059] Based on the system state vector constructed in the preceding steps, this vector encapsulates key physical parameters such as the equivalent heat capacity of the surface fiber coating, the thermal diffusion time constant of the intermediate filament winding structure, and the dominant time constant of the overall winding thermal mass, serving as the initial iterative benchmark for the Runge-Kutta numerical integration method. The third-order thermal inertia model of the inductor is transformed into a standard first-order linear differential equation system. The state variables are defined as the temperature values ​​of each node, and the input variables are the real-time acquired inductor surface temperature and its rate of change. A state-space expression describing the dynamic process of heat transfer between the surface coating, intermediate filaments, and the overall winding is established. The fourth-order Runge-Kutta method (RK4) is used to discretize and solve the differential equation system, with a fixed step size of 1 millisecond to ensure a balance between computational accuracy and real-time performance during the rapid dynamic process of ultrasonic welding. Within each calculation step, slope estimates k1, k2, k3, and k4 are calculated at the current time t, the intermediate time t+0.5h, and the next time t+h, respectively. k1 reflects the instantaneous temperature rise rate under the current state, k2 and k3 reflect the average change trend within a half-step, and k4 reflects the predicted change rate at the end of the full-step. The system state temperature value at the next time step is updated using a weighted average formula. For a 50-millisecond time window, 50 iterations are performed, accumulating and updating the state vector to obtain the discretized temperature rise state component at the end of the 50th step. This component characterizes the rapid thermal conductivity response characteristics in the short term. For a 100-millisecond time window, another 50 iterations are performed until the cumulative step count reaches 100, obtaining the discretized temperature rise state component at the end of the 100th step. This component characterizes the medium-term thermal diffusion equilibrium trend. For a future 200-millisecond time window, 100 iterative calculations are performed until the cumulative step count reaches 200. The discretized temperature rise state component at the end of the 200th step is obtained; this component characterizes the long-term thermal-mass-dominated slow-varying response. During each iteration, the temperature gradient term in the state matrix is ​​updated in real time to ensure that subsequent step calculations are based on the latest thermodynamic state, thereby accurately capturing the nonlinear thermal conduction characteristics of multi-strand wire inductors under complex loads. Through the above multi-step forward extrapolation calculations, the current system state vector is transformed into high-precision discretized temperature rise state components corresponding to three specific time nodes: 50 milliseconds, 100 milliseconds, and 200 milliseconds. This enables the generation of basic data for proactive prediction and feedforward compensation of the inductor's thermal inertia delay effect.

[0060] S3.4: Perform time-series recombination and vector encapsulation processing on the discretized temperature rise state components to generate a future multi-time-domain window temperature rise trend prediction value sequence with clear timestamps. This sequence fully characterizes the expected thermal dynamic trajectory of the inductor under the coupling effect of fast-changing and slow-changing components.

[0061] S3.5: Perform slope differentiation operation based on the predicted value sequence of future multi-time-domain temperature rise trend to extract the characteristic parameter of temperature rise rate of change in each prediction time window, and output the characteristic parameter of temperature rise rate of change to the feedforward channel to trigger the subsequent power gradient pre-adjustment action command generation logic.

[0062] The predicted temperature rise trend sequence with clearly defined timestamps generated in step S3.4 is processed using first-order differencing to quantify the rate of temperature change between each predicted time point. A discretized temperature rise slope calculation model is constructed, and for three key time-domain windows—50 milliseconds, 100 milliseconds, and 200 milliseconds—the ratio of the temperature difference between adjacent predicted points to the corresponding time interval is extracted. Using either the central difference method or the forward difference method, an appropriate numerical differentiation strategy is selected based on the sampling frequency characteristics of the system's real-time control cycle to suppress the amplification effect of high-frequency noise. For the 50-millisecond window, the difference between the predicted temperature at the current moment and the predicted temperature 50 milliseconds later is calculated and divided by the 50-millisecond time span to obtain the short-term temperature rise rate of change index. For the 100-millisecond and 200-millisecond windows, the same differencing logic is performed to obtain the temperature rise rate of change indices in the mid-term and long-term time domains, forming a temperature rise rate of change feature vector containing three time scales. This eigenvector reflects the dynamic response intensity of the inductor's thermal inertia across different time dimensions. The short-time rate of change primarily characterizes the rapid thermal conductivity of the surface fiber coating, while the long-time rate of change reflects the slow cumulative effect of the overall winding thermal mass. Through the aforementioned slope differential operation, the discretized temperature rise state components from the previous step are transformed into temperature rise rate of change characteristic parameters characterizing the thermal dynamic evolution trend. This achieves a quantitative description of the inductor's future thermal state, providing accurate data support for the feedforward channel to identify temperature rise inflection points and generate power gradient pre-adjustment commands.

[0063] like Figure 3 As shown, step S4 involves performing a difference calculation between the current measured temperature and the predicted temperature rise trend sequence to extract the transient disturbance residual signal that cannot be covered by the third-order thermal inertia model of the inductor, and using the transient disturbance residual signal as the input variable of the PI feedback controller to generate a closed-loop correction component. Specifically, this includes: S4.1: Based on the parameter set of the third-order thermal inertia model of the inductor, the predicted value sequence of temperature rise trend is derived and the main input signal of the inductor surface temperature is collected in real time. Time alignment and synchronous sampling processing are performed to obtain the set of measured temperature data and the set of predicted temperature data with the same timestamp reference, so as to provide consistent time domain input conditions for subsequent difference calculation.

[0064] The system receives a sequence of predicted temperature rise trends with clear timestamps generated in step S3, along with raw inductor surface temperature data acquired in real-time by a temperature sensor and preprocessed in step S3.1, serving as dual input sources for time alignment and synchronous sampling. The predicted temperature rise trend sequence is reconstructed using timestamp indexing to extract the predicted temperature reference point corresponding to the current control cycle time t, and the absolute phase position of this time in the system's global clock is locked, establishing a reference time axis for the feedforward channel. Hardware interrupt-triggered capture is performed on the raw inductor surface temperature data acquired in real-time, recording the specific time when the sensor's analog-to-digital conversion completion signal arrives at the main control unit (MCU), and the transmission delay compensation amount relative to the system's global clock is calculated using an internal high-precision timer. Based on the transmission delay compensation amount, linear interpolation or zero-order hold processing is performed on the real-time temperature data, mapping the asynchronously arriving measured temperature data to a system sampling time point completely consistent with the feedforward predicted value, eliminating microsecond-level timing misalignment caused by differences in signal acquisition paths. A time-domain alignment check is performed on the dual-channel data to determine if the timestamp deviation between the measured temperature data and the predicted temperature data is less than a preset synchronization tolerance threshold. If the deviation exceeds the threshold, the current period's data is discarded and the system waits for the next sampling point. If the conditions are met, the data is encapsulated into a standard data pair with a unified time base. Through the above time alignment and synchronous sampling processing, the data streams of the feedforward prediction channel and the feedback measured channel are transformed into a set of measured temperature data and a set of predicted temperature data with the same timestamp reference. This achieves strict synchronization of multi-source heterogeneous thermal signals in the time domain, providing consistent and phase-distortion-free time-domain input conditions for subsequent high-precision residual difference calculations.

[0065] S4.2: Using the measured temperature data set and the predicted temperature data set, perform point-by-point algebraic difference operation to extract the transient disturbance residual signal sequence that characterizes the dynamic characteristics not covered by the third-order thermal inertia model of the inductor. This sequence quantifies the temperature response deviation caused by the sudden change in actual welding load.

[0066] The system receives the measured temperature data set and the predicted temperature data set after time alignment processing in step S4.1. The measured temperature data set includes sampling points of the surface temperature of the inductor winding at the current time and within the historical window. The predicted temperature data set includes predicted future multi-time-domain temperature rise values ​​derived from the third-order thermal inertia model of the inductor at the corresponding timestamp. A point-by-point algebraic difference operation logic is constructed, using the measured temperature value at the same timestamp as the minuend and the corresponding model predicted temperature value as the subtrahend, performing a scalar subtraction operation to obtain the single-point temperature deviation value. This operation strictly follows the principle of linear superposition, aiming to remove the deterministic temperature rise component caused by the inductor's own thermal inertia, thereby separating the non-deterministic residual component caused only by external load mutations or environmental disturbances. For each sampling period k, the transient disturbance residual signal is defined as the difference between the measured temperature and the predicted temperature, and the residual sequence is calculated. The above subtraction operation is performed sequentially on all data points within the entire synchronous sampling window to generate a transient disturbance residual signal sequence containing continuous time series characteristics. This sequence numerically characterizes the degree of thermal response deviation caused by dynamic factors such as load impedance changes and contact pressure fluctuations during actual welding. Its positive or negative sign indicates whether the actual temperature rise leads or lags behind the model prediction. The generated residual signal sequence undergoes integrity verification to ensure no null values ​​or anomalous jumps, forming a standardized transient disturbance residual signal sequence output. Through point-by-point algebraic interpolation, the time-aligned data from the previous step is transformed into a transient disturbance residual signal sequence characterizing dynamic characteristics not covered by the model. This sequence quantifies the temperature response deviation caused by sudden changes in actual welding load, providing a precise error input benchmark for the subsequent PI feedback controller and achieving real-time compensation for the blind spots in the thermal inertia model prediction.

[0067] S4.3: Based on the amplitude and rate of change characteristics of the transient disturbance residual signal sequence, perform residual signal validity judgment and noise filtering to generate pre-processed pure residual control variables and eliminate the influence of sensor high-frequency noise on the false triggering of the feedback loop.

[0068] The transient disturbance residual signal sequence generated in the preceding step S4.2 is received. This sequence contains mixed deviation components caused by sudden changes in welding load, high-frequency noise from sensors, and unmodeled dynamics. Sliding window statistical feature extraction is performed on the transient disturbance residual signal sequence to calculate the local mean and standard deviation of the residual signal at the current time, quantifying the DC offset and AC fluctuation components of the residual signal. Based on the extracted local mean and standard deviation, an adaptive threshold discrimination function is constructed to distinguish between high-frequency random noise caused by electromagnetic interference and effective transient disturbances caused by drastic load changes. An improved Kalman filter is used to estimate the state of the residual signal, treating the residual signal as a system state variable and the sensor measurement as the observation value. The residual estimate is corrected through a predict-update iterative cycle to suppress high-frequency noise components. During the Kalman filtering process, the process noise covariance matrix and the measurement noise covariance matrix are dynamically adjusted. When the residual change rate exceeds a preset sudden change threshold, the process noise weight is increased to quickly track the actual disturbance; when the residual is in a stable range, the measurement noise weight is increased to smooth random jitter. A first-order difference operation is performed on the filtered residual signal to obtain the residual rate of change vector, which is used to characterize the dynamic evolution trend of the disturbance signal. Combining the residual amplitude and the residual rate of change vector, a validity logic judgment is performed. If the residual amplitude is lower than the noise floor threshold and the rate of change is within the normal fluctuation range, it is determined to be invalid noise and set to zero; if the residual amplitude or rate of change exceeds the valid disturbance threshold, the filtered residual value is retained as a clean residual control variable. Through the above adaptive filtering and validity judgment processing, the noisy transient disturbance residual signal obtained in the previous step is transformed into a high signal-to-noise ratio clean residual control variable, achieving the expected technical effect of eliminating the influence of high-frequency sensor noise on the false triggering of the feedback loop while retaining the true load disturbance characteristics.

[0069] For example, the sliding window length is set to 10 sampling points, the sampling frequency is 10kHz, and the window time is 1ms. At a certain welding moment, the acquired transient disturbance residual signal sequence is [r(k-9), ..., r(k)], with a calculated local mean of 0.05℃ and a standard deviation of 0.02℃. The noise floor threshold is set to 3 times the standard deviation, i.e., 0.06℃, the effective disturbance amplitude threshold is 0.5℃, and the rate of change threshold is 10℃ / s. The initial process noise covariance Q is set to 0.01, and the measurement noise covariance R is set to 0.1. When the residual signal r(k) is detected to be 0.04℃, it is determined to be high-frequency noise because it is less than the noise floor threshold of 0.06℃, and is directly set to zero. When the welding load suddenly increases, the residual signal r(k) jumps to 0.8℃, and the first-order differential rate of change reaches 15℃ / s, exceeding the effective disturbance threshold. At this point, the Kalman filter detects an excessive rate of change and automatically adjusts the process noise covariance Q to 0.1, accelerating the tracking speed. After Kalman filtering iterations, the output pure residual control variable is 0.78℃, effectively filtering out the ±0.05℃ high-frequency glitches superimposed on the abrupt signal. This pure residual control variable is then input to the PI controller, avoiding integral saturation and power oscillation caused by noise, significantly improving the stability and response accuracy of power regulation.

[0070] S4.4: Based on the pure residual control variable, the constructed PI feedback controller is invoked to perform proportional gain amplification and integral accumulation calculation to generate an initial closed-loop correction control quantity containing fast response components and steady-state elimination components, thereby achieving preliminary compensation for transient disturbances.

[0071] The PI feedback controller is the core closed-loop control component in this invention for processing transient disturbance residual signals. Its input is the purified residual control variable (i.e., the dynamic deviation between the actual inductor surface temperature and the temperature predicted by the third-order thermal inertia model) after validity discrimination and filtering, and its output is the initial closed-loop correction control quantity. This controller employs a parallel structure with both proportional (P) and integral (I) control paths to achieve rapid response to transient disturbances and eliminate steady-state accumulated errors, respectively. It also features a mechanism for automatically switching the control mode based on the disturbance amplitude, balancing dynamic performance with resistance to integral saturation. This makes it particularly suitable for situations involving frequent load changes and complex dynamics where the model is not yet fully constructed, such as ultrasonic welding.

[0072] The lightweight PI feedback controller consists of the following functional modules: Mode switching discrimination unit: Receives transient disturbance residual signals, calculates their absolute values, and compares them with a preset dynamic disturbance threshold. When the absolute value of the residual exceeds the threshold, it outputs a "high-amplitude disturbance flag," triggering the integral control path; otherwise, it outputs a "low-amplitude disturbance flag," closing the integral path and maintaining only proportional control. This unit ensures that oscillations caused by integral saturation are avoided under small disturbances, and that steady-state deviations are quickly eliminated under large disturbances.

[0073] Proportional path: includes proportional gain coefficient (K) p Registers and multipliers. The proportional path multiplies the current pure residual control variable by K. p This generates an instantaneous adjustment component that is proportional to the magnitude of the deviation. This component has no phase delay and is used to quickly suppress transient temperature jumps, thereby enhancing the system's response speed.

[0074] Integral adjustment path: includes integral gain parameter (K) i The system consists of a register, a discrete integral accumulator, and an anti-saturation limiter. The integral adjustment path performs discrete-time accumulation of the residual signal, and the accumulated amount is multiplied by K. i The integral correction component is then obtained and used to eliminate steady-state temperature errors caused by model parameter mismatch or constant load disturbances. The anti-saturation limiter restricts the integral accumulation value within a preset upper and lower limit range to prevent excessive integration due to long-term large deviations, which could lead to system overshoot oscillation. When the mode switching discrimination unit closes the integral path, the integral accumulator is frozen or cleared to avoid unnecessary historical accumulation.

[0075] Output synthesis and constraint module: It algebraically sums the components of the proportional path and the integral control path (when the integral path is activated) to generate the initial closed-loop correction control quantity. Then, it performs output limiting (based on the power drive unit safety margin) and first-order low-pass smoothing filtering on the control quantity to suppress high-frequency noise and ensure that the command change rate is within the hardware allowable range, and finally outputs a smooth closed-loop correction component.

[0076] Parameter tuning and construction methods: Parameters of a PI feedback controller (proportional gain K) p Integral gain parameter K i (Integral limit, dynamic disturbance threshold, etc.) are tuned through a combination of offline experiments and theoretical calculations. Proportional Gain K p Determination: Based on the open-loop step response experiment of the third-order thermal inertia model of an inductor, the steady-state gain and time constant of the model output temperature under a unit power step were measured. (Based on Ziegler...) Nichols' method or empirical formula for initial setting of K pThis ensures that, when a typical disturbance residual (e.g., 1°C) is applied, the power regulation generated by the proportional component does not exceed 5% of the transducer's rated power. A 10% adjustment was made to ensure system stability. Subsequently, load tests were conducted on actual welding equipment, and K was fine-tuned. p Continue until the temperature overshoot is less than 2°C and the adjustment time meets the process requirements (e.g., return to the target value within 200ms).

[0077] Integral gain parameter K i Determination of K: Based on the premise that proportional control can stabilize the system, gradually increase K. i Simultaneously monitor the rate of change of the integral accumulation. Set K. i The contribution of the integral correction component during a 1-second residual period is designed to not exceed 30% of the proportional component, thus avoiding phase lag caused by excessive integration. Finally, through repeated welding experiments, the minimum K value that ensures the absolute value of the steady-state temperature error is less than 0.5℃ is selected. i Values ​​are set to ensure robustness.

[0078] Integral limit setting: Based on the maximum allowable power adjustment range of the ultrasonic transducer (e.g., ±15% of the rated power) and the single-cycle maximum step size constraint of the power drive unit, the upper and lower limits of the output of the integrator are mapped to 80% of this power adjustment range, reserving adjustment space for the proportional path. For example, if the total allowable power adjustment is ±500W, the integral limit is set to ±400W.

[0079] Setting the dynamic disturbance threshold: By statistically analyzing the residual signals from multiple welding cycles, the standard deviation of the residuals under normal operating conditions is calculated. The dynamic disturbance threshold is set to 3 times the standard deviation (3σ) or an empirical value (e.g., 0.5℃) to balance the requirements of noise suppression and actual disturbance response. When the absolute value of the residual is less than the threshold, the system locks the integral path and uses only proportional control to avoid integral drift caused by sensor noise.

[0080] Build and deploy: The tuned K p K i Parameters such as integral limit value and dynamic disturbance threshold are written to the non-volatile memory of the main control unit. The controller operates with a fixed sampling period (e.g., 10kHz, or 0.1ms), completing residual reading, mode discrimination, proportional / integral calculation, limit filtering, and output update within each period. To reduce computational load, the accumulation operation in the integral control path uses integer or fixed-point decimal arithmetic, and a limit check is performed after each integral operation to ensure the controller is lightweight and real-time.

[0081] Through the above design, the PI feedback controller achieves efficient suppression of transient disturbances caused by inductor temperature rise and precise elimination of steady-state deviations. At the same time, it avoids the problems of integral saturation and complex parameter tuning of traditional PID controllers in noisy environments, and significantly enhances the robustness and engineering practicality of ultrasonic welding power adaptive control.

[0082] The system receives the clean residual control variable after preprocessing in step S4.3. This variable represents the dynamic deviation between the actual temperature rise of the inductor and the predicted value of the third-order thermal inertia model, serving as the sole input source for the PI feedback controller. It calls the pre-installed lightweight proportional-integral control module in the embedded main control unit, reads the value of the clean residual control variable for the current control cycle, and retrieves the stored integral cumulative state value from the previous control cycle. It performs a proportional gain amplification operation, multiplying the current clean residual control variable by a preset proportional gain coefficient Kp to generate a proportional correction component reflecting the instantaneous temperature deviation response speed. This component aims to quickly suppress transient temperature fluctuations caused by sudden changes in welding load. It then performs an integral accumulation operation, adding the product of the current clean residual control variable and the control cycle sampling time Ts to the integral state value of the previous cycle to generate an integral correction term reflecting the persistence of historical temperature deviations. This term is used to eliminate steady-state temperature errors caused by model parameter mismatch or environmental disturbances. The generated integral correction term undergoes anti-saturation limiting processing to determine if its absolute value exceeds a preset integral upper limit threshold. If it does, it is clamped to the threshold boundary to prevent integral saturation caused by long-term large deviations, which could lead to system overshoot oscillation. The anti-saturation-processed integral correction term is multiplied by a preset integral gain coefficient Ki to generate the final integral correction component, which focuses on eliminating systemic temperature drift in the low-frequency range. The aforementioned proportional correction component and integral correction component are algebraically added to synthesize an initial closed-loop correction control quantity that includes fast dynamic response capability and steady-state accuracy maintenance capability. Through the parallel processing mechanism of proportional amplification and integral accumulation, the pure residual control variable obtained in the previous step is transformed into initial closed-loop correction control quantity data with dual regulation characteristics, realizing preliminary compensation for transient disturbances caused by inductor temperature rise and gradual elimination of steady-state deviation, providing a basic control signal for output limiting and smoothing filtering in the subsequent S4.5 step.

[0083] For example, in the main control MCU of the ultrasonic welding equipment, the proportional gain coefficient Kp of the PI controller is set to 0.85, the integral gain coefficient Ki to 0.12, the control cycle sampling time Ts to 2ms, and the integral upper limit threshold to the power equivalent value corresponding to ±15.0℃. At the 50ms of the current welding cycle, the pure residual control variable output by step S4.3 is +2.5℃, indicating that the actual inductor surface temperature is 2.5℃ higher than the model prediction. The system reads the integral cumulative state value of the previous cycle (48ms) as 0.3℃·s. First, the proportional operation is performed to calculate the proportional correction component as 0.85×2.5, obtaining the power adjustment corresponding to 2.125℃. Then, the integral operation is performed to calculate the current integral term increment as 2.5×0.002, obtaining 0.005℃·s, which is added to the state value of the previous cycle to obtain a new integral cumulative state value of 0.3+0.005, or 0.305℃·s. The integral state value of 0.305℃·s is determined to be below the integral upper limit threshold of 15.0℃·s, therefore no limiting is required. The integral correction component is then calculated as 0.12 × 0.305, yielding the power regulation corresponding to 0.0366℃. Finally, the proportional correction component of 2.125 is added to the integral correction component of 0.0366 to generate the initial closed-loop correction control quantity corresponding to the power regulation command of 2.1616℃. If the pure residual control variable suddenly changes to -10.0℃ (a large negative deviation), the proportional correction component is 0.85 × 10.0, which is -8.5. The integral increment is 10.0 × 0.002, i.e., -0.02℃·s, after accumulation, the integral state value becomes 0.285℃·s, and the integral correction component is 0.12 × 0.285, i.e., 0.0342. The initial closed-loop correction control quantity is... 8.5 + 0.0342, or -8.4658. This approach ensures rapid response primarily through proportional control under small disturbances, while introducing integral action to eliminate steady-state error under sustained deviations, significantly improving the dynamic stability and tracking accuracy of the closed-loop control.

[0084] S4.5: Perform output limiting and smoothing filtering on the initial closed-loop correction control quantity to generate the final closed-loop correction component used for power command synthesis.

[0085] The system receives the initial closed-loop correction control quantity generated from the preceding steps. This control quantity includes the proportionally amplified component and the integral cumulative component for the transient disturbance residual signal, serving as the input for subsequent safety constraint processing. An absolute amplitude extraction operation is performed on the initial closed-loop correction control quantity to obtain the original intensity value of the feedback adjustment command within the current cycle, providing benchmark data for subsequent dynamic limiting determination. Based on the linear operating region characteristics of the ultrasonic transducer power drive unit and the thermal shock tolerance threshold of the inductor winding, bidirectional symmetrical output limiting boundary values ​​are set. The positive limiting value corresponds to the maximum allowable power increment, and the negative limiting value corresponds to the maximum allowable power decrement, preventing resonance loss of lockout or thermal breakdown of the insulation layer due to sudden changes in drive current caused by excessive feedback gain. The original intensity value is compared with the bidirectional symmetrical output limiting boundary value using logical operations. If the original intensity value exceeds the positive limiting value, it is truncated to the positive limiting value; if it is below the negative limiting value, it is truncated to the negative limiting value; if it is in between, the original value remains unchanged, generating an intermediate correction control quantity that has undergone hard limiting processing to ensure that the power regulation command is always within the hardware's safe operating range. A first-order low-pass smoothing filter is applied to the hard-limited intermediate correction control quantity. An exponentially weighted moving average method in the discrete time domain is used to construct a smoothing filter transfer function that includes a forgetting factor to suppress high-frequency chatter in the control quantity caused by residual high-frequency noise from the sensor or minor load fluctuations. The final closed-loop correction component after smoothing is calculated using the following formula: in, This is the final closed-loop correction component for the k-th control cycle. The historical value of the final closed-loop correction component in the (k-1)th control cycle. This is the intermediate correction control value after hard limiting in the k-th control cycle. The smoothing coefficient, ranging from 0 to 1, is used to adjust the balance between the system's dependence on historical states and its response speed to new inputs. The smoothing coefficient is determined based on the matching relationship between the mechanical vibration frequency and the control cycle in the ultrasonic welding process. The value is set to 0.85 to balance response speed and smoothness. In the k-th control cycle, assume the intermediate correction control value after hard limiting is 3.2% (corresponding to a digital value of 320), and the final closed-loop correction component of the previous cycle is 2.5% (corresponding to a digital value of 250). Substitute into the formula to calculate: That is, the final closed-loop correction component is 2.605%. If the intermediate correction control value in the next cycle suddenly changes to -4.0% (corresponding to a digital value of -400), while the final value in the previous cycle is 2.605% (corresponding to a digital value of 260.5), then the calculation result is: That is, 1.61425%. As can be seen, although the input signal undergoes a sharp change from positive to negative, after smoothing, the output command changes gradually, avoiding the instantaneous large current impact of the power drive unit switching transistor, and significantly improving the stability of the welding process and the service life of the inductor winding.

[0086] Step S5: Based on the slope change rate of the predicted temperature rise value sequence, identify the state characteristics of the temperature rise transitioning from a gradual transition to an inflection point or entering a plateau region, trigger the feedforward channel to generate a power gradient pre-adjustment action command, and delay the power gradient pre-adjustment action command to the next zero-voltage crossover interval when the vibration phase deviation is detected to exceed the limit. Specifically, it includes: S5.1: Perform first-order difference operation on the predicted temperature rise trend value sequence to obtain the temperature rise slope feature vector that characterizes the rate of temperature change within the future time domain window, as the quantitative basis for subsequent thermal state discrimination.

[0087] The system receives a sequence of predicted temperature rise trends generated in step S3, encompassing three key time-domain windows: 50ms, 100ms, and 200ms. This sequence characterizes the expected thermal dynamic trajectory of the inductor under the coupling effect of fast and slow components. Discrete-time index reconstruction is performed on the predicted temperature rise trend sequence to extract continuous temperature prediction data points within the current control period t and its historical windows, constructing a sliding data buffer for differential operations. Based on the reconstructed discrete temperature prediction data, the forward differencing method is used to calculate the temperature change between adjacent sampling times to quantify the temperature rise rate per unit time interval. The temperature rise slope feature value at the k-th sampling time is calculated. All predicted temperature data points within the future multi-time-domain windows are traversed, and the first-order differencing operation described above is performed sequentially to generate instantaneous temperature rise rate values ​​corresponding to each predicted time node. The calculated instantaneous temperature rise rate values ​​are vector-encapsulated according to timestamp order to form a temperature rise slope feature vector characterizing the evolution of the temperature change rate within the future time-domain window. Boundary condition checks are performed on the temperature rise slope feature vector to ensure the integrity of the first and last data points and prevent slope calculation distortion caused by data truncation. Through the aforementioned first-order difference operation, the temperature rise trend prediction sequence from the previous step is transformed into a temperature rise slope feature vector characterizing the rate of temperature change within the future time-domain window. This vector serves as the quantitative basis for subsequent thermal state determination and power gradient pre-adjustment triggering, achieving a refined characterization of the inductor's thermal dynamic evolution trend.

[0088] S5.2: Based on the temperature rise slope feature vector, identify the critical moment when the temperature rise transitions from the flat zone to the rising inflection point or the steady-state characteristics when it enters the plateau zone, and generate a thermal state switching event marker containing a timestamp and state type.

[0089] The system receives a temperature rise slope feature vector generated in step S5.1, representing the rate of temperature change within a future time window. This vector contains the first derivative data of the temperature rise in a discrete time series, serving as the input benchmark for dynamic thermal state discrimination. A sliding window second-order difference operation is performed on the temperature rise slope feature vector to calculate the change in slope between adjacent sampling points, obtaining a second-order derivative sequence reflecting the temperature rise acceleration characteristics. This sequence quantifies the curvature change trend of the inductor's thermal response curve. An extreme value detection logic is constructed based on this second-order derivative sequence, setting positive and negative bidirectional threshold boundaries to identify inflection point features where the slope changes from a gradual to a sharp increase, or plateau features where high growth transitions to a stable state. When the second-order derivative value is detected to cross zero from the negative range into the positive range and persist for more than a preset time window length, the current moment is determined to be the critical moment when the temperature rise transitions from a gradual region to an inflection point, at which point the rate of heat accumulation inside the inductor begins to accelerate significantly. When the second derivative value is detected to cross from the positive range to the negative range and stabilize within a small fluctuation range near zero, the current moment is determined to be a steady-state characteristic moment when the temperature rise enters the plateau region. At this time, the inductor's heat dissipation and heat generation reach a dynamic balance. For each identified critical moment or steady-state characteristic moment, its precise timestamp in the system's global clock is extracted and combined with the currently detected thermal state type to generate a structured event marker. The timestamp and thermal state type are encapsulated into a standardized thermal state transition event marker data packet, where the state type field is explicitly identified as "rising inflection point" or "plateau region entry," and the timestamp field records the specific microsecond-level moment that triggered the state transition. Through the above-mentioned second derivative extreme value detection processing method, the temperature rise slope feature vector obtained in the previous step is transformed into a thermal state transition event marker containing timestamps and state types. This achieves accurate capture of key nodes in the dynamic evolution of inductor thermal performance, providing a precise trigger basis for the subsequent S5.3 step of calling the power gradient mapping table, effectively solving the problem of untimely power regulation caused by the lag in temperature rise feedback in traditional control.

[0090] S5.3: Based on the thermal state switching event flag, a preset power gradient mapping table is called to perform a lookup operation to generate an initial power gradient pre-adjustment action command corresponding to the current thermal dynamic characteristics, thereby realizing feedforward compensation control for the predicted temperature rise trend.

[0091] The preset power gradient mapping table is a data lookup table pre-stored in the non-volatile memory of the main control unit of the ultrasonic welding equipment in this invention. It is used to map inductor thermal state switching events (such as the temperature rise changing from a gradual transition to an inflection point or entering a plateau region) to corresponding power gradient pre-adjustment action commands (including power adjustment amplitude and direction). This mapping table provides the feedforward channel with a fast and deterministic control decision basis, eliminating the need for complex online calculations, ensuring a response to thermal state changes within microseconds, and achieving advance compensation for predicted temperature rise trends.

[0092] The preset power gradient mapping table adopts a two-dimensional table structure and specifically includes the following elements: Input index column (hot state type): "Inflection point": This indicates that the rate of temperature rise changes from a gradual increase to a sharp increase, and a temperature overshoot is about to occur. "Platform zone entry": This indicates that the rate of temperature rise tends to be stable, and the heat generation and heat dissipation are close to a dynamic equilibrium. Other optional states include "rapidly rising" and "slowly falling", but the basic implementation includes at least the two mentioned above.

[0093] Input auxiliary parameter column (temperature rise slope change rate level): To refine the mapping accuracy, the temperature rise slope change rate (second derivative) can be further graded (e.g., low, medium, and high levels), which can be combined with the thermal state type to form a composite index.

[0094] Output command line (power gradient pre-adjustment action command): includes the power adjustment range (e.g., +X%, -Y%, or a specific power value) and the adjustment direction (increase or decrease). For "rising inflection point", the output command usually appropriately reduces the power (e.g., a slight decrease of 0.5%~2%) to suppress the impending overshoot; for "plateau zone entry", the output command maintains the current power or slightly increases the power (e.g., +0.2%~0.5%) to stabilize the welding energy.

[0095] Optional confidence level or sensitivity factor: used to dynamically adjust the output amplitude under different welding workpieces or working conditions, but in this patent, the preset value is based on the standard working condition calibration.

[0096] The mapping table is constructed through a combination of offline experiments and theoretical simulations. The specific steps are as follows: Offline thermal power response experiment: Under standard welding conditions (specified workpiece material, thickness, welding head pressure, etc.), a typical power step sequence is applied to the inductor, and the predicted temperature rise trend sequence and the actual temperature rise trajectory are recorded simultaneously. Thermal state switching events are identified (e.g., via S5.1). The second derivative extremum detection described in S5.2 is used to record the actual power demand change (i.e., the power gradient adjustment required to maintain temperature stability) when the event occurs. Multiple sets of experiments are repeated to cover different power levels and different heat dissipation environments.

[0097] Data Statistics and Threshold Classification: Experimental data are grouped according to thermal state type (rising inflection point, plateau entry). For each group, the optimal power adjustment range after the event occurs is statistically analyzed (e.g., the power change value that minimizes temperature overshoot or steady-state error). The median or mean of the adjustment range for that group is calculated as the default output command corresponding to that state. If a strong correlation is found between the rate of change of the temperature rise slope and the required adjustment range, the rate of change of the slope is divided into several levels (e.g., low, medium, high), and the corresponding adjustment ranges are statistically analyzed for each level, forming a secondary index.

[0098] Verification and Fine-tuning: Load the constructed mapping table into the actual welding controller and conduct online testing. Observe the feedforward compensation effect. If insufficient or over-compensation occurs, fine-tune the values ​​of the corresponding table entries according to the deviation. Repeat the iteration until the temperature fluctuation meets the process requirements under various typical operating conditions (e.g., overshoot less than ±2℃).

[0099] Solidification and Storage: The finalized mapping table is stored in the flash memory of the main control unit as an array or constant structure for real-time table lookup in step S5.3. The data in the table can be configured to support online hot updates (e.g., modification via a host computer) to adapt to different workpiece or process requirements.

[0100] Through the above construction, the preset power gradient mapping table realizes a fast and deterministic mapping from thermal dynamic characteristics to control actions, avoiding online optimization calculations and significantly improving the real-time performance and interpretability of feedforward control.

[0101] S5.4: Perform real-time zero-crossing tracking processing on the ultrasonic transducer drive voltage waveform to extract the zero-voltage crossing interval boundary parameters of the current vibration cycle, and synchronize them with the system clock to generate a phase-locked reference window.

[0102] The zero-voltage crossover range refers to an extremely short time-domain window in which the voltage amplitude of the ultrasonic transducer drive voltage waveform is below a preset safety threshold in each vibration cycle. This window is typically centered at the zero-crossing point, and its width is determined by the safe commutation time of the power switching devices. Executing power regulation commands within this range can minimize switching losses and avoid resonant lockout.

[0103] The system receives the initial power gradient pre-adjustment command generated in step S5.3. This command includes the power adjustment amplitude and direction determined based on the inductor temperature rise trend prediction, serving as the control input source for phase-locked processing. The high-speed analog-to-digital converter (ADC) within the main control MCU is invoked to sample the ultrasonic transducer drive voltage waveform in real time. The sampling frequency is set to more than 20 times the ultrasonic operating frequency to capture the precise zero-crossing characteristics of the voltage waveform. Sliding window linear interpolation is performed on the acquired discrete voltage sampling point sequence to construct a high-precision voltage continuity function approximation model, used to eliminate the zero-crossing point positioning quantization error caused by discrete sampling. Based on the voltage continuity function approximation model, the Newton-Raphson iterative algorithm is used to solve for the moment when the voltage amplitude is zero, accurately calculating the theoretical timestamp of the voltage crossing from the positive half-cycle to the negative half-cycle or from the negative half-cycle to the positive half-cycle within the current vibration cycle, and extracting the center time parameter of the zero-voltage crossing point. Based on the safe commutation characteristics of IGBT or MOSFET switching devices in the ultrasonic power drive unit, a half-width time threshold for the zero-voltage crossover interval is set, typically 1.5 times the sum of the switching device's turn-off delay time and turn-on delay time. This is used to construct a safe phase window boundary centered on the zero-voltage crossover point. The current value of the system's global clock counter is compared logically with the start and end boundaries of the safe phase window to generate a phase state flag indicating whether the current moment is within the allowable power regulation operation range. If the current moment is within the safe phase window, the initial power gradient pre-adjustment action command is marked as immediately executed; if the current moment is outside the safe phase window, the remaining waiting time from the current moment to the start boundary of the next safe phase window is calculated, and the initial power gradient pre-adjustment action command is suspended until the waiting time ends. Through the above-mentioned real-time zero-crossing tracking and phase window synchronization processing, the initial power gradient pre-adjustment action command generated in the previous step is transformed into a timing alignment control signal strictly constrained by the ultrasonic vibration phase. This achieves the expected technical effect of executing power regulation actions at the moment of minimum electrical stress, effectively avoiding the risk of transducer resonance lock-up and power device thermal breakdown caused by non-zero voltage switching.

[0104] S5.5: The initial power gradient pre-adjustment action command is compared and verified with the phase-locked reference window. If the command trigger time is detected to exceed the zero voltage crossing interval, the delay suspension logic is executed to generate the final composite power adjustment command executed within the safe phase window.

[0105] Step S6: Determine whether the absolute value of the transient disturbance residual signal exceeds a set threshold. If it does, perform integral accumulation processing on the closed-loop correction component; otherwise, only perform proportional amplification processing to generate the final feedback correction command. Specifically, this includes: S6.1: Obtain the transient disturbance residual signal generated by the previous steps, perform absolute value operation on the transient disturbance residual signal to obtain the residual absolute value data that characterizes the current temperature prediction deviation amplitude, and provide a quantitative input benchmark for subsequent mode switching determination.

[0106] S6.2: Based on a preset dynamic disturbance discrimination threshold, the absolute value of the residual data is compared with the dynamic disturbance discrimination threshold by a size comparison logic operation to generate a mode switching control signal containing a high-amplitude disturbance flag or a low-amplitude disturbance flag, thereby clarifying the adjustment strategy type required for the current control cycle.

[0107] S6.3: When the mode switching control signal indicates a high-amplitude disturbance, the integral gain parameter in the PI feedback controller is called to perform cumulative integral adjustment processing on the transient disturbance residual signal to generate an integral correction component for eliminating persistent steady-state temperature deviation, thereby ensuring the tracking accuracy of the system under large load changes.

[0108] In response to the high-amplitude disturbance flag indicated by the mode switching control signal, the system activates the integral adjustment path in the PI feedback controller, reading the transient disturbance residual signal at the current moment as the input variable for the integral operation. The system obtains the integral cumulative state value from the previous control cycle, stored in the non-volatile memory area or high-speed RAM register of the main control unit, representing the accumulated temperature deviation that was not completely eliminated at a historical time. The system multiplies the current transient disturbance residual signal with a preset integral gain coefficient to calculate the incremental integral term for the current cycle. The integral gain coefficient is tuned based on the time constant of the slowly varying component of the inductive thermal inertia model to ensure effective tracking of persistent steady-state deviations without introducing excessive lag. Using the discrete trapezoidal integration method or the forward Euler method, the incremental integral term for the current cycle is algebraically summed with the integral cumulative state value from the previous control cycle to generate the updated original value of the integral correction component. The generated integral correction component's original value undergoes anti-saturation limiting processing. Upper and lower thresholds are set for the integral output, corresponding to the maximum power adjustment margin allowed by the ultrasonic welding power drive unit. This prevents integral saturation caused by long-term large deviations, thus avoiding system overshoot. If the original value of the integral correction component exceeds the upper threshold, it is clamped to the upper limit; if it is below the lower threshold, it is clamped to the lower limit; if it is within the threshold range, the original value remains unchanged, generating a limited integral correction component. This integral correction component is marked as high-priority closed-loop correction data and a timestamp is added, ready to be sent to the subsequent vector synthesis module. Through the above discrete integration operation and anti-saturation limiting processing, the transient disturbance residual signal is transformed into an integral correction component with memory characteristics. This achieves accurate elimination of the persistent steady-state temperature deviation caused by sudden load changes during inductor temperature rise, ensuring the system's tracking accuracy and stability under large load variations.

[0109] S6.4: When the mode switching control signal indicates a low amplitude disturbance, the integral path in the PI feedback controller is locked and only the proportional gain parameter is activated. Pure proportional amplification is performed on the transient disturbance residual signal to generate a proportional correction component for quickly suppressing small fluctuations, thus avoiding system oscillation caused by integral action under small disturbances.

[0110] When the mode switching control signal indicates a low-amplitude disturbance, it means that the deviation between the actual measured value of the current inductor temperature rise and the predicted value of the third-order thermal inertia model is within a small fluctuation range. At this time, the system determines that the main source of interference is high-frequency noise from the sensor or instantaneous micro-vibrations from the welding load, rather than structural thermal accumulation deviation. Therefore, the integral path in the PI feedback controller is locked to prevent low-frequency oscillations caused by error accumulation. The current residual value in the transient disturbance residual signal sequence is read. This residual value characterizes the degree of instantaneous deviation of the measured temperature from the model's predicted trajectory and serves as the direct input variable for proportional control. The proportional gain coefficient pre-stored in the main control unit memory is called. This coefficient has been offline identified and online fine-tuned to ensure sufficient response sensitivity under small disturbances while maintaining the system's phase margin. Pure proportional amplification is performed, multiplying the transient disturbance residual signal by the proportional gain coefficient to generate an initial proportional correction component. Dead-zone filtering is performed on the generated initial proportional correction component. If its absolute value is less than the set minimum action threshold, it is forcibly set to zero to eliminate invalid power fine-tuning commands caused by quantization noise. Output limiting constraints are applied to the proportional correction component after dead-zone filtering to ensure its value range is strictly limited within the minimum adjustment step size allowed by the ultrasonic power drive unit, preventing power overshoot due to excessive proportional gain. The limited proportional correction component is marked as the final proportional correction component, which only contains rapid response information to the current instantaneous deviation and does not contain the cumulative effect of historical errors. Through pure proportional amplification and limiting processing, the transient disturbance residual signal from the previous step is converted into proportional correction component data without phase hysteresis, achieving the expected technical effect of rapidly suppressing small temperature fluctuations and avoiding system oscillations caused by integral saturation under low-amplitude disturbance conditions.

[0111] S6.5: Based on the selection result of the mode switching control signal, select effective data from the integral correction component or the proportional correction component as output variables, generate the final feedback correction command, and transmit the final feedback correction command to the power combining module to complete the construction of the closed-loop correction component.

[0112] The system receives the integral or proportional correction component generated in the preceding steps S6.3 or S6.4, and the mode switching control signal generated in S6.2, as the input data source for this step. Based on the logic level of the mode switching control signal, a multiplexer logic judgment operation is performed. When the signal indicates a high-amplitude disturbance, the integral correction component is locked as a valid output variable, while the proportional correction component is set to zero or masked. When the signal indicates a low-amplitude disturbance, the proportional correction component is locked as a valid output variable, while the integral correction component is set to zero or masked, thus ensuring that only one adjustment mechanism is effective within the same control cycle, avoiding overshoot oscillation caused by dual adjustment. Dynamic limiting processing is performed on the selected valid correction component. Upper and lower threshold values ​​are set to correspond to the maximum allowable adjustment step size and minimum resolution of the ultrasonic power drive unit, respectively, to prevent power command overflow caused by residual abrupt changes. A first-order low-pass filter is applied to the limited correction component to filter out high-frequency quantization noise and electromagnetic interference glitches, generating a smooth and continuous feedback correction value sequence. The feedback correction value sequence is converted into a standardized digital control word and transmitted to the power combining module via the internal bus, serving as the basis for constructing the closed-loop correction components. Through the aforementioned mode selection, amplitude limiting protection, and filtering smoothing processes, the discrete correction values ​​from the previous step are transformed into a final feedback correction command that is physically feasible and free from high-frequency jitter, achieving precise suppression of transient disturbances and a significant improvement in system stability.

[0113] For example, during ultrasonic welding, the main control MCU executes sub-step S6 at a frequency of 10kHz. Assume the current time t... k The mode switching control signal output by S6.2 is a "high-amplitude disturbance flag," indicating that the absolute value of the deviation between the measured temperature and the predicted temperature exceeds the set dynamic threshold of 0.5℃. At this time, the integral correction component calculated by S6.3 is +1.2% of the rated power, and the proportional correction component calculated by S6.4 is +0.3% of the rated power. Based on the mode signal, the system selects the integral correction component +1.2% as the initial output and masks the proportional component. The dynamic limit range of power adjustment is set to ±2.0% of the rated power / cycle. Since +1.2% does not exceed the limit, it passes directly. Subsequently, a first-order low-pass filter with a cutoff frequency of 500Hz is used to smooth +1.2%, resulting in the final feedback correction command +1.18%. This command is sent to the power combining module. If the next time t k+1The mode switching control signal changes to a "low amplitude disturbance flag," and the absolute value of the deviation drops below 0.2℃. At this time, the integral component output by S6.3 retains its historical cumulative value but is masked by the system; the proportional correction component output by S6.4 is -0.1% of the rated power. The system selects -0.1% as the effective output, and after limiting and filtering, generates the final feedback correction command of -0.09%. This process ensures that the integral action can quickly eliminate steady-state error when there is a large deviation, and the proportional action can quickly respond when there is a small deviation, while avoiding oscillations caused by integral saturation, significantly improving the dynamic response speed and steady-state accuracy of inductor temperature rise control.

[0114] Step S7: The power gradient pre-adjustment action command and the feedback correction command are vector-synthesized to generate a composite power adjustment command that includes both feedforward prediction and feedback correction characteristics. Specifically, this includes: S7.1: The power gradient pre-adjustment action command output by the feedforward channel and the closed-loop correction component output by the feedback channel are time-stamp aligned to eliminate timing deviations caused by different paths of feedforward prediction and feedback residual calculation, and generate synchronized feedforward control quantity sequence and synchronized feedback control quantity sequence with unified time base.

[0115] S7.2: Based on the synchronized feedforward control sequence and the synchronized feedback control sequence, a multi-source control signal fusion operation is performed using the weighted vector superposition method to balance the dominant role of feedforward prediction and the compensating role of feedback correction, thereby generating an initial composite power regulation command vector.

[0116] The system receives the synchronized feedforward control sequence and synchronized feedback control sequence after timestamp alignment, serving as the initial input data for multi-source signal fusion. The power gradient pre-adjustment component is extracted from the synchronized feedforward control sequence; this component represents the reference power adjustment magnitude determined based on the future temperature rise trend predicted by the third-order inductor thermal inertia model. The closed-loop correction component is extracted from the synchronized feedback control sequence; this component represents the real-time error compensation value generated by processing the transient disturbance residual signal through a PI controller. A dynamic weighted vector superposition model is constructed, defining the feedforward channel weight coefficient and the feedback channel weight coefficient. The feedforward channel weight coefficient is dynamically adjusted based on the temperature rise prediction confidence level, and the feedback channel weight coefficient is dynamically adjusted based on the residual signal amplitude. The feedforward channel weight coefficient at the current moment is calculated using the following formula: Among them, w ff Here, is the feedforward channel weight coefficient, k is the sensitivity adjustment factor, and C is the model prediction confidence index, which is obtained by normalizing the inverse of the prediction error of the previous cycle. The feedback channel weight coefficient at the current time is determined by the following formula: Among them, w fb The feedback channel weighting coefficients are used to ensure that the sum of the weights of the two channels remains constant at 1, maintaining system energy balance. Weight mapping is performed on the synchronized feedforward control sequence, multiplying the feedforward power gradient pre-adjustment component by the corresponding feedforward channel weighting coefficient to generate a weighted feedforward control vector. Weight mapping is also performed on the synchronized feedback control sequence, multiplying the feedback closed-loop correction component by the corresponding feedback channel weighting coefficient to generate a weighted feedback control vector. Vector addition is performed, linearly superimposing the weighted feedforward control vector and the weighted feedback control vector in the complex or real domain to generate an initial composite power adjustment command vector. An amplitude limiting check is performed on this initial composite power adjustment command vector; if the vector amplitude exceeds the upper limit of the ultrasonic transducer's rated power, it is scaled proportionally to a safe threshold range. A rate-of-change smoothing process is performed on this initial composite power adjustment command vector, using a first-order low-pass filter to suppress high-frequency jitter in the command caused by abrupt weight changes, ensuring the continuity of the power adjustment curve. Through the aforementioned weighted vector superposition and post-processing mechanism, the dominant role of feedforward prediction and the compensating role of feedback correction are organically integrated, transforming them into an initial composite power regulation command vector with dynamic response characteristics and phase safety. This achieves coordinated control of the fast and slow components of inductor temperature rise, significantly improving the real-time performance and stability of power regulation.

[0117] S7.3: Perform ultrasonic transducer resonant phase window constraint verification on the initial composite power adjustment command vector to identify and eliminate invalid power jump points located in the non-zero voltage crossing range, and generate a set of phase-safe and compliant power adjustment commands to be executed.

[0118] The system receives the initial composite power adjustment command vector generated in the preceding step S7.2. This vector contains a feedforward component based on inductor temperature rise prediction and a feedback component based on transient residual correction, serving as the input data source for phase constraint verification. Simultaneously, the system acquires the real-time voltage sampling signal of the ultrasonic transducer drive circuit, capturing the instantaneous voltage waveform sequence of the current vibration cycle using a high-speed analog-to-digital converter at a sampling rate 10 times higher than the carrier frequency. A sliding window zero-crossing detection is performed on the instantaneous voltage waveform sequence to identify the critical time point where voltage polarity reverses, determining the boundary of the zero-voltage crossing interval where the voltage amplitude is below a preset safety threshold within the current cycle. A phase-locked reference window is constructed, mapping the start and end times of the zero-voltage crossing interval to the timestamp coordinates of the system's master clock, forming a set of safe time domains allowing power jumps. Candidate execution times where the power amplitude undergoes a step change are extracted from the initial composite power adjustment command vector, and compared logically with the timestamp coordinates of the phase-locked reference window. If a candidate execution time falls within the zero-voltage crossing interval, the command point is determined to be a phase safety compliance point, and its original power amplitude and timestamp are retained. If the candidate execution time is outside the zero-voltage crossing interval, i.e., during the high-voltage or high-current conduction phase, the command point is determined to be an invalid power transition point, triggering a delay suspension logic. The time difference from the current invalid time to the start of the next zero-voltage crossing interval is calculated and added as a delay offset to the command execution time axis. The execution timing of the power adjustment action is repositioned based on the delay offset, generating a power adjustment command set to be executed after phase alignment correction. Through the aforementioned phase window constraint verification and dynamic delay mechanism, the initial composite power adjustment command generated in the previous step is transformed into a phase-safe and compliant power adjustment command set to be executed, achieving strict synchronization between the power adjustment action and the ultrasonic resonance phase. This eliminates the risk of transducer lockout and electromagnetic interference noise caused by non-zero voltage switching, ensuring the stability and consistency of welding energy output.

[0119] For example, the ultrasonic welding operating frequency is set to 20kHz, corresponding to a vibration period T of 50μs. The main control MCU collects the inverter bridge arm output voltage at a sampling rate of 1MHz, collecting 50 voltage sample points per cycle. The preset zero-voltage crossing safety threshold is 5% of the DC bus voltage, that is, when the absolute value of the voltage amplitude is less than 1.5V, it is determined to be in the zero-voltage range. During a certain welding process, the initial composite power adjustment command vector requires the output power to be adjusted from 800W to 850W at t=12.3ms. The system detects in real time that the current voltage waveform crosses the zero point between t=12.298ms and t=12.302ms, forming a zero-voltage crossing window with a width of 4μs. Since the command execution time of 12.3ms is near the center of this window, it meets the condition |12.3 -12.300|<2μs, and is determined to be phase safe and compliant, so the command is directly retained. At another moment, t=15.6ms, the command requests a power reduction from 850W to 800W. However, the voltage is near its peak at this time, and the nearest zero-voltage crossing window is expected to occur at t=15.625ms. The system determines 15.6ms as an invalid transition point and calculates the delay offset Δt = 15.625 - 15.6 = 0.025ms. The system postpones the execution of the power reduction command to 15.625ms and updates the set of commands to be executed. Through this processing, all power switching actions are strictly limited to within ±2μs of the voltage zero-crossing point, effectively avoiding voltage spikes caused by hard switching losses. The transducer resonant frequency drift is controlled within ±5Hz, significantly improving the electrical safety and welding quality stability of the ultrasonic welding system.

[0120] S7.4: Based on the phase-safe and compliant power adjustment instruction set to be executed, the power drive unit pulse width modulation mapping function is used to perform the conversion processing of digital quantity to duty cycle signal, so as to adapt to the hardware interface protocol of the ultrasonic welding power drive unit and generate the final composite power adjustment instruction pulse sequence.

[0121] The system receives a set of power adjustment instructions to be executed, which has undergone phase safety compliance verification. This set of instructions contains discrete power amplitude requirements and corresponding timestamps. The power amplitude data in the set of instructions is normalized and mapped to the range of the counting register of the pulse width modulation (PWM) generator in the power drive unit, generating a standardized duty cycle control variable. Based on the switching frequency characteristics of the ultrasonic welding power supply inverter bridge arm, a fixed time reference for the PWM carrier period is determined, and the normalized duty cycle control variable is converted into an equivalent high-level duration parameter. Using digital comparator logic, the high-level duration parameter is compared in real time with the incrementally counted triangular or sawtooth wave carrier signal within the system to generate an initial PWM switching drive pulse sequence. For the generated initial PWM switching drive pulse sequence, dead-time insertion processing is performed, inserting a microsecond-level non-conduction protection interval at the instant the complementary switch turns on to prevent shoot-through short circuits between the upper and lower bridge arms. The PWM pulse sequence after dead time insertion is edge-aligned to ensure that the power regulation action's transition edge falls precisely at the center of the zero-voltage crossing interval determined in the previous steps, eliminating phase offset errors caused by the dead time. The optimized PWM pulse sequence is encoded into a digital signal format conforming to the hardware interface protocol and transmitted to the gate drive terminal of the inverter bridge arm through an isolated drive circuit to generate the final composite power regulation command pulse sequence. Through the precise conversion of digital signals to duty cycle signals and phase alignment processing described above, the abstract power regulation command is transformed into a physically executable switching control signal, achieving microsecond-level precise control of ultrasonic welding output power and rapid response of thermal inertia feedforward compensation.

[0122] S7.5: Output the final composite power adjustment command pulse sequence to the ultrasonic welding power drive unit to drive the inverter bridge arm switch to dynamically adjust the output power curve, thereby realizing intelligent distribution and closed-loop control of welding energy under the changing trend of inductor temperature rise.

[0123] Step S8: After each welding operation, the cumulative deviation between the actual temperature rise trajectory and the predicted temperature rise trend sequence is compared. If three consecutive deviations show the same direction, the parameter update process is initiated, and only the surface thermal resistance parameter in the equivalent heat capacity and thermal resistance of the surface fiber coating is incrementally corrected to complete the online fine-tuning of the model. Specifically, this includes: S8.1: Obtain the actual temperature rise trajectory data sequence after the end of the current welding cycle and the temperature rise trend prediction value sequence generated by the third-order thermal inertia model of inductance. Perform time-domain alignment processing on the two to eliminate sampling timestamp deviation and generate a residual time series dataset containing point-by-point temperature difference values, which serves as the basic input for subsequent deviation morphology analysis.

[0124] S8.2: Perform cumulative deviation integration based on the residual time series dataset to calculate the total energy deviation value in the entire welding cycle, and combine the polarity distribution characteristics of the residual signal extracted by the sliding window to generate a deviation morphology identifier vector that represents the direction of systematic deviation predicted by the model.

[0125] S8.3: Logically match and judge the deviation pattern identification vector stored in history with the currently generated deviation pattern identification vector, count the number of welding cycles in which the same offset feature appears consecutively, and if the count value reaches the preset three thresholds, trigger the parameter update enable signal; otherwise, reset the counter and maintain the existing inductor third-order thermal inertia model parameter set unchanged.

[0126] S8.4: In response to the parameter update enable signal, lock the thermal diffusion time constant of the intermediate filament winding structure and the overall winding thermal mass dominant time constant in the parameter set of the third-order thermal inertia model of the inductor, select only the surface thermal resistance parameter in the equivalent heat capacity and thermal resistance of the surface fiber varnish film as the variable to be corrected, and generate the incremental correction value of the surface thermal resistance parameter based on the gradient direction calculated by the deviation morphology identification vector.

[0127] S8.5: The incremental correction value of the surface thermal resistance parameter is superimposed on the original surface thermal resistance parameter, and a boundary constraint check is performed to ensure that the corrected parameter value is within a physically interpretable reasonable range. Finally, the parameter set of the third-order thermal inertia model of the inductor, which has been fine-tuned online, is output for use in the feedforward derivation of the next welding cycle.

[0128] For those skilled in the art, various other corresponding changes and modifications can be made based on the technical solutions and concepts described above, and all such changes and modifications should fall within the protection scope of the claims of this invention.

[0129] Unless otherwise defined, the technical or scientific terms used herein shall have the ordinary meaning as understood by one of ordinary skill in the art to which this application pertains. The terms “first,” “second,” “third,” and similar terms used in this patent application specification and claims do not indicate any order, quantity, or importance, but are merely used to distinguish different components. Similarly, the terms “an” or “a” and similar terms do not indicate a quantity limitation, but rather indicate the presence of at least one. The terms “comprising” or “including” and similar terms mean that the elements or objects preceding “comprising” or “including” encompass the elements or objects listed following “comprising” or “including” and their equivalents, and do not exclude other elements or objects. The “multiple” mentioned in the embodiments of this application refers to two or more. A and / or B indicate three possibilities: A; B; and A and B.

[0130] The above description is merely an exemplary embodiment of this application, but the scope of protection of this application is not limited thereto. Any person skilled in the art can easily conceive of various equivalent modifications or substitutions within the technical scope disclosed in this application, and such modifications or substitutions should all be covered within the scope of protection of this application. Therefore, the scope of protection of this application should be determined by the scope of the claims.

Claims

1. An adaptive control method for ultrasonic welding power based on inductive temperature rise feedback, specifically including: S1: Collect data from ultrasonic welding equipment to generate the original temperature rise time series dataset; S2: Based on the original temperature rise time series dataset, construct the third-order thermal inertia model of the inductor, and calculate the equivalent heat capacity and thermal resistance of the surface fiber coating, the thermal diffusion time constant of the intermediate filament winding structure, and the overall winding thermal mass-dominated time constant to generate the parameter set of the third-order thermal inertia model of the inductor. S3: During the welding operation phase, the inductor surface temperature is collected as the main input signal and input into the inductor thermal inertia identification module constructed based on the parameter set of the third-order thermal inertia model of the inductor to generate a series of predicted temperature rise trends within multiple key time domain windows. S4: Perform difference calculation between the current measured temperature and the predicted temperature rise trend sequence to extract the transient disturbance residual signal that cannot be covered by the third-order thermal inertia model of inductance, and input it into the PI feedback controller to generate closed-loop correction component. S5: Based on the slope change rate of the temperature rise trend prediction value sequence, identify the state characteristics of the temperature rise from a gradual transition to an inflection point or entering a plateau region, trigger the feedforward channel to generate a power gradient pre-adjustment action command, and delay the power gradient pre-adjustment action command to the next zero voltage crossing interval when the vibration phase deviation is detected to exceed the limit. S6: Determine whether the absolute value of the transient disturbance residual signal exceeds the set threshold. If it does, integrate and accumulate the closed-loop correction component; otherwise, only perform proportional amplification and generate a feedback correction command. S7: Vector synthesize the power gradient pre-adjustment action command and the feedback correction command to generate a composite power adjustment command.

2. The ultrasonic welding power adaptive control method based on inductive temperature rise feedback according to claim 1, characterized in that, The process of collecting data from the ultrasonic welding equipment to generate the original temperature rise time series dataset involves, specifically, during the no-load start-up phase of the ultrasonic welding equipment, executing a stepped power excitation sequence and simultaneously collecting the time response curves of multiple temperature sensors arranged on the surface of the wire winding, the inner wall of the skeleton, and the interface of the impregnated varnish curing layer. This yields the original temperature rise time series dataset, which includes the thermal conductivity of the surface fiber varnish film, the thermal diffusion of the intermediate wire bundle winding structure, and the dominant characteristics of the overall winding thermal mass.

3. The ultrasonic welding power adaptive control method based on inductive temperature rise feedback according to claim 1, characterized in that, The multiple key time domain windows are three time domain windows: the next 50 milliseconds, 100 milliseconds, and 200 milliseconds.

4. The ultrasonic welding power adaptive control method based on inductive temperature rise feedback according to claim 1, characterized in that, The PI feedback controller includes a mode switching discrimination unit, a proportional path, an integral adjustment path, and an output synthesis and constraint module.

5. The ultrasonic welding power adaptive control method based on inductive temperature rise feedback according to claim 1, characterized in that, The composite power regulation command has dual characteristics of feedforward prediction and feedback correction.

6. The ultrasonic welding power adaptive control method based on inductive temperature rise feedback according to claim 1, characterized in that, Following S7, the following also includes: S8: After each welding, compare the cumulative deviation pattern between the actual temperature rise trajectory and the predicted temperature rise trend sequence. If the deviations show the same direction for three consecutive times, start the parameter update process and only incrementally correct the surface thermal resistance parameter in the equivalent heat capacity and thermal resistance of the surface fiber coating to complete the online fine-tuning of the model.

7. The ultrasonic welding power adaptive control method based on inductive temperature rise feedback according to claim 1, characterized in that, S3 specifically includes: The analog voltage signal output by the temperature sensor arranged on the surface of the wire winding is sampled at high frequency and digitally filtered to obtain the real-time raw data of the inductor surface temperature after removing noise interference, and the real-time raw data of the inductor surface temperature is converted into a standardized digital temperature measurement value. Based on the digital temperature measurement and the parameter set of the third-order thermal inertia model of the inductor generated in the previous steps, the inductor thermal inertia identification module embedded in the main control unit is called to perform the initialization mapping operation of the state space equation in order to construct the system state vector at the current moment. Using the current system state vector as the iterative benchmark, the third-order heat conduction differential equations are extrapolated in multiple steps to solve the discretized temperature rise state components corresponding to three specific time nodes in the future: 50 milliseconds, 100 milliseconds, and 200 milliseconds. Discretized temperature rise state components are subjected to time-series recombination and vector encapsulation to generate a sequence of predicted temperature rise trends for future multi-time-domain windows with clear timestamps. This sequence fully characterizes the expected thermal dynamic trajectory of the inductor under the coupling effect of fast and slow components. Based on the predicted value sequence of temperature rise trend in future multi-time-domain windows, slope differentiation is performed to extract the characteristic parameter of temperature rise rate of change in each prediction time window, and the characteristic parameter of temperature rise rate of change is output to the feedforward channel to trigger the subsequent power gradient pre-adjustment action command generation logic.

8. The ultrasonic welding power adaptive control method based on inductive temperature rise feedback according to claim 7, characterized in that, The current system state vector includes the equivalent heat capacity of the surface fiber coating, the thermal diffusion time constant of the intermediate filament winding structure, and the overall winding thermal mass-dominant time constant.

9. The ultrasonic welding power adaptive control method based on inductive temperature rise feedback according to claim 1, characterized in that, S4 specifically includes: Based on the temperature rise trend prediction sequence derived from the parameter set of the third-order thermal inertia model of the inductor and the real-time acquired main input signal of the inductor surface temperature, timestamp alignment processing is performed to obtain a set of measured temperature data and a set of predicted temperature data with the same timestamp reference, so as to provide consistent time domain input conditions for subsequent difference calculation. Using the measured temperature data set and the predicted temperature data set, point-by-point algebraic difference operation is performed to extract the transient disturbance residual signal sequence that characterizes the dynamic characteristics not covered by the third-order thermal inertia model of the inductor. This sequence quantifies the temperature response deviation caused by the actual welding load change. Based on the amplitude and rate of change characteristics of the transient disturbance residual signal sequence, residual signal validity judgment and noise filtering are performed to generate pre-processed pure residual control variables and eliminate the influence of sensor high-frequency noise on the false triggering of the feedback loop. Based on the pure residual control variable, the PI feedback controller is called to perform proportional gain amplification and integral accumulation calculation to generate an initial closed-loop correction control quantity containing fast response components and steady-state elimination components, thereby achieving preliminary compensation for transient disturbances. The initial closed-loop correction control quantity is subjected to output limiting and smoothing filtering to generate the final closed-loop correction component used for power command synthesis.

10. The adaptive control method for ultrasonic welding power based on inductive temperature rise feedback according to claim 1, characterized in that, S5 specifically includes: The temperature rise trend prediction value sequence is processed by first-order difference operation to obtain the temperature rise slope feature vector that represents the rate of temperature change within the future time window, which serves as the quantitative basis for subsequent thermal state discrimination. Based on the temperature rise slope feature vector, identify the critical moment when the temperature rise transitions from the flat zone to the rising inflection point or the steady-state characteristics when it enters the plateau zone, and generate thermal state switching event markers containing timestamps and state types. The preset power gradient mapping table is called according to the thermal state switching event flag to perform a lookup operation, so as to generate an initial power gradient pre-adjustment action command corresponding to the current thermal dynamic characteristics, and realize feedforward compensation control for the predicted temperature rise trend. The zero-crossing point real-time tracking processing of the ultrasonic transducer drive voltage waveform is performed to extract the zero-voltage crossing interval boundary parameters of the current vibration cycle, and synchronize them with the system clock to generate a phase-locked reference window; The initial power gradient pre-adjustment action command is compared and verified with the phase-locked reference window. If the command trigger time is detected to be outside the zero-voltage crossing range, a delay suspension logic is executed to generate the final composite power adjustment command that is executed within the safe phase window.

Images