A needle mark adaptive control method and system based on time-series AI prediction and motion control card

Abstract

This invention relates to the field of data acquisition and control technology, and discloses a method and system for adaptive probe mark control based on time-series AI prediction and a motion control card. The method includes: acquiring periodic interrupt signals generated by a hardware timer of the motion control card and forming an original synchronous data frame; performing cubic spline interpolation and motion trend compensation on the original synchronous data frame to generate a unified, equally spaced fused data stream; performing central difference, sliding window cross-correlation, and recursive least squares identification on the fused data stream; outputting a control vector through a lightweight time-series model; completing hardware latch triggering, target position algebraic synthesis, and end-point trajectory interpolation based on the control vector; and outputting an adaptive control command for the probe end. Because this application achieves adaptive correction of the probe end docking position through hardware synchronous acquisition, timestamp alignment, online feature identification, and feedforward predictive control, it improves the consistency of probe marks and the stability of contact control in flying probe testing.

Description

Technical Field

[0001] This invention relates to the field of data acquisition and control technology, and in particular to a needle mark adaptive control method and system based on time-series AI prediction and motion control card. Background Technology

[0002] Currently, in flying probe testing equipment During shaft pressing, to ensure reliable contact between the probe and the pad without damaging the pad, a motion control card typically simultaneously acquires contact status signals from a laser displacement sensor, stepper motor current, and encoder position. A common approach is to use a control system built with a PLC or basic motion control card, sequentially reading values ​​from multiple sensors at a fixed inspection cycle and controlling shutdown based on displacement or current thresholds. In this type of solution, the laser displacement sensor and motor encoder each rely on independent internal clocks to trigger sampling, and data frames are only timestamped at the application layer, lacking hardware-level time-locked loop (TLL) and synchronization triggering mechanisms.

[0003] For example, since the laser sampling rate is usually fixed at 1kHz, while the update frequency of the motor encoder and the current analog-to-digital conversion delay are different, the system cannot guarantee strict alignment of the instantaneous correspondence of multimodal signals on a microsecond scale, resulting in timestamp misalignment when combining multi-source data. At the moment of dynamic contact when the probe approaches the PCB surface at a speed of tens of millimeters per second, the laser displacement sensor can detect displacement changes of tens of micrometers every 1ms, while the motor encoder position update frequency may deviate from 1kHz. In addition, the analog-to-digital conversion and interrupt response have jitter of several to tens of microseconds. Therefore, the displacement, position, and current latched within the same interrupt cycle actually correspond to different real physical moments. When this raw data with timestamp deviations is directly used to calculate derived features such as spring compression speed and acceleration, pseudo-dynamic components are introduced, distorting the input features received by the AI ​​model. Especially in high-frequency sampling, even a small timing misalignment can cause the model's prediction of the contact point to deviate from reality, resulting in overshoot or insufficient thrust, ultimately causing the needle mark depth to exceed the allowable range of 5–20 micrometers, or even piercing the solder pad.

[0004] Therefore, there is an urgent need for a method that can still achieve microsecond-level multimodal timing data synchronization and alignment through hardware triggering and timestamp interpolation, even when the sampling frequencies and delay characteristics of the laser displacement sensor and the stepper motor encoder are inconsistent. This would eliminate the dynamic contact force prediction deviation caused by data misalignment, ensure the timing consistency of the input features of the timing AI model, and thus accurately predict contact behavior and adaptively control the probe docking position, thereby improving the consistency of probe marks and the reliability of the flying probe test process. Summary of the Invention

[0005] To address the aforementioned technical shortcomings, the purpose of this invention is to propose a needle mark adaptive control method based on time-series AI prediction and motion control card. This method aims to solve the technical problem in existing technologies that rely on a single sensor and lack a strict synchronization mechanism, resulting in timestamp misalignment after combining multi-source data and distortion of AI model input features. In particular, under the condition that the sampling frequency and delay characteristics of the laser displacement sensor and the stepper motor encoder are inconsistent, it is impossible to achieve microsecond-level synchronous alignment of multimodal time-series data to avoid dynamic contact force prediction deviation.

[0006] To solve the above technical problems, the present invention adopts the following technical solution: The present invention provides a needle mark adaptive control method based on time-series AI prediction and motion control card.

[0007] The aforementioned needle mark adaptive control method based on time-series AI prediction and motion control card includes: Step S10: Obtain the periodic interrupt signal generated by the motion control card hardware timer, and execute the multi-source sensor data acquisition task based on the periodic interrupt signal using clock phase-locked loop and synchronous trigger latching method, and output the original synchronous data frame; Step S20: Based on the original synchronized data frame, a multimodal timestamp alignment task is performed using a cubic spline interpolation algorithm and a motion trend compensation method, outputting a unified fused data stream containing laser displacement value sequence, motor position value sequence, and motor current value sequence; Step S30: Based on the unified and equally spaced fused data stream, a time-series dynamic preprocessing task is performed using central difference, sliding window cross-correlation, and recursive least squares identification methods to output a multi-dimensional feature vector with physical coupling relationship; Step S40: Based on the multi-dimensional feature vectors with physical coupling relationships, a lightweight time-series model feedforward inference method is used to perform a multi-objective control quantity prediction task and output a control vector; Step S50: Based on the control vector, perform hardware latch triggering, target position algebraic synthesis and end trajectory interpolation, and output probe end adaptive control command.

[0008] Preferably, step S10, which involves acquiring the periodic interrupt signal generated by the motion control card hardware timer, executing the multi-source sensor data acquisition task based on the periodic interrupt signal using a clock-locked loop and synchronous trigger latching method, and outputting the original synchronous data frame, specifically includes: Step S101: The system clock with a main clock frequency of 100MHz inside the motion control card generates a hardware interrupt signal with a basic period of 10μs, and the motion control card sends a periodic synchronization pulse signal with a frequency of 1kHz to the laser displacement sensor, and uses the periodic synchronization pulse signal as the external time reference of the laser displacement sensor. Step S102: The digital phase-locked loop module inside the laser displacement sensor uses the periodic synchronization pulse signal as a reference to adjust the frequency division coefficient of the 40MHz voltage-controlled oscillator inside the laser displacement sensor in real time, so that the sampling time of the laser displacement sensor and the interruption time of the hardware interrupt signal remain in phase-locked for an entire cycle. Step S103: When a hardware interrupt signal arrives every 10μs after phase locking, the hardware synchronously triggers the laser displacement sensor to collect laser displacement data, and triggers the motion control card hardware logic to latch the absolute encoder position value of the motor and the motor current value after analog-to-digital conversion. A unified timestamp is generated by the field programmable gate array, and the laser displacement data, the absolute encoder position value of the motor, the motor current value and the unified timestamp are packaged into an original synchronous data frame and written into the high-speed first-in-first-out buffer.

[0009] Preferably, step S20, which involves performing a multimodal timestamp alignment task based on the original synchronized data frame using a cubic spline interpolation algorithm and a motion trend compensation method, and outputting a unified, equally spaced fused data stream containing laser displacement value sequences, motor position value sequences, and motor current value sequences, specifically includes: Step S201: Take out a batch of original synchronization data frames from the high-speed first-in-first-out buffer used to buffer the original synchronization data frames, and read the original timestamp sequence, laser displacement data sequence, motor absolute encoder position value sequence and motor current value sequence from the original synchronization data frames; Step S202: Using the preset AI model main inference cycle as the uniform resampling interval, find adjacent true sampling points in the original timestamp sequence that satisfy the following relationship for each uniform time point:

[0010] in, For the first A uniform time point, These are the adjacent real sampling points located before the uniform time point. These are the adjacent real sampling points located after the uniform time point; Step S203: Calculate the laser displacement value corresponding to the uniform time point using a cubic spline interpolation function on the laser displacement data sequence, and calculate the motor position value corresponding to the uniform time point using linear interpolation and motion trend compensation term on the motor absolute encoder position value sequence, to obtain a unified fused data stream containing the laser displacement value sequence, the motor position value sequence, and the motor current value sequence; The laser displacement value is calculated according to the following formula:

[0011] in, For the first The laser displacement value corresponding to a uniform time point For real sampling points Laser displacement data at the location, , and They are intervals cubic spline interpolation coefficients; The motor position value is calculated according to the following formula:

[0012] in, For the first The motor position value corresponding to a uniform time point. For real sampling points The absolute encoder position value of the motor at that location. For real sampling points The absolute encoder position value of the motor at that location. For the first Motion trend compensation term corresponding to a uniform time point.

[0013] Preferably, in step S202, the preset AI model main inference cycle is 200μs; within each preset AI model main inference cycle, the laser displacement data sequence and the motor absolute encoder position value sequence are resampled according to the adjacent real sampling points, so that adjacent data points in the unified and equally spaced fused data stream have the same time interval.

[0014] Preferably, step S30, which involves performing a time-series dynamic preprocessing task based on the unified, equally spaced fused data stream using central difference, sliding window cross-correlation, and recursive least squares identification methods to output a multi-dimensional feature vector with physical coupling relationships, specifically includes: Step S301: Perform a central difference operation on the laser displacement value sequence in the unified and equally spaced fused data stream to obtain the spring compression velocity and spring compression acceleration determined by the laser displacement value; and perform a central difference operation on the motor position value sequence in the unified and equally spaced fused data stream to obtain the instantaneous speed of the motor. Step S302: Within the sliding window of the preset window width corresponding to the current uniform time point, extract the first-order difference sequence of laser displacement and the first-order difference sequence of equivalent displacement obtained by converting the motor position value. Calculate the normalized cross-correlation function between the first-order difference sequence of laser displacement and the first-order difference sequence of equivalent displacement. Multiply the offset corresponding to the maximum value of the normalized cross-correlation function by the sampling interval of the uniformly spaced fused data stream to obtain the dynamic hysteresis compensation feature. Step S303: After a contact event is detected based on the laser displacement value sequence and the motor current value sequence, obtain the laser displacement increment sequence and motor current increment sequence corresponding to multiple samples after the contact event. Use the laser displacement increment sequence as the input variable and the motor current increment sequence as the output variable. Use the recursive least squares method with a forgetting factor to identify the online stiffness estimate and bias term in the contact stiffness parameter model online. The online stiffness estimate is used to characterize the slope relationship between the motor current increment and the laser displacement increment, and the bias term is used to characterize the static current offset after the contact event. Step S304: Combine the laser displacement value, the spring compression velocity, the spring compression acceleration, the motor position value, the motor instantaneous velocity, the motor current value, the dynamic hysteresis compensation feature, and the online stiffness estimation value according to a preset feature arrangement order to obtain a multidimensional feature vector with physical coupling relationship.

[0015] Preferably, step S40, which involves performing a multi-objective control quantity prediction task based on the physically coupled multi-dimensional feature vectors using a lightweight time-series model feedforward inference and outputting control vectors, specifically includes: Step S401: The multi-dimensional feature vectors with a number of consecutive preset input time steps before the current uniform time point are combined into a feature vector time sequence, and the feature vector time sequence is input into a lightweight time sequence model deployed in an embedded AI processor. The lightweight time sequence model is a long short-term memory network or a converter network, and is processed by INT8 quantization and operator fusion. Step S402: Within each preset AI model main inference cycle, the lightweight temporal model performs a single feedforward inference on the feature vector temporal sequence and outputs a control vector; The control vector includes a dynamic target compression increment, an adaptive proportional gain, an adaptive integral gain, a predicted overshoot compensation, a cumulative contact probability, and an anomaly flag. The dynamic target compression increment is used to characterize the target displacement compensation amount that the probe tip needs to continue pressing down during the current contact stage. The adaptive proportional gain and the adaptive integral gain are used to refresh the closed-loop control parameters of the servo controller. The predicted overshoot compensation amount is used to characterize the excessive pressing displacement that needs to be pre-compensated before the end stops. The cumulative contact probability is used to characterize the contact confidence obtained by accumulating the contact states at multiple consecutive uniform time points. The anomaly flag is used to characterize whether a smooth rapid retreat trajectory is triggered.

[0016] Preferably, step S50, which involves performing hardware latch triggering, target position algebraic synthesis, and end-point trajectory interpolation based on the control vector, and outputting an adaptive control command for the probe end, specifically includes: Step S501: Compare the cumulative contact probability in the control vector with the preset micro-touch warning threshold. When the cumulative contact probability is greater than the preset micro-touch warning threshold, the embedded AI processor sends a latch trigger signal to the motion control card through the high-speed digital input / output interface, and the motion control card hardware logic latches the current absolute position of the motor to obtain the latch position. Step S502: Perform algebraic synthesis of the latch position, the dynamic target compression increment in the control vector, and the predicted overshoot compensation amount in the control vector to obtain the final target position, and generate an S-shaped smooth deceleration curve interpolation command with the final target position as the endpoint. The final target location is calculated according to the following formula:

[0017] in, For the final target location, For the latch position, The dynamic target compression increment in the control vector is used. This is the predicted overshoot compensation amount with direction sign in the control vector; Step S503: Refresh the adaptive proportional gain and adaptive integral gain in the control vector to the servo controller, and execute the end-loop docking according to the S-shaped smooth deceleration curve interpolation command; when the anomaly flag is true, stop executing the end-loop docking and output the smooth rapid retreat trajectory command; when the anomaly flag is false and the position steady-state error converges to the preset error range, output the probe end adaptive control command.

[0018] This invention also provides a needle mark adaptive control system based on time-series AI prediction and motion control card, comprising: The synchronous acquisition module is used to acquire periodic interrupt signals generated by the hardware timer of the motion control card, and to perform multi-source sensor data acquisition tasks based on the periodic interrupt signals using a clock phase-locked loop and synchronous trigger latching method, and output raw synchronous data frames. The timestamp alignment module is used to perform a multimodal timestamp alignment task based on the original synchronized data frame using a cubic spline interpolation algorithm and a motion trend compensation method, and outputs a unified fused data stream containing laser displacement value sequence, motor position value sequence and motor current value sequence; The temporal dynamic preprocessing module is used to perform temporal dynamic preprocessing tasks based on the unified equally spaced fused data stream using central difference, sliding window cross-correlation and recursive least squares identification methods, and outputs a multi-dimensional feature vector with physical coupling relationship. The feedforward inference module is used to perform multi-objective control quantity prediction tasks based on the multi-dimensional feature vectors with physical coupling relationships using a lightweight time-series model feedforward inference method, and output control vectors. The docking control module is used to perform hardware latching triggering, target position algebraic synthesis and end trajectory interpolation according to the control vector, and output probe end adaptive control commands.

[0019] The present invention also provides a pin mark adaptive control device based on a timing AI prediction and motion control card. The pin mark adaptive control device based on a timing AI prediction and motion control card includes: a memory, a processor, and a pin mark adaptive control program based on a timing AI prediction and motion control card stored in the memory and executable on the processor. When the pin mark adaptive control program based on a timing AI prediction and motion control card is executed by the processor, it implements the above-mentioned method.

[0020] The present invention also provides a computer program product, the computer program product including a pin mark adaptive control program based on a timing AI prediction and motion control card, the pin mark adaptive control program based on a timing AI prediction and motion control card implementing the above method when executed by a processor.

[0021] The beneficial effects of this invention are as follows: This invention generates periodic interrupt signals through a motion control card hardware timer and combines clock phase-locked loop and synchronous trigger latching to package laser displacement data, motor absolute encoder position values, and motor current values ​​into a raw synchronous data frame with a unified timestamp. This reduces the timestamp misalignment problem caused by multi-source data acquisition at the moment of dynamic contact. At the same time, a unified and equally spaced fused data stream is generated through a cubic spline interpolation algorithm and motion trend compensation method, so that the laser displacement value sequence, motor position value sequence, and motor current value sequence participate in subsequent processing under the same time reference. This improves the timing consistency of feature inputs and the accuracy of dynamic contact state representation during probe pressing.

[0022] This invention obtains a multidimensional feature vector containing dynamic hysteresis compensation features and online stiffness estimates by performing central difference, sliding window cross-correlation, and recursive least squares identification on a unified and equally spaced fused data stream. It then outputs dynamic target compression increment, adaptive gain, predicted overshoot compensation, cumulative contact probability, and anomaly flags through a lightweight time-series model. This enables the motion control card to complete hardware latch triggering, target position algebraic synthesis, and end-point trajectory interpolation based on the control vector, reducing the risk of overshoot, insufficient compression, or abnormal contact caused by fixed down-pressure parameters. This improves the consistency of probe marks and the stability of end-point docking control during flying probe testing. Attached Figure Description

[0023] Figure 1This is a flowchart illustrating the first embodiment of a needle mark adaptive control method based on time-series AI prediction and motion control card according to the present invention.

[0024] Figure 2 This is a schematic diagram of the application process of flying probe testing on printed circuit boards according to the first embodiment of the pin mark adaptive control method based on timing AI prediction and motion control card of the present invention. Detailed Implementation

[0025] The technical solutions of the present invention will be clearly and completely described below with reference to the accompanying drawings of the embodiments of the present invention. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments.

[0026] Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0027] Example 1: As Figure 1 The diagram shown is a flowchart of the first embodiment of a needle mark adaptive control method based on time-series AI prediction and motion control card according to the present invention.

[0028] In the first embodiment, the needle mark adaptive control method based on time-series AI prediction and motion control card includes: Step S10: Obtain the periodic interrupt signal generated by the motion control card hardware timer, and execute the multi-source sensor data acquisition task based on the periodic interrupt signal using clock phase-locked loop and synchronous trigger latching method, and output the original synchronous data frame; The periodic interrupt signal generated by the motion control card's hardware timer refers to the hardware clock signal generated by frequency division of the motion control card's internal system clock and used to drive sampling, latching, and timestamp generation. Clock phase-locked loop (PLL) does not simply record the sampling time at the software layer; instead, the motion control card sends periodic synchronization pulses to the laser displacement sensor, causing the laser displacement sensor's internal sampling clock to adjust according to the motion control card's hardware clock reference. Synchronous trigger latching, upon the arrival of the same interrupt, synchronously latches the laser displacement data, the motor absolute encoder position value, and the motor current value after analog-to-digital conversion, assigning a unified timestamp via the field-programmable gate array (FPGA). The resulting raw synchronization data frame is a snapshot of data describing the probe end displacement, drive motor position, and current state using the same hardware time reference, and is written to a high-speed FIFO buffer for subsequent reading.

[0029] The purpose of this step is to move the asynchronous sampling process of multi-source sensors forward to the hardware layer. Laser displacement, motor position, and motor current reflect the geometric changes, actuator motion state, and drive load changes during probe contact, respectively. Only when these three are collected at the same physical moment can subsequent timestamp alignment and feature construction have a true causal basis. By using periodic interruption, sensor phase-locking, and hardware latching to jointly limit the sampling time, the time correspondence between channels within the original synchronous data frame is fixed. When performing cubic spline interpolation and motion trend compensation in step S20, it deals with a data sequence that already has a unified time label, rather than a loose record that requires guessing the sampling order afterward.

[0030] In traditional flying probe testing control, multiple signals are often read sequentially via software polling or ordinary data acquisition cards, lacking a unified clock constraint between the laser displacement sensor, motor encoder, and current acquisition circuit. When the probe is rapidly pressed down, a difference of tens of microseconds in readings can correspond to actual displacement changes of several micrometers, meaning that the displacement, current, and motor position in the same data frame do not originate from the same contact stage. After adopting this step, time drift and channel reading delay are no longer considered uncertain errors in subsequent calculations. The sources of derived quantities such as spring compression speed, acceleration, and contact stiffness are clearer, and the input received by the subsequent lightweight timing model is closer to the actual contact process.

[0031] For example, performing tests using flying probe testing equipment Taking the process of shaft pressing down as an example, inside the motion control card The clock source generates a frequency of The system experiences periodic interruptions, and simultaneously outputs data to the laser displacement sensor via the I / O port. The synchronization pulse sequence. After the digital phase-locked loop enters the locked state, the fifth synchronization pulse after the rising edge of a certain synchronization pulse... When an interrupt occurs, the hardware synchronous trigger circuit activates simultaneously: the laser displacement sensor, upon being triggered, immediately completes a distance measurement, and the reading is... The motion control card's internal logic freezes the current absolute encoder position of the motor. Counting points and latching the motor current value to the analog-to-digital converter. Retrieve them all at once. The FPGA timestamps these three data points. These are combined to form a raw synchronization data frame and pushed into the FIFO. Subsequent interrupts are triggered sequentially, with the timestamp interval of each frame strictly defined as follows: In continuous During its runtime, without phase-locked loop calibration, even a small frequency difference in the internal crystal oscillator of the laser sensor can cause its sampling time to drift cumulatively relative to the reference interrupt for more than [percentage missing]. However, through continuous tracking and correction using a phase-locked loop, this deviation is controlled within... Within this range. In the continuously synchronized data frame sequence obtained in this way, the displacement, position, and current within each frame are all sampled values ​​at the same instant.

[0032] Step S20: Based on the original synchronized data frame, a multimodal timestamp alignment task is performed using a cubic spline interpolation algorithm and a motion trend compensation method, outputting a unified fused data stream containing laser displacement value sequence, motor position value sequence, and motor current value sequence; Multimodal timestamp alignment refers to reconstructing discrete sampling points that do not fall entirely on the main inference cycle to uniform time points with fixed time intervals, while preserving the unified timestamp meaning of the original synchronous data frames. For laser displacement data, a cubic spline interpolation algorithm is used to recover the displacement value at the target time using adjacent sampling points and the smoothness constraint of the entire curve. For motor position data, motion trend compensation is added in addition to the linear position change trend to reflect the continuous change of the actual position of the motor during acceleration or deceleration. After processing, each uniform time point in the fused data stream simultaneously has laser displacement value, motor position value, and motor current value, and the time interval between adjacent data points is consistent.

[0033] This step converts the original synchronous data frames into regular time series that can be directly used for temporal feature extraction and model inference. Laser displacement changes rapidly at the moment of contact; cubic spline interpolation ensures a smooth transition of the displacement curve between sampling points, reducing sawtooth noise when subsequently calculating velocity and acceleration using central difference. Motor position is affected by changes in mechanical inertia and control commands; motion trend compensation reduces the lag in acceleration and deceleration sections of simple interpolation. The resulting unified, equally spaced fused data stream provides a regular, continuous, and physically consistent data foundation for step S30 to calculate dynamic hysteresis compensation features, online stiffness estimates, and multidimensional feature vectors.

[0034] Traditional alignment methods often employ nearest-neighbor sampling or simple linear interpolation. While convenient to implement, these methods can easily transform sampling interval errors into spurious velocity and acceleration fluctuations in highly dynamic scenarios like probe contact. Nearest-neighbor sampling results in step-like displacement changes, linear interpolation exhibits discontinuous rates of change at nodes, and simply processing motor position linearly fails to reflect non-uniform motion. In contrast, this step employs different processing logics for laser displacement and motor position, maintaining the continuity of the laser displacement curve while compensating for the motion trend of the motor position. This ensures that subsequent sliding window cross-correlation does not misinterpret interpolation errors as genuine transmission lag.

[0035] For example, suppose in one inference cycle Internal demand generation and The fused data at different times. The timestamp sequence of the original synchronized data frames is... microseconds. When calculating At that time, find adjacent sampling points in the sequence. and For laser displacement signals, the cubic spline coefficients obtained by solving the entire curve are used in... Calculate on interval The function value at that point is obtained. This value smoothly connects the discrete points before and after the given value and maintains the continuity of the first derivative. For motor position signals, The time encoder value is count, Time for Counting, if the average speed trend calculated in the previous few cycles is each Increase Counting points, at Time-based motion compensation items Approximately Rounding to the nearest whole number A count. Thus, in A set of time-aligned displacement and position values ​​was obtained. After performing similar processing on all target time points, the time interval was obtained to be strictly equal to... An equally spaced sequence.

[0036] Step S30: Based on the unified and equally spaced fused data stream, a time-series dynamic preprocessing task is performed using central difference, sliding window cross-correlation, and recursive least squares identification methods to output a multi-dimensional feature vector with physical coupling relationship; Step S30, the time-series dynamic preprocessing, involves extracting kinematics and contact states from the fused data stream output in step S20. Central difference is used to obtain the spring compression velocity and spring compression acceleration from the laser displacement value sequence, and the instantaneous motor velocity from the motor position value sequence. Sliding window cross-correlation is used to compare the sequential relationship between laser displacement changes and equivalent motor displacement changes within a local time window, obtaining dynamic hysteresis compensation features. Recursive least squares identification, after detecting a contact event, uses the correspondence between laser displacement increments and motor current increments to update contact stiffness parameters and static current offset online. The resulting multidimensional feature vector includes both the current measurement value and derived quantities such as velocity, acceleration, hysteresis, and stiffness. The sliding window width is preferably 10-30 uniform sampling points. Under the condition of a sampling frequency of 5kHz-10kHz, this width strictly corresponds to a physical time window of 2ms-6ms. This window size has been verified by mechanical simulation and precisely covers the key temporal feature region from the probe's contact precursor to the sudden change in stiffness, avoiding redundant calculations.

[0037] This step enables subsequent feedforward inference to obtain not only instantaneous amplitudes of displacement and current, but also a state description of the physical structure of the contact process. Spring compression velocity and acceleration reflect the dynamic changes as the probe approaches the pad, motor instantaneous speed reflects the actuator's motion trend, dynamic hysteresis compensation features reveal the time offset caused by transmission backlash, elastic deformation, or response delay on displacement transmission, and online stiffness estimates characterize the hardness or softness of the measured pad or carrier plate during the current contact process. After this information is organized into a multi-dimensional feature vector with physical coupling relationships, step S40 can determine the contact state evolution within a shorter input sequence, without relying solely on the network to infer all physical relationships from the raw values.

[0038] Traditional flying probe testing control systems often extract only instantaneous values ​​of displacement and current, or simple statistical quantities such as mean and variance, as the basis for judgment. The feature set is static and does not include time-varying dynamic hysteresis and online stiffness information. Hysteresis parameters usually require offline calibration, but transmission clearance gradually drifts with mechanical wear and changes in preload, and fixed calibration values ​​become inaccurate after long-term equipment operation. Contact stiffness is also related to the plating material, thickness, and surface condition of the tested pads; using fixed values ​​cannot adapt to various types of flexibility testing. This invention calculates hysteresis characteristics online using a sliding window cross-correlation method, eliminating the need for manual calibration and adaptively following changes in mechanical state using real-time data. The online stiffness identification using recursive least squares with a forgetting factor updates the stiffness coefficient in the form of a continuous data stream after a contact event, smoothing out measurement noise in current and displacement, and the forgetting factor gradually attenuates the influence of old data, tracking slow stiffness drift. The introduction of these online identification mechanisms allows the feature vector to accurately reflect the system's instantaneous characteristics under different test objects and mechanical states. For example, when the probe contacts the pads of a flexible circuit board, the laser displacement may rise by 0.02 mm within 0.3 ms, while the motor position has not yet changed by the same amplitude due to the transmission gap. In step S30, the first-order difference sequence of the laser displacement and the first-order difference sequence of the equivalent motor displacement are compared within a sliding window of 300 sampling points and corresponding to 6 ms. If the cross-correlation peak corresponds to 12 sampling intervals and the sampling interval of the fused data stream is 200 μs, a dynamic hysteresis compensation feature of approximately 240 μs can be obtained. After the contact event occurs, the recursive least squares identification continuously updates the stiffness using the most recent sets of current increments and displacement increments. If the current estimate is lower than that in the hard pad scenario, the lightweight timing model can determine that the contact object is more flexible and output a more conservative compression amount and overshoot compensation in step S40.

[0039] Step S40: Based on the multi-dimensional feature vectors with physical coupling relationships, a lightweight time-series model feedforward inference method is used to perform a multi-objective control quantity prediction task and output a control vector; Step S40 receives the multi-dimensional feature vectors continuously output from step S30 and assembles them into a feature vector time series sequence according to the preset input time steps prior to the current uniform time point. The lightweight time series model is deployed in an embedded AI processor and can employ a long short-term memory network or a converter network, reducing inference time through INT8 quantization and operator fusion. Within each main inference cycle, the model performs feedforward inference once on the feature vector time series sequence and outputs a control vector. This control vector includes dynamic target compression increment, adaptive proportional gain, adaptive integral gain, predicted overshoot compensation, cumulative contact probability, and anomaly flag. The generation of the predicted overshoot compensation also incorporates physical prior factors such as online stiffness estimation, instantaneous contact velocity, equivalent mass of moving parts, stiffness of elastic elements, and changes in motor current within the near-time window. The preset feature arrangement order is specifically manifested as follows: the extracted laser displacement time series features, spring first-order compression velocity features, motor active current abrupt change features, and online physical stiffness estimates are concatenated and spliced ​​in the form of a one-dimensional time series tensor to form a multimodal fusion feature vector, which serves as the standard input format for the time series network.

[0040] This step unifies contact state identification, compression amount planning, gain adjustment, and overshoot suppression into a single feedforward inference. The dynamic target compression increment indicates the target displacement compensation that the probe still needs to apply pressure to during the current contact phase. Adaptive proportional gain and adaptive integral gain refresh the closed-loop control parameters of the servo controller. The predicted overshoot compensation amount preemptively offsets potential excessive pressure before the probe stops at the end of the circuit. The cumulative contact probability provides a soft decision basis for whether to trigger hardware latching in step S50. The anomaly flag reserves a safety path for the emergency retreat trajectory. Because these quantities are generated simultaneously from the same feature sequence and the same model, they maintain consistent contact state judgments, making it less likely for compression amounts, gains, and compensation directions to contradict each other.

[0041] In traditional flying probe test control systems, the target compression amount and PID gain are usually fixed parameters calibrated offline, or simply switched according to a few preset sets of parameters. When the stiffness of the contact object is unknown or softens, a fixed, large gain can easily lead to overshoot, while a fixed, small gain will slow down the compression process. Furthermore, the preset compression amount often leaves a margin for the worst-case scenario, making it impossible to optimize the probe depth for each weld joint. Simultaneously, overshoot suppression largely relies on the feedback controller detecting displacement exceeding a threshold before retracting; this pure feedback method has an inherent time lag, making it difficult to avoid minor over-compression. In the method proposed in this invention, the time-series model adaptively predicts the dynamic compression increment based on current and historical feature sequences. and variable gain , This ensures that the control parameters match the physical state of the contact; overshoot compensation amount As a feedforward quantity, it is directly superimposed into the target position synthesis to pre-cancele the expected overshoot; the stiffness used in its calculation... and contact speed All data were obtained through online identification or prediction, ensuring the targeted nature of the feedforward compensation. After INT8 quantization and operator fusion, the model can be run on an embedded processor. Its stable cycle operation enables deep embedding of AI computing and real-time motion control, which is difficult to achieve with traditional PC-based or PLC solutions.

[0042] Step S50: Based on the control vector, perform hardware latch triggering, target position algebraic synthesis and end trajectory interpolation, and output probe end adaptive control command.

[0043] Step S50 converts the control vector output in step S40 into an end-effector control action that the motion control card can execute. First, the cumulative contact probability is compared with a preset micro-tactile warning threshold. When the cumulative contact probability exceeds the threshold, the embedded AI processor sends a latch trigger signal to the motion control card via a high-speed digital input / output interface. The motion control card hardware logic latches the current absolute position of the motor, forming a reference position for subsequent target synthesis. Subsequently, the motion control card algebraically synthesizes the latched position, dynamic target compression increment, and predicted overshoot compensation to obtain the final target position, and uses this final target position as the endpoint to generate an S-shaped smooth deceleration curve interpolation command. Simultaneously with interpolation, the adaptive proportional gain and adaptive integral gain are refreshed to the servo controller; if the abnormal flag is true, the downward docking stops and a smooth rapid retreat trajectory command is output. The preset micro-tactile warning threshold is preferably a contact confidence probability value of 90% to 95%. Setting it within this range ensures that the AI ​​inference engine maintains extremely high forward sensitivity to even the weakest physical contact precursors at the probe tip, while effectively filtering false alarms caused by mechanical resonance or high-frequency electrical noise.

[0044] This step completes the closed-loop implementation from model prediction results to hardware motion commands. Accumulated contact probability triggers latching, ensuring that the motor position near the actual contact point is preserved by deterministic hardware actions, preventing latch point drift caused by software latency. Algebraic synthesis of the target position integrates the latch position, continued compression requirements, and overshoot compensation into an executable endpoint, allowing the feedforward prediction results to directly participate in trajectory planning rather than remaining solely at the model output layer. The S-shaped smooth deceleration curve limits sudden changes in end-effector velocity and acceleration, reducing the impact of the probe on the pad surface. The final output probe end-effector adaptive control command includes both the position docking target and the execution conditions for gain refresh and abnormal retraction.

[0045] Traditional docking control often uses fixed displacement or current thresholds for triggering. Signal noise, sensor response delay, and sudden changes in contact stiffness can all affect the triggering timing. Setting the threshold too low may lead to insufficient contact, while setting it too high can easily result in excessively deep pin marks. Furthermore, fixed proportional-integral parameters make it difficult to balance response speed and overshoot suppression between soft and hard pads. This step replaces the single-point hard threshold with cumulative contact probability, judging the contact state based on characteristics from multiple consecutive moments. Hardware latching ensures the physical location at the trigger moment is traceable, and the predicted overshoot compensation is pre-incorporated into the target position. Compared to the control method of overshooting followed by backoff, this step incorporates contact state correction from the starting point of the docking trajectory, making it easier to control the dispersion of pin mark depth within acceptable limits. For example, during a single soft pad contact process, the cumulative contact probability can increase from 0.2 to 0.95 within three consecutive inference cycles, exceeding the preset micro-tactile warning threshold of 0.9. The embedded AI processor then sends a latch trigger signal, and the motion control card hardware logic latches the current absolute encoder position of the motor at count point 23456 within 3μs. The control vector in the same cycle provides a target increment of 480 count points for continued downward pressure, and a predicted overshoot compensation of 25 count points in the opposite direction. After algebraic synthesis, the final target position is 23911 count points. The controller then plans a 40ms S-shaped deceleration curve and refreshes the proportional gain and integral gain. The probe finally stops at a compression depth of approximately 0.013mm from the initial contact point, with a measured needle mark depth of 17μm. If the anomaly flag becomes true during this process, the controller does not continue the downward interpolation but instead outputs a smooth, rapid retreat trajectory, causing the probe to leave the tested pad.

[0046] Furthermore, such as Figure 2 As shown, the host computer acquires the material type metadata of the printed circuit board under test. This material type metadata can be pre-input from the process document or obtained by scanning and identifying the printed circuit board using a front-facing vision camera. The material type metadata includes at least one of the following: board thickness, substrate material, hardness classification, and board warpage state. The host computer inputs the material type metadata into a pre-trained contact curve prediction model to obtain the optimal contact curve corresponding to the current material. The optimal contact curve characterizes the relationship between displacement, compression, and contact stability during probe contact. The motion control card configures the initial feed speed and acceleration / deceleration curves based on the optimal contact curve, thus performing proactive speed adjustment before the probe actually contacts the pads.

[0047] A Z-axis stepper motor drives a probe to press down onto the pad being measured for calibration. During this pressing process, a high-frequency laser displacement sensor acquires the displacement timing of the probe or elastic element, while the stepper motor driver or motion control card acquires the active current timing and absolute position timing of the motor. Simultaneously, the host computer or edge inference unit reads the historical contact data of the current probe. This multimodal data is synchronized according to a unified timestamp, forming the timing input for edge inference. By synchronously acquiring displacement, current, position, and historical contact data, the system can simultaneously reflect the probe's geometric position, changes in contact resistance, and past wear conditions.

[0048] The edge inference unit inputs sheet material type metadata, laser displacement sensor timing data, motor active current timing data, motor absolute position timing data, and historical contact data into the edge inference model. The edge inference model can employ deep neural networks, long short-term memory networks, or temporal attention networks, and is accelerated by quantization through an open neural network exchange format runtime or an edge inference acceleration engine. In one embodiment, the single inference latency is configured to not exceed 1 millisecond to meet the real-time control requirements during the high-speed probe pressing process.

[0049] The edge inference model outputs dynamic target compression, adaptive control gain, and predicted overshoot compensation. Dynamic target compression determines the target compression state after the probe contacts the pad; adaptive control gain adjusts the proportional-integral-derivative control parameters or impulse response gain at the motion control card's underlying layer; and predicted overshoot compensation compensates for docking deviations caused by the Z-axis stepper motor's deceleration inertia and mechanical transmission lag. After receiving these outputs, the motion control card writes the corresponding parameters into the hard real-time motion execution layer.

[0050] The edge inference model also performs anomaly detection. Anomaly detection includes slope change detection, fluctuation amplitude detection, or spectral feature detection of the laser displacement signal and motor active current signal. When the detection results indicate a risk of mechanical resonance, probe bending, or surface obstruction, the edge inference model outputs an active compensation command or an emergency stop command to the motion control card. The active compensation command can be used to reduce the feed rate, decrease the target compression, or adjust the gain; the emergency stop command is used to stop the Z-axis stepper motor when the anomaly risk reaches a preset level.

[0051] When the contact probability output by the edge inference model reaches a preset threshold, the motion control card triggers hardware position latching to record the current absolute position of the Z-axis stepper motor. Subsequently, the motion control card generates the target docking position based on the latched position, dynamic target compression, and predicted overshoot compensation, and switches to multi-order curve interpolation mode to drive the Z-axis stepper motor to decelerate and stop. In one embodiment, the target needle mark depth range is configured to be 5 micrometers to 20 micrometers. Through hardware position latching and multi-order curve interpolation, the system directly injects the edge inference results into the hard real-time execution process of the motion control card, reducing the impact of communication latency and mechanical inertia on the final needle mark depth.

[0052] After the probe docks, the host computer instructs the testing unit to perform electrical performance testing and records the needle mark depth, contact stability, and test result feedback data. If the test results show abnormal contact resistance, test failure, or needle mark deviation from the target range, the data evolution unit uses this feedback data as the basis for subsequent model fine-tuning and parameter updates. The reinforcement learning agent can treat a single needle prick and test cycle as a decision step, and use contact stability, needle mark depth deviation, and test failure rate as reward feedback to fine-tune the contact curve prediction model, edge inference model, or motion control parameters online.

[0053] It is particularly important to emphasize that, in order to ensure the stability and determinism of the closed-loop control system, the predicted overshoot compensation in the algebraic synthesis formula has a clear physical control direction sign, and its positive and negative signs represent different correction logics at the electrical and mechanical levels: When the predicted overshoot compensation amount is negative (-), it means that the lightweight timing model feedforward predicts that the deceleration inertia (mechanical overshoot) of the stepper motor in the forward high-speed motion will cause the final needle mark to exceed the standard. At this time, the control command feedforward at the execution layer to reduce the preset downward feed amount of the Z axis in advance, and force the actuator to make a pre-deceleration or needle lifting trend at the trajectory interpolation end to offset the overshoot energy. When the predicted overshoot compensation is positive (+), it means that the model predicts that insufficient pressing and potential excessive contact resistance may occur due to excessive local stiffness or deformation lag in the sheet metal. The control command compensates by increasing the Z-axis pressing setting. By setting the predicted overshoot compensation with a directional sign, the ambiguity of the control logic is eliminated, achieving micron-level zero overshoot hard real-time precise tracking.

[0054] Optionally, the lightweight time-series model deployed in the embedded AI processor needs to undergo an offline training and calibration preparation phase before being deployed for online real-time inference control, in order to ensure that the model has accurate nonlinear mapping capabilities. This specifically includes the following steps: (1) Sample data acquisition: During the factory calibration or offline configuration stage of the equipment, the control drive shaft carries the probe assembly at different preset pressing speeds (e.g., 50mm / s, 100mm / s, 200mm / s) to conduct actual needle piercing tests on standard test substrates (e.g., FR-4 substrate, ceramic substrate, flexible circuit board) with different physical hardness or material classification; and uses high-precision sensors to collect multiple sets of original time-series signals including laser displacement, motor feedback current and position at high frequency throughout the day to construct an offline training sample set; (2) Acquisition of supervision labels: After each needle puncture test, the actual value of the physical needle mark depth left on the tested pad is measured using an offline high-precision optical microscope or a coordinate measuring machine. The actual value of the needle mark depth is then causally and spatiotemporally aligned with the multimodal temporal feature data in the needle puncture control process and used as the supervision label (Label) corresponding to each group of temporal features in the training sample set. (3) Loss function configuration and training: Input the constructed training sample set into the time series network model to be trained. For the network branch used for regression prediction output (such as prediction of dynamic compression amount, overshoot compensation amount), the mean squared error (MSE) loss function is used for gradient update; for the network branch used for output tactile state or abnormal label classification, the cross-entropy loss function is used for constrained training until the model converges, and the offline construction of the lightweight time series model is completed.

[0055] Example 2: Furthermore, the present invention provides a pin mark adaptive control system based on timing AI prediction and motion control card, which employs a pin mark adaptive control method based on timing AI prediction and motion control card as described in the above embodiments, and can solve the technical problem of pin mark adaptive control based on timing AI prediction and motion control card. The beneficial effects of the pin mark adaptive control system based on timing AI prediction and motion control card provided by the present invention are the same as those of the pin mark adaptive control method based on timing AI prediction and motion control card provided in the above embodiments, and other technical features in the pin mark adaptive control system based on timing AI prediction and motion control card are the same as those disclosed in the methods of the above embodiments, and will not be repeated here.

[0056] Example 3: This invention provides a pinprint adaptive control device based on a timing AI prediction and motion control card. The device includes at least one processor and a memory communicatively connected to the processor. The memory stores instructions executable by the processor, which are then executed to enable the processor to perform the pinprint adaptive control method based on a timing AI prediction and motion control card described in Example 1. The pinprint adaptive control device based on a timing AI prediction and motion control card in this invention can include, but is not limited to, mobile terminals such as mobile phones, laptops, digital radio receivers, PDAs (Personal Digital Assistants), PADs (Portable Application Description), PMPs (Portable Media Players), and in-vehicle terminals (e.g., in-vehicle navigation terminals), as well as fixed terminals such as digital TVs and desktop computers. This pinprint adaptive control device based on a timing AI prediction and motion control card is merely an example and should not limit the functionality or scope of the invention. A pin-mark adaptive control device based on a timing AI prediction and motion control card may include a processing unit (e.g., a central processing unit, a graphics processing unit, etc.) that can perform various appropriate actions and processes according to a program stored in a read-only memory or a program loaded from a storage device into a random access memory. The random access memory also stores various programs and data required for the operation of the pin-mark adaptive control device based on the timing AI prediction and motion control card. The processing unit, read-only memory, and random access memory are interconnected via a bus. An I / O interface is also connected to the bus. Typically, the following systems can be connected to the I / O interface: input devices including touchscreens, touchpads, keyboards, mice, image sensors, microphones, accelerometers, gyroscopes, etc.; output devices including liquid crystal displays (LCDs), speakers, vibrators, etc.; storage devices including magnetic tapes, hard disks, etc.; and communication devices. The communication device allows the pin-mark adaptive control device based on the timing AI prediction and motion control card to communicate wirelessly or wiredly with other devices to exchange data. While a needle-mark adaptive control device based on timing AI prediction and motion control cards with various systems has been described, it should be understood that implementation of all the described systems is not required. More or fewer systems may be implemented alternatively.

[0057] Example 4: This invention also provides a computer program product, including a computer program that, when executed by a processor, implements the steps of the above-described needle mark adaptive control method based on timing AI prediction and motion control card. The computer program product provided by this invention can solve the technical problem of needle mark adaptive control based on timing AI prediction and motion control card. Compared with the prior art, the beneficial effects of the computer program product provided by this invention are the same as those of the needle mark adaptive control method based on timing AI prediction and motion control card provided in the above embodiments, and will not be repeated here.

[0058] In particular, according to the embodiments disclosed in this invention, the processes described above with reference to the flowcharts can be implemented as computer software programs. For example, embodiments of this invention include a computer program product comprising a computer program carried on a computer-readable medium, the computer program containing program code for performing the methods shown in the flowcharts. In such embodiments, the computer program can be downloaded and installed from a network via a communication device, or installed from a storage device, or installed from a read-only memory. When the computer program is executed by a processing device, it performs the functions defined in the methods of the embodiments disclosed in this invention.

[0059] It should be understood that the various parts disclosed in this invention can be implemented using hardware, software, firmware, or a combination thereof. In the description of the above embodiments, specific features, structures, materials, or characteristics may be combined in any suitable manner in one or more embodiments or examples.

[0060] Obviously, those skilled in the art can make various modifications and variations to this invention without departing from its spirit and scope. Therefore, if these modifications and variations fall within the scope of the present invention and its equivalents, the present invention also intends to include these modifications and variations.

Claims

1. A needle mark adaptive control method based on time-series AI prediction and motion control card, characterized in that, The methods include: Step S10: Obtain the periodic interrupt signal generated by the motion control card hardware timer, and execute the multi-source sensor data acquisition task based on the periodic interrupt signal using clock phase-locked loop and synchronous trigger latching method, and output the original synchronous data frame; Step S20: Based on the original synchronized data frame, a multimodal timestamp alignment task is performed using a cubic spline interpolation algorithm and a motion trend compensation method, outputting a unified fused data stream containing laser displacement value sequence, motor position value sequence, and motor current value sequence; Step S30: Based on the unified and equally spaced fused data stream, a time-series dynamic preprocessing task is performed using central difference, sliding window cross-correlation, and recursive least squares identification methods to output a multi-dimensional feature vector with physical coupling relationship; Step S40: Based on the multi-dimensional feature vectors with physical coupling relationships, a lightweight time-series model feedforward inference method is used to perform a multi-objective control quantity prediction task and output a control vector; Step S50: Based on the control vector, perform hardware latch triggering, target position algebraic synthesis and end trajectory interpolation, and output probe end adaptive control command.

2. The needle mark adaptive control method based on time-series AI prediction and motion control card as described in claim 1, characterized in that, Step S10, which involves acquiring the periodic interrupt signal generated by the motion control card hardware timer, executing the multi-source sensor data acquisition task based on the periodic interrupt signal using a clock-locked loop and synchronous trigger latching method, and outputting the original synchronous data frame, specifically includes: Step S101: The system clock with a main clock frequency of 100MHz inside the motion control card generates a hardware interrupt signal with a basic period of 10μs, and the motion control card sends a periodic synchronization pulse signal with a frequency of 1kHz to the laser displacement sensor, and uses the periodic synchronization pulse signal as the external time reference of the laser displacement sensor. Step S102: The digital phase-locked loop module inside the laser displacement sensor uses the periodic synchronization pulse signal as a reference to adjust the frequency division coefficient of the 40MHz voltage-controlled oscillator inside the laser displacement sensor in real time, so that the sampling time of the laser displacement sensor and the interruption time of the hardware interrupt signal remain in phase-locked for an entire cycle. Step S103: When a hardware interrupt signal arrives every 10μs after phase locking, the hardware synchronously triggers the laser displacement sensor to collect laser displacement data, and triggers the motion control card hardware logic to latch the absolute encoder position value of the motor and the motor current value after analog-to-digital conversion. A unified timestamp is generated by the field programmable gate array, and the laser displacement data, the absolute encoder position value of the motor, the motor current value and the unified timestamp are packaged into an original synchronous data frame and written into the high-speed first-in-first-out buffer.

3. The needle mark adaptive control method based on time-series AI prediction and motion control card as described in claim 1, characterized in that, Step S20, which involves performing a multimodal timestamp alignment task based on the original synchronized data frame using a cubic spline interpolation algorithm and motion trend compensation, and outputting a unified, equally spaced fused data stream containing laser displacement value sequences, motor position value sequences, and motor current value sequences, specifically includes: Step S201: Take out a batch of original synchronization data frames from the high-speed first-in-first-out buffer used to buffer the original synchronization data frames, and read the original timestamp sequence, laser displacement data sequence, motor absolute encoder position value sequence and motor current value sequence from the original synchronization data frames; Step S202: Using the preset AI model main inference cycle as the uniform resampling interval, find adjacent true sampling points in the original timestamp sequence that satisfy the following relationship for each uniform time point: in, For the first A uniform time point, These are the adjacent real sampling points located before the uniform time point. These are the adjacent real sampling points located after the uniform time point; Step S203: Calculate the laser displacement value corresponding to the uniform time point using a cubic spline interpolation function on the laser displacement data sequence, and calculate the motor position value corresponding to the uniform time point using linear interpolation and motion trend compensation term on the motor absolute encoder position value sequence, to obtain a unified fused data stream containing the laser displacement value sequence, the motor position value sequence, and the motor current value sequence; The laser displacement value is calculated according to the following formula: in, For the first The laser displacement value corresponding to a uniform time point For real sampling points Laser displacement data at the location, , and They are intervals cubic spline interpolation coefficients; The motor position value is calculated according to the following formula: in, For the first The motor position value corresponding to a uniform time point. For real sampling points The absolute encoder position value of the motor at that location. For real sampling points The absolute encoder position value of the motor at that location. For the first Motion trend compensation term corresponding to a uniform time point.

4. The needle mark adaptive control method based on time-series AI prediction and motion control card as described in claim 3, characterized in that, In step S202, the preset AI model main inference cycle is 200μs; within each preset AI model main inference cycle, the laser displacement data sequence and the motor absolute encoder position value sequence are resampled according to the adjacent real sampling points, so that adjacent data points in the unified and equally spaced fused data stream have the same time interval.

5. The needle mark adaptive control method based on time-series AI prediction and motion control card as described in claim 4, characterized in that, In step S30, the step of performing a time-series dynamic preprocessing task based on the unified and equally spaced fused data stream using central difference, sliding window cross-correlation, and recursive least squares identification methods to output a multi-dimensional feature vector with physical coupling relationships specifically includes: Step S301: Perform a central difference operation on the laser displacement value sequence in the unified and equally spaced fused data stream to obtain the spring compression velocity and spring compression acceleration determined by the laser displacement value; and perform a central difference operation on the motor position value sequence in the unified and equally spaced fused data stream to obtain the instantaneous speed of the motor. Step S302: Within the sliding window of the preset window width corresponding to the current uniform time point, extract the first-order difference sequence of laser displacement and the first-order difference sequence of equivalent displacement obtained by converting the motor position value. Calculate the normalized cross-correlation function between the first-order difference sequence of laser displacement and the first-order difference sequence of equivalent displacement. Multiply the offset corresponding to the maximum value of the normalized cross-correlation function by the sampling interval of the uniformly spaced fused data stream to obtain the dynamic hysteresis compensation feature. Step S303: After a contact event is detected based on the laser displacement value sequence and the motor current value sequence, obtain the laser displacement increment sequence and motor current increment sequence corresponding to multiple samples after the contact event. Use the laser displacement increment sequence as the input variable and the motor current increment sequence as the output variable. Use the recursive least squares method with a forgetting factor to identify the online stiffness estimate and bias term in the contact stiffness parameter model online. The online stiffness estimate is used to characterize the slope relationship between the motor current increment and the laser displacement increment, and the bias term is used to characterize the static current offset after the contact event. Step S304: Combine the laser displacement value, the spring compression velocity, the spring compression acceleration, the motor position value, the motor instantaneous velocity, the motor current value, the dynamic hysteresis compensation feature, and the online stiffness estimation value according to a preset feature arrangement order to obtain a multidimensional feature vector with physical coupling relationship.

6. The needle mark adaptive control method based on time-series AI prediction and motion control card as described in claim 1, characterized in that, Step S40, which involves performing a multi-objective control quantity prediction task based on the physically coupled multi-dimensional feature vectors using a lightweight time-series model feedforward inference and outputting the control vector, specifically includes: Step S401: The multi-dimensional feature vectors with a number of consecutive preset input time steps before the current uniform time point are combined into a feature vector time sequence, and the feature vector time sequence is input into a lightweight time sequence model deployed in an embedded AI processor. The lightweight time sequence model is a long short-term memory network or a converter network, and is processed by INT8 quantization and operator fusion. Step S402: Within each preset AI model main inference cycle, the lightweight temporal model performs a single feedforward inference on the feature vector temporal sequence and outputs a control vector; The control vector includes a dynamic target compression increment, an adaptive proportional gain, an adaptive integral gain, a predicted overshoot compensation, a cumulative contact probability, and an anomaly flag. The dynamic target compression increment is used to characterize the target displacement compensation amount that the probe tip needs to continue pressing down during the current contact stage. The adaptive proportional gain and the adaptive integral gain are used to refresh the closed-loop control parameters of the servo controller. The predicted overshoot compensation amount is used to characterize the excessive pressing displacement that needs to be pre-compensated before the end stops. The cumulative contact probability is used to characterize the contact confidence obtained by accumulating the contact states at multiple consecutive uniform time points. The anomaly flag is used to characterize whether a smooth rapid retreat trajectory is triggered.

7. The needle mark adaptive control method based on time-series AI prediction and motion control card as described in claim 6, characterized in that, Step S50, which involves performing hardware latch triggering, target position algebraic synthesis, and end-point trajectory interpolation based on the control vector, and outputting an adaptive control command for the probe end, specifically includes: Step S501: Compare the cumulative contact probability in the control vector with the preset micro-touch warning threshold. When the cumulative contact probability is greater than the preset micro-touch warning threshold, the embedded AI processor sends a latch trigger signal to the motion control card through the high-speed digital input / output interface, and the motion control card hardware logic latches the current absolute position of the motor to obtain the latch position. Step S502: Perform algebraic synthesis of the latch position, the dynamic target compression increment in the control vector, and the predicted overshoot compensation amount in the control vector to obtain the final target position, and generate an S-shaped smooth deceleration curve interpolation command with the final target position as the endpoint. The final target location is calculated according to the following formula: in, For the final target location, For the latch position, The dynamic target compression increment in the control vector is used. This is the predicted overshoot compensation amount with direction sign in the control vector; Step S503: Refresh the adaptive proportional gain and adaptive integral gain in the control vector to the servo controller, and execute the end-loop docking according to the S-shaped smooth deceleration curve interpolation command; when the anomaly flag is true, stop executing the end-loop docking and output the smooth rapid retreat trajectory command; when the anomaly flag is false and the position steady-state error converges to the preset error range, output the probe end adaptive control command.

8. A needle mark adaptive control system based on a time-series AI prediction and motion control card, applied to the needle mark adaptive control method based on a time-series AI prediction and motion control card as described in any one of claims 1 to 7, characterized in that, The needle mark adaptive control system based on time-series AI prediction and motion control card includes: The synchronous acquisition module is used to acquire periodic interrupt signals generated by the hardware timer of the motion control card, and to perform multi-source sensor data acquisition tasks based on the periodic interrupt signals using a clock phase-locked loop and synchronous trigger latching method, and output raw synchronous data frames. The timestamp alignment module is used to perform a multimodal timestamp alignment task based on the original synchronized data frame using a cubic spline interpolation algorithm and a motion trend compensation method, and outputs a unified fused data stream containing laser displacement value sequence, motor position value sequence and motor current value sequence; The temporal dynamic preprocessing module is used to perform temporal dynamic preprocessing tasks based on the unified equally spaced fused data stream using central difference, sliding window cross-correlation and recursive least squares identification methods, and outputs a multi-dimensional feature vector with physical coupling relationship. The feedforward inference module is used to perform multi-objective control quantity prediction tasks based on the multi-dimensional feature vectors with physical coupling relationships using a lightweight time-series model feedforward inference method, and output control vectors. The docking control module is used to perform hardware latching triggering, target position algebraic synthesis and end trajectory interpolation according to the control vector, and output probe end adaptive control commands.

9. A needle mark adaptive control device based on time-series AI prediction and motion control card, characterized in that, The needle mark adaptive control device based on time-series AI prediction and motion control card includes: a memory, a processor, and a needle mark adaptive control program based on time-series AI prediction and motion control card stored in the memory and executable on the processor. When the needle mark adaptive control program based on time-series AI prediction and motion control card is executed by the processor, it implements a needle mark adaptive control method based on time-series AI prediction and motion control card as described in any one of claims 1 to 7.

10. A computer program product, characterized in that, The computer program product includes a pin mark adaptive control program based on a time-series AI prediction and motion control card. When the pin mark adaptive control program based on a time-series AI prediction and motion control card is executed by the processor, it implements a pin mark adaptive control method based on a time-series AI prediction and motion control card as described in any one of claims 1 to 7.

Images