Intelligent control method for cutting path of wafer sawing machine

Abstract

The application relates to the technical field of precision machining control, and discloses an intelligent control method for a cutting path of a scriber, which comprises the following steps: collecting a load current signal of a main shaft servo motor in real time and storing the load current signal into a sliding window data queue; performing weighted differential calculation to obtain a real-time load change rate representing a load trend, and calculating an average absolute deviation to obtain a shock state quantity representing load discreteness; obtaining a first speed adjustment coefficient and a shock suppression coefficient based on a preset nonlinear gain scheduling relationship respectively; and multiplying the first speed adjustment coefficient and the shock suppression coefficient to couple the first speed adjustment coefficient and the shock suppression coefficient to each other, and then correcting a feed speed setting value in real time. The application constructs an orthogonal double-channel feedforward control mechanism in the same data window, and utilizes multiplication synergy of weighted differential and discreteness statistics to solve the problems of phase lag and response mismatch of conventional feedback control when facing transient impact and micro oscillation of hard and brittle materials.

Description

TECHNICAL FIELD

[0001] The present application relates to an intelligent control method for a dicing saw cutting path, belonging to the technical field of precision machining control. BACKGROUND

[0002] In the current dicing and cutting process of semiconductor wafers, optical glass and hard and brittle materials such as precision ceramics, the constant feed speed motion control strategy is the mainstream, the spindle servo system is configured in the speed loop dominant mode, the driving current is adjusted according to the position command and the encoder feedback deviation, and the set linear speed of the tool relative to the workpiece is maintained. This control logic is based on the assumption that the internal texture of the processed material is uniform, and maintains a constant cutting removal rate to ensure processing efficiency. When processing high-hardness heterogeneous materials such as silicon carbide or sapphire, the linear feedback control strategy has limitations. The hard points or lattice defects are randomly distributed in the material, and the tool is subjected to a sudden load mutation when it touches the material. The existing servo control loop relies on the current amplitude or position following error accumulation to trigger adjustment, which has a phase lag. When the control system detects overload and issues a deceleration command, the tool edge has already caused a high-energy impact on the material, resulting in edge collapse, micro-cracks or edge collapse.

[0003] To solve the above-mentioned deviation and lag, the industry has tried to introduce more complex sensing and prediction mechanisms to compensate for the cutting path. For example, the Chinese patent for invention with the publication number CN118943061B discloses a dicing saw cutting coordinate compensation method and device. A high-speed laser scanner is used to monitor the wafer surface state in real time during the cutting process, obtain wafer surface deformation data, predict future path deviation using a cutting twin model, and then correct the coordinates of the predetermined cutting path. However, this method essentially compensates for the geometric coordinates of the cutting path, and the surface deformation data it relies on is also a result of the cutting stress, not the stress itself. When the tool instantaneously hits a hard point, the compensation delay problem it aims to solve still exists in terms of suppressing instantaneous impact. The edge collapse damage caused by transient impact energy has already occurred before the deformation is detected and the model is predicted. This method focuses on ensuring trajectory accuracy, but still does not provide a solution to how to actively suppress impact load in real time to avoid edge collapse.

[0004] Therefore, how to avoid the lag based on amplitude feedback control without sacrificing system noise performance, and achieve early perception and nonlinear active suppression of transient impact load, has become a technical problem to be solved by the present application. SUMMARY

[0005] To solve the problems raised in the background art, the technical solution of the present application is as follows: an intelligent control method for a dicing saw cutting path, the method comprising:

[0006] Real-time acquisition of the load current signal of the spindle servo motor of the dicing saw at a predetermined sampling frequency, and storage of N historical sampling points of the load current signal in a sliding window data queue;

[0007] In the control cycle, the current feed speed setting value of the dicing machine is acquired, and the following calculation steps are performed in parallel based on N sample point data in the sliding window data queue:

[0008] Step A, a sliding window weighted differential logic is adopted to fit and calculate the trend slope of the N sample point data by setting a weight coefficient that is larger for a sample point closer to the current time, to obtain a real-time load change rate; based on the real-time load change rate, a first speed adjustment coefficient is obtained by querying in a preset first nonlinear gain scheduling relationship; the first nonlinear gain scheduling relationship is configured such that when the absolute value of the real-time load change rate exceeds a first threshold, the first speed adjustment coefficient exponentially decays with the increase of the absolute value of the real-time load change rate;

[0009] Step B, the average absolute deviation value between the N sample point data and the average of the N sample point data is calculated to obtain a oscillation state quantity representing the dispersion of the load current signal; and based on the oscillation state quantity, an oscillation suppression coefficient is obtained by querying in a preset second gain scheduling relationship; the second gain scheduling relationship is configured such that when the oscillation state quantity exceeds an oscillation threshold, the oscillation suppression coefficient decreases with the increase of the oscillation state quantity;

[0010] Step C, the first speed adjustment coefficient and the oscillation suppression coefficient are multiplied to obtain a final speed adjustment coefficient; the final speed adjustment coefficient is used to modify the feed speed setting value in real time, and the modified speed command is sent to the feed shaft servo system for execution.

[0011] Preferably, in step A, the sliding window weighted differential logic calculates the real-time load change rate through the following discretization formula: wherein K(t) is the real-time load change rate, I t-i is the load current value of the i-th sample point calculated from the current time, I t-N is the load current value of the earliest sample point in the sliding window data queue, w i is a preset weight coefficient, and satisfies w0>w1>•••>w N-1 , and ΔT is a sampling time interval.

[0012] Preferably, the first nonlinear gain scheduling relationship specifically includes three consecutive control intervals: a stable zone, when the absolute value of the real-time load change rate is less than the first threshold, a constant unit gain coefficient is output; a suppression zone, when the absolute value of the real-time load change rate is between the first threshold and the second threshold, a gain coefficient attenuated according to a negative exponential function is output; a protection zone, when the absolute value of the real-time load change rate is greater than the second threshold, a constant minimum safety gain coefficient is output, and the minimum safety gain coefficient is not greater than zero point one.

[0013] Preferably, the method further comprises a parameter drift compensation step based on the whole life cycle of the tool: recording the accumulated cutting distance of the current tool of the scriber; establishing a positive linear mapping relationship between the accumulated cutting distance and the first threshold value and the second threshold value; before starting each cutting task, automatically adjusting the values of the first threshold value and the second threshold value according to the current accumulated cutting distance, to offset the increase of background load fluctuation noise caused by tool wear.

[0014] Preferably, the method further comprises a feedforward control step based on the correlation of adjacent path defects: when the final speed adjustment coefficient is lower than a preset event recording threshold value, record the coordinate position of the current tool on the current cutting path as a risk node; when performing the next cutting path operation, map the coordinates of the risk node to the next cutting path to define a virtual risk area; real-time monitor the tool position, and when the tool position enters the pre-reading range of the virtual risk area, preferentially to the first nonlinear gain scheduling relationship, forcibly reduce the feed rate of the scriber to a preset safe speed; if the real-time load change rate does not exceed the first threshold value during passing through the virtual risk area, the label of the virtual risk area is removed.

[0015] Preferably, the method further comprises an edge effect compensation step based on the real-time position of the tool: real-time acquire the position coordinates of the scriber tool on the cutting path, and judge whether the position coordinates are located in a preset cutting edge buffer interval; if located in the cutting edge buffer interval, according to the distance between the current position of the tool and the cutting end point, dynamically reduce the value of the first threshold value according to a preset attenuation function, and the sensitivity of the feed rate adjustment triggered in the cutting edge buffer interval is higher than that in the non-buffer interval.

[0016] Preferably, the method further comprises an asymmetric speed recovery control step: after reducing the feed rate of the scriber by using the final speed adjustment coefficient, the control system enters a dissipation holding state; in the dissipation holding state, only when the real-time load change rate is detected to fall below the first threshold value and a preset stability confirmation condition is met, the feed rate is allowed to be increased; the stability confirmation condition includes that the absolute value of the real-time load current decreases by more than a preset hysteresis ratio or the duration after the feed rate is reduced exceeds a preset minimum time window.

[0017] Preferably, after the dissipation holding state is removed, the process of increasing the feed rate follows a limited climb logic, limiting the upper limit value of the acceleration of the speed recovery, so that the feed rate is recovered to the feed rate set value according to a linear slope, and the upper limit value of the acceleration is less than the maximum physical acceleration of the scriber servo system.

[0018] Preferably, in step B, the oscillation suppression coefficient is applied to the multiplication coupling only when the absolute value of the real-time load change rate is less than the first threshold value, ensuring that the suppression action for high-frequency micro-oscillation is independently performed when the load current signal is in a state where the trend is stable but the micro-dispersity is intense.

[0019] Preferably, the multiplication coupling of the first speed adjustment coefficient and the oscillation suppression coefficient follows the minimum priority principle: if the product of the first speed adjustment coefficient and the oscillation suppression coefficient is less than the minimum running speed coefficient preset by the system, the final speed adjustment coefficient is forcibly set to the minimum running speed coefficient, maintaining the minimum torque required for the servo motor to overcome the static friction.

[0020] Compared with the prior art, the present application has the following advantages:

[0021] 1. Transient response and anti-noise mechanism based on time sequence pattern characteristics The present application constructs a control loop containing a sliding window weighted differential calculation and a nonlinear gain scheduling table, solves the phase lag problem of conventional servo control systems when processing high-frequency step loads, uses a sliding window to perform weighted differential processing on the real-time collected current signal, filters out random electromagnetic noise interference and extracts the steepness characteristics of the load change, as a pre-trigger signal directly input into the gain scheduling table, so that the control system can forcibly attenuate the feed speed command in the initial stage when the load amplitude does not reach the damage threshold but the change rate shows a sudden trend, based on the signal time sequence pattern rather than a single amplitude response mechanism, shortens the response time of the control system to the hard point impact, and blocks the overload energy accumulation process within the safety window before the material fracture stress limit is reached.

[0022] 2. Reconstruction of asymmetric lag logic for system damping characteristics The present application introduces a control logic containing stability confirmation and limited climb asymmetric control in the speed recovery stage, eliminates the risk of secondary oscillation caused by the mismatch between the response speed of the control system and the energy dissipation speed of the mechanical system, and restores the set speed immediately after the control system does not fall back with the load change rate after the feed speed is suppressed, enters the dissipation and maintenance state, until the current absolute value falls back to the amplitude or the low-speed maintenance time meets the preset stability condition, and artificially introduces a one-way time lag and acceleration limit in the control loop for the speed recovery direction, which is equivalent to injecting dynamic damping into the servo drive system at the moment of hard point cutting, ensuring that the tool and workpiece system are in a low-energy level operation state during the mechanical relaxation period after completing high-stress cutting, avoiding the instability of the system dynamic stiffness and the generation of cutting patterns caused by the controller accelerating too early.

[0023] 3、Based on the defect memory and feedforward avoidance of spatial topology correlation The present application utilizes the spatial position correlation between multiple cutting paths to construct a closed loop containing risk node records, coordinate projection mapping and feedforward pre-reading speed reduction iterative control. The system marks the coordinate position triggering gain attenuation in the previous cutting path as a risk node, and maps it to the corresponding control buffer of the subsequent adjacent cutting path. When performing the subsequent cutting task, the control system compares the distance between the tool position and the virtual risk area in real time. Before the tool contacts the potential defect area, it preferentially detaches from the real-time feedback logic and switches to the preset safe speed. By utilizing the physical properties of the continuous distribution of defects in crystal materials along the lattice or growth direction, the single random impact response is converted into predictable spatial feedforward control. The inherent detection delay of feedback control is eliminated when dealing with continuous distribution defects, and zero-impact passing through the known risk area is realized. BRIEF DESCRIPTION OF DRAWINGS

[0024] Fig. 1 The double-channel logic architecture and data flow diagram of the intelligent control method of the present application;

[0025] Fig. 2 The performance comparison of the core algorithm of the present application and the edge collapse and load column chart;

[0026] Fig. 3 The calculation and execution node deployment architecture diagram of the control system of the present application. DETAILED DESCRIPTION

[0027] In order to make the purpose, technical scheme and advantages of the present application clearer, the technical scheme of the present application will be described in detail below. Obviously, the described embodiments are part of the embodiments of the present application, not all the embodiments. Based on the embodiments in the present application, all other embodiments obtained by those of ordinary skill in the art without creative labor belong to the scope of protection of the present application.

[0028] The application provides an intelligent control method for a cutting path of a scriber, which runs in a motion control system of the scriber, such as a CNC or PLC controller, and constitutes a control and regulation closed loop running in real time; in addition to a traditional position servo loop, a double-channel feedforward control mechanism based on spindle load time sequence form analysis, which includes a weighted differential channel for transient impact perception and a dispersion statistical channel for implicit instability perception; in each control cycle, the control system uses the same set of load current sampling points in the sliding window data queue to calculate a real-time load change rate representing a load mutation trend and a shock state quantity representing load fluctuation dispersion in parallel; the system converts the two state quantities into a first speed regulation coefficient and a shock suppression coefficient through a preset nonlinear gain scheduling relationship; by multiplying the two coefficients, the current feed speed setting value of the scriber is corrected in real time, without relying on high-cost sensors or increasing complex physical models, thereby solving the response lag and mismatch problems of conventional control strategies when facing the heterogeneity of hard and brittle materials; in a specific implementation, the execution flow of the intelligent control method is anchored to the servo control cycle of the scriber; the control system acquires the load current signal I(t) of the spindle servo motor of the scriber in real time at a predetermined high sampling frequency, for example, 10 kHz or higher, which is a direct electrical representation of cutting resistance; the latest sampling points acquired are stored in the data queue, maintaining a fixed length containing N historical sampling points, forming a sliding window data queue; the selection of N value is an engineering trade-off, a smaller N value, for example, 20, can improve response sensitivity, while a larger N value, for example, 100, helps to smooth high-frequency electromagnetic noise, and those skilled in the art can calibrate it according to the characteristics of the servo system and the characteristics of the processed material.

[0029] In each control cycle, the control system acquires the current feed speed setting value of the scriber at the same time, and based on the N sampling point data in the sliding window data queue, two calculation steps are started in parallel: Step A, the transient impact perception channel, which is to solve the phase lag problem of hard point impact; instead of using traditional mean filtering, a sliding window weighted differential logic is used, which assumes that the closer the load data point to the current time, the more predictive value it has for the impending mutation; therefore, this logic calculates the trend slope of the N sampling point data by giving a larger weight coefficient w i to the sampling points closer to the current time, thereby obtaining the real-time load change rate K(t); in a specific implementation, the real-time load change rate K(t) is calculated by the following discretization formula: where I t-i is the load current value of the i-th sampling point calculated forward from the current time, I t-N is the load current value of the earliest sampling point in the sliding window data queue, and w i is a preset weight coefficient and satisfies w0>w1>•••>wn-1>wn=0.N-1 ΔT is the sampling time interval; the absolute value of the calculated K(t) value directly represents the steepness of the load increase; the system uses the K(t) value to query the preset first nonlinear gain scheduling relationship to obtain a first speed adjustment coefficient; the scheduling relationship is designed to be a highly sensitive nonlinear form to mutations, and can be specifically divided into three consecutive control intervals: a stable interval: when the system determines that it is normal cutting fluctuation, and outputs a constant unit gain coefficient 1.0; a suppression interval: when between the first threshold T1 and the second threshold T2, the system determines that it is a precursor to encountering a hard point or a defect, at this time, the first speed adjustment coefficient exponentially decays with the increase of , and rapidly decreases in the form of , so as to forcibly suppress the feeding when the impact has not yet formed a peak; a protection interval: when the system determines that a serious impact has occurred, at this time, instead of proportional adjustment, a constant minimum safety gain coefficient is directly output, and the coefficient is not greater than 0.1, such as 0.05, so that the system crawls at the lowest speed or waits for the operator to intervene.

[0030] Step B, i.e. the micro-vibration sensing channel, is to solve the problem of micro-vibration that is ignored in step A, i.e. the problem of no obvious trend but high-frequency and violent micro-vibration; the average absolute deviation value between the data of N sampling points in the sliding window and the mean value in the window is calculated, and the value is the vibration state quantity representing the dispersion of the load current signal; the average absolute deviation instead of the variance is selected, and the calculation power consumption of the square operation is avoided for the consideration of the calculation efficiency of the control system; after obtaining the vibration state quantity, the vibration suppression coefficient is obtained in the preset second gain scheduling relationship; the relationship is configured such that when the vibration state quantity is less than the preset vibration threshold, the coefficient is 1.0, and no suppression effect is generated; when the vibration state quantity exceeds the threshold, the vibration suppression coefficient decreases with the increase of the vibration state quantity, which can decrease linearly or inversely proportionally; in the preferred embodiment, the logic orthogonality of the two channels is ensured, and the vibration suppression coefficient of step B is only allowed to participate in the subsequent multiplication coupling when step A determines that the system is in the stable zone, and this logic configuration is used to independently execute the suppression action for high-frequency micro-vibration in the state that the overall trend is stable but the local dispersion is violent; step C, i.e. the control instruction coupling and output stage, the system performs multiplication coupling on the first speed adjustment coefficient obtained in step A and the vibration suppression coefficient obtained in step B to obtain the final speed adjustment coefficient; this multiplication design ensures that any one channel detects the risk, i.e. the coefficient is less than 1.0, and can suppress the final speed; before output, the system also performs safety check, i.e. the minimum value priority principle: if the final speed adjustment coefficient obtained after multiplication coupling is lower than the minimum running speed coefficient necessary to maintain the servo motor to overcome the static friction, taking 0.02 as an example, the final speed adjustment coefficient is forcibly set to the minimum running speed coefficient, so as to avoid the physical stall of the feed shaft due to the too low speed instruction; finally, the original feed speed setting value obtained from the G code or the upper computer is corrected in real time by using the final speed adjustment coefficient obtained after coupling and verification, and the corrected speed instruction is sent to the feed shaft servo system for execution.

[0031] To further improve the adaptability and stability of the control system in the real production environment, the application can also include a series of additional control logic modules: compensation based on tool full life cycle parameter drift; due to tool wear after long-term use, the overall increase in background load fluctuation noise during cutting will occur, if not compensated, the fixed T1 and T2 thresholds will cause the system to have inconsistent sensitivity on new and old tools, resulting in false positives; the system records the cumulative cutting distance of the current tool in real time; a positive linear mapping relationship between the cumulative cutting distance and the first threshold T1 and the second threshold T2 is established; before starting each cutting task, the control system automatically adjusts the values of T1 and T2 according to the current cumulative cutting distance, dynamically offsetting the control baseline drift caused by tool wear; edge effect compensation based on tool real-time position; the support structure is weakest and the edge collapse resistance is lowest at the cutting edge of the hard and brittle material; therefore, the system obtains the position coordinates of the dicing machine tool on the current cutting path in real time; it is judged whether the coordinates are located within the preset cutting edge buffer interval, which is set as the area 0.5mm away from the cutting end; if the tool enters the buffer area, the system will dynamically and real-time reduce the value of the first threshold T1 according to the distance between the current position of the tool and the cutting end according to the preset attenuation function; this makes the control system have higher triggering sensitivity to load fluctuation in the cutting edge area than in the center area of the material.

[0032] The feedforward control based on the correlation of adjacent path defects considers the internal defects of wafers and other materials. For example, the lattice defects or hard points may be continuously distributed along a specific crystal direction in space, which may cause the risk of repeated occurrence near the same coordinate position of adjacent cutting paths. Therefore, a defect memory mechanism is introduced. When the Nth path is executed, if the final speed adjustment coefficient is lower than the preset event recording threshold value for determining whether a serious risk is encountered, the coordinate position of the current tool on the path is immediately recorded and marked as a risk node. When the N+1th cutting path is executed, the coordinates of all risk nodes recorded on the Nth path are projected and mapped onto the N+1th path to define a plurality of virtual risk areas. When the tool position enters the pre-reading range of these virtual risk areas, the control system will prioritize the real-time feedback logic and forcibly reduce the feed rate to a preset safe speed. In addition, the mechanism also has a dynamic clearing capability. If the detected real-time load change rate K(t) does not exceed the first threshold T1 during the tool passes through the virtual risk area at a safe speed, the system determines that the defect is not continuous at this point, and immediately removes the marking of the virtual risk area to avoid efficiency loss caused by excessive defense. Asymmetric speed recovery control is introduced to solve the problem of secondary shock caused by the mismatch between the energy dissipation speed of the mechanical system and the response speed of the control system. When the system reduces the feed rate due to triggering suppression, the control system immediately enters the dissipation holding state. In this state, the system does not immediately restore the speed just because K(t) falls back, but must wait for the strict stability confirmation condition to be met. The condition is configured to detect that the real-time load change rate has fallen below the first threshold T1, and meet one of the following two conditions: one is that the absolute value of the real-time load current decreases by more than a preset hysteresis ratio, for example, 10% of the peak value; the second is that the duration of the reduced feed rate exceeds a preset minimum time window, for example, 30 ms. Only when the composite condition is met, the system allows to remove the dissipation holding state. After removal, the process of increasing the feed rate is also not step, but follows the limited climbing logic: the system limits the upper limit of the speed recovery acceleration, so that the feed rate is restored to the original feed rate set value according to a gentle linear slope. The upper limit of the acceleration is set to be less than the maximum physical acceleration of the scriber servo system.

[0033] Example 1: This embodiment is a typical application scenario of the intelligent control method and system in the cutting operation of polycrystalline silicon / carbonized silicon composite material. The core challenge of this scenario is the huge difference in physical properties between the hard phase SiC particles and the matrix Si, which leads to a cutting load that presents a dramatic random step mutation and a high-frequency oscillation coexistence working condition. In a specific cutting task, the dicing machine starts cutting at a feed speed setting value of 100 mm / s. In each control cycle, the control system collects the load current signal of the spindle servo motor in real time, for example, 0.1 ms, and stores it in a sliding window data queue with a length of N=50 sampling points. In the initial stage of cutting, the tool is located in the uniform silicon matrix, and the load current signal fluctuates smoothly around a low mean value. Step A, i.e., the weighted differential of the sliding window, calculates the absolute value of the real-time load change rate K(t) which is continuously below the first threshold T1. At the same time, step B, i.e., the average absolute deviation, calculates the oscillation state quantity which is also below the oscillation threshold. The first speed adjustment coefficient and the oscillation suppression coefficient are both constant at 1.0, and the final speed adjustment coefficient after multiplication coupling is also 1.0. The actual execution speed of the feed shaft servo system is the set value of 100 mm / s. When the tool starts to contact a 0.8 mm diameter silicon carbide hard particle at X=10.5 mm coordinate, the cutting resistance increases instantaneously, and the load current signal starts to rise sharply. In the initial stage of signal rising, for example, the first 5 ms, the absolute value of the load has not triggered any overload protection threshold of the traditional servo system. The weighted differential logic of step A is sensitive to this rising trend due to the high weight of the latest data point, and the calculated real-time load change rate K(t) quickly exceeds the first threshold T1, entering the suppression zone of the first nonlinear gain scheduling relationship. The system looks up the table to obtain the first speed adjustment coefficient with exponential decay, such as 0.6. At this time, step B is bypassed because K(t) does not meet the stable zone condition, and the final speed adjustment coefficient is 0.6. The control system has already modified the actual feed speed command to 100x0.6=60 mm / s before the load amplitude reaches the peak value. As the tool further cuts into the hard point, K(t) continues to increase and may exceed the second threshold T2, triggering the protection zone logic, which forces the first speed adjustment coefficient to be set to the minimum safety gain coefficient 0.05, and the feed speed is suppressed to 5 mm / s. The system slowly grinds through the hard point with very low energy instead of hitting it at high speed. During the cutting process of the hard point, another specific working condition may also occur: the tool does not cut smoothly, but induces high-frequency cutting chatter inside the hard phase. At this time, the load current signal may present a kind of overall trend stability, i.e., fluctuating on a certain high platform, with a dramatic local dispersion state. In this state, the weighted differential logic of step A may fall below the first threshold T1 due to the unclear trend, and the first speed adjustment coefficient is restored to 1.0.0; the average absolute deviation calculation of step B will capture this severe dispersion, causing the oscillation state to exceed the oscillation threshold, and thus output an oscillation suppression coefficient less than 1.0, such as 0.7; since the control logic is set to activate the coupling of step B only in the steady region, the final speed adjustment coefficient at this time will become 1.0 x 0.7 = 0.7, and the feed speed is corrected to 70 mm / s; this double-channel orthogonal cooperation solves the limitations of single control logic, and the weighted differential logic of step A solves the overall step impact of the trend, while the dispersion statistical logic of step B solves the microscopic oscillation that the weighted differential logic cannot perceive, both of which together ensure comprehensive suppression of complex loads.

[0034] When the tool completely passes through the hard point and the load current falls, the system does not immediately restore the speed to 100 mm / s; at this time, the asymmetric speed recovery control logic intervenes, and the system enters a dissipation holding state; the control system continuously monitors the load signal until the real-time load change rate K(t) falls below T1, and the absolute value of the load current has decreased by more than the preset hysteresis ratio or the low-speed holding time has exceeded the preset minimum time window; only when this composite stability confirmation condition is met, the system will release the holding state and start the limited climb logic to make the feed speed linearly ramp up to 100 mm / s according to the preset lower acceleration upper limit value, avoiding secondary impact or oscillation of the tool or workpiece due to premature and rapid reacceleration; since the adjustment coefficient triggered below the event recording threshold at X = 10.5 mm, the system has marked this coordinate as a risk node; when the scriber completes this path and moves to the next adjacent cutting path, the feedforward control logic based on the adjacent path defect association is activated; the system projects the X = 10.5 mm coordinate onto the new path to define a virtual risk area; when the tool position is about to enter the pre-reading range of the virtual risk area during the execution of the new path cutting, the control system will no longer rely on the real-time feedback calculation of steps A and B, but will preferentially execute the feedforward logic to forcibly reduce the feed speed to the preset safe speed in advance; this iterative learning based on spatial topological association converts the passive response of the previous cutting into active avoidance this time, enabling the system to achieve a control mode transition from lagging feedback to zero-delay feedforward when facing defect clusters continuously distributed along the lattice.

[0035] Example 2: The following comparative test is set up in this example to quantify the synergy of the double-channel feedforward control mechanism of the present application in suppressing transient impact and high-frequency chatter; the test platform uses an industrial-grade scriber, and the motion controller CNC has the ability to run custom control logic at 10 kHz and high-speed acquisition of the load current signal of the spindle servo motor; the test workpiece selects a silicon substrate with multiple columns of known coordinates pre-embedded, with alumina (Al2O3) hard particles with diameters of 0.5 mm to 0.8 mm, to simulate random hard point impact and chatter caused by non-homogeneous materials encountered during cutting.

[0036] The test sets five control groups, each of which cuts 10 parallel cutting paths on the same workpiece. The basic feed speed setting value is constant at 80 mm / s. The controller monitors and records the spindle peak load current when cutting to the hard particle coordinates in real time. After cutting is completed, the average edge collapse width of each path at the hard particle position is measured using a scanning electron microscope (SEM) as an evaluation index of processing quality. The control logic of the five groups is as follows: Control group 1: uses the standard constant feed speed control of the prior art, i.e. always runs at 80 mm / s; Control group 2: uses a simplified amplitude feedback control, i.e. when the absolute value of the load current exceeds the preset overload threshold, which is set to 80% of the peak value of control group 1, the feed speed is reduced to 10 mm / s; Partial missing group A: step A of the present application is enabled, i.e. only based on the real-time load change rate K(t) and the first nonlinear gain scheduling relationship to adjust the speed, and the oscillation suppression coefficient of step B is forcibly set to 1.0; Partial missing group B: only step B of the present application is enabled, i.e. only based on the oscillation state quantity and the second gain scheduling relationship to adjust the speed, and the first speed adjustment coefficient of step A is forcibly set to 1.0; The sample group of the present application: uses the complete two-channel control method of the present application, i.e. step A and step B are calculated in parallel, and the first speed adjustment coefficient and the oscillation suppression coefficient are multiplied and coupled to jointly correct the final feed speed. The key performance data measured in the test is shown in Table 1.

[0037] Table 1: Performance comparison test data table under different control logics

[0038]

[0039] Analysis of Table 1 data: Control group 1 experienced a high-speed direct impact on the hard point, resulting in the highest peak load current of 7.93A and the most severe edge chipping of 24.6μm. Control group 2, relying on the absolute value of the load, inevitably experienced a lag in adjustment action after the impact. Although the peak current decreased to 6.21A, the damage had already occurred before deceleration, resulting in limited improvement in the edge chipping width of 18.3μm. The peak load current of partially missing group A (3.15A) was significantly lower than the control group, indicating that the K(t)-based feedforward logic could suppress the feed before the impact energy fully accumulated, i.e., through exponentially decaying gain, and the edge chipping width of 8.1μm was also significantly improved. The experimental data for partially missing group B showed a peak current of 7.79A and an edge chipping width of 23.9μm, similar to control group 1. This data suggests that the mean absolute deviation logic is mainly used to sense high-frequency oscillations, but for trend-based oscillations... The step impact is essentially ineffective, which also proves the necessity of step A in suppressing the impact. The data of the sample group of this invention shows that the peak load current of 3.22A is at the same level as that of the partially missing group A, indicating that the impact energy is also suppressed by step A, but the average chipping width of 4.7μm is better than that of 8.1μm in the partially missing group A. This data comparison shows that the dual-channel synergy of this invention is not a simple functional superposition. While step A detects the impact trend through K(t) and performs the main speed reduction, the oscillation state quantity in step B detects the high-frequency chatter generated when the tool cuts into the hard phase and outputs an additional oscillation suppression coefficient. The multiplicative coupling of the two coefficients enables the system to further suppress the micro chatter of the cutting edge during the slow rolling over of the hard point, thereby achieving suppression in both the step impact and high-frequency chatter dimensions, and finally obtaining the best machining quality.

[0040] Example 3: This example combines Figs. 1 to 3 This describes an intelligent control method for the cutting path of a dicing machine, such as... Fig. 1 As shown, the method begins with the spindle servo motor, which acquires the load current signal I(t) in real time. The signal is stored in a sliding window data queue containing N historical sampling points. The data in this queue is processed in parallel by two channels: the first channel performs sliding window weighted differentiation to fit the trend slope and obtain the load change rate K(t). The K(t) value is used to query the first nonlinear gain scheduling table to output the first speed adjustment coefficient, which decays exponentially. The second channel performs mean absolute deviation calculation to statistically analyze the load dispersion and obtain the oscillation state quantity. This state quantity is used to query the second gain scheduling relationship to output the oscillation suppression coefficient. The output coefficients of the two channels are coupled by multiplication and minimum value verification to achieve coefficient coupling and minimum running speed protection. The finally generated corrected speed command is sent to the feed axis servo system for execution. The load change feedback from the physical process forms a closed loop. In addition, the architecture also integrates four additional modules: an edge effect compensation module based on the real-time position of the tool, a parameter drift compensation module based on the entire tool life cycle, an asymmetric speed recovery control module, and a feedforward control module based on the defect association of adjacent paths.

[0041] As shown in Fig. 2 , the vertical coordinate of the figure is the numerical value, and the legend distinguishes between peak load current A and average collapse edge width μm. The data shows that the sample group of the present application has advantages in both peak load current and average collapse edge width compared with the supplementary control groups C and D; as shown in Fig. 3 , the architecture is divided into a computing node and an execution node, where the computing node is a motion control host, i.e., a CNC or PLC kernel, which runs under a high-frequency real-time task cycle and internally includes a data preprocessing area, a double-channel intelligent analysis engine, a dynamic parameter mapping library, a speed decision and recovery controller, and a spatial defect memory component. The computing node receives real-time load current signals from the main shaft servo motor and outputs corrected feed speed instructions. The execution node is a servo drive system, which includes current loop and speed loop control, and is responsible for receiving the instructions and driving the feed shaft servo motor to perform variable speed motion.

[0042] Example 4: To verify the rationality of the selection of the specific algorithm in steps A and B of the present application, the following comparative test groups were additionally set based on the test platform and workpiece of Example 2; all test groups used complete double-channel coupled logic, and the internal core algorithm of step A or step B was modified: supplementary control group C: step A used sliding window ordinary differential logic, i.e., all weight coefficients w i were equal w i =1 / N, and step B used the average absolute deviation (MAD) algorithm of the present application; supplementary control group D: step A used the sliding window weighted differential logic of the present application, and step B used the standard deviation (SD) algorithm instead of the average absolute deviation (MAD); the test results compared with the sample group of the present application (data same as Example 2) are as follows:

[0043] Table 2: Core algorithm comparison test data table

[0044]

[0045] Analysis of Table 2 data: comparing the sample group of the present application 4.7 μm with the supplementary control group C 10.9 μm, there are differences in peak current and collapse edge width; this shows that using weighted differential logic, i.e., giving recent data points higher weights, can capture the trend of impact occurrence earlier and more sensitively compared with ordinary differential logic, thereby triggering more timely speed suppression, confirming the superiority of weighted logic; comparing the sample group of the present application (MAD) with the supplementary control group D (standard deviation), the peak currents 3.22 A vs. 3.19 A and the collapse edge widths 4.7 μm vs. 4.9 μm are at the same level, with no statistical difference; this shows that using average absolute deviation as the oscillation state quantity has physical effects on suppressing high-frequency chatter and improving machining quality comparable to using the more computationally intensive standard deviation algorithm.

[0046] Embodiment 5: This embodiment discloses a parameter calibration procedure for controlling the system, for determining the parameters of dynamic compensation and recovery logic, including the parameters of tool wear compensation, edge effect compensation and asymmetric recovery logic, for determining the length N of sliding window data queue and weight coefficient w i The calibration procedure of the length N includes: collecting multiple sets of load current signals of the dicing saw in an idle state or in a smooth cutting of a standard homogeneous material at a constant speed as baseline noise samples, identifying the main electromagnetic and mechanical noise frequency f noise of the system by performing spectral analysis on the baseline noise samples, setting the length N, and the corresponding total time window , where ΔT is the sampling time interval, which is greater than twice the main noise period, i.e. 2 / f noise , to smooth the high-frequency noise, and the weight coefficient w i is generated using a normalized linear decay function or an exponential decay function, satisfying w0> w1> •••> w N-1 , and the specific form of the decay function is adjusted by the skilled person in the art according to the autocorrelation characteristics of the baseline noise samples, for optimizing the signal-to-noise ratio of the trend fitting; the calibration procedure of the oscillation threshold T osc in step B reuses the baseline noise samples collected when the length N is set or multiple sets of load current signals collected when the standard homogeneous material is smoothly cut at different set feed speeds, for each set of smooth cutting signals, the average absolute deviation value sequence is calculated using the sliding window of length N, and the mean value and the standard deviation of the sequence are counted, the oscillation threshold T osc is set as the high percentile point of the statistical distribution of the average absolute deviation in the smooth state, for example , so that the threshold is higher than the background oscillation level of normal cutting, and the adjustment of the oscillation suppression coefficient is triggered only when high-frequency microscopic chatter occurs, resulting in a dramatic increase in dispersion, the parameter drift compensation model based on tool wear is calibrated, for determining the positive linear mapping relationship between the cumulative cutting distance d and the first threshold T1 and the second threshold T1; a new tool is selected, a reference cutting test is performed on a standard homogeneous workpiece, load data is collected during smooth cutting, and the initial first threshold T 1,new is obtained by the statistical method of mean value plus 3 times standard deviation; the tool is continuously cut under production conditions, and the cumulative cutting distance d is recorded in stages, for example at the nodes of d = 1000m, d = 2000m, d = 3000m; at each distance node, the reference cutting test is repeated, and the smooth load data under the wear state is collected, and the background noise threshold T 1,worn (d) is calculated under the wear state by the same statistical method; a set of data pairs is obtained, and linear regression fitting is applied to the set of data points to obtain the linear mapping relationship where k is the drift coefficient; the compensation model of the second threshold T2 is also calibrated in this way.

[0047] Calibration of the edge effect compensation logic parameters based on the real-time position of the tool, including determining the width L of the cutting-out edge buffer zone buffer and the form of the attenuation function f(x); using a well-maintained tool, cutting in the central region of a standard workpiece, recording the first threshold T1 of its reference; observing the micro-cracks caused by the tool at the cutting-out edge in a standard cutting task through a high-power microscope, measuring the maximum extension length L in the cutting-out direction carck , if the statistical average value of L carck is 0.28 mm; setting the width L buffer of the cutting-out edge buffer zone to be greater than this length and adding a safety margin, setting L buffer = 0.6 mm; within this buffer zone, setting a linear attenuation function f(x) that restores the threshold to normal at x = L buffer , and the threshold is the lowest at x = 0, i.e., the cutting end point, taking 0.5 × T1 as an example, determining the function used to calculate the dynamic local first threshold T 1,local (x): where x is the distance between the current position of the tool and the cutting end point; calibration of the parameters of the asymmetric speed recovery control logic for determining the preset hysteresis ratio P lag , the preset minimum time window t min , and the acceleration upper limit value a rec in the stability confirmation condition; in the impact test, through high-frequency data acquisition, the average time required from the load impact peak point to the system mechanical vibration decay and convergence to the background noise level can be recorded using the spectrum analysis of the accelerometer or current loop signal, if the statistical value of this time is 38 ms, the preset minimum time window t min = 50 ms is set; at the same time, the batch data is analyzed, when the mechanical vibration converges, the absolute value of the real-time load current has all fallen below 15% of the impact peak value, accordingly, the preset hysteresis ratio P lag is set to 15%, in the control logic, t min and P lag satisfy one of them, which constitutes the stability confirmation condition; for the acceleration upper limit value a rec , a rec is set to be less than the maximum physical acceleration a max of the servo system, taking a rec = 0.2 × a max as an example, for gentle speed recovery.

[0048] In a system deployment scenario, when the control method of the present application is first applied to a dicing saw with new uncalibrated characteristics or a new batch of workpiece material, a pre-system procedure can be executed to manage the initial engagement boundary condition of the control logic; specifically, the control system executing the initial down-cut engagement phase in the Z-axis, i.e. the depth of cut axis, will temporarily bypass or mask the weighted differentiation in Step A and the dispersion statistics in Step B, forcing the final speed adjustment factor to remain at 1.0, and the masking state will continue until the controller confirms that the tool has fully entered the workpiece to the preset depth of cut and the X-axis or Y-axis, i.e. the feed axis, has started executing the cutting path instructions. This boundary condition management is used to prevent the inevitable high load transient changes when the tool is vertically entering the material from being misjudged as a cutting anomaly, and to avoid triggering unnecessary feed speed suppression when the cutting task has not yet officially started.

[0049] After the control system enters the cutting path execution phase, the implementation procedure of the feedforward control logic based on the adjacent path defect correlation is as follows: the system maintains a dynamically updated list of risk node coordinates in the controller memory. When the final speed adjustment factor output by Step C is below a preset event recording threshold value, which can be calibrated to 0.5, indicating that speed suppression has occurred, the controller immediately captures the current coordinate position of the tool on the cutting path and stores the coordinate in the risk node coordinate list. Before executing the next adjacent cutting path, the control system iterates through the list and maps all recorded risk node coordinates to the new path, defining a virtual risk area by applying an asymmetric pre-reading range, set to extend from 0.3 mm upstream of the risk node coordinate to 0.1 mm downstream. When the tool travels along the new path guided by the control system, as soon as its real-time position enters the pre-reading range of any virtual risk area, the feedforward control logic takes precedence and forces the feed speed instruction to switch to a preset safe speed, calibrated to 25% of the feed speed set value, and maintains this safe speed until the tool completely exits the virtual risk area. In addition, the system performs dynamic verification on the list. If the real-time load change rate K(t) calculated by Step A does not exceed the first threshold value T1 at any time during the tool's passage through the virtual risk area at safe speed, the system determines that the defect is not continuous or has disappeared on the current path, and immediately removes the risk node coordinate from the list.

[0050] It will be apparent to those skilled in the art that the present application is not limited to the details of the above-described exemplary embodiments, and can be implemented in other specific forms without departing from the spirit or essential characteristics of the present application.

[0051] Finally, it should be noted that the above examples are merely intended to illustrate the technical solutions of the present application and not to limit the present application. Although the present application has been described in detail with reference to the preferred embodiments, those skilled in the art should understand that the technical solutions of the present application can be modified or equivalently replaced without departing from the spirit and scope of the present application.

Claims

1. An intelligent control method for the cutting path of a dicing machine, characterized in that, The methods include: The load current signal of the spindle servo motor of the dicing machine is collected in real time at a predetermined sampling frequency, and N historical sampling points of the load current signal are stored in the sliding window data queue. Within the control cycle, the current feed speed setpoint of the dicing mill is acquired, and based on N sampling points in the sliding window data queue, the following calculation steps are performed in parallel: Step A: Using sliding window weighted differential logic, by setting a larger weight coefficient for sampling points that are closer to the current time, the trend slope of N sampling point data is fitted and calculated to obtain the real-time load change rate. Based on the real-time load change rate, the first speed adjustment coefficient is obtained by querying the preset first nonlinear gain scheduling relationship; the first nonlinear gain scheduling relationship is configured such that when the absolute value of the real-time load change rate exceeds the first threshold, the first speed adjustment coefficient decreases exponentially with the increase of the absolute value of the real-time load change rate. Step B: Calculate the average absolute deviation between the N sampled data points and the mean of the N sampled data points to obtain the oscillation state quantity characterizing the dispersion of the load current signal. Based on the oscillation state quantity, the oscillation suppression coefficient is obtained by querying the preset second gain scheduling relationship; The second gain scheduling relationship is configured such that when the oscillation state quantity exceeds the oscillation threshold, the oscillation suppression coefficient decreases as the oscillation state quantity increases; Step C: Multiply and couple the first speed adjustment coefficient with the oscillation suppression coefficient to obtain the final speed adjustment coefficient; use the final speed adjustment coefficient to correct the feed speed setpoint in real time, and send the corrected speed command to the feed axis servo system for execution; In step A, the sliding window weighted differential logic calculates the real-time load change rate using the following discretization formula: Where K(t) is the real-time load change rate, I t-i To calculate the load current value at the i-th sampling point from the current moment, I t-N w represents the load current value at the earliest sampling point in the sliding window data queue. i The weights are preset and satisfy w0 > w1 > ... > w N-1 ΔT is the sampling time interval.

2. The intelligent control method for the cutting path of a dicing machine according to claim 1, characterized in that, The first nonlinear gain scheduling relationship specifically includes three continuous control intervals: the steady region, where a constant unity gain coefficient is output when the absolute value of the real-time load change rate is less than the first threshold; the suppression region, where a gain coefficient that decays exponentially according to the natural exponential function is output when the absolute value of the real-time load change rate is between the first and second thresholds; and the protection region, where a constant minimum safe gain coefficient is output when the absolute value of the real-time load change rate is greater than the second threshold, and the minimum safe gain coefficient is no greater than 0.

1.

3. The intelligent control method for the cutting path of a dicing machine according to claim 2, characterized in that, The method also includes a parameter drift compensation step based on the entire tool lifecycle: recording the cumulative cutting distance of the current tool of the dicing machine; establishing a positive linear mapping relationship between the cumulative cutting distance and the first threshold and the second threshold; Before each cutting task is started, the values ​​of the first and second thresholds are automatically increased based on the current cumulative cutting distance to offset the increase in background load fluctuation noise caused by tool wear.

4. The intelligent control method for the cutting path of a dicing machine according to claim 1, characterized in that, The method also includes a feedforward control step based on the correlation of defects in adjacent paths: when the final speed adjustment coefficient is lower than the preset event recording threshold, the coordinate position of the current tool on the current cutting path is recorded as a risk node; When executing the operation on the next cutting path, the coordinates of the risk node are mapped onto the next cutting path to define a virtual risk area; The tool position is monitored in real time. When the tool position enters the pre-read range of the virtual risk area, the feed speed of the dicing machine is forcibly reduced to the preset safe speed, taking priority over the first nonlinear gain scheduling relationship. If the real-time load change rate does not exceed the first threshold during the passage through the virtual risk zone, the virtual risk zone will be removed from the list.

5. The intelligent control method for the cutting path of a dicing machine according to claim 1, characterized in that, The method also includes an edge effect compensation step based on the real-time position of the tool: the position coordinates of the dicing machine tool on the cutting path are obtained in real time, and it is determined whether the position coordinates are within the preset cutting edge buffer zone; if they are within the cutting edge buffer zone, the value of the first threshold is dynamically reduced according to the distance between the current position of the tool and the cutting end point, based on the preset attenuation function. The sensitivity of triggering feed speed adjustment within the cutting edge buffer zone is higher than that outside the buffer zone.

6. The intelligent control method for the cutting path of a dicing machine according to claim 1, characterized in that, The method also includes an asymmetric speed recovery control step: after the feed speed of the dicing machine is reduced using the final speed adjustment coefficient, the control system enters a dissipative holding state; in the dissipative holding state, the feed speed is only allowed to be increased when the real-time load change rate is detected to fall below the first threshold and the preset stability confirmation condition is met. Stability verification conditions include: the absolute value of the real-time load current decreases by more than a preset hysteresis ratio or the duration of the feed rate reduction exceeds a preset minimum time window.

7. The intelligent control method for the cutting path of a dicing machine according to claim 6, characterized in that, After the dissipation holding state is released, the process of increasing the feed rate follows the restricted climb logic, which limits the upper limit of the acceleration for speed recovery, so that the feed rate recovers to the feed rate set value by a linear ramp. The upper limit of the acceleration is less than the maximum physical acceleration of the dicing machine servo system.

8. The intelligent control method for the cutting path of a dicing machine according to claim 1, characterized in that, The oscillation suppression coefficient is applied to the multiplicative coupling only when the absolute value of the real-time load change rate is less than the first threshold, ensuring that the suppression action against high-frequency micro oscillations is executed independently when the load current signal is in a state of stable trend but severe micro-dispersion.

9. The intelligent control method for the cutting path of a dicing machine according to claim 1, characterized in that, The multiplicative coupling of the first speed adjustment coefficient and the oscillation suppression coefficient follows the minimum value priority principle: if the product of the first speed adjustment coefficient and the oscillation suppression coefficient is less than the minimum operating speed coefficient preset by the system, the final speed adjustment coefficient will be forcibly set to the minimum operating speed coefficient to maintain the minimum torque required for the feed axis servo motor to overcome static friction.

Images