An edge-computing-based precision fastener production process anomaly identification method

Abstract

The present application relates to the technical field of production process control, in particular to a precision fastener production process anomaly identification method based on edge computing, the present application collects forming load of each station by taking crank angle as a reference, extracts the contact starting phase relative to the strip baseline, determines normal occupancy, stacking or empty position and constructs the occupancy matrix; searches for empty position chain, combines the upstream stacking and continuous stroke to reverse the drop, sticking or retention, generates the part flow trajectory table; extracts the load displacement integral characteristics of the same blank across stations and the same station across blanks, identifies part-level anomalies, station anomalies or machine factors according to the consistency of deviation direction; determines the shunt time according to the part flow trajectory table and the material discharge time of flight, and outputs the cooling, reducing the stroke or maintenance direction according to the station event accumulation, temperature rise rate and station-level anomaly direction.

Description

Technical Field

[0001] This invention relates to the field of production process control technology, specifically to a method for identifying anomalies in the production process of precision fasteners based on edge computing. Background Technology

[0002] Automated production lines for precision fasteners typically include stations for feeding, forming, transfer, inspection, and rejection. Existing control systems establish a flow mapping between workpieces, cavities, inspection results, and rejection positions based on encoder pulses, station trigger signals, and preset cycle times, thereby enabling production process monitoring and rejection of defective parts.

[0003] When issues such as material skipping, slippage, material stacking, material shortage, mold sticking, or sequence transition failure occur during production, the actual cavity occupancy status may differ from the control system's record. Existing methods often compare the occupancy status with sensor signal amplitude and a fixed threshold; however, signal drift, mechanical vibration, mold impact, and load changes can easily cause amplitude range overlap, leading to misjudgments or missed judgments. Incorrect occupancy information entering the cycle shift chain disrupts the part flow mapping, causing misalignment of rejected objects.

[0004] Existing production control methods are insufficient to confirm the cavity occupancy status in real time without relying on amplitude criteria, and are also insufficient to correct part flow mapping, restore rejection direction, and locate production anomalies that cause mapping damage in a timely manner. Therefore, it is necessary to improve the automatic control and anomaly identification methods for the production process of precision fasteners.

[0005] To address this, a method for identifying anomalies in the production process of precision fasteners based on edge computing is proposed. Summary of the Invention

[0006] The purpose of this invention is to provide a method for identifying anomalies in the production process of precision fasteners based on edge computing. By identifying the cavity occupancy through contact start phase and reconstructing the part flow trajectory, the method uses load displacement integral features to distinguish between part-level, station-level and whole-machine anomalies, implements directional flow diversion according to the trajectory, and links cooling, stroke reduction and maintenance.

[0007] To achieve the above objectives, the present invention provides the following technical solution: An edge computing-based method for anomaly identification in the manufacturing process of precision fasteners includes: Using the crank angle as a reference, the forming load at each station is collected stroke by stroke, and the starting angle of the forming load relative to the material baseline is extracted to obtain the contact start phase. Based on the offset direction of the contact start phase relative to the nominal phase, each stroke of each station is judged as normal occupation, overlapping or empty, and arranged into an occupation matrix. Search for empty space chains in the occupancy matrix along the synchronously increasing direction of station number and stroke number. Based on whether there is a stacking position upstream of the station starting point of the empty space chain and the number of strokes continuing from the stacking position, invert it into part drop, sticking, or retention, and generate a part flow trajectory table. Based on the part flow trajectory table, the load displacement integral characteristics of the same blank across workstations and the same workstation across blanks, starting from the contact start phase, are used to form part-level comparison sets and workstation-level comparison sets. The consistency of the deviation direction within the set is judged as part-level anomaly or workstation anomaly. Based on the part flow trajectory table, the diversion time is obtained, and the diversion command is output to the blank corresponding to the part-level anomaly and the merging flag. The interface lubrication deterioration is determined according to the station-level anomaly direction and station temperature rise rate, and the cooling command and the reduction stroke command are output.

[0008] Preferably, the contact initiation phase acquisition process is as follows: After the first piece passes inspection, the median of the forming load at the same station within a continuously set number of strokes is calculated point by point according to the rotation angle, and the calculation result is determined as the initial value of the strip baseline; the median of the forming load at each stroke of the same station that is judged to be in normal position within the rolling window is calculated point by point according to the rotation angle, and the strip baseline is updated stroke by stroke; the forming load at each station is subtracted from the strip baseline of the corresponding station point by point to obtain the station difference sequence; adjacent differences are performed on the station difference sequence along the rotation angle direction, and the first rotation angle position corresponding to the difference that exceeds the noise band for a continuously set number of strokes is retrieved; the fixed time lag of the acquisition path is converted into the rotation angle increment according to the current stroke, and the rotation angle increment is subtracted from the first rotation angle position to obtain the contact initiation phase.

[0009] Preferably, the median of the contact start phase of each stroke that is determined to be normally occupied within the scrolling window at the same workstation is calculated to obtain the nominal phase of the corresponding workstation; the normal zone of each workstation is determined according to the discrete range of the contact start phase within the same scrolling window; the phase difference between the contact start phase of each workstation within the same stroke and the nominal phase of the corresponding workstation is calculated, and the phase difference is sorted by rank according to the degree of advance, and the phase difference corresponding to the median rank is determined as the common mode level of the current stroke, and the offset direction is determined according to the order of the phase difference of each workstation relative to the common mode level; if the offset direction is in the middle, it is determined to be normally occupied; if the offset direction is advanced and the phase difference exceeds the normal zone, it is determined to be overlapped; if the offset direction is lagging and the phase difference exceeds the normal zone, it is determined to be normally occupied and a lag flag is recorded; if no start edge is detected, it is determined to be empty; strokes with lag flags are included in the part-level comparison set according to the correspondence in the occupation matrix.

[0010] Preferably, each newly identified empty cell is used as a seed cell, and subsequent empty cells are connected along the direction of synchronously increasing station number and stroke number to obtain an empty cell chain and the starting point of the empty cell chain; when the upstream station corresponding to the starting point of the empty cell chain is normally occupied within the initial stroke, it is determined by inversion to be a dropped part; when the upstream station is stacked within the initial stroke and the stacking does not continue into subsequent strokes, it is determined by inversion to be a stuck part; when the upstream station is stacked within the initial stroke and the stacking continues into subsequent strokes, it is determined by inversion to be a stuck part. If the stacking stroke count exceeds the set limit, a stop interlock command is output. The part flow trajectory table uses the cut-off stroke number as the primary key and registers the event type, arrival stroke of each station, merging flag, final state, and discharge stroke. When the inversion result is sticking, a binding relationship is established between the primary key of the returned billet and the primary key of the newly fed billet at the same station, and the merging flag is registered. When the inversion result is dropping or being stuck, the final state of the corresponding billet is recorded as exiting the part flow, and the discharge stroke is not registered.

[0011] Preferably, based on the correspondence between the rotation angle and the slider displacement, the displacement increment corresponding to each sampling point within the rotation angle interval from the initial contact phase to the bottom dead center is determined, and the station difference sequence is integrated and accumulated point by point according to the corresponding displacement increment to obtain the load displacement integral feature; based on the rows of the part flow trajectory table, the load displacement integral features corresponding to each arrival stroke of the same billet are extracted to form a part-level comparison set; based on the columns of the part flow trajectory table, the load displacement integral features of each billet corresponding to the same station within the scroll window are extracted to form a station-level comparison set; billets with a merge mark are included in the station-level comparison set; the median of each comparison set is calculated and an insensitive band is set, and the falling... Elements entering the insensitive zone are marked as zero, and the remaining elements are marked with their corresponding deviation directions. When the number of non-zero elements in the part-level comparison set reaches the set number and their deviation directions are consistent, while the deviation directions of the non-zero elements in the station-level comparison set are inconsistent, it is determined to be a part-level anomaly. When the number of non-zero elements does not reach the set number, the corresponding blank is transferred to the re-inspection sequence. When the deviation directions of the non-zero elements in the station-level comparison set are consistent, while the deviation directions of the non-zero elements in the part-level comparison set are inconsistent, it is determined to be a station anomaly, and the station-level anomaly direction is recorded. When the deviation directions of the non-zero elements in both the part-level comparison set and the station-level comparison set are consistent, it is determined to be a machine-wide factor.

[0012] Preferably, the discharge flight time is categorized and calibrated according to the stroke count. The discharge flight time of the corresponding grade is selected based on the current stroke count and converted into a stroke increment. The stroke increment is added to the discharge stroke in the part flow trajectory table to obtain the diversion stroke. The diversion time is determined based on the ejection angle interval within the diversion stroke. The blanks corresponding to part-level anomalies and blanks corresponding to merging marks are identified as diversion objects. When the current stroke count changes, the discharge flight time of the corresponding grade is reselected based on the changed stroke count, and the stroke increment is recalculated. The diversion times that have not yet been executed are rearranged as a whole, and the diversion command is executed.

[0013] Preferably, the generation process of the cooling command and the stroke reduction command includes: accumulating the number of times the part flow trajectory table is inverted into sticking and retention according to the workstation, respectively, to obtain the workstation event accumulation; when the workstation event accumulation increases, the workstation temperature rise rate increases, and the workstation-level abnormal direction occurs continuously and simultaneously, it is determined that the interface lubrication is deteriorated, and the cooling command and stroke reduction command are output simultaneously; when the workstation event accumulation increases but the workstation temperature rise rate does not increase, it is determined that the ejection link and the transfer link are abnormal, and the maintenance direction is output without outputting the stroke reduction command; after outputting the stroke reduction command, the number of stacks, the number of empty chains, and the workstation event accumulation in the scroll window simultaneously fall back as the criteria for restoring strokes, and the strokes are restored step by step according to the gradient; when they do not fall back simultaneously, the mold change maintenance direction is output.

[0014] Compared with the prior art, the beneficial effects of the present invention are as follows: 1. This invention uses crank angle to unify the sampling time base of each station, determines the contact start phase by utilizing the starting edge of the forming load relative to the material baseline, and weakens the influence of speed changes, acquisition lag, and overall machine phase drift by using the rolling updated nominal phase, normal band, and stroke common mode level. Normal occupancy, stacking, and vacancy are determined based on the presence or absence of the starting edge and its advance relationship relative to the common mode, eliminating the need to directly rely on load amplitude to distinguish occupancy states. By combining the vacancy chain with the upstream stacking continuity characteristics to invert dropped parts, stuck bands, and stagnation, the correspondence between billet, station, and stroke can be restored after part flow misalignment, improving the reliability of occupancy identification and part flow reconstruction.

[0015] 2. This invention integrates the load difference from the initial contact phase to the bottom dead center using the slider displacement, forming a load-displacement integral characteristic that considers both the load level and the forming stroke. Based on the part flow trajectory table, it establishes part-level comparison sets for the same blank across different workstations and workstation-level comparison sets for the same workstation across different blanks. Through cross-judgment using the median benchmark, insensitive zone, number of non-zero deviations, and directional consistency, it can distinguish between part-level anomalies, workstation anomalies, machine-wide factors, and blanks awaiting re-inspection. This reduces the interference of single-point fluctuations, workstation differences, and sample dispersion on the judgment results, ensuring that anomaly attribution is consistent with the actual part flow state.

[0016] 3. This invention records the discharge stroke, final state, and merging mark using a part flow trajectory table. Combined with the discharge flight time calibrated by stroke grade, it determines the diversion stroke and diversion time. After stroke changes, it rearranges the diversion times that have not yet been executed, ensuring accurate diversion direction for abnormal blanks and merged blanks formed by adhesion, reducing false rejections and missed rejections. By jointly analyzing the accumulated amount of adhesion and retention, the station temperature rise rate, and the station-level abnormal direction, it can distinguish between interface lubrication deterioration and malfunctions in the ejection and transmission links, outputting cooling, stroke reduction, or maintenance directions accordingly. Based on the decline in abnormal indicators, it restores the stroke count or direction for mold replacement and maintenance, forming a closed-loop abnormality handling system. Attached Figure Description

[0017] Figure 1 A schematic diagram of a method for identifying anomalies in the production process of precision fasteners based on edge computing, provided by the present invention; Figure 2 A schematic diagram of the logical flow of anomaly inversion and component flow trajectory generation based on empty chain search provided by the present invention; Figure 3 This is a schematic diagram of the consistency cross-determination logic flow based on orthogonal comparison sets provided by the present invention. Detailed Implementation

[0018] To make the objectives, technical solutions, and advantages of this invention clearer, the invention will be further described in detail below with reference to the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are for illustrative purposes only and are not intended to limit the invention.

[0019] Example 1: This example is implemented in a multi-station cold heading machine, processing miniature precision fasteners used in laptops and mobile phones. The forming machine has sequentially arranged cutting stations and several forming stations. The spindle completes one stroke per revolution, and each station is simultaneously loaded within the same stroke. The blank is pushed downstream stroke by stroke from the cutting station along with the clamping and transfer. Under normal conditions, the station number and stroke number increase synchronously; this correspondence is the part flow mapping. A discharge passage and a two-position diversion mechanism are provided below the last station.

[0020] The raw data entering the edge computing device consists of only three channels: the equally spaced angle pulses and bottom dead center zero pulses output by the spindle angle encoder; the forming load output by the axial sensitive elements on the die holder support columns of each forming station; and the temperature output by the temperature measuring elements on the outer wall of the die sleeve near the cavity of each station. The analog-to-digital conversion of the load channel is directly triggered by the angle pulses, binding the sampling points to the crank angle one by one. Fluctuations in spindle speed and start / stop acceleration / deceleration do not change the mechanism configuration corresponding to the sampling points; the temperature channel samples over time. These three channels of raw data are no longer used individually. They are all converted step by step through the following data chain and finally leave the edge computing device in the form of diversion commands, cooling commands, reduced stroke commands, and shutdown interlock commands. The execution targets are, in order, the material discharging and diversion mechanism, the cooling and lubrication branches of the corresponding stations, and the main drive.

[0021] Please see Figure 1 This invention provides a method for identifying anomalies in the production process of precision fasteners based on edge computing. The technical solution is as follows: The forming load at each station is collected stroke by stroke using the crank angle as a reference. The starting angle of the forming load relative to the material baseline is extracted to obtain the contact initiation phase. Based on the offset direction of the contact initiation phase relative to the nominal phase, each stroke at each station is judged as normal occupation, overlap, or vacancy, and arranged into an occupation matrix. In the occupation matrix, an vacancy chain is searched along the synchronously increasing direction of the station number and stroke number. The vacancy chain is determined by the presence or absence of overlap at the upstream station of the starting point of the vacancy chain and the continuity of the overlap. The number of strokes is inverted into part dropping, sticking, or retention, generating a part flow trajectory table. Based on the part flow trajectory table, the load-displacement integral characteristics of the same blank across stations and the same station across blanks, from the initial contact phase, are used to construct part-level comparison sets and station-level comparison sets. The consistency of the deviation direction within the sets is judged as either part-level or station-level anomaly. The flow splitting time is obtained based on the part flow trajectory table, and flow splitting commands are output for blanks corresponding to part-level anomalies and merging flags. Interface lubrication degradation is determined based on the station-level anomaly direction and station temperature rise rate, and cooling and stroke reduction commands are output.

[0022] Further, the contact start phase acquisition process is as follows: the initial value of the material baseline is taken from the forming load of the same station within a continuously set number of strokes after the first piece passes inspection, and the median value is taken point by point according to the rotation angle. The forming load of each stroke judged as normally occupied in the same station within the rolling window is taken point by point according to the rotation angle and updated stroke by stroke. The forming load of each station is subtracted from the material baseline of the station to obtain the station difference sequence. Adjacent differences are made along the rotation angle in the station difference sequence. The rotation angle position where the first continuously set number of differences all cross the noise band is searched. The fixed time lag of the acquisition path is converted into the rotation angle increment according to the current stroke and then subtracted from the rotation angle position to obtain the contact start phase.

[0023] Specifically, each transformation on the data chain requires a reference, which is obtained once during the calibration phase after the model is assembled. The calibration order is arranged according to the order in which they are consumed on the data chain.

[0024] First, obtain the angular displacement correspondence table. Using the bottom dead center zero-position pulse as the origin, sequentially number the angular pulses within one spindle revolution to form angular indices. Then, using the crank radius, connecting rod length, and slider guide assembly dimensions, calculate the slider displacement increment between adjacent indices and record it in a table. The slider guide assembly dimensions refer to the assembly eccentricity obtained by measuring the lateral offset of the slider guide trajectory centerline relative to the crank rotation centerline in a direction perpendicular to the crank rotation axis after the main drive components (slider, guide, connecting rod, crank) are assembled or overhauled, when the slider is in the top dead center position. This eccentricity, along with the crank radius and connecting rod length, serves as the input for calculating the slider displacement increment index by index. This corrects for displacement deviations caused by assembly eccentricity in the slider guide, ensuring that the displacement increment recorded in the angular displacement correspondence table reflects the actual stroke of the slider after this assembly, rather than the displacement of the ideally aligned crank-slider mechanism. Specifically, let the crank radius be... The length of the connecting rod is The assembly eccentricity of the slider guide assembly dimension is: The crank angle corresponding to the angle index is Then, the position of the slider relative to the crank rotation center and along the slider guide direction under that rotation angle index. Determined by the following formula: Adjacent corner indices , Corresponding slider displacement increment Take the difference between the slider positions of the two indices, that is... The displacement increment calculated in this way is then recorded as the content of the rotation displacement corresponding table.

[0025] This assembly dimension only changes with the assembly or overhaul of the main drive components; the mold assembly itself does not change this dimension. Therefore, it is measured and recorded at the same time as the other calibration contents in the angular displacement correspondence table, maintaining the principle of "unchanged with mold changes, and re-retrieved after main drive modification." The displacement increment in the table decreases point by point along the downward dead center direction from the angular index. This property is directly used in the load-displacement integral characteristic conversion at the end of the data link. This table remains unchanged with mold changes and is only re-retrieved after main drive modification.

[0026] Next, the fixed time lag of each load channel is calibrated. A step excitation is injected at the channel input, and the time taken for the output to reach steady state from the start is recorded, and registered separately for each workstation. This value is used for time delay compensation of the contact start phase on the data link.

[0027] Next, calibrate the noise band. First, with the spindle stationary and the system powered on, record the static fluctuation range of each load channel within a set time. Then, during the idle stroke, record the repeatability and discrete range of multiple strokes at the same rotation angle index for each channel. Take a multiple of the larger of the two as the noise band for that station. The selection criterion is that the noise band can be used to scan the entire idle stroke without detecting any starting edge. This value is used for starting edge retrieval, and the calibration timing is after each shift's start-up and each mold change.

[0028] In this embodiment, the step excitation refers to the sudden change in load formed by applying a set load value to the axial sensitive element of the die holder support column at the workstation in a time much faster than the response time of the load channel using a mechanical loading device with a synchronous trigger signal output, such as a drop hammer or a pneumatic impact loading head; at the same time as loading, an electrical trigger signal synchronized with the sudden change in load is output as the starting point for recording the response time.

[0029] The initial value of the strip baseline is then taken. After mold assembly and first-piece inspection, the set number of strokes are run continuously. The load samples at the same station are taken point by point according to the corner index, and the median value is written into the strip baseline table. The median value is taken instead of the average value because there may still be individual abnormal strokes that are not detected in this stroke segment, and the median value is not sensitive to a few outliers. The strip baseline is the subtraction of the first conversion of the data link, and is subsequently updated by the output of the link itself. Only the initial value is provided here.

[0030] Using this initial value, the contact initiation phase is calculated for the same stroke segment. The median phase of each station is entered into the nominal phase table, and a multiple of the discrete range of the contact initiation phase within the same stroke segment is entered into the normal zone table. Both are used to determine the offset direction. Next, a pre-formed part of a certain station is pre-placed in the concave mold cavity of that station. A calibration stroke is run, and the difference between the measured contact initiation phase and the nominal phase of that station is the part height phase advance of that station. This is calibrated for each station separately. The correspondence between displacement and rotation angle in the crank mechanism's neighborhood of the bottom dead center is non-linear. The same increment in stack height corresponds to different increments in rotation angle at the contact working points of different stations. Therefore, this advance amount cannot be extended from one station to other stations. The lower limit of the normal zone is taken as a fraction of this advance amount, so that the normal zone and the stack position do not overlap on the phase axis.

[0031] Then, by comparing the load-displacement integral characteristic of the first confirmed stroke with the load-displacement integral characteristic corresponding to the allowable load of the punch, the allowable number of stacked parts in the cavity is calculated. The integer less than this number is taken as the upper limit of the number of stacked strokes, which is used for the shutdown interlock in the attribution process. The physical basis of this upper limit is that the cavity volume and the allowable stress of the punch together limit the number of stacked parts. Exceeding this limit will lead to punch breakage or die cracking, and the stacking is therefore not a freely continuing observation state.

[0032] Specifically, the load-displacement integral characteristic corresponding to the allowable load of the punch is obtained in the following way: the allowable load of the punch is the known design parameter of the punch at this station, which is calculated from the rated allowable stress and the bearing cross-sectional area of ​​the punch when it is selected; on the station difference sequence of the first piece confirmation stroke record, the angle position corresponding to when the load value reaches the allowable load of the punch is retrieved along the angle direction; in the same way as described above, the displacement of the station difference sequence is converted from the contact start phase and accumulated point by point, and the interval from the contact start phase to the angle position is integrated and accumulated, and the accumulated result is determined as the load-displacement integral characteristic corresponding to the allowable load of the punch at this station, and is calibrated and registered for each station; this value is only updated after the punch is replaced or the first piece confirmation stroke is re-executed.

[0033] Finally, three benchmarks are established for the end segment of the data link: Marked billets are placed at several stroke positions, and the time taken from ejection at the final station to reaching the diversion position is recorded. These times are then recorded as a discharge flight time grading table, with values ​​obtained through linear interpolation between positions. The number of positions is determined based on the stroke increment corresponding to the interpolation error between adjacent positions not exceeding one stroke. The discrete range of the load-displacement integral characteristics of the same station during continuous normal operation after the first piece is confirmed is taken as the insensitive zone. A certain multiple of the upper bound of the discrete range of the temperature rise rate at the same station during continuous operation after the equipment reaches thermal equilibrium is taken as the temperature rise rate increase condition. At this point, the benchmarks for each node in the data link are complete, and the link can begin operation.

[0034] Starting from the bottom dead center zero-position pulse, the synchronous conversion of the load channels at each workstation is triggered point by point according to the angle pulse. The sampled values ​​are written into the single-stroke load frame with the workstation number as the row and the angle index as the column. When the next zero-position pulse arrives, the frame is sealed, and the number of strokes and the stroke period of this stroke are obtained from the interval between two adjacent zero-position pulses. Within the same stroke, a section of wire is cut off at the cutting workstation, and this stroke number is used as the cutting stroke number of this section of wire. A new row is added to the part flow trajectory table, the event type is set to empty, and the final state is set to in progress. The single-stroke load frame, the number of strokes of this stroke, and the cutting stroke number constitute the only input for all subsequent processing. The original sample of this stroke is no longer referenced.

[0035] Subtracting the material baseline table from each point of the single-stroke load frame yields the station difference sequence. This step eliminates the inherent load pattern of the station under normal occupancy, making the changes in the packing state within the cavity visible in the sequence. The station difference sequence is the first branching node on the entire data chain: one path continues along the corner index direction to obtain the contact start phase; the other path is temporarily stored and used to calculate the load-displacement integral characteristic after the contact start phase is determined.

[0036] The difference between two adjacent points in the station difference sequence is calculated along the corner index direction to obtain the difference sequence. The difference is used instead of the original value because the criterion is the position where the load begins to appear, not the magnitude of the load. The difference is not sensitive to the overall rise and fall of the sequence, but only to the starting point of the rise. The difference sequence is scanned point by point downwards from the corner index starting point towards the dead center. The index position where the first consecutive set number of differences all cross the noise band of that station is recorded as the starting edge index. In this paper, the starting edge refers to the starting position of the station difference sequence rising from the baseline level along the corner index direction. The requirement of several consecutive crossings, rather than a single crossing, is to eliminate single-point pulse interference. The required number of consecutive crossings ensures that the angle crossed is less than the angle crossed by the shortest contact segment in the first piece confirmation stroke, and greater than the angle crossed by single-point pulse interference. The selection criterion is that both no false detections during the idling stroke and no missed detections during the first piece confirmation stroke are simultaneously met.

[0037] The fixed time lag registered during the calibration phase of this workstation is multiplied by the spindle angular velocity corresponding to the measured stroke of this stroke to obtain the angular increment. This increment is then subtracted from the starting edge index to obtain the contact start phase and written into the contact start phase table. The angular velocity is taken from the measured value of this stroke rather than the set value. Therefore, when the stroke rate is reduced during operation, the compensation amount automatically follows without the need for further calibration. This also holds true after the stroke rate reduction command is output at the end of the data link.

[0038] If no index position meeting the conditions can be found within the entire corner interval, the presence or absence of the starting edge of this stroke at that station is recorded as "no starting edge," and this flag is transmitted downstream along with the contact start phase table. If the station difference sequence of all stations within the same stroke does not cross the noise band within the entire corner interval, the acquisition path is determined to be faulty. This stroke does not output the presence or absence of the starting edge downstream, the corresponding column of the occupancy matrix is ​​set to "unresolved," and this stroke does not participate in any updates to the material baseline table, nominal phase table, or normal band table. The reason for this branch is that using the inability to detect the starting edge as a vacancy criterion is a negative judgment. Channel open circuit, amplification failure, or sensitive component detachment also manifests as no starting edge. If no distinction is made, channel faults will be read as vacancy across all stations, and amplified into a large number of pseudo-vacancy chains through the downstream attribution link.

[0039] Furthermore, the nominal phase is obtained by taking the median of the contact start phase of each stroke judged as normally occupied within the scrolling window at the same workstation. The normal zone is given by the discrete range of the contact start phase within the same scrolling window for each workstation. The difference between the contact start phase of each workstation within the same stroke and the nominal phase of its corresponding workstation is taken, and the order is determined according to the degree of advance. The phase difference corresponding to the median of the order is taken as the common mode level of this stroke. The offset direction is determined based on the order of the workstation relative to the common mode level. If the offset direction is in the middle, it is judged as normally occupied. If it is advanced and exceeds the normal zone, it is judged as overlapping. If it is lagging and exceeds the normal zone, it is judged as normally occupied and a lag mark is added. If there is no starting edge, it is judged as empty. Strokes with lag marks are included in the part-level comparison set along with the occupation matrix.

[0040] Specifically, after the contact start phase table enters this stage, the reference used by this stage itself is refreshed in reverse: each stroke judged as normally occupied within the scrolling window of this station is taken, and the median position is taken point by point according to the corner index to refresh the strip baseline table. The median position of the contact start phase within the same window is taken to refresh the nominal phase table, and the discrete range of the contact start phase within the same window is taken to refresh the normal strip table. Strokes set as unjudged are not included in any of the above refreshes. The lower limit of the scrolling window length ensures that the number of normally occupied strokes accumulated within the window of each station is sufficient to take the median position, and the upper limit ensures that the drift of the nominal phase within the corresponding time period of the window length is less than the normal strip width, which is determined by the drift record of the nominal phase during continuous operation after the first piece confirmation. This forms the first closed loop of this method: the result of the occupation determination is fed back to the reference on which the occupation determination is based. The baseline and nominal phase slowly follow with mold wear and heat, and no manual reset is required after mold change. The benchmark refresh is delayed until the part-level and station-level judgments are completed. Only strokes that are normally occupied, have no lag flags, and are not classified as abnormal are included. When there are insufficient valid samples, unjudged strokes, or continuous abnormalities, the three benchmarks are frozen. The benchmarks are restored after the first piece is reviewed and qualified and normal strokes that meet the lower limit of the rolling window are re-accumulated.

[0041] Take the contact start phase of each effective station along the starting edge within this stroke, and subtract the nominal phase of each station from the phase difference to obtain the phase difference. Rank each phase difference according to its lead time from largest to smallest, and take the phase difference corresponding to the middle rank as the common mode level of this stroke. The common mode is determined by ranking rather than numerical subtraction because the distance from each station to the neutral surface of the frame and the stiffness of the die holder are different. The phase change amplitude caused by the elastic stretching of the same machine in each channel is not equal. Subtracting by equal amount would leave under-deduction at stations with lower stiffness. However, the elastic stretching of the whole machine, batch differences in wire diameter, and overall fluctuations in feed rate all have the same effect on each channel within the same stroke, only changing the absolute value of the phase difference, not the order of phase differences between stations. Therefore, the ranking is not sensitive to whether the gain of each channel is consistent. If the number of valid working positions at the start of this stroke is less than the set number, the common mode level is not reset, and the most recent valid common mode level is used. The set number is more than half of the number of valid working positions at the start, so that the middle-ranked position is not occupied by a working position that is simultaneously abnormal.

[0042] The offset direction is determined with the common mode level as a reference: strokes earlier than the common mode level and exceeding the lower boundary of the normal zone of this station are marked as advanced; strokes later than the common mode level and exceeding the upper boundary of the normal zone are marked as lagging; and the rest are marked as centered. Based on this, a three-state projection is generated and written into the current column of the placeholder matrix: centered strokes are judged as normal placeholders; advanced strokes are judged as stacked strokes, which physically mean that the axial stacking height inside the cavity increases and the punch contacts the material earlier, corresponding to the adhesive part stacking with the new blank after falling back or the retained part inside the cavity stacking with the new blank; lagging strokes are judged as normal placeholders and marked with a lag flag, which physically means that the stacking height is insufficient, corresponding to the blank being cut too short or not fed enough; strokes without a starting edge are judged as empty. Strokes with lag flags are passed downstream along the placeholder matrix and used as one of the inputs for part-level anomaly judgment in the comparison set stage. They do not constitute a separate rejection criterion because insufficient stacking height may also come from the slight sinking of residual grease after being squeezed out of the cavity, which does not constitute a part defect.

[0043] The placeholder matrix is ​​a two-dimensional circular buffer. Rows are workstation numbers, and columns are modulo stroke numbers. The buffer depth is no less than a multiple of the sum of the total number of workstations and the expected longest empty chain length, ensuring that no empty chain is covered by the circular buffer before attribution is completed. The current column of the placeholder matrix represents all the content passed from this stage to the next stage.

[0044] Furthermore, referring to Figure 2 Using each newly identified empty cell as a seed, connect subsequent empty cells along the synchronously increasing direction of station number and stroke number to obtain an empty cell chain and an empty cell chain starting point; if the upstream station of the empty cell chain starting point is in a normal position during the initial stroke, it is inverted as a dropped part; if it is a stacked position and the stacking does not continue to the subsequent stroke, it is inverted as a stuck part; if it is a stacked position and the stacking continues to the subsequent stroke, it is inverted as a stuck part; when the number of stacked position strokes exceeds the set upper limit, a stop interlock command is output; the part flow trajectory table uses the cut-off stroke number as the main key and records the event type, the strokes reached by each station, the merging flag, the final state, and the discharge stroke; when the inversion is a stuck part, the main key of the returned billet and the newly fed billet at the same station is bound as the merging flag; when the inversion is a dropped part or a stuck part, the final state of the corresponding billet is recorded as exiting the part flow, and the discharge stroke is not recorded.

[0045] Specifically, a single cell in the placeholder matrix can only answer whether there are more or fewer parts in the cavity, but it cannot answer where the parts went, and therefore it is insufficient to reconstruct the part flow mapping. The object processed in this step is not the cells themselves, but the arrangement of the cells on the plane formed by the workstation and the stroke.

[0046] Normal occupancy, overlapping, and vacancy are observed states; dropped parts, stuck parts, and lingering parts are event types registered based on the occupancy matrix pattern, which are not equivalent to the unique confirmation of mechanical root causes. When adjacent events overlap, vacancy chains fork or merge, upstream states are not judged, or the same pattern meets multiple conditions, the event type is recorded as pending confirmation and transferred to re-inspection; subsequent states are written back after uniquely matching the established pattern.

[0047] Each newly identified empty cell in this stroke is added to the empty cell linked list as a seed. The starting station number and starting stroke number are taken as the cell coordinates, the chain length is set to the initial value, and the attribution flag is set to undetermined. In each subsequent stroke, the cells of each unterminated chain are checked at the station number plus one and the stroke number plus one. Cells that are empty are merged into the chain and the chain length is increased; cells that are not empty are terminated. Columns set to undetermined do not participate in connection and termination judgments to avoid misinterpreting undetermined cells as empty cells and generating pseudo-chains. The physical basis for this extension direction is: after a billet exits the part flow, its vacancies shift downstream stroke by stroke with the clamping and transfer, which is represented as a diagonally arranged sequence of empty cells on the occupancy matrix.

[0048] The cell containing the smallest workstation number in the chain is taken as the starting point of the empty chain. The occupancy status of its upstream adjacent workstation in the initial stroke is read, and the cause is attributed accordingly. If the upstream workstation is in normal occupancy, it is attributed to part dropping, meaning that the upstream workstation is forming normally but the part has not reached the downstream, and the drop occurs in this transfer link. If it is in a stacked position and the stacking does not continue into subsequent strokes, it is attributed to adhesion, meaning that the part is carried up by the punch and falls back into the prototype cavity in the next stroke and is stacked with the newly fed blank. If it is in a stacked position and the stacking continues into subsequent strokes, it is attributed to retention, meaning that the part is left in the cavity and the ejection channel continues to fail. When the number of stacked position strokes exceeds the upper limit set during the calibration stage, a stop interlock command is output to the main drive.

[0049] The attribution results are then written into the part flow trajectory table. This table uses the cut-off stroke number as the primary key, and the fields include event type, arrival stroke array, merging flag, final state, discharge stroke number, and load displacement integral characteristic array. For items attributed to adhesion, the primary key of the returned blank and the primary key of the newly delivered blank at the same station mutually point to each other as the merging flag. Both share the arrival strokes of subsequent stations and the same discharge stroke, and their final states are both recorded as merged. After being stacked in the cavity, they advance downstream as a single entity. The last station generates only one discharge event, resulting in one less discharge event than the cut-off event in the part flow trajectory table. This difference is the source of the empty chain in the placeholder matrix. For items attributed to dropped or stuck parts, the corresponding blank's final state is recorded as exiting the part flow. The arrival stroke array is not filled in from the start of the empty chain, and the discharge stroke field is left blank. For blanks not involved in the above events, the arrival strokes are filled in sequentially according to station number, and the arrival stroke of the last station is written into the discharge stroke field, with the final state recorded as discharged.

[0050] Distinguishing between adhesion and retention requires determining the occupation status of the next stroke, which introduces an inherent delay of one stroke. If the remaining number of strokes from the current station to the last station is insufficient to cover this delay, it is first registered as pending confirmation, and the corresponding discharge position is written to the re-inspection diversion queue. After the next stroke is confirmed, the event type, final state, and merging flag are written back. Once the diversion record enters the execution window, it is locked, and subsequent write-backs must not change the queue order of locked records and other primary keys.

[0051] In a certain intermediate station, the forming process is normal during a certain stroke, but the part is carried away by the punch. The station is still considered to be in normal position for that stroke. In the next stroke, the cavity of the station is filled with the returned part and the new blank. The contact start phase is advanced and exceeds the lower boundary of the normal band, so it is judged as a stacked position. The downstream adjacent station in the same stroke is judged as an empty position. This empty position becomes the seed of the empty position chain and extends obliquely stroke by stroke. In the attribution link, the upstream station of the starting point is stacked in the initial stroke and does not continue. It is attributed to adhesion. The two primary keys point to the merge flag, share the subsequent arrival stroke and the same discharge stroke, and the final state is recorded as merged. This merged row does not participate in the part-level judgment in the comparison set link but is directly listed as the diversion object. The diversion link schedules the diversion time for it, and because the discharge event of the last station is one less than the cut-off event, it is not reserved for the missing one. The adhesion number of this station is entered into the station event accumulation table and participates in the three-branch judgment of lubrication deterioration.

[0052] If the ejection channel at a certain station fails, the part remains in the cavity; in the next stroke, the station is judged to be stacked, and its downstream is judged to be empty; in the stroke after that, the station is still judged to be stacked, and the stacking continues, which is attributed to stagnation. The final state of the corresponding billet is recorded as exiting the part flow, and the discharge stroke is not registered. The diversion queue does not reserve a position for it; when the number of stacking strokes exceeds the set upper limit, a stop interlock command is output to the main drive; the number of stagnations at this station is entered into the station event accumulation table.

[0053] If a billet falls off at a certain transfer stage, the upstream station remains in normal position during the stroke, and the downstream station determines that there is no position from the next stroke and forms an oblique empty position chain; if the upstream station at the reading start point of the attribution stage is in normal position, it is attributed to a dropped part, and the corresponding final state of the billet is recorded as the exited part flow; the diversion queue neither schedules the diversion time for this primary key, nor causes the subsequent primary keys to be delayed due to its absence.

[0054] At this point, the component flow mapping is reconstructed. The component flow trajectory table is a fan-out node on the entire data chain, and its contents are subsequently referenced in three places: the attribution of load displacement integral characteristics, the scheduling of split times, and the statistics of workstation event accumulation.

[0055] Furthermore, referring to Figure 3Based on the correspondence between rotation angle and slider displacement, the station difference sequence is converted point by point into displacement increments within the rotation angle interval from the initial contact phase to the bottom dead center, and then accumulated to obtain the load-displacement integral feature. The load-displacement integral features of the same billet at each arrival stroke are extracted from the rows of the part flow trajectory table, forming a part-level comparison set. The load-displacement integral features of each billet at the same station within the scroll window are extracted from the columns of the part flow trajectory table, forming a station-level comparison set. Billets with a merged flag are included in the station-level comparison set. Within each comparison set... Take the median and set an insensitive zone. Elements falling into the insensitive zone are recorded as zero, and the remaining elements are recorded as deviation direction. If the number of non-zero elements in the part-level comparison set reaches the set number and the direction is consistent, but the direction is inconsistent in the station-level comparison set, it is judged as part-level anomaly. If the number of non-zero elements is less than the set number, the corresponding blank is transferred to the re-inspection sequence. If the direction of non-zero elements in the station-level comparison set is consistent, but the direction is inconsistent in the part-level comparison set, it is judged as station anomaly and the station-level anomaly direction is recorded. If the direction is consistent in both comparison sets, it is judged as a whole machine factor.

[0056] Specifically, this step retrieves the previously stored station difference sequence and merges it with the part flow trajectory table. First, the station difference sequence is extracted starting from the contact start phase of the stroke at that station and ending at the lower dead center. Then, the angular displacement correspondence table is consulted, and the slider displacement increment corresponding to each index within the interval is multiplied by the difference at that index. This multiplication is then accumulated along the interval to obtain the load-displacement integral characteristic of that stroke at that station. This is written into the corresponding primary key's load-displacement integral characteristic array according to the correspondence given in the part flow trajectory table. The conversion is performed before accumulation, rather than directly accumulating along the angular displacement. This is because the displacement increment in the angular displacement correspondence table decreases point by point along the downward dead center direction from the angular displacement. Since the contact working points of each station are different, the amount obtained by directly accumulating along the angular displacement is not in constant proportion to the actual deformation work, and the comparison set formed across stations will lose its physical comparability.

[0057] After the integral values ​​are entered into the table, the load-displacement integral characteristics of the same primary key at each arrival stroke are taken along the rows of the part flow trajectory table to form the part-level comparison set for that billet. The load-displacement integral characteristics of the same station at each primary key within the scroll window are taken along the columns of the part flow trajectory table to form the station-level comparison set for that station. Billets marked with a merge flag are counted once in the station-level comparison set based on the primary key of the billet brought back. The primary key of newly delivered billets at the same station is not counted repeatedly, avoiding the artificial amplification of the station's directional consistency by having the same load-displacement integral characteristic counted twice. The elements of both sets are taken from the same batch of integral values, and their orientations are orthogonal. This orthogonality is only established after the part flow mapping reconstruction and is a prerequisite for the separation of causes. Before forming the part-level comparison set, the elements are first benchmarked within the workstation according to the median and insensitive zone scale of the normal occupancy scroll window of their respective workstations. This ensures that elements from different workstations represent the deviation direction relative to the normal distribution of their respective workstations. Elements that have not completed the workstation-level calibration are marked as invalid and do not participate in cross-workstation consistency determination. The workstation-level comparison set only contains data from the same workstation.

[0058] Within each comparison set, a median is taken and an insensitive band is set. Elements falling into the insensitive band are marked as zero, those above the median but outside the insensitive band are marked as positive, and those below the median but outside the insensitive band are marked as negative. The criterion is determined by the deviation direction after benchmarking within the workstation, without directly comparing the original load amplitude, thereby reducing the direct dependence of anomaly judgment on material batch variations and channel gain variations; when the channel gain exceeds the calibration range, the corresponding element is marked as invalid. If the number of non-zero elements in the part-level comparison set reaches the set number and the directions are consistent, but the directions of these elements are inconsistent in the workstation-level comparison sets to which they belong, it is judged as a part-level anomaly, the cause of which points to the cutting length or material state of the blank itself; if the number of non-zero elements is less than the set number, no part-level judgment is made, and the corresponding blank is transferred to the re-inspection sequence. The reason is that the number of elements in the part-level comparison set is at most equal to the total number of workstations. When the number of elements is too small, all directions are required to be consistent. The probability of all directions being consistent by chance under pure noise cannot be ignored. The set number is the minimum number that makes this probability lower than the set misjudgment rate and does not exceed the total number of workstations. If the non-zero elements in the station-level comparison set have the same direction, but the directions in the part-level comparison sets to which these elements belong are inconsistent, it is judged as a station-level anomaly, and the station-level anomaly direction is recorded. The cause points to the mold or interface state of that station. If the directions in two sets of comparison sets are simultaneously consistent, and the cause does not belong to a particular blank or a particular station, it is judged as a machine-wide factor, and no part-level rejection is performed; only the material baseline is re-evaluated. If neither set has any cross-boundary elements, it is recorded as no anomaly found. If the directions in both sets are inconsistent, either set has insufficient valid elements, or the same primary key or station has opposite directions, it is recorded as pending confirmation and transferred to re-inspection. When part-level and station-level conditions occur concurrently, they are recorded separately, and a composite anomaly is marked at the intersection.

[0059] When the blank is cut too short, all stations it passes through are judged to be in normal position and a lag mark is set. The load displacement integral characteristics of each station are consistently low within its part-level comparison set, but the directions are inconsistent within the comparison sets of the stations involved. This is judged as a part-level anomaly, and the diversion time is scheduled according to its discharge stroke. When the mold of a certain station is worn, the load displacement integral characteristics of each blank in the rolling window of this station are consistently high within its station-level comparison set, but the directions are inconsistent within the comparison sets of the parts involved. This is judged as a station anomaly, and the station-level anomaly direction is recorded. This direction, together with the event accumulation amount and temperature rise rate of this station, enters the three-branch judgment of lubrication deterioration.

[0060] This step delivers two items downstream: the primary key set of part-level anomalies, which enters the diversion stage along the branch of the part; and the direction of station-level anomalies, which enters the lubrication degradation judgment stage along the branch of the station. The data chain splits into two at this point.

[0061] Furthermore, the discharge flight time is calibrated according to the stroke, and after the current stroke is shifted, it is converted into a stroke increment. This increment is added to the discharge stroke in the part flow trajectory table to obtain the diversion stroke. The diversion time is given by the ejection angle interval within the diversion stroke. The diversion object is the blank corresponding to the part-level abnormality and the blank corresponding to the merging mark. If the current stroke changes, the discharge flight time is re-obtained according to the changed stroke and the stroke increment is recalculated. The diversion times that have not yet been executed are rearranged as a whole, and the diversion command is executed.

[0062] Specifically, the branch of the part first needs to resolve the dimensional alignment. The discharge stroke given in the part flow trajectory table is a sequence number, while the lag of the billet in the discharge path after leaving the last station is a time quantity; the two cannot be directly added together. The ejection mechanism at the last station is driven by a crank, with a fixed ejection stroke and a completion time that decreases with each stroke. The initial velocity of the billet when ejected increases with each stroke, so the discharge flight time varies with each stroke and cannot be treated as a fixed value. Instead, the value is taken from the discharge flight time grading table based on the measured stroke position of this stroke, with linear interpolation between grading positions, and then divided by the stroke period of this stroke to convert it into a stroke increment. The stroke increment is added to the discharge stroke of the corresponding primary key in the part flow trajectory table to obtain the diversion stroke. The ejection angle interval within the diversion stroke is used as the diversion angle interval, and the two together constitute the diversion moment, which is written into the diversion queue along with the target primary key.

[0063] The shift value refers to obtaining the discharge flight time of the corresponding gear from the pre-calibrated discharge flight time classification table according to the measured number of strokes in this stroke, based on the calibrated gear to which the actual stroke belongs. The values ​​between two adjacent gears are obtained by linear interpolation. The conversion into stroke increment refers to dividing the obtained discharge flight time by the stroke cycle of this stroke. The quotient is the stroke increment, where the stroke cycle is determined by the time interval between two adjacent bottom dead center zero position pulses.

[0064] The diversion object is determined by the primary key of the part-level anomaly, as well as the primary key of the returned billet bound to the merge flag and the primary key of the newly delivered billet at the same station; only one diversion record is generated when two bound primary keys correspond to the same discharge entity. In the part flow trajectory table, the primary key whose final state is exiting the part flow does not generate a discharge event at the last station, neither occupying the diversion time nor delaying the diversion time of subsequent primary keys. The reason why the diversion proceeding in stroke counting order after the part flow mapping is disrupted is that it includes non-existent discharge events in the sequence. Here, the final state field is used to explicitly skip these events, aligning the diversion queue with the actual discharge sequence at the last station line by line, thus redirecting the removal action back to the true target part.

[0065] If the measured stroke count changes relative to the previous stroke, the feed flight time is first recalculated based on the changed stroke count, then the stroke increment is recalculated, and all unexecuted records in the diversion queue are rearranged. The diversion command is then executed based on the rearrangement result. The stroke count is an explicit independent variable at the diversion time, and this rearrangement applies to stroke count changes from any source, including the stroke reduction command output by this method itself on another branch. After the diversion command is executed, the final state of the corresponding primary key is written back as diverted; primary keys that should have been written back but were not are transferred to the re-inspection sequence. Diversion records are locked once they reach the latest issuance time calculated by back-calculating the actuator's calibrated response time. Stroke count changes only re-arrange unlocked records; if the execution feedback does not return within the calibrated response time, the process stops, proceeds sequentially according to the estimated result, enters the re-inspection phase, and outputs the diversion mechanism maintenance direction.

[0066] Furthermore, the acquisition of cooling and stroke reduction commands includes: accumulating the number of times the part flow trajectory table is inverted as sticking and the number of times it is inverted as stagnation for each station to obtain the station event accumulation; if the station event accumulation, station temperature rise rate, and station-level abnormal direction are all present simultaneously, it is determined that the interface lubrication is deteriorated, and the cooling and stroke reduction commands are output simultaneously; if the station event accumulation increases but the station temperature rise rate does not increase, it is determined that the ejection and transfer links are abnormal, and a maintenance direction is output, but no stroke reduction command is output; after the stroke reduction command is output, the number of stacks, the number of empty chains, and the station event accumulation in the scroll window all fall back simultaneously as the criteria for restoring strokes, and strokes are restored according to the gradient; if they do not fall back simultaneously, a mold change maintenance direction is output.

[0067] Specifically, the event type is taken from the component flow trajectory table of each workstation branch, and the number of events attributed to adhesion and retention is accumulated separately for each workstation and written into the workstation event accumulation table. The scroll window used for accumulation is based on stroke rather than time: if it were based on time, once the number of strokes is reduced, the number of strokes experienced in the same time window decreases, and the accumulated amount will inevitably decrease. The observed value will be determined by the adjustment action of this branch itself, and the subsequent decline criterion will always be true. At the same time, the rate of increase of the temperature sequence of each workstation within the scroll window is recorded and written into the temperature rise rate table; the rate of increase is taken instead of the absolute temperature value because the absolute temperature rises monotonically during the process of the equipment moving from a cold state to thermal equilibrium, which has no discriminatory power, while the sudden increase in interface friction work after the lubricating film breaks is manifested as an increase in the rate of temperature increase. The term "increase" in this article refers to the value of a certain quantity in the current scroll window exceeding the level limited by its set conditions.

[0068] The cumulative amount of station events, the station temperature rise rate, and the station-level anomaly direction from the previous stage converge here, forming three mutually exclusive branches for judgment. Each of the three maintains its independent meaning and is not numerically synthesized. The simultaneous occurrence of the three is judged as interface lubrication deterioration: an increase in the cumulative amount of events corresponds to the loss of parts, an increase in the temperature rise rate corresponds to the increase in frictional work, and the continuous consistency of the station-level anomaly direction corresponds to the systematic impact of the station on all passing blanks. In this paper, "continuous consistency" means that the direction remains unchanged within a continuously set number of rolling windows. The required number of windows is the shortest window spanned by the perceptible change in mold wear at the station. When the three occur simultaneously, interface lubrication deterioration is registered as a candidate cause. Before outputting control commands, the status of material batch, mold wear, station alignment, cooling circuit, and load acquisition path are verified. If any one of them is abnormal, only the corresponding maintenance direction is output. If all items are normal and the three conditions continue to be met in the next rolling window of the same length, then the constrained control branch is entered. If the event accumulation increases but the temperature rise rate does not, and friction work does not increase while parts are still being lost, it is registered as a candidate for an abnormality in the ejection or transfer process, and a maintenance instruction is output, but a reduction in stroke command is not output. If the temperature rise rate increases but the event accumulation does not, it indicates that the thermal state has not yet reached equilibrium. No process control is performed; instead, the nominal phase and normal band of the station are forcibly re-acquired, so that the phase drift caused by thermal expansion is absorbed by the upstream reference and does not flow into the occupancy judgment.

[0069] Upon entering the constrained control branch, the system outputs primary cooling and primary stroke reduction commands according to the existing process formula. The adjustment amount is constrained by temperature, stroke rate, load, and actuator limits. After each adjustment, a scroll window of the same length is observed before deciding whether to maintain, add, or cancel the adjustment. If there is no execution feedback, temperature or load exceeds limits, or the accumulated event continues to rise, the adjustment stops and a maintenance instruction is output. The aforementioned linkage is only the process strategy for this embodiment and does not imply that individual measures are ineffective under other operating conditions. Stroke rate changes are rearranged according to the aforementioned rules for unexecuted records.

[0070] Subsequently, the simultaneous decline of the number of stacks, empty chain counts, and cumulative station event count within the scroll window is used as the criterion for restoring the stroke count. Stroke counts are restored according to a gradient, with the stroke increment for each restoration and the interval between two adjacent restorations selected based on the criterion that no further increase in the cumulative station event count appears within the scroll window after restoration. If the cumulative station event count rises again within the scroll window after a certain level of restoration, it is determined that the stroke increment or interval for that level is too large: the restoration for that level is cancelled, the stroke count is returned to the value before that level of restoration, and the stroke increment for the next level is halved, the interval between levels is doubled, and restoration is attempted again. This process is repeated step by step until two consecutive levels of restoration do not cause any further increase in the cumulative station event count, thus determining the actually usable gradient. In this embodiment, "decline" refers to a value in the scroll window after adjustment being lower than the corresponding value in the scroll window of the same length before adjustment. The restoration criterion does not consider the rate of temperature rise: even if the rate of temperature rise decreases, using it as evidence of effective control constitutes causal confusion. If the three factors do not fall back simultaneously, it is not determined solely by this that a cold welding transfer layer has been formed. Instead, the failure of the control is registered as an anomaly, the stroke is stopped, and a mold change and maintenance instruction is output. The maintenance personnel then review the cavity surface, cooling and lubrication status, and material status.

[0071] After the instruction is issued, the number of stacks, empty chain counts, and cumulative station events are read. Feedback decline only indicates a decrease in event frequency and does not prove that interface stickiness has been eliminated. The method starts from the placeholder criterion, goes through mapping reconstruction and cause separation, and returns to the placeholder criterion to form a closed loop. The criterion for restoring the stroke is also taken from this readback result, rather than from a separately set independent observation.

[0072] During start-up, stop, acceleration, and deceleration, equal-angle sampling ensures that the meaning of the angle index remains unchanged at any speed. The time delay compensation is recalculated stroke by stroke based on the measured angular velocity of the current stroke, so the phase determination during acceleration and deceleration does not cause system offset. When the number of strokes is lower than the set lower limit, the link only executes until framing, without performing occupancy determination or flow routing. During mold change, all references except the angle displacement correspondence table are re-acquired, the occupancy matrix is ​​cleared, and records in the part flow trajectory table that are in the final state are invalidated. The link restarts from the calibration phase. When the acquisition path fails, the entire column of the occupancy matrix for this stroke is set to unjudged. This column does not participate in the connection and termination determination of the empty chain, nor does it participate in any reference refresh. When the number of effective working positions at the starting edge is insufficient, the most recent effective common mode level is used. If multiple consecutive strokes are insufficient, a prompt to check the acquisition path is made. Records in the part flow trajectory table that are in the final state and merged are retained until the flow routing instruction corresponding to their exit stroke is executed before being released, to ensure that the fan-out node is not cleared before all three references are completed.

[0073] Example 2: This example expands upon the data chain described in Example 1, and only details the newly added links. In Example 1, the station difference sequence is obtained by subtracting the material baseline from the single-stroke load frame. The contact start phase is obtained by differential retrieval and time delay compensation. The difference between the contact start phase and the nominal phase is taken as the common mode level according to the rank. The offset direction is determined and projected as normal occupancy, stacking, or empty space, and written into the occupancy matrix. The state of the empty space chain in the occupancy matrix is ​​attributed to the upstream station of the starting point as part dropping, sticking, or stagnation, generating a part flow trajectory table. The station difference sequence is converted and accumulated according to the rotation displacement correspondence table to obtain the load displacement integral feature. Two sets of comparison sets, part level and station level, are constructed according to the part flow trajectory table. The consistency of the deviation direction within the set is used to judge part level abnormality or station abnormality. The part level abnormality and merging flag are converted into the diversion time through the discharge stroke and discharge flight time and the diversion command is output. The simultaneous appearance of the station event accumulation, station temperature rise rate, and station level abnormality direction indicates interface lubrication deterioration, and cooling command and stroke reduction command are output. The three newly added links in this embodiment are respectively connected to the empty projection point, the contact start phase point, and the component drop attribution point of the chain.

[0074] Furthermore, an inherent phase window is set in the return segment after the bottom dead center and in the corner interval after the cavity load is completely unloaded. The inherent phase window corresponds to the corner interval where the inertial force reversal and transmission clearance flip during the slider return stroke. Adjacent differences are made on the samples of the forming load in this stroke within the inherent phase window. The corner positions where the consecutively set number of differences all cross the noise band are searched to obtain the inherent marker phase. If the inherent marker phase exists, the acquisition path of this stroke at this station is recorded as valid. If there is no starting edge and the acquisition path is valid, it is judged as empty. If the inherent marker phase does not exist, the occupancy status of this stroke at this station is recorded as unjudged, not judged as empty, and the update of the strip baseline, nominal phase and normal strip is not included.

[0075] Specifically, under the conditions of first-piece confirmation, calibrated strokes, and thermal stability, if the original load is detected in the return stroke as a load change caused by inertial reversal or gap flipping, which is repeated within the inherent phase window and whose phase dispersion range does not exceed the allowable bandwidth, then the corresponding phase is taken as the inherent marker phase; this step is not activated for models or operating conditions that do not meet the repeatability conditions, and the occupied state is recorded as unjudged.

[0076] The inherent phase window is defined once during the first piece confirmation stage: the latest corner where unloading is completed in all normal occupancy strokes is taken as the unified starting point of the inherent phase window, and the corner before the top dead center is taken as the ending point; a fixed window boundary is used within the calibration conditions; recalibration is performed when changing molds, performing maintenance, exceeding the calibration stroke range, or when the inherent phase marker is not detected consecutively. The reason for fixing the window boundary is that the corner range occupied by the forming section changes with the occupancy state. Unloading is advanced during stacking, and there is no unloading edge during vacancy. If the window is dynamically demarcated according to the stroke, the starting point of the vacancy stroke will be undefined, while the validity of the channel is a prerequisite for the establishment of vacancy. The window boundary is redefined after changing molds.

[0077] The retrieval is performed on samples where the forming load of this stroke is within the inherent phase window, not on the station difference sequence. This is because the strip baseline is taken from the point-by-point median of the normal stationary stroke; the inherent imprint appears in a stable phase in each stroke and also exists within the baseline, thus being completely canceled out after subtraction. However, this step checks whether the channel itself is still transmitting load changes, which is unrelated to the cavity state; therefore, the original sample without subtraction must be used. The retrieval method is the same as the retrieval along the forming section's starting edge: adjacent differences are made along the corner, and the corner position where the first consecutive set number of differences all cross the noise band is taken as the inherent marker phase.

[0078] Furthermore, the inherent nominal phase is obtained by taking the median of the inherent marker phase of the same station within the scroll window; the difference between the inherent marker phase of each station in this stroke and the inherent nominal phase of the station is taken. If the differences of each station have the same sign and fall into the same interval, it is determined that the corner origin is drifting. The corner index origin is shifted according to the median of the differences of each station, and the contact start phase of this stroke is retaken from the shifted origin; if only the differences of individual stations exceed the same interval, it is determined that the force path of that station is loose, the tightening and maintenance direction of that station is output, and the occupancy status of that station in this stroke is recorded as unjudged; if the cumulative shift of the corner index origin exceeds the set upper limit, the maintenance direction of the encoder connector is output and the stop interlock command is output.

[0079] Specifically, if the inherent marker phase exists, the acquisition path of this stroke at that station is recorded as valid; if there is no starting edge within the forming phase window, it is judged as empty. If the inherent marker phase does not exist, the occupancy status of this stroke at that station is recorded as unjudged, not judged as empty, and is not included in the updating of the material baseline, nominal phase and normal belt. This column also does not participate in the connection and termination judgment of the empty chain.

[0080] The reason for this step is that using the absence of a starting edge as a criterion for a void is a negative judgment. Channel open circuits, amplification failures, or sensitive component detachment also manifest as the absence of a starting edge, which will be read as a void and generate a false void chain in the attribution process. This causes the component flow trajectory table to register a non-existent exit record, and the diversion queue will skip the exit position that should be retained, eliminating the pointer and causing a new error. After adding an inherent flag phase, the establishment of a void requires a positive piece of evidence, thus separating channel faults from genuine voids.

[0081] The inherent nominal phase is obtained by taking the median of the inherent marker phase of the same station within the rolling window. The difference between the inherent marker phase of each station in this stroke and the inherent nominal phase of the station is taken. If the difference of each station has the same sign and falls into the same interval (this interval is a consistency criterion interval independently calibrated from the normal belt: in the first piece confirmation stage, under the baseline condition of confirming no coupling looseness or keyway wear, the difference between the inherent marker phase and the inherent nominal phase of each station within the set number of strokes is continuously recorded, and the interval is recorded as a multiple of the discrete range of the difference between each station. The width of this interval is smaller than the width of the normal belt so that the overall drift of the corner origin can be distinguished from the loosening of the force path of a single station), the corner origin is determined to be drifting. The corner index origin is shifted according to the median of the difference of each station, and the contact start phase of this stroke is retaken from the shifted origin. This judgment is only valid when the temperature rise rate at each station is below the lifting condition. For shifts of the same sign occurring when the temperature rise rate is above the lifting condition, this is classified as thermal imbalance, and the origin of the rotation index is not shifted. Instead, the nominal phase of the corresponding station is recalculated to match the normal phase. The reason for this exclusion condition is that the transmission clearance changes with temperature, and the inherent marked phase of each station also shifts of the same sign during the overall machine temperature rise. If this is confused with the judgment of loose couplings and the origin is repeatedly shifted, the nominal phase will be disrupted.

[0082] The rationale is as follows: a loose coupling between the encoder and the spindle, or wear on the keyway, will cause the zero-position pulse to drift relative to the actual bottom dead center, resulting in a shift in all phase criteria. This pulls the normal band off course, causing all stations to misjudge simultaneously. The inherent marker phase is generated by the mechanism itself; once the origin drifts, the inherent marker phases of each station will change synchronously and in the same direction. If only a few stations deviate from this range, it is not an origin problem but rather a loose force path between the mold base and support column at that station. This indicates a need for tightening and maintenance at that station, and the station's current stroke is recorded as unjudged.

[0083] The cumulative translation of the corner index origin is set to an upper limit. Exceeding the upper limit will output the maintenance direction of the encoder connector and output a stop interlock command. If only software translation is used to follow, a monotonously developing mechanical wear will be continuously masked until the connector fails and the encoder disengages from the spindle. At this point, all phase criteria will collapse simultaneously without any prior warning.

[0084] Furthermore, the station lag count is obtained by accumulating the number of lag flags of strokes not judged as part-level anomalies in the station occupancy matrix; the transfer link drop count is obtained by accumulating the number of times dropped parts are inverted from the part flow trajectory table in the transfer link; the scroll windows used for the above two accumulations are based on strokes, not time; when the drop count and the downstream station lag count rise simultaneously, it is registered as a candidate for handover timing drift; only when the transfer mechanism has an adjustable phase interface and position feedback, and the correction direction has been calibrated, the correction command is output according to the calibrated direction and set step size; when there is no direct feedback condition, only the maintenance direction is output, and automatic correction is not performed; when the cumulative correction amount exceeds the set upper limit, the correction is stopped and the maintenance direction of the transfer link is output; when the drop count rises but the lag count does not rise, it is registered as a candidate for insufficient clamping constraint, the maintenance direction of the transfer link is output, and no transfer phase correction command is output; the simultaneous drop of the station lag count and the drop count of the transfer link is used as the criterion for the completion of correction, and the reduction stroke command is canceled after the drop.

[0085] Specifically, in the placeholder matrix, strokes with a lagging offset direction are marked with a lag flag, indicating insufficient build-up height within the cavity. The lag flag is initially recorded as pending during placeholder projection and backfilled after the part-level determination of the billet is given: those judged as part-level anomalies are removed from the pending list, while those not judged as part-level anomalies are added to the station lag count; the backfill delay does not exceed the number of strokes experienced by the billet from the current station to the last station. The number of dropped parts is obtained by accumulating the number of times the part flow trajectory table is inverted according to the transfer link, resulting in the transfer link dropped part count. The scroll windows used for both accumulations are based on strokes, not time: if based on time, the reduction stroke command output during the correction period of this link would reduce the number of strokes experienced within the same time window, inevitably decreasing both counts, and the subsequent drop criterion would always hold.

[0086] The reason for excluding strokes that have been judged as part-level anomalies is that the cause of lag is not unique. Shorter blank cutting also manifests as lag, and this cause has already been identified in the part-level judgment. If it is included together, the cause of the part will be misread as the cause of the transmission link, causing the correction command to be directed to a mechanism that has not drifted.

[0087] The clamping and handover window for miniature parts is extremely narrow: the parts are light and have low inertia, and once they lose constraint at the moment of handover, they are deflected by oil mist and airflow. The transfer mechanism is driven by the main shaft cam, and wear on the cam and connecting rod causes the handover moment to drift relative to the forming phase; if the handover is too early, the blank has not yet been fully positioned in the cavity and loses constraint. Some parts fall too shallowly and have insufficient stacking height, which manifests as a lag in the downstream station, while other parts fall off directly, which manifests as a dropped part in this transfer stage. The two characteristics are of the same origin.

[0088] Based on this, two judgments are made: if the rise in the drop count of a certain transfer link occurs simultaneously with the rise in the lag count of the downstream station of the same transfer link, it is determined that there is a handover timing drift. A transfer phase correction command is output according to the set step size. The correction amount is measured by the phase of the transfer cam relative to the spindle rotation angle, and the direction is to delay the handover time. During the correction, a reduction stroke command is output simultaneously. The cumulative correction amount is compared with the phase margin left when the clamping jaws exit the envelope (the clamping jaws exiting the envelope refers to the area occupied by the clamping jaws of the transfer mechanism during the process of releasing the constraint on the blank, exiting the cavity, or moving into the interference area of ​​the next station). The angular movement range; during the calibration phase, the transmission mechanism is run unloaded at a set speed, and the position trajectory of the clamping jaws relative to the transmission cam's rotation angle is recorded. The angular difference between the exit termination phase and the start phase of the next action in this trajectory is recorded as the phase margin of this transmission link. During calculation, the cumulative correction amount is converted according to the same phase reference of the transmission cam's rotation angle relative to the spindle and compared with this phase margin. An upper limit is set (taking a fraction of this phase margin to leave a safety margin). Calculation is performed on the same reference. If the upper limit is exceeded, the correction stops and the maintenance direction of this transmission link is output. If the part drop count increases but the lag count does not increase, the part loses its constraint only after it has been in place. The cause is insufficient clamping constraint rather than timing drift. Only the maintenance direction of this transmission link is output, and no transmission phase correction command is output.

[0089] After each correction step is output, the command response and phase position feedback of the phase adjustment mechanism are read to confirm the actual adjustment amount. The inherent flag phase is only used to check the consistency of the load acquisition path and the rotation angle index, and is not used to confirm the phase of the transmission cam. If no direct feedback is obtained or the feedback amount is inconsistent with the command, automatic correction is stopped and the maintenance direction of the transmission link is output. The correction is completed when the station lag count and the part drop count of the transmission link both fall back simultaneously. After the fall, the reduction stroke command is canceled.

[0090] Those skilled in the art will readily understand that the above description is merely a preferred embodiment of the present invention and is not intended to limit the present invention. Any modifications, equivalent substitutions, and improvements made within the spirit and principles of the present invention should be included within the scope of protection of the present invention.

Claims

1. A method for identifying anomalies in the manufacturing process of precision fasteners based on edge computing, characterized in that, include: Using the crank angle as a reference, the forming load at each station is collected stroke by stroke, and the starting angle of the forming load relative to the material baseline is extracted to obtain the contact start phase. Based on the offset direction of the contact start phase relative to the nominal phase, each stroke of each station is judged as normal occupation, overlapping or empty, and arranged into an occupation matrix. Search for empty space chains in the occupancy matrix along the synchronously increasing direction of station number and stroke number. Based on whether there is a stacking position upstream of the station starting point of the empty space chain and the number of strokes continuing from the stacking position, invert it into part drop, sticking, or retention, and generate a part flow trajectory table. Based on the part flow trajectory table, the load displacement integral characteristics of the same blank across workstations and the same workstation across blanks, starting from the contact start phase, are used to form part-level comparison sets and workstation-level comparison sets. The consistency of the deviation direction within the set is judged as part-level anomaly or workstation anomaly. Based on the part flow trajectory table, the diversion time is obtained, and the diversion command is output to the blank corresponding to the part-level anomaly and the merging flag. The interface lubrication deterioration is determined according to the station-level anomaly direction and station temperature rise rate, and the cooling command and the reduction stroke command are output.

2. The method for identifying anomalies in the production process of precision fasteners based on edge computing according to claim 1, characterized in that: The contact start phase acquisition process is as follows: After the first piece passes inspection, the median of the forming load at the same station within a continuously set number of strokes is calculated point by point according to the rotation angle, and the calculation result is determined as the initial value of the strip baseline; the median of the forming load at each stroke of the same station that is judged to be in normal position within the rolling window is calculated point by point according to the rotation angle, and the strip baseline is updated stroke by stroke; the forming load at each station is subtracted from the strip baseline of the corresponding station point by point to obtain the station difference sequence; adjacent differences are performed on the station difference sequence along the rotation angle direction, and the first rotation angle position corresponding to the difference that exceeds the noise band for a continuously set number of strokes is retrieved; the fixed time lag of the acquisition path is converted into the rotation angle increment according to the current stroke, and the rotation angle increment is subtracted from the first rotation angle position to obtain the contact start phase.

3. The method for identifying anomalies in the production process of precision fasteners based on edge computing according to claim 2, characterized in that: For each stroke at the same workstation that is determined to be in normal position within the scrolling window, the median of the contact start phase is calculated to obtain the nominal phase of the corresponding workstation. Based on the discrete range of the contact start phase within the same scrolling window, the normal zone of each workstation is determined. The phase difference between the contact start phase of each workstation within the same stroke and the nominal phase of the corresponding workstation is calculated. The phase differences are sorted by rank according to their advance, and the phase difference corresponding to the median rank is determined as the common mode level of the current stroke. The offset direction is determined according to the order of the phase difference of each workstation relative to the common mode level. Workstations with the offset direction in the middle are determined to be in normal position. If the offset direction is advanced and the phase difference exceeds the normal band, it is judged as overlapping; if the offset direction is lagging and the phase difference exceeds the normal band, it is judged as normal occupancy and the lag flag is recorded. If no starting edge is detected, it is determined to be an empty space; Strokes with hysteresis are included in the part-level comparison set according to the correspondence in the placeholder matrix.

4. The method for identifying anomalies in the production process of precision fasteners based on edge computing according to claim 3, characterized in that: Each newly identified empty cell is used as a seed cell, and subsequent empty cells are connected along the direction where the station number and stroke number increase synchronously to obtain an empty cell chain and its starting point. When the upstream station corresponding to the starting point of the empty cell chain is normally occupied within the initial stroke, it is determined by inversion to be a dropped part. When the upstream station is stacked within the initial stroke and the stacking does not continue into subsequent strokes, it is determined by inversion to be a stuck part. For the purpose of retention; when the number of continuous strokes in the stack exceeds the set upper limit, a stop interlock command is output; the part flow trajectory table uses the cut-off stroke number as the primary key, and registers the event type, the arrival stroke of each station, the merging flag, the final state, and the discharge stroke; when the inversion result is adhesion, a binding relationship is established between the primary key of the returned billet and the primary key of the newly fed billet in the same station, and the merging flag is registered; when the inversion result is part dropping or retention, the final state of the corresponding billet is recorded as exiting the part flow, and the discharge stroke is not registered.

5. The method for identifying anomalies in the production process of precision fasteners based on edge computing according to claim 4, characterized in that: Based on the correspondence between the rotation angle and the slider displacement, the displacement increment corresponding to each sampling point in the rotation angle interval from the contact start phase to the bottom dead point is determined, and the station difference sequence is integrated and accumulated point by point according to the corresponding displacement increment to obtain the load displacement integral characteristics. Based on the rows of the part flow trajectory table, the load-displacement integral features of the same billet at each arrival stroke are extracted to form a part-level comparison set; based on the columns of the part flow trajectory table, the load-displacement integral features of each billet at the same station within the scroll window are extracted to form a station-level comparison set; billets with a merge flag are included in the station-level comparison set; the median of each comparison set is calculated and an insensitive zone is set, elements falling into the insensitive zone are marked as zero, and the remaining elements are marked as the corresponding deviation direction; when the number of non-zero elements in the part-level comparison set reaches a set number and the deviation directions are consistent, while the deviation directions of non-zero elements in the station-level comparison set are inconsistent, it is judged as a part-level anomaly; When the number of non-zero elements does not reach the set number, the corresponding billet will be transferred to the re-inspection sequence; When the deviation directions of non-zero elements in the workstation-level comparison set are consistent, but the deviation directions of non-zero elements in the part-level comparison set are inconsistent, it is determined to be a workstation abnormality, and the direction of the workstation-level abnormality is recorded; when the deviation directions of non-zero elements in both the part-level comparison set and the workstation-level comparison set are consistent, it is determined to be a machine-wide factor.

6. The method for identifying anomalies in the production process of precision fasteners based on edge computing according to claim 5, characterized in that: The discharge flight time is calibrated according to the number of strokes. The discharge flight time of the corresponding grade is selected according to the current stroke and converted into a stroke increment. The stroke increment is added to the discharge stroke in the part flow trajectory table to obtain the diversion stroke. The diversion time is determined according to the ejection angle interval within the diversion stroke. The blanks corresponding to part-level anomalies and blanks corresponding to merging flags are identified as diversion objects; when the current stroke changes, the discharge flight time of the corresponding gear is reselected according to the changed stroke, and the stroke increment is recalculated. The diversion times that have not yet been executed are rearranged as a whole, and the diversion command is executed.

7. The method for identifying anomalies in the production process of precision fasteners based on edge computing according to claim 5, characterized in that: The generation process of the cooling command and the stroke reduction command includes: accumulating the number of times the part flow trajectory table is inverted into sticking and retention according to the workstation, respectively, to obtain the workstation event accumulation; when the workstation event accumulation increases, the workstation temperature rise rate increases, and the workstation-level abnormal direction occurs continuously and simultaneously, it is determined that the interface lubrication is deteriorated, and the cooling command and stroke reduction command are output; when the workstation event accumulation increases but the workstation temperature rise rate does not increase, it is determined that the ejection link and the transfer link are abnormal, and the maintenance direction is output without outputting the stroke reduction command; after outputting the stroke reduction command, the number of stacks, the number of empty chains, and the workstation event accumulation in the scroll window simultaneously fall back as the criteria for restoring strokes, and the strokes are restored step by step according to the gradient; if they do not fall back simultaneously, the mold change maintenance direction is output.

Images