A method for predictive maintenance of manufacturing equipment based on edge computing

Abstract

This invention relates to the field of predictive maintenance of manufacturing equipment and discloses a predictive maintenance method for manufacturing equipment based on edge computing. This method is executed by edge nodes deployed on the manufacturing equipment field side. First, it collects equipment status data, control and process data, and quality proxy data. Then, it generates a joint index of manufacturing conditions based on process stage, load range, recipe number, changeover status, and cycle time segment, and calls the corresponding health baseline data. Subsequently, it determines the reliability of the sampled data, and calculates the degradation evidence value of the target component when the reliable execution conditions are met. When the degradation evidence value is in the pending confirmation range, it executes the corresponding preset diagnostic action using an idle window and updates the degradation evidence value based on the response data. Finally, it determines the maintenance level and recommended maintenance window by combining production cycle time parameters and the available status of spare parts, and outputs maintenance instructions or maintenance warnings. This method can improve the accuracy of degradation identification and the executability of maintenance decisions.

Description

Technical Field

[0001] This invention relates to the field of predictive maintenance of manufacturing equipment, and more specifically to a predictive maintenance method for manufacturing equipment based on edge computing. Background Technology

[0002] Manufacturing equipment typically undergoes multiple operational stages during continuous production, including loading, clamping, cutting, unloading, and changeover. The load levels, cycle times, and actions vary significantly across these stages. As operating time increases, components such as the spindle, feed axis, and clamping mechanism may gradually experience wear, increased clearance, lubrication degradation, or sluggish response, thereby affecting equipment stability and processing quality.

[0003] Existing predictive maintenance solutions typically collect operational data such as vibration, temperature, current, pressure, and location through controllers, sensors, or upper-level systems, and combine this data with empirical thresholds, statistical analysis, or predictive models to determine whether there are any abnormalities in the equipment. Some solutions also incorporate dimensional inspection results, defect inspection results, or maintenance records, and then output early warning information or maintenance recommendations.

[0004] Because existing solutions often process data from different process stages, load ranges, changeover states, and cycle time segments in a unified manner, the state fluctuations formed during equipment switching under normal operating conditions can cause the operating data of the same component to lose comparability at different stages, thereby destroying the basis for degradation identification. This makes it difficult for abnormal judgment results to accurately correspond to the actual degraded components, ultimately resulting in false alarms, missed alarms, and inaccurate maintenance timing. Summary of the Invention

[0005] To address the shortcomings of existing technologies, this invention provides a predictive maintenance method for manufacturing equipment based on edge computing. This method solves the technical problems in existing technologies where operating data from different process stages, load ranges, and changeover states are processed uniformly, leading to fluctuations in normal operating conditions that undermine the criteria for degradation judgment. Consequently, the results of anomaly identification are difficult to accurately correspond to actual degraded components, resulting in false alarms, missed alarms, and inaccurate maintenance timing.

[0006] The above-mentioned technical objective of the present invention is achieved through the following technical solution: A predictive maintenance method for manufacturing equipment based on edge computing, applied to edge nodes deployed on the field side of the manufacturing equipment, the method comprising: S1: Collect equipment status data, control and process data, and quality agent data of the manufacturing equipment within the current sampling window; S2: Based on the process stage, load range, formula number, changeover status and cycle time segment within the current sampling window, generate a joint index of manufacturing conditions corresponding to the current sampling window, and retrieve the health baseline data corresponding to the joint index of manufacturing conditions. S3: Based on the sampling integrity, time synchronization deviation, cross-sensor physical consistency, and quality proxy data synchronization within the current sampling window, determine the data reliability index of the current sampling window. When the data reliability index is not lower than the reliability threshold, perform degradation analysis on the current sampling window. When the data reliability index is lower than the reliability threshold, mark the current sampling window as a window to be confirmed. S4: When performing degradation analysis, based on the deviation of equipment status data from health baseline data, the deviation of quality proxy data from health baseline data, and the time correlation between equipment status deviation sequence and quality proxy deviation sequence, combined with the pre-established component correlation relationship, the target component and the corresponding degradation evidence value are determined. S5: When the degradation evidence value is in the pending confirmation range, determine the idle window corresponding to the manufacturing equipment, and issue a preset diagnostic action command to the target component within the idle window, so that the target component performs the preset diagnostic action and generates response data. Update the degradation evidence value according to the deviation relationship between the response data and the health baseline data. S6: Based on the updated degradation evidence value, current production cycle parameters, and spare parts availability, determine the maintenance level and recommended maintenance window for the target component; S7: Outputs maintenance instructions or maintenance warnings corresponding to the maintenance level and recommended maintenance window from the edge node.

[0007] Preferably, the equipment status data includes vibration RMS value, temperature rise rate, current RMS value, pressure fluctuation value, and position following error; the control and process data includes process stage, load range, formula number, changeover status, and cycle time segment; and the quality proxy data includes dimensional deviation statistics, surface defect count, single-piece processing time offset, rework rate fluctuation, and scrap rate fluctuation.

[0008] Preferably, in step S2, generating a joint index of manufacturing conditions corresponding to the current sampling window includes: The process stage is encoded as the first field, the load range as the second field, the formula number as the third field, the changeover status as the fourth field, and the cycle time segment as the fifth field. The first, second, third, fourth, and fifth fields are combined according to a fixed field order to obtain the manufacturing condition composite index; Health baseline data includes historical equipment status baseline data, historical quality proxy baseline data, and historical time-related baseline data, which are associated with the manufacturing condition joint index.

[0009] Preferably, in step S3, determining the data reliability index for the current sampling window includes: The sampling completeness score is determined based on the number of valid sampling points and the number of sampling points that should be sampled within the current sampling window. The time synchronization score is determined based on the maximum time deviation after multi-source data alignment within the current sampling window and the maximum allowable time deviation. Based on whether the actual changes in the status parameters of each device within the current sampling window are consistent with the pre-established physical relationships, the cross-sensor physical consistency score is determined. The quality proxy data synchronization score is determined based on the difference between the actual lag of the quality proxy data corresponding to the current sampling window and the normal lag of the historical time-related baseline data. Based on the sampling integrity score, time synchronization score, cross-sensor physical consistency score, and quality proxy data synchronization score, the data reliability index is determined according to the weighted rules obtained from offline calibration.

[0010] Preferably, the component association relationship is a correspondence table between component identifiers and monitoring parameter identifiers, determining the degradation evidence value corresponding to the target component, including: The deviation of equipment status is determined based on the difference between the equipment status data and the historical equipment status baseline data and the dispersion of the historical equipment status baseline data; The deviation of the quality agency is determined based on the difference between the quality agency data and the historical quality agency baseline data and the dispersion of the historical quality agency baseline data; The time correlation quantity is determined based on the degree of correlation between the equipment status deviation sequence and the quality agent deviation sequence within the allowable lag interval; Based on the correspondence table, equipment status deviation, quality agent deviation, and time correlation, the degradation evidence value is determined according to the evidence fusion rules obtained from offline calibration.

[0011] Preferably, the interval to be confirmed is limited by a first evidence threshold and a second evidence threshold, and the first evidence threshold is less than the second evidence threshold; When the degradation evidence value is less than the first evidence threshold, the target component is determined to be in a non-maintenance state; When the value of degraded evidence is greater than or equal to the first evidence threshold and less than the second evidence threshold, the value of degraded evidence will be determined to be in the range to be confirmed. When the value of degraded evidence is greater than or equal to the second evidence threshold, proceed directly to step S6.

[0012] Preferably, in step S5, determining the idle window corresponding to the manufacturing equipment includes: Read the changeover status and cycle time segment from the control and process data; When the changeover status is a changeover shutdown status, the corresponding time period will be designated as an idle window; When the cycle segment is the interval between adjacent processing cycles, and the duration of the interval is not less than the execution time of the diagnostic action, the corresponding time period is determined as the idle window. Before designating the corresponding time period as an idle window, verify the safety conditions, which include the spindle not executing cutting commands, the feed axis being in a safe position, the hydraulic clamping state being stable, and the tool not being in the workpiece contact area.

[0013] Preferably, the preset diagnostic action command is selected according to the execution method of the target component; When the target component is a linear actuator, the preset diagnostic action is a small reciprocating displacement action within the safe range; When the target component is a rotary actuator, the preset diagnostic action is a short-time no-load speed scan. When the target component is a pressure actuator, the preset diagnostic action is a short-term pressure holding action; The response data includes startup response data, steady-state response data, and recovery response data.

[0014] Preferably, in step S5, updating the degradation evidence value based on the deviation between the response data and the health baseline data includes: The startup response deviation is determined based on the difference between the startup response data and the historical active diagnostic baseline data. The steady-state response deviation is determined based on the difference between the steady-state response data and the historical active diagnostic baseline data. The deviation of the recovery response is determined based on the difference between the recovery response data and the historical active diagnostic baseline data; Based on the degradation evidence value before the update, the deviation of the startup response, the deviation of the steady-state response, and the deviation of the recovery response, the updated degradation evidence value is determined according to the update fusion rules obtained from offline calibration.

[0015] Preferably, the maintenance levels include Level 1 maintenance, Level 2 maintenance, and Level 3 maintenance; The maintenance level is determined based on the updated degradation evidence value, the third interval threshold, and the fourth interval threshold, with the third interval threshold being less than the fourth interval threshold. When the updated degradation evidence value is less than the threshold of the third interval, the maintenance level is determined to be Level 1 maintenance, and the recommended maintenance window is determined to be the next replacement downtime period. Level 1 maintenance corresponds to online inspection confirmation and lubrication status review. When the updated degradation evidence value is greater than or equal to the third interval threshold and less than the fourth interval threshold, the maintenance level is determined to be Level 2 maintenance, and the recommended maintenance window is determined to be the first idle window after the current batch ends. Level 2 maintenance corresponds to planned downtime adjustment, lubrication replenishment, or replacement of worn parts. When the updated degradation evidence value is greater than or equal to the threshold of the fourth interval, the maintenance level is determined to be Level 3 maintenance, and the recommended maintenance window is determined to be the current idle window. Level 3 maintenance corresponds to immediate shutdown and maintenance. When the availability status of spare parts changes to unavailability, the recommended maintenance time windows for Level 1 and Level 2 maintenance are adjusted to the first idle window after the spare parts arrive, and a spare parts missing prompt is output for Level 3 maintenance.

[0016] In summary, the present invention has the following main beneficial effects: This application simultaneously collects equipment status data, control and process data, and quality proxy data at the edge node side. It then generates a joint index of manufacturing conditions based on process stage, load range, recipe number, changeover status, and cycle time segment. This accurately maps the current sampling window to the corresponding health baseline, preventing normal fluctuations under different conditions from being misjudged as equipment degradation. Furthermore, by assessing the reliability of sampling completeness, time synchronization deviation, cross-sensor physical consistency, and quality proxy data synchronization before degradation analysis, low-reliability data such as sensor disconnections, timing misalignments, and changeover disturbances are preemptively eliminated, thereby improving the accuracy and stability of degradation identification results.

[0017] By pre-establishing component relationships, the deviation of equipment status, the deviation of quality agent, and the time correlation between the two are collectively transformed into degradation evidence values ​​for the target component. This achieves the goal of shifting from machine-level anomaly judgment to component-level degradation localization, making degradation identification results closer to the actual fault location. Furthermore, when the degradation evidence value is in the pending confirmation range, corresponding preset diagnostic action commands are issued to the target component during the idle window of the manufacturing equipment. This achieves the effect of secondary confirmation of suspected degradation status without interfering with normal production, thereby reducing false alarms and missed alarms caused by relying solely on passively collected data and improving the reliability of predictive maintenance results.

[0018] By combining updated degradation evidence values, current production cycle parameters, and spare parts availability at the edge nodes to determine maintenance levels and recommended maintenance windows, the goal is to directly generate actionable maintenance decisions on the manufacturing floor. This avoids the problem of simply outputting abstract warning information without timely implementation. Furthermore, by assigning Level 1, Level 2, and Level 3 maintenance to different maintenance timings and handling methods, the maintenance schedule is coordinated with batch production rhythms, changeover downtime periods, and spare parts readiness status. This ensures equipment reliability while reducing the risk of production interruptions due to unplanned downtime, improving the continuous operation capability of manufacturing equipment and maintenance management efficiency. Attached Figure Description

[0019] Figure 1 This is a flowchart of the method of the present invention; Figure 2 This is a block diagram of industrial predictive maintenance according to the present invention. Detailed Implementation

[0020] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0021] Example 1 refer to Figure 1 A predictive maintenance method for manufacturing equipment based on edge computing, applied to edge nodes deployed on the field side of the manufacturing equipment, the method comprising: S1: Collect equipment status data, control and process data, and quality agent data of the manufacturing equipment within the current sampling window; S2: Based on the process stage, load range, formula number, changeover status and cycle time segment within the current sampling window, generate a joint index of manufacturing conditions corresponding to the current sampling window, and retrieve the health baseline data corresponding to the joint index of manufacturing conditions. S3: Based on the sampling integrity, time synchronization deviation, cross-sensor physical consistency, and quality proxy data synchronization within the current sampling window, determine the data reliability index of the current sampling window. When the data reliability index is not lower than the reliability threshold, perform degradation analysis on the current sampling window. When the data reliability index is lower than the reliability threshold, mark the current sampling window as a window to be confirmed. S4: When performing degradation analysis, based on the deviation of equipment status data from health baseline data, the deviation of quality proxy data from health baseline data, and the time correlation between equipment status deviation sequence and quality proxy deviation sequence, combined with the pre-established component correlation relationship, the target component and the corresponding degradation evidence value are determined. S5: When the degradation evidence value is in the pending confirmation range, determine the idle window corresponding to the manufacturing equipment, and issue a preset diagnostic action command to the target component within the idle window, so that the target component performs the preset diagnostic action and generates response data. Update the degradation evidence value according to the deviation relationship between the response data and the health baseline data. S6: Based on the updated degradation evidence value, current production cycle parameters, and spare parts availability, determine the maintenance level and recommended maintenance window for the target component; S7: Outputs maintenance instructions or maintenance warnings corresponding to the maintenance level and recommended maintenance window from the edge node.

[0022] In step S1, the edge node collects equipment status data, control and process data, and quality proxy data of the manufacturing equipment within the current sampling window.

[0023] The equipment status data includes RMS vibration values, temperature rise rate, RMS current values, pressure fluctuation values, and position tracking error. RMS vibration values ​​are collected by vibration sensors mounted on the spindle housing and feed axis support; the temperature rise rate is collected by temperature sensors mounted on the spindle bearing housing, servo motor housing, and the outer wall of the hydraulic circuit; the RMS current values ​​are output by the spindle driver and servo driver; the pressure fluctuation values ​​are collected by pressure sensors in the hydraulic clamping circuit; and the position tracking error is determined by the difference between the commanded position output by the CNC controller and the position feedback value.

[0024] The control and process data include process stage, load range, recipe number, changeover status, and cycle time segment. The process stage is determined by the process segment number currently being executed by the CNC controller; the load range is determined by the range in which the spindle load rate and feed axis load rate are located; the recipe number is determined by the process recipe file corresponding to the current machining program; the changeover status is output by the manufacturing execution system; and the cycle time segment is determined by the current machining cycle segment in the loading, clamping, cutting, unloading, and idling waiting phases.

[0025] The quality control data includes dimensional deviation statistics, surface defect counts, single-piece processing time offset, rework rate fluctuation, and scrap rate fluctuation. Dimensional deviation statistics are output from downstream online measurement stations; surface defect counts are output from visual inspection stations; single-piece processing time offset is determined by the difference between the current processing cycle time and the baseline processing time under the same conditions; rework rate fluctuation and scrap rate fluctuation are output by the manufacturing execution system according to batch statistics.

[0026] To ensure that multi-source data can be processed on a unified time basis, edge nodes use the CNC controller clock as the master clock and append a unified timestamp to all collected data. The current sampling window preferably uses a complete processing cycle; however, when a certain process stage lasts for a long time and the processing status is stable, a fixed-duration rolling window can also be used. Regardless of the partitioning method, the start and end boundaries of the current sampling window are aligned with the CNC controller clock.

[0027] All data sources are based on data links available at the manufacturing site and do not rely on fictitious data. Driver current, controller cycle time, pressure data, and position feedback are raw operating data from the equipment side; dimensional deviations, surface defects, rework rates, and scrap rates are existing outputs in the production line quality chain.

[0028] In step S2, the edge node generates a manufacturing condition joint index corresponding to the current sampling window based on the process stage, load range, recipe number, changeover status, and cycle time segment within the current sampling window, and retrieves the health baseline data corresponding to the manufacturing condition joint index.

[0029] Specifically, the process stage is encoded as the first field, the load range as the second field, the formula number as the third field, the changeover status as the fourth field, and the cycle time segment as the fifth field. Then, the first, second, third, fourth, and fifth fields are combined according to a fixed field order to obtain the manufacturing condition joint index. This fixed field order remains unchanged within the same equipment model, the same control system configuration, and the same process line.

[0030] In this embodiment, a manufacturing condition joint index is used instead of a single global threshold or a unified prediction model because the operating data of manufacturing equipment varies significantly across different process stages, load ranges, and cycle time segments. For example, spindle current and vibration levels are typically higher in the roughing stage than in the finishing stage, and cycle time fluctuations are typically higher in the changeover stage than in the normal cutting stage. If a unified baseline or threshold is used directly without distinguishing between operating conditions, normal operating condition transitions can easily be misjudged as degradation signals. Therefore, this embodiment first maps the current sampling window to the corresponding manufacturing condition joint index and then calls the healthy baseline data under that operating condition, ensuring that degradation analysis is always performed within the same operating condition range.

[0031] The health baseline data is constructed from historical samples collected when the equipment is in a confirmed health state. A confirmed health state is preferably a state where the equipment has completed inspections, calibrations, or maintenance, and is in continuous, stable production with downstream quality inspection results meeting standards. For each set of manufacturing condition joint indexes, edge nodes establish historical equipment status baseline data, historical quality proxy baseline data, and historical time-related baseline data. The historical equipment status baseline data includes the mean, dispersion, and allowable fluctuation range of each equipment status parameter under the corresponding operating condition; the historical quality proxy baseline data includes the mean, dispersion, and allowable fluctuation range of each quality proxy parameter under the corresponding operating condition; and the historical time-related baseline data includes the normal lag range for equipment status offsets to propagate to quality proxy offsets.

[0032] To ensure the stability of baseline data under the same operating conditions, healthy baseline data is preferably obtained from the statistical analysis of multiple consecutive healthy batches, rather than from a single batch. For a joint index of the same manufacturing condition, if the number of healthy samples is insufficient, the edge node marks the condition as a baseline incomplete state. Before the baseline samples meet the preset number, the condition will not be directly used for the formal determination of degradation evidence values, but will only be used as a temporary comparison basis to avoid misjudgment caused by insufficient baseline.

[0033] In step S3, the edge node determines the data reliability index of the current sampling window based on the sampling integrity, time synchronization deviation, cross-sensor physical consistency, and quality proxy data synchronization within the current sampling window.

[0034] This implementation method first determines the reliability of the data before degradation analysis, which is a crucial step that distinguishes it from common methods that directly predict data after acquisition. In manufacturing environments, situations such as short-term sensor disconnection, fieldbus refresh delays, natural cycle time lags in downstream quality inspection stations, and instantaneous disturbances caused by model changeovers are quite common. If the reliability of the data is not determined first, missing data, timing misalignments, or operational disturbances can easily be misidentified as equipment degradation.

[0035] The sampling integrity score is determined based on the ratio between the number of valid sampling points and the number of sampling points that should be sampled within the current sampling window. If a sensor experiences packet loss within the current sampling window, the corresponding missing points are not counted in the valid sampling point count.

[0036] The time synchronization score is determined based on the maximum time deviation after aligning multi-source data within the current sampling window. Therefore, the edge node uses the CNC controller clock as the master clock, mapping driver reporting, sensor sampling, and quality inspection station outputs to a unified time axis, and calculating the maximum time deviation within the same sampling window. The maximum allowable time deviation is jointly determined by the controller sampling period, the fieldbus refresh period, and the edge node buffer delay.

[0037] The cross-sensor physical consistency score is determined based on whether the actual changes in the status parameters of each device within the current sampling window are consistent with the pre-established physical relationships. These pre-established physical relationships are not arbitrarily set based on experience, but are jointly determined by the structural relationships of the devices, the driving relationships, and the collaborative change patterns of historical samples under confirmed healthy conditions, and are pre-stored in the edge nodes in the form of a rule table. The rule table includes a rule identifier, associated parameter groups, expected change direction, applicable operating conditions, and judgment result fields. For example, when the cutting load increases and the machining program is not switched, the effective value of the spindle current and the effective value of the spindle vibration will not decrease significantly simultaneously; during hydraulic clamping, there is a stable correspondence between the pressure fluctuation value and the duration of the action. The edge nodes select applicable rules based on the current manufacturing condition joint index and perform consistency checks on the parameter change relationships within the current sampling window.

[0038] The synchronicity score of the quality proxy data is determined based on the difference between the actual lag in the quality proxy data and the normal lag in the historical time-related baseline data. Since the quality inspection station is located downstream of the processing station, the quality proxy data has an objective time lag relative to the equipment status data. This implementation does not require the equipment status data and the quality proxy data to change synchronously at the same time; instead, it treats this lag as a normal characteristic of the manufacturing process and constrains its synchronicity through the normal lag amount.

[0039] The data credibility index is expressed by the following formula: in, Indicates the first The data reliability index for the current sampling window; Indicates the first The sampling integrity score of the current sampling window; Indicates the first The time synchronization score of the current sampling window; Indicates the first The cross-sensor physical consistency score for the current sampling window; Indicates the first The quality proxy data synchronization score for the current sampling window; , , and This represents the weight coefficient for the corresponding score, and the sum of the coefficients is 1.

[0040] The weighting coefficients are obtained through offline calibration. Specifically, confirmed healthy samples, known abnormal samples, and known false positive samples are selected to reduce the false positive rate of healthy samples while maintaining the abnormal sample recognition rate as a constraint. The distinguishing contributions of the four scores are statistically analyzed to determine the weighting rules. The credibility threshold is also determined through offline calibration, with the principle being that the false positive rate of healthy samples is controlled within a preset range, while ensuring that the false negative rate of known abnormal samples does not exceed a preset range.

[0041] When the data credibility index is not lower than the credibility threshold, a degradation analysis is performed on the current sampling window; when the data credibility index is lower than the credibility threshold, the current sampling window is marked as a window to be confirmed, instead of directly outputting a degradation conclusion. This can remove sampling windows with obvious data quality defects from the formal judgment chain.

[0042] In step S4, the edge node determines the target component and the corresponding degradation evidence value based on the device status deviation of the device status data relative to the health baseline data, the quality proxy deviation of the quality proxy data relative to the health baseline data, and the time correlation between the device status deviation sequence and the quality proxy deviation sequence, combined with the pre-established component association relationship.

[0043] The component relationships are established using a correspondence table between component identifiers and monitoring parameter identifiers. This correspondence table is jointly established by the equipment structure list, sensor configuration relationships, and maintenance records, and stored in the edge nodes in a structured table format. This correspondence table remains fixed for the same equipment model. For the vertical CNC machining center in this embodiment, candidate components include the spindle bearing, the feed axis lead screw pair, and the hydraulic clamping unit. The equipment status parameters corresponding to the spindle bearing include the effective value of spindle vibration, the spindle temperature rise rate, and the effective value of spindle current; the corresponding quality proxy parameters include the statistical value of dimensional deviation and the count value of surface defects. The position parameters corresponding to the feed axis lead screw pair include the position following error, the effective value of feed axis current, and the effective value of feed axis vibration; the corresponding quality proxy parameters include the statistical value of dimensional deviation and the offset of single-piece machining time. The equipment status parameters corresponding to the hydraulic clamping unit include the pressure fluctuation value and the temperature rise rate; the corresponding quality proxy parameters include the count value of surface defects, the rework rate fluctuation, and the scrap rate fluctuation.

[0044] To ensure consistent comparison of parameters with different dimensions, edge nodes first normalize the deviation of each parameter from the corresponding operating condition health baseline. The normalized deviation is expressed by the following formula: in, Indicates the first Within the current sampling window, the first Normalized deviation of each parameter; Indicates the first Within the current sampling window, the first The actual values ​​of each parameter; The manufacturing condition composite index is Time Each parameter corresponds to the mean of the health baseline; The manufacturing condition composite index is Time Each parameter corresponds to the dispersion of the health baseline; This represents the lower limit of dispersion, used to prevent distortion of the normalization result due to an excessively small denominator. The lower limit of dispersion is taken as the protection lower limit of the minimum dispersion of stable operating states among healthy samples of similar equipment.

[0045] After obtaining the normalized deviation of each parameter, the edge nodes, according to the component association, summarize the normalized deviations of equipment status parameters belonging to the same candidate component to obtain the equipment status deviation; and summarize the normalized deviations of quality proxy parameters belonging to the same candidate component to obtain the quality proxy deviation. Parameter summarization uses parameter weights obtained from offline calibration. The parameter weights are determined based on the consistency contribution of each parameter to the maintenance conclusion in the pre-maintenance sample, post-maintenance recovery sample, and known degradation sample.

[0046] Considering that device state deviation typically lags behind quality proxy deviation, this implementation further calculates the time correlation between the device state deviation sequence and the quality proxy deviation sequence. The time correlation is represented by the maximum normalized cross-correlation value within the allowable lag interval, as shown in the following formula: in, Indicates the first Target component within the current sampling window Time-related quantities; Indicates target component The maximum allowed hysteresis window length; This indicates the window length used to calculate cross-correlation; Indicates target component The device status deviates from the sequence; Indicates target component Quality agent deviation sequence; and These represent the mean values ​​of the corresponding sequences within the cross-correlation window. The length of the cross-correlation window is determined based on the number of consecutive processing cycles, and the maximum allowable hysteresis window length is determined jointly based on the propagation path of the target component's influence, the workpiece transport time, and the downstream detection cycle time. The degradation evidence value is represented by the fusion result of equipment state deviation, quality proxy deviation, and time correlation, as shown in the following formula: in, Indicates the first Target component within the current sampling window The value of evidence of degradation; Indicates target component The deviation of the equipment status; Indicates target component Quality deviation; , and This represents the corresponding fusion coefficient, and the sum of the coefficients is 1. The fusion coefficient is determined through offline calibration, with the principle of maximizing the distinguishability between the healthy state, the pending confirmation state, and the direct maintenance state.

[0047] For each target component, the edge node pre-sets a first evidence threshold and a second evidence threshold, with the first evidence threshold being less than the second evidence threshold. When the degradation evidence value is less than the first evidence threshold, the target component is determined to be in a non-maintenance state; when the degradation evidence value is greater than or equal to the first evidence threshold and less than the second evidence threshold, the degradation evidence value is determined to be in the pending confirmation range; when the degradation evidence value is greater than or equal to the second evidence threshold, proceed directly to step S6. The first and second evidence thresholds are jointly determined by the statistical results of healthy samples, early degradation samples, and samples that continuously deteriorate before maintenance.

[0048] This implementation does not directly feed all data into a single unified model. Instead, it first locks in a health baseline under the same operating conditions through a joint index of manufacturing conditions. Then, it uses equipment status deviation, quality proxy deviation, and time correlation to form a degradation evidence value, thereby incorporating manufacturing condition differences and downstream quality lag factors into the same judgment chain. This constitutes one of the essential differences from the unified prediction model or single signal threshold judgment route in common existing technologies.

[0049] In step S5, when the degradation evidence value is in the pending confirmation range, the edge node determines the idle window corresponding to the manufacturing equipment and issues a preset diagnostic action command to the target component within the idle window, so that the target component performs the preset diagnostic action and generates response data. The degradation evidence value is updated according to the deviation relationship between the response data and the health baseline data.

[0050] The idle window is determined as follows: The edge node reads the changeover status and cycle time segment from the control and process data. When the changeover status is a changeover stop status, the corresponding time period is designated as an idle window. When the cycle time segment is an interval between adjacent machining cycles, and the duration of this interval is not less than the execution time of the diagnostic action, it is also designated as an idle window. Before designating the corresponding time period as an idle window, the edge node also verifies safety conditions, including that the spindle is not executing a cutting command, the feed axis is in a safe position, the hydraulic clamping state is stable, and the tool is not in the workpiece contact area. Only when all of the above conditions are met simultaneously will the edge node execute the preset diagnostic action within that time period.

[0051] To support proactive diagnostic actions for different target components, this implementation pre-establishes a component diagnostic library within the edge nodes. The component diagnostic library stores at least the correspondence between target component type, corresponding preset diagnostic actions, diagnostic action execution duration, and safe action boundaries. The component diagnostic library presets different types of diagnostic actions based on the execution method of the target component. For linear actuators, the preset diagnostic action is a small-amplitude reciprocating displacement action within a safe range; for rotary actuators, the preset diagnostic action is a short-term unloaded speed scanning action; and for pressure actuators, the preset diagnostic action is a short-term pressure holding action.

[0052] When the target component is a feed axis leadscrew pair, the edge node invokes a micro-amplitude reciprocating displacement action within the safety range. Specifically, during the diagnostic action issuance phase, the edge node sends a micro-amplitude displacement command to the servo driver, causing the feed axis to move a preset distance along the safety direction; during the diagnostic action holding phase, the current position is maintained, and servo current, position following error, and local vibration data are continuously collected; during the diagnostic action release phase, the feed axis is controlled to return to the initial safe position along the original path, and the current change, position following error change, and return stabilization time during the return process are continuously collected. The preset distance is jointly determined by the safety range length, servo resolution, and the safety requirement of not interfering with the workpiece.

[0053] When the target component is the spindle bearing, the edge node invokes a short-time no-load speed scan. Specifically, during the diagnostic action issuance phase, the spindle accelerates from a stationary state to a preset diagnostic speed; during the diagnostic action holding phase, the spindle maintains the preset diagnostic speed under no-cutting load conditions; and during the diagnostic action release phase, the spindle decelerates from the preset diagnostic speed back to a stationary state. The edge node collects spindle current, RMS vibration value, characteristic frequency band vibration amplitude, deceleration coasting time, and temperature rise rate during these three phases. The preset diagnostic speed is determined based on the spindle structure's safety range, no-load operation limitations, and the diagnostic sensitive frequency band.

[0054] When the target component is a hydraulic clamping unit, the edge node invokes a short-term pressure holding action. Specifically, during the diagnostic action issuance phase, a pressure holding command is issued to the hydraulic control unit to establish a preset pressure in the hydraulic circuit; during the diagnostic action holding phase, the preset pressure is maintained, and the pressure holding change, temperature rise change, and valve control response time are continuously collected; during the diagnostic action release phase, the pressure holding state is released, and the pressure recovery time and fall curve are continuously collected. The preset pressure is determined based on the rated operating pressure, safety margin, and no-load conditions of the hydraulic clamping unit.

[0055] For the three types of target components mentioned above, the response data uniformly includes initiation response data, steady-state response data, and recovery response data. Initiation response data reflects the dynamic characteristics at the beginning of the diagnostic action, steady-state response data reflects the stability characteristics during the action's maintenance phase, and recovery response data reflects the recovery characteristics after the diagnostic action is released. Edge nodes compare the response data with historical active diagnostic baseline data under the same manufacturing condition's joint index to obtain the initiation response deviation, steady-state response deviation, and recovery response deviation.

[0056] The updated degradation evidence value is expressed using the following formula: in, Indicates the first Target component within the current sampling window The updated degraded evidence value; Indicates the degraded evidence value before the update; Indicates target component The retention factor; Indicates the deviation from the startup response; This indicates the deviation from the steady-state response; Indicates the recovery of the target component. The retention factor; Indicates the deviation from the startup response; This indicates the deviation from the steady-state response; Indicates the deviation of the recovery response; , and This represents the corresponding fusion coefficient, and the sum of the coefficients is 1. All the above coefficients were obtained through offline calibration. For the feed axis lead screw pair, the recovery phase is more sensitive to mechanical clearance, lubrication status, and local resistance, so the weight of the recovery response deviation is higher; for the spindle bearing, the vibration and characteristic frequency band amplitude in the steady-state phase are more sensitive, so the weight of the steady-state response deviation is higher; for the hydraulic clamping unit, the holding and recovery phases are more sensitive to pressure relief and valve control hysteresis, so the steady-state response deviation and recovery response deviation have higher weights.

[0057] The proactive diagnostic confirmation in this embodiment is not simply adding another test. Instead, it only performs a preset diagnostic action on the target component during a free window that does not affect normal production when the degradation evidence value is in the pending confirmation range. This action updates the original degradation evidence value in response to the data. Compared with common existing technologies that rely solely on passive data collection and directly output fault risks, this embodiment forms a new technical chain through proactive diagnostic confirmation triggered by the pending confirmation range during a free window. This constitutes the second essential difference between this embodiment and existing technologies.

[0058] In step S6, the edge node determines the maintenance level and recommended maintenance window for the target component based on the updated degradation evidence value, the current production cycle parameters, and the availability of spare parts.

[0059] Current production cycle parameters include the number of remaining parts in the current batch, the average processing time per part, the next changeover schedule time, and the duration of the current idle window. Spare parts availability status is output by the spare parts management system, including both available and unavailable statuses.

[0060] Edge nodes pre-set a third interval threshold and a fourth interval threshold for each target component, with the third interval threshold being less than the fourth interval threshold. These thresholds are obtained statistically based on the distribution of the updated degradation evidence values ​​in historical maintenance samples.

[0061] The edge node first determines the maintenance level based on the updated degradation evidence value, then selects the executable time period based on the maintenance level, the current production cycle parameters, and the available status of spare parts, and determines the earliest time period that meets the conditions as the recommended maintenance window.

[0062] When the updated degradation evidence value is less than the threshold of the third interval, the maintenance level is determined to be Level 1 maintenance, and the recommended maintenance window is determined to be the next replacement downtime period. Level 1 maintenance corresponds to online inspection confirmation and lubrication status review.

[0063] When the updated degradation evidence value is greater than or equal to the third interval threshold and less than the fourth interval threshold, the maintenance level is determined to be Level 2 maintenance, and the recommended maintenance window is determined to be the first idle window after the current batch ends. Level 2 maintenance corresponds to planned downtime adjustment, lubrication replenishment, or replacement of worn parts.

[0064] When the updated degradation evidence value is greater than or equal to the threshold of the fourth interval, the maintenance level is determined to be Level 3 maintenance, and the recommended maintenance window is set to the current idle window. Level 3 maintenance corresponds to immediate shutdown and repair.

[0065] When the availability status of spare parts changes to unavailability, the recommended maintenance time windows for Level 1 and Level 2 maintenance are adjusted to the first idle window after the spare parts arrive, and a spare parts missing prompt is output for Level 3 maintenance.

[0066] This approach does not simply output the failure probability or remaining lifespan. Instead, it combines the updated degradation evidence value with the current production cycle parameters and spare parts availability at the edge node to directly generate an executable maintenance level and recommended maintenance window. Compared to existing technologies that only provide fault warnings or generate maintenance plans uniformly from the cloud, this implementation directly completes the maintenance decision-making loop at the edge node, which constitutes the third essential difference from existing technologies.

[0067] In step S7, the edge node outputs maintenance instructions or maintenance warnings corresponding to the maintenance level and recommended maintenance window. These instructions or warnings can be output synchronously through the manufacturing execution system, the equipment's human-machine interface, and the maintenance management system. For Level 1 maintenance, the edge node outputs an inspection confirmation instruction; for Level 2 maintenance, it outputs a planned shutdown maintenance instruction; and for Level 3 maintenance, it outputs an immediate shutdown inspection warning.

[0068] Edge nodes can also record actual maintenance results, post-maintenance recovery status, and subsequent operational status. For target components confirmed by maintenance, corresponding samples before and after maintenance can be included in the offline calibration dataset for periodic updates of weight coefficients, thresholds, and component correlations. However, the above update process does not change the core steps of this implementation method and is only performed during offline maintenance, without affecting the online execution of steps S1 to S7 by the edge nodes.

[0069] Example 2 This embodiment shares the same overall architecture as Embodiment 1, except that the manufacturing equipment is a continuous conveyor assembly equipment. Candidate components include a conveyor servo unit, a rotary indexing unit, and a pneumatic pressing unit. The equipment status data collected by the edge nodes still includes vibration RMS values, temperature rise rate, current RMS values, pressure fluctuation values, and position tracking errors. Control and process data still include process stages, load ranges, recipe numbers, changeover status, and cycle time segments. Quality proxy data still includes dimensional deviation statistics, surface defect counts, single-piece processing time offset, rework rate fluctuations, and scrap rate fluctuations. The determination methods for the manufacturing condition joint index, health baseline data, data reliability indicators, degradation evidence values, idle windows, preset diagnostic actions, and maintenance levels are consistent with Embodiment 1.

[0070] When the target component is a conveyor servo unit, the edge node invokes a micro-amplitude reciprocating displacement action within the safe range during the idle window; when the target component is a rotary indexing unit, the edge node invokes a short-time no-load speed scanning action; when the target component is a pneumatic press unit, the edge node invokes a short-time pressure holding action. Therefore, this invention is not limited to vertical CNC machining centers, nor to any particular equipment structure. As long as the equipment possesses obtainable equipment status data, control and process data, and quality proxy data, and has available idle windows and executable preset diagnostic actions, predictive maintenance can be achieved using the method of this invention.

[0071] In this application, the pre-established physical relationships are jointly determined by the equipment structure relationships, driving relationships, and the collaborative change patterns of historical samples under confirmed health conditions, and are pre-stored in the edge nodes in the form of a rule table. The rule table includes at least a rule identifier, associated parameter group, expected change direction, applicable operating conditions, and judgment result field.

[0072] The weighting rules and reliability thresholds for the data reliability index are obtained through offline calibration. During offline calibration, the sample set is defined as confirmed healthy samples, known abnormal samples, and known false alarm samples. The weighting rules and reliability thresholds for the data reliability index are determined with the constraint of reducing the false positive rate of healthy samples and maintaining the identification rate of abnormal samples.

[0073] The component relationships are established by the equipment structure list, sensor configuration relationships, and maintenance records, and are stored in the form of a correspondence table between component identifiers and monitoring parameter identifiers, which remains fixed under the same equipment model.

[0074] The maximum allowable time deviation is determined by the controller sampling period, the fieldbus refresh period, and the edge node buffer delay; the maximum allowable hysteresis window length is determined by the target component's influence propagation path, the workpiece conveying time, and the downstream detection cycle time; the lower limit of dispersion is taken as the protection lower limit of the minimum dispersion of stable operating states in the healthy samples of similar equipment; the cross-correlation window length is determined based on the number of continuous processing cycles.

[0075] The fusion coefficient of the degraded evidence value, the retention coefficient of the updated degraded evidence value, and the response fusion coefficient were all obtained through offline calibration. The determination principle was to maximize the distinguishability between healthy states, pending confirmation states, and directly maintained states in historical samples.

[0076] The first evidence threshold, the second evidence threshold, the third interval threshold, and the fourth interval threshold are all determined based on the statistical results of healthy samples, early degradation samples, samples that continued to deteriorate before maintenance, and samples that recovered after maintenance.

[0077] The component diagnostic library stores at least the correspondence between the target component type, the corresponding preset diagnostic action, the execution duration of the diagnostic action, and the safe action boundary. For linear actuators, the preset diagnostic action is a small-amplitude reciprocating displacement action within the safe range; for rotary actuators, the preset diagnostic action is a short-time unloaded speed scanning action; and for pressure actuators, the preset diagnostic action is a short-time pressure holding action.

[0078] Although embodiments of the invention have been shown and described, it will be understood by those skilled in the art that various changes, modifications, substitutions and alterations can be made to these embodiments without departing from the principles and spirit of the invention, the scope of which is defined by the appended claims and their equivalents.

Claims

1. An edge-computing-based manufacturing equipment predictive maintenance method characterized by, The method, applied to edge nodes deployed on the field side of manufacturing equipment, includes: S1: Collect equipment status data, control and process data, and quality agent data of the manufacturing equipment within the current sampling window; S2: Based on the process stage, load range, formula number, changeover status and cycle time segment within the current sampling window, generate a joint index of manufacturing conditions corresponding to the current sampling window, and retrieve the health baseline data corresponding to the joint index of manufacturing conditions. S3: Based on the sampling integrity, time synchronization deviation, cross-sensor physical consistency, and quality proxy data synchronization within the current sampling window, determine the data reliability index of the current sampling window. When the data reliability index is not lower than the reliability threshold, perform degradation analysis on the current sampling window. When the data reliability index is lower than the reliability threshold, mark the current sampling window as a window to be confirmed. S4: When performing degradation analysis, based on the deviation of equipment status data from health baseline data, the deviation of quality proxy data from health baseline data, and the time correlation between equipment status deviation sequence and quality proxy deviation sequence, combined with the pre-established component correlation relationship, the target component and the corresponding degradation evidence value are determined. S5: When the degradation evidence value is in the pending confirmation range, determine the idle window corresponding to the manufacturing equipment, and issue a preset diagnostic action command to the target component within the idle window, so that the target component performs the preset diagnostic action and generates response data. Update the degradation evidence value according to the deviation relationship between the response data and the health baseline data. S6: Based on the updated degradation evidence value, current production cycle parameters, and spare parts availability, determine the maintenance level and recommended maintenance window for the target component; S7: Outputs maintenance instructions or maintenance warnings corresponding to the maintenance level and recommended maintenance window from the edge node.

2. The edge computing-based predictive maintenance method for manufacturing equipment according to claim 1, characterized by, The equipment status data includes vibration RMS value, temperature rise rate, current RMS value, pressure fluctuation value, and position following error. The control and process data includes process stage, load range, formula number, changeover status, and cycle time segment. The quality agency data includes dimensional deviation statistics, surface defect count, single-piece processing time offset, rework rate fluctuation, and scrap rate fluctuation.

3. The edge-computing-based predictive maintenance method for manufacturing equipment according to claim 2, characterized by, In step S2, a joint index of manufacturing conditions corresponding to the current sampling window is generated, including: The process stage is encoded as the first field, the load range as the second field, the formula number as the third field, the changeover status as the fourth field, and the cycle time segment as the fifth field. The first, second, third, fourth, and fifth fields are combined according to a fixed field order to obtain the manufacturing condition composite index; Health baseline data includes historical equipment status baseline data, historical quality proxy baseline data, and historical time-related baseline data, which are associated with the manufacturing condition joint index.

4. The edge-computing-based predictive maintenance method for manufacturing equipment according to claim 3, characterized by, In step S3, the data reliability index for the current sampling window is determined, including: The sampling completeness score is determined based on the number of valid sampling points and the number of sampling points that should be sampled within the current sampling window. The time synchronization score is determined based on the maximum time deviation after multi-source data alignment within the current sampling window and the maximum allowable time deviation. Based on whether the actual changes in the status parameters of each device within the current sampling window are consistent with the pre-established physical relationships, the cross-sensor physical consistency score is determined. The quality proxy data synchronization score is determined based on the difference between the actual lag of the quality proxy data corresponding to the current sampling window and the normal lag of the historical time-related baseline data. Based on the sampling integrity score, time synchronization score, cross-sensor physical consistency score, and quality proxy data synchronization score, the data reliability index is determined according to the weighted rules obtained from offline calibration.

5. A predictive maintenance method for manufacturing equipment based on edge computing according to claim 4, characterized in that, The component association relationship is a correspondence table between component identifiers and monitoring parameter identifiers, which determines the degradation evidence value corresponding to the target component, including: The deviation of equipment status is determined based on the difference between the equipment status data and the historical equipment status baseline data and the dispersion of the historical equipment status baseline data; The deviation of the quality agency is determined based on the difference between the quality agency data and the historical quality agency baseline data and the dispersion of the historical quality agency baseline data; The time correlation quantity is determined based on the degree of correlation between the equipment status deviation sequence and the quality agent deviation sequence within the allowable lag interval; Based on the correspondence table, equipment status deviation, quality agent deviation, and time correlation, the degradation evidence value is determined according to the evidence fusion rules obtained from offline calibration.

6. A predictive maintenance method for manufacturing equipment based on edge computing according to claim 5, characterized in that, The interval to be confirmed is limited by a first evidence threshold and a second evidence threshold, and the first evidence threshold is less than the second evidence threshold; When the degradation evidence value is less than the first evidence threshold, the target component is determined to be in a non-maintenance state; When the value of degraded evidence is greater than or equal to the first evidence threshold and less than the second evidence threshold, the value of degraded evidence will be determined to be in the range to be confirmed. When the value of degraded evidence is greater than or equal to the second evidence threshold, proceed directly to step S6.

7. A predictive maintenance method for manufacturing equipment based on edge computing according to claim 6, characterized in that, In step S5, the available window corresponding to the manufacturing equipment is determined, including: Read the changeover status and cycle time segment from the control and process data; When the changeover status is a changeover shutdown status, the corresponding time period will be designated as an idle window; When the cycle segment is the interval between adjacent processing cycles, and the duration of the interval is not less than the execution time of the diagnostic action, the corresponding time period is determined as the idle window. Before designating the corresponding time period as an idle window, verify the safety conditions, which include the spindle not executing cutting commands, the feed axis being in a safe position, the hydraulic clamping state being stable, and the tool not being in the workpiece contact area.

8. A predictive maintenance method for manufacturing equipment based on edge computing according to claim 7, characterized in that, The preset diagnostic action command is selected according to the execution method of the target component; When the target component is a linear actuator, the preset diagnostic action is a small reciprocating displacement action within the safe range; When the target component is a rotary actuator, the preset diagnostic action is a short-time no-load speed scan. When the target component is a pressure actuator, the preset diagnostic action is a short-term pressure holding action; The response data includes startup response data, steady-state response data, and recovery response data.

9. A predictive maintenance method for manufacturing equipment based on edge computing according to claim 8, characterized in that, In step S5, the degradation evidence value is updated based on the deviation between the response data and the healthy baseline data, including: The startup response deviation is determined based on the difference between the startup response data and the historical active diagnostic baseline data. The steady-state response deviation is determined based on the difference between the steady-state response data and the historical active diagnostic baseline data. The deviation of the recovery response is determined based on the difference between the recovery response data and the historical active diagnostic baseline data; Based on the degradation evidence value before the update, the deviation of the startup response, the deviation of the steady-state response, and the deviation of the recovery response, the updated degradation evidence value is determined according to the update fusion rules obtained from offline calibration.

10. A predictive maintenance method for manufacturing equipment based on edge computing according to claim 9, characterized in that, The maintenance levels include Level 1 maintenance, Level 2 maintenance, and Level 3 maintenance; The maintenance level is determined based on the updated degradation evidence value, the third interval threshold, and the fourth interval threshold, with the third interval threshold being less than the fourth interval threshold. When the updated degradation evidence value is less than the threshold of the third interval, the maintenance level is determined to be Level 1 maintenance, and the recommended maintenance window is determined to be the next replacement downtime period. Level 1 maintenance corresponds to online inspection confirmation and lubrication status review. When the updated degradation evidence value is greater than or equal to the third interval threshold and less than the fourth interval threshold, the maintenance level is determined to be Level 2 maintenance, and the recommended maintenance window is determined to be the first idle window after the current batch ends. Level 2 maintenance corresponds to planned downtime adjustment, lubrication replenishment, or replacement of worn parts. When the updated degradation evidence value is greater than or equal to the threshold of the fourth interval, the maintenance level is determined to be Level 3 maintenance, and the recommended maintenance window is determined to be the current idle window. Level 3 maintenance corresponds to immediate shutdown and maintenance. When the availability status of spare parts changes to unavailability, the recommended maintenance time windows for Level 1 and Level 2 maintenance are adjusted to the first idle window after the spare parts arrive, and a spare parts missing prompt is output for Level 3 maintenance.

Images