An ai technology-based device predictive maintenance method

Abstract

The application provides an AI technology-based equipment predictive maintenance method, comprising: collecting running response data of multiple devices in a cooperative operation process in real time; detecting a production line running state to identify a naturally occurring working condition switching event, taking the identified working condition switching event as a quasi-intervention signal, and extracting a response data segment within a predetermined time window before and after the occurrence of the quasi-intervention signal from the running response data; inputting the collected response data segment into a pre-constructed causal decoupling model, enabling the causal decoupling model to determine the transmission relationship between the response abnormalities of each device through comparative analysis of the response data segments before and after the quasi-intervention signal, and separate the independent degradation characteristic components of each device based on the transmission relationship; predicting the remaining useful life of each device according to the degradation characteristic components of each device, and outputting a maintenance decision. The application overcomes the defect that the root cause device cannot be accurately identified due to compensation system interference, and significantly improves the accuracy of cooperative production line predictive maintenance.

Description

Technical Field

[0001] This invention belongs to the field of artificial intelligence operation and maintenance technology, specifically relating to a predictive maintenance method for equipment based on AI technology. Background Technology

[0002] In the field of intelligent manufacturing, automated assembly lines widely employ multiple industrial robots, CNC machine tools, press-fitting units, and other equipment to collaboratively complete complex processes. To adapt to the demands of flexible production, these lines are typically equipped with vision guidance systems and force compensation algorithms. When the actual position or output force of a piece of equipment deviates from the commanded value, the control system automatically adjusts the motion parameters or force of the cooperating equipment to maintain the final process accuracy. This compensation mechanism allows the production line to tolerate a certain degree of positional error and equipment performance fluctuations, thereby ensuring continuous production. However, with long-term operation, the internal mechanical components (such as reducer gears, bearings, and lead screw nut pairs) will gradually wear down. Early degradation usually manifests as a slight drift in output position or force, which falls precisely within the range that the compensation system can correct.

[0003] In existing technologies, equipment maintenance for multi-machine collaborative production lines mainly relies on independent monitoring data of individual devices. This includes collecting vibration signals, servo motor current, temperature, and other characteristic parameters from each device, and then assessing the equipment's health status using threshold judgments or time-series prediction models. However, this monitoring method ignores the "masking" effect of compensation systems on degradation signals: when a device experiences early degradation, its own vibration and current characteristics may not change significantly, but the compensation system forces other collaborating devices to exert extra effort, leading to increased load and more pronounced current fluctuations in the latter. Therefore, traditional methods often point warnings to healthy devices with increased load, while the truly root cause of degradation is missed due to mild symptoms. More seriously, with the increasing complexity of visual compensation algorithms and force control strategies, this coupling interference effect becomes more pronounced, making it difficult for existing predictive maintenance methods based on independent monitoring of single devices to accurately identify the root cause device in collaborative production lines. Summary of the Invention

[0004] In view of the above-mentioned defects or deficiencies in the prior art, this application provides a predictive maintenance method for equipment based on AI technology, including: Real-time acquisition of operational response data of multiple devices during collaborative operation, including drive current, output torque, deviation between actual position and commanded position of each device, and relative pose correction amount between each device; The system detects the production line's operating status to identify naturally occurring operating condition switching events. These events are then used as quasi-intervention signals. Response data segments within a predetermined time window before and after the occurrence of these quasi-intervention signals are extracted from the operating response data. The collected response data fragments are input into a pre-built causal decoupling model. The causal decoupling model determines the transmission relationship between the response anomalies of each device by comparing and analyzing the response data fragments before and after the intervention signal. Based on the transmission relationship, the degradation feature components of each device are separated from the influence of other devices. Based on the time series of degradation characteristic components of each device at different times, the remaining service life is predicted, and maintenance decisions are output.

[0005] According to the technical solution provided by the present invention, the monitoring of production line operating status to identify naturally occurring operating condition switching events includes: Real-time acquisition of production line operation status parameters, including current product model identifier, tool identifier, set cycle time, and workpiece gripping success rate; When one or more preset conditions are detected, it is determined to be a working condition switching event; the preset conditions include changes in robot trajectory parameters caused by product model switching, a step change in clamping force caused by automatic tool replacement, a sudden change in motion speed during production cycle adjustment, or the start and end of a retry sequence after workpiece gripping failure.

[0006] According to the technical solution provided by the present invention, after extracting the response data fragments within a predetermined time window before and after the occurrence of the quasi-intervention signal, a preprocessing step is further included: The extracted response data segments are timestamped and the time when the quasi-intervention signal occurs is marked as time zero. The response data segments before time zero are taken as the first state sequence, and the response data segments after time zero are taken as the second state sequence.

[0007] According to the technical solution provided by the present invention, the step of determining the transmission relationship between response anomalies of various devices through comparative analysis of response data fragments using the causal decoupling model includes: For each device, compare the maximum mean difference of each operational response data item in the first state sequence and the second state sequence; if the maximum mean difference of any operational response data item exceeds the first threshold corresponding to the device, then mark the device as having a response anomaly. For each first device marked as having an abnormal response, it is paired with all other second devices. The time-series cross-correlation function between the data fluctuations in the second state sequence of the first device and the corresponding data fluctuations in the first state sequence of the second device is calculated, and the maximum correlation coefficient and its corresponding lag time are recorded. If the maximum correlation coefficient exceeds the second threshold and the lag time matches the inherent delay of the production line control compensation, it is determined that at the moment the quasi-intervention signal occurs, the abnormal response of the first device is caused by the degradation of the second device transmitted through control compensation, and the transmission relationship is recorded.

[0008] According to the technical solution provided by the present invention, separating the degradation feature components of each device independently of the influence of other devices based on the transmission relationship includes: For a pair of devices with a transitive relationship, the data fluctuation component corresponding to the transitive relationship in the second state sequence of the first device is subtracted, and the remaining part is used as the independent degradation feature component of the first device at the moment when the quasi-intervention signal occurs. For devices that are determined to have abnormal responses but are not the source of the transmission relationship, the difference between their first state sequence and second state sequence is directly used as the independent degradation feature component of the device at the moment when the quasi-intervention signal occurs. For the second device identified as the source of the transmission relationship, the difference between its first state sequence and its second state sequence is directly used as the independent degradation feature component of the device at the moment the quasi-intervention signal occurs, and it is retained as the identifier of the source of the transmission relationship.

[0009] According to the technical solution provided by the present invention, predicting the remaining service life of each device based on the time series composed of degradation characteristic components at different times includes: Once the independent degradation feature components of each device are obtained at the current moment, the independent degradation feature components are appended and stored according to the device identifier and time to form the time series of degradation feature components of each device; The degradation trend function is obtained by fitting the time series of degradation characteristic components of each device. The time required to reach the failure threshold is calculated by extrapolating the degradation trend function to the preset failure threshold, and then used as the predicted remaining life of the device.

[0010] According to the technical solution provided by the present invention, the output maintenance decision includes: For each device, the number of times the device acts as the source of a transmission relationship or is affected by other devices is counted based on the transmission relationship, and the contribution of the device is calculated. The device with the largest contribution is identified as the root cause device, the device with a contribution less than the third threshold is identified as a device without anomalies, and the rest are identified as subordinate devices; maintenance instructions are generated only for the root cause device, and coupling interference prompts are generated for subordinate devices. Based on the predicted remaining lifespan of the root cause equipment, a three-level early warning is generated. When the early warning level reaches the second level or above, a temporary task rescheduling plan is generated to transfer the process of the root cause equipment to other equipment or manual workstations.

[0011] According to the technical solution provided by the present invention, during the process of extracting response data segments within a predetermined time window before and after the occurrence of the quasi-intervention signal, if a sufficient number of working condition switching events are not captured after a continuous detection time exceeding a preset duration, an active detection mode is triggered, including: Send short-duration, small-amplitude additional motion commands to the production line control system. The amplitude is limited to the allowable range of the equipment's static accuracy, and the duration does not exceed the set duration. Mark the start time of the additional motion command as the quasi-intervention signal occurrence time, and extract response data segments within the predetermined time window before and after the condition switching event in the same way as the condition switching event; After the active detection command is completed, the original control command of the production line is restored, and the response data fragment generated by the active detection is marked as a human intervention sample for use by the causal decoupling model.

[0012] According to the technical solution provided by the present invention, during the real-time acquisition of operational response data from various devices, when continuous data loss or a sustained value exceeding a preset reasonable range is detected from a certain sensor, a data recovery process is triggered, including: Based on the established inter-device transmission relationship in the causal decoupling model, find the real-time data of other sensors that are coupled with the faulty sensor; Using the aforementioned coupling relationship and data from other sensors, an approximate replacement value for the faulty sensor is calculated through fitting. The approximate substitution value is used in subsequent calculations, and a sensor anomaly record is generated in the system log to prompt maintenance personnel to check or replace it during the next shutdown.

[0013] According to the technical solution provided by the present invention, the process of constructing a time series based on the degradation characteristic components of each device at different times further includes: If the independent degradation feature component at the current time is less than the independent degradation feature component at the previous time, and the absolute value of the difference between the two is greater than the fourth threshold, an abnormal jump is determined to have occurred. Revise the current independent degradation feature component to the sum of the component from the previous time step and a set growth step.

[0014] Compared with existing technologies, the advantages of this invention are as follows: This invention utilizes naturally occurring operating condition switching events on the production line as quasi-intervention signals, obtaining causal comparison data without active disturbance, thus avoiding impact on normal production. By comparing and analyzing response data fragments before and after the event using a causal decoupling model, the transmission relationship of response anomalies between devices can be accurately identified, thereby separating the independent degradation characteristic components of each device from the mixed signals masked by the compensation system. Based on the time series prediction of these components, the remaining lifespan can be accurately pinpointed to the root cause device of the actual degradation, avoiding misdirection of warnings to healthy devices. This invention fundamentally overcomes the shortcomings of existing technologies that cannot accurately identify the root cause device due to interference from the compensation system, significantly improving the accuracy of predictive maintenance of collaborative production lines. Attached Figure Description

[0015] Other features, objects, and advantages of this application will become more apparent from the following detailed description of non-limiting embodiments with reference to the accompanying drawings: Figure 1 A flowchart illustrating the steps of the AI-based predictive maintenance method for equipment provided in Example 1; Figure 2 This is a flowchart of the steps of the AI-based predictive maintenance method for equipment provided in Example 2; Figure 3 This is a flowchart of the steps of the AI-based predictive maintenance method for equipment provided in Example 3; Figure 4 The flowchart illustrates the steps of the AI-based predictive maintenance method for equipment provided in Example 4. Detailed Implementation

[0016] The present application will now be described in further detail with reference to the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are merely illustrative of the invention and not intended to limit it. Furthermore, it should be noted that, for ease of description, only the parts relevant to the invention are shown in the accompanying drawings.

[0017] It should be noted that, unless otherwise specified, the embodiments and features described in this application can be combined with each other. This application will now be described in detail with reference to the accompanying drawings and embodiments.

[0018] Example 1 Please refer to Figure 1 This embodiment proposes a predictive maintenance method for equipment based on AI technology, including: S100: Real-time acquisition of operational response data of multiple devices during collaborative operation. The operational response data includes the drive current, output torque, deviation between actual position and commanded position of each device, and relative pose correction amount between each device.

[0019] Specifically, in step S100, "multiple devices" refers to a group of devices that collaboratively complete a certain process on the same automated production line. For example, one industrial robot is responsible for gripping workpieces, another for assembly, and a pressing unit performs pressing. Each device is equipped with an independent servo drive system, thus enabling real-time acquisition of drive current and output torque. The drive current directly reflects the load on the servo motor; an increase in current may indicate increased resistance or the need for additional force. The output torque is the actual torque value output by the motor, and its fluctuations can characterize wear or jamming of mechanical transmission components. The deviation between the actual position and the commanded position is obtained through the device's own encoder or grating ruler. A larger deviation indicates poorer motion accuracy, which may be related to lead screw wear or increased bearing clearance. The relative pose correction between devices is provided by a vision guidance system or laser tracker at the top of the production line. This correction reflects the intensity of control intervention applied to compensate for positional deviations between devices, and its fluctuations can reveal that the positioning error of a certain device is being actively compensated by other devices.

[0020] In multi-machine collaborative production lines, when a piece of equipment experiences early mechanical degradation, its own changes in drive current, positional deviation, and other characteristics are often masked by the active adjustments of the compensation system. However, other cooperating equipment will fluctuate in load (manifested as increased drive current and output torque) and pose correction to compensate for the errors caused by this degradation. Therefore, this method does not monitor a single piece of equipment in isolation, but rather collects the above four types of data from all equipment simultaneously, providing a data foundation for revealing hidden transmission relationships through cross-equipment comparative analysis. The collected data is grouped and stored by equipment, meaning each equipment maintains an independent time-series database while retaining a global timestamp index to facilitate the subsequent extraction of synchronous response data fragments from all equipment at the precise intervention signal time point.

[0021] To facilitate understanding, the following explanation uses a reducer housing press-fitting production line as an example. This production line includes: robot R1 (responsible for picking up the housing from the loading station and placing it into the press-fitting fixture), robot R2 (responsible for picking up the bearing and moving it to the press-fitting position), press-fitting unit P (performing interference fit), and robot R3 (responsible for removing the finished product). During production line operation, this method uses a sampling frequency of 1kHz to collect the drive current and output torque of R1, R2, and R3 in real time through the servo driver feedback channels of each device; it reads the command position and actual feedback position for each motion cycle through the controller of each device and calculates the position deviation; simultaneously, the vision guidance system reads the pose correction of the robot end effector relative to the target workpiece every 50ms. All data is stored with the device identifier as the first index and the timestamp as the second index, ensuring that independent time-series data for each device can be quickly extracted during each event analysis.

[0022] S200: Detects the production line operating status to identify naturally occurring operating condition switching events, uses the identified operating condition switching events as quasi-intervention signals, and extracts response data segments within a predetermined time window before and after the occurrence of the quasi-intervention signal from the operating response data.

[0023] Specifically, in step S200, naturally occurring operating condition switching events refer to sudden changes in the operating state of the production line that can occur without human intervention during normal production. These events disrupt the steady-state operating point of the equipment, triggering a series of compensatory responses from the control system, resulting in comparable differences in the operational response data before and after the event. Therefore, they can serve as natural "quasi-intervention signals." This step identifies these events in real time by monitoring the status flags in the production line control system. Once an event is detected, its occurrence time is marked as time zero. Then, data segments within a predetermined time window before and after this time are extracted from the operational response data collected in step S100, serving as the basic units for subsequent analysis.

[0024] S300: Input the collected response data fragments into the pre-constructed causal decoupling model, so that the causal decoupling model can determine the transmission relationship between the response anomalies of each device by comparing and analyzing the response data fragments before and after the intervention signal, and separate the degradation feature components of each device independently of the influence of other devices based on the transmission relationship.

[0025] Specifically, in step S300, the causal decoupling model is a pre-trained machine learning model whose core function is to extract the true degradation signal of each device from the observation data mixed with compensation interference. The model construction process is as follows: First, based on the control compensation logic between multiple devices, a directed graph structure is established, where nodes represent each device and directed edges represent possible compensation transmission paths; second, key parameters in the model are trained using historical normal operation data and known fault data, enabling the model to learn the normal compensation transmission rules between devices, thereby identifying abnormal transmission relationships and separating degradation components from the measured data. The core function of the model is to determine whether the abnormal response of a certain device is caused by the degradation of other devices through compensation transmission by comparing the differences between the two state sequences of the same device before and after the quasi-intervention signal, combined with the temporal correlation between different devices. Taking a single robot as an example: if its drive current exhibits abnormal fluctuations after a change in operating conditions, while another robot already had a slight positional deviation before the change, and the temporal correlation of the fluctuations between the two robots matches the compensation delay of the control system, then the model determines that the abnormality in the former is caused by the degradation of the latter. The model then subtracts the transmitted fluctuation component from the data of the former, and the remaining part is its own degradation characteristic component. After this processing, each device obtains an independent degradation characteristic component unaffected by coupling with other devices, providing a clean data foundation for subsequent lifetime prediction.

[0026] S400: Based on the time series of degradation characteristic components of each device at different times, predict its remaining service life and output maintenance decisions.

[0027] Specifically, in step S400, whenever a natural operating condition switching event occurs on the production line, this method obtains a set of independent degradation characteristic components of each device at the current moment through steps S100 to S300. These components are stored according to device identification and time order to form a time series of degradation characteristic components for each device. As the production line continues to operate, this series continuously extends, reflecting the evolution trend of the degradation degree of each device. This method performs trend fitting and extrapolation on the time series of degradation characteristic components for each device to obtain the time required for its degradation characteristic components to reach a preset failure threshold, which is the predicted remaining service life of the device. Based on this, by combining the prediction results of all devices and the transmission relationships recorded in step S300, the root cause device with severe degradation and negative impact on other devices is identified, and maintenance instructions, early warning information, and temporary task rescheduling schemes are generated for the root cause device.

[0028] Further, in step S200, monitoring the production line operating status to identify naturally occurring operating condition switching events includes the following steps S210-S220: S210: Real-time acquisition of production line operation status parameters, including current product model identifier, tool identifier, set cycle time, and workpiece grabbing success rate.

[0029] Specifically, in step S210, this step communicates with the production line PLC or Manufacturing Execution System (MES) via industrial Ethernet to read the production line's operating status parameters in real time. Among these parameters, the product model identifier is used to distinguish the process parameters of different workpieces; the tool identifier is used to identify the currently used end effector such as a gripper or suction cup; the set cycle time reflects the production line's operating speed; and the workpiece gripping success rate is calculated by the vision system or sensors. These parameters are stored in the PLC's registers or the MES's database. This method polls and reads these parameters at a frequency of no less than 10Hz to ensure timely capture of status changes.

[0030] Taking the reducer housing press-fit production line as an example, the PLC of this production line is equipped with dedicated registers: D100 stores the current product model (e.g., 001 represents type A, 002 represents type B), D101 stores the current tool number (e.g., 1 represents the gripper, 2 represents the press-fit head), D102 stores the set cycle time (in seconds), and D103 stores the success rate (percentage) of the last 100 grips.

[0031] S220: When one or more preset conditions are detected, it is determined to be a working condition switching event; the preset conditions include changes in robot trajectory parameters caused by product model switching, a step change in clamping force caused by automatic tool replacement, a sudden change in motion speed during production cycle adjustment, or the start and end of a retry sequence after workpiece gripping failure.

[0032] Specifically, in step S220, the status parameters acquired in step S210 are continuously monitored. Once a change in a parameter value that conforms to a preset rule is detected, a working condition switching event is determined to have occurred. The specific rules are as follows: Product Model Switching: A product model switching event is identified when the value of the product model identifier changes. This event will cause a jump in the motion trajectory parameters (such as point coordinates, velocity, and acceleration) of each robot.

[0033] Automatic tool change: A tool change event is detected when the value of the tool identifier changes. This event will cause a step change in the clamping force command.

[0034] Production cycle time adjustment: When the rate of change of the set cycle time exceeds 5% (e.g., from 10 seconds to 9 seconds), it is determined to be a cycle time adjustment event. This event will cause a sudden change in the robot's movement speed.

[0035] Workpiece gripping failure retry: When the workpiece gripping success rate drops instantly below 80% and a gripping retry sequence (multiple consecutive gripping attempts) occurs, it is determined as a retry event. This event usually indicates that there is a positional deviation or workpiece abnormality in the gripping process.

[0036] In the aforementioned reducer housing press-fitting production line, for example, if the value of D100 changes from 001 to 002, this method immediately determines that a product model switching event has occurred; if the value of D101 changes from 1 to 2, it determines that a tool replacement event has occurred. If multiple operating condition switching events are detected simultaneously within a predetermined time window, the response data segment corresponding to each event is extracted according to the chronological order of their occurrence, and each segment is used as an independent quasi-intervention sample.

[0037] Specifically, in actual production line operation, multiple operating condition switching events may occur consecutively within a short period, such as product model switching accompanied by automatic tool replacement. This method sets a predetermined time window (e.g., 1 second). If multiple events are detected within this window, they are not merged into a single sample but processed separately. Specifically, each event is sequentially labeled as a quasi-intervention signal based on its actual timestamp, and response data fragments within the predetermined time windows before and after each event are extracted. Each fragment is independently input into a causal decoupling model for analysis, outputting corresponding degradation feature components. The significance of this approach is that different events may trigger different degradation characteristics in different equipment; analyzing them separately yields richer degradation information and improves prediction accuracy.

[0038] In the press-fitting production line example above, for instance, a product model switch (time t1) and an automatic tool change (time t2, t2>t1) occur successively within 0.5 seconds. This method extracts the first response data segment using t1 as the quasi-intervention signal and the second response data segment using t2 as the quasi-intervention signal. The two segments are then processed independently. If the time difference between these two events is extremely small (e.g., less than the window length), the two segments may partially overlap, but this does not affect independent analysis because each segment is aligned with its own event zero point, reflecting the equipment response under different disturbance conditions.

[0039] Furthermore, in step S200, after extracting the response data segments within a predetermined time window before and after the occurrence of the quasi-intervention signal, the process further includes the following preprocessing steps S201-S202: S201: Timestamp align the extracted response data segments and mark the time when the quasi-intervention signal occurred as time zero.

[0040] Specifically, since the operational response data of each device collected in step S100 comes from different data sources, their respective time bases may have slight offsets. In order to accurately compare the data changes of each device before and after the intervention signal occurs, step S201 first timestamps all extracted response data segments. Specifically, the time of occurrence of the condition switching event identified in step S200 is taken as the global reference zero point, and the timestamps of all data points before and after this time point are subtracted from this reference zero point, so that the time axis of the entire data segment is recalibrated as a time series centered on time zero. In this way, the data of all devices are synchronized in time, establishing a unified time reference system for subsequent state comparison analysis.

[0041] In the aforementioned reducer housing press-fitting production line example, for instance, the absolute time of a product model change event is 10:30:15.200 on June 9, 2025. This step marks this time as time zero, so the timestamps of drive current data points before this time are converted to negative values ​​(e.g., -5.000 seconds, -4.995 seconds, etc.), and the timestamps of data points after this time are converted to positive values ​​(e.g., +0.010 seconds, +0.020 seconds, etc.). S202: Take the response data segment before time zero as the first state sequence and the response data segment after time zero as the second state sequence.

[0042] Specifically, in step S202, after timestamp alignment, the response data segments are divided into two sequences according to time zero: all data points before time zero constitute the first state sequence, representing the steady-state operation of the equipment before the condition switching event; all data points after time zero constitute the second state sequence, representing the response state of the equipment during the compensation and adjustment process of the control system after the event. The lengths of the two sequences are determined by the predetermined time window set in step S200, for example, a 5-second window for the first sequence and a 10-second window for the second sequence. Through this division, this method obtains two state samples of the same equipment before and after the same event, providing a natural "control group" and "experimental group" for subsequent comparative analysis.

[0043] It should be noted that the first and second state sequences obtained from the division also need to be smoothed and filtered to remove spikes caused by random noise.

[0044] Further, in step S300, the step of enabling the causal decoupling model to determine the transmission relationship between the response anomalies of each device through comparative analysis of response data fragments includes the following steps S310-S330: S310: For each device, compare the maximum mean difference of each running response data in the first state sequence and the second state sequence; if the maximum mean difference of any running response data exceeds the first threshold corresponding to the device, then mark the device as having a response anomaly.

[0045] Specifically, in step S310, the Maximum Mean Discrepancy (MMD) is a nonparametric statistic used to measure the degree of difference between two probability distributions. Unlike simple mean comparisons, MMD can simultaneously capture differences in mean, variance, and higher-order moments between distributions, making it suitable for detecting scenarios where degradation signals are "suppressed" under the action of a compensation system, resulting only in changes in fluctuation patterns. For each device, its first and second state sequences are considered as two sample sets. The mean vector of each sample set in the regeneration kernel Hilbert space (RKHS) is calculated, and the distance between the two mean vectors is the MMD value. The larger the MMD value, the more significant the change in the distribution of the device's operational response data before and after the event.

[0046] This embodiment calculates the MMD value for each operational response data point (drive current, output torque, position deviation, and pose correction) for each device. If the MMD value of any item exceeds the first threshold corresponding to that device, the device is marked as having an abnormal response. The first threshold can be set based on the 99th percentile of the MMD values ​​in historical normal operation data, or it can be determined through cross-validation. For example, for the drive current of robot R1, its MMD value before and after an event is usually less than 0.05 under normal operating conditions, so the first threshold can be set to 0.08.

[0047] Table 1 below shows the MMD calculation results and anomaly markings for various operational response data of each device after a product switching event.

[0048] Table 1. MMD calculation results of the operating response data of each device

[0049] As shown in Table 1, the MMD value of the position deviation of robot R1 (0.11) exceeds the threshold, and the MMD values ​​of the drive current and output torque of robot R2 (0.19 and 0.15, respectively) also exceed the threshold. Therefore, this step marks robots R1 and R2 as having abnormal responses. All indicators of the pressing unit P do not exceed the threshold, and no abnormality is marked.

[0050] S320: For each first device marked as having an abnormal response, pair it with all other second devices, calculate the time-series cross-correlation function between the data fluctuations in the second state sequence of the first device and the corresponding data fluctuations in the first state sequence of the second device, and record the maximum correlation coefficient and its corresponding lag time.

[0051] Specifically, in step S320, for each device marked as having an abnormal response (referred to as the first device), it is paired with all other devices (including devices not marked as abnormal, referred to as the second devices). For each pair of devices, the temporal cross-correlation function between the data fluctuations in the second state sequence of the first device and the corresponding data fluctuations in the first state sequence of the second device is calculated. The reason for using the second state sequence of the first device (after the event) and the first state sequence of the second device (before the event) is that: if the abnormality of the first device is caused by the degradation of the second device through compensation transmission, then the degradation of the second device already exists before the event (manifested as fluctuations in its first state sequence), while the abnormal response of the first device occurs after the event (manifested as fluctuations in its second state sequence), and there is a time difference between the two determined by the control system delay.

[0052] The formula for calculating the time-series cross-correlation function is:

[0053] In the formula, This is the second state sequence data of the first device. This corresponds to the first state sequence data of the second device. for The arithmetic mean of the sequence for The arithmetic mean of the sequence For the time period afterward. Calculate a series of... The correlation coefficients under the given values ​​are taken as the maximum correlation coefficients, and the corresponding lag times are recorded. .

[0054] Table 2 below shows the time-series cross-correlation results calculated by pairing the marked anomalies robots R1 and R2 with other devices.

[0055] Table 2. Timing cross-correlation calculation results between the faulty device and each paired device.

[0056] As shown in Table 2, the correlation coefficient between the drive current fluctuation of robot R2 and the position deviation fluctuation of the first state sequence of robot R1 is as high as 0.87, with a lag time of 0.21 seconds, which is very close to the inherent delay of 0.2 seconds of the production line vision compensation system. The correlations of other pairs are all low. Therefore, this step records the maximum correlation coefficient of this pair as 0.87 and the lag time as 0.21 seconds.

[0057] S330: If the maximum correlation coefficient exceeds the second threshold and the lag time matches the inherent delay of the production line control compensation, it is determined that at the moment the quasi-intervention signal occurs, the abnormal response of the first device is caused by the degradation of the second device transmitted through control compensation, and the transmission relationship is recorded.

[0058] Specifically, in step S330, a second threshold is first set, for example, 0.7, to indicate a strong correlation. Secondly, it is necessary to determine whether the lag time matches the inherent delay of the production line control compensation. The inherent delay refers to the total time from when the control system detects the position deviation to when it issues a compensation command, and then to when the actuator responds; it can usually be obtained through offline calibration or analysis of historical data. The matching criterion can be that the absolute difference between the lag time and the inherent delay is less than a tolerance value (e.g., 0.05 seconds), or the relative error is less than 10%.

[0059] For the pairing of robot R2 and robot R1 in Table 2, the maximum correlation coefficient 0.87 > 0.7, and the difference between the lag time of 0.21 seconds and the inherent delay of 0.2 seconds is 0.01 seconds, which is less than the tolerance value of 0.05 seconds. Therefore, the pairing is deemed successful. This step determines that at the moment the quasi-intervention signal occurs, the abnormal response of robot R2 is caused by the degradation of robot R1 transmitted through control compensation. The transmission relationship is recorded as: robot R1 → robot R2, that is, robot R1 is the root cause, and robot R2 is the subordinate device affected by the coupling.

[0060] For pairings between robot R1 and other devices, the maximum correlation coefficients did not exceed the second threshold, therefore no transmission relationships were recorded. This means that the abnormal response of robot R1 itself (increased positional deviation) is directly caused by its own degradation, rather than by transmission from other devices.

[0061] The above determination accurately reflects the actual situation: wear on the reducer of robot R1 led to a decrease in its positional accuracy. After the vision system detected this deviation, it compensated by adjusting the motion trajectory of robot R2, resulting in an abnormal increase in the drive current of robot R2. This method successfully separated this hidden causal chain from the mixed signal.

[0062] Further, in step S300, the step of separating the degradation feature components of each device independently of the influence of other devices based on the transmission relationship includes the following steps S340-S360. It should be noted that steps S340-S360 are parallel steps following step S330.

[0063] S340: For a pair of devices with a transitive relationship, the data fluctuation component corresponding to the transitive relationship in the second state sequence of the first device is subtracted, and the remaining part is used as the independent degradation feature component of the first device at the moment when the quasi-intervention signal occurs.

[0064] Specifically, in step S340, the data fluctuation component corresponding to the transmission relationship refers to the portion of the signal in the second state sequence of the first device that is strongly correlated with the data fluctuation of the first state sequence of the second device. This embodiment uses a linear regression method to estimate the amplitude of this fluctuation component: using the data fluctuation of the first state sequence of the second device as the independent variable and the data fluctuation of the second state sequence of the first device as the dependent variable, a linear regression model is fitted, and the predicted value obtained is the transmission contribution. Subtracting this predicted value from the original second state sequence data of the first device, the remaining part is the independent degradation characteristic component after deducting the compensation for transmission effects. This component reflects the true degree of degradation of the first device itself; if the first device is healthy, the remaining part should approach zero.

[0065] In the aforementioned reducer housing press-fitting production line example, step S330 has determined that the abnormal response (increased drive current) of robot R2 is caused by the positional deviation degradation of robot R1 transmitted through visual compensation. This step performs the following operation: using the positional deviation data of robot R1's first state sequence as the independent variable X, and the drive current data of robot R2's second state sequence as the dependent variable Y, a linear regression model is established. ,in This is the transfer gain coefficient. It is estimated using the least squares method. That is, for every 1mm increase in the positional deviation of robot R1, the drive current of robot R2 will increase by approximately 2.3A. Calculate the predicted value of the regression model. This refers to the data fluctuation component corresponding to the transmission relationship. Subtract the actual drive current of robot R2 from... The remaining part is obtained. The value of this remaining part is small and shows no obvious trend change, indicating that robot R2 itself has not degenerated, and its anomaly is entirely caused by the degradation compensation of robot R1. Therefore, the independent degradation feature component of robot R2 is this remaining part.

[0066] S350: For a device that is determined to have an abnormal response but is not the source of the transmission relationship, the difference between its first state sequence and second state sequence is directly used as the independent degradation feature component of the device at the moment when the quasi-intervention signal occurs.

[0067] Specifically, in step S310, some devices are marked as having response anomalies, but in step S330 they are not determined to be the source of any transmission relationship (i.e., no other device's anomaly is caused by this device). The possible situation for these devices is that their response anomalies are entirely due to their own degradation, rather than being transmitted from other devices. For these devices, step S350 directly uses the difference between their first and second state sequences as their independent degradation feature components. This difference can be a comprehensive difference in various operational response data, such as using Euclidean distance or a weighted sum. Since the anomalies of these devices are caused by their own reasons, the difference between their preceding and following state sequences should directly reflect their degree of degradation, without the need for subtraction. S360: For the second device identified as the source of the transmission relationship, the difference between its first state sequence and second state sequence is directly used as the independent degradation feature component of the device at the moment the quasi-intervention signal occurs, and its identifier as the source of the transmission relationship is retained.

[0068] Specifically, for the device identified as the source of the transmission relationship in step S330 (i.e., the root cause of abnormalities in other devices), this step directly uses the difference between its first state sequence and second state sequence as its independent degradation feature component. Similar to S350, this difference directly reflects the change in the device's own state before and after the event. The difference is that step S360 requires retaining the identifier of the device as the source of the transmission relationship; this identifier will be used in subsequent maintenance decisions. The reason for not performing any deduction operation on the source device is that the degradation of the source device is the original driving factor, and the changes in its observed data are its true degradation signal, which should not be weakened or modified.

[0069] In the press-fitting production line example above, robot R1 is identified as the source of the transmission relationship (R1→R2). This step directly uses the difference between the first and second state sequences of robot R1 as its independent degenerate feature component. Specifically, the change in robot R1's position deviation before and after the event is calculated: the average position deviation in the first state sequence is 0.15mm, and the average position deviation in the second state sequence is 0.28mm, with a difference of 0.13mm. This difference is used as the independent degenerate feature component of robot R1. Simultaneously, the system records an identifier "Source Equipment: Robot R1" for use in subsequent steps.

[0070] After processing in steps S340, S350, and S360, each device obtains an independent degradation characteristic component: the component of robot R1 (the source) is relatively large (0.13 mm), directly reflecting its decreased positional accuracy; the component of robot R2 (affected by transmission) is very small (the remaining part after subtraction), indicating its own health; the components of other devices not marked as abnormal (such as pressing unit P) are close to zero. These independent degradation characteristic components eliminate coupling interference between devices, laying a data foundation for accurately predicting remaining lifespan.

[0071] Further, in step S400, predicting the remaining service life of each device based on the time series composed of degradation characteristic components at different times includes the following steps S410-S430: S410: Whenever the independent degradation feature component of each device is obtained at the current time, the independent degradation feature component is appended and stored according to the device identifier and time to form the time series of degradation feature components of each device.

[0072] Specifically, in step S410, after each naturally occurring operating condition switching event, the independent degradation characteristic components of each piece of equipment at the current event moment are calculated. Each component is a scalar value (e.g., change in position deviation, residual after current correction, etc.), representing the degree of degradation of the equipment at that moment. This step appends and stores these components to an industrial database or time-series data platform, using the equipment identifier as the primary key and the event occurrence time as the timestamp. As the production line continues to operate, each time an operating condition switching event occurs, a new set of degradation characteristic components is obtained, thereby constructing a time series of degradation characteristic components evolving over time for each piece of equipment. This series reflects the complete process of the equipment gradually moving from a healthy state to failure.

[0073] In the example of the reducer housing press-fitting production line described above, it is assumed that multiple product model switching events occurred over several consecutive days. After each event, the independent degradation characteristic component (i.e., position deviation difference) of robot R1 was calculated and recorded as follows: Day 1: 0.02mm, Day 5: 0.05mm, Day 8: 0.07mm, Day 12: 0.10mm, Day 15: 0.13mm, Day 19: 0.17mm, Day 23: 0.22mm, Day 28: 0.28mm, Day 32: 0.35mm. The independent degradation characteristic component of robot R2 (residual current after deducting transmission effects) was close to zero each time, thus not showing a significant upward trend. These time series are stored separately to provide a basis for subsequent fitting.

[0074] S420: Fit the time series of degradation characteristic components for each device to obtain the degradation trend function.

[0075] Specifically, in step S420, the locally weighted regression (LOWESS) method is used to fit the time series of degradation characteristic components for each device. Unlike ordinary global regression (such as linear regression and multinomial regression), locally weighted regression does not require pre-assuming a global functional form of the degradation trajectory (such as linear, exponential, etc.). Instead, for each predicted point, it uses data points in its neighborhood for weighted regression, with points closer to the target having greater weight. This method is particularly suitable for the degradation process of industrial equipment because the degradation rate may exhibit non-stationary characteristics due to changes in operating conditions, maintenance behavior, etc., and locally weighted regression can flexibly capture local trends, avoiding overfitting or underfitting.

[0076] The calculation process of locally weighted regression is as follows: For each time point Using it as the center, select a bandwidth parameter. (Typically, 20% to 40% of the total data points are used) to determine all time points within the neighborhood. satisfy For each point in the neighborhood Calculate weights ,in Kernel function, This represents the normalized distance. This embodiment uses a cubic kernel function. (when Otherwise, it is 0). (Weighted) We perform weighted least squares linear regression on the data points in the neighborhood to obtain the fitted value for that point. After iterating through all time points, a smooth degradation trend curve can be obtained. Bandwidth parameters The bandwidth that minimizes the prediction error is determined by cross-validation.

[0077] In the example above, the time series of the degradation feature components of robot R1 exhibits an approximately exponential upward trend. After fitting with local weighted regression, the degradation trend function is obtained. able to at any time Provide estimates of the degradation eigencomponents.

[0078] S430: Extrapolate the degradation trend function to the preset failure threshold, and calculate the time required to reach the failure threshold as the predicted remaining life of the device.

[0079] Specifically, in step S430, a failure threshold is first set according to the equipment type and process requirements. For example, regarding the degradation characteristic component of robot R1's positional deviation, when this value exceeds 0.8 mm, it is considered that it can no longer meet the assembly accuracy requirements, i.e., it is declared a failure. The degradation trend function obtained in step S420... Extend outwards and solve the equation. The value obtained This is the absolute time it takes for the device to reach its failure threshold. Subtracting the current operating time (or the time of the last event) from this absolute time gives the remaining service life.

[0080] Further, in step S400, the output maintenance decision includes the following steps S440-S460: S440: For each device, count the number of times the device acts as the source of a transmission relationship or is affected by other devices based on the transmission relationship, and calculate the device's contribution.

[0081] Specifically, in step S440, the total number of times each device acts as a source of transmission relationships in historical events, and the total number of times it acts as a device affected by transmission, are counted. The contribution rate can be calculated as: Contribution Rate = Number of Sources × Weight 1 + Number of Affected Devices × Weight 2, where Weight 1 is greater than Weight 2. Alternatively, it can be simplified to Contribution Rate equal to the Number of Sources. This embodiment uses the number of sources as the contribution rate because the source devices are the ones that truly need maintenance.

[0082] S450: The device with the largest contribution is identified as the root cause device, the devices with a contribution less than the third threshold are identified as devices without abnormalities, and the rest are identified as subordinate devices; maintenance instructions are generated only for the root cause device, and coupling interference prompts are generated for subordinate devices.

[0083] Specifically, in step 450, a third threshold is set. For example, if the third threshold is 0, then among the devices with a contribution greater than 0, the one with the largest contribution is the root cause device; devices with a contribution equal to 0 are devices without abnormalities; if there is a device with a contribution greater than 0 but not the largest, it is designated as a subordinate device (but in the case of a single root cause, there is usually only one source). After classification, this step only generates specific maintenance instructions (such as replacing parts, lubrication, calibration, etc.) for the root cause device; and generates coupling interference prompts for subordinate devices (i.e., devices affected by the transmission), informing maintenance personnel that the abnormality of the device is caused by the root cause device and does not require repair, thus avoiding unnecessary maintenance.

[0084] S460: Generate a Level 3 warning based on the predicted remaining lifespan of the root cause device. When the warning level reaches Level 2 or above, generate a temporary task rescheduling plan to transfer the process of the root cause device to other equipment or manual workstations.

[0085] Specifically, in step S460, the predicted remaining lifespan of the root cause equipment is first categorized into three warning levels. For example: a remaining lifespan greater than 30 days is a Level 3 warning, recommending planned maintenance; a remaining lifespan between 7 and 30 days is a Level 2 warning, recommending proactive maintenance and starting preparation of spare parts; and a remaining lifespan less than 7 days is a Level 1 warning, recommending immediate shutdown for repair. When the warning level reaches Level 2 or higher, to reduce production quality risks or sudden downtime losses caused by root cause equipment degradation, this step automatically generates a temporary task rescheduling plan. The rescheduling plan reallocates the processes undertaken by the root cause equipment based on existing production line resources (such as backup equipment, manual workstations, and idle periods of other collaborating equipment). The rescheduling instruction is sent to the production line control system (PLC or MES) via industrial Ethernet, and the system automatically switches the task allocation.

[0086] Example 2 Based on the above embodiment 1, please refer to Figure 2 This embodiment provides another predictive maintenance method for equipment based on AI technology. The same content as in Embodiment 1 will not be repeated here; the difference lies in: In step S200, when extracting response data segments within a predetermined time window before and after the occurrence of the quasi-intervention signal, if a sufficient number of working condition switching events are not captured after a continuous detection time exceeding a preset duration, an active detection mode is triggered, including the following steps S203-S205. It should be noted that the sequence numbers of steps S203-S205 are only for ease of explanation and are not limited to being located after steps S201-S202.

[0087] S203: Sends short-duration, small-amplitude additional motion commands to the production line control system. The amplitude is limited to the allowable range of the equipment's static accuracy, and the duration does not exceed the set duration.

[0088] Specifically, in step S203, if a sufficient number of natural operating condition switching events are not captured after continuous monitoring for a preset duration (e.g., 48 hours), this method automatically triggers an active detection mode. The core of the active detection mode is to actively apply a small disturbance to artificially construct a quasi-intervention signal. This step sends an additional motion command to the production line control system (such as a PLC or robot controller), which superimposes a short-term offset onto the current motion trajectory. To ensure that normal production is not interfered with or equipment is not damaged, the amplitude of the additional motion command is strictly limited to the allowable range of static accuracy for each piece of equipment. For example, the position offset of the robot end effector does not exceed ±0.05mm, and the force offset of the pressing unit does not exceed 1% of the rated pressure. The duration does not exceed the set duration; in this embodiment, it is set to 0.3 seconds to ensure rapid recovery after detection within a very short time. The waveform of the additional motion command adopts a smooth pulse shape (such as a half-sine wave) to avoid impact.

[0089] S204: Mark the start time of the additional motion command as the quasi-intervention signal occurrence time, and extract the response data segments within the predetermined time window before and after the operation condition switching event in the same way.

[0090] Specifically, in step S204, the start time of the additional motion command is manually marked as the quasi-intervention signal occurrence time, which is completely equivalent to a naturally occurring condition switching event. This step, in the same manner as step S200, extracts response data segments within a predetermined time window before and after this moment from the operational response data. The lengths of the preceding and following windows are consistent with those of natural events. This processing ensures that the data segments generated by active detection can be seamlessly integrated into the subsequent causal decoupling model analysis process.

[0091] S205: After the active detection command is completed, restore the original control command of the production line and mark the response data fragment generated by the active detection as a manual intervention sample for use by the causal decoupling model.

[0092] Specifically, in step S205, after the additional motion command is executed, this step immediately sends a recovery command to the production line control system to cancel the previously superimposed offset, restoring the equipment to its original motion trajectory and control parameters. Simultaneously, the entire response data segment generated by this active detection is labeled as a "human intervention sample," distinguishing it from the "natural samples" generated by natural operating condition switching events. The labeling information includes metadata such as the detection time, the applied equipment, and the disturbance amplitude. This labeling helps evaluate the effectiveness of the active detection in subsequent analysis and can also be used for sample weighting during model training. The causal decoupling model treats human intervention samples in exactly the same way as natural samples, thus ensuring the consistency of predictions.

[0093] Example 3 Based on the above embodiment 1, please refer to Figure 3 This embodiment provides another predictive maintenance method for equipment based on AI technology. The same content as in Embodiment 1 will not be repeated here; the difference lies in: During the real-time acquisition of operational response data from each device in step S100, if continuous data loss or a value exceeding a preset reasonable range is detected from a certain sensor, a data recovery process is triggered, including the following steps S101-S103: S101: Based on the established inter-device transmission relationship in the causal decoupling model, find the real-time data of other sensors that are coupled with the faulty sensor.

[0094] Specifically, when a sensor is detected to have continuously lost data (e.g., no new data for more than 1 second) or continuously exceed a preset reasonable range (e.g., drive current exceeds twice the rated value for more than 0.5 seconds), it is determined that the sensor has malfunctioned or the data transmission is abnormal. At this time, step S101 first extracts the established inter-device transmission relationships from the causal decoupling model. These transmission relationships include which sensors' data are coupled between devices (e.g., there is a compensation transmission relationship between the position deviation of robot R1 and the drive current of robot R2). Based on these coupling relationships, real-time data from other sensors that have direct or indirect coupling relationships with the malfunctioning sensor are searched as a basis for alternative estimations.

[0095] S102: Using the coupling relationship and data from other sensors, an approximate replacement value for the faulty sensor is calculated by fitting.

[0096] Specifically, in step S102, using the real-time data from other sensors found in step S101, and combining it with the coupling parameters stored in the causal decoupling model, an approximate replacement value for the faulty sensor is calculated through least-squares fitting. The coupling parameters reflect the linear or nonlinear relationship between the data from the two sensors and can be pre-calibrated using regression analysis in historical normal data segments. When a fault occurs, the current real-time data from other sensors is substituted into the regression model, and the calculated predicted value is the approximate replacement value for the faulty sensor. If the coupling relationship is nonlinear, the corresponding nonlinear regression model is used.

[0097] S103: Use the approximate substitution value for calculations in subsequent steps, and generate a sensor anomaly record in the system log to prompt maintenance personnel to check or replace it during the next shutdown.

[0098] Specifically, in step S103, the approximate substitute value calculated in step S102 is input into the subsequent preprocessing, causal decoupling analysis, and lifetime prediction processes, enabling the entire method to continue operating during sensor failure and avoiding prediction interruptions due to data loss. Simultaneously, a sensor anomaly record is generated in the system log, including at least the faulty sensor identifier, fault occurrence time, fault type (loss or exceeding limits), the substitute value calculation method used, and the substitute value data. This record can be pushed to the maintenance management platform via industrial Ethernet, prompting maintenance personnel to perform a physical inspection or replacement of the sensor during the most recent planned downtime. Once the sensor resumes normal data output, the method automatically switches back to using real data and stops generating substitute values.

[0099] Example 4 Based on the above embodiment 1, please refer to Figure 4This embodiment provides another predictive maintenance method for equipment based on AI technology. The same content as in Embodiment 1 will not be repeated here; the difference lies in: In step S400, the process of constructing a time series based on the degradation feature components of each device at different times also includes the following steps S401-S402: S401: If the independent degradation feature component at the current time is less than the independent degradation feature component at the previous time, and the absolute value of the difference between the two is greater than the fourth threshold, an abnormal jump is determined to have occurred.

[0100] Specifically, in step S401, after obtaining the independent degradation characteristic component of a device at the current moment, this step compares it with the component of the device at the previous moment. If the current component is less than the previous component, and the decrease amplitude (i.e., the difference between the previous component and the current component) exceeds a preset fourth threshold, an abnormal jump is determined to have occurred. The fourth threshold should be set according to the normal fluctuation range of the device and the measurement noise level. For example, it can be set to 3 times the standard deviation of the difference between two adjacent points in historical data, or directly set to an empirical value. If the decrease amplitude is less than or equal to the fourth threshold, it is considered to be normal random fluctuation and no correction is required; if it exceeds the threshold, a correction step is performed.

[0101] S402: Revise the independent degradation feature component at the current time step to the sum of the component at the previous time step and a set growth step.

[0102] Specifically, in step S402, after an abnormal jump is detected, this step no longer uses the original calculated value, but replaces it with a corrected value. The correction rule is: the component at the current time step is corrected to the component at the previous time step plus a set growth step size. The growth step size can be estimated based on the average slope of historical normal rising segments, or set to a very small positive number to ensure that the corrected sequence remains monotonically non-decreasing and conforms to physical laws. This correction only applies to abnormal points and does not affect the values ​​of other normal points.

[0103] The above description is merely a preferred embodiment of this application and an explanation of the technical principles employed. Those skilled in the art should understand that the scope of the invention involved in this application is not limited to technical solutions formed by specific combinations of the above-described technical features, but should also cover other technical solutions formed by arbitrary combinations of the above-described technical features or their equivalents without departing from the inventive concept. For example, technical solutions formed by substituting the above features with (but not limited to) technical features with similar functions disclosed in this application.

Claims

1. A predictive maintenance method for equipment based on AI technology, characterized in that, include: Real-time acquisition of operational response data of multiple devices during collaborative operation, including drive current, output torque, deviation between actual position and commanded position of each device, and relative pose correction amount between each device; The system detects the production line's operating status to identify naturally occurring operating condition switching events. These events are then used as quasi-intervention signals. Response data segments within a predetermined time window before and after the occurrence of these quasi-intervention signals are extracted from the operating response data. The collected response data fragments are input into a pre-built causal decoupling model. The causal decoupling model determines the transmission relationship between the response anomalies of each device by comparing and analyzing the response data fragments before and after the intervention signal. Based on the transmission relationship, the degradation feature components of each device are separated from the influence of other devices. Based on the time series of degradation characteristic components of each device at different times, the remaining service life is predicted, and maintenance decisions are output.

2. The AI-based predictive maintenance method for equipment according to claim 1, characterized in that, The monitoring of production line operating status to identify naturally occurring operational condition switching events includes: Real-time acquisition of production line operation status parameters, including current product model identifier, tool identifier, set cycle time, and workpiece gripping success rate; When one or more preset conditions are detected, it is determined to be a working condition switching event; the preset conditions include changes in robot trajectory parameters caused by product model switching, a step change in clamping force caused by automatic tool replacement, a sudden change in motion speed during production cycle adjustment, or the start and end of a retry sequence after workpiece gripping failure.

3. The AI-based predictive maintenance method for equipment according to claim 1, characterized in that, After extracting the response data segments within a predetermined time window before and after the occurrence of the quasi-intervention signal, a preprocessing step is also included: The extracted response data segments are timestamped and the time when the quasi-intervention signal occurs is marked as time zero. The response data segments before time zero are taken as the first state sequence, and the response data segments after time zero are taken as the second state sequence.

4. The AI-based predictive maintenance method for equipment according to claim 3, characterized in that, The causal decoupling model determines the transmission relationship between anomalies in each device's response through comparative analysis of response data fragments, including: For each device, compare the maximum mean difference of each operational response data item in the first state sequence and the second state sequence; if the maximum mean difference of any operational response data item exceeds the first threshold corresponding to the device, then mark the device as having a response anomaly. For each first device marked as having an abnormal response, it is paired with all other second devices. The time-series cross-correlation function between the data fluctuations in the second state sequence of the first device and the corresponding data fluctuations in the first state sequence of the second device is calculated, and the maximum correlation coefficient and its corresponding lag time are recorded. If the maximum correlation coefficient exceeds the second threshold and the lag time matches the inherent delay of the production line control compensation, it is determined that at the moment the quasi-intervention signal occurs, the abnormal response of the first device is caused by the degradation of the second device transmitted through control compensation, and the transmission relationship is recorded.

5. The AI-based predictive maintenance method for equipment according to claim 4, characterized in that, The process of separating the degradation feature components of each device, independent of the influence of other devices, based on the transmission relationship includes: For a pair of devices with a transitive relationship, the data fluctuation component corresponding to the transitive relationship in the second state sequence of the first device is subtracted, and the remaining part is used as the independent degradation feature component of the first device at the moment when the quasi-intervention signal occurs. For devices that are determined to have abnormal responses but are not the source of the transmission relationship, the difference between their first state sequence and second state sequence is directly used as the independent degradation feature component of the device at the moment when the quasi-intervention signal occurs. For the second device identified as the source of the transmission relationship, the difference between its first state sequence and its second state sequence is directly used as the independent degradation feature component of the device at the moment the quasi-intervention signal occurs, and it is retained as the identifier of the source of the transmission relationship.

6. The AI-based predictive maintenance method for equipment according to claim 1, characterized in that, The method of predicting the remaining service life of each device based on the time series composed of degradation characteristic components at different times includes: Each time the independent degradation feature component of each device is obtained at the current moment, the independent degradation feature component is appended and stored according to the device identifier and time to form the time series of degradation feature components of each device; The degradation trend function is obtained by fitting the time series of degradation feature components for each device. The time required to reach the failure threshold by extrapolating the degradation trend function is used as the predicted remaining life of the device.

7. The AI-based predictive maintenance method for equipment according to claim 5, characterized in that, The output maintenance decision includes: For each device, the number of times the device acts as the source of a transmission relationship or is affected by other devices is counted based on the transmission relationship, and the contribution of the device is calculated. The device with the largest contribution is identified as the root cause device, the device with a contribution less than the third threshold is identified as a device without anomalies, and the rest are identified as subordinate devices; maintenance instructions are generated only for the root cause device, and coupling interference prompts are generated for subordinate devices. Based on the predicted remaining lifespan of the root cause equipment, a three-level early warning is generated. When the early warning level reaches the second level or above, a temporary task rescheduling plan is generated to transfer the process of the root cause equipment to other equipment or manual workstations.

8. The AI-based predictive maintenance method for equipment according to claim 1, characterized in that, During the extraction of response data segments within a predetermined time window before and after the occurrence of the quasi-intervention signal, if a sufficient number of condition switching events are not captured after a continuous detection time exceeding a preset duration, an active detection mode is triggered, including: Send short-duration, small-amplitude additional motion commands to the production line control system. The amplitude is limited to the allowable range of the equipment's static accuracy, and the duration does not exceed the set duration. Mark the start time of the additional motion command as the quasi-intervention signal occurrence time, and extract response data segments within the predetermined time window before and after the condition switching event in the same way as the condition switching event; After the active detection command is completed, the original control command of the production line is restored, and the response data fragment generated by the active detection is marked as a human intervention sample for use by the causal decoupling model.

9. The AI-based predictive maintenance method for equipment according to claim 1, characterized in that, During the real-time acquisition of operational response data from various devices, if continuous data loss or a value exceeding a preset reasonable range is detected from a certain sensor, a data recovery process is triggered, including: Based on the established inter-device transmission relationship in the causal decoupling model, find the real-time data of other sensors that are coupled with the faulty sensor; Using the aforementioned coupling relationship and data from other sensors, an approximate replacement value for the faulty sensor is calculated through fitting. The approximate substitution value is used in subsequent calculations, and a sensor anomaly record is generated in the system log to prompt maintenance personnel to check or replace it during the next shutdown.

10. The AI-based predictive maintenance method for equipment according to claim 1, characterized in that, The process of constructing a time series based on the degradation feature components of each device at different times also includes: If the independent degradation feature component at the current time is less than the independent degradation feature component at the previous time, and the absolute value of the difference between the two is greater than the fourth threshold, an abnormal jump is determined to have occurred. Revise the current independent degradation feature component to the sum of the component from the previous time step and a set growth step.

Images