Data-adaptive device production process defect identification method and system

Abstract

The application relates to the technical field of industrial equipment monitoring and defect diagnosis, and particularly discloses a data-adaptive equipment production process defect identification method and system. A multi-dimensional delay dynamic benchmark is constructed, and a joint decision based on a delay gradient and fluctuation entropy is adopted to realize intelligent classification and rapid routing of sudden failures, environmental disturbances and composite hidden dangers. The system adopts a three-layer nested diagnosis mechanism of abnormal dimension screening, cause-effect chain tracing and defect mode matching to accurately locate the root cause and drive a three-order adaptive response closed loop of instantaneous inhibition, parameter compensation and model correction to minimize production interruption. Meanwhile, the system periodically calculates a production line health index to realize self-evaluation and collaborative optimization of key parameters, so that the monitoring model can continuously evolve with changes in the equipment state and the environment, and finally form an integration of real-time sensing, intelligent diagnosis and adaptive optimization, thereby significantly improving the defect identification accuracy, system stability and overall operation efficiency.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of industrial equipment monitoring and defect diagnosis, and in particular to a data-adaptive equipment production process defect identification method and system. BACKGROUND

[0002] In the field of industrial automation and intelligent manufacturing, real-time monitoring and defect identification of equipment production processes are key links to ensure product quality and production efficiency. The monitoring methods widely used at present mainly rely on setting fixed thresholds for single or a few operating parameters to perform out-of-limit alarm, or on statistical process control methods to perform trend monitoring. Although these methods can alarm obvious abnormalities, their defect identification capabilities are heavily dependent on prior experience and static models, and they are difficult to adapt to parameter drift caused by dynamic running state of equipment, process switching, environmental disturbance and component aging. Existing technologies generally lack the ability to perceive the coordinated changes of multiple physical quantities and multiple time dimensions, resulting in a high rate of missed judgment of early implicit defects, intermittent failures and complex abnormalities. In addition, traditional methods often use direct alarm or simple shutdown strategies after abnormality judgment, lack intelligent classification and root cause tracing of abnormal types, and cannot provide clear diagnostic guidance for maintenance personnel, nor do they have the ability to adaptively suppress and compensate, resulting in frequent production interruptions and low maintenance efficiency.

[0003] Existing defect identification systems are often relatively flat in decision logic, and cannot effectively distinguish between sudden hard failures, gradual drifts and complex hidden dangers with multiple sources of coupling. Their response mechanism is usually one-way and fixed, either resetting parameters or triggering shutdown, and lacks an intelligent response strategy of staging, verification and gradual escalation. At the same time, the system generally lacks self-evaluation and optimization capabilities based on long-term running data, and the monitoring model and threshold parameters are fixed once set, and cannot evolve with changes in equipment state and production environment, resulting in a gradual decline in system monitoring performance over time.

[0004] The above existing technologies face the core difficulties of poor adaptability, high false alarm rate, unclear diagnosis, over-reaction or insufficient response, and lack of continuous optimization capability in actual application. SUMMARY

[0005] In order to overcome the problems of the prior art that the existing defect identification method relies on fixed threshold, lacks multi-dimensional dynamic perception ability, is difficult to distinguish and adapt to complex abnormalities, the response mechanism is single and solidified, and the system lacks self-optimization ability, the present application provides a data adaptive equipment production process defect identification method and system. By constructing a multi-dimensional delay benchmark for dynamic perception, combining the joint decision of delay gradient and fluctuation entropy to realize intelligent classification of abnormalities, and based on three-layer nested diagnosis and three-order adaptive response mechanism for precise root cause positioning and progressive repair, finally relying on the production line health index to realize the optimization of system parameters. The sudden failure, parameter drift and complex hidden danger in the equipment production process are accurately identified, intelligently classified, adaptively inhibited and the whole process management of system continuous evolution is realized.

[0006] The technical solution of the present application is as follows:

[0007] According to an aspect of the present application, a data adaptive equipment production process defect identification method is provided, comprising:

[0008] Switching the device state as a trigger, collecting multi-dimensional original delay data and modeling, generating initial multi-dimensional delay benchmark and its tolerance range representing the health state of the device; in the subsequent production rhythm, the actual delay of the current rhythm is calculated in real time and compared with the benchmark, if the deviation exceeds the safety threshold, a first-level abnormal event is triggered;

[0009] For the first-level abnormal event, the delay gradient and fluctuation entropy of the current rhythm are calculated; according to the comparison results of the delay gradient and fluctuation entropy with the preset threshold, the first-level abnormal event is classified and routed to different processing paths, the paths include: triggering instantaneous inhibition response, triggering benchmark re-modeling, triggering deep diagnosis, and determining as false alarm and restoring monitoring;

[0010] For the complex hidden danger routed to the deep diagnosis path, three-layer nested diagnosis is performed, including: screening out the core abnormal dimension set that deviates significantly; based on the transfer entropy, analyzing the causal flow direction between each dimension in the core abnormal dimension set, tracing back to the source abnormal dimension and abnormal propagation path; matching the core abnormal dimension set, the source abnormal dimension and the propagation path characteristics with the historical defect mode library, and outputting specific defect type instructions;

[0011] According to the defect type instruction, a progressive three-order adaptive response is performed: the first order executes instantaneous inhibition operation and verifies immediately; if invalid, enter the second order, compensate and adjust the control or process parameters and verify; if still invalid, enter the third order, locally correct the dimension statistical characteristics of the initial multi-dimensional delay benchmark and verify; if the three-order response fails, manual intervention is triggered;

[0012] Periodic statistics of the defect triggering frequency, defect judgment success rate and response success rate are used to calculate the production line health index; according to the comparison result of the production line health index and the health threshold value, the collaborative optimization suggestion of the system parameters and the model is generated, and is applied in the next time of benchmark modeling.

[0013] As a further selection of the application method, the device state switching includes one of the following events: the device cold start or hot start is completed and the self-check is passed, the mold replacement is completed and it is confirmed that the installation is in place, the process formula is switched and the new parameters are successfully downloaded to the controller; after entering the benchmark modeling period, the system reserves a preheating period, and starts to collect multi-dimensional original delay data after the device runs stably.

[0014] As a further selection of the application method, the multi-dimensional original delay data includes the following five dimensions: electrical drive response time, mechanical transmission delay, control instruction execution time, sensor feedback period, system inertia delay under load mutation; the initial multi-dimensional delay benchmark calculation formula is: , wherein is the initial five-dimensional delay benchmark value, , , , , is the preprocessed data mean of the electrical drive response time , the mechanical transmission delay , the control instruction execution time , the sensor feedback period , and the system inertia delay under load mutation , , , , , is the initial dynamic weight coefficient.

[0015] As a further selection of the application method, for the current beat , the delay gradient is: ; wherein is the delay of the current beat, is the delay of the last beat, is the nominal time period of the beat;

[0016] The calculation of the fluctuation entropy is: ; wherein, is the number of intervals, is the interval index, is the frequency of the delay value falling in the interval.

[0017] As a further choice of the application method, the four-path classification is performed according to the comparison result of the delay gradient and the fluctuation entropy with the preset threshold, and specifically includes:

[0018] If the delay gradient is higher than the high gradient threshold and the fluctuation entropy is lower than the low entropy threshold, it is determined as a sudden hard failure, and a transient suppression response is triggered;

[0019] If the delay gradient is lower than the low gradient threshold and the fluctuation entropy is higher than the high entropy threshold, it is determined as an environmental disturbance or a parameter drift, and a reference re-modeling is triggered;

[0020] If the delay gradient is higher than the high gradient threshold and the fluctuation entropy is higher than the high entropy threshold, it is determined as a complex hidden danger, and a deep diagnosis is triggered;

[0021] If the delay gradient is lower than the low gradient threshold and the fluctuation entropy is lower than the low entropy threshold, it is determined as a false alarm, and the monitoring is resumed.

[0022] As a further choice of the application method, in the three-layer nested diagnosis:

[0023] When screening the core abnormal dimension set, the standard deviation of the delay of each dimension in the sliding window centered on the abnormal beat is calculated, and the gradient and the entropy of the standard deviation are simultaneously screened based on whether the gradient and the entropy exceed the dimension abnormal threshold;

[0024] When tracing the cause-effect chain, the cause-effect net influence between the dimensions in the core abnormal dimension set is calculated by using the transfer entropy analysis method, and the source abnormal dimension and the abnormal propagation path are determined;

[0025] When matching the defect mode, the similarity between the current feature and the historical defect mode library is matched, if the matching is successful, the defect type instruction is output, otherwise the defect mode library is updated by manual judgment.

[0026] As a further choice of the application method, in the three-order adaptive response:

[0027] The first-order transient suppression includes sending an emergency stop signal, cutting off the power supply of the fault module or switching to a safe backup parameter, and immediately collecting the next beat delay data for verification;

[0028] The second-order parameter compensation includes adjusting the control loop gain or the process parameter, and verifying whether it is restored to stability after several beats;

[0029] The third-order model local correction includes updating the statistical characteristics of the corresponding dimension in the initial multi-dimensional delay reference according to the defect root cause, and verifying under the corrected reference.

[0030] As a further choice of the application method, the calculation formula of the production line health index is: ; wherein , , is a weight coefficient, and ; represents normalizing the defect trigger frequency to the interval [0, 1], is a defect judgment success rate, is a third-order response success rate.

[0031] As a further option of the application method, the collaborative optimization suggestion is generated based on a line health index being lower than a health threshold, analyzing weak links in the defect trigger frequency, the judgment success rate and the response success rate, automatically adjusting a safety threshold, a decision threshold or a reference parameter, and applying after being confirmed by a human being when triggering reference modeling in the next device state switching.

[0032] Another aspect of the application provides a data-adaptive device production process defect identification method system, which comprises:

[0033] A reference modeling and anomaly detection module is configured to collect multi-dimensional original delay data and model the same based on device state switching as a trigger, to generate an initial multi-dimensional delay reference and a tolerance range thereof representing a device health state; in a subsequent production beat, to calculate an actual delay of the current beat in real time and compare the same with the reference, and to trigger a first-level anomaly event if a deviation exceeds a safety threshold;

[0034] A decision routing module is configured to calculate a delay gradient and a fluctuation entropy of the current beat after the first-level anomaly event is triggered; and to classify and route the first-level anomaly event to different processing paths according to a comparison result of the delay gradient and the fluctuation entropy with a preset threshold, the paths including triggering an instantaneous suppression response, triggering reference remodeling, triggering a deep diagnosis and determining as a false alarm and resuming monitoring;

[0035] A deep diagnosis module is configured to perform three-layer nested diagnosis on a complex hazard routed to the deep diagnosis path, including: screening out a core abnormal dimension set that deviates significantly; analyzing a causal flow direction among dimensions in the core abnormal dimension set based on a transfer entropy, and tracing a source abnormal dimension and an abnormal propagation path; matching features of the core abnormal dimension set, the source abnormal dimension and the propagation path with a historical defect mode library, and outputting a specific defect type instruction;

[0036] An adaptive response execution module is configured to execute a gradual three-stage adaptive response according to the defect type instruction: a first stage performs an instantaneous suppression operation and verifies immediately; if invalid, a second stage is entered to compensate and adjust a control or process parameter and verify; if still invalid, a third stage is entered to locally correct a dimension statistical feature of the initial multi-dimensional delay reference and verify; if all the three stages fail, a human intervention is triggered.

[0037] A system cooperative optimization module is configured to periodically count the defect triggering frequency, the defect judgment success rate and the response success rate, calculate the production line health index, generate the cooperative optimization suggestion of the system parameters and the model according to the comparison result of the production line health index and the health threshold, and apply the cooperative optimization suggestion in the next benchmark modeling.

[0038] The application has the following beneficial effects:

[0039] The application significantly improves the accuracy and classification precision of defect identification through multi-dimensional delay dynamic modeling and gradient-entropy joint decision mechanism. The traditional method relies on fixed threshold, and the recognition rate of complex and intermittent faults is low. However, the method can distinguish sudden hard faults, environmental disturbances and complex hidden dangers based on real-time calculation of gradient and fluctuation entropy based on five-dimensional delay benchmark, and realize four-path intelligent routing. In practical application, the false alarm rate can be reduced by about 40%-50%, the detection rate of latent defects can be increased by more than 30%, and the defect type judgment success rate can be stabilized at more than 92% after continuous optimization.

[0040] The system adopts three-layer nested diagnosis and three-order adaptive response strategy, realizes closed-loop processing from abnormal identification to autonomous inhibition, and greatly reduces manual intervention and production interruption time. Through abnormal dimension screening, causal chain tracing and defect mode matching, the system can locate the root cause and sequentially execute instantaneous inhibition, parameter compensation and model correction, and gradually upgrade the response. Implementation data shows that the average fault repair time can be shortened by about 50% by using the strategy, the three-order response success rate can reach 85%-90%, and the unplanned downtime of the production line caused by defects can be reduced by about 35%.

[0041] The application has the ability of continuous self-evaluation and cooperative optimization, and realizes long-term stability and adaptive evolution of the system by periodically calculating the production line health index and dynamically adjusting the benchmark and threshold. The fixed system parameters of the traditional system can easily lead to performance degradation in the later period, while the system of the application can always adapt to equipment degradation and environmental changes by adjusting the weight, threshold and benchmark model according to the value optimization every period. Long-term operation data shows that the health index of the system can still be maintained at more than 0.8 after continuous operation for 6 months, the defect triggering frequency gradually decreases, and the overall reliability of the system is improved by about 25%.

[0042] In summary, the application fundamentally solves the core difficulties of the traditional threshold monitoring method, such as poor adaptability, high false alarm and missed alarm rate, unclear diagnosis and rigid response. The application realizes the fundamental change from the static alarm relying on artificial experience to the dynamic self-learning, self-adaptation and self-optimization driven by data, and significantly improves the overall stability, product quality and operation efficiency of the production system. BRIEF DESCRIPTION OF DRAWINGS

[0043] Figure 1 It is a whole schematic diagram of the data adaptive equipment production process defect identification method.

[0044] Figure 2 Flowchart of the method for identifying process defects of data-adaptive equipment production S100;

[0045] Figure 3 Flowchart of the method for identifying process defects of data-adaptive equipment production S200;

[0046] Figure 4 Flowchart of the method for identifying process defects of data-adaptive equipment production S300;

[0047] Figure 5 Flowchart of the method for identifying process defects of data-adaptive equipment production S400;

[0048] Figure 6 Flowchart of the method for identifying process defects of data-adaptive equipment production S500. DETAILED DESCRIPTION

[0049] The technical solutions in the embodiments of the present application will be described clearly and completely below with reference to the drawings in the embodiments of the present application. Obviously, the described embodiments are only some of the embodiments of the present application, but not all the embodiments of the present application. Based on the embodiments in the present application, all other embodiments obtained by those skilled in the art without creative labor fall within the scope of protection of the present application.

[0050] In a modern intelligent manufacturing system, stable operation of production equipment is the core to guarantee product quality and production efficiency. If subtle defects in the process cannot be identified and inhibited in time and accurately, batch quality accidents and production interruptions are likely to occur. Traditional methods relying on fixed thresholds or periodic inspection are difficult to adapt to complex working conditions such as dynamic degradation of equipment, environmental disturbance and sudden failure, and have inherent defects such as response lag, high false alarm rate and inability to locate the root cause. The technical core of the present application is to build a closed-loop system for identifying defects of equipment that can learn and optimize itself by multi-dimensional delayed dynamic modeling, gradient-entropy joint decision-making, three-layer nested diagnosis and three-order adaptive response. The theoretical basis of the present application is based on multi-dimensional time series analysis, information entropy theory and adaptive control theory. By fusing, feature extraction and pattern recognition of multi-source heterogeneous delayed data generated during equipment operation, accurate positioning, intelligent classification and progressive repair of defects are realized.

[0051] The core theory is as follows:

[0052] Initial five-dimensional delayed benchmarking: Modeling:

[0053] Specifically, the overall response delay of a device within a single production cycle is jointly determined by multiple key dimensions. Let the five dimensions of delay data collected within a cycle be: electrical drive response time , mechanical transmission to position delay , control command execution time consumption , sensor feedback period , and system inertia delay under load mutation . The initial five-dimensional delay benchmark is: ; where , , , , is the dynamic weight coefficient of each dimension, satisfying . The initial weight can be set according to the device mechanism or expert experience, and is adaptively adjusted in the optimization section of S500.

[0054] Delay gradient calculation:

[0055] The delay gradient is used to measure the degree of delay change, i.e. the change rate of delay within a unit time or between adjacent cycles. For the current cycle , its delay gradient is defined as: ; where is the delay of the current cycle, is the delay of the previous cycle, is the nominal time period of the cycle. A high gradient value usually indicates a sudden failure or state mutation.

[0056] Fluctuation entropy calculation:

[0057] Fluctuation entropy is used to quantify the randomness and uncertainty of the delay sequence, reflecting the degree of system state disorder. First, the probability distribution of delay is calculated within a time window containing historical cycles. The delay value range is divided into equal intervals, and the frequency of delay values falling into each interval is counted. Then the fluctuation entropy is calculated by the information entropy formula: . The higher the fluctuation entropy , the more random and irregular the delay fluctuation, which may be caused by continuous environmental disturbance or slow parameter drift; a low entropy value indicates that the fluctuation is concentrated and regular, which may indicate a specific failure mode.

[0058] Production line health index calculation:

[0059] is an index for evaluating the health state of the system, and the value range is [0, 1], and the higher the value is, the healthier the system is. Defect trigger frequency , defect type judgment success rate and third-order response success rate together determine: ; wherein, , , is a weight coefficient, and . represents the normalization processing of the defect trigger frequency , which is mapped to the interval [0, 1], and the higher the frequency is, the greater the negative contribution to is. and are success rate ratios, which directly contribute to .

[0060] The specific embodiments of the application will be described in detail below.

[0061] Example one:

[0062] Please refer to Figure 1 , which shows a data adaptive device production process defect identification method provided by an embodiment of the application, and the method comprises:

[0063] S100: triggered by device state switching, a multi-dimensional delay reference is constructed and primary anomaly detection is performed.

[0064] S200: based on the joint decision of delay gradient and fluctuation entropy, four-path classification and routing are performed on the primary anomaly.

[0065] S300: for complex hidden dangers, three-layer nested deep diagnosis of anomaly dimension screening, cause-effect chain tracing and defect mode matching is performed.

[0066] S400: according to the defect type instruction, three-order adaptive response of transient suppression, parameter compensation and model correction is performed.

[0067] S500: periodically evaluate the health index of the production line, and perform collaborative optimization of system parameters and models based on the evaluation results.

[0068] The specific scheme is as follows:

[0069] In a data adaptive device production process defect identification method, S100 establishes a dynamic reference that can represent the running characteristics of the device in a stable and healthy state, and sets a sensitive and reliable initial anomaly detection mechanism.

[0070] Referring to Figure 2 , which shows a flowchart of an exemplary data-adaptive device production process defect identification method S100 of the present application, the contents of which include:

[0071] S110: Trigger reference modeling period and collect multi-dimensional raw delay data.

[0072] In the present application, the system continuously monitors the running state of the device, and when a restart signal is detected, it enters the reference modeling period. The goal of the reference modeling period is to capture the fingerprint data of the device in a known good state.

[0073] In one possible implementation, the restart signal is defined as one of the following three events:

[0074] (1) The device cold start or hot start is complete, and the self-check program passes.

[0075] (2) The mold replacement process is completed, and it is confirmed that the new mold is installed in place.

[0076] (3) The process recipe is switched, and the new recipe parameters have been successfully downloaded to the controller.

[0077] In one possible implementation, after entering the reference modeling period, the system does not immediately start modeling, but reserves a warm-up period, and only after the device is fully stable, it starts collecting data for the first stable production period, and simultaneously collects high-speed five-dimensional raw delay signals.

[0078] Specifically, the five-dimensional raw delay signals include:

[0079] Electrical drive response time , i.e. the time from the controller issuing a drive instruction to the motor current reaching the target value. Mechanical transmission to position delay , i.e. the time difference from the start of the drive mechanism to the end of the execution mechanism reaching the target position. Control instruction execution time , i.e. the time spent from the host computer issuing a control instruction to the programmable logic controller executing unit starting to process the instruction, which is obtained by network packet analysis. Sensor feedback period , i.e. the time interval between two adjacent effective sensor feedback signals of the key physical quantity. System inertia delay under load mutation , i.e. the additional delay time from the start of load change to the recovery of spindle speed in a specific test beat with a preset step increase in system load rate.

[0080] S120: Generate initial five-dimensional delay reference .

[0081] In the present application, the original multi-dimensional delay data collected is pre-processed to generate a reference value for subsequent beat comparison.

[0082] In one possible implementation, the initial five-dimensional delay reference The generating step includes:

[0083] The delay data of each dimension is pre-processed by cleaning, aligning and fusing.

[0084] The mean value of the pre-processed data of each dimension is calculated , , , , .

[0085] The initial dynamic weight coefficient is set according to the prior knowledge of the influence of each dimension delay on the final product quality , , , , .

[0086] The initial five-dimensional delay reference value is calculated , which is calculated by the following formula: .

[0087] The mean value and the standard deviation of each dimension delay are recorded to form a five-dimensional reference vector and a tolerance range vector , The tolerance range of the system inertia delay under the load mutation, respectively, is the electrical driving response time, the mechanical transmission to the delay, the control instruction execution time, the sensor feedback period and the load mutation.

[0088] After modeling, the system flag is switched to enter the normal monitoring mode.

[0089] S130: Real-time comparison and first-level abnormal event determination.

[0090] In the present application, the actual delay of the current beat is calculated by the system in real time and compared with the initial five-dimensional delay reference value .

[0091] In one possible implementation, for the beat , according to the same weighted fusion formula in S120, the five-dimensional real-time data collected in the current beat is used to calculate . The absolute deviation is calculated. The safety threshold is preset, and the safety threshold Determine the workflow.

[0092] Specifically, based on the above implementation method, the workflow judgment step includes:

[0093] like If the current cycle is considered to be running normally, the system will only record the data and update the data used to calculate the fluctuation entropy. The historical sequence window.

[0094] like If an abnormal cycle occurs, the current abnormal cycle will be immediately marked as a Level 1 abnormal event. The system will freeze the output of the current abnormal cycle and trigger an interrupt, transferring control and related data to the S200 process.

[0095] In a data-adaptive equipment manufacturing process defect identification method, S200 utilizes gradients. Entropy It can quickly identify first-level anomalies and route them to different processing pipelines.

[0096] Please refer to Figure 3 The diagram illustrates a flowchart of an exemplary data-adaptive equipment manufacturing process defect identification method S200 of this application, the contents of which include:

[0097] S210: Freeze output and calculate characteristic metrics and .

[0098] In this invention, upon receiving a Level 1 anomaly event signal from S100, the output of the current anomaly clock cycle is physically or logically isolated. Two key characteristic indicators used for decision-making are then calculated: the delay gradient of the current clock cycle. And the fluctuation entropy that reflects the recent disorder of the system state .

[0099] In one possible implementation, for the first Delay gradient of each beat The calculation is as follows: ;in, This is a delay to the normal tempo.

[0100] In one possible implementation, for the first The entropy of the beat The calculation is as follows: ;in, For the number of intervals, For range indexes, from 1 to... , For the delay value to fall into the first Frequency within a certain interval.

[0101] S220: based on the value of the four-path joint decision.

[0102] In the present application, a high gradient threshold and a low gradient threshold are set, as well as a high entropy threshold and a low entropy threshold . The calculated and are compared with the preset thresholds to form four decision branches.

[0103] In one possible implementation, the decision logic of the four decision branches is as follows:

[0104] 1. If and :

[0105] It is determined as a sudden hard failure, which is manifested as a sudden and dramatic increase in delay, but the overall system has been running smoothly and orderly in recent time, corresponding to a clear component damage.

[0106] The specific action is to send a structural defect type instruction to S400. The instruction contains all the data of the abnormal beat and the preliminary classification label of the sudden hard failure, indicating that S400 starts the targeted first-order transient suppression and second-order parameter compensation.

[0107] 2. If and :

[0108] It is determined as an environmental disturbance or parameter drift, which is manifested as a delay that does not suddenly change greatly, but the recent fluctuations are very random and disordered, which may be caused by slow temperature rise between machines, change of hydraulic oil viscosity, and sporadic interference of sensors, etc.

[0109] The specific action is to feedback to S100 to suggest relaxing and restarting the reference calibration. The system considers that the current initial five-dimensional delay reference may not be applicable to the changed environment, and starts a new reference modeling cycle to adjust the initial dynamic weight coefficient to establish an initial five-dimensional delay reference that conforms to the current working condition. And suggest S100 to relax the safety threshold .

[0110] 3. If and :

[0111] It is determined as a complex hidden danger. The complex hidden danger is the most complex, which has a sudden delay growth and is accompanied by recent instability of the system. It may be that the main fault triggers a chain reaction, or multiple potential problems break out at the same time.

[0112] Specific action: send composite defect to be further investigated instruction to S300. The composite defect to be further investigated instruction contains all relevant data within the abnormal beat and the recent window, requiring S300 to conduct in-depth, multi-dimensional root cause analysis.

[0113] 4. If and :

[0114] Determine that the false alarm or transient interference has recovered, and the system state has spontaneously recovered to normal.

[0115] Specific action: revoke the first-level abnormal event mark of the current beat, unfreeze the output, and notify S100 to resume using the original baseline and threshold for monitoring.

[0116] In a data-adaptive device production process defect identification method, S300 is a specialized diagnostic room for composite problems. S300 uses a progressive three-layer analysis to locate abnormal dimensions, analyze causal relationships, and finally match specific defect patterns.

[0117] Please refer to Figure 4 , which shows the flowchart of an exemplary data-adaptive device production process defect identification method S300 of the present application, the content of which includes:

[0118] S310: First layer nested judgment - abnormal dimension screening.

[0119] In the present application, abnormal dimension screening is to locate the core dimension set that causes the current composite defect anomaly from five dimensions.

[0120] In one possible implementation, the system extracts a time sequence sliding window centered on the abnormal beat . For each beat in the window, calculate the deviation of each dimension's delay from the corresponding mean value in the corresponding dimension baseline value , and standardize to obtain the standardized value of each dimension. Then, calculate the gradient and entropy of the standardized deviation sequence of each dimension within the time sequence sliding window. Set the dimension abnormality threshold, if the gradient and entropy of a certain dimension exceed the dimension abnormality threshold at the same time, mark it as significantly deviating from the dimension. All marked dimensions constitute the core abnormal dimension set .

[0121] S320: Second layer nested judgment - causal chain tracing.

[0122] In the present application, after the core abnormal dimension set is determined, the causal relationship between the dimensions in the core abnormal dimension set is further analyzed to find the source abnormal dimension and the affected dimensions.

[0123] In one possible implementation, the dimension delay time series in the core abnormal dimension set is analyzed based on a transfer entropy tracing method. The transfer entropy tracing method measures the degree of reduction in uncertainty of one time series about the future state of another time series, and can be used to infer the causal direction.

[0124] Specifically, the transfer entropy tracing method calculates the transfer entropy from any two dimensions and and to to , and the transfer entropy from to to . The greater the transfer entropy value, the stronger the causal impact. The direction with the greatest causal net impact indicates the main causal flow. The dimension is determined as the causal source dimension. At the same time, according to the significant causal strength, the propagation path of the abnormality from the causal source dimension to other dimensions is outlined.

[0125] S330: Third layer of nested judgment-defect mode matching.

[0126] In the present application, the features obtained from the first two layers of analysis, i.e., the core dimension set, the causal source, and the propagation path, are matched with the historical accumulated defect mode library to find the most consistent known defect type.

[0127] In one possible implementation, the system calculates the similarity between the features obtained from the current analysis and the features of each record in the library defect mode library. If the highest similarity exceeds the matching threshold, it is determined that the matching is successful, and the corresponding specific defect type instruction is output to S400. If the matching fails, a new defect mode is generated by manual judgment, and the corresponding specific defect type instruction is output to S400.

[0128] In one data adaptive device production process defect identification method, S400 receives the explicit defect instruction from S200 or S300, and adopts a progressive and verified response strategy to eliminate defects at the fastest speed and with the least cost.

[0129] Please refer to Figure 5 , which shows the flowchart of an exemplary data adaptive device production process defect identification method S400 of the present application, the contents of which include:

[0130] ​S410: Perform first stage response - Instantaneous suppression.

[0131] In the present invention, for all defects that require immediate action, the fastest means is first taken to prevent the situation from getting worse or producing defective products.

[0132] In one possible implementation, according to the defect type instruction, the instantaneous suppression measures include:

[0133] Sending an emergency stop signal to the corresponding actuator.

[0134] Cutting off the power supply of a suspected faulty module.

[0135] Forcibly switching the process parameters to a safe backup value.

[0136] Specifically, after the measures are executed, the system waits for a very short time and forcibly performs effect verification. The verification method is to immediately collect the five-dimensional delay data of the next beat, calculate , and check whether is satisfied. If the verification is successful, it means that the instantaneous suppression has solved the problem, the process is terminated, and the event and response are archived. If the verification fails, the second stage response is entered.

[0137] S420: Perform second stage response - Parameter compensation.

[0138] In the present invention, when the instantaneous suppression is ineffective or the defect itself belongs to the performance deviation category, an attempt is made to compensate for the deviation by adjusting the control parameters.

[0139] In one possible implementation, parameter compensation compensates for existing performance deviations by adjusting control loops or process parameters, trying to restore stability in the existing state of the system. Effect verification is performed, and if successful, the process is terminated and archived. If it fails, the third stage is entered.

[0140] For example, for a servo motor gain drift defect, the strategy is to increase the current loop proportional gain by 5%. The system automatically modifies the corresponding parameters and allows the device to run for several beats to adapt to the new parameters. Then, effect verification is forcibly performed: calculate the average delay of these 5 beats, check whether it returns to the normal range, and whether the fluctuation is reduced. If successful, the process is terminated and archived. If it fails, the third stage is entered.

[0141] S430: Perform third stage response - Model local correction.

[0142] In the present invention, when the current two stages are ineffective, it indicates that the performance of the device has changed permanently or long-term to some extent, and the initial five-dimensional delay reference and its weights are no longer suitable for the current state. In this stage, the five-dimensional delay reference is adjusted online and locally.

[0143] In a possible implementation, the model local correction is to update the statistical characteristics of the corresponding dimension in the initial five-dimensional delay reference according to the defect root cause. After the correction, the system runs under the corrected reference and forces the effect verification to evaluate whether the delay is stable in the corrected expected range. If successful, the model parameters in the system configuration library are updated, and the process terminates. If the third-order response is executed successfully but the verification fails, it is determined as an unknown serious defect that cannot be automatically repaired, triggering the highest level of audible and light artificial alarm, and all data, analysis logs and response records are packaged and provided to the maintenance personnel.

[0144] In a data adaptive device production process defect identification method, S500 jumps out of the perspective of single defect processing, evaluates from the perspective of longer time period and system overall performance, and optimizes key parameters based on data driving to realize self-evolution of the system.

[0145] Please refer to Figure 6 which shows a flowchart of an exemplary data adaptive device production process defect identification method S500 of the present application, the content of which includes:

[0146] S510: Periodic statistics and line health index Calculation.

[0147] In the present application, the system sets two trigger conditions: fixed period or cumulative production time reaches threshold. When any condition is met, the health evaluation process is started.

[0148] In a possible implementation, the health evaluation process is based on the following indicators, including:

[0149] Defect trigger frequency That is, the total number of first-level abnormal events / total production beats in the evaluation period.

[0150] Defect type judgment success rate That is, the proportion of correct defects instructions output by S200 or S300 after manual review or final effect confirmation.

[0151] Three-order response success rate That is, the number of times of successful verification of response action in a certain order / the total number of times of triggering S400 in S400.

[0152] According to the formula Calculation .

[0153] S520: Generate and confirm collaborative optimization proposal, and execute optimization.

[0154] In the present application, the calculated is compared with the preset health threshold Comparison is made. If , the system is considered healthy, and only a record is made for this round.

[0155] If , optimization is made. The system analyzes each sub-index and locates the weak link.

[0156] For example, if is too low, it means that the decision threshold , , etc. is not reasonable; if is too high, it means that the reference or safety threshold is too sensitive. The system automatically generates a collaborative optimization proposal in the form of a structured report.

[0157] In one possible implementation, all optimization proposals need to be reviewed and approved by a maintenance engineer at a manual confirmation interface. The engineer can accept, modify, or reject the proposal. The final optimization instruction after confirmation is officially written into the system configuration library.

[0158] In one possible implementation, the new parameters written into the system configuration library do not take effect immediately and affect the current operation. They are marked as to be applied. When the system next detects an event that triggers reference reconstruction, such as device startup, mold change, or process switching, the following operations will be performed:

[0159] Read the latest final optimization instruction from the system configuration library.

[0160] Apply the final optimization instruction as an initial constraint condition or priori knowledge to the new round of reference modeling period.

[0161] Embodiment II:

[0162] A data-adaptive device production process defect identification system, comprising:

[0163] A reference modeling and anomaly detection module for collecting multi-dimensional raw delay data and modeling with device state switching as a trigger, generating an initial multi-dimensional delay reference representing the health status of the device and its tolerance range; in the subsequent production cycle, the actual delay of the current cycle is calculated in real time and compared with the reference, if the deviation exceeds the safety threshold, a first-level anomaly event is triggered;

[0164] A decision routing module for calculating the delay gradient and fluctuation entropy of the current cycle after the first-level anomaly event is triggered; according to the comparison result of the delay gradient and fluctuation entropy with the preset threshold, the first-level anomaly event is classified and routed to different processing paths, including triggering transient suppression response, triggering reference re-modeling, triggering deep diagnosis, and determining as false alarm and resuming monitoring;

[0165] The deep diagnosis module is configured to perform three-layer nested diagnosis on the complex hidden danger routed to the deep diagnosis path, including: screening out a core abnormal dimension set that is significantly deviated; analyzing the causal flow direction among dimensions in the core abnormal dimension set based on transfer entropy to trace a source abnormal dimension and an abnormal propagation path; and matching features of the core abnormal dimension set, the source abnormal dimension and the propagation path with a historical defect mode library to output a specific defect type instruction.

[0166] The adaptive response execution module is configured to perform a gradual three-stage adaptive response according to the defect type instruction: a first stage performs a transient suppression operation and immediately verifies; if invalid, a second stage is entered to perform compensation adjustment on a control or process parameter and verify; if still invalid, a third stage is entered to locally correct a dimension statistical feature of an initial multi-dimensional delay reference and verify; if all three stages fail, manual intervention is triggered.

[0167] The system coordination optimization module is configured to periodically count a defect triggering frequency, a defect judgment success rate and a response success rate, calculate a production line health index, generate a coordination optimization suggestion for system parameters and models according to a comparison result of the production line health index and a health threshold, and apply the suggestion in next time reference modeling.

[0168] Those skilled in the art should understand that the embodiments of the present application provide methods, systems or computer program products. Therefore, the present application can be in the form of a complete hardware embodiment, a complete software embodiment or an embodiment combining software and hardware aspects. Moreover, the present application can be in the form of a computer program product implemented on one or more computer storage media (including but not limited to disk storage, CDROM, optical storage, etc.) containing computer program codes. The solutions in the embodiments of the present application are implemented in various computer languages, such as the object-oriented programming language Java and the interpreted scripting language JavaScript.

[0169] The present application is described with reference to flowcharts and / or block diagrams according to the methods, devices (systems) and computer program products of the embodiments of the present application. It should be understood that each flow and / or block in the flowcharts and / or block diagrams and the combination of the flows and / or blocks in the flowcharts and / or block diagrams are realized by computer program instructions. These computer program instructions are provided to the processor of a general-purpose computer, a special-purpose computer, an embedded processor or other programmed data processing equipment to produce a machine, so that the instructions executed by the processor of the computer or other programmed data processing equipment produce a device for realizing the functions specified in the flowchart flow or multiple flows and / or the block diagram block or multiple blocks.

[0170] These computer program instructions can also be stored in a computer- readable memory that can direct a computer or other programmable data processing apparatus to function in a particular manner, such that the instructions stored in the computer-readable memory produce an article of manufacture including instructions which implement the function specified in the flowchart or flowsheets or block or blocks of the block diagrams.

[0171] These computer program instructions can also be loaded onto a computer or other programmable data processing apparatus to cause a series of operational steps to be performed on the computer or other programmable apparatus to produce a computer implemented process such that the instructions which execute on the computer or other programmable apparatus provide steps for implementing the functions specified in the flowchart or flowsheets or block or blocks of the block diagrams.

[0172] Although preferred embodiments of the application have been described herein, substitutions and alterations are possible in view of the disclosure of this application without departing from the spirit and scope of the application. Therefore, it is intended that the appended claims encompass all such substitutions and alterations as fall within the scope of the application. The disclosures of these prior art documents herein are incorporated by reference herein in their entirety.

[0173] Obviously, numerous modifications and variations of the present application are possible in light of the above teachings. It is therefore to be understood that within the scope of the appended claims and their equivalents, the application can be practiced otherwise than as specifically described herein.

Claims

1. A data-adaptive method for identifying defects in equipment manufacturing processes, characterized in that, include: Triggered by equipment state switching, multi-dimensional raw delay data is collected and modeled to generate an initial multi-dimensional delay benchmark and its tolerance range that characterizes the health status of the equipment. In subsequent production cycles, the actual delay of the current cycle is calculated in real time and compared with the benchmark. If the deviation exceeds the safety threshold, a first-level abnormal event is triggered. For Level 1 abnormal events, calculate the delay gradient and fluctuation entropy of the current tick. Based on the comparison results of the delay gradient and fluctuation entropy with the preset threshold, the first-level abnormal events are classified and routed to different processing paths. The paths include: triggering instantaneous suppression response, triggering baseline remodeling, triggering deep diagnosis, and determining it as a false alarm and restoring monitoring. For complex hidden dangers routed to the deep diagnostic path, a three-layer nested diagnosis is performed, including: screening out the core abnormal dimension set that deviates significantly; analyzing the causal flow between each dimension in the core abnormal dimension set based on the propagation entropy, tracing the source abnormal dimension and the abnormal propagation path; matching the characteristics of the core abnormal dimension set, the source abnormal dimension and the propagation path with the historical defect pattern library, and outputting specific defect type instructions. Based on the defect type instruction, execute a progressive three-order adaptive response: the first order performs an instantaneous suppression operation and verifies it immediately; if it is ineffective, proceed to the second order to compensate and adjust the control or process parameters and verify them; if it is still ineffective, proceed to the third order to locally correct and verify the dimensional statistical characteristics of the initial multi-dimensional delay baseline; if all three orders fail, trigger manual intervention. The production line health index is calculated by periodically analyzing the defect trigger frequency, defect judgment success rate, and response success rate. Based on the comparison between the production line health index and the health threshold, suggestions for co-optimization of system parameters and models are generated and applied in the next baseline modeling.

2. The data-adaptive equipment production process defect identification method according to claim 1, characterized in that, The equipment status switching includes one of the following events: the equipment completes a cold start or hot start and passes the self-test; the mold is replaced and confirmed to be installed in place; the process formula is switched and the new parameters are successfully downloaded to the controller; after entering the benchmark modeling period, the system reserves a preheating period, and starts collecting multi-dimensional raw delay data after the equipment runs stably.

3. The data-adaptive equipment production process defect identification method according to claim 1 or 2, characterized in that, The multi-dimensional raw delay data includes the following five dimensions: electrical drive response time, mechanical transmission arrival delay, control command execution time, sensor feedback cycle, and system inertial delay under load mutation; the initial multi-dimensional delay benchmark calculation formula is: ,in, This is the initial five-dimensional delay baseline value. , , , , These are the electrical drive response times. Delay in mechanical transmission positioning Control command execution time Sensor feedback cycle System inertial delay under sudden load changes The mean of the preprocessed data, , , , , These are the initial dynamic weighting coefficients.

4. The data-adaptive equipment production process defect identification method according to claim 1, characterized in that, For the current beat The delay gradient for: ;in For the delay of the current beat, For the delay of the previous beat, The nominal time period of the beat; The fluctuation entropy The calculation is as follows: ;in, For the number of intervals, For range index, For the delay value to fall into the first Frequency within a certain interval.

5. The data-adaptive equipment production process defect identification method according to claim 4, characterized in that, The four-path classification is performed based on the comparison results of the delay gradient and fluctuation entropy with a preset threshold, specifically including: If the delay gradient is higher than the high gradient threshold and the fluctuation entropy is lower than the low entropy threshold, it is determined to be a sudden hard fault, triggering an instantaneous suppression response. If the delayed gradient is lower than the low gradient threshold and the fluctuation entropy is higher than the high entropy threshold, it is determined to be an environmental disturbance or parameter drift, triggering baseline remodeling. If the delay gradient is higher than the high gradient threshold and the fluctuation entropy is higher than the high entropy threshold, it is judged as a compound hidden danger and deep diagnosis is triggered. If the delay gradient is lower than the low gradient threshold and the fluctuation entropy is lower than the low entropy threshold, it is determined to be a false alarm, and monitoring is restored.

6. The data-adaptive equipment manufacturing process defect identification method according to claim 1, characterized in that, In the aforementioned three-layer nested diagnosis: When screening the core abnormal dimension set, the standardized deviation of the delay of each dimension within the sliding window centered on the abnormal beat is calculated, and the screening is based on whether the gradient and entropy of the standardized deviation simultaneously exceed the dimension abnormality threshold. When tracing the causal chain, the transmission entropy analysis method is used to calculate the net causal impact between each dimension in the core anomaly dimension set, and to determine the source anomaly dimension and the anomaly propagation path. During defect pattern matching, the current feature is compared with the historical defect pattern library for similarity. If the match is successful, the defect type instruction is output; otherwise, the defect pattern library is manually judged and updated.

7. The data-adaptive equipment production process defect identification method according to claim 1, characterized in that, In the third-order adaptive response: The first-order transient suppression includes sending an emergency stop signal, cutting off the power to the faulty module or switching to safe backup parameters, and immediately collecting the next cycle delay data for verification; The second-order parameter compensation includes adjusting the control loop gain or process parameters, and verifying whether stability is restored after several cycles. The third-order model local correction involves updating the statistical characteristics of the corresponding dimensions in the initial multi-dimensional delay benchmark based on the root causes of the defects, and then running the verification under the corrected benchmark.

8. The data-adaptive equipment production process defect identification method according to claim 1, characterized in that, The production line health index The calculation formula is: ;in, , , These are the weighting coefficients, and ; Indicates the defect triggering frequency Normalize it to map it to the [0,1] interval. To improve the success rate of defect identification. This represents the success rate of a third-order response.

9. The data-adaptive equipment production process defect identification method according to claim 8, characterized in that, The collaborative optimization suggestions are generated when the production line health index is below the health threshold. They analyze the weak links in the defect triggering frequency, judgment success rate and response success rate, automatically adjust the safety threshold, decision threshold or benchmark parameters, and apply them when the next equipment state switch triggers benchmark modeling after manual confirmation.

10. A system for identifying defects in equipment manufacturing processes using a data-adaptive method as described in any one of claims 1-9, characterized in that, The system includes: The benchmark modeling and anomaly detection module is used to collect and model multi-dimensional raw delay data triggered by equipment state switching, and generate an initial multi-dimensional delay benchmark and its tolerance range that characterizes the health status of the equipment. In subsequent production cycles, the actual delay of the current cycle is calculated in real time and compared with the benchmark. If the deviation exceeds the safety threshold, a first-level anomaly event is triggered. The decision routing module is used to calculate the delay gradient and fluctuation entropy of the current tick after a first-level abnormal event is triggered; based on the comparison results of the delay gradient and fluctuation entropy with preset thresholds, the first-level abnormal event is classified and routed to different processing paths, including: triggering instantaneous suppression response, triggering baseline remodeling, triggering deep diagnosis, and determining it as a false alarm and restoring monitoring. The deep diagnostic module is used to perform a three-layer nested diagnosis on complex hidden dangers routed to the deep diagnostic path, including: filtering out the core abnormal dimension set that deviates significantly; analyzing the causal flow between the dimensions in the core abnormal dimension set based on the propagation entropy, tracing the source abnormal dimension and the abnormal propagation path; matching the characteristics of the core abnormal dimension set, the source abnormal dimension and the propagation path with the historical defect pattern library, and outputting specific defect type instructions. The adaptive response execution module is used to execute a progressive three-order adaptive response according to the defect type instruction: the first order performs an instantaneous suppression operation and verifies it immediately; if it is ineffective, it enters the second order to compensate and adjust the control or process parameters and verify them; if it is still ineffective, it enters the third order to locally correct and verify the dimensional statistical characteristics of the initial multi-dimensional delay baseline; if all three orders fail, manual intervention is triggered. The system collaborative optimization module is used to periodically count the defect triggering frequency, defect judgment success rate, and response success rate, and calculate the production line health index. Based on the comparison results between the production line health index and the health threshold, it generates collaborative optimization suggestions for system parameters and models, and applies them in the next benchmark modeling.

Images