Quartz continuous melting fault prediction and quality risk identification method based on deep learning

Abstract

The application discloses a quartz continuous melting fault prediction and quality risk identification method based on deep learning, and belongs to the field of quartz continuous melting continuous production process monitoring, deep learning fault prediction and quality risk identification. In order to solve the problems of strong coupling of multi-source process data, quality detection lag and few abnormal samples, the application realizes the technical effects of predicting potential faults in advance, identifying the quality risk level and positioning the main influencing factors through lag response association, dynamic process influence graph, graph time sequence prediction model and shape risk calibration.

Description

Technical Field

[0001] This invention relates to the fields of continuous production process monitoring of quartz continuous melting, deep learning fault prediction and quality risk identification, and particularly to a method for quartz continuous melting fault prediction and quality risk identification based on deep learning. Background Technology

[0002] Quartz continuous melting production is characterized by long continuous operation time, slow changes in thermal state, and delayed manifestation of quality defects. During production, furnace temperature, current, voltage, feed rate, drawing speed, furnace pressure, cooling parameters, raw material batches, and manual operation records all collectively influence bubbles, streaks, crystallization, dimensional fluctuations, and equipment malfunctions. Current production sites typically rely on fixed threshold alarms, single-variable trend judgments, or manual experience to identify anomalies. While these methods can alert to parameters that have already deviated significantly, they struggle to describe the coupled effects between multiple process variables.

[0003] Because quality inspection results such as bubbles, streaks, crystallization, and dimensional fluctuations lag behind production parameters, and electrode aging, abnormal feeding, and abnormal cooling may not manifest as quality fluctuations for a considerable period, traditional methods struggle to accurately correlate subsequent quality results with the production window where the anomaly occurred. Furthermore, the limited number of severe fault and defect samples makes directly training a single classification model susceptible to class imbalance, leading to delayed warnings, false alarms, missed alarms, and unclear anomaly causes.

[0004] Therefore, there is a need for a method for predicting quartz continuous melting failures and identifying quality risks that can overcome the shortcomings of the existing technologies. Summary of the Invention

[0005] One objective of this invention is to propose a deep learning-based method for predicting quartz continuous melting faults and identifying quality risks. Addressing the challenges of existing technologies that struggle to correlate multi-source time-series process fluctuations with delayed quality detection results, and the limited number of severe anomaly samples and difficulty in locating the causes of anomalies, this invention proposes a technical solution involving learnable delayed correlation of multi-source process windows, construction of a dynamic process influence diagram constrained by process mechanisms, outputting risk prediction results using a graph time-series prediction model, and generating risk levels and confidence scores through conformal risk calibration. This invention offers the technical advantages of predicting potential faults in advance, reducing false alarms and missed alarms, and outputting a ranking of key influencing factors.

[0006] This invention provides a deep learning-based method for predicting quartz continuous melting faults and identifying quality risks, comprising: S1. Synchronize furnace temperature, current, voltage, feeding amount, drawing speed, furnace pressure, cooling parameters, raw material batches and manual operation records in time, segment working conditions and slice sliding windows to generate a multi-source process window set; S2. Input the delayed quality inspection results and equipment failure records into the learnable lag response module to determine the effective production window corresponding to each quality defect type and equipment failure type, and generate sample association records with lag contribution weights. S3. Using the process variables, raw material batches, operation events in the multi-source process window associated with the sample association record, as well as the quality defect items and equipment failure items in the sample association record as nodes, a dynamic process influence diagram containing sample learning diagrams, calibration diagrams and online inference diagrams is generated based on process mechanism constraints and sample statistical relationships. The online inference diagram does not include the current window confirmation status, true label, detection results and contribution weight. S4. Input the dynamic process influence diagram and multi-source process window into the graph time series prediction model to generate future failure probability, quality defect type probability, risk level candidate value, prediction time domain parameter and explanation intermediate quantity; S5. Input the future failure probability, the quality defect type probability, the risk level candidate value, the prediction time domain parameter, and the explanation intermediate quantity into the conformal risk calibration module, output the quality risk level, risk confidence, early warning lead time, and the sorting of major influencing factors, and output the status pending manual review, the reason for review triggering, and the abnormal output flag under the abnormal branch.

[0007] Optionally, S1 includes: The furnace temperature, current, voltage, feeding amount, drawing speed, furnace pressure and cooling parameters are resampled using a unified time reference, and operating condition labels are generated according to the equipment operation section, batch change section, feeding disturbance section and drawing speed adjustment section. Establish event coding fields for raw material batches and manual operation records; The multi-source process window set is formed by slicing with a preset window length and a preset step size, and the corresponding window is marked as a window to be reviewed when the missing rate of any data field is greater than a preset missing rate threshold.

[0008] Optionally, S2 includes: Candidate hysteresis interval sets are set for each of the following: bubbles, streaks, crystallization, size fluctuations, electrode aging, abnormal feeding, and abnormal cooling. The learnable hysteresis response module calculates the normalized contribution weight of each candidate hysteresis interval based on the detection time, quality defect type, equipment failure type, and the multi-source process window set. According to the target retention ratio, the quantile corresponding to 1 minus the target retention ratio is read from the calibration sample contribution weight set in ascending order as the lag screening threshold. The production window with a contribution weight not less than the lag screening threshold is determined as the effective production window. The detection result, fault record, window identifier, quality defect type, equipment fault type and contribution weight are written into the sample association record. Furthermore, the candidate lag interval set is generated by the time difference distribution between historical quality inspection time, fault record time, and production process window time, as well as the preset quartz continuous melting process experience boundary. The normalized contribution weight is calculated by an attention network using the time difference feature of the candidate lag interval, the embedding of the quality defect type or the equipment fault type, the embedding of the operating condition label, and the detection confidence field. The sum of the contribution weights of the candidate lag intervals corresponding to the same detection result or the same equipment fault record is one.

[0009] Optionally, S3 includes: Initial edge weights are generated based on the thermoelectric coupling relationship between furnace temperature and electrical parameters, the material disturbance relationship between feeding amount and furnace pressure, the forming relationship between drawing speed and dimensional fluctuation, and the thermal history relationship between cooling parameters and crystallization risk. The initial edge weights and the statistical correlation coefficients calculated from the associated records of the samples are input into the dynamic graph module of the process mechanism constraint, and the initial edge weights are dynamically corrected by graph attention weights to obtain the dynamic process influence graph containing node fields, edge type fields, edge weight fields and time version fields.

[0010] Optionally, S4 includes: The graph temporal prediction model includes a graph attention network, a temporal Transformer encoder, and a multi-task output layer; The graph attention network extracts variable coupling features based on the dynamic process influence graph; The temporal Transformer encoder extracts long- and short-term fluctuation features based on the variable coupling characteristics and the multi-source process window set. The multi-task output layer outputs the future failure probability, the quality defect type probability generated by the independent sigmoid channels for each quality defect type, the risk level candidate value, the prediction time domain parameters, and the explanatory intermediate quantity. The explanatory intermediate quantity includes graph attention weights, temporal Transformer attention weights, and the prediction output gradient contribution value. The future failure probability includes the failure type probability generated by the independent sigmoid channels for each device failure type and the total failure probability obtained by fusing the failure type probabilities. The prediction time domain parameters include the current window end time, the prediction duration, and the prediction target time. Furthermore, the training of the graph time series prediction model includes: using the sample association records as supervision samples, constructing a multi-task loss function according to the quality defect type, equipment failure type and quality risk level, wherein the quality defect type task and the equipment failure type task both adopt the label-by-label binary cross-entropy loss; Sample weights are assigned to severe fault samples and severe defect samples, and the sample weights are determined jointly based on the reciprocal of the category frequency and the contribution weight; When multiple quality results or multiple device failure results are associated with the same window, the corresponding task losses in the multi-task loss function are weighted and summed to obtain the trained graph time series prediction model.

[0011] Optionally, S5 includes: The conformal risk calibration module constructs a set of non-consistent scores for faults, a set of non-consistent scores for defects, and a set of non-consistent scores for risk levels based on the predicted scores output by the multi-task output layer and the real quality results and real fault results of recent calibration samples. Among them, the quality defect type task constructs a multi-label non-consistent score according to the probability of each label, and the equipment fault type task constructs a multi-label non-consistent score according to the probability of each fault type. The dynamic grading threshold is determined according to the quantile corresponding to the target coverage rate; During online inference, the defect prediction uncertainty component and the fault prediction uncertainty component, which are independent of the current real label, are calculated based on the current quality defect type probability and the current future failure probability, and the risk level candidate value is corrected using the defect prediction uncertainty component and the fault prediction uncertainty component. The corrected risk level candidate value is compared with the dynamic grading threshold to generate the quality risk level, and the risk confidence is generated according to the proportion of samples in the risk level inconsistency score set that are not less than the current risk level inconsistency score. The current risk level inconsistency score is determined by the inverse of the current corrected candidate probability corresponding to the final output quality risk level. Furthermore, when the number of recent calibration samples is less than the preset minimum number of calibration samples, or when the risk confidence level is less than the preset confidence threshold, the conformal risk calibration module will mark the corresponding prediction result as pending manual review, and retain the variable influence path output by the process mechanism constraint dynamic graph module and the ranking of the main influencing factors determined by the explanatory intermediate quantity via S5; Furthermore, the ranking of the main influencing factors is generated in the following way: the graph attention weights, time series Transformer attention weights and prediction output gradient contribution values ​​in the graph time series prediction model are normalized to generate node influence components, node-level time window influence components and output sensitivity components obtained by mapping the time window attention according to the node contribution or node gradient contribution within the corresponding window. When the sum of node contributions within a time window is zero, the node allocation coefficient of the corresponding time window is set to uniform distribution of nodes within the same window and written into the interpretation degradation flag. When the normalized sum of node influence components, node-level time window influence components, or output sensitive components is zero, the corresponding normalized component is set to a zero vector and written into the interpretation component missing flag. The node influence component, the node-level time window influence component, and the output sensitivity component are weighted and fused according to the preset interpretation weights to obtain the final influence factor score, and the main influence factors are output in descending order of the final influence factor score.

[0012] The beneficial effects of this invention are: 1. By using a learnable lag response module to inversely correlate delayed quality inspection results and equipment failure records to the effective production window, it is possible to avoid judging risk based solely on parameters near the inspection time, thus creating a trainable sample correlation between quality defects and corresponding process fluctuations.

[0013] 2. By incorporating thermoelectric coupling, material disturbance, forming relationship, and thermal history relationship into the dynamic graph edge weights through the process mechanism constraint dynamic graph module, and combining it with dynamic correction of the graph attention network, it is possible to extract the variable influence path under strong coupling of multi-source variables, thereby improving the interpretability of anomaly cause localization.

[0014] 3. By jointly outputting the future failure probability, quality defect type probability, quality risk level, risk confidence, early warning lead time, and ranking of major influencing factors through a multi-task graph time series prediction model and a conformal risk calibration module, the reliability of risk classification results can be improved under the condition of a small number of severe abnormality samples. Attached Figure Description

[0015] The accompanying drawings are provided to further illustrate the invention and form part of the specification. They are used in conjunction with embodiments of the invention to explain the invention and do not constitute a limitation thereof. In the drawings: Figure 1 This is a schematic diagram of a deep learning-based method for predicting quartz continuous melting faults and identifying quality risks.

[0016] Figure 2 This is a flowchart of step S5, the conformal risk calibration module, of the present invention. Detailed Implementation

[0017] The present invention will now be described in further detail with reference to the accompanying drawings. These drawings are simplified schematic diagrams, illustrating only the basic structure of the invention, and therefore only show the components relevant to the invention.

[0018] refer to Figure 1A deep learning-based method for predicting quartz continuous melting failures and identifying quality risks includes: S1. Synchronize furnace temperature, current, voltage, feeding amount, drawing speed, furnace pressure, cooling parameters, raw material batches and manual operation records in time, segment working conditions and slice sliding windows to generate a multi-source process window set; S2. Input the delayed quality inspection results and equipment failure records into the learnable lag response module to determine the effective production window corresponding to each quality defect type and equipment failure type, and generate sample association records with lag contribution weights. S3. Using the process variables, raw material batches, operation events in the multi-source process window associated with the sample association record, as well as the quality defect items and equipment failure items in the sample association record as nodes, a dynamic process influence diagram containing sample learning diagrams, calibration diagrams and online inference diagrams is generated based on process mechanism constraints and sample statistical relationships. The online inference diagram does not include the current window confirmation status, true label, detection results and contribution weight. S4. Input the dynamic process influence diagram and multi-source process window into the graph time series prediction model to generate future failure probability, quality defect type probability, risk level candidate value, prediction time domain parameter and explanation intermediate quantity; S5. Input the future failure probability, the quality defect type probability, the risk level candidate value, the prediction time domain parameter, and the explanation intermediate quantity into the conformal risk calibration module, output the quality risk level, risk confidence, early warning lead time, and the sorting of major influencing factors, and output the status pending manual review, the reason for review triggering, and the abnormal output flag under the abnormal branch.

[0019] In this specific embodiment, S1 includes: The data processing server of the quartz continuous melting production line reads production data from the furnace body temperature control system, power supply cabinet, feeding controller, drawing servo controller, furnace pressure acquisition unit, cooling water control unit, raw material batch management terminal and operation record terminal. The original data record includes the acquisition time, acquisition source identifier, furnace temperature value, heating current value, heating voltage value, feeding set amount, actual feeding amount, drawing speed, furnace pressure value, cooling water flow rate, cooling water inlet temperature, cooling water outlet temperature, raw material batch number, operation event type and operator confirmation flag. The sampling period of furnace temperature, electrical parameters, furnace pressure and cooling parameters is set to 1 second in this specific embodiment. Raw material batch and manual operation record are written to the event log queue according to the event occurrence time. The data processing server uses the time base provided by the production line clock server as the unified time base, and performs resampling on continuously collected data. The timestamps after resampling are calculated according to the formula. Confirmed, among which This is the original data collection time. This is the start time for modeling this batch. To ensure a consistent sampling interval, which is set to 10 seconds in this specific embodiment, For resampling timestamps, falling into the same Multiple records are selected based on the principle of closest time, and no records fall into the same category. The fields are padded with linear interpolation while retaining the interpolation flag field; For each resampling time point, the data processing server generates a process status record, which includes a window primary key, timestamp, continuous process field, discrete event field, batch field, missing value flag field, and outlier flag field. The server reads the valid range table calibrated before deployment for furnace temperature, current, voltage, feed rate, drawing speed, furnace pressure, and cooling parameters. If a continuous field exceeds the corresponding upper or lower limit in the valid range table, the field is replaced with the nearest boundary value within the upper or lower limit, and the boundary clipping result is recorded in the outlier flag field. The data processing server generates operating condition tags according to the equipment operation segment, batch change segment, feeding disturbance segment, and pulling speed adjustment segment. The equipment operation segment is determined by the pulling speed being non-zero and the furnace pressure continuously meeting the effective range. The batch change segment is determined by the 30-minute time interval after the raw material batch number changes. The feeding disturbance segment is determined by the absolute value of the deviation of the actual feeding amount from the feeding set amount being not less than the feeding disturbance threshold. The pulling speed adjustment segment is determined by the absolute value of the change in pulling speed within 5 consecutive minutes being not less than the pulling speed adjustment threshold. If multiple operating condition tags are triggered simultaneously, the main operating condition tag is written in the order of batch change segment, feeding disturbance segment, pulling speed adjustment segment, and equipment operation segment, and the remaining tags are written to the auxiliary operating condition field. For raw material batches and manual operation records, the data processing server establishes an event coding field, mapping the raw material batch number to a batch embedding index, and mapping records of feed rate adjustment, pulling speed adjustment, cooling parameter adjustment, furnace pressure intervention, and manual confirmation to event type codes. At each resampling time point, the server counts the occurrence frequency of each event type within the preceding 30-minute event window, the interval between the most recent event and the current time, and the operation confirmation flag, thus obtaining an event feature vector. ; The data processing server uses a preset window length. and preset step size In this specific embodiment, a sliding window slice is used to record the process status. Set to 120 minutes. Set to 10 minutes, the first The time range of each window is Within the window, continuous fields are constructed with statistical features based on mean, standard deviation, maximum value, minimum value, slope, and last value. Event fields are constructed with event features based on occurrence frequency, most recent interval, and event type. Batch fields are constructed with batch features based on the main batch number and batch change flag within the window. The data processing server calculates the missing data field rate for each window. ,in Count the fields marked with missing information within the window. The total number of fields to be collected within the window is specified. The preset missing rate threshold in this specific implementation is determined by the 95th percentile of the missing rate of 30 consecutive normal production batches. If the value exceeds the threshold, the window to be reviewed is written into the window status field and the window is retained but not included in the automatic closed-loop calibration. For windows that pass the missing rate check, the data processing server determines the minimum value of the field from the training samples. Maximum value Dimensionless protection quantity Normalization is performed on continuous features, and the normalized values ​​are calculated according to the formula. Received, among which For the first The first window One original feature, , and and Having the same unit To normalize the denominator to the same dimensions, the normalization process involves dividing the difference in the same unit by the reference value within the same unit range. This field is determined by the sensor resolution and prevents the denominator from being zero. This indicates that the value is truncated to the range of 0 to 1, and then normalized. It is a dimensionless numerical value, and the continuous features, event features, batch features, operating condition labels and window status fields are combined to form a multi-source process window record; The data processing server writes all multi-source process window records into a multi-source process window set. The set fields include window identifier, start and end time, main operating condition label, auxiliary operating condition label, normalized process features, event feature vector, batch index, missing rate, pending review flag, and data version number. The version number of this set is then passed to the learnable hysteresis response module of S2 as input for subsequent sample association.

[0020] In this specific embodiment, S2 includes: The learnable hysteresis response module reads the multi-source process window set written by S1, and reads the quality inspection results and equipment fault records that appear late from the quality inspection database and equipment maintenance database. The quality inspection result record includes inspection time, product segment number, bubble mark, stripe mark, crystallization mark, size fluctuation mark, inspection confidence field and inspection equipment number. The equipment fault record includes fault record time, electrode aging mark, feeding abnormality mark, cooling abnormality mark, maintenance confirmation field and fault confirmation time. The learnable hysteresis response module sets candidate hysteresis interval sets for bubbles, streaks, crystallization, size fluctuations, electrode aging, feeding anomalies, and cooling anomalies. In this specific embodiment, bubbles correspond to 2h to 36h, streaks correspond to 1h to 24h, crystallization corresponds to 4h to 72h, size fluctuations correspond to 0.5h to 12h, electrode aging corresponds to 12h to 120h, feeding anomalies correspond to 0.5h to 18h, and cooling anomalies correspond to 1h to 36h. Each interval is discretized into a candidate hysteresis interval index with a 30min particle size. For a quality inspection result or equipment fault record, a learning hysteresis response module can be used to measure the inspection time or fault record time. Based on this, calculate the end time of the candidate production window. and Time difference Only retain The window that falls into the corresponding candidate lag interval set is used as the candidate association window. If the candidate association window is empty, the status of pending manual review is written and the automatic sample association of the detection result is stopped. The learnable hysteresis response module inputs candidate hysteresis interval indices, time difference features, defect type embeddings, fault type embeddings, S1-generated condition label embeddings, and detection confidence fields into the attention network. The attention network scores according to the formula. Received, among which For the first The window feature vector corresponding to each candidate lag interval This represents the type and confidence feature vector of the current detection result or fault record. , , and These are the parameters learned from the training samples; The learnable lag response module converts attention scores into normalized contribution weights, which are calculated using the formula... Received, among which For the first The contribution weights of each candidate lag interval are calculated, and the sum of the contribution weights of all candidate lag intervals corresponding to the same detection result or fault record is one. If the detection confidence field is lower than the detection confidence threshold, then all candidate lag intervals corresponding to the detection result are considered as one. Write low-confidence samples to a cache instead of using them for automatic model training; The learnable hysteresis response module determines the hysteresis screening threshold based on the quantiles of the calibration samples. Specifically, it reads the contribution weight set of confirmed quality inspection results and fault records from the most recent 90 days and sets the target retention ratio as follows. Sort the set of contribution weights from low to high, and take... The corresponding quantile is used as the lag screening threshold. ,when Not less than When this happens, the corresponding production window is determined as the effective production window, ensuring that the candidate window with the highest contribution weight is retained. When all... All less than At that time, retain the largest The corresponding window will then be set to "Pending Manual Review" in the associated status field. For each valid production window, the learnable hysteresis response module generates a sample association record. The sample association record includes an association identifier, a window identifier, a detection result identifier, a fault record identifier, a defect type, a fault type, a candidate hysteresis interval index, a time difference, a contribution weight, a detection confidence field, a manual review status, and a source data version. The association identifier is generated jointly by the window identifier, the result identifier, and the candidate hysteresis interval index. When multiple quality inspection results or multiple equipment fault records are associated with the same production window, the learning hysteresis response module retains all associated records and generates multi-label fields. Bubbles, streaks, crystallization, dimensional fluctuations, electrode aging, abnormal feeding, and abnormal cooling in the multi-label fields are represented by 0 or 1, respectively. For production windows that do not show records of quality defects such as bubbles, streaks, crystallization, or dimensional fluctuations, nor records of equipment faults such as electrode aging, abnormal feeding, or abnormal cooling, and which do not generate records of downtime, downgraded production, product scrapping, rework, or manual handling within the preset quality inspection observation period and equipment fault observation period, the learning hysteresis response module generates normal sample associated records. The associated identifier is generated by the window identifier and observation period number. The inspection result identifier and fault record identifier are set to null values. The quality defect type field and equipment fault type field are both set to 0. The quality risk level is set to normal risk, and the contribution weight is set to 1. When one observation period covers multiple production windows... When the window duration is normalized, the contribution weight is determined according to the coverage ratio of each window. The normal sample association record is allowed to enter the S4 training sample and the S5 recent calibration sample. The quality risk level label is generated according to the risk level mapping table based on the confirmed quality defect degree, equipment failure confirmation status and production handling results. The mapping table writes samples without defects and without confirmed failures as normal risk, slight bubbles, slight streaks or slight size fluctuations that have not triggered handling as concern risk, samples with crystallization trend, continuous size exceeding limit, confirmed feeding abnormality or cooling abnormality but not yet stopped as warning risk, and samples that lead to shutdown, downgraded production, product scrapping, rework or repair confirmation of electrode aging, feeding abnormality, cooling abnormality as serious risk. When there are multiple confirmed results in the same window, the highest risk level is selected and written according to the priority of serious risk, warning risk, concern risk and normal risk, while the contribution weight corresponding to each label is reserved for the weighted use of S4 training loss. For low-confidence detections, fault records that have not passed maintenance confirmation, or samples marked as pending review windows by S1, the learnable hysteresis response module can write the associated status as pending manual review and isolate the record from the automatic training sample table to the review sample table. The review sample table still retains the window features, hysteresis interval index, and contribution weight, which are used to write back the calibration sample after manual confirmation. The learnable lag response module also maintains a lag interval parameter table. The parameter table fields include defect type, lower limit of candidate interval, upper limit of candidate interval, discrete granularity, screening threshold, target retention ratio, sample time range, low confidence threshold, source of engineering experience boundary, and parameter version number. The parameter table is updated monthly based on historical confirmed results. When updating, the time difference distribution between the historical confirmed quality inspection time or fault record time and the candidate production window end time is read, and the 5th to 95th percentiles are taken as the data boundary. The intersection with the preset quartz continuous melting process experience boundary is taken to generate the candidate lag interval. If the number of confirmed samples for any defect type or fault type is less than 30, the previous parameter version or preset engineering experience boundary is used and a freeze flag is written in the parameter version number. Before outputting the associated records of the samples, the learning lag response module can perform duplicate association deduplication. For the same detection result and the same window within the same candidate lag interval, only the record with the largest contribution weight is retained. If the contribution weights of two records are the same, the record with the higher detection confidence field is retained. If the detection confidence fields are still the same, the record with the time difference closer to the center of the corresponding candidate lag interval is retained. The deduplication result is written to the association deduplication status field. When electrode aging, feeding abnormality, and cooling abnormality occur simultaneously within the same effective production window, the learning hysteresis response module does not merge fault types. Instead, it retains an independent contribution weight for each fault type in the sample association record and generates a window-level fault combination code. This fault combination code is mapped to a multi-label fault label vector according to the three fault types: electrode aging, feeding abnormality, and cooling abnormality. Each component takes either 0 or 1. The multi-label fault label vector is used by S3 to establish equipment fault item nodes and by S4 to construct the equipment fault type task loss function. The learnable hysteresis response module performs source consistency checks on quality inspection results and equipment fault records. If conflicting defect types occur within the same inspection batch for the same product segment number, the inspection equipment number and inspection confidence field are read, the result with the highest inspection confidence field is retained, and the conflict result is written to the conflict sample cache. If the equipment fault record lacks a maintenance confirmation field, the fault type is written to the unconfirmed fault field and the record is prohibited from being used as a positive sample of a serious fault. The learnable hysteresis response module maintains the sample lifecycle status in the sample association table. The status includes automatic association, pending manual review, review passed, review rejected, and version freeze. After manual review is passed, the original contribution weight is retained and manual confirmation time is added. After manual review is rejected, the contribution weight is reset to zero and written to the rejection reason field. The version freeze status is used to ensure that the source of the supervision samples read during S4 training is consistent with the source of the calibration samples read during S5 calibration. To avoid frequent fluctuations in the hysteresis interval due to short-term noise, the learning hysteresis response module can calculate the interval stability coefficient for each defect type. The interval stability coefficient is obtained by dividing the difference in the screening threshold between two consecutive parameter versions by the baseline of the screening threshold range of the previous version. If the interval stability coefficient exceeds the version update threshold, only candidate parameter versions are generated and enter the manual confirmation queue. Before confirmation, the previous stable parameter version is still used. The learnable hysteresis response module generates an association quality summary field before output. The association quality summary field includes the number of automatic associations, the number of samples awaiting manual review, the number of low-confidence detections, the number of unconfirmed faults, the number of parameter version freezes, and the number of conflicting samples. If the proportion of samples awaiting manual review exceeds the review proportion threshold, the association table of this batch of samples will be marked as a low automated confidence state. The learnable hysteresis response module writes the sample association records into the sample association table and outputs the effective production window set, contribution weight vector, label field, review status field and association quality summary field. The sample association table serves as the basis for generating nodes and edges in the S3 dynamic process influence diagram and as the source of supervised samples for training the S4 graph time series prediction model.

[0021] If the candidate association window is empty, the effective production window set is empty, or the sample association record generation fails, the learning hysteresis response module can write the corresponding window or result into the sample association failure status, generate a review trigger reason, and prohibit the record from entering the automatic training, automatic calibration, or automatic alarm closed-loop confirmation; only after manual review is passed can the sample association table and the recently calibrated sample table be written back according to the review result.

[0022] In this specific embodiment, S3 includes: The process mechanism constraint dynamic graph module reads the sample association records and effective production window set output by S2, and establishes nodes for process variables, raw material batches, operation events, quality defect items and equipment failure items in the multi-source process windows associated with the sample association records. The process variable node includes variable name, variable type, unit, window statistical characteristics and normalization fields. The raw material batch node includes batch index, supply batch number and batch change flag. The operation event node includes event type, event time difference and event count fields. The quality defect item node and equipment failure item node include label type, confirmation status and contribution weight fields in the training graph and calibration graph. To avoid introducing posterior labels for the current window during the online inference phase, the process mechanism constraint dynamic graph module generates training graphs, calibration graphs, and online inference graphs according to their intended use. The training graph reads the confirmation status and contribution weights from the confirmed sample association records. The calibration graph only reads historical records of samples that have completed quality testing and entered the recent calibration sample table. The node types and edge structures of the online inference graph are determined by the historical sample association records and process prior edges. Only the quality defect type, equipment failure type, process prior edges, and edge weight prior values ​​calculated from the current window process variables, raw material batches, operation events, and operating condition labels are retained. The confirmation status, true label, test completion result, or contribution weight fields of the current window are not written. If the online graph construction request contains the above posterior fields, the field is set to empty and a label isolation flag is written in the graph version summary. The process mechanism constraint dynamic graph module reads the process prior edge weight table to generate initial edge weights. The fields of the process prior edge weight table include source node type, target node type, edge type, initial edge weight, applicable operating conditions, unit description, calibration source, and version number. A record is as follows: source node type furnace temperature, target node type electrical parameters, edge type thermoelectric coupling, initial edge weight 0.75, applicable operating conditions equipment operating segment, calibration source is the operating records of 30 consecutive stable batches, and version number is the current production line configuration version. For the thermoelectric coupling relationship between furnace temperature and electrical parameters, the material disturbance relationship between feeding amount and furnace pressure, the forming relationship between drawing speed and dimensional fluctuation, and the thermal history relationship between cooling parameters and crystallization risk, the process mechanism constraint dynamic graph module writes the prior edge weights into the candidate edge set and marks the candidate edges that do not conform to the current working condition label as mismatched edges. The mismatched edges are not deleted, but their edge weights are entered into the lower limit constraint field. The process mechanism constraint dynamic graph module calculates statistical correlation coefficients based on sample correlation records and calculates normalized mutual information for continuous process variables and quality defect items. Calculate the absolute value of the Pearson correlation between continuous process variables. The statistical correlation components are calculated according to the formula. Received, among which and The statistical weights are defined as follows, and the sum of the two is equal to one. The statistical weights are determined by the principle of minimizing the verification error of the calibration samples over the past 90 days. For each candidate edge, the process mechanism constraint dynamic graph module will initialize the edge weights. Statistical correlation components and sample contribution weight mean The fusion is based on the prior values ​​of the edge weights, and the fusion rule is as follows: ,in , and The edge weight fusion parameters are obtained from the validation samples. The value ranges from 0 to 1, and the larger the value, the stronger the influence of the source node on the target node. The dynamic graph module constrained by process mechanism dynamically corrects the prior values ​​of edge weights and node features through a graph attention mechanism. and Input attention scoring function and according to the formula The corrected edge weights are obtained, where To point to the target node The set of source nodes, and The graph attention parameters obtained during training; When the candidate edge's adjusted edge weight When the edge is less than the edge retention threshold, the process mechanism constraint dynamic graph module writes the edge into the low-impact edge cache and retains the reason field. When there are multiple edge types between the same source node and target node, the module retains the first two edge types in descending order of the modified edge weight and marks the remaining edge types as candidate explanation edges to avoid unexplainable duplicate impact paths in the dynamic graph. The dynamic process mechanism constraint graph module writes a time version field to the dynamic graph. The time version is determined by the window end time, the operating condition label and the model configuration version. Each time version of the dynamic process influence graph includes a node table, an edge table, an edge weight table and a node feature table. The node table records the node identifier, node type and source field. The edge table records the source node, target node and edge type. The edge weight table records the initial edge weight, statistical related components, average contribution weight and corrected edge weight. In order to preserve the variable influence path for S5, the process mechanism constraint dynamic graph module calculates the path influence value from the process variable node to the quality defect item node or equipment failure item node in each time version. The path influence value is obtained by multiplying the corrected edge weights on the path. If the path length exceeds 3 edges, the expansion stops. If the influence values ​​of two paths are the same, the path containing the process mechanism prior edge is selected as the variable influence path. The process mechanism constraint dynamic graph module performs version consistency checks on the dynamic graph. If there are inconsistencies in the node table version, edge table version, and sample association table version within the same time version, the time version is marked as an untrainable version and the original edge weights are retained. Untrainable versions can still enter the manual review and interpretation process of S5, but will not enter the automatic training samples of S4. The process mechanism constraint dynamic graph module also writes edge weight source descriptions into the edge weight table. When the modified edge weight is mainly contributed by the initial edge weight, the source description is recorded as process prior dominant. When the modified edge weight is mainly contributed by the statistically relevant components, the source description is recorded as sample correlation dominant. When the modified edge weight is mainly contributed by the mean of the contribution weights, the source description is recorded as lag association dominant. This source description is passed to S5 along with the variable influence path. When a quality defect node has no incoming edges in the current time version, the process mechanism constraint dynamic graph module reads the incoming edge structure from the previous valid time version of the same working condition label as the back-off graph structure, and multiplies the corrected edge weights of all back-off edges by the back-off attenuation coefficient. The back-off attenuation coefficient is obtained by interpreting the path stability calibration on the validation set and written into the graph version index. The process mechanism constraint dynamic graph module performs redundancy compression on the path of node influence. If the source node, target node and intermediate node of two paths are of the same type but different time versions, the current time version path is retained and the historical path is written into the path history field. If the two paths are only different in edge type, the path with the higher ranking of the modified edge weight product is retained and the compressed edge type is written into the candidate explanation field. Before the dynamic graph is output, the process mechanism constraint dynamic graph module calculates the graph integrity score. The graph integrity score is obtained by weighting the effective node ratio, effective edge ratio and sample association coverage ratio. All three ratios are dimensionless values ​​between 0 and 1. If the graph integrity score is lower than the graph output threshold, the graph version can only be used for manual review and interpretation and will not be included in the automatic training and automatic risk level output links. The process mechanism constraint dynamic graph module outputs a dynamic process influence graph, a variable influence path table, a graph version index, and a graph version summary. The dynamic process influence graph serves as the input structure for the S4 graph attention network, and the variable influence path table serves as the output source for S5 to retain the basis for locating the cause of anomalies even in low confidence or manual review states.

[0023] In this specific embodiment, S4 includes: The graph time series prediction model service reads the dynamic process influence diagram output by S3, the multi-source process window set output by S1, and the sample association record output by S2. The model version field consists of the training date, sample association table version, dynamic graph version, and calibration parameter version. The inference input includes continuous process feature matrix, event feature vector, batch embedding, operating condition label embedding, dynamic graph node table, dynamic graph edge table, and edge weight table. The graph attention network in the graph time-series prediction model uses the node feature matrix for each time version. and the corrected edge weight matrix Using the input, calculate the variable coupling features. ,in This represents a graph attention network with edge-weighted input. During training, confirmed quality defect nodes and equipment failure nodes from the training graph are read for supervised learning. During calibration, only historical labels from the calibration graph that have entered the recent calibration sample table are read. During online inference, the online inference graph is read, and this graph only retains label type nodes, process prior edges, and edge weight prior values ​​calculated from the current window's process variables. It does not include the current window confirmation status, true labels, detection completion results, and contribution weight fields. The output is... Each row corresponds to a coupled representation of a node. The coupled representation of a node retains the node type, window identifier, and graph version index for subsequent backtracking of influencing factors. The graph-based time series prediction model concatenates the node-coupled representation of each sliding window with the window statistical features into a time series input vector. And form a sequence according to the window end time. In this specific embodiment, the sequence length Set to 12 windows, corresponding to a 120-minute historical span. If there are fewer than 12 historical windows, fill the average vector under the same working condition label and write it into the fill flag field. Temporal Transformer encoder for sequences Output hidden state after adding time location encoding and working condition location encoding The time position encoding is used to distinguish between short-term disturbances and long-term drifts, and the working condition position encoding is used to distinguish between batch change section, feeding disturbance section, speed adjustment section and equipment operation section. In this specific embodiment, the encoder adopts a 4-layer self-attention structure and 8 attention heads, the hidden dimension is set to 128, and the hidden state of the last window is written into the prediction state cache. Multi-task output layer based on hidden state The hidden state is calculated by separately calculating the future failure probability, the probability of quality defect type, the candidate value of risk level, the prediction time-domain parameter, and the explanatory intermediate quantity. The dimensionless process features normalized by S1 and the edge weights normalized by S3 are encoded by the model and are dimensionless vectors. The future failure probability includes the probability of each equipment failure type and the total failure probability. The probability of each equipment failure type is calculated as follows: get, This indicates one of the equipment failure types: electrode aging, abnormal feeding, or abnormal cooling. The total failure probability is calculated as follows: The resulting data shows that the probability of quality defect types is calculated using a single label channel. get, This indicates a quality defect type, such as bubbles, streaks, crystallization, or dimensional fluctuations. Risk level candidate values ​​are listed as follows: The predicted time-domain parameters include the current window end time. Predicting target time and predicted duration , Read from the prediction time-domain configuration in the model version registration table and match it with the window end time. The intermediate quantities are jointly written into the prediction result record. The intermediate quantities are explained by the node coupling representation, graph attention weights, temporal Transformer attention weights and output gradients, which are generated and written into the explanation cache for S5 to generate the final influencing factor score. The training samples for the graph time series prediction model come from the sample association records generated by S2. The training labels include future fault labels, quality defect type labels, and quality risk level labels. The training samples are divided into training set, validation set, and out-of-time test set according to production batch. When dividing, windows of the same raw material batch only enter one set to avoid leakage of windows of the same batch between training and validation. Multi-task loss function according to formula Build, in Label-by-label binary cross-entropy loss for equipment fault types and according to calculate, The first fault obtained by S2 window-level fault combination coding mapping Each device fault type label, For each label of quality defect type, use binary cross-entropy loss and... calculate, For the first A 0 or 1 label for each quality defect type, For risk level cross-entropy loss, , and The task weights are determined by balancing the average recall of the three tasks on the validation set, and the sum of the task weights is one. For severe fault samples and severe defect samples, the graph time series prediction model training service sets sample weights, first by adjusting the frequency ratio of each category. Convert to dimensionless category scarce components Then follow the formula The sample weights are obtained, where The median of the frequency proportions of all categories. The dimensionless contribution weights are assigned to the output of S2. To prevent dimensionless frequency protection during division by zero, In this specific implementation, the weight is set to 5. The larger the sample weight, the greater the contribution of the sample to the loss function. In this specific embodiment, severe fault samples are those whose repair confirmation field is confirmed and whose handling field records shutdown, downgraded production, prohibition of automatic handling, or two or more consecutive batches with increased quality risk, such as electrode aging, abnormal feeding, or abnormal cooling. Severe defect samples are those whose quality inspection level field reaches severe, or whose quality handling field records product scrapping, rework, downgraded use, or continuous dimensional exceedances, such as bubbles, streaks, crystallization, or dimensional fluctuations. The judgment fields of severe fault samples and severe defect samples, together with the manual review status, detection confidence field, and source data version in the S2 sample association record, are written into the training sample table. Samples that fail confirmation are only entered into the review sample table and are not considered severe positive samples.

[0024] When multiple quality results are associated with the same window, the training service performs a weighted sum of the corresponding task losses, and the window loss is calculated according to the formula. Received, among which An index for quality results or fault results associated with the same window. For the corresponding contribution weight, For the corresponding sample weights, The task loss for this association result is determined by the association status of "awaiting manual review". If the association result is not included in the automatic training loss, it will not be included in the automatic training loss. During training, the graph time series prediction model service uses the risk level recall rate, severe failure recall rate, and false positive rate of the validation set as stopping conditions. If there is no improvement after 5 consecutive training rounds, the current model version is frozen, and the model parameters, graph attention parameters, Transformer attention parameters, task weights, class frequency table, and training data version are written into the model version registration table. During online inference, if the input window is marked as a window to be reviewed by S1, or if the number of valid edges related to the window in the dynamic graph of S3 is less than the minimum edge number threshold, the graph time series prediction model will still output the prediction score but will set the automatic handling flag to prohibited, and only allow S5 to generate a state to be manually reviewed. If the input meets the integrity condition, the inference result will enter the conformal risk calibration module of S5. Before each online inference, the graph time series prediction model service reads the model version registration table. If the graph version index of the current window does not belong to the graph version range registered during the training of the model version, the graph version difference flag is written into the prediction result record, and the previous stable model version is called for parallel inference. If the highest probability category of the candidate value of the quality risk level is different between the current model and the previous stable model, the automatic handling flag is set to prohibited and the reason for review is output. To enable S5 to generate an interpretable final influencing factor score, the graph time series prediction model service performs a gradient backpropagation once for each risk output channel, calculates the gradient contribution value of the prediction output relative to the node coupling representation, and saves the gradient contribution value, graph attention weight, and time series Transformer attention weight to the interpretation cache. The interpretation cache is indexed according to window identifier, model version, and risk output channel for S5 to generate a sort of major influencing factors. For windows where the probability of quality defect type and the probability of future failure are both lower than their respective alarm thresholds, but the candidate value of risk level shows a warning risk, the graph time series prediction model service does not directly discard the results, but writes the window into the divergence sample cache. The divergence sample cache fields include three types of prediction output, trigger threshold, dynamic process influence graph version and input window summary, which are used for subsequent manual confirmation and model retraining. The graph time series prediction model service performs time leakage checks on training samples. If the detection time of a sample's associated record is earlier than the end time of its effective production window, the record is written to the time anomaly cache and excluded from the training set. If a raw material batch spans the training set and the validation set, the entire batch is moved to the set with a later time to keep the model evaluation results consistent with the actual continuous production deployment conditions. The model version registration table also stores a list of input fields and field order hash values. If the field order hash value of the real-time window is inconsistent with the model version registration table during online inference, the graph time series prediction model service stops the automatic inference of that window and generates an abnormal prediction result record. The abnormal prediction result record is written with a field version inconsistency flag, window identifier, current window end time and model version, and the automatic handling flag is set to prohibited. The review reason is written as field version inconsistency. This allows S5 to set the quality risk level to review pending, the risk confidence level to 0, the early warning advance to invalid, and the main influencing factor ranking to uncalculable or only retain the previous stable model reference explanation, to prevent furnace temperature, current, voltage or cooling fields from being misplaced into the model and triggering an automatic alarm closed loop. The graph time series prediction model service outputs prediction result records, with fields including window identifier, current window end time, graph version index, model version, total failure probability, per-device failure type probability, quality defect type probability, risk level candidate value, prediction duration, prediction target time, explanation intermediate quantity, graph attention weight, time series Transformer attention weight, prediction output gradient contribution value, and automatic handling flag. These prediction result records serve as inputs for S5 to generate quality risk level, risk confidence, early warning lead time, and ranking of major influencing factors.

[0025] In this specific embodiment, S5 includes: The conformal risk calibration module reads the prediction result record output by S4, the variable influence path table output by S3, and the recent calibration sample table. The recent calibration sample table consists of production windows that have completed the quality inspection observation cycle and the equipment failure observation cycle and have confirmed the failure status or no failure status. It includes the observed completed windows corresponding to the normal sample association records. The fields include window identifier, prediction score, actual quality result, actual failure result, risk level manual confirmation value, inspection completion time and calibration version number. The conformal risk calibration module constructs a set of mutually isolated, inconsistent calibration sample scores based on model predicted scores, actual quality results, and actual failure results, for the probability of quality defect types. The calibration samples used defective multi-label inconsistency scores, according to... Calculation, where For the first The calibration sample at the ... A true label of 0 or 1 on each quality defect type, This represents the label-by-label probability given by the model for that label. A smaller value indicates that the multi-label defect prediction is closer to the confirmed quality result; for equipment failure type probability The calibration samples used fault multi-label inconsistency scores. ,in For the first The calibration sample at the ... A 0 or 1 true fault label for each device fault type; for risk level candidate values. The calibration samples use risk level inconsistency scores. ,in This represents the actual risk level. To give the model the probability of the true risk level, the inconsistency scores of the three calibration samples are written into the defect calibration field, the fault calibration field, and the risk level calibration field, respectively. The conformal risk calibration module determines a dynamic grading threshold based on the target coverage rate. In this specific implementation, the target coverage rate is... Set to 0.9, calibration sample number is Time-division index by Determine the dynamic grading threshold. For the set of inconsistent risk level scores The first one after ascending order A value, if Greater than Then, the highest risk level inconsistency score is taken, and the calibration status is marked as insufficient sample warning, resulting in a defect inconsistency score set. and the set of inconsistent scores for faults Updated only by confirmed historical calibration samples, used to calibrate the weights of the uncertainty components for defect prediction and the uncertainty components for fault prediction; In this specific implementation, the corrected candidate values ​​for each risk level are first re-normalized to the same probability space through the same calibration mapping. The dynamic grading threshold Tc is applicable to the inconsistent scores of the current risk level corresponding to normal risk, risk of concern, warning risk and severe risk, and generates a unique output level according to the comparison order of severe risk first, warning risk second, risk of concern third, and normal risk last. The threshold version table records the unified threshold and its applicable working condition labels.

[0026] For the current online prediction results, the conformal risk calibration module does not read the current window's actual quality label, actual fault label, or manually confirmed risk level. Instead, it generates online calculable prediction uncertainty components based on the current prediction probability. The defect prediction uncertainty components are calculated according to... Calculation, the uncertainty component of fault prediction according to Calculation, where For the current window The probability of each quality defect type For the current window Probability of each equipment failure type and The value ranges from 0 to 1. A larger value indicates that the current defect or fault prediction is closer to the judgment boundary of the corresponding anomaly label. θd and θf in the formula are the lower limits for triggering defect and fault attention, respectively, determined by the 95th percentile of the probability of the anomaly label corresponding to the normal window in the recent calibration samples. I(·) is an indicator function, taking 1 if the condition is met and 0 otherwise. When the probabilities of all defect types and all fault types are both lower than the corresponding lower limits for triggering attention, the prediction uncertainty component is set to 0. The low anomaly probability window does not automatically increase the candidate channel to a level no lower than the warning level due to uncertainty correction. The correction rule is as follows: ,in and For the recent calibration samples , The dimensionless corrected weights obtained from the correlation calibration with the risk level error, and the sum of the two is not greater than one. To calibrate the upper limit of correction, the corrected risk level candidate values ​​will be arranged from low to high risk level. When the aforementioned prediction uncertainty component is greater than 0, it is added to a candidate channel at or above the warning level and renormalized; when all the aforementioned prediction uncertainty components are 0, the calibration correction is set to 0 and the candidate channel at or above the warning level remains unchanged; the conformal risk calibration module compares the corrected risk level candidate value with the dynamic grading threshold to generate a quality risk level, and the current risk level inconsistency score is calculated according to... Calculation, where For the first The revised risk level candidate probability; when the severity level corresponds to Not greater than Outputting severe risks in a timely manner, when the warning level corresponds to Not greater than Timely output of early warning risks, when the attention level corresponds to Not greater than In cases of emergency, output "risk of concern"; in other cases, output "normal risk". The comparison order is: severe risk, warning risk, risk of concern, and normal risk. Risk confidence scores are generated solely based on a set of inconsistent risk level scores, according to the formula. Generate, where For the first Risk level inconsistency scores of recently calibrated samples The current prediction result is a risk level inconsistency score calculated based on the revised risk level candidate values. An indicator function that takes the value 1 if the condition is met and takes the value 0 if the condition is not met. The value ranges from 0 to 1. A larger value indicates that the inconsistency score of the current risk level is smaller in recent calibration samples, and that the current risk level prediction result is more consistent with the confirmed quality results. If it is 0, then Set to 0 and trigger the pending manual review status; The current prediction result is based on the risk level inconsistency score calculated from the corrected risk level candidate values. The final output quality risk level of S5 is used as the calculation object, and s* is denoted as the index of the final output quality risk level according to the priority of severe risk, early warning risk, attention risk, and normal risk. When the highest probability risk category is inconsistent with the final output quality risk level according to the priority of severe risk, early warning risk, attention risk, and normal risk, the conformal risk calibration module reads the corrected candidate probability corresponding to the final output quality risk level to calculate the current risk level inconsistency score, and writes the highest probability category and the final output category into the candidate level reference field at the same time.

[0027] The advance warning time is determined by the current window end time and the effective time range of the model prediction. During online inference, the conformal risk calibration module only reads the current window end time from the S4 prediction result record. Predicted duration and predicted target time According to the formula Calculate the lead time for early warning, where Written by S4 and The units are all hours. If the value is less than 0, it is set to 0 and the record is written to the timing anomaly cache; during training or calibration playback, the module allows reading the effective production window end time in the S2 sample association record as the playback window end time, but this field is not included in the online inference early warning lead calculation; The ranking of key influencing factors is jointly generated by graph attention weights, temporal Transformer attention weights, and the contribution value of the predicted output gradient. The conformal risk calibration module first maps the temporal Transformer attention from the time window dimension to the node dimension, and then adjusts the time window... internal nodes The absolute contribution of the gradient or the node coupling represents the change in the calculation of the node assignment coefficient. Press again The node-level time window attention summary value is obtained, where For nodes In the time window The amount of contribution, For time window Transformer attention weights, when When it is 0, Set the nodes to be evenly distributed within the same window and write the explanation degradation flag; The conformal risk calibration module normalizes the three types of node-level components respectively, and the node influence components are normalized according to... The node-level time window influence components are obtained according to... The output sensitive components are obtained according to... Received, among which For nodes The sum of related graph attention weights To predict the gradient contribution of the output to the node features, when , or When the value is 0, the corresponding component is set to a 0 vector and an explanation component missing flag is written into the risk identification record; The conformal risk calibration module reads a preset interpretation weight table and performs weighted fusion of the three normalized components. The final influencing factor score is calculated according to the formula. Received, among which , and The consistency rate of human cause localization of confirmed abnormal samples is obtained by calibrating the three factors, and the sum of these three factors is one. The value ranges from 0 to 1, with larger values ​​indicating higher node values. The stronger the contribution to the current risk output, the higher the final influencing factor score is generated by S5 and used to rank the main influencing factors; The conformal risk calibration module scores according to the final influencing factors. Generate a ranking of the main influencing factors from largest to smallest. If two nodes... If the values ​​are the same, the node with the larger path influence value in the S3 variable influence path table will be output first. If the path influence values ​​are still the same, the nodes will be output in a fixed order of furnace temperature, electrical parameters, feeding amount, drawing speed, furnace pressure, cooling parameters, raw material batch and manual operation event to ensure that the sorting results are stable and verifiable. When the number of recent calibration samples is less than the preset minimum number of calibration samples, the conformal risk calibration module will mark the corresponding prediction result as pending manual review. In this specific embodiment, the preset minimum number of calibration samples is determined by no less than 20 confirmed samples for each risk level, when the risk confidence level... When the value is less than the preset threshold of 0.6, the module will also output a status pending manual review and disable automatic alarm closed-loop confirmation. The conformal risk calibration module maintains a dynamic grading threshold version table. The threshold version table fields include calibration sample start and end time, sample quantity, target coverage, non-consistency score quantile, risk level unified threshold, applicable risk level, applicable operating condition label, update trigger time and failure flag. If the sample time range of the most recent threshold version does not overlap with the current production batch, the module writes a cross-batch calibration prompt in the risk identification record. When both severe risk and warning risk conditions are met in the same window, the conformal risk calibration module determines a unique quality risk level in the order of severe risk first, warning risk second, attention risk third, and normal risk last. The covered lower-level candidate results are written into the candidate level lookup field to avoid multiple conflicting risk levels being generated in the same window. The conformal risk calibration module performs a stability check on the ranking of major influencing factors. If the factor ranked first has the highest final score... If the difference between the score of the second-ranked final influencing factor and the score of the explanatory factor is less than the explanatory threshold, then the first two influencing factors are marked as co-primary causes. The explanatory threshold is obtained by calibration of the consistency rate of human cause localization. In this specific implementation, the threshold is set to 0.05. While awaiting manual review, the conformal risk calibration module retains the variable impact path from the S3 output and the intermediate quantities explained from the S4 output, and sets the quality risk level to "review pending." The review record includes the window identifier, model version, calibration version, inconsistency score, dynamic grading threshold, risk confidence level, variable impact path, ranking of major influencing factors, and reason for review triggering. When the S4 abnormal prediction result record contains a field version inconsistency flag, the conformal risk calibration module sets the risk confidence level to 0, sets the early warning advance to invalid, sets the ranking of major influencing factors to uncalculable or only retains the reference explanation from the previous stable model, and disables automatic alarm closed-loop confirmation. After manual review is completed, this record is written back to the recent calibration sample table for the next version threshold update. Anomaly priority is implemented in the following order: inconsistent field versions, mismatched graph versions, insufficient graph integrity or insufficient number of valid edges, failed sample association, pending review window, insufficient calibration samples, and insufficient risk confidence. If a previous priority has already triggered the pending manual review status, the next priority will only be added to the review reason list and will not cover the pending review position of the quality risk level. Inconsistent field versions and mismatched graph versions take precedence over automatic model inference, insufficient graph integrity and insufficient number of valid edges take precedence over automatic risk classification, and insufficient risk confidence takes precedence over automatic alarm closed-loop confirmation.

[0028] When the manual review results are inconsistent with the automatic quality risk level, the conformal risk calibration module does not directly overwrite the historical risk identification record, but generates a calibration correction record. The calibration correction record includes the original risk level, the manually confirmed risk level, the correction time, the confirmation mark of the correction personnel, the corresponding inconsistency score, and the ranking of the main influencing factors. The next time the threshold is updated, only the calibration correction record with the confirmation mark will be used. The conformal risk calibration module performs consistent encapsulation on the content output by the on-site early warning interface. The quality risk level, risk confidence, early warning lead time, and main influencing factors in the same risk identification record use the same calibration version number. If the calibration version number is updated during the calculation, the current window will still use the calibration version frozen before the calculation started, and the new calibration version will be delayed until the next window takes effect. The conformal risk calibration module ultimately outputs a risk identification record, with fields including window identifier, quality risk level, risk confidence level, early warning lead time, future failure probability, quality defect type probability, ranking of main influencing factors, variable influence path, status pending manual review, reason for review trigger, abnormal output flag, automatic handling flag, calibration sample version, and model version. The risk identification record is written to the production quality risk database and sent to the on-site early warning interface; among them, pending review, invalid, incalculable, and previous stable model reference interpretation are all used as valid set values ​​for the corresponding output fields under abnormal branches.

[0029] The above description is only a preferred embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any equivalent substitutions or modifications made by those skilled in the art within the scope of the technology disclosed in the present invention, based on the technical solution and inventive concept of the present invention, should be covered within the scope of protection of the present invention.

[0030] This invention creates a continuous technical chain between process fluctuations, subsequent quality results, equipment failure records, and risk output by performing time synchronization, windowing, lag correlation, dynamic graph modeling, graph time series prediction, and conformal calibration on continuous production data of quartz continuous melting. This enables the output of verifiable risk identification results for quality inspection lag and multivariate nonlinear coupling problems.

[0031] This invention further incorporates a learnable hysteresis response module, a process mechanism constraint dynamic graph module, and a conformal risk calibration module into the conventional deep learning prediction process. This enables the model to output not only the prediction probability but also the effective production window, variable influence path, dynamic grading threshold, and risk confidence level, making it more suitable for identifying the quality risk of quartz continuous melting under small sample abnormal conditions.

Claims

1. A method for predicting quartz continuous melting faults and identifying quality risks based on deep learning, characterized in that, include: S1. Synchronize furnace temperature, current, voltage, feeding amount, drawing speed, furnace pressure, cooling parameters, raw material batches and manual operation records in time, segment working conditions and slice sliding windows to generate a multi-source process window set; S2. Input the delayed quality inspection results and equipment failure records into the learnable lag response module to determine the effective production window corresponding to each quality defect type and equipment failure type, and generate sample association records with lag contribution weights. S3. Using process variables, raw material batches, operation events, quality defect items and equipment failure items in the multi-source process window associated with the sample association record as nodes, a dynamic process influence diagram containing sample learning diagram, calibration diagram and online inference diagram is generated based on process mechanism constraints and sample statistical relationships. The online inference diagram does not write the current window confirmation status, true label, test results and contribution weight. S4. Input the dynamic process influence diagram and multi-source process window into the graph time series prediction model to generate future failure probability, quality defect type probability, risk level candidate value, prediction time domain parameter and explanation intermediate quantity; S5. Input the future failure probability, quality defect type probability, risk level candidate value, prediction time domain parameter and interpretation intermediate quantity into the conformal risk calibration module, and output the quality risk level, risk confidence, early warning lead time and the ranking of main influencing factors. When there are inconsistencies in field versions, mismatches in graph versions, insufficient graph integrity, insufficient number of valid edges, empty candidate association windows, insufficient valid production windows, failure to generate sample association records, insufficient calibration samples, or insufficient risk confidence, output the status pending manual review, the reason for review triggering, and the abnormal output flag, and write the corresponding output fields as pending review, invalid, or uncalculate.

2. The method for predicting quartz continuous melting faults and identifying quality risks based on deep learning according to claim 1, characterized in that, S1 includes: The furnace temperature, current, voltage, feeding amount, drawing speed, furnace pressure and cooling parameters are resampled using a unified time reference, and operating condition labels are generated according to the equipment operation section, batch change section, feeding disturbance section and drawing speed adjustment section. Event coding fields are established for raw material batches and manual operation records; the multi-source process window set is formed by slicing with a preset window length and a preset step size, and the corresponding window is marked as a window to be reviewed when the missing rate of any data field is greater than the preset missing rate threshold.

3. The method for predicting quartz continuous melting faults and identifying quality risks based on deep learning according to claim 1, characterized in that, S2 include: Candidate hysteresis interval sets are set for each of the following: bubbles, streaks, crystallization, size fluctuations, electrode aging, abnormal feeding, and abnormal cooling. The learnable hysteresis response module calculates the normalized contribution weight of each candidate hysteresis interval based on the detection time, quality defect type, equipment failure type, and the multi-source process window set. According to the target retention ratio, the quantile corresponding to 1 minus the target retention ratio is read from the calibration sample contribution weight set in ascending order as the lag screening threshold. The production window with a contribution weight not less than the lag screening threshold is determined as the effective production window. The detection result, fault record, window identifier, quality defect type, equipment fault type and contribution weight are written into the sample association record. For production windows that do not have quality defect records, equipment failure records, or production handling records within the preset quality inspection observation period and equipment failure observation period, a normal sample association record is generated. The quality defect type field and the equipment failure type field are both written as 0, the quality risk level is written as normal risk, the contribution weight is set to 1 or normalized according to the window coverage ratio, and the normal sample association record is allowed to enter the S4 training sample and the S5 recent calibration sample.

4. The method for predicting quartz continuous melting faults and identifying quality risks based on deep learning according to claim 3, characterized in that, The candidate lag interval set is generated by the time difference distribution between historical quality inspection time, fault record time and production process window time, and the preset quartz continuous melting process experience boundary. The normalized contribution weight is calculated by an attention network using the time difference feature of the candidate lag interval, the embedding of the quality defect type or the equipment fault type, the embedding of the operating condition label, and the detection confidence field. The sum of the contribution weights of the candidate lag intervals corresponding to the same detection result or the same equipment fault record is one.

5. The method for predicting quartz continuous welding failures and identifying quality risks based on deep learning according to claim 1, characterized in that, S3 include: Initial edge weights are generated based on the thermoelectric coupling relationship between furnace temperature and electrical parameters, the material disturbance relationship between feeding amount and furnace pressure, the forming relationship between drawing speed and dimensional fluctuation, and the thermal history relationship between cooling parameters and crystallization risk. The initial edge weights and the statistical correlation coefficients calculated from the associated records of the samples are input into the dynamic graph module of the process mechanism constraint, and the initial edge weights are dynamically corrected by graph attention weights to obtain the dynamic process influence graph containing node fields, edge type fields, edge weight fields and time version fields. The node types and edge structures of the online inference graph are determined by historical sample association records and process prior edges. The current window only provides process variables, raw material batches, operation events and operating condition labels to update the edge weight prior values, and does not read the detection results, fault confirmation results or contribution weights of the current window.

6. The method for predicting quartz continuous melting faults and identifying quality risks based on deep learning according to claim 1, characterized in that, S4 include: The graph temporal prediction model includes a graph attention network, a temporal Transformer encoder, and a multi-task output layer; The graph attention network extracts variable coupling features based on the dynamic process influence graph; The temporal Transformer encoder extracts long- and short-term fluctuation features based on the variable coupling characteristics and the multi-source process window set. The multi-task output layer outputs the future failure probability, the quality defect type probability generated by the independent sigmoid channels for each quality defect type, the risk level candidate value, the prediction time-domain parameters, and the explanatory intermediate quantity. The explanatory intermediate quantity includes graph attention weights, temporal Transformer attention weights, and the prediction output gradient contribution value. The future failure probability includes the failure type probability generated by the independent sigmoid channels for each device failure type and the total failure probability obtained by fusing the failure type probabilities. The prediction time-domain parameters include the current window end time, the prediction duration, and the prediction target time.

7. The method for predicting quartz continuous melting faults and identifying quality risks based on deep learning according to claim 6, characterized in that, The training of the graph time series prediction model includes: Using the associated records of the samples as supervisory samples, a multi-task loss function is constructed according to the quality defect type, equipment failure type and quality risk level. The quality risk level is mapped from the confirmed quality defect degree, equipment failure confirmation status and production handling result to normal risk, attention risk, warning risk or serious risk. Both the quality defect type task and the equipment failure type task adopt the label-by-label binary cross-entropy loss. Sample weights are assigned to severe fault samples and severe defect samples. Severe fault samples are equipment fault samples that have been confirmed through repair and have led to automatic handling prohibition, production downgrade or shutdown. Severe defect samples are quality defect samples that have reached the severe level of defect or have led to product scrapping, rework or downgrade. The sample weights are determined together based on the reciprocal of the category frequency and the contribution weight. When multiple quality results or multiple device failure results are associated with the same window, the highest risk level among the quality risk levels is written into the training label, and the corresponding task losses in the multi-task loss function are weighted and summed to obtain the trained graph time series prediction model.

8. The method for predicting quartz continuous melting faults and identifying quality risks based on deep learning according to claim 6, characterized in that, S5 include: The conformal risk calibration module constructs a set of non-consistent scores for faults, a set of non-consistent scores for defects, and a set of non-consistent scores for risk levels based on the predicted scores output by the multi-task output layer and the real quality results and real fault results of recent calibration samples. Among them, the quality defect type task constructs a multi-label non-consistent score according to the probability of each label, and the equipment fault type task constructs a multi-label non-consistent score according to the probability of each fault type. The dynamic grading threshold is determined according to the quantile corresponding to the target coverage rate; During online inference, the defect prediction uncertainty component and the fault prediction uncertainty component, which are independent of the current real label, are calculated based on the current quality defect type probability and the current future failure probability, and the risk level candidate value is corrected using the defect prediction uncertainty component and the fault prediction uncertainty component. After the corrected risk level candidate values ​​are normalized to the same probability space through the same calibration mapping, they are compared with the dynamic grading thresholds applicable to the current risk level inconsistency scores of normal risk, risk of concern, risk of warning, and severe risk to generate the quality risk level. The risk confidence score is generated based on the proportion of samples in the risk level inconsistency score set that are not less than the current risk level inconsistency score. The current risk level inconsistency score is determined by the inverse of the current corrected candidate probability corresponding to the final output quality risk level.

9. The method for predicting quartz continuous melting faults and identifying quality risks based on deep learning according to claim 8, characterized in that, When any of the following conditions occur: inconsistent field versions, mismatched graph versions, graph integrity score less than graph output threshold, insufficient number of effective edges, presence of windows awaiting review, empty candidate association windows, insufficient effective production windows, failure to generate sample association records, number of recent calibration samples less than preset minimum calibration sample number, or risk confidence level less than preset confidence threshold, the conformal risk calibration module will mark the corresponding prediction result as pending manual review, prohibit automatic alarm closed-loop confirmation, generate review trigger reason and abnormal output flag, write the quality risk level as pending review, write the early warning advance as invalid, and write the ranking of the main influencing factors as uncalcifiable or previous stable model reference interpretation when the influencing factors are incalculable.

10. The method for predicting quartz continuous melting faults and identifying quality risks based on deep learning according to claim 8, characterized in that, The ranking of the main influencing factors in step S5 is generated as follows: The graph attention weights, time-series Transformer attention weights, and prediction output gradient contribution values ​​in the graph time-series prediction model are normalized to generate node influence components, node-level time-window influence components, and output sensitivity components obtained by mapping the time window attention according to the node contribution or node gradient contribution within the corresponding window. When the sum of node contributions within a time window is zero, the node allocation coefficient of the corresponding time window is set to uniform distribution of nodes within the same window and written into the interpretation degradation flag. When the normalized sum of node influence components, node-level time window influence components, or output sensitive components is zero, the corresponding normalized component is set to a zero vector and written into the interpretation component missing flag. The node influence component, the node-level time window influence component, and the output sensitivity component are weighted and fused according to the preset interpretation weights to obtain the final influence factor score, and the main influence factors are output in descending order of the final influence factor score.

Images