Quartz tube heat processing quality prediction method based on time sequence neural network

Abstract

The application discloses a quartz tube heat processing quality prediction method based on a time sequence neural network and belongs to the field of intelligent quality monitoring of quartz glass heat processing. In order to solve the problems of asynchronous multi-source process data, difficult identification of processing stages, difficult characterization of long-term heat accumulation and lagging defect early warning, the application constructs a variable time lag diagram with stage constraints through material space synchronization and stage identification, forms a long-term heat history state by using causal expansion convolution and heat dose gating recurrent unit, performs multi-task and multi-lead quality prediction, and reversely maps key evidence along the time lag diagram, so as to realize risk output before defect formation and tracing of the formation stage, material position and time window.

Description

Technical Field

[0001] This invention relates to the field of intelligent quality monitoring of quartz glass thermal processing, and in particular to a method for predicting the quality of quartz tube thermal processing based on a time-series neural network. Background Technology

[0002] The drawing, flaring, necking, welding, annealing, and cooling processes of quartz tubes involve multiple process variables, including flame temperature, heating position, rotational speed, feed rate, internal pressure, and cooling state. These variables are coupled together under the influence of axial movement, circumferential rotation, and continuous thermal action, and exhibit propagation time delays caused by equipment spatial position, material movement, and stage dwell time. Existing production processes typically collect equipment control quantities, temperature or thermal images, and action events, and inspect wall thickness, ellipticity, weld quality, and residual stress after processing.

[0003] Existing final inspection methods can only determine quality after processing is completed, making it difficult to pinpoint the material segment, circumferential location, processing stage, and duration of defect formation. When using prediction models with single-moment parameters or free learning time delays, asynchronous data is prone to material misalignment, process switching can easily generate cross-stage pseudo-correlation, and local thermal shock, circumferential cyclic action, and long-term thermal accumulation are difficult to simultaneously reach a unified state, resulting in insufficient stability of early warning and final inspection traceability.

[0004] Therefore, there is a need for a method to predict the heat treatment quality of quartz tubes that can overcome the shortcomings of the existing technology. Summary of the Invention

[0005] One objective of this invention is to propose a method for predicting the quality of hot processing of quartz tubes based on temporal neural networks. Addressing the problems in existing technologies such as difficulty in synchronizing multi-source process time-series data according to the location of the processed material, lack of utilization of stage transfer and dwell time constraints, lack of physical boundaries for propagation delays between variables, difficulty in modeling long-term thermal history, and difficulty in early identification of quality risks, the invention proposes a method that first establishes material spatial time-series data through velocity segment integration; then identifies processing stages and forms a material segment-stage dual index under the constraints of legal process transfers and stage dwell time; furthermore, determines the nominal time delay based on equipment spatial relationships, material velocity, rotation cycle, and stage dwell time; learns the time delay residuals only within the allowed interval and switches connection edges according to stage masks; finally, forms a long-term thermal history state using causal dilated convolution and heat dose-gated cyclic units, and performs multi-task, multi-advance prediction and evidence-based reverse mapping, enabling the prediction results to correspond to specific process parameters, processing stages, material locations, and formation time windows.

[0006] This invention provides a method for predicting the heat treatment quality of quartz tubes based on a time-series neural network, comprising: S1. Acquire multi-source process timing data and action event data, integrate the feed speed and rotation speed piecewise, determine the material axial coordinate and circumferential phase, and map the multi-source process timing data and action event data to the material segment, circumferential sector and sampling time to obtain material spatial timing data; S2. Based on the action event data and the process status data in the material space time series data, identify the processing stage under the constraints of process transfer relationship and stage dwell time, obtain the stage identification result and stage mask, and organize the material space time series data according to material segment and processing stage to obtain dual-index time series data. S3. Based on the dual-index time series data and the stage mask, construct a variable time delay graph with stage constraints. Determine the nominal time delay of the connecting edge according to the distance between the acquisition position of the process variable and the position of the processed material, the material movement speed, the rotation cycle and the stage dwell time. Learn the time delay residual within the allowable range and switch the connecting edge according to the processing stage to obtain the graph time series characteristics. S4. Input the graph time-series features into a time-series neural network, extract local impact features and circumferential periodic features through causal dilated convolution, recursively deduce the thermal dose state, and control the cycle unit update with the processing stage and the thermal dose state to obtain the long-term thermal history state. S5. Input the long-term thermal history state into the multi-advance prediction unit and output the risk probability or quality prediction value of the target quality defect.

[0007] Optionally, S1 includes: accumulating the feed rate according to the sampling time interval to obtain the axial displacement of the material, and accumulating the rotational speed according to the sampling time interval and converting it by one rotation cycle to obtain the circumferential phase; The current integration segment is terminated and the next integration segment is established at the action events corresponding to the clamping direction reversal, feed pause or process switching. The start and end times, motion direction and coordinate reference of each integration segment are recorded. The material segment is determined based on the preset axial interval to which the material's axial coordinate belongs, and the circumferential sector is determined based on the preset phase interval to which the circumferential phase belongs. Resampling is performed at a preset common sampling time based on the timestamps of various process data. For continuously changing quantities, adjacent valid sample values ​​are interpolated, and for hold-type state quantities, the previous valid sample value is used for hold. When the time difference between the most recent valid sample value and the common sampling time exceeds the allowable time deviation determined by the sampling period of the corresponding sensor, the corresponding data at the common sampling time is marked as invalid.

[0008] Optionally, S2 includes: establishing a set of process states including drawing, flaring, necking, welding, annealing and cooling stages according to the hot working process specification, as well as process transfer relationships that allow transfer between adjacent process states; Candidate stage boundaries are formed based on the triggering time of action events, and the stage dwell time range is determined based on the duration statistical interval of each process status in the historical qualified processing records. The process state observation model corresponding to each processing stage calculates the matching probability of each sampling time belonging to the corresponding processing stage based on the flame working state, rotation state, feed state, tube pressure state and cooling state. The logarithm of the matching probability contained in the candidate stage sequence is taken and accumulated to obtain the sequence score. The candidate stage sequence that satisfies the process transfer relationship and the stage dwell time range and has the largest sequence score is selected as the stage identification result. The graph connection edges allowed to be used in the corresponding processing stage are recorded as valid to form the stage mask. When there is no candidate stage sequence that satisfies the process transfer relationship and the stage dwell time range, the corresponding sampling interval is marked as an unknown stage to generate the stage identification result, and the stage-related connection edges of the corresponding sampling interval are marked as invalid in the stage mask.

[0009] Optionally, S3 includes: establishing process variable nodes for flame temperature, heating position, rotation speed, feed speed, tube pressure and cooling parameters respectively; establishing quality nodes for target quality defects; establishing directed edges from process variable nodes to quality nodes and directed edges between process variable nodes that characterize the coupling relationship of process variables; and controlling whether the directed edges participate in graph feature aggregation through the stage mask. For directed edges that propagate along the material's axial movement, the propagation time within the same integration segment is determined by the axial distance from the corresponding acquisition position to the material segment position and the material's movement speed. When the feed is paused or the absolute value of the material's movement speed is less than the equipment's minimum stable feed speed, the corresponding axially propagating directed edges are disabled. When the clamping direction is reversed or the process is switched, the propagation of directed edges across integration segments is prohibited. For a directed edge that repeats its action with circumferential rotation, the phase propagation time is determined by the circumferential phase difference and the rotation period, and the nominal time delay is determined by combining the stage dwell time. The residual allowable intervals for each directed edge are determined based on the equipment calibration error, speed fluctuation range, and stage boundary error. The sum of the nominal time delay and the time delay residual constrained by the interval is taken as the actual time delay of the directed edge, and the weights of the directed edges are learned.

[0010] Optionally, S4 includes: the thermal dose state is generated recursively according to the sampling time, and the recursive input of the thermal dose state includes the duration of the temperature exceeding the set value of the quartz tube softening temperature, the range or variance of the temperature of each circumferential sector within the same material segment, the cumulative amount of the pressure inside the tube with respect to the sampling time, and the rate of change of temperature with respect to time within the cooling interval. Each recursive input is normalized and weighted according to the dimensional statistics of the training data. The fusion result is then recursively combined with the thermal dose state at the previous sampling time according to the attenuation coefficient corresponding to the processing stage to obtain the thermal dose state at the current sampling time. The corresponding processing stage embedding is obtained by querying the stage coding table based on the stage identification result. The heat dose state is used as the state node input of the variable time delay diagram. The heat dose state and the processing stage embedding are used together to determine the input gate and forget gate of the loop unit. Furthermore, the causal dilated convolution uses convolutional connections that only reference data from the current sampling time and previous sampling times, sets an increasing dilation interval along the time dimension, and ensures that at least one convolutional receptive field covers the number of sampling points corresponding to one rotation cycle. The output of the causal dilated convolution, the node features aggregated by the variable time delay graph, the processing stage embedding, and the thermal dose state are input into the loop unit. The current input write amount is controlled by the input threshold, and the amount of the previous long-term thermal history state is retained by the forget threshold, so as to update the long-term thermal history state.

[0011] Optionally, S5 includes: setting up quality classification tasks and quality regression tasks for uneven wall thickness, excessive ellipticity, insufficient welding and residual stress respectively, and setting prediction outputs for multiple preset time intervals before each quality inspection time. The quality classification task outputs the risk probability of the corresponding target quality defect forming within each preset time interval, and the quality regression task outputs the wall thickness deviation, ellipticity, weld quality characterization value or residual stress characterization value. From historical quality inspection data independent of the temporal neural network to be trained, wall thickness deviation, ellipticity, weld quality characterization value or residual stress characterization value associated with material segment and circumferential sector are obtained as target quality indicators. The quality deviation of the target quality indicator relative to the quality target value is calculated, and the maximum correction amount of the remaining adjustable process after each sampling time is determined according to the historical qualified correction record. The first sampling time when the absolute value of the quality deviation is still greater than the corresponding quality qualified limit allowable deviation after deducting the maximum correction amount is determined as the time when the defect is irreversibly formed. The preset time interval is determined according to the historical duration statistics between the quality inspection time and the time when the defect is irreversibly formed. Furthermore, it also includes the reverse mapping of key evidence: based on the sensitivity of the prediction results to the long-term thermal history state and the connection edge weights determined by training in the variable time delay graph, the contribution value of each graph node to the prediction results is determined. Based on the sum of the nominal time delay of the connecting edge and the time delay residual constrained by the allowable interval, the sampling time corresponding to the contribution value is shifted back to the time of action of the process variable, and the position of action of the process variable and the processing stage are determined according to the material segment identifier, circumferential sector and stage identification results. The continuous sampling intervals whose contribution values ​​reach the evidence threshold determined by the contribution values ​​of the training samples and the preset quantile level are merged into a forming time window, and the process parameters corresponding to the continuous sampling interval are output. Furthermore, the training of the temporal neural network includes: generating supervision labels corresponding to material segments, circumferential sectors, and defect types based on the detection location and detection time of the quality inspection data; The supervision labels are moved forward to the corresponding sampling time according to each preset time interval to form multi-advance training labels; A label mask that does not calculate the loss for undetected material segments and undetected circumferential sectors is used to jointly calculate the quality classification loss, quality regression loss, time delay residual interval constraint loss, and adjacent lead prediction consistency loss. The parameters of the connection edge weights, time delay residuals, causal dilated convolution, recurrent units, and multi-task and multi-lead prediction units of the variable time delay graph are updated according to the joint loss. Furthermore, when there are missing or abnormal samples in the multi-source process time series data, a data validity mask is generated based on the valid sensor sampling flag and equipment action event. The graph node input corresponding to the invalid data is set to a preset fill value and the corresponding connection edge is prohibited from participating in graph feature aggregation. The credibility flag of the prediction result is determined based on the effective sampling ratio of the current material segment and the effective sampling ratio range in the training samples, and the credibility flag is output together with the risk probability and quality prediction value.

[0012] The beneficial effects of this invention are: 1. By integrating the feed rate and rotational speed in segments and dividing the integration segments at the reversal, pause and process switching points, process data with different sampling rates can be mapped to a unified material segment, circumferential sector and sampling time, reducing the impact of material missegmentation on quality modeling.

[0013] 2. By constraining the variable time delay diagram through process transfer relationships, stage dwell time, and stage mask, and aligning the variable effects with physical nominal time delay and constrained time delay residuals, cross-stage pseudo-correlation can be suppressed; the heat dose state and stage embedding jointly control the cyclic memory, which is beneficial for characterizing local shocks, circumferential cycles, and long-term heat accumulation.

[0014] 3. By predicting multiple defects and multiple lead times and contributing evidence through reverse mapping along a variable time delay diagram, risk probabilities and quality predictions can be output before quality inspection, and key process parameters, action stages, material locations, and formation time windows can be given, providing consistent data for early warning and final inspection traceability. 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 flowchart of the method for predicting the heat treatment quality of quartz tubes based on a time-series neural network, as described in this invention.

[0016] Figure 2 This is a flowchart for constructing the variable time delay graph for the S3 stage constraints of this 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 Figures 1-2 A method for predicting the heat treatment quality of quartz tubes based on temporal neural networks includes: S1. Acquire multi-source process timing data and action event data, integrate the feed speed and rotation speed piecewise, determine the material axial coordinate and circumferential phase, and map the multi-source process timing data and action event data to the material segment, circumferential sector and sampling time to obtain material spatial timing data; S2. Based on the action event data and the process status data in the material space time series data, identify the processing stage under the constraints of process transfer relationship and stage dwell time, obtain the stage identification result and stage mask, and organize the material space time series data according to material segment and processing stage to obtain dual-index time series data. S3. Based on the dual-index time series data and the stage mask, construct a variable time delay graph with stage constraints. Determine the nominal time delay of the connecting edge according to the distance between the acquisition position of the process variable and the position of the processed material, the material movement speed, the rotation cycle and the stage dwell time. Learn the time delay residual within the allowable range and switch the connecting edge according to the processing stage to obtain the graph time series characteristics. S4. Input the graph time-series features into a time-series neural network, extract local impact features and circumferential periodic features through causal dilated convolution, recursively deduce the thermal dose state, and control the cycle unit update with the processing stage and the thermal dose state to obtain the long-term thermal history state. S5. Input the long-term thermal history state into the multi-advance prediction unit and output the risk probability or quality prediction value of the target quality defect.

[0019] In this specific embodiment, S1 includes: This specific implementation method focuses on the same processing batch within a continuous quartz tube drawing, flaring, necking, welding, annealing, and cooling production line. The data acquisition industrial control computer uses a 100... The ms common sampling period receives the temperature and temperature measurement position output by the flame thermometer, the rotational speed output by the spindle encoder, the feed speed output by the servo driver, the pipe pressure output by the pressure transmitter, the cooling valve opening, and batch start, general process switching, drawing start and end, flaring start and end, necking start and end, welding start and end, annealing start and end, cooling start and end, flame start and stop, clamping direction reversal, and feed pause action events. Action events are generated by the process code, actuator start position, and end position of the production line controller when the state changes. The event type, previous process code, subsequent process code, and monotonically increasing event sequence number are written. Each original record is written with batch identifier, sensor identifier, equipment timestamp, sampled value, engineering unit, sampling validity flag, and action event sequence number. Before acquisition, the timestamp of each device is corrected using the industrial control computer's synchronous clock, and the record order conflict under the same timestamp is eliminated by the action event sequence number. The data acquisition industrial control computer follows the data acquisition process within each integration segment. Calculate the axial coordinates of the material, and according to Calculate the circumferential phase, where, Indicates the first The material axial coordinates corresponding to each common sampling moment are given in mm. The coordinate reference is the clamping reference surface at the starting point of the current integration segment. This represents the axial coordinate of the material at the previous common sampling time, in mm. This indicates the first [number] read by the servo driver and assigned a positive or negative sign according to the clamping direction. The feed rate at any given moment, in mm / s. This indicates the common sampling period, which is 0.1 s in this specific implementation. Indicates the first Converted to time The circumferential phase, in rad. This indicates the circumferential phase at the previous moment, in rad. Indicates the number read by the spindle encoder Angular velocity at each moment, in rad / s This represents the sampling sequence number within the current integration segment, incremented by time. π is a dimensionless value. This represents the operation of taking the phase modulo over one rotation period; When the action event record shows a reverse clamping direction, a feed pause for three consecutive common sampling cycles, or a process switch, the data acquisition industrial control computer terminates the current integration segment and generates a new integration segment identifier at the event trigger moment. The segment record saves the start and end times, motion direction, starting point axial coordinate, starting point circumferential phase, and coordinate reference version. When the clamping direction is reversed, the last valid coordinate before the reversal is used as the new segment reference, but the feed speed sign is changed. When the feed is paused, the axial coordinate is maintained and the valid rotation phase continues to accumulate. When the process switches, the coordinate mapping version of the old segment is frozen at the same time, so that subsequent diagram connections will not cross the direction or process boundary and reference the wrong material position. The material mapping module divides the material into segments with a width of 5 mm along the axial coordinates and into twelve circumferential sectors with a phase of 30°, according to... and Determine the mapping index, where, Indicates the first The material segment number at each moment is a non-negative integer. This indicates the first value obtained according to the aforementioned axial coordinate recursive formula. The material's axial coordinates at a common sampling time are shown in mm. This indicates the axial coordinate reference of the first integration segment in the current batch, in mm. This indicates the length of the material segment, which is 5 mm in this specific embodiment. Indicates the first The circumferential sector number at each moment, with values ​​ranging from 0 to 11. This represents the total number of circumferential sectors, which is 12 in this specific embodiment and is dimensionless. This indicates the first phase obtained according to the aforementioned circumferential phase formula. The circumferential phase at each moment, in rad. This represents the aforementioned pi, which is dimensionless. This indicates a floor operation; records that exceed the effective length of the current batch are written to the boundary exception queue and do not enter subsequent training and inference. The resampling module performs resampling to a preset common sampling time based on the timestamps of various process data. For continuously changing quantities, it uses interpolation of adjacent valid sample values. Specifically, it employs... Perform linear interpolation of adjacent valid sample values, where, Indicates the time of common sampling The interpolation results should be retained for the original engineering units. and Representing timestamps and Adjacent valid sampled values ​​from the same sensor and in units of same, , and Both represent time in seconds as given by the synchronous clock. Indicates strict earlier than The previous valid sampling time, Indicates strictly later than The next effective sampling time, and satisfying and When duplicate timestamps appear on the same sensor, first retain one record according to the sampling validity flag, action event sequence number, and absolute value of synchronization error from high priority to low priority, and write the remaining records to the deduplication log. If a strictly increasing time synchronization cannot be formed after deduplication, do not perform interpolation, mark the corresponding common sampling point as invalid, and enter the existing filling and quality record branch. For holdable state variables such as flame start / stop, clamping direction, and process status, the previous valid sampling value is used to hold them, and interpolation between different integration segments is prohibited. The validity determination module is based on And only when Furthermore, the original sample validity flag is set to true to generate a data validity mask, where, Indicates the first Type of sensor in the first A validity flag for each common sampling time, with a value of 0 or 1. This indicates the most recent effective sampling time of the sensor from the common sampling moment, in seconds. Indicates the first A common sampling time, measured in seconds. This indicates the first [item] read from the device configuration table. The rated sampling period of this type of sensor is expressed in seconds. This represents the allowable time deviation coefficient, which is taken as 1.5 and is dimensionless in this specific embodiment. This indicates the data source sequence number corresponding to flame temperature, heating position, rotation speed, feed speed, tube pressure, or cooling parameters. If the condition is not met, the corresponding data at the common sampling time is marked as invalid, and the corresponding value is set as the median of the effective samples in the sensor training set, while retaining the zero-value validity flag to prevent the filled value from being interpreted as the real measurement by the network. The material mapping module uses batch identifier, integration segment identifier, material segment number, circumferential sector number, and common sampling time to form a joint key. It writes the original value, resampled value, engineering unit, data validity mask, action event bitmap, axial coordinate, and circumferential phase into the material space time series record. When the joint key is repeated, the record with the smallest absolute value of synchronization error is retained and the remaining records are written to the deduplication log. Finally, the material space time series data is formed in ascending order of material segment number, circumferential sector number, and common sampling time. Together with the coordinate mapping version and integration segment boundary, it is written into the batch time series library for S2 to read. The data acquisition configuration table stores the rated sampling period, engineering units, range, resolution, time deviation coefficient, and most recent calibration time using the sensor identifier as the key. Before the batch starts, the channel is verified according to the standard temperature source, standard pressure source, and encoder zero-position record. Channels whose verification values ​​exceed the allowable error of the equipment are not included in the current batch, and the data validity mask is continuously set to zero. The configuration table version is saved with the material space time sequence record to ensure that the same conversion relationship is used for training and inference. After the segmented integration is completed, the coordinate verification module compares the endpoint coordinates of each integration segment with the cumulative displacement register of the servo drive, and compares the cumulative circumferential phase with the encoder full circle count. If the axial difference exceeds 1 mm or the circumferential difference exceeds 3°, the entire integration segment is marked as pending verification and prohibited from entering the model training. If the limit is not exceeded, the integration result and comparison residual are retained, thereby enabling the next integration segment established at the reversal, pause and process switching points to have a traceable coordinate reference. The resampling quality record saves the common sampling time, the effective sampling time before and after, the interpolation or hold method, the time difference, the allowable time deviation, and the final validity flag. For continuously changing quantities, the record interpolated by adjacent effective sampling values ​​is written with an interpolation flag. For hold-type state quantities, the record held by the previous effective sampling value is written with a hold flag. When the most recent effective sampling value exceeds the allowable time deviation, the corresponding data of the common sampling time is marked as invalid. The quality record and the material space time series data use the same concatenation key and are provided to S2 together.

[0020] In this specific embodiment, S2 includes: The stage identification module reads the material space timing data and action event bitmap written by S1, and establishes six processing stages, namely drawing, flaring, necking, welding, annealing and cooling, as well as unknown stages, based on the current quartz tube heat treatment process specification. The process transfer table stores the allowable flag, the shortest dwell time, the longest dwell time and the specification version with the current stage and the candidate next stage as the joint key. Allowed transfers include keeping the same stage and transferring to adjacent stages as specified in the process specification. Transfers that do not comply with the specification, such as jumping directly from cooling to welding, are prohibited. The stage identification module extracts the duration of each stage from the processing records of the same type of quartz tubes that have passed final inspection in the past six months. After removing abnormal records that are below the 1st percentile or above the 99th percentile, the 5th percentile and 95th percentile are written as the lower limit and upper limit of the corresponding stage residence time range, respectively. Each welding stage record in the residence time parameter table saves the stage code, 30 s lower limit, 180 s upper limit, statistical sample number, process specification version and effective time. The trigger time of the action event is used as the candidate stage boundary, and two common sampling periods before and after the trigger time form a window that allows fine adjustment. For each common sampling moment, the stage identification module assembles the flame working indicator, normalized rotational speed, normalized feed rate, normalized tube pressure, cooling valve opening, and action event bitmap into an observation vector, and then... The stage matching probability is obtained, where, Indicates the first The common sampling time belongs to the first The matching probability of the class at the known processing stage takes a value of Furthermore, the sum of the probabilities of each stage at the same moment is 1. The phase observation model is for the first stage. The output of the nth observation vector Similar to dimensionless logarithmic values, This represents an exponential function with the natural constant as its base and its input being dimensionless logarithmic values. This represents the stage category index in the summation. This indicates the number of known processing stage categories, which is 6 in this specific implementation and is dimensionless. The category index represents the drawing, flaring, shrinking, welding, annealing or cooling. The stage observation model uses a two-layer fully connected classifier and is trained by annotating action events in historical qualified records and process specifications. The training set and validation set are divided into 8:2 according to the processing batch. Candidate sequence decoding employs dynamic programming with residence time constraints, according to... Cumulative sequence score, where, Indicates as of the date A common sampling time and currently in the stage And have been stationed Maximum dimensionless sequence score for each sampling period and These represent the candidate stage at the previous moment and its number of sampling periods, respectively. This indicates that the maximum score is selected from the candidate prestates that satisfy the following branch predicates. Represents the natural logarithm function. This indicates the first stage matching probability obtained according to the aforementioned stage matching probability formula. At the [time]th moment Class-stage matching probability; when When the time belongs to the same stage, keep it, And only verify Not exceeding the current stage's maximum stay limit; when This is a cross-stage switch, only when Located between the previous stage's residency limit and upper limit, with the process transfer table allowing it and the action event conditions meeting the requirements. ; and All are positive integers. The dwell count of the first known stage of the batch is initialized to one sampling period. The initial score is given by the candidate stage limited by the first action event of the batch. Thus, early samples in the same stage are not blocked by the shortest dwell condition, and cross-stage switching is still constrained by the previous dwell range. When the candidate stage boundary is within the action event fine-tuning window, dynamic programming allows boundary movement but imposes a dimensionless negative penalty of 0.2 per cycle on the number of sampling cycles deviating from the event trigger time. For candidate sequences with the same score, the one with the smallest boundary offset is selected first, followed by the one with fewer stage switching times. If no candidate sequence simultaneously satisfies the process transfer relationship and dwell time range exists in a continuous interval, the stage identification module marks the interval as an unknown stage. It also saves the last credible stage before entering the unknown stage, the entry time of the unknown stage, and the valid stage events within the unknown interval arranged by monotonically increasing event numbers. It clears the known stage dwell count output externally, sets the mask of all stage-related connection edges to 0, and records the action event number and constraint field that caused the unsolvable problem. During recovery, events within the unknown interval are replayed sequentially from the last credible stage according to their event numbers. Batch re-initialization events have the highest priority, and all events before them to be replayed are cleared. Events are only updated in the pre-replay state if the process transfer predicate table allows a transition from the current pre-replay state to the subsequent stage. Events with clearly defined process routes do not participate in the stage and are handled according to existing legal skip rules. Events with duplicate or reversed sequence numbers, conflicting pre- and post-order codes, or where the subsequent stage cannot be uniquely determined are all written to the conflict log and remain in the unknown stage. The unique stage obtained after replay is used as a recovery candidate. It exits the unknown stage only when at least eight of the ten consecutive common sampling periods are valid, the probability of the recovery candidate stage is maximized, and the continuous observations after the last accepted event still satisfy the stage predicate. The dwell count is then initialized to the number of consecutive sampling periods matching the candidate after the last accepted event, with a minimum of one sampling period, and dynamic programming is restarted. If the candidate is not unique, replay conflicts are not eliminated, or the continuous observation condition is not met, the unknown stage and all-zero stage mask continue to be output to avoid forcibly classifying abnormal conditions into known stages. The stage identification module generates stage identification result records based on the optimal candidate sequence. Record fields include batch identifier, material segment number, circumferential sector number, common sampling time, stage code, matching probability, number of dwell sampling cycles, candidate boundary identifier, and procedure version. It also generates a stage mask according to the stage-edge type configuration table. In the The phase at the moment allows the first The value is 1 if the edge connection in the graph is not in an unknown phase at that moment, and 0 otherwise. This represents the stage mask value and is either 0 or 1. This indicates the index of the connection edge type between process variables and quality nodes or process variables in S3. This indicates the common sampling time sequence number. The stage-edge type configuration table is established by process engineers based on the published procedures and is updated with the procedure version. The data organization module uses material segment number and stage code to form a first-level joint index, and circumferential sector number and common sampling time to form a second-level ordered index. Sensor values, action event bitmap and data validity mask in S1 are written to the same dual-index record along with stage identification results, stage matching probability and stage mask. When the stage is unknown, the material space measurement value is still retained but the stage-related connection edge is marked as invalid. Finally, dual-index time series data and stage boundary table are generated and written to the batch feature library for S3 to read continuously by material segment, stage and time. Before the observation vector is generated, the stage identification module truncates the continuous quantity according to the effective range of the same equipment channel in the training set, and then performs dimensionless scaling based on the median and interquartile range of the historical qualified samples in the same stage. Channels with zero data validity mask are filled with the scaled zero value and a missing indicator bit is added. If less than three of the five states of flame, feed, rotation, pressure and cooling are valid, the observation model is not called and unknown stage candidates are directly generated. The process transfer predicate table stores, item by item, allowed adjacent stages, dedicated action events, alternative observation predicates, and fine-tuning windows for two common sampling periods before and after, using the procedure version and batch process route as keys. The process transfer parameters are as follows: batch initialization to drawing requires a batch start event and feed is valid; drawing to flaring requires a flaring start event; drawing to necking requires a necking start event; flaring to necking requires a necking start event; flaring or necking to welding requires a welding start event; welding to annealing requires a welding end event or annealing start event; and annealing to cooling requires a cooling start event. If the controller does not have dedicated start or end bits, the preceding and following process codes in the general process switching event must be consistent with the transfer record, and the corresponding parameters must be satisfied for at least three consecutive sampling periods within the fine-tuning window. Alternate observation predicates include: flaring requires a decrease in feed rate and effective rotation; necking requires the feed direction to be consistent with the procedure setting and the internal pressure to be within the necking range; welding requires flame operation and feed rate to be below the welding threshold; annealing requires the annealing temperature zone to be effective and the welding execution position to be closed; and cooling requires the cooling valve opening to exceed the procedure threshold. Batches not participating in a certain stage are explicitly skipped by the participation flag in the process route. The transfer table directly connects the two participating stages before and after this stage and uses the general process switching event for verification. Unknown stages are not output due to legitimate skipping. Stage transfers that do not match the current process route, event type, and alternative observation predicates are considered prohibited. When the procedure is released or the historical qualified records accumulate to 500 batches, the residence range is recalculated and a new version is generated. The stage result verification module writes the score difference between the best candidate sequence and the second best candidate sequence into the stage confidence field. When the score difference is lower than the 5th percentile of the validation set, the best stage is retained but the low confidence flag is set to one. The stage mask is still generated strictly according to the best stage and the procedure version. Subsequently, S3 uses the low confidence flag as a node feature without widening the connection edges, so that the stage identification result, stage mask and dual-index time series data share the same material segment, stage, sector and sampling time index.

[0021] In this specific embodiment, S3 includes: The graph construction module takes the dual-indexed time-series data output by S2 as input, and establishes process variable nodes for flame temperature, heating position, rotation speed, feed rate, tube pressure, and cooling parameters, respectively. It also establishes quality nodes for wall thickness inhomogeneity, ellipticity deviation, insufficient welding, and residual stress, and establishes thermal history state nodes for each batch-material segment-circumferential sector. The graph aggregation at each time step only reads the thermal dose state of S4 solidified in the previous time step. And a long-term thermal history summary, without reading data that has not been updated at the current moment. This forms an acyclic execution sequence of S3 aggregation, S4 update, and re-aggregation at the next moment; the node record stores the node type, material segment number, circumferential sector number, sampling time, standardized feature vector, data validity flag, and stage code; the directed edge record stores the source node, target node, propagation source, equipment acquisition location, nominal time delay, residual allowable interval, time delay residual, connection edge weight, and stage allowable bitmap. For the propagation along the axial motion of the material, the first For each directed edge, the graph construction module first transforms the center coordinates of the material segment to the equipment calibration coordinate system based on the coordinate reference version of the integral segment saved in S1. This transformation uses the calibration coordinates of the clamping reference surface in the equipment coordinate system as the translation amount, and determines the positive direction of the material coordinates based on the current direction of movement, thus obtaining... ; then according to Determine the signed axial propagation time, where, Indicates the first The edge at the 1st The axial propagation time at each common sampling moment, in seconds. This represents the center coordinates of the processed material segment after conversion using the current coordinate reference version, in mm. This represents the non-negative integer index of the currently processed physical material segment read from the S1 material mapping result, and is jointly bound to the current batch identifier, the integral segment coordinate reference version, and the material mapping version. Indicates the first item in the equipment calibration table The stripe corresponds to the device coordinates of the sensor's operating position, in mm. This represents the material feed rate recorded in S1 and labeled in the positive direction of the equipment coordinate system, in mm / s. Indicates the index of the directed edge. Indicates the index of the common sampling time; only when , This axial propagation edge is enabled when the source time corresponds to the same physical material mapping and does not cross stages, integration segments, or edge configuration version boundaries. This indicates the minimum stable feed rate of the equipment, which is 0.5 mm / s in this specific embodiment. It represents absolute value operations, and the symbol "≥" indicates a relationship greater than or equal to. When the feed is paused, the speed is below the threshold, the coordinate transformation verification is not completed after the clamping direction is reversed, or the signed propagation time is negative, the reason for the direction being unreachable is written and the edge is disabled. The future meeting position is not converted into the historical source time using absolute distance. For the first repetitive action with circumferential rotation A directed edge, the graph construction module according to Determine the phase propagation time, where, Indicates the first The edge at the 1st The phase propagation time at each moment, in seconds. This represents the phase difference from the sensor's active phase along the current rotation direction to the center of the target's circumferential sector, converted to... And the unit is rad. This indicates the non-negative complete rotation cycle read by the current edge configuration version, which does not change with the increase of cache length or stage dwell time: the most recently applied edge is 0, the fixed cycle applied edge reads a fixed value from the device thermal response calibration table, and the multi-cycle accumulation mode expands each allowed cycle into an independent directed edge with a stable edge identifier and aggregates them together with the normalized connection edge weights trained on the validation set. π is a dimensionless value. This represents the angular velocity recorded by S1, in rad / s. This represents the minimum stable angular velocity of the device, which is 0.2 rad / s in this specific embodiment. This represents the absolute value operation. This indicates that the maximum value is selected from the two inputs; the current stage's dwell time, the traceable duration of the integration segment, and the effective historical cache duration are only used as the upper limit of the availability of the edge for this cycle. If the phase propagation time corresponding to the configured cycle exceeds any of the upper limits or If so, disable the edge of that cycle and record the unavailable phase delay flag, without changing the maximum number of backtracking cycles; The graph construction module sets the circumferential propagation time to 0 s for edges affected only by axial propagation, and sets the axial propagation time to 0 s for edges affected only by circumferential periodicity. For edges affected by both axial and circumferential forces, the module uses the sum of both propagation times and forms the graph. As the upper limit for allowed backtracking; only when season Otherwise, disable that edge without truncating the physical propagation time to the allowed backtracking upper limit; where, Indicates the first The edge at the 1st The maximum allowed backtracking time at any given moment, in seconds. This indicates the number of sampling periods that have been resident in the current stage as output by S2. This indicates the aforementioned common sampling period, in seconds. This indicates the effective historical cache duration (in seconds) for edges with the same configuration version within the current integration segment. This indicates the current stage read from the S2 residence time parameter table. The maximum stay time, in seconds. This represents the untruncated nominal physical time delay, measured in seconds (s). and These represent the times obtained according to the aforementioned axial propagation formula and phase propagation formula, respectively, in seconds. This indicates that the minimum value among the listed times is taken. Indicates a less than or equal to relation; when Physical transmission time exceeds If the source time crosses the stage boundary, edge configuration version boundary, or unverifiable material mapping boundary, disable the edge and record the reason for disabling it. The residual boundary is jointly determined by the equipment calibration error, speed fluctuation range, phase calibration error, angular velocity fluctuation, and stage boundary error; the graph construction module first follows... Calculate the axial error components and follow the... Calculate the circumferential error components, and then... The residual half-width is obtained; among which, and These represent the axial propagation error component and the circumferential propagation error component, respectively, both in seconds (s). This indicates the allowable half-width of the time-delay residual, in seconds. This indicates the equivalent time error of the clock joint calibration, expressed in seconds (s). and These are 0 or 1 indicators, representing whether the edge contains axial propagation and circumferential propagation, respectively. and These represent the coordinates of the sensor's operating location and the converted coordinates of the material segment's center, respectively, in mm. and These represent the aforementioned signed feed rate and minimum steady-state feed rate, respectively, in mm / s. This represents the 95th percentile absolute fluctuation of the feed rate, expressed in mm / s. This indicates the joint calibration error of axial position and coordinate transformation, in mm. This indicates the phase calibration error, in rad. This indicates the aforementioned phase difference, in rad. This represents the absolute fluctuation of the 95th percentile of angular velocity, expressed in rad / s. and These represent the aforementioned angular velocity and minimum steady angular velocity, respectively, in rad / s. This represents the 95th percentile error at the stage boundary, expressed in seconds (s). This represents the absolute value operation. This indicates that the maximum value is selected from the two inputs; the axial side, circumferential side, and mixed side are selected by the corresponding error components of the indication quantity, and the resulting half-width is limited to 0.1 s to 5 s after equipment verification. The time-delay learning module is formed first. and Then according to The actual time delay is obtained; among which, and These represent the lower and upper bounds of the actual non-negative feasible interval with time delay, respectively, in seconds. This represents the time obtained according to the aforementioned nominal time delay formula, in seconds. This represents the allowable half-width obtained according to the aforementioned residual half-width formula, in seconds (s). This represents the upper limit obtained according to the aforementioned formula for allowed backtracking time, in seconds. This indicates that the largest value is selected. This indicates that the smallest value is selected. This indicates the actual sampling delay, expressed in seconds (s). Indicates the first update that occurs during model training. Dimensionless original residual parameters of the strip edge This indicates that its input is mapped to The function; if If the source time is not in the same stage, integration segment, or edge configuration version, then the edge is disabled and fixed after training is completed. Half-width version and model version; During graph feature aggregation, the graph construction module follows... Generate target node features; where, Represents the target node In the Dimensionless plot temporal feature vector at each time step Indicates the target node index. Indicates the source node index. Indicates that it is from the source node Point to target node The set of directed edges, This represents the 0 or 1 stage mask generated by S2. Indicates the actual source time The binding 0 or 1 lag validity flag is set to 1 only if both interpolation endpoints surrounding the source time are valid and belong to the same stage, the same integration segment, and the same side configuration version; otherwise, it is set to 0 and interpolation is disabled. This represents the softmax-normalized edge weights obtained during training and among the valid incoming edges to the same target node. This represents the dimensionless edge-type linear transformation matrix obtained through training. This represents the dimensionless eigenvector of the source node obtained by linear interpolation at the actual source time. Indicates the first A common sampling time, measured in seconds. This represents the time obtained according to the aforementioned actual time delay formula, in seconds. Represents the ReLU activation function; the side evidence record synchronously saves the two interpolation endpoints and The reasons for the judgment; When all incoming edges of a target node are disabled at the current time due to stage, integration segment, direction unreachability, cycle unavailability, or data validity, the graph construction module enters a safe rollback state. It sets the graph temporal features of the node to the zero center vector of the training samples in the same stage, writes the empty neighborhood flag and all reasons for disabling, and prohibits using the node features at this time as credible evidence or triggering normal warnings of the neural network. After at least one incoming edge re-satisfies the stage, direction, cycle, material mapping, and validity predicates at three consecutive common sampling times, it exits the safe rollback state. Before recovery, the empty neighborhood flag is still retained for S4 and S5 downgrade processing. Under normal circumstances, the actual edge identifiers, nominal time delays, restricted time delay residuals, connection edge weights, and source sampling times are saved. Finally, the graph temporal features, empty neighborhood flags, and edge evidence records of each material segment, circumferential sector, and common sampling time are written into the graph feature library for S4 to read, and the edge-level index used for S5 evidence reverse mapping is retained. The node standardization parameters are obtained by statistically analyzing valid samples of the same node type and processing stage in the training set. The parameter table stores the mean, standard deviation, engineering units, effective range and model version. During runtime, the measured value is first subtracted from the corresponding mean and divided by the standard deviation, and then truncated to a dimensionless range of -5 to +5. When the standard deviation is lower than the sensor resolution, the resolution is used instead to avoid different engineering units from being directly aggregated and summed in the same graph. Candidate directed edges are generated based on the spatial reachability of equipment and the variable interaction relationships in the process specification. Candidate edges are deleted when there is no spatial propagation path between the source node and the target node, the stage to which they belong is not allowed, or they have never been simultaneously valid in historical training samples. For edges with the same source target and propagation source that are repeated, only the one with the smallest calibration error is retained. The retained edges generate stable edge identifiers based on the source node, target node, propagation source, and configuration cycle number and write them into the edge configuration version. The static spatial reachability flag of the candidate edge is only used to form the candidate set. During inference, the aforementioned axial signed direction reachability, physical material mapping, stage boundary, and configuration version determination must still be performed at each sampling time. If any dynamic predicate is not satisfied, the corresponding edge is disabled. During model training, the edge weights and constrained time-delay residuals are updated based on the validation set quality prediction loss. After each training round, edges with normalized weights below 0.001 for five consecutive rounds and whose validation loss increment does not exceed 0.1% are removed. When validation losses are the same, the model version with the smallest absolute residual value is retained first. The final edge set, weights, residual parameters, and allowed bits for each stage are then determined. Figure 1 And it is fixed; during the inference phase, only the fixed version is read and no new cross-phase connections are added online.

[0022] In this specific embodiment, S4 includes: The temporal neural network reads the graph temporal features output by S3 and forms a sliding sequence of 512 sampling points according to the common sampling times of the same batch, material segment, and circumferential sector; in the first... At this moment, S3 first uses the data solidified in the previous moment. After completing the graph aggregation with the long-term thermal history state summary, S4 updates it with the effective original values ​​at this moment. and The updated results are only available for S5 and later versions. The graph is aggregated and read at each time point, and reverse recalculation at the same time point is prohibited to avoid circular dependencies. The input records include graph aggregation features, empty neighborhood flags, stage codes, flame temperature, pipe pressure, cooling status and data validity mask. When the sequence has less than 512 points, the mean of the same stage in the training set is added to the front end and the validity flag of the added position is set to zero. The network only outputs the state corresponding to the real sampling position. The causal dilated convolution uses four layers of one-dimensional convolution with a kernel length of 3 and dilation intervals of 1, 2, 4, and 16 respectively, and follows the... Extracting local impact and periodic features, among which, Indicates the first The layer in the first The dimensionless output vector at each sampling time. Represents the current sequence input vector. This represents the convolutional layer index and takes values ​​from 1 to 4. Represents the kernel position index and takes values ​​from 0 to 1. , This indicates the kernel length, which is 3. The training result is the first... Layer There are dimensionless convolution weight matrices. Indicates the first The layer expansion intervals are 1, 2, 4, and 16 in sequence. This represents the dimensionless bias vector obtained during training. This represents the ReLU activation function, where only the first term is referenced. Data from the previous sampling time; The convolution configuration table calculates the required number of sampling points based on the 95th percentile of the training batch rotation period and the 0.1 s sampling period, and sets the receptive field of the last convolutional layer to be no less than this number of sampling points. In this specific implementation, the effective training rotation period is 2 s to 4 s, the 95th percentile is 3.8 s, corresponding to 38 sampling points. When the four-layer dilated convolution uses dilation intervals of 1, 2, 4, and 16, the receptive field is 47 sampling points, thus covering the 40 sampling points required for the longest 4 s rotation period within the effective training range. If the median of the rotation period for ten consecutive periods exceeds the effective training range of 2 s to 4 s, an out-of-distribution flag is output and the current configuration is retained without changing the causal convolutional structure online. The thermal dose calculation module generates a four-dimensional recursive input at each common sampling time, according to... , and Calculate the duration of temperature exceedance, cumulative pressure, and cooling temperature change rate, among which, Indicates the first The integral increment of temperature exceeding the limit at each moment, in °C·s. and These represent the surface temperatures of the quartz tube measured at the current and previous moments, respectively, in °C. This indicates the softening temperature setting value read from the parameter table of the same batch of quartz materials; in this specific embodiment, it is 1660 ℃. This means selecting the largest value from the two inputs. Indicates the positive part of the temperature difference. This represents the cumulative pressure increment, expressed in kPa·s. This indicates the current pressure inside the pipe, in kPa. This represents the rate of temperature change within the cooling zone, expressed in °C / s, and is only valid during the cooling phase. This indicates a common sampling period of 0.1 s. Indicates the sampling time index; The same material segment in the first The circumferential temperature difference input at any given time is taken as the range of the effective temperatures of twelve circumferential sectors. If there are fewer than nine effective sectors, the filler value is not used and the effective flag of that component is set to zero. The heat dose calculation module generates separate values ​​for temperature exceeding the limit, circumferential temperature difference, pressure accumulation, and cooling temperature change rate. When the first When all required primitive values ​​of the class input are valid and the stage is applicable ,otherwise ; then according to Normalization, in which, Indicates the first Class input in the first The dimensionless value after validity gating at each moment. Indicates whether the 0 or 1 component is valid. This indicates the corresponding original unit input. and These represent the known stages in the training set. Next The mean and standard deviation of the class of valid inputs and the units are... Similarly, when the standard deviation is less than the device resolution, the device resolution is used instead. This represents four types of recursive input indices. Indicates the index of the processing stage. This indicates that the input will be truncated to... The calculation; the unknown stage uses independently versioned unknown stage statistics, and must not borrow the fill value from the previous known stage; According to the state of heat dosage Recursion, where, Indicates the first The dimensionless thermal dose state at a given moment. This indicates the dimensionless state of the circumferential sector state key within the same batch of physical materials at the previous moment. The state key is set to 0 when it first appears. Indicates the current processing stage The attenuation coefficient is calibrated and limited by historical qualified batches according to the minimum prediction error. , The aforementioned first The input is a valid 0 or 1 flag. Representation phase Next The dimensionless fusion weights of the normalized input are calibrated on the validation set by the heat dose parameter calibration module (described later) and satisfy the condition that the sum of the absolute values ​​of the four weights in the same stage equals 1. This indicates the first normalization result obtained according to the aforementioned normalization formula. Dimensionless input This represents the recursive input index from 1 to 4. Indicates the processing stage index output by S2; all 0 time command The co-occurrence state credibility flag is set to 0; when partially invalid, only the valid components participate in the summation and the credibility flag is set to low; the unknown stage uses a validation set calibrated by an independent unknown stage and fixed with the parameter version. and ,in, Represents the dimensionless decay coefficient of the unknown stage and is restricted to... , Indicates the unknown stage Dimensionless fusion weights Choose values ​​from 1 to 4, where the sum of the absolute values ​​of the four weights equals 1; freeze if there is no corresponding version with unknown stage parameters. Set the state trust flag to 0 and record the reason for the missing parameters. For cyclic gating and reading of the S3 state node in the next time step; The stage encoding table stores 16-dimensional stage embeddings using processing stage and procedure version as keys. For unknown stages, it stores independently trained unknown stage embeddings; if such an embedding does not exist, a fixed zero vector is used with a low-confidence flag. The temporal neural network concatenates the causal dilated convolution output, S3 graph aggregated node features, 16-dimensional stage embeddings, and heat dose status into the current input vector, and then... and Determine the threshold value, where, and They represent the first The input gate vector and forget gate vector at time n, and each component is located at... , This represents the current concatenated dimensionless input vector. This represents the dimensionless long-term thermal history state vector of the previous moment. , , and This represents the dimensionless gated weight matrix obtained during training. and This represents the dimensionless bias vector obtained during training. This indicates that each component is mapped to... The function; Cyclic unit preorder And only if the current time-mapping aggregation is non-empty, at least one thermal dose input is valid, and the phase is not an unknown low-confidence state, otherwise let Then according to Update the long-term hot history status; among which, Indicates the first The 128-dimensional dimensionless long-term thermal history state vector at each moment. The write gate represents a 0 or 1 state. and These represent the forget gate and input gate vectors, respectively. This represents the state vector at the previous moment. This represents the aforementioned dimensionless input vector after concatenation. This represents element-wise multiplication of vectors of the same dimension. and This represents the dimensionless candidate state weight matrix obtained during training. This represents the dimensionless candidate state bias vector obtained during training. This indicates that the candidate state components are restricted to... Hyperbolic tangent function; deterministic freezing when state write gate is zero The reason for the block is recorded, and the populated value does not generate a state update; The temporal neural network writes the long-term thermal history state into the state cache according to the batch identifier, physical material segment number, circumferential sector number, and common sampling time; when the state key of each batch-physical material segment-circumferential sector first appears, Take 0, Take the 128-dimensional initial vector obtained from training and fixed with the model version. If the initial vector is unavailable, take the zero vector. Regardless of whether the initial vector is available, unconditionally set the cold start flag to 1 and the continuous valid history count to 0. Only when 512 consecutive common sampling periods meet the following conditions: graph aggregation is non-empty, stage is known and not low confidence, at least three types of heat dose input are valid, material mapping is unique, and data validity predicate is met, and the continuous interval covers at least one complete valid rotation period, is the cold start flag cleared. The 512-point threshold, rotation period coverage threshold, and convolution configuration version are written together into the cold start configuration table. If the cold start is not lifted, S5 must at least set manual verification, prohibit the output of the reliable status, and prohibit the triggering of reliable alarm latches. When switching integration segments, the status must not be directly inherited by the new segment number. Instead, the batch identifier, material segment coordinate mapping, circumferential sector, and coordinate reference version saved by S1 are used to verify whether the physical material before and after the switch is the same. When the feed is paused or the process is switched but the material mapping is continuous, the state before the switch is frozen and the recursion continues after recovery. When the clamping direction is reversed, the last valid coordinate before the reversal and the new segment coordinate reference are used to establish a one-to-one mapping and inherit the corresponding physical material state. When the material segment is changed, the mapping is not unique, there is a cross-batch, the coordinate reference cannot be verified, or the model version or convolution configuration version is switched, a new state key is established and the cold start flag and count are reset. It is forbidden to inherit the state from different physical material segments or incompatible model versions. The cache simultaneously saves the stage code, integral segment identifier, material mapping version, heat dose state, convolution configuration version, threshold summary, empty neighborhood flag, out-of-distribution flag, component validity flag, cold start flag, continuous valid historical count and state blocking reason. When the input is invalid for 20 consecutive sampling cycles, the previous valid state is frozen, the state confidence flag is set to 0 and the cold start flag and count are reset. After recovery, the above 512-point release predicate is re-executed. The convolution configuration selection module forms a candidate set based on the sequence of convolution kernel length, number of layers, and expansion interval. First, configurations with a receptive field smaller than the 95th percentile sampling point number of one rotation cycle of the training batch or a single-window inference latency exceeding 20 ms are filtered out. Then, the configurations are sorted from smallest to largest according to the joint loss of the four quality tasks in the validation set, and the one with the smallest loss is selected. If the losses are the same, the one with fewer parameters is selected. The receptive field, inference latency, number of parameters, and model version of the selected configuration are written into the convolution configuration table. The normalization of thermal dose input is clearly centered on the mean of the same processing stage and input category in the training set. The original input is subtracted from the mean and then divided by the standard deviation of the same statistical set as the normalization denominator. The resulting dimensionless result is truncated to -5 to +5. The mean, standard deviation, device resolution, training data version and stage code are written into the dimensional statistics table. When the running data exceeds the truncated range, the truncated value is retained and the thermal dose anomaly flag is set to one. The thermal dose parameter calibration module forms candidate combinations of attenuation coefficients and four types of fusion weights at each stage on the validation set. It sorts these combinations according to the sum of the four types of standardized regression errors at the next quality detection time, and filters out combinations with attenuation coefficients that exceed the limit, weight absolute values ​​that are not equal to 1, or absolute values ​​of consecutive recursive states that exceed 20. It selects the feasible combination with the smallest error sum as the current version. If no feasible combination exists and a previously validated parameter exists, the new version release is stopped and the parameter is used. If there is no historical version for the first deployment, the neural network parameters are disabled, an independently validated process rule baseline is enabled, and only the "model not released" status, rule threshold alarms, and manual review requests are output. The state of no model must not be written as normal quality prediction. At least 200 valid labeled samples are added and the system is retrained before the release gate is executed.

[0023] In this specific embodiment, S5 includes: The multi-lead time prediction unit reads the 128-dimensional long-term thermal history states from the S4 state cache and establishes classification output heads and regression output heads for four types of target quality defects: uneven wall thickness, excessive ellipticity, insufficient welding, and residual stress. In this specific embodiment, four preset time intervals of 5 s, 10 s, 20 s, and 30 s from the quality inspection time are set. Each classification output head outputs the risk probability of the corresponding defect forming within the corresponding time interval, and each regression output head outputs the wall thickness deviation (mm), ellipticity deviation (mm), dimensionless welding quality characterization value, or residual stress (MPa), respectively. During operation, before calling each output head, a lead time availability mask is generated for each head according to the prediction time, target lead time, current process route, minimum available stage remaining time saved in the lead time configuration table, expected end time of adjustable stage, planned quality inspection coverage endpoint, and consistency between the procedure version and the lead time configuration version. Only when the prediction time plus the target lead time is not later than the available lead time can a lead time availability mask be generated. The output header is only made available when the expected end time of the adjustment phase and the planned quality inspection coverage end time, the remaining time of the current phase is not less than the corresponding minimum value, and the versions are consistent; the output header with a mask of 0 writes "Lead amount unavailable" and the reason, does not output risk probability, quality prediction value, neural network warning or key evidence, and participates in state merging according to the priority of "prohibit neural network prediction or warning" but does not suppress other available lead amount output headers; the available output header writes the batch identifier, material segment number, circumferential sector number, prediction time, lead amount, defect type, probability, prediction value, availability mask and model version into the prediction result record; The preset time interval is determined by historical quality detection data independent of the network to be trained. The label construction module reads the results of ultrasonic thickness measurement, roundness measurement, weld strength detection, and polarized stress detection, and maps them to material segments and circumferential sectors according to the detection location. Calculate the quality deviation, where, Indicates the first The absolute deviation of a quality indicator from its target value, with the same unit as the quality indicator. This refers to the measured values ​​of wall thickness, ellipticity, weld quality, or residual stress from historical inspection records. This indicates the corresponding quality target value given in the process specification, and the unit is... same, This represents an index of four categories of quality indicators; The label building module counts the adjustable processes that have not yet ended after each sampling time and their maximum correction amount from the historical qualified correction records of the same type of quartz tube, and uses them as... Determine the moment when the defect becomes irreversibly formed, among which, Indicates the first The earliest sampling time of irreversible formation of quality defects, in seconds. This indicates the candidate sampling time, in seconds. This indicates the aforementioned quality deviation. Indicates from time The remaining adjustable processes then apply to the first... The historical maximum correction value of the quality index and its unit is The same, obtained from the 95th percentile absolute correction of the qualified correction record. Indicates the permissible deviation of the quality acceptance limit and the unit is equal to The same and read from the process specification, This indicates that the first sampling time that meets the conditions is selected. If the conditions are not met throughout the entire process, the sample will not generate an irreversible defect label. The preset time interval is determined by the lead time for irreversible defect formation in historical quality inspection data: The tag construction module calculates the difference between the quality inspection time and the aforementioned irreversible defect formation time for each valid historical record. After removing outliers below the 1st percentile or above the 99th percentile, it summarizes the valid records of the four quality indicators and calculates the 20th, 40th, 60th, and 80th percentiles. To align with the production scheduling cycle, the four percentiles are rounded to the nearest 5-second grid and limited to the range of 5 to 30 seconds. If the rounded values ​​are repeated, a 5-second grid value with an increased value and a sample size of no less than 200 is used. In this specific implementation, the four percentiles are 6.2 seconds, 11.4 seconds, 18.6 seconds, and 28.7 seconds, respectively. According to the above determination rules, 5 seconds, 10 seconds, 20 seconds, and 30 seconds are obtained. Four target lead times are defined. The lead time configuration table stores historical data versions, percentile values, rounding rules, defect types, target lead times, number of forward points corresponding to the common sampling period, remaining duration of the minimum available stage, number of samples, process specification version, and effective batch for each item. When the remaining duration from a certain prediction time to the end of quality inspection or the current adjustable stage is less than the target lead time, the label mask of the target lead time is set to zero. Defect-lead time combinations with fewer than 200 samples use the previous verified configuration. When deploying for the first time without a historical configuration, the output header is not trained or published. The training data generation module generates supervisory labels corresponding to material segments, circumferential sectors, defect types, and measured quality values ​​based on the quality inspection location and inspection time; and predicts the time... Target lead time Passing the exam Class defects, if the moment of irreversible formation of the defect has been determined. and Then set the classification label of all target lead times to 1 and write it to the irreversible formation flag; if If the category label is 1 and the irreversible formation flag is 0; if , or up to It has complete testing and process traceability records and no first-order records have been formed within this period. If it is a defect, the category label is 0; if If the inspection exceeds the coverage area, crosses incomparable procedure versions, fails to inspect a material segment or sector, or cannot be determined at an irreversible time and lacks a complete record of defect-free items, the label mask will be 0 and no positive or negative label will be generated; among these, The predicted time is expressed in seconds. This indicates the target lead time of 5 s, 10 s, 20 s, or 30 s, in seconds. Indicates the defect type index. The aforementioned first The time of irreversible formation of this type of defect is measured in seconds. , and These represent greater than or equal to, less than, and less than or equal to relationships, respectively. The "irreversible formation" flag serves only as a supervisory label for the output header of the irreversible formation state (described later), and is not used as a direct trigger for runtime model input or online latching. Regression labels are generated only when the corresponding detection value is valid and the label mask is 1, and are associated with the same... , The material segment and sector are bound together by a joint key. After the training batch is grouped by production date and raw material batch, the training set, validation set and test set are divided in a 7:2:1 ratio. Network training adopts Jointly update parameters, where, Indicates dimensionless joint loss. This represents the dimensionless average of the binary cross-entropy of the four types of defects, calculated at positions where the label mask is 1. This represents the dimensionless average Huber loss, calculated by dividing each regression error by the standard deviation of the corresponding index in the training set. Represents the normalized time-delay offset over all valid edges. Calculate the average of the non-constantly zero soft constraint losses. This represents the actual time delay obtained from S3. This represents the nominal time delay obtained from S3, both in seconds (s). This indicates the allowable half-width of the residual obtained from S3, in seconds. This represents the numerical stability constant, which is 0.01s in this specific embodiment, and is used to make... Keep positive values. This represents the dimensionless hinge loss when the probability of adjacent lead time violates the relationship that the risk should not decrease when the proximity detection time is near. The non-negative dimensionless weights represent the classification loss. This represents the non-negative dimensionless weight of the regression loss. The non-negative dimensionless weights representing the time-delay soft constraint loss are: The non-negative dimensionless weights representing the lead consistency loss are all determined by a grid search of the validation set and sum to 1. In this specific implementation, , , and We take values ​​of 0.40, 0.30, 0.15, and 0.15 respectively; the feasible interval parameterization of S3 is responsible for the hard boundary. It is responsible for suppressing time delays from unjustifiably approaching the boundary and regularizing them towards the nominal value; The classification probability is calibrated on the validation set using isometric regression. The risk threshold for each type of defect and lead time is the probability value in the validation set that minimizes the false alarm rate while maintaining a recall rate of at least 0.90. When the risk probability is not lower than the threshold, a warning flag is written into the neural network; when it is lower, a normal flag is written into the neural network. The higher the probability, the higher the risk of the target defect forming. The prediction unit also establishes an irreversible formation state output head for each type of defect. It is trained using the aforementioned irreversible formation flag as a supervised label and its probability and threshold are calibrated on the validation set. During operation, this output head only reads the prediction time and the previous S4 state, the current regression prediction deviation, the quality allowable deviation given by the process specification, and the remaining maximum correction amount calculated based on the unfinished adjustable process. It does not read the post-detection results or the actual irreversible formation time. Only the corresponding lead time output head is available, the S4 cold start has been deactivated, and the current state is displayed. If the state is reliable and the probability of irreversible state formation is not lower than its verification threshold for three consecutive common sampling periods, and the current regression prediction deviation minus the remaining maximum correction amount is still greater than the quality allowable deviation, then the prediction service will latch the alarm formed by batch-material segment-circumferential sector-defect type joint key, and make this alarm take priority over the ordinary risk threshold state. The latch record is fixed with the trigger time, model version, material joint key, probability of irreversible state formation, regression prediction deviation, remaining maximum correction amount, quality allowable deviation and corresponding edge evidence snapshot, until the corresponding final inspection record confirms that it is qualified, it is released by authorized manual review or the batch ends, and it is closed. The person, time and basis for release are written into the audit record. The regression output is saved according to its respective engineering unit after denormalization. If the classification and regression conclusion conflict, the original output of both types is retained and the review flag is set to one without one output covering the other. For missing or anomalous sampling, the prediction unit defines the current material segment. Sensor set within the prediction window and common sampling time set Only when and At that time, according to Calculate the effective sampling ratio; where, Material segment The effective sampling ratio and its value , Indicates the generation of the first generation by S1 Type of sensor in the first A validity flag of 0 or 1 at any given time. and These represent the sensor and time indices in the aforementioned set, respectively. and Let represent the number of elements in the two non-empty sets, respectively. This represents the cardinality of a set, with the symbol "≥" indicating a greater than or equal to relation; it is not counted when any set is empty. Set the state confidence flag to 0 and record the reason for the empty sensor configuration set or prediction window set; the 5th to 95th percentiles of the effective proportion of training samples constitute the confidence interval. The neural network output state is merged according to the priority of "prohibit neural network prediction or warning, manual review, low confidence, and confidence": continuous invalidity, empty set or no published model, prohibit writing to neural network risk probability, quality prediction value and neural network warning; unknown stage, low confidence stage, empty neighborhood of key target, out of distribution, abnormal heat dose, cold start or classification and regression conflict, at least manual review is required; when the effective proportion is out of bounds or some inputs are invalid, set to low confidence; only when none of the above anomalies exist, When the S4 state is within the trusted range and the S4 state is trusted, the trusted flag 1 is output. The rule baseline alarm field is independent of the neural network output state. When there is no published model, the verified process rule is still called to calculate the rule threshold alarm and the manual review field is set to 1. The result record saves the neural network output state, rule alarm state, triggering reason and release condition. The key evidence reverse mapping module first forms a map for each valid edge. Then according to Calculate the edge-level contribution; where, Indicates the first The edge at the 1st The dimensionless message vector that enters the target node aggregation at each time point. , , , , and These represent the stage mask, hysteresis validity flag, edge weight, edge transformation matrix, time delay alignment source feature, and actual time delay as defined in S3. This indicates the non-negative dimensionless contribution of the edge to the current output. Represents the message vector of the first generation. One portion, Indicates the message component index. This indicates the risk probability or standardized quality prediction value of the current defect type and lead time. This indicates the gradient of the output message component on that side. This represents absolute value operations and summation iteration. All components; invalid edges and All are 0; The evidence reverse mapping module first processes each valid edge according to... Determine the time of source action, where, Indicates the first The edge at the 1st The source action time corresponding to each evidence sampling time is expressed in seconds. Indicates the first A common sampling time, measured in seconds. This indicates the actual time delay of the edge stored in S3, in seconds. The module retains the edge identifier, source node, target node, material segment, circumferential sector, stage, integration segment, edge configuration version, time delay version, and source action time for each edge, and quantizes the source action time into a source time bucket based on a common sampling period width. Subsequently, only edges with identical source node, material segment, circumferential sector, stage, integration segment, configuration version, and source time bucket are included in the set. ,according to The node-source time contribution is formed, where, Indicates the source node Falling into The set of valid edge-sampling time pairs with consistent source time buckets and joint keys. This represents the non-negative dimensionless aggregation contribution under the joint bond. Indicates the source node index. Indicates the source time bucket index. This indicates the aforementioned edge-level contribution; summation only iterates through the edges. Summing across edges with different actual time delays is prohibited before the source action time is shifted back; the evidence threshold is taken as the 95th percentile of the contribution distribution of training sample nodes-source time with the same defect type and the same lead amount. Adjacent source time buckets that reach the threshold are merged into a time window when the interval does not exceed two common sampling periods. The joint key, related edge identifier, edge-by-edge contribution, aggregate contribution, source record and time delay version are recorded and saved. If the threshold is not reached, the "no key evidence" flag is output. The prediction service writes the available risk probabilities, quality prediction values ​​for each engineering unit, risk thresholds, neural network output status, rule baseline alarms, formed latched alarms, credibility flags, verification flags, formation time windows, and corresponding process parameters into the quality prediction result table, and pushes them to the production early warning interface and final inspection traceability interface according to material segment and circumferential sector. The interface displays the neural network result area, rule alarm area, and manual verification area respectively. When there is no published model, the neural network result area displays "not published" and does not display probabilities or prediction values. The rule baseline configuration table uses process route, processing stage, original quantity, and rule version as joint keys to store trigger thresholds, release thresholds, judgment direction, number of consecutive trigger cycles, number of consecutive release cycles, original quantity validity requirements, parameter source, and effective batch. The release threshold for upper limit rules is strictly lower than The trigger threshold and the release threshold of the lower limit rule are strictly higher than the trigger threshold. The two form a hysteresis interval that has been independently verified and determined. In this specific implementation, the number of consecutive release cycles is ten common sampling cycles. When multiple rules are triggered at the same time, they are displayed in descending order of security level in the configuration table. If any high-level rule is not released, the rule alarm is maintained. When switching rule versions, the latched alarm continues to be judged according to the old version at the time of triggering until it is released or manually confirmed. The new version is used only when the new trigger quantity is new. The rule alarm area displays the rule version, trigger threshold, release threshold, trigger quantity, direction, and trigger reason. It is only released when the relevant original quantity is valid and meets the corresponding release direction and continuously reaches the configured number of release cycles, or after authorized manual confirmation. The model package saves the training data version, advance configuration version, stage procedure version, parameter checksum, and effective batch. The model release module filters candidate models based on the following criteria: test set classification recall rate not less than 0.90, classification calibration error not exceeding 0.05, standardized regression mean absolute error not exceeding 0.20, and single-material segment inference latency not exceeding 50 ms. Recall rate is a benefit-based threshold used only for qualified selection; calibration error, regression error, inference latency, and joint loss are all cost-based indicators, with lower values ​​being better. Models that pass the thresholds are sorted in ascending order of joint loss for reliable samples, and the first model with a memory usage not exceeding 256 MB is selected. For the MB version, when the joint loss is the same, the one with the smallest sum of absolute values ​​of time-delay residuals is selected. A validation cycle in online monitoring is defined as twenty consecutive production batches that have ended and all final inspection labels have arrived after the same model version has taken effect. Within the cycle, there must be no fewer than 200 material segment samples with a label mask of 1, no fewer than 20 positive samples for each type of defect, and the prediction record and final inspection label must be bound together by batch, material segment, circumferential sector, defect type, and lead time. If the number of samples is less than the above, the cycle is not closed and the number of violations is not accumulated. The denominator of the recall rate is the number of valid positive samples of that type. Calibration error and regression error are calculated only for samples with a corresponding label mask of 1. At the end of the cycle, the model version, sample list, indicator denominator, evaluation time point, and label data version are written into the validation record. If any release threshold is violated in two consecutive closed validation cycles, a rollback candidate is selected before the start of the next batch, and the rollback is performed. Before rollback, verify the compatibility of the model package hash, input sensor mode, stage procedure version, convolution configuration, and lead time configuration. Then, verify the candidate model against the two most recent closed cycles to ensure it still meets the release threshold. Reset the continuous violation count when any cycle passes all thresholds, a new model version takes effect, or the label data version changes. If a qualified previous version exists, switch at the batch boundary using an atomic version pointer and write it to the audit log. If the previous version is missing, package verification fails, is incompatible, or the verification still exceeds the limit, the model must not take effect and immediately enter the rule baseline mode, only outputting rule alerts and manual review requests and writing the model status as "no qualified rollback version". If the candidate is empty and a continuously qualified current version exists, do not release a new version and continue to output the current version, version status, and trust flag. The model is also prohibited from taking effect during the first deployment or when the version library is empty. Only after supplementing samples and passing all release gates can the neural network prediction be output.

[0024] 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.

[0025] This invention uses the material space synchronization results and stage identification results to define the organization of temporal features, so that the temperature, motion, pressure and cooling information on the equipment time axis are converted into the processing history corresponding to the quartz tube material segment and circumferential sector; the variable time delay diagram further completes the time alignment according to the physical propagation source of the connecting edge, so the prediction results can keep in correspondence with the actual material action position and action time.

[0026] This invention addresses the issue of hot processing of quartz tubes by modifying the internal structure of the time-series model: it utilizes stage mask switching graph connections, allowing thermal dose states to simultaneously participate in graph node representation and cyclic gating, and uses long-term thermal history states for multi-task, multi-advance output, and evidence backtracking. This structure closes stage constraints, thermal accumulation memory, and quality localization within the same processing chain, thereby supporting pre-defect warning and post-processing traceability.

Claims

1. A method for predicting the heat treatment quality of quartz tubes based on temporal neural networks, characterized in that, include: S1. Acquire multi-source process timing data and motion event data, integrate the feed speed and rotation speed piecewise, determine the material axial coordinate and circumferential phase, and map the multi-source process timing data and motion event data to the material segment, circumferential sector and sampling time to obtain material spatial timing data; S2. Based on the process status data in the action event data and material space time series data, identify the processing stage under the constraints of process transfer relationship and stage dwell time, obtain the stage identification result and stage mask, and organize the material space time series data according to material segment and processing stage to obtain dual-index time series data. S3. Construct a variable time delay graph with stage constraints based on dual-index time series data and stage mask. Determine the nominal time delay of the connecting edge according to the distance between the acquisition location of the process variable and the location of the processed material, the material movement speed, the rotation cycle and the stage dwell time. Learn the time delay residual within the allowable interval and switch the connecting edge according to the processing stage. Take the sum of the nominal time delay and the time delay residual with interval constraints as the actual time delay of the directed edge to obtain the graph time series characteristics. During graph feature aggregation, the graph construction module follows the procedure as follows: ； Generate target node features; where, Represents the target node In the Dimensionless plot temporal feature vector at each time step Indicates the target node index. Indicates the source node index. Indicates that it is from the source node Point to target node The set of directed edges, This represents the 0 or 1 stage mask generated by S2. Indicates the actual source time The binding 0 or 1 lag validity flag is set to 1 only if both interpolation endpoints surrounding the source time are valid and belong to the same stage, the same integration segment, and the same side configuration version; otherwise, it is set to 0 and interpolation is disabled. This represents the softmax-normalized edge weights obtained during training and among the valid incoming edges to the same target node. This represents the dimensionless edge-type linear transformation matrix obtained through training. This represents the dimensionless eigenvector of the source node obtained by linear interpolation at the actual source time. Indicates the first A common sampling time, measured in seconds. This indicates the actual sampling delay, expressed in seconds (s). Represents the ReLU activation function; S4. Input the graph time-series features into the time-series neural network, extract local impact features and circumferential periodic features through causal dilated convolution, recursively deduce the heat dose state, and control the update of the loop unit with the processing stage and heat dose state to obtain the long-term thermal history state. S5. Input the long-term thermal history status into the multi-advance prediction unit and output the risk probability or quality prediction value of the target quality defect.

2. The method for predicting the heat treatment quality of quartz tubes based on a time-series neural network according to claim 1, characterized in that, S1 includes: The feed rate is accumulated according to the sampling time interval to obtain the axial displacement of the material, and the rotational rate is accumulated according to the sampling time interval and converted into the circumferential phase by one rotation cycle. The current integration segment is terminated and the next integration segment is established at the action events corresponding to the clamping direction reversal, feed pause, or process switching. The start and end times, movement direction, and coordinate reference of each integration segment are recorded. The material segment is determined according to the preset axial interval to which the material axial coordinate belongs, and the circumferential sector is determined according to the preset phase interval to which the circumferential phase belongs. Resampling is performed at the preset common sampling time according to the timestamp of each process data. For continuously changing quantities, adjacent valid sample values ​​are interpolated, and for held state quantities, the previous valid sample value is used for holding. When the time difference between the most recent valid sample value and the common sampling time exceeds the allowable time deviation determined by the sampling period of the corresponding sensor, the corresponding data of the common sampling time is marked as invalid.

3. The method for predicting the heat treatment quality of quartz tubes based on a time-series neural network according to claim 1, characterized in that, S2 includes: Based on the hot working process specification, a set of process states including drawing, flaring, necking, welding, annealing, and cooling stages are established, along with process transfer relationships that allow transitions between adjacent process states. Candidate stage boundaries are formed based on the trigger times of action events, and the stage dwell time range is determined based on the statistical interval of the duration of each process state in historical qualified processing records. The process state observation model corresponding to each processing stage calculates the matching probability of each sampling time belonging to the corresponding processing stage based on flame working state, rotation state, feed state, tube pressure state, and cooling state. The logarithm of the matching probabilities contained in the candidate stage sequence is taken and accumulated to obtain a sequence score. The candidate stage sequence that satisfies the process transfer relationship and the stage dwell time range and has the highest sequence score is selected as the stage identification result, and the graph connection edges allowed for the corresponding processing stage are marked as valid to form the stage mask. When there is no candidate stage sequence that satisfies the process transfer relationship and the stage dwell time range, the corresponding sampling interval is marked as an unknown stage to generate the stage identification result, and the stage-related connection edges of the corresponding sampling interval are marked as invalid in the stage mask.

4. The method for predicting the hot working quality of quartz tubes based on a time-series neural network according to claim 2, characterized in that, S3 include: Process variable nodes are established for flame temperature, heating position, rotation speed, feed rate, tube pressure, and cooling parameters, respectively. Quality nodes are established for the target quality defect. Directed edges from process variable nodes to quality nodes and directed edges between process variable nodes characterizing the coupling relationship of process variables are established. The stage mask controls whether the directed edges participate in graph feature aggregation. For directed edges propagating along the material's axial movement, the propagation time is determined within the same integration segment by the axial distance from the corresponding acquisition position to the material segment position and the material's movement speed. This propagation time is considered during feed pauses or material... When the absolute value of the motion speed is less than the minimum stable feed speed of the equipment, the corresponding axial propagation directed edge is disabled, and the propagation of the directed edge across the integral section is prohibited when the clamping direction is reversed or the process is switched. For the directed edge that repeatedly acts with circumferential rotation, the phase propagation time is determined by the circumferential phase difference and the rotation period, and the nominal time delay is determined by combining the stage dwell time. The residual allowable interval of each directed edge is determined according to the equipment calibration error, speed fluctuation range and stage boundary error. The sum of the nominal time delay and the time delay residual constrained by the interval is taken as the actual time delay of the directed edge, and the weight of the directed edge is learned.

5. The method for predicting the heat treatment quality of quartz tubes based on a time-series neural network according to claim 1, characterized in that, The thermal dose status in step S4 is generated recursively according to the sampling time. The recursive input of the thermal dose status includes the duration of the temperature exceeding the set value of the quartz tube softening temperature, the range or variance of the temperature of each circumferential sector within the same material segment, the cumulative amount of pressure inside the tube with respect to the sampling time, and the rate of change of temperature with respect to time within the cooling zone. Each recursive input is normalized and weighted according to the dimensional statistics of the training data. The fusion result is then recursively combined with the thermal dose state at the previous sampling time according to the attenuation coefficient corresponding to the processing stage to obtain the thermal dose state at the current sampling time. The corresponding processing stage embedding is obtained by querying the stage coding table based on the stage identification result. The heat dose state is used as the state node input of the variable time delay diagram. The heat dose state and the processing stage embedding are used together to determine the input gate and forget gate of the loop unit.

6. The method for predicting the hot working quality of quartz tubes based on a time-series neural network according to claim 5, characterized in that, The causal dilated convolution employs convolutional connections that reference only the data from the current sampling time and previous sampling times, sets an increasing dilation interval along the time dimension, and ensures that at least one convolutional receptive field covers the number of sampling points corresponding to one rotation cycle. The output of the causal dilated convolution, the node features aggregated by the variable time delay graph, the processing stage embedding, and the thermal dose state are input into the loop unit. The input threshold controls the current input write amount, and the forget threshold controls the amount of the previous long-term thermal history state retained, so as to update the long-term thermal history state.

7. The method for predicting the hot working quality of quartz tubes based on a time-series neural network according to claim 6, characterized in that, S5 includes: setting quality classification tasks and quality regression tasks for uneven wall thickness, excessive ellipticity, insufficient welding, and residual stress, respectively, and setting prediction outputs for multiple preset time intervals before each quality inspection time; outputting the risk probability of the corresponding target quality defect forming within each preset time interval through the quality classification task, and outputting wall thickness deviation, ellipticity, welding quality characterization value, or residual stress characterization value through the quality regression task; obtaining wall thickness deviation, ellipticity, welding quality characterization value, or residual stress characterization value associated with the material segment and circumferential sector as target quality indicators from historical quality inspection data independent of the temporal neural network to be trained; calculating the quality deviation of the target quality indicator relative to the quality target value; determining the maximum correction amount of the remaining adjustable process after each sampling time based on historical qualified correction records; determining the first sampling time when the absolute value of the quality deviation minus the maximum correction amount is still greater than the corresponding quality qualified limit allowable deviation as the time of irreversible defect formation; and determining the preset time interval based on the historical duration statistics between the quality inspection time and the time of irreversible defect formation.

8. The method for predicting the heat treatment quality of quartz tubes based on a time-series neural network according to claim 7, characterized in that, It also includes the reverse mapping of key evidence: based on the sensitivity of the prediction results to the long-term thermal history state and the connection edge weights determined by training in the variable time delay graph, the contribution value of each graph node to the prediction results is determined. Based on the sum of the nominal time delay of the connecting edge and the time delay residual constrained by the allowable interval, the sampling time corresponding to the contribution value is shifted back to the time of action of the process variable. The position of action of the process variable and the processing stage are determined according to the material segment identifier, circumferential sector and stage identification results. The continuous sampling intervals where the contribution value reaches the evidence threshold determined by the contribution value of the training sample and the preset quantile level are merged into a time window, and the process parameters corresponding to the continuous sampling interval are output.

9. The method for predicting the heat treatment quality of quartz tubes based on a time-series neural network according to claim 7, characterized in that, The training of the temporal neural network includes: generating supervision labels corresponding to material segments, circumferential sectors, and defect types based on the detection location and detection time of the quality inspection data; shifting the supervision labels to the corresponding sampling time according to each preset time interval to form multi-lead training labels; using a label mask that does not calculate the loss for undetected material segments and undetected circumferential sectors, jointly calculating the quality classification loss, quality regression loss, time-delay residual interval constraint loss, and adjacent lead prediction consistency loss, and updating the parameters of the connection edge weights, time-delay residuals, causal dilated convolutions, recurrent units, and multi-lead prediction units of the variable time-delay graph based on the quality classification loss, quality regression loss, time-delay residual interval constraint loss, and adjacent lead prediction consistency loss.

10. The method for predicting the heat treatment quality of quartz tubes based on a time-series neural network according to claim 9, characterized in that, When there are missing or abnormal samples in the multi-source process time series data, a data validity mask is generated based on the valid sensor sampling flag and equipment action event. The graph node input corresponding to the invalid data is set to the preset fill value and the corresponding connection edge is prohibited from participating in graph feature aggregation. The credibility flag of the prediction result is determined based on the effective sampling ratio of the current material segment and the effective sampling ratio range in the training samples, and the credibility flag is output together with the risk probability and quality prediction value.

Images