A multi-parameter intelligent sensing method and system for quartz tube melting

Abstract

The application discloses a quartz tube melting multi-parameter intelligent sensing method and system, and belongs to the field of quartz tube melting intelligent sensing. In order to solve the problems that multi-source sensing signals lack unified time reference and correlation calibration, and it is difficult to identify multivariable coupling abnormalities in time, the application realizes the technical effect of layered discrimination of working condition instability and sensing abnormalities through unified time base acquisition, calibration compensation, time alignment, dynamic coupling graph construction and graph time sequence abnormality identification.

Description

Technical Field

[0001] This invention relates to the field of intelligent sensing in quartz tube melting, and more particularly to a multi-parameter intelligent sensing method and system for quartz tube melting. Background Technology

[0002] The quartz tube melting process typically requires simultaneous monitoring of multiple process parameters, including temperature, furnace pressure, gas flow rate, traction speed, and online tube diameter. Due to the presence of high-temperature radiation, dust, electromagnetic interference, and equipment vibration on-site, different sensors exhibit variations in sampling frequency, response time, range boundaries, and temperature drift characteristics. Consequently, the dispersed measurement results cannot directly form a unified basis for judging the operating conditions.

[0003] Existing measurement methods mostly rely on fixed threshold alarms for single parameters, lacking unified time-base processing, response delay compensation, temperature drift correction, and reliability assessment for multi-source signals. They also struggle to express the coupling relationships between temperature and furnace pressure, gas flow rate and online pipe diameter, and traction speed and online pipe diameter. Therefore, operational instability is often difficult to identify in a timely manner before quality defects such as bubbles, streaks, or wall thickness deviations form.

[0004] Therefore, there is a need for a multi-parameter intelligent sensing method and system for quartz tube melting that can overcome the shortcomings of the existing technology. Summary of the Invention

[0005] One objective of this invention is to propose a multi-parameter intelligent sensing method and system for quartz tube melting. Addressing the problems of existing technologies lacking a unified time base and associated calibration, and the difficulty in timely identifying multivariate coupling anomalies, this invention proposes a technical solution for unified time base acquisition, calibration compensation, time alignment, dynamic coupling graph construction, and graph timing anomaly identification of multi-source sensing sequences. This invention has the technical effect of distinguishing between sensing anomalies and process coupling anomalies and outputting the anomaly type.

[0006] This invention provides a multi-parameter intelligent sensing method for quartz tube melting, comprising: S1, collecting temperature, furnace pressure, gas flow rate, traction speed, online tube diameter, and sensor self-diagnostic status at the quartz tube melting site, and writing a unified time base timestamp, original sampling time, sampling period, response time, range, ambient temperature, vibration status, and electromagnetic interference status for each data acquisition path, thereby obtaining a multi-source original sequence and a set of status attributes; S2, performing response delay compensation, temperature drift correction, and range reliability judgment based on the multi-source original sequence and the set of status attributes, generating a calibration sequence and a reliability matrix; S3, performing calibration on the calibration sequence according to a unified event time window. Alignment is performed to generate a missing mask matrix, and the asynchronous segments are filled in by combining the missing mask matrix and the confidence matrix to generate an aligned time series matrix; S4, a dynamic coupling graph is generated with temperature, furnace pressure, gas flow rate, traction speed and online pipe diameter as nodes, according to the preset process coupling relationship and the deviation correlation relationship between real-time parameters calculated by the aligned time series matrix; S5, the aligned time series matrix and the dynamic coupling graph are input into the graph time series model, time window encoding is performed in the nodes and deviation information is transmitted along the dynamic edge to obtain single parameter residuals, parameter pair residuals and operating condition anomaly scores, and the anomaly type is output according to the hierarchical calibration threshold determined by the operating condition layer.

[0007] Optionally, S1 includes:

[0008] Temperature, furnace pressure, gas flow rate, traction speed and online pipe diameter are used as sensing channels respectively. The device number, installation location, upper limit of range, lower limit of range, calibration time, sampling period and response time of the sensing channels are written into the status attribute set.

[0009] Ambient temperature, vibration status, and electromagnetic interference status are associated with the same event time window using a unified time base timestamp.

[0010] The sensor's self-diagnostic status is converted into a communication status marker, a range status marker, and a calibration status marker, which, together with the sampled values ​​of the corresponding sensing channels, form a multi-source raw sequence.

[0011] Optionally, S2 includes:

[0012] The response delay compensation amount is determined based on the slope of the sampled value change and the response time of the sensing channel, and the sampled value corresponding to the original sampling time is corrected according to the response delay compensation amount.

[0013] The temperature drift correction amount is determined based on the ambient temperature and the calibration deviation of the same sensing channel within the historical stable window. The historical stable window is the time window of an event marked as a stable operating condition in the calibration stage or in historical production records. The temperature drift correction amount is used to correct the sampled value after response delay compensation.

[0014] The reliability coefficients are determined based on the distance between the sampled value and the range boundary, the communication status marker, the range status marker, the calibration status marker, the vibration status, and the electromagnetic interference status. The reliability matrix is ​​composed of the reliability coefficients of each sensing channel.

[0015] Furthermore, the reliability coefficient includes a basic reliability value and a synchronization correction value. The basic reliability value is determined based on the distance between the sampled value and the range boundary, the communication status marker, the range status marker, the calibration status marker, the vibration status, and the electromagnetic interference status. The synchronization correction value is determined based on the synchronization deviation, which is the time deviation of the sampling time after the response delay compensation of the sensing channel relative to the unified event time window alignment index. The unified event time window alignment index is an index established according to the window start point and window length of the unified event time window.

[0016] When the reliability coefficient of any sensing channel determined by the distance between the sampled value and the range boundary, communication status flag, range status flag, calibration status flag, vibration status, electromagnetic interference status, and synchronization deviation is less than the preset reliability threshold, a status flag is generated according to the factors that cause the reliability coefficient to be less than the preset reliability threshold. Specifically, a time reference anomaly flag is generated when the synchronization deviation exceeds the preset synchronization deviation threshold, a sensor drift anomaly flag is generated when the temperature drift correction exceeds the preset drift threshold or the calibration status flag is abnormal, and a data reliability anomaly flag is generated when the distance between the sampled value and the range boundary is less than the preset range distance threshold, the vibration status is abnormal, or the electromagnetic interference status is abnormal. The generated status flags are then written into the status attribute set.

[0017] Furthermore, the stratified calibration threshold is determined by acquiring and associating the product specifications, melting stage, and furnace temperature range of each event time window as stratification labels, and grouping historical stable windows according to the stratification labels. The historical stable windows are event time windows marked as stable operating conditions in the calibration stage or historical production records.

[0018] For each group of historical stable windows, calculate the single-parameter inconsistency score based on the single-parameter residuals, and calculate the parameter pair inconsistency score based on the parameter pair residuals.

[0019] The coverage rate preset by group is used as the target coverage rate. Under the target coverage rate, the quantile values ​​of the single-parameter inconsistency score and the quantile values ​​of the parameter-to-inconsistency score are determined respectively.

[0020] Map the quantile values ​​of single-parameter inconsistency scores to single-parameter thresholds, and map the quantile values ​​of parameter-pair inconsistency scores to coupled edge thresholds;

[0021] The current event time window calls the corresponding single-parameter threshold and coupling edge threshold according to the same hierarchical label, compares the single-parameter residual of the current window with the single-parameter threshold, and compares the parameter pair residual of the current window with the coupling edge threshold to obtain the anomaly discrimination value;

[0022] When the single-parameter residual exceeds the corresponding single-parameter threshold, the anomaly type includes the single-parameter anomaly of the corresponding sensing channel. When the parameter pair residual exceeds the corresponding coupling edge threshold, the anomaly type includes the process coupling anomaly of the corresponding process prior edge. When there are time base anomaly markers, sensor drift anomaly markers, or data reliability anomaly markers in the state attribute set, and only the parameter pair residuals of the marked sensing channel exceed the corresponding coupling edge threshold, the anomaly type is the anomaly represented by the corresponding state marker. When multiple anomaly discrimination conditions are met simultaneously, a composite anomaly type is output. When neither the single-parameter residual nor the parameter pair residual exceeds the corresponding threshold and the state marker does not exist in the state attribute set, the anomaly type is empty.

[0023] Furthermore, it also includes a model update step: when the anomaly type of the event time window is empty, the single parameter residual is not greater than the corresponding single parameter threshold, and the parameter pair residual is not greater than the corresponding coupling edge threshold, the event time window is determined to be a stable operating condition and written into the stable sample library.

[0024] When the number of newly written samples in the stable sample library reaches the preset update number, the stable sample library is used to redetermine the hierarchical calibration threshold, while keeping the node input, edge input, and output types of the graph time series model unchanged;

[0025] If the redefined stratified calibration threshold causes the number of abnormal outputs in the historical stable window to exceed the preset number of reviews, the stratified calibration threshold before the update is retained.

[0026] Optionally, S3 includes:

[0027] An alignment index is established using the window start point and window length of a unified event time window, and the calibration sequence is mapped to the alignment index.

[0028] When any sensor channel lacks a sampled value within the target window, the missing position is recorded in the missing mask matrix, and the filler value is generated using the calibration values ​​of adjacent windows, the calibration values ​​of other sensor channels within the same window, and the corresponding elements in the confidence matrix.

[0029] The original calibration values ​​and the padding values ​​are arranged according to the alignment index to form an alignment timing matrix.

[0030] Optionally, S4 includes:

[0031] Set temperature and furnace pressure, gas flow rate and online pipe diameter, traction speed and online pipe diameter, and temperature and online pipe diameter as process priors respectively;

[0032] For each process prior edge, the hysteresis correlation of the nodes at both ends of the edge in the current window and the historical stable window is calculated according to the alignment time matrix. The historical stable window is the event time window marked as a stable condition in the calibration stage or in the historical production record. The historical stable window benchmark value is the statistical benchmark value of the corresponding node calibration value.

[0033] Within a scrolling window containing the current event time window, calculate the deviation covariance based on the deviation sequence of the nodes at both ends of the edge relative to the historical stable window benchmark value, and calculate the confidence product of the nodes at both ends of the edge.

[0034] The corresponding edge weights are updated based on the product of the hysteresis correlation, the deviation covariance, and the confidence level, and a dynamic coupled graph is formed by the nodes, process prior edges, and the updated edge weights.

[0035] Optionally, S5 includes:

[0036] The graph time series model uses the time window of the event to be discriminated as the target event time window, and performs time window encoding on the sequence of each node in the aligned time series matrix to obtain the node time series representation.

[0037] The coupling representation is obtained by transmitting node temporal representation and deviation information between adjacent nodes based on the edge weights of the dynamic coupling graph;

[0038] Based on the coupling characterization, predict the node parameters of the target event time window, and reconstruct the parameter edge deviation between the two ends of the same process prior edge in the same target event time window;

[0039] The measured edge deviation is determined based on the difference in measured parameters between the two nodes of the corresponding process prior edge within the target event time window.

[0040] The difference between the predicted node parameters and the measured node parameters of the target event time window is determined as the single-parameter residual. The difference between the reconstructed parameter pair deviation and the measured edge deviation corresponding to the same target event time window and the same process prior edge is determined as the parameter pair residual. The operating condition anomaly score is generated from the single-parameter residual and the parameter pair residual.

[0041] On the other hand, the present invention also provides a multi-parameter intelligent sensing system for quartz tube melting, comprising:

[0042] The data acquisition module is used to collect multi-source raw sequences from the quartz tube melting site and write them into the state attribute set; the calibration compensation module is used to perform response delay compensation, temperature drift correction, and range reliability judgment, and generate calibration sequences and reliability matrices; the time alignment module is used to generate missing mask matrices and alignment time series matrices according to a unified event time window; the dynamic graph construction module is used to generate dynamic coupling graphs based on preset process coupling relationships and real-time parameter deviation correlations; the anomaly identification module is used to output single-parameter residuals, parameter pair residuals, operating condition anomaly scores, and anomaly types based on the alignment time series matrix, dynamic coupling graphs, and hierarchical calibration thresholds determined by operating condition stratification.

[0043] The beneficial effects of this invention are:

[0044] 1. This invention enables data from different types of sensors to be calibrated and compared within the same event time window by writing a unified time base timestamp and state attribute set to multi-source acquired data, and by performing response delay compensation, temperature drift correction and range reliability judgment.

[0045] 2. This invention processes different sampling frequencies, asynchronous sampling, and missing segments through a missing mask matrix, a confidence matrix, and an alignment time sequence matrix, enabling subsequent dynamic graph construction and anomaly identification to be based on consistent time sequence input.

[0046] 3. This invention identifies single-parameter anomalies, process coupling anomalies, and state marking anomalies by using process prior edges, real-time parameter deviation correlations, and hierarchical calibration thresholds, which helps to detect instability trends in operating conditions before quality defects are formed. Attached Figure Description

[0047] 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:

[0048] Figure 1 This is a flowchart of a multi-parameter intelligent sensing method for quartz tube melting.

[0049] Figure 2 This is a flowchart of step S5 of the present invention, which describes the timing anomaly identification and hierarchical threshold discrimination. Detailed Implementation

[0050] 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.

[0051] refer to Figures 1-2A multi-parameter intelligent sensing method for quartz tube melting includes: S1, collecting temperature, furnace pressure, gas flow rate, traction speed, online tube diameter, and sensor self-diagnostic status at the quartz tube melting site, and writing a unified time base timestamp, original sampling time, sampling period, response time, range, ambient temperature, vibration status, and electromagnetic interference status for each data acquisition path to obtain a multi-source original sequence and a set of status attributes; S2, performing response delay compensation, temperature drift correction, and range reliability judgment based on the multi-source original sequence and the set of status attributes to generate a calibration sequence and a reliability matrix; S3, aligning the calibration sequence according to a unified event time window. S4. A missing mask matrix is ​​generated, and asynchronous segments are filled in by combining the missing mask matrix and the confidence matrix to generate an aligned timing matrix; S5. A dynamic coupling graph is generated with temperature, furnace pressure, gas flow rate, traction speed and online pipe diameter as nodes, according to the preset process coupling relationship and the deviation correlation relationship between real-time parameters calculated by the aligned timing matrix; S6. The aligned timing matrix and the dynamic coupling graph are input into the graph timing model, time window encoding is performed in the nodes and deviation information is transmitted along the dynamic edge to obtain single parameter residuals, parameter pair residuals and operating condition anomaly scores, and the anomaly type is output according to the hierarchical calibration threshold determined by the operating condition layer.

[0052] In this specific embodiment, S1 includes:

[0053] A data acquisition module runs on the industrial computer of the quartz tube melting production line. The system clock of the industrial computer is synchronized with the main clock of the production line through IEEE1588PTP to form a unified time base. The time resolution of the unified time base is set to 1ms and used to write the same source timestamp to all acquired data.

[0054] Temperature, furnace pressure, gas flow rate, traction speed, and online pipe diameter are defined as five sensing channels, respectively denoted as temperature channel, furnace pressure channel, flow rate channel, traction channel, and pipe diameter channel. Each of the five sensing channels is bound to a unique device number and installation location code and is fixedly configured within the system.

[0055] The data acquisition module establishes a status attribute set for each sensing channel. The status attribute set writes and saves the device number, installation location code, upper range limit, lower range limit, calibration time, sampling period and response time for each sensing channel. The upper and lower range limits are taken from the sensor nameplate or calibration file of the channel, the calibration time is taken from the most recent valid calibration record of the channel, the sampling period is taken from the acquisition task scheduling period of the channel, and the response time is taken from the sensor technical specifications of the channel and used as a fixed parameter for subsequent compensation.

[0056] When any sensing channel generates a sample, the data acquisition module writes the sampled value, along with the unified time base timestamp and the original sampling time, into the multi-source raw sequence. The unified time base timestamp is defined as the time value generated by the PTP synchronization system clock when the acquisition module receives and completes the verification of a frame of data. The original sampling time is defined as the internal sampling time field carried in the sensor message and corresponds one-to-one with the sampled value. When the sensor message does not carry the internal sampling time field, the original sampling time is set to the unified time base timestamp to ensure that no items are missing in subsequent timing processing.

[0057] The data acquisition module maps a unified time base timestamp to a unified event time window to complete cross-channel archiving. The event time window length is set to [value missing]. And the window start point is set to It also retrieves the unified time base timestamp of the moment when the data acquisition module starts and completes clock synchronization, and the window index corresponding to any sample. according to Calculation, where This represents the event time window index and is a non-negative integer. This indicates the unified time base timestamp written for this sample. Indicates the starting point of the unified event time window. This indicates the window length of the uniform event time window. The floor operator is used to map continuous time to discrete window indices;

[0058] Ambient temperature is collected by ambient temperature sensors arranged on the outer wall of the furnace body and written into the status attribute set according to the same window index. Corresponding to the corresponding event time window, the vibration state is calculated by the root mean square acceleration of the triaxial accelerometer installed on the furnace body support within each event time window and compared with the threshold. The binary result is obtained by comparison and written into the state attribute set. The electromagnetic interference state is calculated by the electromagnetic field strength sensor within each event time window, and compared with the threshold. The binary results are obtained by comparison and written into the state attribute set. The above-mentioned ambient temperature, vibration state and electromagnetic interference state are used as window-level states and share the same window index with all sensor channel samples within the window, thereby realizing state association under a unified time base.

[0059] The sensor self-diagnostic status is obtained by parsing the diagnostic words returned by each sensing channel device along with the data frame, resulting in communication status flags, range status flags, and calibration status flags. The communication status flag is used to characterize the online status of the device and the data frame verification result. The range status flag is used to characterize whether the sampled value triggers the upper and lower limit protection of the range or the internal overflow of the sensor. The calibration status flag is used to characterize whether the calibration coefficient on the device side is valid and consistent with the calibration time recorded in the status attribute set. All three types of status flags use binary encoding, with 0 representing normal and 1 representing abnormal. They are written together with the sampled value of the corresponding sensing channel, the unified time base timestamp, the original sampling time, the sampling period, and the response time into a multi-source raw sequence to form a multi-source raw sequence and a status attribute set containing sampled values, time information, and status attributes.

[0060] In this specific embodiment, S2 includes:

[0061] The calibration compensation module indexes each event time window. Upon arrival, the original sampled values, unified time base timestamps, and original sampling times belonging to the event time window are extracted from the multi-source original sequence according to the sensing channel. The sampling period, response time, upper limit of range, lower limit of range, calibration time, communication status flag, range status flag, and calibration status flag of the sensing channel are read from the status attribute set. At the same time, the ambient temperature, vibration status, and electromagnetic interference status associated with the event time window are read as inputs for calibration compensation and reliability judgment.

[0062] For each sensing channel, two adjacent samples are arranged in ascending order of their original sampling times. The calibration compensation module divides the difference between adjacent sample values ​​by the difference between adjacent original sampling times to obtain the slope of the sample value change. The response time of the sensing channel is set as the response delay compensation amount. The sample value corresponding to the original sampling time is compensated for the response delay by using a linear extrapolation method of "slope of change multiplied by response delay compensation amount" to obtain the sample value after response delay compensation. At the same time, "original sampling time minus response delay compensation amount" is determined as the sampling time after response delay compensation for subsequent synchronization deviation calculation.

[0063] Temperature drift correction establishes a first-order linear drift model for each sensing channel, using a historical stability window as the data source for model parameters. The historical stability window is taken from the stable sample library, specifically the last 500 stable operating condition event time windows. The calibration compensation module calculates the deviation of the channel's response delay compensated sample value relative to the channel's stable reference value within the aforementioned historical stability window. The stable reference value is defined as the median of the channel's response delay compensated sample value within the last 500 stable operating condition event time windows. Subsequently, with ambient temperature as the independent variable and the deviation as the dependent variable, least squares fitting is performed to obtain the drift coefficient and bias term, which are then fixed as the channel's drift model parameters. During runtime, within the current event time window, the temperature drift correction amount is calculated based on the current ambient temperature and the channel's drift model parameters, and the response delay compensated sample value is corrected to generate a calibration sequence. The absolute value of the temperature drift correction amount exceeds the channel's measurement range. At that time, write a sensor drift anomaly flag into the status attribute set;

[0064] The range reliability assessment generates a reliability coefficient for each sensing channel within each event time window and aggregates them to form a reliability matrix. The reliability coefficient is composed of a base reliability value and a synchronization correction value, and they are multiplied and fused together.

[0065] ;

[0066] in, Indicates the sensing channel Index in the event time window The credibility coefficient within the range of values ​​is Indicates the sensing channel Index in the event time window The credibility base value and its range are within. Indicates the sensing channel Index in the event time window The synchronous correction value within the range of [value] is [value]. This indicates the sensor channel index and corresponds to the temperature channel, furnace pressure channel, flow channel, traction channel, or pipe diameter channel. This represents the event time window index calculated according to S1;

[0067] Among them, the basic value of credibility Within this window, the baseline confidence level of each sample is determined by taking the minimum value according to the worst-case scenario principle. The baseline confidence level of a single sample is obtained by multiplying the range boundary distance factor, communication status marker factor, range status marker factor, calibration status marker factor, vibration status factor, and electromagnetic interference status factor. The range boundary distance factor is 0 when the sampled value exceeds the upper limit of the range or falls below the lower limit of the range, and is 0 when the distance from the sampled value to the nearest range boundary is less than the range span. The factor is 0.5 when the communication status is abnormal and 1 when the communication status is normal. When the range status is abnormal, the factor is 0.2 when the range status is abnormal and the data reliability abnormality flag is written into the status attribute set. When the range status is normal, the factor is 1. When the calibration status is abnormal, the factor is 0.5 when the calibration status is abnormal and the sensor drift abnormality flag is written into the status attribute set. When the calibration status is normal, the factor is 1. When the vibration status is abnormal, the factor is 0.7 when the vibration status is abnormal and the data reliability abnormality flag is written into the status attribute set. When the vibration status is normal, the factor is 1. When the electromagnetic interference status is abnormal, the factor is 0.7 when the electromagnetic interference status is abnormal and the data reliability abnormality flag is written into the status attribute set. When the electromagnetic interference status is normal, the factor is 1.

[0068] Synchronization correction value The synchronization deviation is determined by the synchronization error. The synchronization error is defined as the absolute time difference between the sampled time after response delay compensation for each sample within the window and the expected sampling time of the alignment index of the channel within the window, and the maximum value is taken according to the worst-case principle. The expected sampling time of the alignment index is obtained by "aligning the sampled time after response delay compensation with the window start point as a reference according to the sampling period". When the synchronization error is not greater than 20ms, the synchronization correction value is 1. When the synchronization error is greater than 20ms but not greater than 200ms, the synchronization correction value decreases linearly to 0.2 according to the synchronization error. When the synchronization error is greater than 200ms, the synchronization correction value is 0.2 and a time base abnormality mark is written in the status attribute set.

[0069] When the reliability coefficient of any sensing channel When the data reliability threshold is less than 0.6, the calibration compensation module retains the time reference anomaly flag, sensor drift anomaly flag, and data reliability anomaly flag written to that channel within the event time window and uses them as input for subsequent time alignment and anomaly detection. Simultaneously, it compares the calibration sequence of each channel with the data reliability flag written to that channel within the event time window. The resulting credibility matrix is ​​output to step S3.

[0070] In this specific embodiment, S3 includes:

[0071] The time alignment module indexes each event time window. Upon arrival, read all calibration sampling points of each sensor channel within the event time window from the calibration sequence output in step S2, and then... Determined unified event time window start point With window length Establish an alignment index, and assign each calibration sampling point to the corresponding sampling time after response delay compensation, satisfying the alignment index. The event time window;

[0072] When multiple calibration sampling points exist for each sensing channel within the same event time window, the time alignment module takes the midpoint of the window. The calibration sampling point with the smallest absolute time difference from the midpoint of the window is selected as the window calibration value for that channel in that window, thereby compressing the asynchronous high-frequency sampling of each channel into a discrete sequence of "one window calibration value per channel per window".

[0073] Indexing each sensor channel within the target event time window When no calibration sampling points are present, the channel is marked as missing at its position within the window, and the missing mark is written into the missing mask matrix, denoted as . And its elements A value of 1 indicates a sensing channel. Index in the event time window Missing, with a value of 0 indicating no missing, where the sensor channel index... Consistent with step S2 and corresponding to one of the following: temperature channel, furnace pressure channel, flow channel, traction channel, or pipe diameter channel;

[0074] when At that time, the time alignment module generates a padding value based on the calibration values ​​of the same channel windows in adjacent windows and the calibration values ​​of the other channels windows in the same window, and writes it into the alignment timing matrix. The adjacent window takes the channel at the index... Leftmost non-missing window index index of the nearest non-missing window to the right And the search scope is set to An event time window index, if and If both exist, linear interpolation is performed using the window index distance to obtain the time-filled candidate value. If only one exists... Or only exist The existing one-sided window calibration value will then be used as a candidate value for time completion.

[0075] The remaining channels in the same window are filled using a process-coupled channel set. Generate cross-channel fill candidate values, where The process coupling relationship is fixed and determined, and in this embodiment, it is taken as... temperature Furnace pressure, pipe diameter Furnace pressure temperature flow Pipe diameter Traction Pipe diameter Pipe diameter Flow rate, traction, temperature The cross-channel fill-in candidate value is obtained by weighting the confidence of the window calibration values ​​of each channel in the same window within the coupled channel set. The confidence weight is taken from the corresponding element in the confidence matrix and only the channels marked as not missing in the missing mask matrix are used in the calculation.

[0076] The time alignment module fuses the time-padding candidate values ​​and cross-channel padding candidate values ​​according to their confidence level to obtain the padding value, and writes the padding value into the alignment time series matrix. The fusion calculation uses the following formula:

[0077] ;

[0078] in, Indicates the sensing channel Index in the event time window The fill value, Indicates time-padding fusion coefficient and takes This indicates that the time is used to supplement the credibility weight and take... When only one side of the adjacent window exists, the value of the corresponding side is taken. This indicates that the candidate values ​​are filled in by time. This indicates that the confidence weights are filled across channels and the set of coupled channels is taken. Reliability coefficient of the internally involved calculation channel The average value, This indicates that candidate values ​​are filled across channels. and Each represents the confidence coefficient of the corresponding sensing channel and the corresponding event time window in the confidence matrix. Represents the set of coupled channels Channel index within;

[0079] When the time-filled candidate value does not exist, Set to 0 and only use cross-channel padding candidate values ​​to complete padding; when no cross-channel padding candidate value exists, Set to 1 and use only time-based candidate values ​​for padding; if neither type of candidate value exists, then... Set to null and keep So that the missing mask matrix can be used to identify the location as unavailable in subsequent graph time series model inputs;

[0080] The time alignment module indexes each sensor channel within each event time window. Window calibration values ​​and fill values ​​are indexed by sensor channel. Index of Event Time Window The one-dimensional index arrangement forms an aligned timing matrix, and a missing mask matrix is ​​output synchronously to indicate the positions in the aligned timing matrix obtained by the actual sampling and the positions obtained by padding.

[0081] In this specific embodiment, S4 includes:

[0082] The dynamic graph building module indexes each event time window. Generate dynamic coupling graph The dynamic coupling diagram It consists of a set of nodes, a set of process prior edges, and a set of edge weights. The set of nodes is fixed to include temperature nodes, furnace pressure nodes, gas flow nodes, traction speed nodes, and online pipe diameter nodes. The set of process prior edges is fixed to include the edges between temperature nodes and furnace pressure nodes, the edges between gas flow nodes and online pipe diameter nodes, the edges between traction speed nodes and online pipe diameter nodes, and the edges between temperature nodes and online pipe diameter nodes.

[0083] The dynamic graph building module reads the index of each node in the event time window from the aligned time sequence matrix. The window calibration value and the alignment timing matrix are denoted as ,in Indicates the sensing channel Index in the event time window The window calibration value and when the position is generated by padding. Take the padding value; if the position is empty, the position is considered unusable and will be removed in subsequent calculations.

[0084] To introduce a "historical stability window baseline" in edge weight updates, the dynamic graph construction module reads the indexes of the 500 most recent event time windows marked as stable operating conditions from the stability sample library and constructs a set of historical stability windows for each node. Extract the corresponding data from the historical stable window set. The available values ​​are calculated and the median is used as the historical stable window reference value for that node, thereby defining the node deviation in the current window and the rolling window as "the difference between the window calibration value and the historical stable window reference value".

[0085] In each process prioritization edge The dynamic graph construction module simultaneously calculates three types of quantities: lag correlation, bias covariance, and confidence product, and updates the edge in the window accordingly. The edge weights, where the lagged correlation quantities are in the set of lagged orders. Internally calculate nodes one by one Deviation sequence and nodes The Pearson correlation coefficients between the deviation sequences are calculated, and the one with the largest absolute value is taken as the lag correlation for that side. The deviation sequences are taken from the time window index containing the current event. The scrolling window is included and the length of the scrolling window is fixed at 30 event time window indices. When calculating the correlation coefficient, sample points with null values ​​at any end node in the corresponding window position are removed, and when the number of valid sample points is less than 10, the hysteresis correlation is set to 0.

[0086] Deviation covariance for nodes within the same rolling window With nodes The sample covariance of the deviation sequence is calculated, and the absolute value is taken as the deviation covariance of that edge. When calculating the covariance, the same sample point elimination rule as the lag correlation quantity is used, and when the number of effective sample points is less than 10, the deviation covariance is set to 0. This deviation covariance is then used according to the same process prior edge within the historical stable window set. The distribution of the deviation covariance Quantile values ​​are truncated at the upper limit and linearly normalized to ensure the stability of their numerical range;

[0087] The confidence product is obtained directly from the nodes in the confidence matrix. With nodes Index in the current event time window The product of the credibility coefficients is set to 0 when either credibility coefficient is null.

[0088] After obtaining the above three types of quantities, the dynamic graph construction module updates the process prior edges using the following formula. Index in the event time window Boundary rights

[0089] ;

[0090] in Indicates the process prior edge Index in the event time window The edge weights and their range of values ​​are This represents the index of one end node of the process prior edge, and its value is taken from the temperature node, furnace pressure node, gas flow rate node, traction speed node, and online pipe diameter node. Represents the index of the other end node of the process prior edge and the value rules are the same as... Consistent Indicates the event time window index. Representing an edge In the window The normalized result of the lagged correlation and the range of values ​​is Representing an edge In the window The normalized result of the deviation covariance and the range of values ​​is Representing an edge In the window The product of credibility and the range of values ​​is It is obtained by multiplying the confidence coefficients of the two endpoints in the confidence matrix. This represents the fusion weight of lagged correlation quantities and its value is... This represents the weights of the bias covariance fusion and takes values ​​of 1. This represents the confidence product fusion weight with a value of 0.2. This function represents a truncation function that sets the part of the result within the parentheses that is less than 0 to 0 and the part that is greater than 1 to 1.

[0091] After updating the edge weights of all process prior edges, the dynamic graph construction module encapsulates the node set, process prior edge set, and edge weight set into a dynamically coupled graph. The information is then output to step S5 for transmitting deviation information along the dynamic edge and for identifying abnormal operating conditions.

[0092] In this specific embodiment, S5 includes:

[0093] The anomaly detection module indexes events using an event time window. The corresponding time window of the event to be judged is used as the time window of the target event, from the steps Output alignment timing matrix Extract each sensor channel in the most recent The sequence of window calibration values ​​on the event time window index is read synchronously to mask unusable positions, and the dynamic coupling graph output from step 44 is also read synchronously. Read the process prior edge set and its position in the window The right of the border;

[0094] The graph temporal model adopts a deterministic structure of "intra-node time window encoder + dynamic graph message passer + dual-head reconstruction output", wherein the intra-node time window encoder is a single-layer gated cyclic unit network with hidden dimensions. The length of each sensing channel is The window calibration value sequence is encoded and the channel is output at the target event time window index. The node timing representation is as follows: during encoding, the corresponding input value is set to zero for the missing position marked as missing in the missing mask matrix, and the update gate at that position is forced to zero to prevent missing values ​​from entering the state;

[0095] Dynamic graph message passer index in target event time window With dynamic coupling diagram The edge weights are used as transfer coefficients to transfer the node timing representation and the "deviation relative to the historical stable window benchmark value" between the nodes at both ends of each process prior edge. The transferred information from adjacent nodes is weighted and summed, then concatenated with the timing representation of the current node. A coupled representation is obtained through linear transformation and hyperbolic tangent activation. The output dimension of the linear transformation is set to... Furthermore, the weight parameters are fixed after the model training is completed;

[0096] The dual-headed reconstruction output includes node output headers and edge output headers. The node output header maps the coupling representation of each node to the node's index within the target event time window. Predict the node parameters and output them as The output head outputs the reconstruction parameters between the two nodes of each process prior edge, and outputs the edge deviation as follows: ,in and All are deterministic outputs of the model;

[0097] The anomaly detection module uses the alignment timing matrix. Read the target event time window index Measured node parameters The single-parameter residual is generated using the "absolute difference between predicted node parameters and measured node parameters", and the parameter pair residual is generated using the "absolute difference between the side deviation of reconstructed parameters and the side deviation of measured parameters". The measured side deviation is defined as the index of the time window of the same target event. Differences in measured node parameters between the two ends of the prior edge of the same process ;

[0098] The anomaly identification module fuses single-parameter residuals and parameter pair residuals to generate an anomaly score for the operating condition and calculates it using the following formula:

[0099] ;

[0100] in, Indicates the index of the target event time window The abnormal operating condition score is used to perform comparable ranking of different windows. Indicates the number of sensor channels and takes This indicates a sensor channel index that corresponds to one of the following: temperature channel, furnace pressure channel, flow rate channel, traction channel, or pipe diameter channel. Indicates the sensing channel Index in the target event time window The single-parameter residual, Indicates the sensing channel In layered labels Single parameter threshold under, Indicates the number of prior edges in the process and takes This indicates the process prior edge index, and its value is taken from the four edges: temperature-furnace pressure, flow rate-pipe diameter, traction-pipe diameter, and temperature-pipe diameter. Indicates the process prior edge Index in the target event time window The parameters affect the residuals. Indicates the process prior edge In layered labels The coupling edge threshold under the condition, Indicates the index of the target event time window. The associated hierarchical labels are uniquely identified by the product specifications, melting stage, and furnace temperature range ternary combination of this window;

[0101] The stratified calibration thresholds are generated offline by the anomaly detection module based on a historical stability window and stored permanently when the system goes online. Specifically, this is done for each stratified label. Collect the corresponding stable sample library The historical stability window is used to calculate the single-parameter inconsistency score on each sensing channel using the single-parameter residual sequence within that window, and then the target coverage is obtained. The quantile value below is used as the single-parameter threshold for this channel. On each process prior edge, the parameters within that window are used to calculate the parameter-to-inconsistency score of the residual sequence and the target coverage is obtained. The quantile value below is used as the coupling edge threshold for that edge. Both single-parameter inconsistency scores and parameter-pair inconsistency scores are calculated by dividing the residual by the absolute deviation of the median residual of the group and adding a stability term. The calculation method is determined to make residuals of different dimensions comparable when the threshold is determined;

[0102] The exception type output uses a hierarchical discrimination logic and is consistent with the state flag written into the state attribute set in step S2. Specifically, it is based on the target event time window index. There exists a single-parameter residual in any sensing channel that exceeds the corresponding single-parameter threshold. The output corresponds to a single parameter anomaly in the sensor channel, and the target event time window index is updated accordingly. There exists a parameter on any process prior edge whose residual exceeds the threshold of the corresponding coupling edge. When the process coupling anomaly corresponding to the process prior edge is output, when there is a time reference anomaly mark, sensor drift anomaly mark, or data reliability anomaly mark in the state attribute set, and the parameter pair residual that exceeds the coupling edge threshold only appears on the process prior edge containing the marked sensing channel, the anomaly represented by the corresponding state mark is output. When multiple of the single parameter anomaly, process coupling anomaly, and state mark anomaly are satisfied at the same time, a composite anomaly type is output. When neither the single parameter residual nor the parameter pair residual exceeds the corresponding threshold and the state mark does not exist in the state attribute set, the anomaly type is empty.

[0103] The model update process and anomaly output are executed in a closed loop, using a stable sample library as a carrier. Specifically, this is done when the target event time window index... When the anomaly type is empty and all single-parameter residuals are not greater than the corresponding single-parameter threshold and all parameter pair residuals are not greater than the corresponding coupling edge threshold, the target event time window index is used. Aligned time-series data, dynamic coupling graph edge weights, and hierarchical labels are written into the stable sample library and marked as stable operating conditions. When the number of newly written samples in the stable sample library reaches the preset update number... At that time, only the updated stable sample library is used to redetermine the single-parameter thresholds and coupling edge thresholds under all hierarchical labels, while keeping the node input type, edge input type, and output type of the graph time series model unchanged. When the redetermined hierarchical calibration threshold causes the number of windows output as abnormal in the historical stable window set to exceed the preset number of reviews, The updated results are discarded while the previous stratified calibration threshold is retained to ensure the consistency of the stable sample library and the consistency of anomaly detection.

[0104] 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.

[0105] The technical solution of this invention transforms scattered field sensor signals into comparable, correlated, and discriminable operating condition data by unifying the time base, calibration compensation, time alignment, and dynamic graph timing identification through continuous processing relationships, enabling the identification of coupling anomalies between temperature, furnace pressure, gas flow rate, traction speed, and online pipe diameter.

[0106] This invention further introduces a reliability coefficient, a status marker, a hierarchical calibration threshold, and a stable sample library update mechanism, enabling low-reliability sensing data, sensor drift, time reference anomalies, and real process coupling anomalies to be handled according to different anomaly types, thereby better adapting to the high-interference working conditions at the quartz tube melting site.

Claims

1. A method for multi-parameter intelligent sensing during quartz tube melting, characterized in that, include: S1. Collect temperature, furnace pressure, gas flow rate, traction speed, online pipe diameter, and sensor self-diagnostic status at the quartz tube melting site. Write a unified time base timestamp, original sampling time, sampling period, response time, range, ambient temperature, vibration status, and electromagnetic interference status for each data acquisition channel to obtain a multi-source original sequence and state attribute set. S2. Perform response delay compensation, temperature drift correction, and range reliability judgment based on the multi-source original sequence and state attribute set to generate a calibration sequence and reliability matrix. S3. Align the calibration sequence according to a unified event time window to generate a missing mask matrix. Combine the missing mask matrix and reliability matrix to fill in asynchronous segments and generate an aligned time sequence matrix. S4. Using temperature, furnace pressure, gas flow rate, traction speed, and online pipe diameter as nodes, generate a dynamic coupling diagram based on the preset process coupling relationship and the deviation correlation between real-time parameters calculated by the alignment time series matrix; S5. Input the alignment time series matrix and the dynamic coupling diagram into the graph time series model, perform time window encoding within the nodes and transmit deviation information along the dynamic edges to obtain single parameter residuals, parameter pair residuals, and operating condition anomaly scores, and output the anomaly type according to the hierarchical calibration threshold determined by the operating condition stratification.

2. The method for multi-parameter intelligent sensing in quartz tube melting according to claim 1, characterized in that, S1 includes: Temperature, furnace pressure, gas flow rate, traction speed, and online pipe diameter are used as sensing channels. The device number, installation location, upper and lower range limits, calibration time, sampling period, and response time of each sensing channel are written into the status attribute set. Ambient temperature, vibration status, and electromagnetic interference status are associated with the same event time window using a unified time base timestamp. The sensor self-diagnostic status is converted into communication status markers, range status markers, and calibration status markers, which, together with the sampled values ​​of the corresponding sensing channels, form a multi-source raw sequence.

3. The method for multi-parameter intelligent sensing in quartz tube melting according to claim 2, characterized in that, S2 includes: The response delay compensation amount is determined based on the slope of the sampled value change and the response time of the sensing channel, and the sampled value corresponding to the original sampling time is corrected according to the response delay compensation amount. The temperature drift correction amount is determined based on the ambient temperature and the calibration deviation of the same sensing channel within the historical stable window. The historical stable window is the event time window marked as a stable operating condition in the calibration stage or historical production records. The temperature drift correction amount is used to correct the sampled value after response delay compensation. The reliability coefficient is determined based on the distance between the sampled value and the range boundary, communication status mark, range status mark, calibration status mark, vibration status, and electromagnetic interference status. The reliability matrix is ​​composed of the reliability coefficients of each sensing channel.

4. The method for multi-parameter intelligent sensing in quartz tube melting according to claim 1, characterized in that, S3 includes: establishing an alignment index based on the window start point and window length of a unified event time window, mapping the calibration sequence to the alignment index; when any sensing channel lacks a sampled value within the target window, recording the missing position in the missing mask matrix, and generating a filler value using the calibration values ​​of adjacent windows, the calibration values ​​of other sensing channels within the same window, and the corresponding elements in the confidence matrix; arranging the original calibration values ​​and the filler values ​​according to the alignment index to form an alignment time sequence matrix.

5. The method for multi-parameter intelligent sensing in quartz tube melting according to claim 1, characterized in that, S4 includes: Set temperature and furnace pressure, gas flow rate and online pipe diameter, traction speed and online pipe diameter, and temperature and online pipe diameter as process priors respectively; For each process prior edge, the hysteresis correlation of the nodes at both ends of the edge in the current window and the historical stable window is calculated according to the alignment time series matrix. The historical stable window is the event time window marked as a stable operating condition in the calibration stage or historical production record. The historical stable window benchmark value is the statistical benchmark value of the corresponding node calibration value. Within the rolling window containing the current event time window, the deviation covariance is calculated according to the deviation sequence of the nodes at both ends of the edge relative to the historical stable window benchmark value, and the confidence product of the nodes at both ends of the edge is calculated. The corresponding edge weight is updated according to the hysteresis correlation, the deviation covariance and the confidence product, and a dynamic coupling graph is formed by the nodes, the process prior edge and the updated edge weight.

6. The method for multi-parameter intelligent sensing in quartz tube melting according to claim 1, characterized in that, S5 includes: The graph time series model uses the time window of the event to be discriminated as the target event time window, and performs time window encoding on the sequence of each node in the aligned time series matrix to obtain the node time series representation. The coupling representation is obtained by transmitting node timing representation and deviation information between adjacent nodes based on the edge weights of the dynamic coupling graph. The node parameters of the target event time window are predicted based on the coupling representation, and the parameter pair deviation between the two ends of the same process prior edge in the same target event time window is reconstructed. The measured edge deviation is determined based on the difference in measured parameters between the two ends of the corresponding process prior edge in the target event time window. The difference between the predicted node parameters and the measured node parameters of the target event time window is determined as the single parameter residual. The difference between the reconstructed parameter pair deviation and the measured edge deviation corresponding to the same target event time window and the same process prior edge is determined as the parameter pair residual. The operating condition anomaly score is generated from the single parameter residual and the parameter pair residual.

7. The method for multi-parameter intelligent sensing in quartz tube melting according to claim 3, characterized in that, The reliability coefficient includes a base reliability value and a synchronization correction value. The base reliability value is determined based on the distance between the sampled value and the range boundary, communication status markers, range status markers, calibration status markers, vibration status, and electromagnetic interference status. The synchronization correction value is determined based on the synchronization deviation, which is the time deviation of the sampling time relative to the unified event time window alignment index after compensation for the response delay of the sensing channel. The unified event time window alignment index is an index established according to the window start point and window length of the unified event time window. When the reliability coefficient of any sensing channel determined by the state label, vibration state, electromagnetic interference state, and synchronization deviation is less than the preset reliability threshold, a state label is generated according to the factors that cause the reliability coefficient to be less than the preset reliability threshold. Among them, a time reference anomaly label is generated when the synchronization deviation exceeds the preset synchronization deviation threshold; a sensor drift anomaly label is generated when the temperature drift correction exceeds the preset drift threshold or the calibration state label is abnormal; and a data reliability anomaly label is generated when the distance between the sampled value and the range boundary is less than the preset range distance threshold, the vibration state is abnormal, or the electromagnetic interference state is abnormal. The generated state labels are written into the state attribute set.

8. The method for multi-parameter intelligent sensing in quartz tube melting according to claim 7, characterized in that, The stratified calibration threshold is determined as follows: The product specifications, melting stage, and furnace temperature range of each event time window are acquired and associated as stratification labels. Historical stable windows are grouped according to these stratification labels. The historical stable windows are event time windows marked as stable operating conditions in the calibration stage or historical production records. For each group of historical stable windows, a single-parameter inconsistency score is calculated based on the single-parameter residual, and a parameter-to-parameter inconsistency score is calculated based on the parameter-to-parameter residual. A preset coverage rate for each group is used as the target coverage rate. Under the target coverage rate, the quantile values ​​of the single-parameter inconsistency score and the parameter-to-parameter inconsistency score are determined. The quantile values ​​of the single-parameter inconsistency score are mapped to single-parameter thresholds, and the quantile values ​​of the parameter-to-parameter inconsistency score are mapped to coupling edge thresholds. The current event time window calls the corresponding single-parameter threshold and coupling edge threshold according to the same hierarchical label, compares the single-parameter residual of the current window with the single-parameter threshold, and compares the parameter pair residual of the current window with the coupling edge threshold to obtain the anomaly discrimination value; When the single-parameter residual exceeds the corresponding single-parameter threshold, the anomaly type includes single-parameter anomaly of the corresponding sensing channel. When the parameter pair residual exceeds the corresponding coupling edge threshold, the anomaly type includes process coupling anomaly of the corresponding process prior edge. When there are time base anomaly markers, sensor drift anomaly markers, or data reliability anomaly markers in the state attribute set, and only the parameter pair residuals of the marked sensing channel exceed the corresponding coupling edge threshold, the anomaly type is the anomaly represented by the corresponding state marker. When multiple anomaly discrimination conditions are met simultaneously, a composite anomaly type is output. When neither the single-parameter residual nor the parameter pair residual exceeds the corresponding threshold and the state marker does not exist in the state attribute set, the anomaly type is empty.

9. A multi-parameter intelligent sensing method for quartz tube melting according to claim 8, characterized in that, It also includes a model update step: when the anomaly type of the event time window is empty, the single parameter residual is not greater than the corresponding single parameter threshold, and the parameter pair residual is not greater than the corresponding coupling edge threshold, the event time window is determined as a stable condition and written into the stable sample library; when the number of newly written samples in the stable sample library reaches the preset update number, the stable sample library is used to redetermine the hierarchical calibration threshold, and the node input, edge input, and output types of the graph time series model remain unchanged; If the redefined stratified calibration threshold causes the number of abnormal outputs in the historical stable window to exceed the preset number of reviews, the stratified calibration threshold before the update is retained.

10. A multi-parameter intelligent sensing system for quartz tube melting, used to execute the multi-parameter intelligent sensing method for quartz tube melting as described in any one of claims 1 to 9, characterized in that, include: The data acquisition module is used to collect multi-source raw sequences from the quartz tube melting site and write them into the status attribute set; The calibration compensation module is used to perform response delay compensation, temperature drift correction, and range confidence judgment, and to generate calibration sequences and confidence matrices. The time alignment module is used to generate a missing mask matrix and an aligned time sequence matrix according to a unified event time window; The dynamic graph construction module is used to generate dynamic coupling graphs based on preset process coupling relationships and the correlation between deviations between real-time parameters. The anomaly identification module is used to output single-parameter residuals, parameter pair residuals, anomaly scores, and anomaly types based on the alignment time series matrix, dynamic coupling diagram, and hierarchical calibration thresholds determined by operating condition stratification.

Images