A welding pipe unit abnormality detection method based on contrast learning

Abstract

The application discloses a kind of based on contrast learning's weld pipe unit abnormality detection method, to solve the time scale drift of multiple sensing channels, cross-station alignment and the lack of physical consistency, transient interference causes detection instability and inaccurate positioning problem, the present application is through the time scale correction, time to cumulative length mapping and encoder boundary constraint, neural residual correction and alignment confidence evaluation, fixed length grid constructs physical consistent contrast sample, memory bank and dynamic threshold detection and causal online fine-tuning rollback, realizes cross-station unified alignment and station level positioning, improves the technical effects of abnormality detection accuracy and robustness and maintains online stability.

Description

TECHNICAL FIELD

[0001] The present application relates to the field of industrial anomaly detection, and particularly relates to a welding pipe unit anomaly detection method based on contrast learning. BACKGROUND

[0002] The welding pipe unit belongs to a continuous production line, and is connected in series at workstations such as forming, welding, sizing, and cutting. On-site signals such as line speed, encoder pulse, weld temperature, and start-stop and cutting states are usually collected. Affected by start-stop, cutting, joint passing, and load fluctuations, the line speed frequently changes, the same pipe section does not appear at different workstations at the same time, and "time scale mismatch" is easily generated, which reduces the effect of anomaly detection and positioning based on a time window.

[0003] Existing technologies usually use a unified time base and a PLC encoder to calibrate the length, or use speed normalization, resampling, and setting a fixed propagation time delay to align across workstations. Some methods use signal alignment methods such as cross-correlation and DTW, and threshold determination and automatic encoder deep learning for detection. However, there are still deficiencies in physical consistency and online stability.

[0004] (1) The features and detection based on a time window are not equivalent at variable line speeds. The time span corresponding to the same cumulative length changes with the line speed, which causes mismatch between training samples and detection samples, weakens the correlation across workstations, and easily causes false positives and false negatives.

[0005] (2) The alignment relies on static length calibration of the encoder or a fixed propagation time delay, and it is difficult to correct the channel time tag drift and speed-dependent propagation differences in real time, which causes inconsistency between the length coordinates and boundary conditions.

[0006] (3) Transient interference such as start-stop, cutting, and joint events lacks effective shielding and alignment confidence measurement, and unstable samples are easily introduced into model training and online detection, which causes performance fluctuations and inaccurate positioning.

[0007] Therefore, the existing technology still lacks a welding pipe unit anomaly detection method that can overcome the above deficiencies, which is a problem to be solved by those skilled in the art. SUMMARY

[0008] One purpose of the present application is to propose a welding pipe unit abnormality detection method based on contrast learning, aiming at the problems of time scale mismatch caused by line speed change, difficulty in cross-station alignment and unstable detection under transient interference in the prior art, a technical scheme is proposed, which unifies time base collection and event shielding, channel differentiable time scale correction, causal update, time to cumulative length mapping and encoder boundary calibration meeting the advection constraint, combined with physical distance calculation of propagation delay and cross-station alignment on a fixed length grid, causal neural residual correction and alignment confidence, contrast sample construction and memory bank and dynamic threshold detection, online fine-tuning based on confidence and anchor point deviation, The present application has the technical effects of realizing unified length scale and stable alignment under variable line speed, improving abnormality detection accuracy and positioning accuracy and maintaining online stability.

[0009] According to an embodiment of the present application, a welding pipe unit abnormality detection method based on contrast learning is provided, characterized by comprising the following steps:

[0010] S1, collect time series data and process data of multiple sensing channels, and timestamp label with unified time base, output multiple sensing channel data, including line speed, encoder pulse, event flag and event window;

[0011] S2, input multiple sensing channel data, establish a differentiable time scale correction layer for each sensing channel, map the original time to the corrected time using multiplicative coefficients and additive offsets, and output the corrected time of each sensing channel and the multiplicative coefficients and additive offsets;

[0012] S3, input the corrected time and line speed, construct a differentiable mapping from time to cumulative length coordinates, calculate the propagation delay based on the physical distance between stations, and align all sensing channels on a fixed length grid, output the cumulative length coordinates, propagation delay and aligned samples on the fixed length grid;

[0013] S4, input the cumulative length coordinates and propagation delay as the main input, and input the corrected time, line speed, encoder pulse, event flag and event window as auxiliary input, establish a neural residual correction module, correct the cumulative length coordinates and propagation delay, and output the corrected cumulative length coordinates, corrected propagation delay and alignment confidence;

[0014] S5, input the fixed length grid alignment sample and alignment confidence, and set the cumulative length window on the fixed length grid based on the event window shielding, construct and output the positive and negative sample sets of contrast learning training under the premise of meeting the advection constraint;

[0015] S6, inputting the sample set, training the feature representation network, establishing the memory bank, generating the anomaly score according to the distance or similarity with the memory bank, adopting the dynamic threshold and aligning the output window boundary with the encoder pulse pair, and outputting the anomaly score, the anomaly detection result, the station-level positioning and the dynamic threshold;

[0016] S7, inputting the alignment confidence, the anchor point deviation calculated from the time-to-accumulation length mapping and the propagation delay, and the anomaly detection and quality inspection results, fine-tuning the multiplicative coefficient and the additive offset and writing back to the time-to-accumulation length mapping, purifying the memory bank and updating the dynamic threshold according to the quality inspection result, and outputting the updated multiplicative coefficient and additive offset, time-to-accumulation length mapping, dynamic threshold and memory bank state.

[0017] Optionally, step S1 is specifically:

[0018] The multi-sensor channel at least includes a line speed time series, an encoder pulse, and event markers for marking start-stop, cutting, joint passing, and weld temperature peak;

[0019] The time series data and process data of the multi-sensor channel are time-stamped with a unified time base, so that the time stamps of each acquisition card and the control unit are monotonous and consistent and continuous in the start-stop phase;

[0020] The line speed time series and the encoder pulse are subjected to anti-interference filtering and de-bouncing processing, and the packet loss and saturation state are recorded;

[0021] The event markers are generated by rule triggering, wherein the start-stop event marker is triggered by the line speed time series crossing the set speed threshold and meeting the set duration, the cutting event marker is triggered by the rising edge of the cutting control signal and is confirmed by the encoder pulse count, the joint passing event marker is triggered by the peak of the weld temperature signal exceeding the set temperature threshold and is confirmed by the encoder pulse, and the weld temperature peak event marker is triggered by the local extreme value of the weld temperature signal meeting the set condition;

[0022] The event markers record the event window with the start time stamp and the end time stamp, and are output together with the data marked by the unified time base.

[0023] Optionally, step S2 is specifically:

[0024] The multiplicative coefficient and the additive offset meet the homology constraint and the smoothing constraint in the accumulation length dimension, and are constrained by the encoder pulse and the event marker as end anchor points, and the update of the multiplicative coefficient and the additive offset is reduced or frozen within the event window;

[0025] Inputting the multi-sensor channel data, mapping the original time to the corrected time by linear combination of the multiplicative coefficient and the additive offset for each sensor channel;

[0026] The multiplicative coefficient and the additive offset are determined by jointly considering a least square objective of endpoint anchor consistency and a smoothing penalty of cumulative length dimension, the endpoint anchors are determined by time points given by encoder pulses and event markers, and the revised time is consistent with the original record at the endpoint anchors;

[0027] The smoothing penalty takes cumulative length as independent variable, limits the variation rate of multiplicative coefficient and additive offset along cumulative length, and sets upper and lower limits for multiplicative coefficient and additive offset respectively;

[0028] The homogenous constraint adopts the way of in-group parameter sharing or in-group parameter deviation penalty, so that the multiplicative coefficient and the additive offset of channels of the same acquisition card or the same control unit are consistent or the deviation thereof is limited;

[0029] The update of multiplicative coefficient and additive offset adopts a causal strategy, only uses data no later than the current time for update, and freezes update or reduces update step within the event window.

[0030] Optionally, step S3 is specifically:

[0031] The mapping satisfies the constraint of advection partial differential equation and takes encoder pulses as boundary conditions;

[0032] Taking revised time and linear speed as input, a differentiable time-to-cumulative length coordinate mapping is established by numerical integration, so that the cumulative length monotonically increases with the revised time and satisfies the constraint of advection partial differential equation;

[0033] The calibration coefficient of the mapping is determined according to the length calibration of the encoder pulse, and the encoder pulse is taken as the boundary condition to calibrate the value of cumulative length at the pulse time;

[0034] The data falling into the event window is not involved in the estimation of the calibration coefficient and the update of the boundary condition, so as to avoid transient interference on the mapping stability;

[0035] The propagation time delay is calculated according to the physical distance between each station, and the length offset of each station is determined, a reference station is selected, so that the difference between the cumulative length of the same pipe segment at any station and the cumulative length at the reference station is equal to the corresponding physical distance;

[0036] The samples of each sensing channel are uniformly aligned on a fixed length grid, the fixed length grid adopts a preset length interval, the grid starting point is aligned with the nearest encoder pulse, and linear interpolation or spline interpolation is adopted to complete gridding and cross-station alignment;

[0037] When the cumulative length difference of different station sample windows falls within the set physical tolerance range, it is determined that the alignment of the same pipe segment is successful, and the cumulative length coordinate, propagation time delay and aligned sample on the fixed length grid are output.

[0038] Optionally, step S4 is specifically:

[0039] The parameter update and inference of the neural residual correction module adopts a causal strategy, and amplitude constraints and smoothing constraints are imposed on the cumulative length dimension, while the alignment confidence is calculated;

[0040] A neural residual correction module with a causal structure is established, taking cumulative length coordinates and propagation time delays as main inputs, and taking correction times, linear speeds, encoder pulses, event flags, and event windows as auxiliary inputs;

[0041] The module performs residual compensation on cumulative length coordinates and propagation time delays, and only executes parameter update and inference based on data no later than the current correction time, and reduces the parameter update step or freezes the parameters within the event window;

[0042] The residual is set as the independent variable of the cumulative length, and the amplitude constraint is set to limit the upper and lower limits of the absolute value of the residual, and the smoothing constraint is realized by the second-order difference penalty in the cumulative length dimension to limit the rapid change of the residual;

[0043] Different stations share the same parameters in the shared parameter part to reflect mechanical coupling, and use station-specific parameters in the independent parameter part to compensate for station differences;

[0044] The alignment confidence is calculated based on the difference between the cumulative length coordinates before and after correction, the difference between the propagation time delays before and after correction, and the consistency index of adjacent length grids, and monotonically decreases with the increase of the above-mentioned differences;

[0045] The corrected cumulative length coordinates, corrected propagation time delays, and alignment confidence are output.

[0046] Optionally, step S5 is specifically:

[0047] Taking the alignment samples and alignment confidence on the fixed length grid as input, and taking the event window as the shielding basis, setting the cumulative length window on the fixed length grid determined by the window length and the step, and shielding the cumulative length window falling into the event window and the preset length range before and after it;

[0048] According to the alignment confidence threshold, only the cumulative length window with alignment confidence not lower than the threshold is retained for sample construction;

[0049] The positive samples are constructed under the premise of satisfying the advection constraint, and are selected from the same cumulative length window, including samples from the same sensing channel at different time periods or different line speeds, and samples from different sensing channels or different stations, wherein the positive samples across stations are limited to have a corresponding propagation time delay difference within a set physical tolerance range, and the start and end positions of the cumulative length window are confirmed by encoder pulses;

[0050] When constructing the negative samples, candidate windows are selected within a cumulative length neighborhood, sampling weights are set according to cumulative length distances, the weights monotonically decrease with the increase of the cumulative length distance, and windows with an alignment confidence lower than a threshold and windows shielded by event windows are removed;

[0051] For the negative samples across stations, windows with a propagation time delay difference not satisfying the physical tolerance range or a cumulative length offset exceeding a set neighborhood radius are preferentially selected;

[0052] The cumulative length neighborhood radius, window length, step size and alignment confidence threshold are set according to the physical distance between stations, line speed statistics and encoder pulse calibration, and are fine-tuned online according to the statistical results of the alignment confidence in the production process;

[0053] The positive sample and negative sample sets for training are output.

[0054] Optionally, step S6 specifically comprises:

[0055] The positive sample and negative sample sets are used as training inputs, and the alignment confidence is used as a sample weight to train a feature representation network;

[0056] The similarity calculation is performed in a fixed length grid, the normalized feature representation is used, and the cumulative length window is used for aggregation, a physically consistent similarity comparison loss is adopted, the similarity of the positive samples is increased, and the similarity of the negative samples is decreased;

[0057] To ensure physical consistency, a constraint is added to the loss, so that the feature representations of the same cumulative length window at different line speeds and different stations remain consistent, and samples not satisfying the advection partial differential equation constraint or having an alignment confidence lower than a threshold are reduced in weight or removed;

[0058] A memory bank is established and is managed in a partitioned manner according to a fixed length grid and a process state, the samples in the memory bank satisfy an alignment confidence not lower than a set threshold and do not fall into an event window;

[0059] The distance or similarity to the memory bank is calculated for a to-be-detected sample in the corresponding partition, and an abnormality score is generated, the abnormality score increases as the distance increases or the similarity decreases;

[0060] A dynamic threshold is set for different line speeds and different stations, the dynamic threshold is adaptively updated according to the statistics of the memory bank partition and the statistics of the alignment confidence, and the calculation window of the dynamic threshold is boundary-aligned by the encoder pulse;

[0061] Finally, an abnormality detection result is output based on the abnormality score and the dynamic threshold, and a station-level positioning is generated in combination with the cumulative length coordinate and the propagation time delay.

[0062] Optionally, step S7 is specifically:

[0063] The alignment confidence of the sliding statistics and the endpoint anchor point deviation calculated from the mapping of the time to the cumulative length and the propagation time delay are taken as the triggering basis, wherein the triggering condition of the alignment confidence is that the alignment confidence is lower than a set threshold in a continuous preset number of cumulative length windows of the fixed length grid, and the triggering condition of the endpoint anchor point deviation is that the absolute value of the difference between the cumulative length at the endpoint anchor point and the nominal cumulative length determined by the encoder pulse exceeds a set tolerance;

[0064] When any triggering condition is met, the multiplicative coefficient and the additive offset are fine-tuned online using a causal strategy, the update process uses an incremental least square with a cumulative length dimension smoothing penalty or a small step gradient update, and the parameter upper and lower limits and the homologous constraints are obeyed, the update step is frozen or reduced in the event window;

[0065] After the fine-tuning is completed, the update result is written back to the mapping of the time to the cumulative length, so that the mapping satisfies the boundary condition at the endpoint anchor point and maintains monotonicity, and the dynamic threshold is re-estimated using a quantile strategy according to the distribution of the abnormality score in each fixed length grid partition;

[0066] The samples are back-labeled on the cumulative length coordinate according to the quality inspection result, the memory bank is purified, the samples judged as abnormal or unqualified are removed, and the samples close to the event window are down-weighted or removed;

[0067] If the alignment confidence does not recover to above the threshold or the endpoint anchor point deviation still exceeds the tolerance after the fine-tuning, or there is a sign of instability that the abnormality score moves up as a whole, then the last stable parameter group is rolled back, and the dynamic threshold and the memory bank partition before the rollback are restored;

[0068] The multiplicative coefficient and the additive offset after the update or the rollback, the mapping of the time to the cumulative length, the dynamic threshold and the memory bank state are output.

[0069] The beneficial effects of the present application are:

[0070] (1) Line speed invariance and time scale unification: Through the advection PDE time-cumulative length mapping of differentiable digital twinning, superimposing differentiable clock, communication delay correction and causal estimation of multiplicative time scale drift, the same pipe segment is converted to a unified length scale at different line speeds, significantly alleviating time scale mismatch, improving the consistency of training and detection samples, and reducing false positives and false negatives;

[0071] (2) Cross-station high-precision alignment and positioning: With encoder pulses as boundary conditions, combined with physical modeling of propagation delay and neural residual correction with amplitude and smoothing constraints, and with physically consistent contrastive learning and alignment confidence filtering samples, the alignment accuracy and robustness of multiple channels and cross-station are improved, enabling more accurate station-level anomaly positioning;

[0072] (3) Online stability and adaptive capability: With causal update and event window shielding, combined with dynamic threshold, memory pool purification and parameter fine-tuning / rollback mechanism, the detection stability and rapid adaptation can be maintained under transient disturbances such as start-stop, cutting, joint passing and long-term time scale drift, reducing maintenance costs and improving long-term operation reliability. BRIEF DESCRIPTION OF DRAWINGS

[0073] The accompanying drawings are included to provide a further understanding of the application, and constitute a part of the specification, together with the embodiments of the application, to explain the application, and do not constitute a limitation of the application. In the drawings:

[0074] Figure 1 A flowchart of an abnormality detection method for a welded pipe unit based on contrastive learning is provided. DETAILED DESCRIPTION

[0075] The application will now be described in further detail with reference to the drawings. These drawings are simplified schematic diagrams that only schematically illustrate the basic structure of the application, and therefore only show the components related to the application.

[0076] REFERENCE Figure 1 An abnormality detection method for a welded pipe unit based on contrastive learning, characterized by comprising the following steps:

[0077] S1, collect time series data and process data of multiple sensing channels, and timestamp label with a unified time base, output multiple sensing channel data, including line speed, encoder pulse, event flag and event window;

[0078] S2, input the multi-sensing channel data, establish a differentiable time scale correction layer for each sensing channel, map the original time to the corrected time using multiplicative coefficients and additive offsets, and output the corrected time of each sensing channel and the multiplicative coefficients and additive offsets;

[0079] S3, constructing a differentiable mapping from time to cumulative length coordinate with corrected time and linear velocity as inputs, calculating propagation delay according to physical distance between stations, and aligning each sensing channel on a fixed length grid to output cumulative length coordinate, propagation delay and aligned samples on the fixed length grid;

[0080] S4, establishing a neural residual correction module with cumulative length coordinate and propagation delay as main inputs and corrected time, linear velocity, encoder pulse, event flag and event window as auxiliary inputs to correct cumulative length coordinate and propagation delay, and output corrected cumulative length coordinate, corrected propagation delay and alignment confidence;

[0081] S5, taking fixed length grid aligned samples and alignment confidence as inputs and event window as shielding basis, setting cumulative length window on the fixed length grid and performing event window shielding, constructing and outputting positive and negative sample sets for contrastive learning training under the premise of satisfying the advection constraint;

[0082] S6, taking sample sets as inputs, training feature representation network, establishing memory bank and generating abnormal score according to distance or similarity to memory bank, adopting dynamic threshold and aligning calculation window boundary with encoder pulse, and outputting abnormal score, abnormal detection result, station level positioning and dynamic threshold;

[0083] S7, taking alignment confidence, anchor point deviation calculated from time to cumulative length mapping and propagation delay, and abnormal detection and quality inspection results as inputs, fine-tuning multiplicative coefficient and additive offset and writing back to time to cumulative length mapping, purifying memory bank and updating dynamic threshold according to quality inspection results, and outputting updated multiplicative coefficient and additive offset, time to cumulative length mapping, dynamic threshold and memory bank state.

[0084] In the embodiment, one specific implementation of S1 is:

[0085] Uniform time base acquisition and timestamp labeling are adopted, and linear velocity and encoder signals are subjected to anti-interference preprocessing and event window generation to ensure that timestamps of multiple acquisition cards and control units are monotonous and consistent and continuous in start-stop phase, and the collected multiple sensing channels at least include linear velocity time sequence, encoder pulse, cutting control signal and weld temperature signal, and event flags and event windows of start-stop, cutting, joint passing and weld temperature peak are generated under regular triggering, wherein the discrete time grid of uniform time base is defined as: ;

[0086] wherein is the th uniform time base time point, is the start time of uniform time base, is a non-negative integer discrete time index, to unify the sampling interval and to be positive;

[0087] First-order exponential smoothing is performed on the line speed on the unified grid to suppress transient noise and jitter, and the filtering formula is: ;

[0088] wherein is the filtered line speed at time , is the filtered line speed at the last grid time, is the original line speed sample at time , is a smoothing factor between 0 and 1;

[0089] At the same time, the encoder pulse is counted to remove burrs, and the deburring formula is: ;

[0090] wherein is the deburred cumulative pulse count at time , is the cumulative pulse count at the last grid time, and are the original pulse counts at times and , is an indicator function and takes the value 1 when the condition in the brackets is true, otherwise it takes the value 0;

[0091] During sampling, the loss of packets and saturation of each channel are recorded for quality control. The loss of packets record includes the ratio statistics of the expected sample number and the actual received sample number, and the saturation record includes the out-of-boundary markers of the line speed sampling upper and lower thresholds and , the generation of event flags follows the rule trigger and is confirmed by the encoder pulse, the start-stop event is triggered by the line speed crossing the set speed threshold and meeting the minimum duration , the cutting event is triggered by the rising edge of the cutting control signal and is confirmed within the confirmation time window with the minimum number of confirmation pulses , the joint is confirmed by the local extreme value of the weld temperature signal meeting the temperature threshold within the confirmation time window with the minimum number of confirmation pulses , all events are recorded with start and end time stamps as event windows and are output together with unified time base data, and the event window is uniformly represented as: ;

[0092] wherein is the a time window of an event, with the start and end timestamps of the window respectively and the window length can be denoted as and keep the continuity of the timestamps during the start-stop phase.

[0093] In this embodiment, one specific implementation of S2 is:

[0094] A differentiable timestamp correction layer is established for each sensing channel to map the original time under the unified time base to the corrected time and impose a smoothing constraint in the cumulative length dimension, while combining the endpoint anchors formed by the encoder and event flag to impose a consistency constraint and adopt a causal strategy for online updating. To balance the implementability and stability, this embodiment sets the correction mapping to an affine form that varies slowly with the cumulative length, specifically: ;

[0095] where denotes the corrected time of channel at the unified time base index , denotes the original time of the unified time base, denotes a channel in the channel index set, is a non-negative integer time index, denotes the multiplicative coefficient of channel at the cumulative length position , denotes the additive offset of channel at the cumulative length position , denotes the cumulative length coordinate obtained by numerical integration of the filtered line speed of step S1 and serves as the independent variable of the smoothing constraint;

[0096] To make the corrected time consistent with the original record at the endpoint anchors and suppress rapid fluctuations along the cumulative length, endpoint consistency least squares and cumulative length smoothing penalty are jointly used, and the objective function is written as:

[0097] ;

[0098] where is the total cost function, is the endpoint anchor set of channel and is composed of the encoder pulse time and the start-stop boundary of the event flag, is the time position of the th needle point, is the weighting coefficient of the th needle point and can be taken as zero within the event window to achieve freezing, is channel the correction time at the needle point, the original record time at the needle point, and the smoothing penalty coefficient of the multiplicative coefficient and the additive offset, the discrete gradient operator of the cumulative length , the square of the two-norm to measure the amplitude of the change along the cumulative length direction;

[0099] To ensure monotonicity and physical reasonableness, box constraints are imposed on the parameters: ;

[0100] where and are the lower and upper limits of the multiplicative coefficient, and are the lower and upper limits of the additive offset, and the above constraints can ensure that the correction mapping is monotonically increasing over time ;

[0101] The homologous constraints of the parameters are realized by sharing parameters within the group or penalizing deviations within the group. The groups are divided according to the acquisition card or control unit and ensure that the and of the channels in the same group are consistent or their deviations are limited;

[0102] The online update follows a causal strategy, that is, at time only the data increment not later than is used to solve or small-step gradient update, and the update step is frozen or significantly reduced within the event window , while when the anchor point falls into the event window, the corresponding weight is set to avoid transient interference, and finally the channel-level correction time sequence that satisfies the endpoint consistency, cumulative length smoothness, homologous consistency and monotonicity is obtained .

[0103] In this embodiment, one specific implementation of the S3 is:

[0104] Based on the correction time and the filter line speed, a differentiable mapping from time to cumulative length that satisfies the advection constraint is constructed, the encoder pulse is used as the boundary condition for calibration, and after the propagation delay is calculated according to the physical distance between stations, the samples of each sensing channel are uniformly aligned on a fixed length grid, where the cumulative length mapping adopts a speed-driven integral form: ;

[0105] where denotes the cumulative length coordinate at the correction time point , denotes the th correction time point ( time index of non-negative integer, denotes the integral start time, denotes the length calibration coefficient, denotes the filtered linear velocity at time and denotes the differential element of time;

[0106] The encoder pulse is taken as the boundary condition to ensure the consistency of the cumulative length and the pulse count, and the boundary constraint is: ;

[0107] wherein denotes the time when the th encoder pulse occurs, denotes the reference pulse time, denotes the pulse index, denotes the nominal length increment corresponding to a single pulse;

[0108] To avoid the transient interference of events such as start-stop, cutting, joint passing, etc., the data falling into the event window is not involved in the calibration coefficient estimation and boundary condition update, and only the data in the stable interval is reserved for mapping and alignment;

[0109] The propagation delay is solved in a length-conserved manner according to the physical distance between stations, and the reference station is denoted as , and any station is denoted as , and the delay integral equation at the modified time point is: ;

[0110] wherein denotes the delay from the reference station to the station , denotes the physical distance of the station relative to the reference station, denotes the time independent variable of integration;

[0111] The cross-station unified alignment is completed on a fixed length grid, and the starting point of the grid is aligned with the nearest encoder pulse and defined as: ;

[0112] wherein denotes the th grid position, denotes the index of the nearest encoder pulse to the current processing interval, denotes the grid index, denotes the preset grid length interval, and each channel signal is mapped to in the cumulative length coordinate in a linear or spline interpolation manner, and the cross-station alignment is successfully determined by the physical tolerance, and if the grid position of the reference station is considered to correspond to the same pipe section as the grid position of the station ; ;

[0113] wherein denotes the i-th grid position of the station, denotes the length tolerance threshold for alignment, the final output is the cumulative length coordinate satisfying the advection constraint and the encoder boundary condition, the propagation time delay calculated by the distance between stations, and the cross-station aligned sample on the fixed length grid.

[0114] In this embodiment, one specific implementation of the S4 is:

[0115] Based on the uncorrected cumulative length and the propagation time delay on the cross-station fixed length grid, a causal neural residual correction module is constructed and only uses data no later than the current correction time for inference and update, freezes or significantly reduces the update step within the event window to suppress the transient interference brought by start-stop, cutting and joint passing, and the residual compensation adopts an additive form to maintain interpretability and physical consistency: ;

[0116] wherein denotes the corrected cumulative length of the i-th grid position of the station, denotes the uncorrected cumulative length of the corresponding position, denotes the residual function of the cumulative length direction and takes the grid position as the independent variable and is controlled by the parameters and denotes the corrected propagation time delay of the station at position , denotes the uncorrected propagation time delay, denotes the residual function of the propagation time delay, denotes the station index, denotes the i-th fixed length grid position on the cumulative length coordinate axis, is the grid index integer, denotes the set of cross-station shared parameters to represent mechanical coupling, denotes the specific parameter set of the station to compensate for the station difference.

[0117] ​​​​​​​​To avoid over-correction and ensure the smoothness of cumulative length direction, the strong amplitude and second-order difference smoothness constraints are imposed on the residual: ;

[0118] wherein denotes absolute value, and are the amplitude upper bounds of cumulative length residual and time delay residual, denotes the smoothness penalty term, and are the smoothness weights of cumulative length residual and time delay residual, denotes the second-order difference operator along the cumulative length grid, and is or collectively;

[0119] To quantify the alignment quality and provide weights for subsequent sample basket selection and detection, the alignment confidence is defined as: ;

[0120] wherein denotes the alignment confidence of the station at position , denotes the exponential function, and are the penalty coefficients of cumulative length difference and time delay difference, is the penalty coefficient of adjacent grid consistency, denotes the amplitude of cumulative length difference before and after correction, denotes the amplitude of propagation time delay difference before and after correction, denotes the second-order inconsistency degree of cumulative length after correction on adjacent grids, and monotonically decreases with the increase of the inconsistency degree;

[0121] In summary, through the synergy of sharing and specific parameters, the causal update strategy and the amplitude-smoothness dual constraints, stable and reliable correction results are obtained, and finally the corrected cumulative length , the corrected propagation time delay and the alignment confidence are output.

[0122] In the embodiment, one specific implementation of the S5 is:

[0123] Taking the aligned samples and the alignment confidence on the fixed-length grid as input, and taking the event window as the shielding basis, a sliding window is set on the cumulative length axis, stable segments are screened, and a physically consistent comparative learning sample set is constructed, and the window start and end and center are defined as: ;

[0124] wherein ​Indicates the first The starting position of each cumulative length window, Indicates the end position of this window. Indicates the center position of the window. Indicates a fixed-length raster at the index Cumulative length coordinates at the location For window indexing, For the raster index corresponding to the step size, The length of the window;

[0125] Map the event window to the cumulative length axis and buffer and mask the events before and after it. The event length range is denoted as: ;

[0126] The shielding indication is: for any ;

[0127] in Indicates the first The interval of an event on the cumulative length axis and These represent the start and end timestamps of the event, respectively. Represents the mapping from time to cumulative length. Indicates the first Is the window available? This indicates an indicator function that takes the value 1 if the condition within the parentheses is true, otherwise takes the value 0. and Representing the start and end positions of the event length interval, respectively. This indicates the length of the front and rear shields on the cumulative length axis;

[0128] To quantify window alignment quality, the set of raster indexes contained in the window is denoted as: Define window confidence. ;

[0129] in Indicates the first The set of raster indexes within a window Indicates the number of elements in the set, Indicates workstation index, Indicates workstation At grid position Alignment confidence, Indicates the first A fixed-length grid position, Represents a set of workstations;

[0130] The window filter set is denoted as: ;

[0131] in Indicates workstation The set of available windows This represents the alignment confidence threshold. Positive samples, under the premise of satisfying the advection constraint, are selected from the same cumulative length window and include samples from the same channel across time periods or across line velocities, samples from different channels, and samples from across workstations. Among them, positive samples from across workstations are constrained by the physical tolerance of the propagation delay difference, and the difference between the average window delay and the threshold is denoted as: ;

[0132] in Indicates workstation In the Average corrected propagation delay per window Indicates workstation In the grid Corrected propagation delay, Indicates workstation With workstation The magnitude of the average propagation delay difference in the window, Indicates another workstation and ;

[0133] Encoder pulses are used to confirm window boundaries to ensure physical consistency. Negative samples are selected as candidate windows within their cumulative length neighborhood, with distances weighted in a monotonically decreasing manner, and low-confidence and masked windows are removed. The neighborhood and weights are defined as follows: ;

[0134] in Indicates the first The cumulative length neighborhood of each window Indicates the candidate negative sample window index, Represents the neighborhood radius, Represents the sampling weights before normalization. Indicates the attenuation coefficient of the weight;

[0135] The actual sample set is denoted as: In the cross-station negative samples, priority is given to window pairs that do not meet the propagation delay tolerance. ;

[0136] in The physical tolerance threshold representing the propagation delay difference, as mentioned above These parameters are set based on the physical distance to the workstation, line speed statistics, and encoder calibration, and can be fine-tuned online according to confidence level statistics. The dynamic threshold can be updated using quantiles. ;

[0137] in Indicates the new confidence threshold, Quantile function The quantile level is represented, and the final output is the positive and negative sample set for training and the reference of the window station to support subsequent partition management and dynamic threshold calculation.

[0138] In this embodiment, one specific implementation of S6 is:

[0139] The positive and negative sample sets are used as training inputs and are weighted by the alignment confidence. First, the normalized feature representation is generated in the fixed-length grid and is summarized by the cumulative length window. Then, the memory bank is established according to the grid and process state partition. The abnormal score is given based on the similarity to the memory bank. The dynamic threshold that aligns the statistical boundary with the length is used to realize online detection and station-level positioning. The normalized embedding is denoted as:

[0140] wherein denotes the normalized feature vector of the sample , denotes the feature representation network, denotes the two-norm, denotes the embedding dimension, and during training, the samples that do not meet the advection constraint or have a confidence lower than the threshold are down-weighted or removed to ensure physical consistency.

[0141] The window-weighted summary representation is:

[0142] wherein denotes the summary embedding of the station in the first cumulative length window, denotes the grid index set covered by the window , denotes the alignment confidence of the station at the grid position , denotes the normalized embedding of the station at the position , denotes the first fixed-length grid position, and the similarity is defined as the inner product:

[0143] wherein denotes the normalized inner product similarity of the vector and the vector , denotes the normalized embedding;

[0144] The abnormal score is given by the maximum similarity of the memory bank partition:

[0145] wherein ​​​​​representing the abnormal score of a sample representing the abnormal score of a sample representing the work station and the process state corresponding memory partition, representing the partition in which the embedded representative is embedded and the partition is managed according to a fixed length grid and the process state;

[0146] The dynamic threshold value is aligned with the statistical window boundary according to the encoder pulse and is updated by quantile number: ;

[0147] wherein representing the work station and the process state at time the abnormal threshold value, representing the quantile function, representing the quantile level, representing the abnormal score of a historical sample, representing the time stamp of the score, and respectively represent the occurrence time of the first and the second encoder pulse to align the statistical boundary with consistent length, representing the statistical window width defined by pulse count;

[0148] The online detection is determined by the threshold value to output: ;

[0149] wherein representing the abnormal determination of a sample , representing the indicator function and taking 1 when the condition in the bracket is true, otherwise taking representing the time stamp of a sample ;

[0150] And the work station level positioning is given by combining the cumulative length coordinate and the correction propagation delay: ;

[0151] wherein representing the positioning cumulative length coordinate of a sample at a work station , representing the differentiable mapping of time to cumulative length, representing the correction propagation delay of a work station so as to realize the unified length scale positioning of the same pipe section at different work stations.

[0152] In the embodiment, one specific implementation of the S7 is:

[0153] Based on sliding statistics, alignment confidence and anchor point deviation trigger causal online fine-tuning and closed-loop updates of thresholds and memory. Simultaneously, updates are frozen or downweighted within the event window while maintaining the monotonicity of the time-to-cumulative length mapping. The trigger criterion is either insufficient weighted confidence or excessive endpoint anchor point deviation. ;

[0154] in The indicator value for whether to trigger fine-tuning For indicator functions, For the most recent consecutive The index set of cumulative length windows For workstation In the window Average alignment confidence For trigger threshold, For the encoder pulse index set participating in the verification, For parameters Mapping from the determined time to the cumulative length For the first encoder pulse timing, For the nominal length increment of a single pulse, For anchor point deviation tolerance;

[0155] when The time-matter strategy uses only data no later than the current time to perform small-step incremental updates on the multiplicative and additive parameters of the channel differentiable timescale correction layer, employing an objective and update law with smoothing penalty and box-constrained projection:

[0156] ;

[0157] in For the cost of the alliance, For the multiplicative and additive parameter vectors of aggregation, When the pulse time falls within the event window, the weight of the pulse point can be set to 0 to freeze the update. For the smoothing weights of the cumulative length dimension, For discrete gradient operators along the cumulative length, For the second norm, For the updated parameters, For step size, For the projection operator to the feasible region, Constrain the feasible region for the parameter box and for the channel Multiplicative coefficient function With additive offset function Apply upper and lower bounds and ( For channel indexing, for parameter upper limit) ;

[0158] Homogeneous constraints are implemented in practice by sharing within group or bias penalty to ensure parameters consistency within the same acquisition card / control unit or bias limited, event window is denoted as and freeze or significantly reduce update step within this interval to suppress transient impact;

[0159] Fine-tuned results are written back to mapping After keeping monotonic and encoder boundary approximation satisfied, dynamic threshold re-estimation is performed by quantile method in statistical window aligned with encoder pulse according to distribution of historical anomaly score: ;

[0160] where is the station and process state at time , dynamic threshold, is the quantile function, is the quantile level, is the anomaly score of historical sample, is its timestamp, and is the statistical window boundary aligned with pulse count, is the pulse count window width.

[0161] Meanwhile, according to the quality inspection results, the memory bank is back-labeled and purified in cumulative length, and samples that are judged as abnormal or unqualified and near the event interval are removed and their neighborhood is down-weighted to improve robustness;

[0162] If there are signs of non-recovery or instability, rollback is performed, and the rollback criterion is:

[0163] ;

[0164] where is the indication quantity of whether to rollback, is the recovery threshold, is the recovery tolerance, and are the average anomaly scores in the same length aligned statistical window before and after updating, is the allowed average upward movement amplitude, if rollback is triggered, the last stable parameter set is restored and the dynamic threshold and memory bank partition are rolled back synchronously, and finally the updated or rolled back parameters , mapping , dynamic threshold and memory bank state are output.

[0165] The above merely describes preferred specific embodiments of the present application, but the protection scope of the present application is not limited thereto, and any person skilled in the art, according to the technical solution and inventive concept of the present application, makes equivalent replacement or change within the technical range disclosed by the present application, which should be covered within the protection scope of the present application.

Claims

1. A contrast learning-based pipe welder unit anomaly detection method, characterized in that, The method comprises the following steps: S1, collecting time series data of multiple sensing channels and process data, and time stamping with a unified time base, outputting multiple sensing channel data, including line speed, encoder pulse, event flag and event window thereof; S2, taking the multiple sensing channel data as input, establishing a differentiable time correction layer for each sensing channel, mapping the original time to the corrected time by using multiplicative coefficients and additive offsets, and outputting the corrected time of each sensing channel and the multiplicative coefficients and additive offsets thereof; S3, taking the corrected time and line speed as input, constructing a differentiable mapping from time to cumulative length coordinates, calculating the propagation delay according to the physical distance between stations, and aligning each sensing channel on a fixed length grid, outputting the cumulative length coordinates, the propagation delay and the aligned samples on the fixed length grid; S4, taking the cumulative length coordinates and the propagation delay as the main input, and taking the corrected time, the line speed, the encoder pulse, the event flag and the event window as auxiliary input, establishing a neural residual correction module to correct the cumulative length coordinates and the propagation delay, and outputting the corrected cumulative length coordinates, the corrected propagation delay and the alignment confidence; S5, taking the fixed length grid alignment samples and the alignment confidence as input, and taking the event window as a shielding basis, setting a cumulative length window on the fixed length grid and performing event window shielding, constructing and outputting a set of positive and negative samples for comparative learning training under the premise of satisfying the constraint of the advection partial differential equation; S6, taking the sample set as input, training a feature representation network, establishing a memory bank and generating an anomaly score according to the distance or similarity to the memory bank, using a dynamic threshold and calculating the window boundary alignment with the encoder pulse, and outputting the anomaly score, the anomaly detection result, the station level positioning and the dynamic threshold; S7, taking the alignment confidence, the anchor point deviation calculated from the mapping from time to cumulative length coordinates and the propagation delay, and the anomaly detection and quality inspection results as input, fine-tuning the multiplicative coefficients and additive offsets and writing them back to the mapping from time to cumulative length, purifying the memory bank and updating the dynamic threshold according to the quality inspection result, and outputting the updated multiplicative coefficients and additive offsets, the mapping from time to cumulative length, the dynamic threshold and the memory bank state.

2. The abnormality detection method for a welded tube mill set based on contrastive learning according to claim 1, characterized in that, Step S1 is specifically: The multiple sensing channels at least include line speed time series, encoder pulses and event flags for marking start-stop, cutting, joint passing and weld temperature peak; The time series data and process data of the multiple sensing channels are time stamped with a unified time base, so that the time stamps of each acquisition card and the control unit are monotonous and consistent and continuous in the start-stop stage; The line speed time series and the encoder pulses are subjected to anti-interference filtering and de-bouncing processing, and the packet loss and saturation state are recorded; The event flags are generated by rules, wherein the start-stop event flag is triggered by the line speed time series crossing the set speed threshold and meeting the set duration, the cutting event flag is triggered by the rising edge of the cutting control signal and confirmed by the encoder pulse count, the joint passing event flag is triggered by the peak of the weld temperature signal exceeding the set temperature threshold and confirmed by the encoder pulse, and the weld temperature peak event flag is triggered by the local extreme value of the weld temperature signal meeting the set condition. The event mark records the event window with a start time stamp and an end time stamp, and is output together with the data marked with the unified time base.

3. The abnormality detection method of a welded tube mill set based on contrastive learning according to claim 1, characterized in that, Step S2 is specifically: The multiplicative coefficient and the additive offset satisfy the homogeneity constraint and the smoothing constraint in the cumulative length dimension, and are constrained by the encoder pulse and the event mark as end-point anchor points, and the update of the multiplicative coefficient and the additive offset is reduced or frozen in the event window; The original time of each sensing channel is mapped to the corrected time through a linear combination of the multiplicative coefficient and the additive offset; The multiplicative coefficient and the additive offset are determined by jointly considering the least square objective of the end-point anchor consistency and the smoothing penalty in the cumulative length dimension, the end-point anchor points are determined by the time points given by the encoder pulse and the event mark, and the corrected time is consistent with the original record at the end-point anchor points; The smoothing penalty takes the cumulative length as the independent variable, limits the change rate of the multiplicative coefficient and the additive offset along the cumulative length, and sets upper and lower limits for the multiplicative coefficient and the additive offset respectively; The homogeneity constraint uses the in-group parameter sharing or in-group parameter deviation penalty to keep the multiplicative coefficient and the additive offset of the channels of the same acquisition card or the same control unit consistent or the deviation limited; The update of the multiplicative coefficient and the additive offset adopts a causal strategy, and only uses data no later than the current time for update, and freezes or reduces the update step in the event window.

4. The abnormality detection method of a welded tube mill set based on contrast learning according to claim 1, characterized in that, Step S3 is specifically: The mapping satisfies the constraint of the advection partial differential equation and takes the encoder pulse as the boundary condition; A differentiable mapping from the corrected time to the cumulative length coordinate is established by numerical integration using the corrected time and the linear speed as input, so that the cumulative length monotonically increases with the corrected time and satisfies the constraint of the advection partial differential equation; The calibration coefficient of the mapping is determined according to the length calibration of the encoder pulse, and the encoder pulse is taken as the boundary condition to calibrate the value of the cumulative length at the pulse time; The data falling into the event window is not involved in the estimation of the calibration coefficient and the update of the boundary condition, so as to avoid transient interference on the mapping stability; The propagation time delay is calculated according to the physical distance between each station, and the length offset of each station is determined, a reference station is selected, and the difference between the cumulative length of the same pipe segment at any station and the cumulative length at the reference station is equal to the corresponding physical distance; The samples of each sensing channel are uniformly aligned on a fixed length grid, the fixed length grid uses a preset length interval, the grid starting point is aligned with the nearest encoder pulse, and linear interpolation or spline interpolation is used to complete the gridding and cross-station alignment; When the cumulative length difference of different station sample windows falls within the set physical tolerance range, it is determined that the same pipe segment is successfully aligned, and the cumulative length coordinate, the propagation time delay and the aligned sample on the fixed length grid are output.

5. The abnormality detection method of a welded tube mill set based on contrastive learning according to claim 1, characterized in that, Step S4 is specifically: The parameter update and inference of the neural residual correction module adopt a causal strategy, and amplitude constraint and smoothing constraint are applied in the cumulative length dimension, and the alignment confidence is calculated. The neural residual correction module with a causal structure is established by taking the accumulated length coordinate and the propagation time delay as main inputs, and taking the correction time and the line speed, the encoder pulse, the event flag and the event window as auxiliary inputs; The module performs residual compensation on the accumulated length coordinate and the propagation time delay, and performs inference and parameter updating only based on data no later than the current correction time, and reduces the parameter update step or freezes the parameters within the event window; The residual is set as the independent variable of the accumulated length, the amplitude constraint is set to limit the upper and lower limits of the absolute value of the residual, and the smoothing constraint is realized by the second-order difference penalty in the accumulated length dimension to limit the rapid change of the residual; The same parameters are used in the shared parameter part to reflect the mechanical coupling, and the station-specific parameters are used in the independent parameter part to compensate for the station differences for different stations; The alignment confidence is calculated based on the difference between the accumulated length coordinates before and after correction, the difference between the propagation time delays before and after correction, and the consistency index of adjacent length grids, and monotonically decreases with the increase of the above differences; The corrected accumulated length coordinate, the corrected propagation time delay and the alignment confidence are output.

6. The abnormality detection method of a welded tube mill set based on contrastive learning according to claim 1, characterized in that, Step S5 specifically comprises: Taking the alignment samples on the fixed length grid and the alignment confidence as inputs, and taking the event window as a shielding basis, an accumulated length window determined by the window length and the step is set on the fixed length grid, and the accumulated length window falling into the event window and the preset length range before and after the event window is shielded; According to the alignment confidence threshold, only the accumulated length window with an alignment confidence not lower than the threshold is retained for sample construction; Under the premise of satisfying the advection constraint, positive samples are constructed, the positive samples are selected from the same accumulated length window, including samples from the same sensing channel at different time periods or different line speeds, and samples from different sensing channels or different stations, wherein the cross-station positive samples limit the corresponding propagation time delay difference to be within a set physical tolerance range, and the start and end positions of the accumulated length window are confirmed by the encoder pulse; When constructing negative samples, candidate windows are selected in the accumulated length neighborhood, sampling weights are set according to the accumulated length distance, the weights monotonically decrease with the increase of the accumulated length distance, and windows with an alignment confidence lower than the threshold and windows shielded by the event window are removed; For cross-station negative samples, windows with a propagation time delay difference not satisfying the physical tolerance range or an accumulated length offset exceeding a set neighborhood radius are preferentially selected; The accumulated length neighborhood radius, the window length, the step and the alignment confidence threshold are set according to the physical distance between stations, the line speed statistics and the encoder pulse calibration, and are fine-tuned online according to the statistical results of the alignment confidence in the production process; The positive sample and negative sample sets for training are output.

7. The abnormality detection method of a welded tube mill set based on contrastive learning according to claim 1, wherein, Step S6 specifically comprises: Taking the positive sample and negative sample sets as training inputs, and taking the alignment confidence as the sample weight to train the feature representation network; The similarity calculation is performed in the fixed length grid, the normalized feature representation is used, and the accumulated length window is used for aggregation, the physically consistent similarity comparison loss is used to increase the similarity of the positive samples and reduce the similarity of the negative samples; To ensure physical consistency, constraints are added to the loss to keep the same cumulative length window consistent in different line speeds and different stations, and samples that do not meet the advection partial differential equation constraint or have a confidence level below the threshold are reduced in weight or removed; A memory bank is established and managed in fixed length grids and process states, and the samples in the bank meet the alignment confidence threshold and do not fall into the event window; The distance or similarity between the test sample and the memory bank is calculated in the corresponding partition, and an abnormal score is generated. The abnormal score increases as the distance increases or the similarity decreases; Dynamic thresholds are set for different line speeds and different stations. The dynamic threshold is updated adaptively according to the statistics of the memory bank partition and the alignment confidence, and the calculation window of the dynamic threshold is boundary-aligned with the encoder pulse; Finally, the abnormal detection result is output based on the abnormal score and the dynamic threshold, and the station-level positioning is generated based on the cumulative length coordinate and the propagation time delay.

8. The abnormality detection method of a welded tube mill set based on contrast learning according to claim 1, characterized in that, Step S7 is specifically: The alignment confidence of the sliding statistics and the endpoint anchor point deviation calculated from the mapping of the time to the cumulative length and the propagation time delay are used as the triggering basis, wherein the triggering condition of the alignment confidence is that the alignment confidence is below the set threshold in a continuous preset number of cumulative length windows in the fixed length grid, and the triggering condition of the endpoint anchor point deviation is that the absolute value of the difference between the cumulative length at the endpoint anchor point and the nominal cumulative length determined by the encoder pulse exceeds the set tolerance; When either triggering condition is met, the multiplicative coefficient and the additive offset are fine-tuned online using a causal strategy. The update process uses incremental least squares or small step gradient updates with a cumulative length dimension smoothing penalty, and complies with parameter upper and lower limits and homology constraints. The update step is frozen or reduced within the event window; After fine-tuning, the updated results are written back to the mapping of time to cumulative length, so that the mapping satisfies the boundary condition at the endpoint anchor point and maintains monotonicity. Meanwhile, based on the distribution of the abnormal score in each fixed length grid partition, the dynamic threshold is re-estimated using a quantile strategy; According to the quality inspection results, the samples are labeled on the cumulative length coordinate, the memory bank is purified, the samples judged as abnormal or unqualified are removed, and the samples near the event window are reduced in weight or removed; If the fine-tuned alignment confidence does not recover to above the threshold or the endpoint anchor point deviation still exceeds the tolerance, or there are signs of instability in the overall upward shift of the abnormal score, roll back to the last stable parameter set, and restore the dynamic threshold and memory bank partition before rollback; The output includes the multiplicative coefficient and the additive offset, the mapping of time to cumulative length, the dynamic threshold, and the memory bank state after updating or rollback.

Images