PVC guardrail weight on-line monitoring system and method

Abstract

The application discloses a PVC guardrail weight online monitoring system and method, and relates to the technical field of plastic extrusion monitoring.The system comprises the following steps: constructing a piece event window with cutting triggering as an anchor point, injecting a small disturbance inversion arrival path kernel, and establishing a version applicable domain with a formula and a mold;calculating an upstream signal to obtain a forward prediction by replaying the arrival path kernel, forming a reverse prediction by using the geometric chain unit length mass, and screening high credible samples across the link gate;implementing dynamic weighing waveform dynamic weighing gate for the symmetric weighing waveform, driving the bias and drift double state self-learning by the high credible samples;mapping the deviation to the feeding, traction, mold temperature and diameter vacuum three knobs to solve the minimum change, performing geometric priority and short window verification, and realizing the minimum recall boundary by the protection state machine and MES audit chain.The system realizes cross-formula rapid copying, reduces misadjustment and shutdown, considers geometric stability and weight consistency, and improves the tracing and auditing capabilities.

Description

Technical Field

[0001] This invention relates to the field of plastic extrusion monitoring technology, specifically to an online weight monitoring system and method for PVC guardrails. Background Technology

[0002] The online extrusion production of PVC guardrails is characterized by continuous operation, long cycle time, and strong coupling: the melt self-feeding-barrel-die head-sizing and cooling-traction-cutting are all transmitted in one line. The slight fluctuation in traction speed, the slow drift of sizing vacuum and mold temperature, as well as the batch and seasonal temperature and humidity changes of the formulation, will introduce significant viscoelastic memory and arrival time delay.

[0003] Current practices largely rely on end-point weighing and SPC control, or trigger parameter tuning using single-channel thresholds for feeding signals / traction speeds. They generally lack a uniquely identifiable time-length mapping per piece, failing to transmit upstream perturbations to the cutting end to form a verifiable arrival path. Furthermore, conventional moving averages / low-pass filtering easily mix mechanical noises such as cutting impacts and conveyor vibrations with actual quality changes, leading to sample contamination and learning distortion. More critically, current correction methods often rely on empirical approaches, applying a one-size-fits-all approach to feeding, traction, or mold temperature without setting geometrically prioritized hard constraints, frequently amplifying fluctuations in profile outer diameter / wall thickness. The lack of cross-link verification and traceable auditing results in poor reproducibility after time synchronization drift, seasonal drift, and mold changes, further increasing out-of-tolerance rates, rework, scrap, and excessively wide recall boundaries, leading to operational costs.

[0004] To address the issues of continuous processes and multi-source disturbances, as well as signal path time-scale mismatches, cross-component energy leakage, and channel common-cause errors, the core technical problem this invention aims to solve is: without interrupting the cycle time, to construct an online weight monitoring and correction method using a unique event window per component as the carrier, an arrival path kernel as the physical kernel, dual-track prediction + gating as the truth-detection mechanism, bias / drift dual states as the learning object, and geometry priority + minimum modification of three knobs as the execution strategy. This method ensures that upstream quality changes can be accurately recalculated to the cutting end and correlated with the unit length quality of independent geometric links. The quantity forms a cross-verification; when shift changes, formula / mold changes or seasonal changes cause working condition deviations, the contaminated samples can still be screened out at the part level by restricted perturbation identification, cross-link gating, consistency index and time synchronization, avoiding mislearning and over-parameter tuning; at the same time, when continuous anomalies accumulate, the protection state machine triggers deceleration, stricter rejection and learning freeze, and writes the entire link of prediction-actual measurement-gating-correction-state transition into MES to form a chain audit, so as to ensure that the minimum verifiable and rollback parameter tuning combination is output without destroying the key geometry and to achieve cross-production line replication. Summary of the Invention

[0005] (a) Technical problems to be solved To address the shortcomings of existing technologies, this invention provides an online weight monitoring system and method for PVC guardrails. It constructs a per-piece event window using cutting triggers as anchor points, injects small disturbances to invert the arrival path kernel, and establishes a version applicability domain with the formula and mold. Forward predictions are obtained by recalculating upstream signals using the arrival path kernel, and inverse predictions are formed using the mass per unit length of the geometric chain. Cross-link gating filters high-confidence samples are used. Dynamic checkweighing gating is implemented on the symmetrical weighing waveform, and high-confidence samples drive bias and drift dual-state self-learning. Deviations are mapped to the three knobs of feeding, traction, mold temperature, and sizing vacuum to solve for minimum changes, performing geometric priority and short-window verification, and minimizing the recall boundary using a protected state machine and MES audit chain. This enables rapid cross-formula replication, reduces misadjustments and downtime, and solves the technical problems described in the background art.

[0006] (II) Technical Solution To achieve the above objectives, the present invention provides the following technical solution: Methods for online monitoring of the weight of PVC guardrails include, Using the cutting trigger as the anchor point, a unique part number is generated in conjunction with the length encoding. A part-by-part event window is constructed, small disturbances are injected and the response at the cutting end is collected. The arrival path kernel from upstream to the cutting end is obtained through weighted inversion, and a version applicable domain is established with the formula and mold. Based on the arrival path kernel, the upstream quality change is recalculated into the event window to obtain the part-level positive prediction. At the same time, the inverse prediction is obtained by using the unit length mass and the length of the part. The consistency index is calculated and combined with the time offset and the weighing side warning to form a cross-link gating conclusion. Dynamic check-repeat gating is implemented for symmetrical re-waveforms. Low-confidence samples are re-weighed in parallel without being included in the learning process, while high-confidence samples drive bias and drift dual-state self-learning. The learning gain and forgetting factor are adjusted according to the gating score and working condition distance to generate part-level release or rejection conclusions and states. The part weight deviation is mapped to the feed response, traction fine-tuning, mold temperature and sizing vacuum three knobs to solve the minimum change combination, and geometric priority constraints and short window verification are performed. If it does not converge, it will back down and trigger the protection state machine. The entire process is written into MES with part number to form audit and recall boundaries.

[0007] Furthermore, the item-by-item event window only allows time boundary fine-tuning between two adjacent items according to a preset fine-tuning limit and prohibits overlap; all items in the same shift form a directed event chain according to the cutting trigger order for beat self-checking and time synchronization verification; the unique item number is generated by combining the cutting trigger time, item length and shift identifier and runs through subsequent steps.

[0008] Furthermore, a small perturbation with zero integral and limited amplitude and energy is injected under stable operating conditions, and the response at the cutting end is recorded; the arrival path kernel is obtained by weighted total variation canonical inversion, and the arrival path kernel, formula identifier, mold family and environmental parameters are used to establish a version applicable domain and archive it; The weighting function is composed of time decay, beat consistency and waveform edge indication. The version application domain is defined by the working condition vector and the metric matrix and is used for production change review.

[0009] Furthermore, the reverse prediction employs the fusion of the geometric chain main channel and the mass flow ratio backup channel. The geometric chain calculates the mass per unit length using online geometry and temperature-corrected density, while the backup channel estimates the mass per unit length using robust truncated mean. The fusion weight is determined by the geometric measurement confidence level. The two channels are independent of each other on the sensing path and time-aligned to the event window. The mass per unit length and the length of the component are accumulated in the same window to form an inverse prediction and participate in subsequent gating.

[0010] Furthermore, cross-link gating includes: constructing a symmetric consistency index composed of forward and reverse predictions, introducing weighing-side early warning and time offset constraints, and locating synchronous offset within a preset search radius using cross-correlation peak values, which can be refined by interpolation. The gating thresholds include a consistency threshold, a warning amplitude threshold, and a time offset threshold. All thresholds are archived at the version level and loaded with each shift. When any constraint is not met, a downgrade is marked and a light correction and synchronization check are triggered.

[0011] Furthermore, the dynamic checkweighing gating includes: constructing a waveform reliability index obtained by monotonically mapping the ratio of platform coverage to change intensity within the event window, and combining its consistency index and reweighing consistency in a convex manner to form a gating score; The re-weighing channel is implemented through a bypass weighing platform and decoupled from the main cycle. The re-weighing label is associated with the gating score and stored. When re-weighing is not triggered, neutral consistency is assigned to maintain the fusion scale.

[0012] Furthermore, the bias and drift dual-state update is driven by the learning gain obtained from the gating score mapping. The drift channel is coupled with the operating condition distance of the version application domain to establish a forgetting factor. When the gating score does not reach the threshold or the time offset exceeds the limit, the update is frozen. The learning gain is given by the monotonic mapping and is constrained by the gating threshold. The forgetting factor decreases monotonically with the working condition distance and only acts on the drift channel. Measurement and traceability records are retained during the freeze period. The output includes component-level release or rejection conclusions and dual-state parameters.

[0013] Furthermore, the part weight deviation is mapped to the three knobs of feeding response, traction fine adjustment, mold temperature and sizing vacuum according to the source weight. The minimum change combination is obtained according to the weight matrix and quality sensitivity, and projected to the geometric zero influence subspace through the geometric sensitivity matrix. The upper limit, lower limit and step size of the three knobs are constrained by the equipment file. When the geometric priority constraint is activated, only slow timescale compensation on the feeding side is allowed. The verification window covers the next two parts and records the cause code.

[0014] Furthermore, the protection state machine uses the risk index composed of amplitude, phase, low confidence ratio and trend, as well as the sliding window batch risk, as trigger quantities, and adopts double threshold hysteresis to determine deceleration, stricter rejection, pause learning and rollback; The predictions, measurements, gating, corrections, and status are written to the manufacturing execution system on a piece-by-piece basis and linked by hashes to form an audit chain and recall boundary. The fields include part number, event window, dual-track prediction, time offset, gating score, execution modification, and status.

[0015] PVC fence weight online monitoring system, including, The event modeling module uses the cutting trigger as the anchor point, generates a unique part number with the length encoding, constructs a per-part event window, injects small disturbances and collects the response from the cutting end, obtains the arrival path kernel from upstream to the cutting end through weighted inversion, and establishes a version applicable domain with the formula and mold. The dual-track gating module recalculates the upstream quality change to the event window based on the arrival path kernel to obtain a part-level positive prediction. At the same time, it obtains an inverse prediction based on the unit length mass and the length of the part, calculates the consistency index, and combines the time offset and the weighing side warning to form a cross-link gating conclusion. The checksum learning module implements dynamic checksum gating on the re-counting waveform. Low-confidence samples are re-weighed in parallel and not included in the learning process. High-confidence samples drive bias and drift dual-state self-learning. The learning gain and forgetting factor are adjusted according to the gating score and working condition distance to generate part-level release or rejection conclusions and states. The deviation correction and protection module maps the part weight deviation to the minimum change combination of the three knobs of feeding response, traction fine adjustment, mold temperature and sizing vacuum according to the source. It performs geometric priority constraints and short window verification. If it does not converge, it backs down and triggers the protection state machine. The entire process is written to the MES with the part number to form the audit and recall boundary.

[0016] (III) Beneficial Effects This invention provides an online weight monitoring system and method for PVC guardrails, which has the following beneficial effects: Using the per-piece event window as a carrier, and in conjunction with the arrival path core and version applicable domain, a transmission baseline is established to eliminate cross-piece aliasing and timescale mismatch, so that upstream changes form a unique mapping and traceability at the cutting end; during production change and seasonal changes, version review is used to achieve verification and safe rollback, improving the replicability across formulas and molds.

[0017] By combining prediction and verification through dual-track prediction and cross-link gating, upstream recalculation and geometric link are paired independently. At the same time, consistency and time synchronization constraints are introduced, and learning is only allowed when the two tracks match and the weighing waveform is healthy. This significantly reduces the probability of dirty samples entering the model and reduces false alarms and misadjustments.

[0018] By combining dynamic checkweighing gating with bias and drift dual-state self-learning, zero-point errors and slow operating condition drift are managed separately. Under the drive of high-reliability samples, the bias is stably corrected and the drift is smoothly followed. When the operating condition deviates, the old information is adaptively forgotten, so that the new steady state is quickly converged after the production change, while minimizing the impact on the cycle time.

[0019] By combining minimum modification correction with geometry priority and three-knob solution, the weight deviation is transformed into the minimum executable combination of feeding, traction, mold temperature and sizing vacuum. Geometric zero-influence projection and short window verification are used to avoid secondary disturbances to the outer diameter and wall thickness. The non-convergence backoff mechanism ensures a safe closed loop for parameter tuning.

[0020] The protection state machine takes over the risk index and batch trend. When there are continuous out-of-bounds, an increase in the low confidence ratio, or a divergence between the two tracks, it automatically slows down, tightens the rejection and pauses learning. When there is a synchronization anomaly, it rolls back to the most recent stable level to suppress the spread of anomalies and parameter drift, and maintains the system operating within the diagnostic and controllable domain. Attached Figure Description

[0021] Figure 1 This is a schematic diagram of the online weight monitoring method for PVC guardrails according to the present invention; Figure 2 This is a schematic diagram of the online weight monitoring system for PVC guardrails according to the present invention. Detailed Implementation

[0022] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0023] Please see Figure 1 This invention provides an online weight monitoring system and method for PVC guardrails, including, In the continuous process of online extrusion-sizing-traction-cutting of PVC guardrails, step one, the weight deviation is not solely caused by single-point fluctuations in feeding or traction, but rather by the combined effect of multiple segments of transmission and timescale mismatch along the upstream to the cutting end. If a uniquely identifiable event window in the time-length domain is not first established for each piece, and an arrival path baseline is not obtained under stable operating conditions, any subsequent weight prediction and verification based on upstream signals will lose its reference surface.

[0024] Without disrupting the production line rhythm, an event window is constructed for each piece around the cutting trigger moment. With a unique part number, the part has a unique mapping in the time and length domains, and is linked to a directed event chain within the same shift. Complete the self-check of the beat and solidify the time synchronization benchmark.

[0025] PVC melt exhibits significant viscoelastic memory during cooling, sizing, and traction processes. The arrival time and amplitude of upstream disturbances on the final weight drift due to fluctuations in traction speed and changes in the cooling field. If an event window anchored at the cutting trigger is not established, an irresolvable mapping ambiguity will arise between the upstream signal and the final weighing. Furthermore, shift changes, tool replacements, and slight speed fluctuations can cause false cycle time anomalies, necessitating the use of a directed event chain within the same shift to distinguish between true out-of-synchronization and tolerable micro-drifts. Therefore, using the cutting trigger time as the anchor point... +Length encoding A unique part number is constructed, and the start and end timestamps, length, and synchronization error of the part are converged into a restricted window to form a unique carrier for subsequent recalculation.

[0026] At the moment of cutting trigger Once reliably captured, the length of the corresponding part is given as the part length in real time by the encoder. Considering the slow drift of the traction speed within the window, the average traction speed is used. The effective transmission rate is approximated within the window, allowing limited fine-tuning only around the boundaries of adjacent parts to absorb encoder quantization errors and cutter hysteresis without altering the physical attribution of the parts. The window definition follows rigid constraints of non-overlapping and auditable rollback.

[0027]

[0028] Among them: Event Window ; Define the first The unique range of an element in the time domain; a closed interval; the trigger time for cutting. Window anchor point; strictly ascending timestamp on the production clock axis; part length The length reference of this component is a positive real number, provided by a length encoder; average traction speed. The effective transmission rate within the window is approximately a positive real number, determined by the traction speed. exist Obtained by internal weighted average;

[0029] Among them: average traction speed Function: To determine Leading edge backtracking timescale; value range: positive real number, implemented by discrete integration of the traction encoder. In implementation, the production sampling cycle... Down, ,in Leading edge fine-tuning amount Trailing edge fine adjustment amount Compensation for sensing and execution lag, small real numbers, and limited absolute values; Allow for fine-tuning of limits ;limited The maximum amplitude is used to prevent cross-part migration. The encoder quantization step distance and tool hysteresis are calibrated offline as constants or operating condition-related functions.

[0030] When used, this definition locks time, length, and part number into a one-to-one mapping, preventing upstream signals from shifting between parts; the limited fine-tuning absorbs noise without destroying the boundary, allowing traceability and auditability to coexist; the non-overlapping window constraint provides prior consistency for subsequent sample gating.

[0031] Same shift identifier Next, all event windows Connect them according to the time sequence to form a directed event chain. To eliminate false out-of-sync errors caused by slight speed fluctuations, a beat self-check index is constructed. The link consistency is judged by the deviation between the predicted cycle time and the actual cycle time, thereby achieving online self-correction without interrupting production, as detailed below:

[0032] Where: rhythm self-check index ; Quantification The consistency between the predicted and measured cycle time, a non-negative real number, where a smaller value indicates better consistency; traction speed. Instantaneous linear velocity, sampled by the traction encoder; cutting trigger time sequence. Define the measured beat rate and strictly use ascending timestamps; implement a directed event chain for each shift. ; Bearing Temporal relationships and consistency states; directed acyclic graphs consisting of nodes and adjacent directed edges; beat consistency threshold. Determine the rhythm self-check index The threshold for qualification; a small positive number, optimized offline and associated with equipment capacity constraints.

[0033] When using, Not exceeding the threshold When the length and velocity fields are consistent, the event chain is stable. If the limits are exceeded continuously, a cycle time anomaly warning can be issued without stopping the production line, which can be used to trigger subsequent time synchronization verification and degradation protection. The event chain provides a reliable timing reference for the dual-track consistency in the second step.

[0034] Inject slight, short-duration, and safe small disturbances within the steady-state operating window. Collect the response from the cutting end. The path kernel is estimated through physically constrained regularized inversion. And generate a baseline version Scope of application This strongly binds the path mechanism to the formula and mold family.

[0035] The core of upstream-to-end recalculation lies in understanding how upstream quality changes propagate to the cutting end within the coupled field of cooling and traction. Directly relying on natural disturbances is both uncontrollable and subject to operational drift, making it difficult to obtain a reproducible path kernel. Therefore, it is necessary to inject small, energy-constrained disturbances with zero integrals under the constraints of shift-window-operating conditions. These disturbances should not alter the average output but should produce a identifiable response. Subsequently, a time-invariant (local) but operational condition-dependent (cross-version) arrival path kernel is obtained using regularized inversion with prior knowledge. This kernel is then bound to the versioning of the recipe and mold family for rapid review or rollback after production changes or seasonal shifts.

[0036] Disturbance injection prioritizes waveforms with opposing positive and negative effects that have no biased impact on average output, thus avoiding changes to the target weight average. The injection location is preferably on the feed side (extruder feed or metering pump fine-tuning), with traction-side torque micro-jumps used as a redundancy channel when necessary. To ensure production line safety and auditability, the integral, peak value, and energy of the disturbance are constrained, and the time width is limited to the current stability window.

[0037] Where: disturbance input Small-signal excitation, bounded measurable function;

[0038] Among them: disturbance amplitude ; number of whole periods (Use positive integers to ensure zero integration); starting time ; time width ; For window functions. When When it is an integer ,and Amplitude and energy are also calculated in the same way. Constraint selection. Disturbance start time. Define the starting point of the disturbance window, which falls within a qualified event window. Internal; Disturbance duration The duration of the disturbance is a small positive number and does not exceed [a certain value]. Width; Peak value limit Limit instantaneous amplitude to ensure safety; small positive number, linked to the equipment response curve. Disturbance energy boundary. Limit injected energy to avoid driving average drift; small positive number, derived from offline verification and device steady-state margin assessment.

[0039] When used, the zero-integral constraint ensures that the average output is not disturbed; the peak value and energy double limit avoids the excitation of nonlinear segments; the time width constraint ensures the location and traceability of the disturbance, thus creating conditions for kernel estimation with high signal-to-noise ratio.

[0040] In the corresponding event window Internal recording cut end response It includes the mass flow trajectory after weighing-side filtering and the transient state of the cutting trigger neighborhood; it also records the environment and operating condition vectors. (Such as extruder barrel temperature, die temperature, sizing vacuum, ambient temperature and humidity, etc.), and the cycle time self-check results from the previous step. As a reference for weight design, the influence of abnormal beat samples is suppressed.

[0041] To obtain a physically interpretable and noise-resistant arrival path kernel, a weighted inversion with a total variation smooth prior is employed:

[0042] Among them: arrival path kernel estimation ; A real function describing the causal kernel of the quality response from the upstream perturbation to the cutting end. ; nuclear candidate Optimize variables, bounded variation function, satisfy physical constraints such as causality and nonnegativity (optional);

[0043] Among them: nuclear candidates ; Path shape function, shape parameters Control the speed of the ascent before the peak. Among them: scale parameter Controlled expansion; ; Start delay Earliest arrival time marker Normalization factor Matching dimension or total gain (usually 1 or determined by calibration). Regularization strength ; Controls the trade-off between smoothness and fit, a positive real number, adjusted with signal-to-noise ratio and window width; kernel derivative The rate of change of the kernel is characterized and used for total variation penalty to suppress high-frequency noise. Cutting end response The explained variable, derived from the weighing side signal; convolution symbol ; indicates that and The time-shifted superposition integral follows a local approximation of the linear time-invariant assumption; the weighting function ; Reduce the contribution of mechanical impact and abnormal cutting rhythm during the cutting process. A bounded function, which can be followed and Attenuation; where: make exist Intrasampling, constructing the Toeplitz convolution matrix Difference operator approximate .have:

[0044] Where: kernel discrete vector ; Discretization; nonnegative vectors (causality and nonnegativity are optional constraints). Weight matrix. To suppress cutting transients and beat anomalies, a diagonal matrix with the following elements:

[0045] Where: difference operator Total variational canonical discretization, first-order forward difference matrix. Numerical solution using ADMM: Introduction of alternating least squares and soft threshold operators Iteration, nonnegativity with projection Implementation. ADMM convergence and step size. With stopping criteria (relative residual < (This is recorded as a parameter file.)

[0046] To achieve a strong binding with the formula and mold family, and to possess falsifiability during production change / seasonal changes, a baseline version is defined. The corresponding applicable scope is the working condition ellipsoidal domain:

[0047] Where: Baseline version ; Index group Its applicable domain, natural numbers; working condition vector ; Incorporate characteristics such as formula labeling, mold family, and environmental conditions. ; Central vector ;Version Representative working conditions, The metric matrix is ​​determined by statistics of the stable segment or by experts. Define the weights and correlations for each working condition dimension; a symmetric positive definite matrix; radius. ; Defines the leniency of the applicable boundaries; Positive real number, related to risk appetite.

[0048] When used, weighted total variation inversion avoids overfitting to high-frequency mechanical noise and maintains the physical interpretability of the kernel shape; the ellipsoidal application domain strongly binds the version to the operating conditions, and any deviation from the range due to production changes or seasonal shifts triggers a review or downgrade; kernel As a transmission operator from upstream to the cutting end, it provides a defined physical baseline for the arrival end recalculation in the second step.

[0049] Event Window Only within the boundaries of two adjacent items is it permitted to do no more than Fine-tuning must be performed and must not overlap; when the working condition vector Detached from the present When a seasonal transition occurs in the environmental state, it must be verified using the latest set of perturbation tests. If the review fails, the status will be downgraded. The learning-related interfaces are frozen, retaining only the ability to allow, remove, and trace until the review is passed.

[0050] Output Item Event Window , reach path kernel Baseline version Scope of application and with a unique part number (Triggered by the cutting moment) +piece length +Shift indicator The code serves as the primary key across steps, providing the window-core-part number triplet to the second step.

[0051] Step 2: Using item-level time windows +Reaching the path kernel +Version Applicability Using the baseline, the upstream quality flow changes are transmitted to the cutting end via the physical kernel and accumulated on a piece-by-piece basis to obtain a positive prediction. At the same time, an inverse prediction channel of unit length quality-geometry-length is independently established. The two predictions are verified by consistency index and time synchronization under the same piece window to form a cross-link gating conclusion, thereby unifying the prediction and verification at the single piece scale and providing a highly reliable sample source for subsequent learning admission and release decisions.

[0052] In the event window of each item Internally, the upstream mass flow changes are transmitted through the arrival path core. Recalculate to the cutting end and integrate on a per-piece basis to obtain the part-level positive predicted quality. Simultaneously constructing inverse prediction quality that is physically independent of it. This results in a dual-track outcome.

[0053] Relying solely on end-of-line weighing will only detect deviations after they have occurred, making timely intervention difficult; conversely, relying solely on upstream mass flow will result in distortions due to arrival timescale shifts caused by cooling fields, traction speed regulation, and cutter lag. Therefore, a core arrival path is required. Mapping the tiny upstream increments to the cutting end, in the part-level window. Internal accumulation enables executable forward prediction; simultaneously, a separate inverse prediction, independent of the shared sensor link, must be established to avoid common-source errors and form a physically independent cross-verification basis. Therefore, subsequent gating no longer relies on single-channel thresholds, but instead uses dual-track mutual verification to define the sample credibility level.

[0054] In the document window Inside, to reach the path core Upstream mass flow offset Perform convolution to obtain the arrival quality flow at the cut end, then integrate within the same window and superimpose the version baseline term to form the part-level positive prediction quality.

[0055] To ensure successful implementation, convolution uses a causal kernel and is time-step driven in the discrete implementation, while integration is precisely truncated at the component window boundaries, thereby avoiding cross-component energy leakage.

[0056] Where: positive prediction quality ;No. The component is based on the prediction quality of the upstream channel, a positive real number; event window. ; Define the first The integral boundary of the component in the time domain is derived from the binding constraint in step one; Reaching the path kernel ; Causal transport kernel from upstream perturbation to the cutting end, The nonnegative real function originates from the perturbation identification in step one and is constrained by the version domain; upstream mass flow offset. The quality flow increment relative to the version baseline is a measurable real function, defined as follows: Convolution symbol The time superposition operation under the linear time-invariant approximation has a well-defined integrable function, and its discrete implementation is completed in fixed time steps. The base version has a unit length quality. ;Version The steady-state mass per unit length within the applicable domain is a positive real number, estimated from historical stable segments and influenced by... Constraints; Part Length ;No. The length reference of the component; a positive real number derived from the length encoding; version baseline quality stream. ;Version Steady-state mass flow reference value; positive real number, in Effective internally.

[0057] In use, a structure combining kernel convolution, window integration, and baseline superposition preserves the physical properties of transmission while avoiding cross-component aliasing; when the traction speed drifts slightly, the integration boundary is locked within the component window. It is predicted that it will not be interfered with by adjacent components; baseline term This ensures that the forecast remains aligned with the stable output even at zero increments, improving zero-bias consistency under undisturbed conditions.

[0058] To construct a channel physically independent of the upstream link, inverse prediction does not reuse upstream quality flow, but instead uses quality per unit length as the core variable. Quality per unit length In engineering, this is achieved by monitoring the mass flow on the upstream metering side. With traction speed The ratio sequence is subjected to robust filtering on one side and verified by combining the geometric profile of the mold family and the sizing vacuum stable section; This estimation explicitly eliminates the influence of the weighing side and the arrival path kernel, thereby decoupling it from the forward prediction on the sensing path and forming the following component-level inverse prediction quality:

[0059] Inverse prediction quality ;No. Predicted quality based on the quality of a link per unit length; a positive real number. Quality per unit length. ; in the Robust unit length mass estimate within the time neighborhood of the corresponding item; positive real number, derived from The length of the component is obtained through filtering on one side and geometric verification. ; Length factor; Positive real number, derived from length encoding and related to Align with classmates.

[0060] In use, the reverse channel avoids the common-cause error between the convolution kernel and the weighing side, especially when the upstream sensor is stable but there is mechanical jitter at the end. It can still maintain a stable reference; at the same time, It is sensitive to gradual changes in mold geometry and can provide early drift signals near version boundaries, providing redundant information for subsequent gating and light calibration.

[0061] Main channel defines the mass per unit length ;in Density is corrected for temperature. The cross-sectional area is determined by inversion of the outer diameter, wall thickness, or key geometry of the profile, measured by an online geometer.

[0062] The backup channel employs robust one-sided filtering in the event of geometric missing channels. .final:

[0063] Wherein: effective density Temperature field correction, positive real number, derived from material curve lookup table; cross-sectional area. The geometric chain core quantity is a positive real number, estimated from the laser outer diameter plus wall thickness or characteristic geometry, with geometrically reliable weights. This determines the primary / backup channel fusion ratio. It is determined by the confidence level of the geometry instrument and the calibration status. For robust mass per unit length; Therefore, the inverse channel is decoupled from the kernel convolution channel on the sensing path, within the same window. Positive internal forecasting With reverse prediction Implement cross-link gating, combine weighing-side early warning with time-synchronized minor calibration, and generate highly reliable conclusions for access learning / release and anomaly downgrade triggering.

[0064] A single threshold is insufficient to cover all anomalies: two predictions that are close but deviate from reality, or one that is phase-displaced due to time mismatch, can mislead the learning admission process. Therefore, it is necessary to construct a symmetrical and scale-independent consistency index to suppress amplitude effects and introduce a time synchronization quantification index to ensure consistency in arrival time scale; then, the waveform health of the weighing side is superimposed to form a three-dimensional gating, thereby logically blocking dirty samples from entering the model.

[0065] To avoid false consistency caused by amplitude amplification, a symmetric and bounded consistency index is adopted. The index evaluates the similarity between two predictions; it depends only on the ratio rather than the absolute magnitude, and is naturally resistant to scaling changes. The consistency index is then used to evaluate the similarity between two predictions. With the weighing side warning amplitude Joint thresholding results in a binary gating conclusion that is highly reliable / learnable and degraded / verifiable:

[0066] Where: Consistency index Quantify the symmetric similarity between two predictions. The larger the value, the more consistent the prediction; positive prediction quality. ; Component-level prediction from kernel convolution channels, positive real numbers; Inverse prediction quality ; Part-level prediction from the unit length quality channel, positive real number; numerical stability term To prevent the denominator from approaching zero and causing divergence, use extremely small positive numbers, set according to the range; Weighing side warning amplitude The quantification of the health status of the weighing waveform of this item is a non-negative real number, output by the rapid screening pre-indicator in step three or the lightweight judgment in this step; consistency threshold. Gating The lower limit requirement; Warning threshold Gating The upper limit requirement; positive real numbers.

[0067] In use, this index is insensitive to amplitude scaling and can reliably identify anomalies that appear similar but have shifted magnitudes. Joint gating with weighing-side health avoids the trap of both being incorrect yet closely related, ensuring that samples entering the learning pool possess both physical and signal reliability. The gating results simultaneously serve the release logic, reducing false releases and rejections. By synthesizing the dual-track results with weighing health on the same criterion plane, a single-item-level, interpretable, and auditable admission conclusion is formed.

[0068] When the consistency index When the edge meets the standard or the weighing side does not issue a warning but is still suspected of phase misalignment, a time synchronization metric is introduced to perform cross-correlation extreme value alignment between the predicted manifold at the arrival end and the response at the weighing end within a window to quantify the time offset at the component level; if the offset exceeds the limit, a baseline light calibration and synchronization check are performed, and the component is downgraded.

[0069]

[0070] Where: Time offset estimation ; Measurement The arrival time standard deviation of the component between the two channels. Real numbers; maximum search radius Limit the search range for time offset to a positive real number, less than the width of the component window; This is a time offset variable; Upstream quality flow The input sequence transmitted by the kernel; the response at the cutting end. The instantaneous response of the weighing side after shock filtering is sampled from the component window; the arrival path core. Causal kernels in convolution operations; Nonnegative real functions; event window Cross-correlation integral interval; time offset threshold ,determination Is it acceptable to use a positive real number?

[0071] In use, cross-correlation alignment can quantify phase mismatch without introducing additional model degrees of freedom, avoiding misjudging amplitude error as model inaccuracy; when Light calibration can update only the synchronization bias without modifying the time. The shape is maintained to keep the version stable; continuous over-limit will trigger version rollback and re-identification to ensure the entire line can quickly reconverge after seasonal changes or traction chain replacement.

[0072] When the consistency index Weighing side warning amplitude And time offset When the sample is deemed highly reliable, an item-level prediction is output. The output should include the dual-track consistency level and the admission / release conclusion; otherwise, it should be marked as downgraded, along with a reason code that triggers a minor calibration / synchronization check / baseline review. All outputs should be uniquely numbered. The primary key is written to the data bus for direct use by the dynamic checkduration gating and dual-state self-learning in the third step.

[0073] Step 3: Item-level event window + Dual-track prediction + The cross-link gating conclusion is used as a priori. A dynamic gating and bias / drift dual-state self-learning mechanism is constructed for end-point weighing: First, a rapid morphological screening is performed on each weighing waveform, and parallel re-weighing is conducted to isolate dirty samples. Then, a negative feedback method is used to update the dual-state parameters only on high-confidence samples, so that the fixed bias and slow drift are estimated and converged separately. When the working condition crosses the version application domain... In the event of time synchronization mismatch, the learning process is automatically frozen, while the release / rejection judgment and traceability capabilities are retained.

[0074] Without sacrificing cycle time, a regular and rapid screening process is performed on each weighing waveform to generate a waveform reliability index, which is then combined with the consistency index output from step two. Consistent with parallel re-weighing, a component-level gating score is formed, based on which only high-confidence samples are allowed to enter the self-learning channel, while low-confidence samples undergo parallel re-weighing and are only used for release / rejection, without entering the learning process.

[0075] The mechanical impact during the cutting of PVC guardrails, the micro-vibration of the conveyor rollers, and the occasional particles in the incoming material can cause spikes, burrs, and short window distortion in the weighing waveform.

[0076] If not in the item-level window first Removing these transient morphological anomalies and directly comparing the weighing results with the dual-track prediction for learning can misinterpret short-term mechanical disturbances as systemic changes in mass, contaminating the model's estimation of bias / drift. Therefore, it is necessary to construct a waveform reliability metric that is sensitive to spikes but robust to plateaus and consistent across links. The consistency of the two scales converges into a single gating score, thereby strictly limiting the learning entry point to samples that are signal-healthy and physically self-consistent.

[0077] In the document window First, the original signal is subjected to anti-impact pre-filtering to obtain the decoupled signal. Subsequently, the ratio of total variation intensity to platform coverage was used as an inverse indicator of morphological stability, and compressed to a monotonically decreasing mapping. The waveform confidence index is obtained from the interval. .

[0078] Among them, the plateau coverage emphasizes the contribution of the plateau segment to the integral through a smooth soft threshold function, avoiding the dominance of pure amplitude.

[0079] Where: Waveform credibility index ; Quantification The morphological health of the weighing waveform; The larger the value, the more reliable it is; the proportionality coefficient The effect of adjusting the total variation per piece on the index is a positive real number, which is adjusted online according to the rigidity of the weighing platform and the conveying status; event window. Integration interval and component-level attribution boundary, closed interval, anchoring and fine-tuning constraints from step one; decoupling signal The weighing signal after shock filtering is a real function. The filter is initially disclosed as a linear-phase finite impulse response. The passband covers the target fundamental frequency, and the stopband suppresses the cutting of the main lobe of the shock. time derivative ; Characterizes regions of drastic waveform changes; measurable function; soft threshold function Emphasizing platform segment coverage; The forms of full disclosure are: Platform threshold Distinguish between the plateau segment and the low-amplitude segment; positive real numbers, derived from the stable segment calibration; numerical stability constant. To prevent the denominator from approaching zero; to use extremely small positive numbers.

[0080] When used, the waveform reliability index It is highly sensitive to spikes and jagged edges, and positively correlated with the width and smoothness of the platform segment. It can quickly characterize the waveform health without calculating variance / standard deviation. It exponentially maps the impact of natural shear extreme anomalies on the gating score, ensuring that learnable parts can still be stably selected even in scenarios with heavy cutting impact.

[0081] In obtaining the waveform credibility index Then, the consistency index with the output of step two. and parallel rebalancing consistency Perform convex combination to obtain component-level gating score. Consistency of reweighing Calculated based on the symmetrical consistency measure between the initial weighing mass and the reweighing mass, and after time-synchronized minor calibration, a neutral value is given based on the equipment's prior knowledge when there is no reweighing. The score exceeds the threshold and the weighing side issues a warning. If the limit is not exceeded, the item enters the self-learning channel; otherwise, it enters the parallel channel that only allows / rejects items, where:

[0082] Where: Gating score Learning admission scores that aggregate evidence from three sources; Weight Adjusting the relative contributions of the three sources of evidence, non-negative real numbers, satisfying Consistency Index Dual-track prediction of symmetric proximity. Consistency of reweighing The degree of symmetry between the initial weighing and the reweighing. When there is no reweighing, take the prior arithmetic mean value of the equipment.

[0083] Let the initial quality be determined Re-weighing quality ,but

[0084] Among them, re-weighing consistency To measure the reproducibility of the repeated weighings Its form and consistency index Isomorphism and independent physical meaning facilitate the integration of gating scoring at the same scale. Numerical stability term To prevent the denominator from approaching zero and causing divergence; When in use, gate scoring By compressing the three-dimensional evidence of signal health, model self-consistency, and equipment redundancy into a single metric, it can be weighted online to adapt to different formulations / molds, and can also improve the waveform credibility index when specific anomalies (such as conveyor vibration) dominate. Weighted suppression of mislearning; and Joint thresholding can achieve robust learning admission without increasing the risk of equipment downtime.

[0085] Under the premise of including only high-confidence samples, the fixed bias and slow drift are separated into two states that can be independently tuned and updated in the form of negative feedback; when the operating conditions are far from the version domain or synchronization anomalies continue, the forgetting factor is automatically reduced and the update is frozen, so that the estimate remains at a traceable stable level.

[0086] The weight error of PVC guardrails includes both the fixed bias of the weighing link zero point / calibration and the slow drift caused by the feeding-traction-cooling coupling, which changes slowly with the evolution of working conditions. If a single parameter absorbs both, it will inevitably lead to parameter tuning conflicts: increasing the learning rate to follow the slow drift will push the bias estimate away from the steady state; conversely, tightening the learning rate to stabilize the bias will result in a loss of response to drift. Therefore, it is necessary to explicitly separate these factors at the model level and link the update gain to the degree of abrupt changes in working conditions and the gating score, achieving a trade-off between fast response and steady-state stability.

[0087] For the first Component, define fixed bias state With drift state Step two fusion prediction quality With reliable measured quality The difference is the weighing residual. weighted by gating Driven by a dual-state update; where the drift state uses a decay factor. To maintain gradual variability, the bias state does not decay to lock the zero point. Component-level updates are written as:

[0088] Wherein: fixed bias state Absorbs systematic errors of zero point / calibration type that do not change with operating conditions over a long period of time; real number, small amplitude, slow variation; drift state. Characterized by weight drift that evolves slowly with operating conditions; a real number; the time constant is significantly larger than the cycle time; attenuation factor. Maintaining the drift-gradual nature, Approaching 1; Gated learning gain ; by gating score This value is derived from mapping and is only taken as a larger value when the sample is highly reliable. Offset step size ; Control bias update magnitude, a positive real number, typically less than Drift step size Control the drift following speed; a positive real number, greater than... Weighing residual ; equals the measured quality minus the fusion prediction quality; a real number, calculated and used only on high-confidence samples.

[0089] When used, dual-state updates decouple zero-point correction from slow following. Suppressing high-frequency noise injection into the drift channel without affecting bias zeroing; gated learning gain Learning occurs only during gating, isolating dirty samples from the mechanism and preventing model divergence caused by mislearning; the drift step size is adjusted when a new steady state of the formulation is formed. To ensure fast convergence, the bias step size is... Stabilize at zero point.

[0090] To adapt to production changes and seasonal variations, historical weights need to adaptively decrease as operating conditions deviate, thus preventing the old steady-state concept from dragging down the new steady-state; simultaneously, updates should be frozen when synchronization offsets or consistency remain low for an extended period. The square root of the version domain quadratic form should be used as the operating condition distance metric. Constructor-level forgetting factor Monotonically decreases with working distance:

[0091] Among them: forgetting factor Adjusting the influence of historical samples on current learning. The smaller the value, the faster the forgetting; proportionality coefficient Set the sensitivity of working distance to the rate of forgetting, a positive real number; Working distance ; Measures the deviation of the current working condition vector from the version center; a non-negative real number, defined as the set in step one. The quadratic form used The square root of.

[0092] Where: working condition vector , No. The process and environmental characteristic vectors corresponding to the moment of the part (such as extruder temperature zone, melt pressure, die temperature, sizing vacuum, traction speed, ambient temperature and humidity, etc., assembled after time alignment); value range: .

[0093] Version Center ,Version Representative operating condition center obtained from statistics during the stable phase; value range: Archived in versions according to recipe and mold family.

[0094] metric matrix Define the importance and relevance weights of each working condition dimension; value range: symmetric positive definite matrix. Acquisition method: diagonal weighted form is acceptable. or contraction estimate of the inverse covariance matrix ,in For version The sample covariance of the stable segment, For regularization parameters.

[0095] When using it, when the working condition vector Forgetting factor when approaching version boundary Automatic reduction allows the influence of old samples to decay rapidly, ensuring a quick and local reconstruction of a new steady state after production change; if the consistency index Long-term decline or If the limit is exceeded, the forgetting factor will... Clamp to the lower limit and freeze the gating learning gain This ensures parameter stability and audit traceability when synchronization / model is questionable.

[0096] The weighing platform must be rigidly vibration-isolated and equipped with an inlet guide to stabilize the material landing point; each shift, standard parts are used to complete the zero point and platform threshold calibration. Fast calibration; parallel re-weighing channel is implemented through a bypass weighing platform and decoupled from the main line cycle time; gating threshold. Warning threshold With the upper / lower limits of the forgetting coefficient in version Archived by unit, and refilled uniformly when production changes or seasons change.

[0097] Output approval / rejection conclusions and sample confidence levels (including) , (with complex scale labels), two-state estimation Its health status (update range, freeze flag), and with a unique item number. Write it into MES; this set serves as direct evidence for the bias breakdown-minimum change recommendation-state machine trigger in step four, and supports per-item computation to minimize the recall boundary.

[0098] Step 4: In the Item-Level Events Window + Dual-track prediction +Gate control score + Two-state estimation Under the constraints, the part weight deviation is decomposed into four attributable sources, and the minimum modification combination with geometric priority is obtained in the three knobs (feed response, traction fine adjustment, mold temperature / sizing vacuum). When there are continuous abnormalities or insufficient evidence, the protection state machine is triggered to implement speed reduction, stricter rejection, pause learning and rollback.

[0099] The mapping of part-level deviation-source category-three knobs is established as a physically constrained minimum modification problem, and univariate short-window verification and non-convergence backoff are implemented under the geometry priority rule.

[0100] The dual-track result and time synchronization amount output in step two And the two states in step three This indicates that the weight deviations are not from the same source; if a single knob is used for compensation directly, it often leads to geometric fluctuations or the accumulation of secondary disturbances. Therefore, it is necessary to first decouple the deviations to the four types of sources at the evidentiary level, then restrict the changes to the reachable domain of the three knobs, and use the geometry priority rule to set the key geometry as a hard constraint. Finally, short-window verification is used to ensure convergence and rollback.

[0101] In the document window Internally, construct the residual-evidence vector. Its components include evidence from multiple sources, such as weighing residuals, dual-track deviation, and time synchronization anomalies, which originate from... .

[0102] To avoid amplifying a single piece of evidence into a sole causal relationship, a sparse convex decomposition with nonnegativity and probabilistic simplex constraints is employed to transform the residual-evidence vector. Projecting onto a convex combination of the four source base dictionaries yields the source weights. The four categories correspond to the metering link error process coupling fluctuation medium / mechanical interference model and synchronization problem.

[0103]

[0104] Where: Source weight vector ;Quantify the relative contributions of the four sources to the current bias, a four-dimensional non-negative vector with the sum of its components equal to 1; Residual-evidence vector ; Gathering case-by-case evidence, typical examples , real number vector; dictionary matrix The standard evidence morphology base from four sources is obtained offline by training through stimulus experiments and historical stable periods and can be slightly adaptive online. It is a real matrix with 4 columns and full rank. Regularization coefficient Suppress overfitting caused by non-indicative multi-source collinearity and encourage sparse attribution; Weighing residual ; equals the difference between measured quality and fusion prediction; consistency index Dual-track symmetry proximity; Time offset estimation Cross-link arrival time standard deviation; ; Weighing side warning amplitude Waveform health alarm quantity, a non-negative real number; When using it, after introducing nonnegativity and probability constraints, the source weight vector It is interpretable and auditable; when When significant, the model and synchronization problem components will increase rather than be misattributed to process fluctuations; when the weighing side warning amplitude... Increase and Consistency Index High, generally indicating media / mechanical interference, avoids misadjustment of the feeding side; this decomposition provides a weighting benchmark for source selection in the solution of the three knobs.

[0105] Set control increment These correspond to the feeding response, traction fine-tuning of mold temperature, and sizing vacuum, respectively. The effective mass error at the part level is considered. (Remove fixed bias) The impact is the compensated quantity, and the source weight vector is used as the source weight vector. The preference matrix of the three knobs is reweighted to construct a weighted minimum norm problem with geometric hard constraints. When there is a critical geometric conflict, a geometric priority projection is performed, allowing only changes in the geometrically zero-influence subspace to enter, while other changes are rolled back to the feed-side slow time-scale compensation.

[0106] To facilitate online computation, we present the closed-form solution when the geometric constraints are not activated (equivalent to the weighted minimum norm solution), and the orthogonal projection correction when the constraints are activated, yielding the weighted minimum norm solution. :

[0107] Where: geometric priority projection solution The weighted minimum norm solution The changes projected onto the geometrically zero-effect subspace are used for online verification and execution; control increments. The setting change of the three knobs is a real vector, and its components are constrained by the upper / lower limits of the device. Feeding rate modification Upstream mass flow setting or feed gain fine-tuning, real number, small and smooth; traction modification Fine-tuning of traction speed; real numbers, with more stringent geometric constraints on amplitude; mold temperature / vacuum modification. The rate of change of mold temperature or sizing vacuum fine-tuning is constrained by thermal / pneumatic inertia. weight matrix Characterizing the cost and availability of the three knobs, typically a symmetric positive definite matrix, with elements based on the source weight vector. Online adjustment of equipment load; symmetric positive definite matrix; mass sensitivity row vector The sensitivity of the mass to the three knobs under local linearization; Real matrix, maintained online by small perturbation identification or historical regression; geometric sensitivity matrix The sensitivity of key geometric features to the three knobs; Real matrix, Controlled geometric dimension; identity matrix Identity operators in projection operations; effective mass error The quality error that needs to be immediately offset on this part is defined as follows: ;Real number.

[0108] When using it, the weighted minimum norm solution Achieving weighted minimum norm correction for quality errors, while using geometrically prioritized projection solutions. Projecting geometric effects to zero strictly prioritizes geometry; as geometric constraints tighten, the projection automatically pushes more changes toward feed rate changes. The slow time-scale compensation avoids adverse disturbances to the cross section caused by traction and mold temperature; the closed solution facilitates fast online execution and supports the immediate strategy of single-variable short-window verification → backing down if it does not converge.

[0109] The protection state machine is driven by risk measurement and evidence consistency. When anomalies are clustered, it implements slowdown, stricter elimination, pause learning and rollback. At the same time, the entire chain of prediction-testing-gating-correction is written into MES to form a falsifiable hash audit chain and minimize the recall boundary.

[0110] When multiple consecutive out-of-bounds events occur, the proportion of low-confidence samples increases, or the two tracks deviate for a long time, continued learning and frequent parameter tuning will amplify uncertainty and pollute the baseline version. Therefore, a controlled state machine logic is needed to map risk aggregation into deterministic actions and ensure that all actions are traceable and reversible. In addition, in recall scenarios, the boundary of the minimum necessary impact range should be given to avoid over-recall.

[0111] Constructor-level feature vectors By aggregating key evidence (magnitude, phase, credibility, proportion, and trend), a risk index is obtained using a monotonic S-shaped mapping. And aggregate them into batch risks within a sliding window. State machines Batch risk The threshold combination triggers a slowdown, stricter elimination, pauses learning, and rolls back to the most recent stable setting.

[0112]

[0113] Among them: risk index ; for the first A dimensionless measure of overall risk. The larger the value, the higher the risk; S-shaped function Monotonic compression mapping, publicly available in the form of Weight vector Feature contribution coefficients, real vectors, initialized offline and slightly adaptively online; Feature vector ;polymerization , , Low credibility ratio Warning amplitude , Normalized combination; batch risk Risk is averaged within the sliding window, and the window width is tied to the beat; low reliability percentage. Proportion of low-confidence samples within the window; Time offset threshold Synchronization tolerance; positive real number, from step two.

[0114] When using it, use the risk index For particle size and batch risk The dual-scale criterion for trends can both quickly respond to abnormal eruptions and suppress occasional fluctuations; when the risk index Batch risk Simultaneously, when the boundary is exceeded, the state opportunity executes a combination of slowing down and pausing learning; if If the boundary is exceeded for an extended period, a rollback and synchronous review will be triggered first to avoid affecting... Error altered.

[0115] All and the All related predictions, measurements, gating, corrections, and status actions are based on... The primary key is written to MES to form a timestamped audit entry; the entry hash is concatenated with the hash of the previous destination to form an immutable chained log. When a recall demand arises, the minimum recall boundary set is calculated based on the combined threshold of the operating condition domain and the risk index, covering what is necessary without expanding the scope of impact.

[0116]

[0117] Among them: recall collection The set of part numbers requiring review or recall, a discrete set; working condition distance. ; item The distance between the working condition and the version center is defined using the quadratic form metric from step one, and is a non-negative real number; version radius. ;Version The applicable boundary radius, a positive real number, comes from step one; risk threshold. The minimum risk threshold for a recall action. ; Actual measured quality ; item The weighing result is a positive real number; positive prediction. ; item Kernel convolution channel prediction, positive real number; recall bias threshold ; The quality deviation threshold that triggers a recall, a positive real number.

[0118] In practice, chained hashing ensures that any subsequent modifications can be verified; the recall set is constrained by both the operational domain and risk threshold to ensure that only high-risk items within the same domain as the current issue are covered, achieving a regulatory-friendly recall with minimal impact; combined with source weight vectors... With geometric priority projection solution The traces left behind can reproduce the experimental path and clarify the attribution of responsibility.

[0119] Any geometric-first projection solution Online verification is only allowed within a single-variable short window: first, perform small-step trials on the single knob with the lowest cost, then monitor the other two windows. With critical geometric drift; if the verification error does not decrease monotonically or geometric constraints are triggered, immediately roll back and select the next knob or combination; if convergence fails three times, freeze the current batch of learning, maintain the slowdown and enter synchronous review; all rollback actions and reason codes are written to MES.

[0120] The final output includes three types of executable and traceable objects: a three-knob minimum modification suggestion and a geometry-first projection solution. Execution records; protection status and transfer trajectory (including deceleration ratio, removal threshold, and learning freeze flag); recall set. With audit chain hash. The above objects are... Primary key and steps two and three It enables horizontal integration, supporting subsequent process review and cross-formula migration.

[0121] Please see Figure 2 This invention provides an online weight monitoring system for PVC guardrails, comprising: The event modeling module uses the cutting trigger as the anchor point, generates a unique part number with the length encoding, constructs a per-part event window, injects small disturbances and collects the response from the cutting end, obtains the arrival path kernel from upstream to the cutting end through weighted inversion, and establishes a version applicable domain with the formula and mold. The dual-track gating module recalculates the upstream quality change to the event window based on the arrival path kernel to obtain a part-level positive prediction. At the same time, it obtains an inverse prediction based on the unit length mass and the length of the part, calculates the consistency index, and combines the time offset and the weighing side warning to form a cross-link gating conclusion. The checksum learning module implements dynamic checksum gating on the re-counting waveform. Low-confidence samples are re-weighed in parallel and not included in the learning process. High-confidence samples drive bias and drift dual-state self-learning. The learning gain and forgetting factor are adjusted according to the gating score and working condition distance to generate part-level release or rejection conclusions and states. The deviation correction and protection module maps the part weight deviation to the minimum change combination of the three knobs of feeding response, traction fine adjustment, mold temperature and sizing vacuum according to the source. It performs geometric priority constraints and short window verification. If it does not converge, it backs down and triggers the protection state machine. The entire process is written to the MES with the part number to form the audit and recall boundary.

[0122] Those skilled in the art will recognize that the units and algorithm steps of the various examples described in conjunction with the embodiments disclosed herein can be implemented in electronic hardware, or a combination of computer software and electronic hardware. Whether these functions are implemented in hardware or software depends on the specific application and design constraints of the technical solution. Those skilled in the art can use different methods to implement the described functions for each specific application, but such implementation should not be considered beyond the scope of this application.

[0123] Those skilled in the art will clearly understand that, for the sake of convenience and brevity, the specific working processes of the systems, devices, and units described above can be referred to the corresponding processes in the foregoing method embodiments, and will not be repeated here.

[0124] In the several embodiments provided in this application, it should be understood that the disclosed systems, apparatuses, and methods can be implemented in other ways. For example, the apparatus embodiments described above are merely illustrative; for instance, the division of units is only a logical functional division, and in actual implementation, there may be other division methods. For example, multiple units or components may be combined or integrated into another system, or some features may be ignored or not executed. Furthermore, the coupling or direct coupling or communication connection shown or discussed may be through some interfaces; the indirect coupling or communication connection between apparatuses or units may be electrical, mechanical, or other forms.

[0125] The units described as separate components may or may not be physically separate. The components shown as units may or may not be physical units; that is, they may be located in one place or distributed across multiple network units. Some or all of the units can be selected to achieve the purpose of this embodiment according to actual needs.

[0126] The above description is merely a specific embodiment of this application, but the scope of protection of this application is not limited thereto. Any variations or substitutions that can be easily conceived by those skilled in the art within the scope of the technology disclosed in this application should be included within the scope of protection of this application. Therefore, the scope of protection of this application should be determined by the scope of the claims.

Claims

1. A method for online monitoring of the weight of PVC guardrails, characterized in that: include, Using the cutting trigger as the anchor point, a unique part number is generated in conjunction with the length encoding. A part-by-part event window is constructed, small disturbances are injected and the response at the cutting end is collected. The arrival path kernel from upstream to the cutting end is obtained through weighted inversion, and a version applicable domain is established with the formula and mold. Based on the arrival path kernel, the upstream quality change is recalculated into the event window to obtain the part-level positive prediction. At the same time, the inverse prediction is obtained by using the unit length mass and the length of the part. The consistency index is calculated and combined with the time offset and the weighing side warning to form a cross-link gating conclusion. Dynamic check-repeat gating is implemented for symmetrical re-waveforms. Low-confidence samples are re-weighed in parallel without being included in the learning process, while high-confidence samples drive bias and drift dual-state self-learning. The learning gain and forgetting factor are adjusted according to the gating score and working condition distance to generate part-level release or rejection conclusions and states. The part weight deviation is mapped to the feed response, traction fine-tuning, mold temperature and sizing vacuum three knobs to solve the minimum change combination, and geometric priority constraints and short window verification are performed. If it does not converge, it will back down and trigger the protection state machine. The entire process is written into MES with part number to form audit and recall boundaries.

2. The method for online monitoring of the weight of PVC guardrails according to claim 1, characterized in that: The item-by-item event window only allows time boundary fine-tuning between two adjacent items according to a preset fine-tuning limit and prohibits overlap; all items in the same shift form a directed event chain according to the cutting trigger order for beat self-checking and time synchronization verification; the unique item number is generated by combining the cutting trigger time, item length and shift identifier and runs through subsequent steps.

3. The method for online monitoring of the weight of PVC guardrails according to claim 2, characterized in that: Under stable operating conditions, a small perturbation with zero integral, limited amplitude and energy is injected and the response at the cutting end is recorded; the arrival path kernel is obtained by weighted total variation canonical inversion, and the arrival path kernel, formula identifier, mold family and environmental parameters are used to establish a version applicable domain and archive it; The weighting function is composed of time decay, beat consistency and waveform edge indication. The version application domain is defined by the working condition vector and the metric matrix and is used for production change review.

4. The method for online monitoring of the weight of PVC guardrails according to claim 3, characterized in that: The reverse prediction uses the fusion of the geometric chain main channel and the mass flow ratio backup channel. The geometric chain calculates the mass per unit length using online geometry and temperature-corrected density, and the backup channel estimates the mass per unit length using robust truncated mean. The fusion weight is determined by the geometric measurement confidence level. The two channels are independent of each other on the sensing path and time-aligned to the event window. The mass per unit length and the length of the component are accumulated in the same window to form an inverse prediction and participate in subsequent gating.

5. The method for online monitoring of the weight of PVC guardrails according to claim 4, characterized in that: Cross-link gating includes: constructing a symmetric consistency index composed of forward and reverse predictions, introducing weighing side early warning and time offset constraints, and locating synchronous offset within a preset search radius using cross-correlation peak values, which can be refined by interpolation. The gating thresholds include a consistency threshold, a warning amplitude threshold, and a time offset threshold. All thresholds are archived at the version level and loaded with each shift. When any constraint is not met, a downgrade is marked and a light correction and synchronization check are triggered.

6. The method for online monitoring of the weight of PVC guardrails according to claim 5, characterized in that: Dynamic checkweighing gating includes: constructing a waveform reliability index obtained by monotonically mapping the ratio of platform coverage to change intensity within the event window, and combining its consistency index and re-weighing consistency in a convex manner to form a gating score; The re-weighing channel is implemented through a bypass weighing platform and decoupled from the main cycle. The re-weighing label is associated with the gating score and stored. When re-weighing is not triggered, neutral consistency is assigned to maintain the fusion scale.

7. The method for online monitoring of the weight of PVC guardrails according to claim 6, characterized in that: The bias and drift dual-state update is driven by the learning gain obtained by the gating score mapping. The drift channel is coupled with the working condition distance of the version application domain to establish a forgetting factor. The update is frozen when the gating score does not reach the threshold or the time offset exceeds the limit. The learning gain is given by the monotonic mapping and is constrained by the gating threshold. The forgetting factor decreases monotonically with the working condition distance and only acts on the drift channel. Measurement and traceability records are retained during the freeze period. The output includes component-level release or rejection conclusions and dual-state parameters.

8. The method for online monitoring of the weight of PVC guardrails according to claim 7, characterized in that: The part weight deviation is mapped to the three knobs of feeding response, traction fine adjustment, mold temperature and sizing vacuum according to the source weight. The minimum change combination is obtained according to the weight matrix and mass sensitivity, and then projected to the geometric zero influence subspace through the geometric sensitivity matrix. The upper and lower limits and step size of the three knobs are constrained by the equipment file. When the geometric priority constraint is activated, only slow time-scale compensation on the feeding side is allowed. The verification window covers the next two items and records the reason code.

9. The method for online monitoring of the weight of PVC guardrails according to claim 8, characterized in that: The protection state machine uses the risk index composed of amplitude, phase, low confidence ratio and trend, and the risk of the sliding window batch as the triggering quantity, and adopts double threshold hysteresis to determine deceleration, stricter rejection, pause learning and rollback. The predictions, measurements, gating, corrections, and status are written to the manufacturing execution system on a piece-by-piece basis and linked by hashes to form an audit chain and recall boundary. The fields include part number, event window, dual-track prediction, time offset, gating score, execution modification, and status.

10. A PVC guardrail weight online monitoring system, characterized in that: include, The event modeling module uses the cutting trigger as the anchor point, generates a unique part number with the length encoding, constructs a per-part event window, injects small disturbances and collects the response from the cutting end, obtains the arrival path kernel from upstream to the cutting end through weighted inversion, and establishes a version applicable domain with the formula and mold. The dual-track gating module recalculates the upstream quality change to the event window based on the arrival path kernel to obtain a part-level positive prediction. At the same time, it obtains an inverse prediction based on the unit length mass and the length of the part, calculates the consistency index, and combines the time offset and the weighing side warning to form a cross-link gating conclusion. The checksum learning module implements dynamic checksum gating on the re-counting waveform. Low-confidence samples are re-weighed in parallel and not included in the learning process. High-confidence samples drive bias and drift dual-state self-learning. The learning gain and forgetting factor are adjusted according to the gating score and working condition distance to generate part-level release or rejection conclusions and states. The deviation correction and protection module maps the part weight deviation to the minimum change combination of the three knobs of feeding response, traction fine adjustment, mold temperature and sizing vacuum according to the source. It performs geometric priority constraints and short window verification. If it does not converge, it backs down and triggers the protection state machine. The entire process is written to the MES with the part number to form the audit and recall boundary.

Images