An adaptive early warning method for industrial monitoring parameters based on equipment operation profiles

By constructing hierarchical operational profiles and conditional baselines, and using conditional quantile networks to generate monitoring parameter vectors, the problem of baseline drift and missed reporting in industrial equipment monitoring was solved, achieving accurate early warning of equipment status and collaborative optimization of parameters.

CN123113372APending Publication Date: 2026-11-17湖南医药学院 +1
View PDF 0 Cites 0 Cited by

Patent Information

Application Number
CN202611449321.9
Authority / Receiving Office
CN · China
Patent Type
Applications(China)
Current Assignee / Owner
Filing Date
2026-09-17
Publication Date
2026-11-17

AI Technical Summary

Technical Problem

In existing technologies, industrial equipment monitoring suffers from missed reports due to the normal baseline update criteria drifting in the same direction as the monitored degradation state in scenarios with frequent formula changes and intermittent processes. Furthermore, the monitoring parameter settings suffer from contradictions between high-frequency sampling and narrow thresholds, and lack a unified risk budget for multivariate sensitivity allocation.

Method used

By acquiring multi-source sensing data, we construct phase-sequence features and data quality reliability vectors, decompose operating conditions into formula factors and operating segment factors, construct hierarchical operating profiles and conditional baselines, use conditional quantile networks to generate monitoring parameter vectors, perform hysteresis limit judgment and residual displacement accumulation calculation, generate graded early warning conclusions, and update and calibrate parameters.

Benefits of technology

It achieves the ability to avoid missed reports when equipment is slowly degrading, distinguish between normal operating condition drift and fault evolution, reduce conflicting configurations between monitoring parameters, and has engineering feasibility and real-time solution capabilities.

✦ Generated by Eureka AI based on patent content.

Smart Images

  • Figure CN123113372A_ABST
    Figure CN123113372A_ABST
Patent Text Reader

Abstract

This invention relates to the field of industrial equipment monitoring technology and discloses an adaptive early warning method for industrial monitoring parameters in equipment operation profiling. This method addresses the technical problem in related technologies where the normal baseline update criterion and the degradation state are of the same origin, leading to missed detections. The method decomposes the operating condition into a formula factor and an operating segment factor, synthesizes a conditional baseline through outer product, maintains a following baseline and a frozen anchor baseline, projects the center displacements of both onto the operating condition subspace to obtain residual displacements, and performs cumulative checks to determine the drift type and open / close the profile update gate. Then, based on the conditional quantile calibrated by coverage, and under a false alarm budget that tightens with the value of work-in-process, it jointly generates monitoring upper and lower limits, hysteresis bands, duration, and sampling parameters. This invention enables collaborative adaptive early warning of monitoring parameters in intermittent process scenarios where the combined samples of formula and operating segments are sparse.
Need to check novelty before this filing date? Find Prior Art

Description

Technical Field

[0001] This invention relates to the field of industrial equipment monitoring technology, and more specifically, to an adaptive early warning method for industrial monitoring parameters of equipment operation profile. Background Technology

[0002] Industrial equipment monitoring typically uses vibration, temperature, current, vacuum, speed, load, and environmental parameters to determine equipment status. Related technologies have evolved from single-parameter fixed thresholds to multi-source fusion health index modeling, deep time series prediction, and dynamic threshold adjustment. This has led to the development of technical routes such as operating condition gating networks, Bayesian dynamic thresholds, rolling statistical dynamic thresholds, and multi-granularity context correction, which are used to identify equipment operating status from signals such as vibration, temperature, and current and provide corresponding alarm prompts. In intermittent process scenarios where formulas are frequently switched and coating valves are periodically opened and closed according to production rhythm, the formula and the running section form a discrete Cartesian product combination. The historical samples confirmed to be normal are relatively sparse in most combinations. The monitoring quantity has a segmented structure with alternating coating and blank phases as the valve is opened and closed. The duty cycle changes with the formula. The scrap and re-threading losses from unplanned downtime accumulate with the length of the processed work-in-process. Therefore, multiple constraints are put forward on the setting of monitoring parameters that take into account both sensitivity and robustness. The related technologies have technical problems. The criteria for determining whether to update the normal baseline are calculated by the normal baseline itself, which causes the baseline to drift in the same direction as the fault when the equipment slowly degrades, resulting in missed reports. At the same time, the monitoring upper and lower limits, hysteresis band, and over-limit duration are related to the sampling frequency but are set independently. This can easily lead to contradictory combinations such as high-frequency sampling with long delay criteria and narrow thresholds without hysteresis band. Furthermore, the sensitivity allocation among multiple variables lacks verifiable basis under a unified risk budget. Summary of the Invention

[0003] This invention provides an adaptive early warning method for industrial monitoring parameters of equipment operation profile, which solves the technical problem in related technologies where the normal baseline update criterion and the monitored degradation state are from the same source, leading to drift and missed reporting.

[0004] This invention provides an adaptive early warning method for industrial monitoring parameters based on equipment operation profiles, comprising the following steps: S1: Acquire multi-source sensing data, coating valve on / off signal, encoder phase and formula number, aggregate phase by phase along on / off edge and calculate missing step and noise variable by variable to obtain phase feature sequence and data quality confidence vector; S2 receives the phase-sequence feature sequence and the data quality confidence vector, decomposes the working condition into the formula factor and the running segment factor, and constructs the hierarchical operating profile and condition baseline through factor prototype outer product synthesis and sample size shrinkage. S3, derived from the conditional baseline, follows the baseline and freezes the anchoring baseline, projects the center displacement onto the working condition subspace to separate the residual displacement, accumulates and checks according to the direction consistency, and determines the drift type and opens and closes the image update gate to obtain the cumulative residual displacement. S4, the hierarchical operation profile, conditional baseline and residual displacement accumulation, along with the target risk share, are fed into the conditional quantile network. After coverage calibration and closed-loop allocation of risk shares under the overall false alarm budget that tightens with the value of work-in-process, a monitoring parameter vector is generated. S5, based on the monitoring parameter vector, performs dual-channel evidence verification of hysteresis over-limit judgment and residual displacement accumulation on each phase segment of the phase characteristic sequence, synthesizes alarms according to phase effective timing and clearing hysteresis, and obtains graded early warning conclusions and sampling execution instructions; S6 collects the manual and maintenance confirmation results of the graded early warning conclusions and sorts them by type. It is used for anchoring baseline updates, fault memory writing, training set governance and budget weight correction, and updates the hierarchical operation profile, anchoring baseline, conditional quantile network parameters and overall false alarm budget.

[0005] Furthermore, each component of the data quality confidence vector represents the data quality confidence of the corresponding monitored variable, and the data quality confidence is synthesized by the product of three factors; The first factor is one minus the sampling missing ratio, the second factor is one minus the normalized time synchronization deviation, and the third factor is the natural exponent of the normalized noise intensity after taking the negative value. The time synchronization deviation is obtained by using the on / off edge of the coating valve on / off signal as the common event reference.

[0006] Furthermore, the formulation factor is a formulation anchor point obtained by clustering historical formulations in a normalized attribute space spanned by areal density setting, slurry viscosity, coating line speed and valve cycle duty cycle. The degree of belonging of the current phase segment on the formulation factor is obtained by soft allocation based on the distance from the coordinates of the current formulation in the normalized attribute space to the center of each formulation anchor point. The running segment factor consists of standby, speed increase, stable coating, speed decrease and roll change. The condition center of the conditional baseline is synthesized from the main effects of the population baseline, the formulation factor and the running segment factor, and the outer product of the bilinear interaction term after sample size shrinkage. When the normal sample size of the combination is insufficient, the condition center degenerates into the sum of the population baseline and the two types of main effects.

[0007] Furthermore, when the profile update gate is opened, the following baseline is weighted according to the data quality confidence and corrected by the main effect, bilinear interaction term and individual bias of the preset learning rate correction factor. The anchoring baseline is frozen as a frozen copy of the following baseline and is only completely replaced at the end of the observation period after normal operating condition drift is confirmed, maintenance is completed and stable operation is confirmed, and sensor replacement or recalibration is completed. The center displacement is the difference between the condition centers synthesized by the following baseline and the frozen anchor baseline under the same formulation factor and the same running segment factor.

[0008] Furthermore, a working condition subspace design matrix is ​​constructed, the columns of which consist of the components of the formula factor, the components of the operating section factor, the ambient temperature and humidity, and the feed temperature. The data quality reliability vector is used as the weight and a ridge regularization term is added. Weighted least squares are used to regress the center displacement to the working condition subspace design matrix to obtain the working condition interpretation coefficient. The residual displacement is obtained by subtracting the product of the working condition subspace design matrix and the working condition interpretation coefficient from the center displacement.

[0009] Furthermore, the residual displacement is locally normalized according to the robust residual displacement scale corresponding to the current formula, and the dot product of the normalized residual displacement vector and the unit vector of its recent moving average direction is used as the direction consistency. A sequential probability ratio test is performed on the two hypotheses that the residual displacement is zero-mean noise and that the residual displacement increases slowly along the continuous direction. The log-likelihood ratio is accumulated segment by segment and weighted according to the direction consistency to obtain the cumulative residual displacement. When the cumulative residual displacement exceeds the preset test threshold, it is judged as fault evolution. The profile update gate is opened when the cumulative residual displacement does not exceed the preset inspection threshold, the average data quality reliability meets the preset data entry requirements, the data is not currently in an alarm state, is not marked as a fault or maintenance state, and is not in the structural change observation period.

[0010] Furthermore, the input of the conditional quantile network consists of the embedding vectors of the recipe identifier, the running segment identifier, and the individual equipment identifier, the time-series coding vector obtained by one-dimensional causal convolution of the phase-series feature sequence, the data quality confidence vector, the statistical summary of the residual displacement, and the target risk share concatenated together. The embedding vector of the recipe identifier and the embedding vector of the running segment identifier are multiplied by the outer product and then passed backward. The final layer of the network outputs the upper and lower quantiles of the standardized residuals. The training samples of the conditional quantile network are only taken from the phase segments where the image update gate is open and confirmed to be normal. Its training objective consists of quantile loss, monotonicity penalty term, and consistency regularization term that constructs augmented samples by setting the statistical summary of the residual displacement to zero. Coverage calibration adjusts the upper and lower quantiles on the calibration set confirmed to be normal by the ratio of the actual over-limit ratio to the target risk share, so that the over-limit ratio under normal conditions matches the target risk share.

[0011] Furthermore, the overall false alarm budget is weighted by the current coated length and the length of the electrode remaining in the oven and tightens as the value of work-in-process increases. The expected risk share of each monitored variable is formed by its safety importance, the number of false alarms and missed alarms in the recent feedback window, and the reliability of data quality. Under the constraint that the sum of the weighted risks of each monitored variable equals the overall false alarm budget, and with the preset lower and upper bounds of the share, the risk share of each monitored variable is obtained by a closed-form solution. The multiplier corresponding to the budget constraint is obtained by finding the root of the constraint equation through bisection. The monitoring parameter vector consists of the monitoring upper and lower limits, hysteresis band, effective duration of phase, and sampling and waveform retention budget. The monitoring upper and lower limits are located at the condition center of the conditional baseline, and the upper and lower offsets are multiplied by the conditional scale of the conditional baseline by the upper and lower quantiles after coverage calibration. The sampling and waveform retention budget increases with the increase of the cumulative residual displacement.

[0012] Furthermore, the mutation channel of the dual-channel evidence calculation measures the effective timing of the phase of the over-limit state. The effective timing of the phase is only accumulated in the phase segment with the same phase type as the initial phase of the current over-limit. In the blank phase segment, it is suspended and the existing accumulated value is maintained and not cleared to zero. The slow drift channel of the dual-channel evidence calculation takes the normalized ratio of the residual displacement accumulation relative to the preset test threshold as evidence. The two channels of evidence are recursively fused to form cumulative evidence. After the cumulative evidence reaches the preset evidence threshold for each level and is maintained for a preset duration according to the effective phase timing, the corresponding level of warning conclusion is output. The warning conclusion is cleared after the cumulative evidence is lower than the preset evidence threshold for the corresponding level minus the preset clearing backlash and is maintained for a preset recovery time.

[0013] Furthermore, the type-based flow divides the confirmation results into five categories: when the confirmation is a normal operating condition drift, it is marked as normal confirmation and allowed to follow the baseline update, and after the confirmation duration, the anchoring baseline is changed, and the corresponding phase segment enters the training set and calibration set; When a fault is confirmed to be real, the corresponding phase segment and its original waveform are written into the fault memory, prohibited from entering the training set and calibration set, and the anchoring baseline is frozen to maintain the version before the fault. When a false alarm is confirmed to be false, the unit false alarm handling cost weight of the corresponding monitoring variable is increased, and the corresponding phase segment is incorporated into the calibration set. When a missed report is confirmed, the expected risk share and safety importance of the corresponding monitoring variable are increased, the base value of its sampling and waveform retention budget is increased, and the missed report segment is written into the fault memory bank; when a sensor is confirmed to be faulty, replaced, or recalibrated, a structural change mark is written and an observation period is started.

[0014] The beneficial effects of this invention are as follows: The system maintains the online tracking baseline and the anchored baseline, which is replaced as a whole only after confirmation. The center displacement of the two is projected onto the working condition subspace spanned by known external factors such as formula factors, operating section factors, and ambient temperature and humidity, and feed temperature. The residual displacement that cannot be explained by the working condition is checked separately for directional consistency accumulation. The criterion is independent of the updated baseline itself. The profile update channel can be closed before the cumulative amount of residual displacement exceeds the test threshold. This avoids the normal baseline drifting in the same direction as the equipment slowly degrades, which may cause false negatives. It can also distinguish the three different types of baseline movement: normal working condition drift, fault evolution, and structural change, so that maintenance personnel can handle them according to their types. By using conditional quantiles calibrated with coverage to ensure that the probability of each monitored variable exceeding the limit under normal conditions matches the allocated risk share, and under the overall false alarm budget constraint that tightens in real time with the value of work-in-process, the risk share of each monitored variable is solved jointly using a closed-form solution. This then collaboratively generates monitoring upper and lower limits, hysteresis band, effective phase duration, and sampling and waveform retention budgets, enabling multiple monitoring parameters to change jointly under the same conditional baseline and the same budget. This reduces contradictory configuration combinations between parameters and allows for real-time solution and parameter distribution on edge monitoring devices with limited computing power, demonstrating engineering feasibility. Attached Figure Description

[0015] Figure 1 This is an overall structural diagram of an industrial monitoring parameter adaptive early warning method for equipment operation profiling according to the present invention; Figure 2 This is a flowchart of an industrial monitoring parameter adaptive early warning method for equipment operation profiling according to the present invention; Figure 3 This invention provides a flowchart of an adaptive early warning method for industrial monitoring parameters based on equipment operation profiles. Figure 1 ; Figure 4 This invention provides a flowchart of an adaptive early warning method for industrial monitoring parameters based on equipment operation profiles. Figure 2 . Detailed Implementation

[0016] The subject matter described herein will now be discussed with reference to exemplary embodiments. It should be understood that these embodiments are discussed only to enable those skilled in the art to better understand and implement the subject matter described herein, and changes may be made to the function and arrangement of the elements discussed without departing from the scope of this specification. Various processes or components may be omitted, substituted, or added as needed in the examples. Furthermore, some features described in the examples may be combined in other examples.

[0017] At least one embodiment of the present invention discloses an adaptive early warning method for industrial monitoring parameters of equipment operation profile, such as... Figure 1 The diagram shows the overall structure of the method of this invention. This invention consists of three layers: an industrial field data layer, an adaptive early warning processing unit, and an edge monitoring execution unit. The industrial field data layer aggregates multi-source sensor data, coating valve on / off data, encoder phase data, formula and PLC process control data, equipment ledgers and maintenance records, and work-in-process length information to form an operation monitoring signal, which is then sent to the adaptive early warning processing unit. The adaptive early warning processing unit sequentially executes phase alignment and quality calculation, factorized profile and conditional baseline synthesis, dual-timescale residual displacement accumulation and update gate, risk budget constraint parameter vector generation, dual-channel evidence calculation and hierarchical early warning, and confirmation diversion and feedback update along the operation profile and parameter bus, sending monitoring parameters and early warning commands to the edge monitoring execution unit. The edge monitoring execution unit completes parameter vector distribution and sampled waveform retention, dual-channel judgment and hierarchical early warning output, and the confirmation results are fed back to the confirmation diversion and feedback update stage by the manual and maintenance confirmation interface, thereby updating the hierarchical operation profile, anchored baseline, conditional quantile network parameters, and overall false alarm budget. Specific steps are as follows: Figures 2 to 4 As shown, it includes: S1: Acquire multi-source sensing data, coating valve on / off signal, encoder phase and formula number, aggregate phase by phase along on / off edge and calculate missing step and noise variable by variable to obtain phase feature sequence and data quality confidence vector; Specifically, the acquired data comes from four channels: the first channel is equipment-side sensing data, including triaxial vibration acceleration at the drive and non-drive ends of the coating roller, vibration of each bearing seat in the vacuum pump group, bearing temperature of the pump and fan, motor winding temperature, motor current and torque, pump outlet vacuum, and fan air pressure and air volume; the second channel is process and control signals provided by the programmable logic controller, including formula number, coating valve on / off signal, roller surface linear speed, tension setting and actual measurement, start / stop and speed increase / decrease commands, coated length of the master roll, and length of the electrode sheet remaining in the oven; the third channel is environmental measurement points, including workshop temperature and humidity and slurry inlet temperature; the fourth channel is equipment ledger and maintenance system, including equipment model, installation location, sensor configuration, criticality, maintenance and calibration history; The sampling rate of each channel differs from the clock source. The vibration channel operates in the kilohertz range, while the temperature and vacuum channels operate in the second range. The process signals are refreshed with the control cycle. This step does not use a fixed duration window for aggregation. Instead, it uses the rising and falling edges of the coating valve on / off signal as phase boundaries to divide the time axis into a coating phase segment and a blank phase segment. Each phase segment is treated as an independent statistical unit. Under intermittent coating processes, the distribution of monitored quantities in the same equipment under the same health condition changes with the valve status. In the coating phase segment, the slurry passes through the coating head and the pump unit bears the adsorption load. In the blank phase segment, the load is unloaded and the excitation structure of the transmission system changes. If a fixed duration window is used, the window will cross the on / off edges and mix the two distributions. The mixing ratio is determined by the valve cycle duty cycle, which changes with the formula. The statistics under the same health condition will shift with formula switching, and the deviation calculated in subsequent steps will not be able to characterize the equipment health condition. Within each phase segment, the root mean square time domain, peak factor, kurtosis, frequency band energy distribution, spectral entropy, and energy proportion of the envelope spectrum at the characteristic frequencies and lower harmonics of the bearing outer ring, inner ring, rolling elements, and cage are extracted for the vibration channel. The characteristic frequencies are calculated online from the speed encoder readings and bearing geometric parameters. The effective mean of the speed is taken within the phase segment. The encoder phase is also used to divide the vibration samples into bins according to the circumferential position of the roller surface to distinguish between local defects on the roller surface and bearing defects. For channels such as temperature, vacuum, wind pressure, air volume, current, and torque, the effective mean, range, and linear trend slope are extracted within the phase segment. The above features are robustly standardized to form the phase-specific feature sequence of the phase segment. This standardization is performed independently on each monitored variable and is independent of the operating condition. It is only used to remove dimensions and unify the magnitude. The position and scale used are the robust position and robust scale of each variable, taken from the anchored baseline version and updated with each version change, so that the features of different periods are comparable on the same scale. The positional differences related to the operating condition are not handled here, but are characterized by S2 on this standardized scale using the overall benchmark and main effects of factors. Phase-separated aggregation also produces an additional quantity, namely the current valve cycle duty cycle; this quantity is calculated from the time interval between adjacent on and off edges, and then enters S2 as a component of the formulation factor, so that the duty cycle change is explicitly expressed, rather than implicit in the feature drift. In one embodiment of the present invention, while obtaining the phase-sequence feature, this step calculates the data quality reliability for each monitored variable to distinguish the influence of high-quality channels from low-quality channels in subsequent steps. The calculation covers three items: the sampling missing ratio within the phase segment, the time synchronization deviation of the channel relative to the common time reference, and the noise intensity. The time synchronization deviation is obtained by robustly estimating the difference between the response time of each channel near the coating valve on / off edge and the reference time, using the coating valve on / off edge as the common event reference, and then normalizing it according to the dynamic response time constant of the channel. This time constant is taken from the sensor manual or obtained by identifying the first-order response time constant of the channel online using the coating valve on / off edge as a known step excitation. The noise intensity is expressed as the ratio of the robust scale of the high-frequency differential sequence within the phase segment to the robust scale of the signal itself. This ratio is dimensionless and facilitates comparison between different physical quantities. Data quality reliability is synthesized by the product of three factors. The first factor is one minus the sampling missing ratio within the phase segment. The second factor is one minus the normalized time synchronization deviation. The third factor is the natural exponent of the normalized noise intensity after taking the negative value. Data quality reliability is denoted as a scalar between zero and one for each monitored variable in each phase segment. Its variable subscript traverses all monitored variables. The phase segment index is a discrete time-series index according to the phase segment number. This synthesis method is not sensitive to dimensions and can be compared horizontally between different physical quantities. The three terms are multiplicated rather than weighted, so that any deterioration of one term independently lowers the confidence level and is not compensated by the good values ​​of the other two terms; when a channel has no data at all in a phase segment, the sampling missing ratio approaches one and the confidence level approaches zero; when the channel data is complete, the synchronization is good and the noise level is low, the confidence level approaches one; the normalized benchmark for the normalized time synchronization deviation and noise intensity is determined by the robust quantile of the channel in the historical phase segment that has been confirmed to be normal. Therefore, this benchmark is recalibrated when the channel is changed and is not written into the configuration as a fixed constant. This step also handles the situation where critical measurement points are interrupted. When communication is interrupted at a measurement point that has entered the safety interlock, this method does not stop the operation of subsequent steps. Instead, it sets the confidence level of the variable to near zero, causing its weight in subsequent distance, deviation, and evidence calculations to decrease accordingly. At the same time, a separate data quality alarm is output for maintenance personnel to handle. For variables that require an alarm upon measurement point failure according to safety specifications, the alarm is sent directly to the operation terminal without subsequent evidence accumulation. The output of this step is a phase-sequence feature sequence and the corresponding data quality confidence vector for each phase segment.

[0018] S2 receives the phase-sequence feature sequence and the data quality confidence vector, decomposes the working condition into the formula factor and the running segment factor, and constructs the hierarchical operating profile and condition baseline through factor prototype outer product synthesis and sample size shrinkage. Specifically, this step receives the phase-sequence feature sequence and data quality confidence vector output by S1, and organizes the current state of the equipment into a hierarchical operational profile. This profile consists of five layers: the static identity layer records the equipment model, specifications, installation location, sensor configuration, criticality level, and maintenance calibration history; the operating context layer records the current formula number, areal density setting, slurry viscosity, coating line speed, valve cycle duty cycle, current operating segment category, and ambient temperature and humidity and slurry inlet temperature; the short-term dynamic layer records the phase-sequence features output by S1 and their trends and rates of change across multiple phase segments, as well as the correlation structure between variables; the long-term memory layer stores individual biases, degradation trends, fault memory indexes, and structural change markers for each equipment; and the confidence layer records the data quality confidence vector output by S1 and the subsequently obtained matching uncertainty. This method does not treat the operating context as a whole to match a pre-established normal state prototype. Instead, it decomposes the operating condition into two categories: formulation factors and operating segment factors. The formulation factor is not a single formulation number or a single physical quantity. Instead, it is a number of formulation anchor points obtained by robustly clustering historical formulations in a normalized attribute space spanned by areal density setting, slurry viscosity, coating line speed, and valve cycle duty cycle. The number of anchor points is much smaller than the actual number of formulations. The formulation number is only used to index the coordinates of the current formulation in the attribute space. The degree of belonging of the current phase segment to the formulation factor is obtained by soft allocation based on the distance from the current formulation coordinates to the center of each anchor point. The closer the distance, the higher the degree of belonging, and the sum of the degrees of belonging of each anchor point is one. When switching formulations, the current coordinates move continuously in the attribute space, and the degree of belonging transitions continuously between adjacent anchor points. The degree of belonging of newly put into production is distributed by the neighboring anchor points according to the landing point of its attribute coordinates. Therefore, it is not necessary to establish separate anchor points for it to obtain the degree of belonging. The operating segment factor consists of five categories: standby, speed-up, stable coating, speed-down, and roll change. Its degree of belonging is determined by the start / stop and speed-up / down commands, as well as the rate of change of the roller surface linear speed and tension. The larger the amplitude of the rate of change of linear speed and the rate of change of tension, the higher the degree of belonging of the speed-up or speed-down segment. When the amplitude of both is close to zero and the linear speed is maintained near the set value, the degree of belonging of the stable coating segment is higher. The sum of the degrees of belonging of each operating segment is also one, and continuous transition between adjacent operating segments is allowed, thus describing the current operating condition with two sets of degrees of belonging together with the formulation factor. The technical basis for using factor decomposition is that the operating conditions in this embodiment are composed of the product of the formula and the operating segment. The number of combinations of dozens of formulas multiplied by five types of operating segments is much greater than the distribution density of the confirmed normal historical phase segments on each combination. If an independent normal state prototype is established for each combination, the center and scale estimation on most combinations is unreliable. If the combinations are merged into a coarse-grained prototype to meet the sample size requirements, the normal interval is widened so that early degradation does not enter the judgment range. Factor decomposition reduces the number of parameters from the level of the number of combinations to the level of the number of factors, so that a limited number of historical samples can be shared among multiple combinations. In one embodiment of the present invention, based on factor decomposition, this step maintains the formulation factor prototype and the running segment factor prototype respectively; each factor prototype stores the main effect of each monitored variable on the factor, that is, the positional shift of the factor relative to the overall baseline, and the robust scale within the factor; in addition, a set of bilinear interaction terms are maintained to represent the additional shift unique when a certain formulation and a certain running segment co-occur, such as the transmission resonance lift unique to a certain high viscosity formulation in the acceleration segment; the set of factor prototypes maintained in this step is the following baseline, and the condition center and condition scale output by S2 are synthesized from the following baseline, and its frozen copy at the time of version change is used as the anchoring baseline of S3; The overall baseline, the main effects of the two types of factors, the bilinear interaction term, and the individual bias were all calibrated on historical phases that were confirmed to be normal and whose data quality reliability met the requirements for inclusion in the database. The calibration was weighted by the data quality reliability and adopted a loss robust to outliers. It was solved step by step according to the additive factor model. First, the weighted robust center of all normal samples was taken as the overall baseline. Then, after deducting the overall baseline, the main effects of the two types of factors were estimated by weighting by the degree of belonging. Subsequently, after deducting the two types of main effects, the bilinear interaction term was estimated by combining the residuals. Finally, the residual systematic bias of the same equipment individual was shrunk to zero and taken as the individual bias. The robustness scale within the factor was taken as the robust dispersion of the corresponding sample after deducting the center. The overall baseline here is the overall position on the S1 standardized scale. It is different from the dimensionless coarse standardization division of each variable in S1. The two are not deducted repeatedly. The main effects of the factors describe the operating condition deviation of each formulation anchor point and each operating segment relative to the overall baseline. The conditional center of the current combination is synthesized by the following formula:

[0019] in, To monitor variables In the phase segment The condition center; This serves as the overall baseline for the variable across all confirmed normal samples; Formulating factors The main effects of this variable; For the running segment factor The main effects of this variable; Formulating factors With running segment factor Bilinear interaction terms; and These are the degrees of belonging of the current phase segment to the formula factor and the running segment factor, respectively, and the sum of each with respect to the subscript is one. The coefficient of contraction for the interaction term; This represents the individual bias of the device on this variable; and These are the summation indices for the formula factor and the running segment factor, respectively;

[0020] The shrinkage coefficient of the interaction term in the above formula is given by the effective normal sample size of the current formula and running segment combination. Its value is equal to the effective normal sample size of the combination divided by the sum of the effective normal sample size of the combination and the shrinkage baseline sample size. The effective normal sample size of the combination is the number of historical phase segments in the current combination that have been confirmed as normal and whose data quality reliability meets the requirements for inclusion in the database. When including them, each phase segment is weighted according to its average data quality reliability. The shrinkage baseline sample size is used to determine the transition interval of the interaction term from being suppressed to fully participating. This value form is a common practice of empirical Bayes shrinkage. The more samples in the combination, the closer the shrinkage coefficient is to one, and the fewer samples, the closer it is to zero. The shrinkage coefficient constrains the credibility of the interaction terms with the size of the supporting sample. When there are sufficient historical samples for a combination, the shrinkage coefficient approaches one, the interaction terms fully participate in the synthesis of the condition center, and the condition center approximates the actual position of the combination. When a combination has never appeared or the sample is insufficient, the shrinkage coefficient approaches zero, and the condition center degenerates into the sum of the overall baseline and the two types of main effects, that is, it is extrapolated from the performance of the formula in other operating segments and the performance of the operating segment in other formulas. Accordingly, when a new formula is put into production for the first time, usable condition centers can still be obtained without accumulating samples in all operating segments of the formula in advance. The shrinkage baseline sample size is obtained by taking the sample size corresponding to the robustness of the interaction term estimation and the robustness of the main effect estimation reaching the same order of magnitude on the confirmed normal historical data. For example, the default value of the shrinkage baseline sample size in this embodiment is the number of phase segments of a single formula in a stable coating segment of the production line in one shift. The composition of the conditional scale includes both the internal fluctuations of the factors and the differences between the factor centers. Its value is the square root of the sum of the following two types of weighted contributions: The first type of contribution is a weighted sum of the formula factors according to their affiliation, and the contribution of each formula factor is the square of the factor's internal robustness scale, plus the square of the difference between the factor's main effect and the mean of the formula's main effect weighted by affiliation; The second type of contribution is a weighted sum of the running segment factors according to their affiliation, and its composition is the same as that of the formula factors; The factor's internal robustness scale and main effect are both obtained from the above calibration, and the mean of the main effect is the average of the main effects weighted by the current affiliation. The squared terms corresponding to the differences between factor centers in the above synthesis play a role in the transition of operating conditions. When the formula is switched or the speed is increased or decreased, the degree of attribution is distributed among adjacent factors. The difference between the main effects causes the condition scale to temporarily increase, and the normal range is widened accordingly, so that the transition of operating conditions itself does not exceed the limit. After the transition is completed, the degree of attribution is re-concentrated and the condition scale falls back. This treatment makes the baseline at the boundary of the operating condition change continuously without the jump when switching the baseline according to the operating condition level. The standardized residuals are given by the condition center and condition scale. Their value is equal to the sum of the phase characteristics of the variable minus the condition center, divided by the condition scale and a numerically stable term to prevent the denominator from being zero. The phase characteristics are the corresponding characteristic values ​​output by S1. The matching uncertainty is the bounded weighted sum of three parts truncated to between zero and one. The three parts are the difference between the maximum attribution degree of the current phase segment on the two types of factors and one, the difference between the contraction coefficient of the current combination and one, and the average insufficiency of the data quality confidence of each variable. An increase in any of these parts will increase the matching uncertainty. The matching uncertainty is used in S4 to slow down parameter following and improve the allocation of observation resources. The output of this step is the hierarchical running profile vector, condition center, condition scale, standardized residuals, and matching uncertainty. The condition center and condition scale are collectively called the conditional baseline. The running profile vector is composed of static identity embedding, attribution degree of the two types of factors, short-term dynamic features, individual memory summary, and data quality confidence vector concatenated together.

[0021] S3, derived from the conditional baseline, follows the baseline and freezes the anchoring baseline, projects the center displacement onto the working condition subspace to separate the residual displacement, accumulates and checks according to the direction consistency, and determines the drift type and opens and closes the image update gate to obtain the cumulative residual displacement. Specifically, this step receives the conditional centers, conditional scales, and running profile vectors output by S2, and maintains two sets of baselines in the form of factor prototypes. The following baseline, which is the set of factor prototypes used by S2 to synthesize conditional centers, is tracked online with a small learning rate. That is, when the profile update gate is opened, it is weighted according to the degree of belonging and the data quality confidence, and the main effects, interaction terms, and individual biases of the factors are slowly corrected. The anchoring baseline is a frozen reference version, whose content is exactly the same as the following baseline structure, but it is only replaced as a whole when three types of version change events occur: first, it is confirmed by manual or process as a normal operating condition drift and the drift has lasted for a specified confirmation period; second, maintenance work is completed and a period of stable operation is confirmed; third, the observation period after sensor replacement or recalibration ends. Each version change generates a new version number and records the basis and time of the version change so that any judgment at any time can be traced back to the baseline version on which it was based at that time. When a production line or equipment is first connected and there is no confirmed normal history, the anchoring baseline is initially established as follows: If there is equipment of the same model and similar installation environment, its factor prototype is inherited as the initial anchoring baseline and marked as the inherited version. Then, during the observation period, the confirmed normal phase segment of this equipment is gradually corrected with a small learning rate. If there is no inheritable object, it is first operated with the manufacturer's safety boundary and the safety boundary monitoring limit given by the process specification. After this equipment accumulates a confirmed normal phase segment of the same order of magnitude as the shrinkage reference sample size, the first formal anchoring baseline version is calibrated accordingly. The position and scale used in robust standardization in S1 are taken from the robust statistics of the inherited version or the first batch of confirmed normal phase segments before the formal anchoring baseline is generated. During the observation period, the anchoring baseline version is prohibited from being changed to avoid solidifying the initial transition state into a normal reference. The baseline updated online cannot be used as a reference for whether it is contaminated. Therefore, two sets of baselines are introduced: the following baseline and the anchored baseline. If the current standardized residual is used as the criterion for allowing updates, then when the equipment slowly degrades, the position offset of each phase segment is partially absorbed by the following baseline, the standardized residual remains at a low level, the condition for allowing updates continues to be met, the degraded sample is continuously written into the normal state estimate, and the normal interval moves in the same direction as the degradation. The anchored baseline does not participate in online updates and can therefore be used as an independent reference. Based on this, the center displacement is calculated; for each monitored variable, the difference between the condition center synthesized by the following baseline and the anchoring baseline under the same formula factor attribution degree and the same operating segment factor attribution degree is taken to form the center displacement vector; comparison is made under the same set of attribution degrees to exclude the positional differences caused by the operating conditions themselves from the displacement, so that the displacement only reflects the movement of the baseline itself; the displacement calculation is completed once in each phase segment and accumulated separately according to the phase type, and the displacements of the coated phase and the blank phase are maintained separately; In one embodiment of the present invention, after obtaining the center displacement vector, this step needs to distinguish between two different types of movement. The first type is normal operating condition drift, such as when the viscosity increases due to batch change of slurry, the transmission torque, motor current and pump outlet vacuum move synchronously in the same direction, and the ambient temperature rises, causing the bearing temperature to rise overall. This type of movement should be learned, otherwise the baseline will lag for a long time and generate false alarms. The second type is fault evolution, such as when pitting occurs in the vacuum pump bearing, the energy ratio of the envelope spectrum in the octave band of the bearing's characteristic frequency rises in the same direction as the bearing housing temperature, while the operating condition variables have not changed. This type of movement should not be learned. The distinction between the two types of movement is whether the direction of movement can be explained by the operating conditions. This step constructs a subspace design matrix for the operating conditions, whose column vectors consist of formula factor components, operating section factor components, ambient temperature and humidity, feed temperature, and the sensitivity coefficients between these quantities and each monitored variable, calibrated on confirmed normal data. Subsequently, using data quality reliability as weight and adding a ridge regularization term, the central displacement vector is regressed to the subspace design matrix using weighted least squares. The weighting matrix is ​​a diagonal matrix, with diagonal elements representing the data quality reliability of each variable. The ridge regularization term is used to ensure solution stability when there is collinearity among the operating condition variables. The obtained operating condition explanation coefficient vector represents the part of the displacement that can be explained linearly by the operating conditions. Subtracting the product of the design matrix and the operating condition explanation coefficient vector from the central displacement vector yields the residual displacement, which is the component of the displacement that cannot be explained by the operating conditions. The ridge regularization coefficients are determined by taking the values ​​that minimize the robustness of the residuals in the operating condition interpretation based on confirmed normal historical data. This calibration is re-executed as the number of columns in the operating condition subspace increases or decreases. Weighted least squares uses data quality reliability as a weight to avoid channels with intermittent communication or high noise levels dominating the solution of the operating condition interpretation coefficients. Both weighted least squares and ridge regularization are conventional regression methods. The regressed object is the center displacement of the following baseline relative to the anchoring baseline, and the design matrix consists of operating condition factor components and known external factors. The residual displacement is then subjected to cumulative testing; the key feature of slow degradation is that its residual displacement increases slowly in roughly the same direction over continuous phase segments, for example, the proportion of energy in the octave band of the bearing characteristic frequency in the envelope spectrum rises in the same direction as the bearing housing temperature, while the direction of random fluctuations alternates between adjacent phase segments; to take advantage of this difference, this step first performs local normalization on the residual displacement according to the robust scale of the residual displacement corresponding to the current formulation to eliminate non-stationarity between different formulations, and then calculates the directional consistency of the normalized residual displacement for each phase segment. The directional consistency is taken as the inner product of the current normalized residual displacement vector and the unit vector of its recent moving average direction. When the direction is continuous, the inner product is close to one and is given a weight gain close to one; when the direction alternates, the inner product is close to zero and is given a weight gain close to zero. Based on this, this step performs a sequential probability ratio test on the null hypothesis and the alternative hypothesis. The null hypothesis is that the residual displacement is zero-mean noise, and the alternative hypothesis is that the residual displacement increases slowly along a certain continuous direction. The noise covariance of the null hypothesis is estimated on confirmed normal data, and the direction of the alternative hypothesis is the direction given by the recent directional consistency. This step accumulates the log-likelihood ratio of the two hypotheses segment by phase and weights them according to directional consistency. The resulting cumulative amount is the cumulative residual displacement. When the cumulative residual displacement exceeds the test threshold, it is judged as fault evolution. If it does not exceed the threshold for a long period of time, it is judged as no fault evolution. The sequential probability ratio test and its boundary determination are all conventional statistical methods. The test threshold is given by the allowable false alarm rate and allowable false alarm rate through the sequential test boundary relationship. The allowable false alarm rate and false alarm rate are further determined by the allowable inspection frequency and safety requirements. They are not manually set as independent configuration items. The test object is the residual displacement and its directional consistency after the operating condition interpretation is deducted, rather than the standardized residual itself. Therefore, the quantity on which the judgment is based is independent of the online updated following baseline. The cumulative test makes the residual displacement generated by the slow degradation in each phase segment, although not enough to exceed the limit on its own, continue to grow in the cumulative amount because of its consistent direction; the random fluctuation, due to the alternating direction, has a gain close to zero and does not accumulate under the weighting of the consistent direction; accordingly, the quantity on which the decision is based is not a function of the updated object, but the movement of the updated object relative to the frozen reference, and the homology dependency between the decision quantity and the updated object is eliminated. Based on the above quantities, this step outputs the drift type judgment result; when the interpretable component of the working condition is the main component of the displacement and the cumulative amount does not exceed the inspection threshold, it is judged as normal working condition drift, and the baseline is allowed to continue to be updated. After the drift continues until the confirmation time is reached and confirmed by manual or process, the anchoring baseline is changed; when the residual displacement is the main component of the displacement and the cumulative amount exceeds the inspection threshold, it is judged as fault evolution, the profile update gate is closed, the baseline correction is stopped, the anchoring baseline remains the version before the fault, and the features of the current and subsequent phase segments are entered into the confirmation cache; when the displacement undergoes a step with consistent phase type across multiple phases at the time of sensor replacement or calibration, and the step coincides with the time of the structural change record in the equipment ledger, it is judged as structural change, a structural change mark is written into the profile and it enters the observation period. During the observation period, the anchoring baseline is prohibited from being changed to avoid interpreting the measurement point offset as equipment degradation; The opening and closing state of the profile update gate is determined by multiple conditions: the cumulative residual displacement does not exceed the test threshold, the average data quality confidence meets the data entry requirements, there are currently no valid alarms, the manual and control systems have not marked faults or maintenance status, and it is not in the structural change observation period. When all conditions are met, the update gate is open, and the factor main effect is corrected online incrementally following the baseline. That is, an increment is added to the formula factor main effect, which is the product of the main effect learning rate, the update gate opening and closing state, the formula factor attribution degree, the variable data quality confidence, and the difference between the current phase feature and the condition center. The update gate opening and closing state is 1 for open and 0 for closed. Therefore, when the update gate is closed, the increment is zero and it remains unchanged following the baseline. The main effects of the running segment factors, bilinear interaction terms, and individual biases are corrected in the same online incremental form, and their learning rates are set independently. The correction increment of the interaction terms is also multiplied by a shrinkage coefficient to prevent the interaction terms on the sparse combination of samples from being driven by a small number of samples. The output of this step is the residual displacement, the cumulative residual displacement, the drift type discrimination result, and the profile update gate opening and closing state.

[0022] S4, the hierarchical operation profile, conditional baseline and residual displacement accumulation, along with the target risk share, are fed into the conditional quantile network. After coverage calibration and closed-loop allocation of risk shares under the overall false alarm budget that tightens with the value of work-in-process, a monitoring parameter vector is generated. Specifically, this step receives the phase-sequence output by S1, the hierarchical running profile vector, condition center, condition scale, standardized residuals and matching uncertainty output by S2, and the residual displacement, residual displacement accumulation and update gate state output by S3, to generate parameters for each monitoring variable that can be directly sent to the edge monitoring device; the core quantity for parameter generation is the conditional quantile of the standardized residuals under a given risk share; the condition center and condition scale have been given by the factor synthesis of S2, and the learning step in this step is only used to characterize the shape of the standardized residual distribution as the formula, running segment and data quality change, so that the input and output boundaries of the learning step are clear and the required sample size matches the confirmed normal sample size that can be obtained in this embodiment; unlike the practice of directly generating monitoring limits from operating conditions or directly generating parameter correction quantities from the network, in this step the condition center and condition scale are always given by the factor outer product synthesis of S2, and the network only outputs the conditional quantile of the standardized residuals, the two are separated, and what the network learns is the distribution shape rather than the monitoring limits themselves; The input to the conditional quantile network consists of five parts: the first part is a low-dimensional vector obtained by embedding the recipe identifier through the embedding layer; the second part is a low-dimensional vector obtained by embedding the running segment identifier through the embedding layer; the third part is a low-dimensional vector obtained by embedding the individual equipment identifier through the embedding layer; the fourth part is a vector obtained by temporal coding of the phase-sequence feature sequence; and the fifth part is a data quality confidence vector, matching uncertainty, statistical summary of residual displacement, and target risk share. The data flow between layers is as follows: Each of the three embedding layers maps discrete identifiers into low-dimensional continuous vectors. The recipe embedding vector and the running segment embedding vector are outer-productted and expanded row-wise to obtain a vector representing the interaction between the two types of factors. This vector, along with the two embedding vectors themselves, is passed forward, giving the network a structural pathway to express the interaction between the recipe and the running segment. The phase-sequence features are fed into a one-dimensional causal convolution stack. The dilation rate of each layer in the stack increases sequentially, ensuring that the receptive field of the deepest layer covers several complete coating beats. The output of each layer is normalized and non-linearly activated before being fed into the next layer. The convolution only uses the current and earlier phase segments in the temporal direction, excluding future phase segments. The phase type is fed in parallel as an additional input channel, allowing the coated phase and the blank phase to be distinguished within the same network. Separate without mixing; the output of the final layer of the convolution stack is aggregated along the time direction to obtain the temporal coding vector; after the above vectors are concatenated with the fifth part, they enter two fully connected layers, and the final layer outputs two scalars, which are the upper quantile and lower quantile of the standardized residuals; the target risk share enters the concatenation in scalar form, so that the same set of network parameters covers all risk levels, without the need for separate training for each risk level; for example, in this embodiment, the default value of the number of layers of the one-dimensional causal convolution is the number of layers required to cover at least three complete coating beats in the deepest receptive field, the default value of the dilation rate is to increase layer by layer in powers of two from the first layer, and the default value of the output dimension of the three embedding layers of formula, running segment and individual equipment is the minimum dimension that makes the same type of equipment separable in historical data and does not produce overfitting; The training process is constrained by the operating conditions of this embodiment; training samples are only taken from phase segments whose image update gate is open and which are confirmed to be normal by manual or process control. Phase segments confirmed to be faulty, phase segments in the pending confirmation cache, and phase segments in the structural change observation period are not included in the training set. This sample selection rule is used to prevent the network from learning a degenerate distribution and to connect the judgment result of S3 with the learning link of this step; samples are sampled hierarchically according to the combination of formula and running segment, so that combinations with fewer samples have a chance to appear in each round of training; each sample randomly samples several risk share values ​​during training to participate in the forward calculation, so that the network is supervised throughout the entire risk range; The training objectives consist of three parts:

[0023] in, For training objectives; For training sample index; This represents the risk share sampled in this forward calculation; For the sample Standardized residuals; This represents the upper quantile of the network output; This indicates the operation of taking the positive part; The quantile level corresponding to the upper quantile; This is a monotonicity penalty term; For consistency regularization; and The weights of the two terms are given; the target term for the lower quantile is symmetrical to the above formula, and its quantile level is half of the risk share.

[0024] The first term is the quantile loss, which applies asymmetric weights to overestimation and underestimation, causing the network output to converge to the conditional quantile at the specified quantile level, rather than the conditional mean, thus corresponding the risk share to the probability of exceeding the limit. The monotonicity penalty term imposes a constraint on the output of the same input under two adjacent risk shares. Its value is the sum of the squares of the positive parts of the amount of reverse increase in the upper quantile when the risk share increases. The upper quantile should not increase when the risk share increases. If a reverse increase occurs, it is included in the penalty. This constraint ensures that the quantile is monotonic with the share in the subsequent risk share solution process, thus making the share allocation problem have a definite solution. The consistency regularization term is the square of the difference between the network output of the original sample and the augmented sample. The augmented sample is obtained by setting the statistical summary of the residual displacement in the input to zero, so that the quantile is mainly determined by the working condition and data quality, avoiding the passive widening of the normal range during the degradation process. The network output also needs to undergo coverage calibration to compensate for the distribution differences between the training set and the current operating state. The calibration set takes the most recent confirmed normal phase segment, and the proportion of the absolute value of the standardized residual that exceeds the actual limit of the network output quantile is calculated on it. The ratio of this actual limit-exceeding proportion to the target risk share is used as a multiplicative calibration coefficient and applied to the quantile of the network output. When the actual limit-exceeding proportion is higher than the target risk share, the coefficient amplifies the quantile, and vice versa. The calibration coefficient is maintained in groups according to phase type and data quality, and is updated on a rolling basis with the calibration set. After calibration, the limit-exceeding proportion under normal conditions is consistent with the set risk share, and the risk share thus has a verifiable meaning. When the calibration coefficient continues to deviate and exceed the specified range, network retraining is triggered. In one embodiment of the present invention, after obtaining calibrable conditional quantiles, this step allocates sensitivity among multiple monitoring variables. The allocation is based on the overall false alarm budget, which in this embodiment varies with the work-in-process status. When the master roll begins coating, the scrap length caused by unplanned downtime is limited. As the coated length and the length of electrode sheets remaining in the oven accumulate, the scrap and re-threading losses caused by downtime increase. Therefore, the unit false alarm handling cost is weighted by the current coated length and the length of electrode sheets remaining in the oven, and the overall false alarm budget tightens as the value of the work-in-process increases. The expected risk share of each variable is formed by the safety importance of the variable, the number of false alarms and missed alarms of the variable in the recent feedback window, the current data quality reliability, and the matching uncertainty. Variables with high safety importance and recent missed alarms receive a higher expected share. The risk share problem is solved under budget constraints and univariate upper and lower bounds, meaning that, under the condition that the total weighted risk does not exceed the budget, the share of each variable should be as close as possible to its expected share. The solution to this problem has a closed-form solution:

[0025] in, To assign to variables Risk share; Its expected share of risk; This represents the unit cost of handling false alarms for this variable in its current work-in-process status. Penalty weight for share deviation from expected value; The Lagrange multipliers corresponding to the budget constraint; and These are the lower and upper bounds of the univariate share. For the current overall false alarm budget; This means restricting the value to within the given upper and lower bounds; To monitor the number of variables;

[0026] The left side of the above equation decreases monotonically with the multiplier. The multiplier can be obtained by finding the root of the constraint equation using the bisection method. The root-finding process only involves scalar operations and can be completed within one control cycle of the edge monitoring device. When the weighted risk calculated according to the lower bound of each variable still exceeds the budget, it indicates that there is a conflict between the budget and the safety lower bound. At this time, the pre-set safety priority parameters are used and the configuration conflict is reported. No solution is performed for unsolvable cases. The penalty weight is determined so that the share of variables with high safety importance decreases less than that of other variables when the budget is tightened. The parameter vector consists of four types of parameters; the upper and lower limits of monitoring are synthesized from the condition center, condition scale, and calibrated condition quantiles, i.e., the condition center is used as the position, and the upper and lower offsets are obtained by multiplying the condition scale by the upper and lower quantiles under the corresponding risk share; the hysteresis band is taken as the bounded weighted sum of the condition scale and noise intensity, which increases as both increase, and is clipped to between the lower and upper limits of the hysteresis band. This is used to prevent frequent reversals of the over-limit state when the monitored feature fluctuates around the limit. When the operating condition changes and the condition scale increases or the channel noise is too high, the hysteresis band is widened accordingly to suppress false penetrations. The effective duration of the phase is determined by multiplying the base duration by an adjustment factor and then trimming it to between its lower and upper bounds. The numerator of this adjustment factor increases with increasing noise intensity, while the denominator increases with increasing importance and data quality weighted average of the standardized residual amplitudes and variable importance. Thus, when noise is high, the persistence criterion is appropriately lengthened to suppress false alarms, while when there is a significant deviation from the overall trend or when variable importance is high, the persistence criterion is shortened to respond as early as possible. The base duration, noise intensity, average standardized residual amplitudes, and variable importance are all quantities given in the preceding steps. The sampling and waveform retention budget is determined by multiplying the base value by an amplification factor of not less than one and then clipping it to between its lower and upper bounds. This amplification factor increases with the normalized ratio of the residual displacement accumulation to the test threshold and with the matching uncertainty. The upper bound of the sampling and waveform retention budget is jointly determined by the sensor capability, network bandwidth, and edge storage quota, where the residual displacement accumulation and the test threshold are derived from S3, and the matching uncertainty is derived from S2. This step drives the observation density based on the ratio of the cumulative residual displacement to the test threshold, rather than the current deviation amplitude. In the early stage of slow degradation, the deviation amplitude is still low, while the cumulative residual displacement has already begun to rise. Therefore, the retention of high-frequency waveforms is increased before the deviation takes shape, preserving the original data for subsequent manual review. When the total budget reaches the upper limit of the resources of the edge device, the variables are allocated according to their importance and current evidence, prioritizing the variables that enter the safety interlock, while the lower limits of the remaining variables are retained. The above four types of parameters for all variables constitute the candidate parameter vector. The actual issued vector smooths the candidate vector. The smoothing coefficient is determined by the average data quality confidence and matching uncertainty. When the data quality is low or the matching uncertainty is high, the smoothing coefficient decreases and the parameters follow more slowly. For parameters that require discrete levels, such as sampling and waveform retention budget, they are mapped to the nearest allowable level after smoothing. The output of this step is the monitoring parameter vector of each monitored variable.

[0027] S5, based on the monitoring parameter vector, performs dual-channel evidence verification of hysteresis over-limit judgment and residual displacement accumulation on each phase segment of the phase characteristic sequence, synthesizes alarms according to phase effective timing and clearing hysteresis, and obtains graded early warning conclusions and sampling execution instructions; Specifically, this step receives the monitoring parameter vector and calibrated conditional quantiles output by S4, the phase-separated characteristics output by S1, the standardized residuals output by S2, and the cumulative residual displacement output by S3; the evidence verification is divided into two channels, corresponding to the amplitude deviation under intermittent coating process and the displacement deviation under slow degradation, respectively. The mutation channel handles amplitude deviations. For each monitored variable, the current phase characteristic is compared with the upper and lower limits containing the hysteresis band. When the characteristic exceeds the sum of the upper limit and the hysteresis band or exceeds the difference between the lower limit and the hysteresis band, the variable is determined to be in an out-of-limit state. The immediate evidence after the out-of-limit is determined by the out-of-limit magnitude, variable importance, and data quality confidence. That is, the part of the standardized residual magnitude that exceeds the calibrated conditional quantile is taken as the magnitude term, multiplied by the variable importance weight and data quality confidence, and then normalized by the weighted sum of all variables. A single variable with low data quality confidence is not enough to form high-level evidence on its own, but its out-of-limit can form evidence together with the out-of-limit of other variables with good data quality in the same direction. The duration of exceeding the limit is measured using phase-based effective timing. The timer only accumulates within the phase segment of the same type as the initial phase of the current exceedance. During the blank phase, the timer is suspended and maintains the existing accumulated value. It continues to accumulate after returning to the same type of phase segment without being reset. In intermittent coating processes, the load is removed during the blank phase and the monitored quantity returns to near the baseline. If the measurement is continuously based on physical time, the continuity criterion will be repeatedly interrupted by the blank phase and cannot be met. The slowly evolving deviation will not trigger an alarm for a long time. The timer is reset when the formula is switched and the operating segment is switched, and the reset event and the accumulated value at that time are recorded together to avoid misjudgment caused by accumulation across operating conditions. In one embodiment of the present invention, the slow drift channel handles displacement-type deviations; this channel does not pass through upper and lower limits and hysteresis bands, but directly uses the normalized ratio of the residual displacement accumulation output by S3 relative to the test threshold as evidence; the deviation caused by slow degradation in each phase segment is less than the hysteresis band width for a period of time. If it is uniformly passed through the abrupt change channel, the deviation will be absorbed by the hysteresis band and will never enter the over-limit state; the slow drift channel allows this part of the information to be included in the evidence synthesis; The evidence from the two channels is fused through exponential weighting to form cumulative evidence. That is, the cumulative evidence of the current phase segment is equal to the evidence history retention coefficient multiplied by the cumulative evidence of the previous phase segment, plus one minus the retention coefficient and then multiplied by the weighted sum of the evidence from the two channels in the current phase segment. The weighted sum of the evidence from the two channels is the sum of the instantaneous evidence from the mutation channel and the evidence from the slow drift channel multiplied by their respective weights, and the sum of the two weights is one. The evidence from the slow drift channel is the normalized ratio of the cumulative residual displacement to the test threshold. The evidence retention coefficient determines the time required for a weak but persistent deviation to reach the alarm condition. It is determined by ensuring that the accumulated evidence on confirmed normal data does not exceed the minimum level threshold, while ensuring that the historical confirmed fault samples are accumulated before they can be detected. The weights of the two channels are determined according to the target failure mode of the equipment. For vacuum pump sets and circulating fans, where slow wear is the main failure mode, the weight of the slow drift channel is higher than that of the sudden change channel. For transmission mechanisms, where impact and jamming are the main failure modes, the weights of the two channels are close. The grading determination uses a fixed grading evidence threshold. Since the conditional quantile has already been calibrated for coverage, the accumulated evidence has a meaning corresponding to the probability of exceeding the limit, and the grading threshold does not need to change with the profile, thus reducing the amount of online adjustment required. After the accumulated evidence reaches a certain level threshold and remains there for the effective duration of the phase corresponding to that level, a warning conclusion for that level is output, with levels being prompt, warning, and severe in order. The alarm clearing requires the accumulated evidence to be lower than the level threshold minus the clearing hysteresis, and to maintain a specified recovery time. The clearing hysteresis is used to prevent the alarm state from repeatedly starting and stopping near the threshold. When multiple level conditions are met simultaneously, the highest level is output. This step also sets up a tiered access mechanism; when a variable entering a safety interlock exceeds its limit, multiple high-importance variables exceed their limits in the same direction within the same phase type, or the residual displacement shows a consistent step change across phases before the structural change marker is confirmed, the level is directly upgraded to the corresponding level without waiting for the duration; in addition, this step converts the sampling and waveform retention budget output by S4 into execution instructions and sends them to the edge monitoring device. The instructions include the sampling level of each variable, the retention duration of the original waveform, and the upload priority; execution is carried out within the upper limit of the resource according to the importance of the variable and the current evidence, with variables entering the safety interlock given priority, and the lower limit of the remaining variables retained; the warning conclusion, the evidence that triggered the conclusion, the baseline version used at that time, and the parameter vector are all written into the record.

[0028] S6 collects the manual and maintenance confirmation results of the graded early warning conclusions and sorts them by type. It is used for anchoring baseline update, fault memory writing, training set governance and budget weight correction, and updates the hierarchical operation profile, anchoring baseline, conditional quantile network parameters and overall false alarm budget. Specifically, this step receives the graded early warning conclusions and their records output by S5, and waits for confirmation results from the operation end, process end, and maintenance end. The confirmation results are divided into five categories, each leading to a different update path. Different types of confirmation results have different meanings for profiles, fault memories, and budgets. If retraining is performed in the same way, operating condition information and fault information will be mixed together. The first type is confirmed as normal operating condition drift, such as when the operating range of a new formula is confirmed to be reasonable after the process is confirmed, or when the displacement of torque and vacuum degree after the batch change of slurry has been explained by the process. In this case, the phase segment is marked as confirmed as normal, and it is allowed to continue to update with the baseline. After the drift continues until the confirmation time, the anchoring baseline is updated, and a new version number is generated. At the same time, the sample of this segment is added to the training set and coverage calibration set of the conditional quantile network. The second type is confirmed as a real fault. In this case, the phase segment and its original waveform are written into the fault memory, the failure mode and location are marked, and it is prohibited from entering the training set and calibration set. The anchoring baseline is kept at the version before the fault, so that the baseline before and after the fault can be traced back after maintenance. The differences are as follows: The third category is confirmed as a false alarm. In this case, the unit false alarm handling cost weight of the variable in the current work-in-process status is increased, or its base duration is extended, and the segment is incorporated into the coverage calibration set to correct the calibration coefficient; The fourth category is confirmed as a missed alarm. In this case, the expected risk share and safety importance of the variable are increased, the base value of its sampling and waveform retention budget is increased, and the missed alarm segment is written into the fault memory bank as a real fault; The fifth category is confirmed as a sensor fault, replacement, or recalibration. In this case, a structural change mark is written into the profile and an observation period is started. During the observation period, the anchoring baseline is prohibited from being changed, and the time synchronization deviation normalization benchmark and noise intensity normalization benchmark of the channel are recalibrated. After maintenance is completed, this step requires the equipment to run stably for a period of time and be manually confirmed before the maintenance follow-up baseline is upgraded to the new anchor baseline version, so as to avoid solidifying the transitional state during the maintenance and commissioning process into the normal reference. In addition to being used for backtracking, the samples in the fault memory bank are also used to check for missed reports, that is, to replay historical fault segments under the new parameter vector and check whether they can form sufficient accumulated evidence before the original fault occurred. In one embodiment of the present invention, this step performs governance on the training set and calibration set of the conditional quantile network; the training set only accepts phase segments whose image update gate is open and which are confirmed to be normal, and the sources include the first type of confirmation results and confirmed normal segments that have been continuously run in normal operation without triggering alarms; the phase segments in the pending confirmation cache do not enter the training set before obtaining confirmation, and after obtaining confirmation, they are classified into the training set or the fault memory according to their type; the training set maintains the sample count according to the combination of recipe and running segment, and this count is also used as the input of the shrinkage coefficient in S2. Therefore, in the process of accumulating samples, the participation of the interaction term of the new recipe and the fit of the network to the combination increase synchronously; when the coverage calibration coefficient continuously deviates from and exceeds the specified range in multiple phase types, or when the number of new recipes reaches the specified value, the network is triggered to retrain. The retraining uses the governed training set, and the coverage calibration is re-executed after the retraining is completed. The overall false alarm budget is adjusted based on the false alarm and missed alarm statistics within the feedback window. Specifically, the overall false alarm budget for the next feedback window equals the current overall false alarm budget plus an increment, then trimmed to between the lower and upper bounds of the budget. This increment is the budget learning rate multiplied by the following difference: the product of the missed alarm feedback weight and the missed alarm rate of this window minus the product of the false alarm feedback weight and the false alarm rate of this window. When missed alarms increase, this difference is positive, and the budget is adjusted upwards; when false alarms increase, this difference is negative, and the budget is adjusted downwards. This adjustment is a regular recursive update based on feedback error adjustments. The adjustment target is the total overall false alarm budget, which varies with the value of work-in-process. The correction applies to the total budget, while the share allocation among the variables is still solved by equation (3) under the new budget. Therefore, the budget adjustment is still constrained by the weighted relationship between safety importance and work-in-process value. When underreporting increases, the budget is increased, so that high-importance variables can obtain higher sensitivity. When false alarms increase, the budget is decreased, so that the overall criterion tends to be robust. The feedback weight makes the weight of underreporting items higher than the weight of false alarm items in safety-critical scenarios. The output of this step is the updated hierarchical operational profile, anchored baseline version, conditional quantile network parameters, and overall false alarm budget. The operational profile and anchored baseline are fed back to S2 to participate in the next round of conditional baseline synthesis, the network parameters and budget are fed back to S4 to participate in the next round of parameter vector generation, and the fault memory and structural change markers are fed back to S3 to participate in drift type discrimination.

[0029] In one embodiment of the present invention, the above method is applied to the operation monitoring of the electrode vacuum adsorption pump group and the oven circulating fan group on the electrode coating and drying linkage production line of a power battery manufacturing enterprise; the production line switches between multiple formulas according to orders within a shift, the coating valve is turned on and off according to the production rhythm, and the high-frequency waveform retention quota of the edge monitoring device is shared with the process network; the system divides the coating phase and blank phase according to the on and off edge of the coating valve according to S1, and outputs phase characteristics and data quality reliability according to the phase segment. The initial data example is shown in Table 1, and each row in the table corresponds to the aggregation result of a coating phase segment; Table 1. Initial data example of phase-separated characteristics and data quality reliability;

[0030] The system decomposes formulations F-03 and F-07 into formulation factors according to S2, and decomposes the speed-up and stable coating into operating segment factors. Since formulation F-07 is a newly introduced formulation on this production line, there are insufficient historical samples of its combination with the speed-up segment, the shrinkage coefficient of the interaction term is low, and the condition center is mainly obtained by extrapolation from the overall baseline and two types of main effects, while the matching uncertainty increases. In the third phase segment, the speed-up process is underway, and the attribution degree of the two types of factors is distributed among adjacent factors. The difference term of the factor center in the condition scale temporarily widens the normal range. Therefore, the root mean square of vibration and the increase in current in this phase segment do not exceed the limit. During the subsequent several hours of continuous operation, the system calculates the center displacement of the following baseline relative to the anchored baseline according to S3. After the displacement is projected onto the operating condition subspace, most of the displacement of current and wind pressure can be explained by the linear velocity and viscosity of formula F-07. However, the energy proportion of the pump group vibration envelope spectrum in the octave band of the characteristic frequency of the bearing outer ring and the displacement of the pump bearing temperature cannot be explained by the operating condition, constituting the residual displacement. The value of this residual displacement in a single phase segment is insufficient to exceed the limit, but its direction remains consistent in continuous phase segments, thus continuously increasing in the cumulative amount of sequential verification. When the cumulative amount of residual displacement exceeds the verification threshold, the system judges it as a fault evolution, the profile update gate is closed, the following baseline stops being corrected, the anchored baseline remains in the version before the degradation began, and the subsequent phase segment features enter the pending confirmation cache. Meanwhile, the system regenerates the parameter vector according to S4. Since the coated length of the master roll and the length of the electrode sheet remaining in the oven have both accumulated to a high level, the unit false alarm handling cost has increased, the overall false alarm budget has been tightened, and the risk share of most variables has been reduced accordingly. However, the safety importance of the pump vibration channel is high and there is a recent missed alarm record, so its expected risk share is relatively high. In the closed solution, the reduction of its share is limited due to its large penalty weight, thus maintaining sensitivity under the condition of budget tightening. At the same time, the sampling and waveform retention budget is increased according to the ratio of the residual displacement accumulation to the inspection threshold, and the retention time of the original waveform of the pump vibration channel is increased, which preserves the early degradation waveform for subsequent manual review. In S5, the root mean square of pump unit vibration has not yet exceeded the upper limit of the hysteresis band, and the sudden change channel has not formed immediate evidence; the slow drift channel continuously contributes evidence based on the cumulative residual displacement; after the two evidences are recursively fused, the cumulative evidence successively reaches the prompt level and the warning level, and remains in the same type of phase segment for the corresponding effective duration of the phase. During the blank phase, the timer is suspended and not cleared, so the continuous criterion is met and the warning level warning conclusion is output; based on this, the maintenance personnel retrieved the original waveform and confirmed the characteristic frequency of the bearing outer ring and its harmonic components in the envelope spectrum. After maintenance, it was confirmed to be the early stage of the outer ring peeling of the pump unit drive end bearing; In S6, this section is identified as a real fault and written into the fault memory. It is prohibited from entering the network training set and coverage calibration set. The anchoring baseline retains the version before the fault for backtracking. After the bearing is replaced, the equipment runs stably for a period of time and is manually confirmed. The following baseline is upgraded to the new anchoring baseline version. At the same time, the channel is marked with a structural change due to sensor disassembly and installation and enters the observation period. The number of missed alarms increases in this feedback window. The overall false alarm budget is adjusted upward according to the aforementioned budget correction method. The expected risk share of the pump vibration channel and the waveform retention budget base value are increased simultaneously, and the next round of operation begins.

[0031] The embodiments of the present invention have been described above. However, the embodiments are not limited to the specific implementation methods described above. The specific implementation methods described above are merely illustrative and not restrictive. Those skilled in the art can make more equivalent embodiments under the guidance of the present embodiments, and all of them are within the protection scope of the present embodiments.

Claims

1. An adaptive early warning method for industrial monitoring parameters based on equipment operation profiles, characterized in that, Includes the following steps: S1: Acquire multi-source sensing data, coating valve on / off signal, encoder phase and formula number, aggregate phase by phase along on / off edge and calculate missing step and noise variable by variable to obtain phase feature sequence and data quality confidence vector; S2 receives the phase-sequence feature sequence and the data quality confidence vector, decomposes the working condition into the formula factor and the running segment factor, and constructs the hierarchical operating profile and condition baseline through factor prototype outer product synthesis and sample size shrinkage. S3, derived from the conditional baseline, follows the baseline and freezes the anchoring baseline, projects the center displacement onto the working condition subspace to separate the residual displacement, accumulates and checks according to the direction consistency, and determines the drift type and opens and closes the image update gate to obtain the cumulative residual displacement. S4, the hierarchical operation profile, conditional baseline and residual displacement accumulation, along with the target risk share, are fed into the conditional quantile network. After coverage calibration and closed-loop allocation of risk shares under the overall false alarm budget that tightens with the value of work-in-process, a monitoring parameter vector is generated. S5, based on the monitoring parameter vector, performs dual-channel evidence verification of hysteresis over-limit judgment and residual displacement accumulation on each phase segment of the phase characteristic sequence, synthesizes alarms according to phase effective timing and clearing hysteresis, and obtains graded early warning conclusions and sampling execution instructions; S6 collects the manual and maintenance confirmation results of the graded early warning conclusions and sorts them by type. It is used for anchoring baseline updates, fault memory writing, training set governance and budget weight correction, and updates the hierarchical operation profile, anchoring baseline, conditional quantile network parameters and overall false alarm budget.

2. The adaptive early warning method for industrial monitoring parameters of equipment operation profile according to claim 1, characterized in that, Each component of the data quality confidence vector represents the data quality confidence of the corresponding monitored variable, and the data quality confidence is synthesized by the product of three factors. The first factor is one minus the sampling missing ratio, the second factor is one minus the normalized time synchronization deviation, and the third factor is the natural exponent of the normalized noise intensity after taking the negative value. The time synchronization deviation is obtained by using the on / off edge of the coating valve on / off signal as the common event reference.

3. The industrial monitoring parameter adaptive early warning method for equipment operation profiling according to claim 1, characterized in that, The formulation factor is a formulation anchor point obtained by clustering historical formulations in a normalized attribute space spanned by areal density setting, slurry viscosity, coating line speed and valve cycle duty cycle. The degree of belonging of the current phase segment on the formulation factor is obtained by soft allocation based on the distance from the coordinates of the current formulation in the normalized attribute space to the center of each formulation anchor point. The running segment factor consists of stop standby, speed increase, stable coating, speed decrease and roll change. The condition center of the conditional baseline is synthesized from the main effects of the population baseline, the formulation factor and the running segment factor, and the outer product of the bilinear interaction term after sample size shrinkage. When the normal sample size of the combination is insufficient, the condition center degenerates into the sum of the population baseline and the two types of main effects.

4. The adaptive early warning method for industrial monitoring parameters of equipment operation profile according to claim 1, characterized in that, The following baseline is weighted according to the data quality reliability when the image update gate is opened, and the main effect, bilinear interaction term and individual bias of the preset learning rate correction factor are used to freeze the anchor baseline as a frozen copy of the following baseline. It is only completely replaced when the drift under normal operating conditions is confirmed, maintenance is completed and stable operation is confirmed, and the observation period ends after sensor replacement or recalibration. The center displacement is the difference between the condition centers synthesized by the following baseline and the frozen anchor baseline under the same formulation factor and the same running segment factor.

5. The adaptive early warning method for industrial monitoring parameters of equipment operation profile according to claim 4, characterized in that, A subspace design matrix for the operating conditions is constructed, the columns of which consist of the components of the formulation factor, the components of the operating section factor, the ambient temperature and humidity, and the feed temperature. The data quality reliability vector is used as the weight and a ridge regularization term is added. The center displacement is regressed to the subspace design matrix for the operating conditions using weighted least squares to obtain the operating condition interpretation coefficient. The residual displacement is obtained by subtracting the product of the subspace design matrix for the operating conditions and the operating condition interpretation coefficient from the center displacement.

6. The adaptive early warning method for industrial monitoring parameters of equipment operation profile according to claim 5, characterized in that, The residual displacement is locally normalized according to the robust residual displacement scale corresponding to the current formula, and the dot product of the normalized residual displacement vector and the unit vector of its recent moving average direction is used as the direction consistency. A sequential probability ratio test is performed on the two hypotheses that the residual displacement is zero-mean noise and that the residual displacement increases slowly along the continuous direction. The log-likelihood ratio is accumulated segment by segment and weighted according to the direction consistency to obtain the cumulative residual displacement. When the cumulative residual displacement exceeds the preset test threshold, it is judged as fault evolution. The profile update gate is opened when the cumulative residual displacement does not exceed the preset inspection threshold, the average data quality reliability meets the preset data entry requirements, the data is not currently in an alarm state, is not marked as a fault or maintenance state, and is not in the structural change observation period.

7. The adaptive early warning method for industrial monitoring parameters of equipment operation profile according to claim 1, characterized in that, The input of the conditional quantile network consists of the embedding vectors of the recipe identifier, the running segment identifier, and the individual equipment identifier, the time-series coding vector obtained by one-dimensional causal convolution of the phase-series feature sequence, the data quality confidence vector, the statistical summary of the residual displacement, and the target risk share concatenated together. The embedding vector of the recipe identifier and the embedding vector of the running segment identifier are multiplied by the outer product and then passed backward. The network's final layer outputs the upper and lower quantiles of the standardized residuals. The training samples of the conditional quantile network are only taken from the phase segments where the image update gate is open and confirmed to be normal. Its training objective consists of quantile loss, monotonicity penalty term, and consistency regularization term that constructs augmented samples by setting the statistical summary of the residual displacement to zero. Coverage calibration adjusts the upper and lower quantiles on the calibration set confirmed to be normal by the ratio of the actual over-limit ratio to the target risk share, so that the over-limit ratio under normal conditions matches the target risk share.

8. The industrial monitoring parameter adaptive early warning method for equipment operation profiling according to claim 7, characterized in that, The overall false alarm budget is weighted by the current coated length and the length of electrode remaining in the oven and tightens as the value of work-in-process increases. The expected risk share of each monitored variable is formed by its safety importance, the number of false alarms and missed alarms in the recent feedback window, and the reliability of data quality. Under the constraint that the sum of the weighted risks of all monitored variables equals the overall false alarm budget, and with the preset lower and upper bounds of the share, the risk share of each monitored variable is obtained by a closed-form solution, and the multiplier corresponding to the budget constraint is obtained by finding the root of the constraint equation through bisection. The monitoring parameter vector consists of monitoring upper and lower limits, hysteresis band, effective phase duration, and sampling and waveform retention budget. The monitoring upper and lower limits are located at the condition center of the conditional baseline, and the upper and lower offsets are calculated by multiplying the conditional scale of the conditional baseline by the upper and lower quantiles after coverage calibration. The sampling and waveform retention budget increases with the increase of residual displacement accumulation.

9. The adaptive early warning method for industrial monitoring parameters of equipment operation profile according to claim 1, characterized in that, The mutation channel of the dual-channel evidence calculation measures the effective timing of the phase in the over-limit state. The effective timing of the phase is accumulated only in the phase segment with the same phase type as the initial phase of the current over-limit. It is suspended in the blank phase segment and the existing accumulated value is maintained and not cleared to zero. The slow drift channel of the dual-channel evidence calculation takes the normalized ratio of the residual displacement accumulation to the preset test threshold as evidence. The two channels of evidence are recursively fused to form cumulative evidence. After the cumulative evidence reaches the preset evidence threshold for each level and is maintained for a preset duration according to the effective phase timing, the corresponding level of warning conclusion is output. The warning conclusion is cleared after the cumulative evidence is lower than the preset evidence threshold for the corresponding level minus the preset clearing backlash and is maintained for a preset recovery time.

10. The industrial monitoring parameter adaptive early warning method for equipment operation profiling according to claim 1, characterized in that, The type-based triage divides the confirmation results into five categories: when the confirmation is a normal operating condition drift, it is marked as normal confirmation and allowed to follow the baseline update. After the confirmation duration is reached, the anchoring baseline is changed and the corresponding phase segment is entered into the training set and calibration set. When a fault is confirmed to be real, the corresponding phase segment and its original waveform are written into the fault memory, prohibited from entering the training set and calibration set, and the anchoring baseline is frozen to maintain the version before the fault. When a false alarm is confirmed to be false, the unit false alarm handling cost weight of the corresponding monitoring variable is increased, and the corresponding phase segment is incorporated into the calibration set. When a missed report is confirmed, the expected risk share and safety importance of the corresponding monitoring variable are increased, the base value of its sampling and waveform retention budget is increased, and the missed report segment is written into the fault memory bank; when a sensor is confirmed to be faulty, replaced, or recalibrated, a structural change mark is written and an observation period is started.