On-line detection method and device for forming defects of activated carbon

Abstract

The invention discloses an activated carbon forming defect online detection method and device, and belongs to the technical field of material defect detection.The method comprises the steps that technological parameters are collected in real time and preprocessed, a parameter sequence is generated, self-sensing recognition of a new technology is conducted, lagged starting of a quality control vacuum period is avoided, and once the new technology is recognized, the quality control vacuum period is started. An unknown defect pre-recognition mechanism is triggered immediately, novel anomalies which do not conform to known defects are recognized, and an unknown defect candidate set is generated; constructing a multi-stage hierarchical detection network, performing adaptive detection on the unknown defect candidate set, establishing a reference model, screening samples, updating the reference model, and forming a defect identification model adaptive to a new process, so that the influence of insufficient samples on the detection accuracy is effectively relieved; and setting three-dimensional quantitative indexes, judging a transition period, automatically switching to a full-model detection mode when the indexes meet preset requirements, and objectively evaluating the recognition capability of the detection system on new process defects and the process suitability through the quantitative indexes.

Description

Technical Field

[0001] The invention relates to an online detection method and device for activated carbon molding defects, and belongs to the technical field of material defect detection. Background Art

[0002] The desulfurization and denitrification activated carbon molding process involves multiple steps such as raw material mixing, extrusion, and drying. It is easy to cause defects such as surface cracks, internal bubbles, and dimensional deviations due to fluctuations in process parameters or insufficient equipment precision. Traditional detection methods rely on manual sampling or offline equipment analysis, which have limitations such as low efficiency, strong subjectivity, and inability to provide real-time feedback. It is difficult to meet the quality control needs of large-scale continuous production. With the development of industrial automation and intelligent manufacturing technology, machine vision inspection has gradually become the mainstream means of industrial defect detection due to its advantages of high precision, non-contact, and quantification.

[0003] However, existing technologies do not consider the unknown defect blind spots caused by the introduction of new processes in the desulfurization and denitrification activated carbon molding process. For example, if the current process flow changes and microwave drying and supercritical fluid activation are introduced, which are not present in the production line, especially when the new process produces defect types that do not exist in traditional processes, such as radial microcracks caused by microwave thermal stress and pore collapse caused by supercritical fluid activation, the existing detection model relies entirely on historical defect sample training and lacks the ability to recognize "new defect patterns" that are not included in the database, such as internal stress concentration caused by non-contact heating and abnormal pore structure caused by the new activation process. Manual re-labeling of samples and iteration of the model are required, which can easily lead to a quality control vacuum period in the production line during the transition period of the new process introduction. New defective products caused by fluctuations in process parameters cannot be identified and intercepted in time, which can easily cause batches of unqualified products to flow into subsequent processes, affecting the molding quality and adsorption performance stability of the desulfurization and denitrification activated carbon. In addition, the existing technology has not established a correlation prediction mechanism between process parameters and new defects, and cannot reduce the defect incidence through process regulation during the model iteration period, further exacerbating the quality control risk during the transition period. Summary of the Invention

[0004] In view of the shortcomings of the existing technology, the purpose of the present invention is to provide an online detection method and device for activated carbon molding defects. By constructing a multi-level hierarchical detection network and combining dynamic weight allocation, Bayesian active learning and pseudo-labeling technology, the problems of insufficient samples and false alarms during the transition period of the new process are solved, and the adaptive evolution and quantitative judgment of the detection system are realized.

[0005] To achieve the above object, the present invention provides the following technical solutions:

[0006] The online detection method for activated carbon molding defects includes:

[0007] Real-time collection and preprocessing of process parameters to generate parameter sequences for self-perception identification of new processes. Once a new process is identified, the unknown defect pre-identification mechanism is immediately triggered to identify new anomalies that do not conform to known defects and generate a candidate set of unknown defects.

[0008] Constructing a multi-level hierarchical detection network, performing adaptive detection on the unknown defect candidate set, establishing a benchmark model, screening samples and updating the benchmark model to form a defect recognition model adapted to the new process;

[0009] Set three-dimensional quantitative indicators to judge the transition period. When all indicators meet the preset requirements, it automatically switches to the full model detection mode.

[0010] Specifically, the self-perception identification step includes:

[0011] Establish a multi-dimensional sensor matrix, collect and pre-process data, and generate a multi-dimensional time series data set;

[0012] Set up a sliding window and calculate multi-dimensional key indicators within each window;

[0013] Through fast Fourier transform, the dominant frequency is extracted and the energy proportion of each frequency band is calculated;

[0014] For any two types of parameters, a bivariate VAR model is constructed, the Granger causal coefficient is calculated, and a causal matrix is ​​generated based on the causal threshold;

[0015] The extracted features are integrated to generate feature vectors, and the process fingerprint vector is obtained by dimensionality reduction through principal component analysis;

[0016] Define the necessary features and mutually exclusive conditions of known processes to form a production rule base. Use the rule engine to pre-screen the process fingerprint vector, calculate the rule matching score, set the matching threshold, and determine whether the current process is a known process.

[0017] If the rule matching score is greater than the matching threshold, it is determined to be a known process; otherwise, it is determined to be a process to be confirmed, and a pre-response mechanism is triggered.

[0018] Specifically, the pre-response mechanism includes:

[0019] For the process to be confirmed, the full amount of data in real time and historical periods is automatically collected, encrypted and saved, and an initial label is generated at the same time;

[0020] The historical traditional process steady-state data is obtained as a training set, and the normal process feature space boundary is constructed based on the isolation forest algorithm;

[0021] Calculate the outlier value of the process to be confirmed. If the outlier value exceeds the boundary of the normal process feature space, it is determined to be a high-confidence unknown process. Otherwise, it is temporarily stored as a low-risk anomaly and continuously monitored.

[0022] Specifically, the pre-response mechanism also includes:

[0023] Calculating the similarity between the process to be confirmed and the historical templates in the process library to be confirmed using weighted Euclidean distance;

[0024] If the similarity exceeds the preset similarity threshold, the similarity review mechanism is triggered, including: matching similar templates, determining the cluster to which the current sample belongs through K-means clustering, calculating the historical defect rate of templates within the cluster, generating a pre-classification report, and sending it to experts for review to classify it as a new process or abnormal fluctuation;

[0025] Otherwise, a deep intervention mechanism is triggered, including: calculating the contribution of the process fingerprint vector to the outlier to generate a ranking diagram, explaining the impact of simulation parameter adjustment based on counterfactuals, analyzing whether the anomaly conforms to the new process mechanism, and classifying it as a new process or abnormal fluctuation based on the set confirmation conditions of the new process;

[0026] If it is determined to be an abnormal fluctuation, the current data is added to the isolated forest noise sample set, the model boundary is retrained, and an equipment maintenance work order is triggered;

[0027] If it is determined to be a new process, the core features are extracted to generate the initial rules to inject into the rule base, the history is traced back to re-label the data to expand the training set, and the known process model corresponding to the maximum similarity is selected through the K-nearest neighbor algorithm. Continuous monitoring is carried out for three consecutive cycles. If it is determined to be a new process, the new process signal is output. Otherwise, it returns to the deep intervention mechanism for re-evaluation.

[0028] Specifically, the unknown defect pre-identification mechanism includes:

[0029] Based on the multi-dimensional time series data set, a three-dimensional coordinate system for the device is established to map the sensor detection values ​​to physical locations;

[0030] For each type of parameter, perform equal interval binning, count the number of samples in each interval, generate marginal histogram, and for parameter pairs, calculate joint histogram;

[0031] Based on the mutual information formula, the MI value of any parameter pair is calculated, and a mutual information threshold is set, and the parameter pairs with MI values ​​less than the mutual information threshold are eliminated to generate a correlation matrix;

[0032] Assigning different weights to the parameters in the association matrix based on a preset weight allocation threshold;

[0033] The filtered parameter features are standardized and linearly combined according to the corresponding weights to generate a fusion feature vector;

[0034] The cosine similarity between the fused feature vector and the traditional defect template in the known defect library is calculated, and samples with similarity less than the traditional similarity threshold are retained and marked as potential unknown defects.

[0035] Specifically, the unknown defect pre-identification mechanism also includes:

[0036] Using the forest model, for the sample of the potential unknown defect, an outlier score is calculated, and a defect outlier threshold is set. Once the calculated outlier score exceeds the defect outlier threshold, it is determined to be a high outlier sample;

[0037] Use the KD tree algorithm to calculate the LOF value. Once the LOF value is greater than the preset local outlier threshold, the sample is determined to be a local outlier.

[0038] Constructing an unknown defect candidate set, marking candidate samples in the unknown defect candidate set, and classifying the candidate samples into high risk level, medium risk level, and low risk level by setting a secondary threshold of outlier risk;

[0039] Set the secondary thresholds for adjusting the screening threshold, namely the first adjustment threshold and the second adjustment threshold. Time period, count the sample size in the unknown defect candidate set;

[0040] If the sample size is lower than the first adjustment threshold, the screening threshold is lowered and the reduction coefficient is set to; if the sample size is between the first adjustment threshold and the second adjustment threshold, the original screening threshold is maintained; if the sample size is higher than the second adjustment threshold, the screening threshold is increased and the increase coefficient is set to;

[0041] At the same time, the relaxation coefficient is set to in the new process learning mode, the tightening threshold is set to during stable production, and the threshold adjustment log is recorded.

[0042] Specifically, the step of adaptive detection includes:

[0043] Construct a multi-level hierarchical detection network, including high-level, mid-level, and low-level layers, to generate three-level detection results;

[0044] Establish a benchmark model using historical process parameters and defect characteristics, calculate the posterior distribution entropy and prediction probability variance of the parameters of the benchmark model, and calculate a comprehensive score through weighted summation;

[0045] Each week, the top samples, and mark them;

[0046] For unlabeled samples, if the high-level layer detection confidence exceeds the label threshold and the intermediate layer label is consistent with the high-level layer result, a pseudo label containing the defect type and confidence is generated;

[0047] The high-level neural network parameters are divided into a frozen general layer and a new process-specific layer that is allowed to be updated. A hybrid loss function is set based on the soft labels annotated by experts. The stochastic gradient descent algorithm is used to update only the parameters of the new process-specific layer and then merge it with the general layer.

[0048] Formulate a risk-graded response strategy based on the risk level and the three-level detection results;

[0049] When the number of consecutive occurrences of high-risk samples at the same workstation reaches the threshold and the three-level detection results are consistent, the process-end parameters are automatically adjusted, the process fingerprint vector is recalculated, and the normal process feature space boundary is updated.

[0050] Specifically, the multi-level hierarchical detection network includes:

[0051] The multi-level hierarchical detection network includes primary layer, intermediate layer and advanced layer:

[0052] The primary layer uses the constructed defect rule base to calculate the comprehensive confidence;

[0053] If the comprehensive confidence exceeds the matching threshold, the defect type is output; otherwise, no match is output;

[0054] After the intermediate layer normalizes the fused feature vector, it constructs an undirected graph based on the cosine similarity between samples and sets the initial label vector;

[0055] Update all node labels through an iterative formula, take the maximum value of the label vector after propagation for unlabeled samples as the confidence, and the corresponding label as the soft label;

[0056] The advanced layer uses a deep neural network, freezes the first three convolutional layers based on the known process detection model, and adds two fully connected layers to process cross-modal fusion features and output defect confidence;

[0057] Design a hybrid loss function that combines the cross entropy loss of expert-labeled samples and the KL divergence loss of pseudo-labeled samples, and only update the parameters of the newly added fully connected layer;

[0058] A dynamic weight allocation mechanism is set up. The proportion adjustment coefficient is normalized based on the proportion of correctly detected samples in each layer every week. After multiplying it by the preset adjustment step size, it is added to the original weight and normalized to obtain the adjusted weight, and weight constraints are set.

[0059] Specifically, the steps for determining the transition period include:

[0060] Construct a three-dimensional quantitative indicator system to calculate the detection accuracy, parameter fluctuation coefficient, number of process adjustments, model detection accuracy, and update amplitude, and output a three-dimensional indicator score table;

[0061] If any dimension fails to meet the standard, continue to implement the multi-level hierarchical detection strategy and increase the amount of active learning sample screening;

[0062] When the three-dimensional indicators simultaneously meet the preset indicator thresholds, the transition period end signal is triggered, the primary layer is deactivated, the weights of the intermediate and advanced layers are merged, and the full model detection mode is switched to. The high-level deep neural network directly detects the unknown defect candidate set, and the intermediate layer performs a secondary verification of the confidence fuzzy samples. The detection results are synchronized with the process fingerprint vector generation step to update the normal process feature space boundary;

[0063] If continuous missed detection occurs, the Bayesian active learning of adaptive detection is triggered in reverse to supplement the training data, and the full model detection result report and process fingerprint vector update record are output.

[0064] An online detection device for activated carbon molding defects includes: a data acquisition module, an identification module, and a transition detection module;

[0065] The data acquisition module is used to collect process parameters in real time and perform preprocessing to generate a multi-dimensional time series data set;

[0066] The identification module includes a process identification unit and a defect identification unit;

[0067] The process identification unit is used to calculate the similarity between the current parameter sequence and the historical process library through a dynamic time warping algorithm to identify process mode switching;

[0068] The defect recognition unit is used to perform mutual information analysis and feature fusion on the multi-dimensional time series data, filter out known defects, perform feature outlier analysis using the isolation forest algorithm and the LOF algorithm, screen abnormal samples to generate a set of unknown defect candidates, and dynamically adjust the screening threshold based on the production stage;

[0069] The transition detection module includes a layered detection unit and a transition management unit;

[0070] The hierarchical detection unit is used to construct a multi-level hierarchical detection network, adopt a dynamic weight allocation mechanism, combine Bayesian active learning and pseudo-labeling technology to update the baseline model, and perform adaptive detection of unknown defects;

[0071] The transition management unit is used to set three-dimensional quantitative indicators, determine the end conditions of the transition period, and automatically switch to the full model detection mode.

[0072] Beneficial effects of the present invention:

[0073] The dynamic time warping algorithm is used to achieve self-perception recognition of new processes in seconds, eliminating the lag of the quality control vacuum period, allowing the detection system to enter the early warning state at the initial stage of new defects, and shortening the first identification time of new defects from hundreds of sample annotation cycles of traditional methods to a short time after the process is started; through the combination of the isolation forest algorithm and the multi-level hierarchical detection network, the initial screening of unknown defects under zero-sample conditions is achieved to prevent new defects from flowing into subsequent processes. The feature fusion and outlier analysis mechanism based on the three-dimensional coordinate system can accurately locate new defects such as radial microcracks caused by microwave thermal stress that cannot be identified by traditional models; at the same time, the three-dimensional quantitative indicator system objectively evaluates the system's adaptability to new processes, and automatically switches to full detection when the detection accuracy, process adaptability, and model stability meet the standards. The model detection mode avoids the subjectivity of manual experience judgment and improves the efficiency of transition period judgment. Bayesian active learning and pseudo-labeling technology solves the problem of model iteration in small sample scenarios. Only a small number of labeled samples are needed to update the model, avoiding damage to the original detection capability. At the same time, the correlation prediction mechanism between process parameters and defects is established through parameter fluctuation coefficients and causal matrices. The defect rate can be reduced through process control during model iteration. Experimental data show that the defect rate in the transition period of the new process is effectively reduced, which effectively guarantees the molding quality and adsorption performance stability of desulfurization and denitrification activated carbon. By constructing a full-process solution covering new process identification, unknown defect pre-identification, adaptive detection and transition period quantitative evaluation, the quality control blind spot problem of existing technologies when introducing new processes is effectively solved. BRIEF DESCRIPTION OF THE DRAWINGS

[0074] Figure 1 Schematic diagram of the online detection method for activated carbon molding defects;

[0075] Figure 2 This is a flow chart of self-perception recognition in the present invention;

[0076] Figure 3 This is a flow chart of the unknown defect pre-identification mechanism of the present invention;

[0077] Figure 4 This is the structural diagram of the online detection device for activated carbon molding defects. DETAILED DESCRIPTION

[0078] The technical solution of the present invention is described in detail below through the accompanying drawings and specific embodiments. It should be understood that the embodiments of the present invention and the specific features in the embodiments are detailed descriptions of the technical solution of the present invention, rather than limitations on the technical solution of the present invention. In the absence of conflict, the embodiments of the present invention and the technical features in the embodiments can be combined with each other.

[0079] Example 1:

[0080] refer to Figures 1 to 3As shown, this embodiment introduces an online detection method for activated carbon molding defects, including the following steps:

[0081] Step S1: Deploy a sensor matrix at key workstations of the activated carbon forming production line, such as the extrusion outlet and the drying section entrance, to collect process parameters such as pressure, temperature, power, and flow in real time, and pre-process the collected data to generate a continuous parameter sequence. Use the dynamic time warping algorithm to calculate the similarity between the current parameter sequence and the historical process library, perform self-perception identification of the new process, and identify the process mode switch, thereby avoiding the delayed start during the vacuum period of quality control, and ensuring that the detection and warning state is entered at the early stage of the generation of new defects. Once a new process is identified, the unknown defect pre-identification mechanism is immediately triggered to identify new anomalies that do not conform to known defects, and screen outlier unknown anomaly samples from the production data stream to form a preliminary unknown defect candidate set, which is automatically marked and enters the review process, realizing the initial screening of abnormal defects under zero-sample conditions, shortening the first identification time of new defects, and avoiding unknown defects from directly flowing into subsequent processes, providing an accurate target sample set for transition period detection;

[0082] In this embodiment, since the new process is different from the traditional process due to the physical field, it is easy to give rise to new defects, and the detection model relies on historical sample training and lacks prior knowledge of defects that are not included in the database, thus forming a transition period. At the same time, the traditional supervised learning model cannot be directly generalized to new defect types when facing zero-sample or small-sample scenarios. During the initial inspection, it is easy to "misjudge normal products as unknown defects" or "miss new defects". It takes at least hundreds of labeled samples to gradually converge the accuracy. The dynamic time warping algorithm can complete the identification in a short time after the new process is started, avoiding the quality control vacuum period caused by the lag of manual judgment.

[0083] Specifically, the steps of self-perception recognition include:

[0084] A multidimensional sensor matrix is ​​established at key locations on the activated carbon molding production line. All sensors are precisely time-aligned through a time synchronization bus to ensure that the collected multi-source signals strictly correspond in the time dimension, avoiding feature analysis errors caused by asynchronous acquisition. This allows for the acquisition of raw data that fully reflects the physical changes in the production process. The collected data is also pre-processed, using industrial edge computing nodes to perform real-time noise reduction on the raw sensor signals. Sliding median filtering is used to remove impulse noise, and Fourier transforms are used to filter out power frequency interference. This ensures that the input signal meets the signal-to-noise ratio requirements for material defect detection, thereby generating a multi-dimensional time series dataset with timestamps.

[0085] Key locations include the extrusion molding section, microwave drying section, and supercritical activation section. In the extrusion molding section, an annular array pressure sensor is used to monitor the pressure distribution balance, and a laser rangefinder is used to measure the diameter fluctuation of the embryo in real time to capture the spatial non-uniformity of the material's plastic deformation. In the microwave drying section, a distributed infrared thermal imager is used to construct a three-dimensional temperature field, and a microwave power sensor is used to monitor the power oscillation characteristics to identify the thermal stress concentration caused by non-uniform heating. In the supercritical activation section, optical fiber strain gauges and density sensors are used to measure the channel deformation and activation efficiency.

[0086] To adapt to the dynamic changes of the process, a variable-length sliding window is applied to the multi-dimensional time series data set. The window length is automatically adjusted according to the process stability. Within each window, multi-dimensional key indicators are calculated, including the range, kurtosis, and autocorrelation attenuation coefficient of the parameters, to obtain the characteristics of the parameters in the time series. The range is the difference between the maximum and minimum values ​​of the parameter within the window, which is used to reflect the range of parameter variation. The kurtosis is used to measure the steepness of the waveform and determines whether the waveform is abnormal by calculating the degree to which the signal deviates from the normal distribution. The autocorrelation attenuation coefficient reflects the periodic changes of the signal by calculating the correlation of the signal at different times.

[0087] Perform a fast Fourier transform on the multi-dimensional time series data set to obtain the amplitude spectrum, traverse the frequency domain amplitude spectrum, and find the frequency with the largest amplitude , and defined as the dominant frequency. The frequency band width is set based on the physical mechanism of the new process. The characteristic frequency band is pre-delineated with the dominant frequency as the center. The ratio of the energy sum in the characteristic frequency band to the energy sum of the entire frequency band is calculated, thereby calculating the energy proportion of each frequency band to identify the characteristic frequency signal of a specific process and obtain the frequency domain characteristics.

[0088] For any two types of parameters, a bivariate VAR model is constructed to describe the parameter lag association, and the Granger causality coefficient between the parameters is calculated to quantify the driving relationship between the parameters. A causal threshold is set. Once the Granger causality coefficient is greater than the causal threshold, there is a causal influence between the two parameters, and a causal matrix is ​​generated.

[0089] The extracted features are integrated to generate feature vectors, and the dimension is reduced through principal component analysis to reduce the complexity of subsequent calculations and generate process fingerprint vectors to extract composite information that can characterize process characteristics;

[0090] Define the necessary features and mutually exclusive conditions of known processes to form a production rule base. Use the rule engine to pre-screen the extracted process fingerprint vectors, count the number of matching rules that match the production rule base, and calculate the rule matching score of the current process based on the ratio of the number of matching rules to the total number of rules in the production rule base. Then set a matching threshold to determine whether the current process is a known process and perform preliminary screening.

[0091] If the rule matching score is greater than the matching threshold, indicating that the current process matches the known process, it is judged as a known process, and the rationality of the process parameters is verified and the corresponding detection strategy is activated; otherwise, the current process characteristics are insufficiently matched with the known process rules, and it is judged as a process to be confirmed, triggering the pre-response mechanism.

[0092] Specifically, since the production rule base only contains historically defined process types, it cannot cover future newly added process variants or extreme working conditions. At the same time, the matching threshold is essentially a manually set probability threshold, and the boundary of the threshold is fuzzy. When the score is close to the threshold, insufficient matching may occur due to sensor noise and short-term process fluctuations. At this time, directly judging it as a new process will introduce a pseudo-new process, which is actually the edge state of a known process. In addition, the known process produces progressive feature offsets due to factors such as raw material batches and equipment aging, resulting in the rule matching score being slightly lower than the matching threshold. Before defining the process to be confirmed as a new process, a pre-response mechanism is used to ensure that the insufficient matching is due to real innovation in process characteristics, rather than data noise or feature extraction errors, thereby eliminating short-term interference and avoiding misjudgment.

[0093] Pre-response mechanisms include:

[0094] Automatically save current and past processes for confirmed processes All data within a time period, including raw sensor signals, pre-processed multi-dimensional time series datasets, and process fingerprint vectors, are encrypted and stored in HDF5 format. The file name contains the process identification timestamp and key feature summary. Initial labels are automatically generated, including suspected process type, abnormal feature summary, and equipment status.

[0095] Obtain historical traditional process steady-state data, which are manually labeled as defect-free and cover more than three equipment states, such as full load / half load / startup phase, as a training set. Based on the isolation forest algorithm, construct the normal process feature space boundary, including randomly selecting feature dimensions and split values, recursively splitting the data until a single sample is isolated or the maximum depth is reached, building 100 isolation trees to form a forest model, calculating the average path length of each sample, that is, the number of splits in which the sample is isolated in the isolation tree, converting the path length into an outlier score using the outlier score calculation formula, and selecting the 95% quantile as the normal process feature space boundary. The 95% quantile, that is, setting the unknown threshold is the maximum value of the first 5% outlier scores in the normal samples. If the outlier scores of 95% of the normal samples do not exceed 0.3, then , samples exceeding this value are judged as abnormal; among them, for normal samples, due to their dense distribution in historical data, more splits are required to be isolated, while abnormal samples are sparsely distributed and are often isolated within fewer splits;

[0096] Calculate the outliers of the process to be confirmed. When the outliers exceed the unknown threshold, it is determined to be beyond the boundary of the normal process feature space and marked as a high-confidence unknown process. Samples that significantly deviate from the normal process mode are screened to prepare for subsequent cross-modal verification. Otherwise, they are temporarily stored as low-risk anomalies and continuously monitored.

[0097] Based on the dynamic time warping algorithm and using weighted Euclidean distance, a weight of 0.8 was assigned to the defined key features, such as microwave energy and supercritical pressure, and a weight of 0.2 was assigned to the defined auxiliary features. The similarity between the unconfirmed process and the historical templates in the unconfirmed process library was calculated. The unconfirmed process library stores process feature templates that have not been defined for nearly a year. Each template contains a feature vector, expert annotation status, first detection timestamp, and equipment status. The expert annotation status includes unconfirmed, excluded, and to be learned.

[0098] If the similarity exceeds the preset similarity threshold, it indicates that a suspected similar process has been found in the process library to be confirmed, triggering the similarity review mechanism to ensure the accurate identification and knowledge reuse of similar processes. Otherwise, it indicates that it is a new unknown process or an extreme anomaly, triggering the deep intervention mechanism. Through interactive feature mining, it analyzes whether the anomaly conforms to the new process mechanism and confirms whether it is a new process based on multiple conditions.

[0099] For the similarity review mechanism, based on the expert annotation status of historical templates, unconfirmed or to-be-learned similar templates are prioritized for matching. K-means clustering is performed on similar templates to determine the cluster to which the current sample belongs. The historical defect rate of templates within the cluster is calculated, and a pre-classification report is generated. A review work order is then pushed to process experts, including a comparison of historical process curves for similar templates and the results of equipment status consistency verification, to facilitate expert decision-making and classify it as a new process or abnormal fluctuation.

[0100] For deep intervention mechanisms, the contribution of the process fingerprint vector of the process to be confirmed to the outlier is calculated, and a ranking diagram is generated. Based on counterfactual explanations, the impact of parameter adjustments on the outlier is simulated. Based on the contribution and physical location, the anomaly is analyzed to see whether it conforms to the new process mechanism. Based on the confirmation conditions set for the new process, it is classified as a new process or an abnormal fluctuation.

[0101] If it is an abnormal fluctuation, the current data is added to the noise sample set of the isolation forest, the model boundary is retrained, that is, the boundary of the normal process feature space, and an equipment maintenance work order is triggered. The full data is returned for storage;

[0102] If it is a new process, the core features are extracted through the algorithm, the initial rules are generated and injected into the rule base, and the historical data is re-labeled and the training sample set is expanded. The most similar known process model is selected through the K-nearest neighbor algorithm, and continuous monitoring is carried out for three consecutive cycles. If it is determined to be a new process, a new process confirmation signal is output to trigger the subsequent unknown defect pre-identification process. Otherwise, it returns to the deep intervention mechanism for re-evaluation. Among them, by calculating the similarity between the new process and the known process model, the known process model corresponding to the maximum similarity is defined as the most similar known process model.

[0103] Specifically, the unknown defect pre-identification mechanism includes:

[0104] Based on the multi-dimensional time series data set collected in real time, a three-dimensional coordinate system for the device is established, and the sensor detection values ​​are mapped to the physical location to complete the spatial coordinate mapping;

[0105] For each type of parameter, perform equal interval binning, count the number of samples in each interval, and generate a marginal histogram. At the same time, for any two types of parameters, that is, parameter pairs, calculate the joint histogram. Based on the mutual information formula, calculate the MI value of any parameter pair, and set the mutual information threshold. Parameter pairs with MI values ​​less than the mutual information threshold are eliminated to generate the final cross-modal association matrix.

[0106] At the same time, the weight allocation threshold is set, and the key parameters whose MI value is greater than the weight allocation threshold are assigned weights. , give weight to auxiliary parameters whose MI value is between the mutual information threshold and the weight allocation threshold ,Wherein, the weights are determined by the analytic hierarchy process, and process experts are invited to score the importance of the parameters;

[0107] The filtered parameter features are standardized and linearly combined according to the corresponding weights to generate a fusion feature vector;

[0108] Calculate the cosine similarity between the fused feature vector and the traditional defect templates in the known defect library. Keep samples with similarity less than the traditional similarity threshold and mark them as potential unknown defects. Keep them in the potential unknown defect set to eliminate interference from known defects.

[0109] Perform feature outlier analysis and abnormal sample screening to locate new defect clues. This includes using the forest model trained in the pre-response mechanism to calculate the outlier score for samples of potential unknown defects and set a defect outlier threshold. Once the calculated outlier score exceeds the defect outlier threshold, the sample is identified as a high outlier.

[0110] Use KD tree algorithm to search The nearest neighbors are found and the reachable distance is calculated, thereby calculating the local reachable density and LOF value. Once the LOF value is greater than the preset local outlier threshold, the sample is determined to be a local outlier.

[0111] The screened abnormal samples are stored in a structured manner according to industrial standards to generate a set of unknown defect candidates. Each candidate sample includes corresponding basic information, feature information, and evidence chain. Basic information includes sample ID, process type, and equipment status. Feature information includes process fingerprint vectors, fused feature vectors, and outlier detection results. The evidence chain includes original signal segments, spatiotemporal positioning, and correlation matrix.

[0112] Label candidate samples in the unknown defect candidate set, and classify them into high-risk, medium-risk, and low-risk levels by setting a secondary threshold for outlier risk. Files are named and stored according to the rules.

[0113] The screening threshold needs to be adjusted dynamically at different production stages to avoid overloading or missing detection of unknown defect candidate sets, so as to improve the effective sample rate and dynamically optimize the quality of unknown defect candidate sets; set the secondary thresholds for screening threshold adjustment, namely the first adjustment threshold and the second adjustment threshold, and adjust the threshold every Sample size in the unknown defect candidate set for period statistics , according to the sample size If the sample size is lower than the first adjustment threshold, the screening threshold is lowered and the reduction coefficient is set to If the sample size is between the first adjustment threshold and the second adjustment threshold, the original screening threshold is maintained; if the sample size is higher than the second adjustment threshold, the screening threshold is increased and the increase coefficient is set to ;

[0114] At the same time, the threshold is relaxed in the new process learning mode, and the relaxation coefficient is set to , tighten the threshold when stabilizing production, set the tightening threshold to , and record the threshold adjustment log to ensure the quality of the unknown defect candidate set.

[0115] Step S2: Construct a multi-level hierarchical detection network to perform adaptive detection on the unknown defect candidate set. To address the problem of false positives and negatives caused by insufficient sample size during the transition period, dynamic weight allocation is adopted. Initially, the primary layer rule detection is mainly used to quickly control the risk of obvious defects flowing in. As new samples accumulate, the weights of the intermediate and advanced layers are gradually increased to balance the detection of known defects and the exploration of unknown defects, avoiding false positives and negatives. At the same time, the process parameters and defect characteristics in the historical data are used to establish a benchmark model. Samples are screened based on Bayesian active learning. For unlabeled unknown defect candidate samples, pseudo-label technology is used to automatically generate soft labels. The training data is quickly expanded and the benchmark model is updated. Only the local parameters related to the new process in the model are updated to avoid catastrophic forgetting of the original defect recognition capability. When a small number of samples are added, the model can be iteratively optimized to form an unknown defect recognition model that adapts to the new process, effectively alleviating the impact of insufficient samples on detection accuracy.

[0116] Specifically, the steps of adaptive detection include:

[0117] Due to the small sample size in the early stages of a new process, a single model is prone to missing known defects or misjudging unknown defects. Therefore, a multi-level hierarchical detection network is constructed to achieve a dynamic balance between the detection of new and old defects. The network also outputs the judgment results and recommended response strategies, thereby generating a three-level detection result.

[0118] Since there are many unlabeled samples in the unknown defect candidate set and random labeling is inefficient, Bayesian active learning is used to screen samples. The process parameters and defect characteristics in the historical data are used to establish a benchmark model. The posterior distribution entropy of the benchmark model parameters is calculated as epistemic uncertainty. The MC Dropout method is used to calculate the prediction probability variance through 100 forward propagations as prediction uncertainty. The comprehensive score is calculated according to the epistemic uncertainty weight threshold and the prediction uncertainty weight threshold.

[0119] Sort by comprehensive score every week and select the top A sample, along with three-level detection results, spatiotemporal positioning evidence, and original signal fragments, is sent to experts for annotation;

[0120] Given the scarcity of labeled samples in the early stages of a new process, directly training the full benchmark model is prone to overfitting and may undermine the original detection capabilities. Therefore, pseudo-labeling technology and local parameter updates are used. For unlabeled samples, if the high-level detection confidence exceeds the preset label threshold and the mid-level label is consistent with the high-level result, a pseudo-label containing the defect type and confidence level is automatically generated.

[0121] The high-level neural network parameters are divided into a general layer and a new process-specific layer. The general layer is shared with known defect detection and has frozen parameters. The new process-specific layer is used to process cross-modal fusion features and allows parameter updates. Based on the hard-labeled samples annotated by experts and the soft labels of the pseudo-labeled samples, a hybrid loss function is set. The stochastic gradient descent algorithm is used to update only the parameters of the new process-specific layer. After the update, it is merged with the general layer parameters to generate a detection model adapted to the new process.

[0122] In order to balance risk control and production efficiency during the transition period of the new process, a risk classification response strategy was formulated. Based on the risk level and the three-level test results, for high-risk samples, if the primary layer determines that the defect is a known defect, the corresponding test strategy is triggered. If it is an unknown defect, the process end is directly shut down for inspection. For medium-risk samples, they are marked as key observation samples and the inspection frequency is increased. For low-risk samples, they are included in the dynamic observation pool and the feature change trend is analyzed daily. If there is no abnormality aggravation within the preset time, the inspection priority is lowered.

[0123] When the number of consecutive occurrences of high-risk defects at the same workstation reaches a threshold and the three-level detection results are consistent, the process parameters are automatically adjusted, the process fingerprint vector is recalculated, and the normal process feature space boundary is updated.

[0124] Specifically, the multi-level hierarchical detection network includes a primary layer, a middle layer, and a high-level layer:

[0125] In the early stages of a new process, the proportion of known defects is high and they need to be quickly intercepted to prevent them from flowing into the next process. At this time, the primary layer uses known process rules and known defect threshold rules to form a defect rule library. The Rete algorithm is used to optimize rule matching efficiency, convert feature vectors into fact nodes, and perform parallel matching according to rule priority, such as giving priority to high-risk defects. The weighted voting method is used to calculate the comprehensive confidence of rule matching. Once the comprehensive confidence exceeds the preset matching threshold, the match is determined to be successful. At this time, the defect type is output and the corresponding process adjustment strategy is triggered. Otherwise, it is determined to be a match failure and no match is output.

[0126] Given the lack of labeled samples in the early stages of a new process, the intermediate layer uses semi-supervised learning to expand the detection boundary, standardize the fused feature vector, eliminate the dimension effect, and construct an inter-sample association matrix based on the cosine similarity between samples. Similarities exceeding the preset value are retained as edges, and an undirected graph is constructed using the feature vectors of samples in the unknown defect candidate set as nodes.

[0127] The high-risk samples labeled by experts are used as seed nodes. The initial label vector is set and the iterative formula is used to update all node labels. The labels of neighboring samples are propagated to the current sample through the normalized similarity matrix while retaining the anchoring effect of the initial label vector. For unlabeled samples, the maximum value of the propagated label vector is taken as the confidence level, and the corresponding label vector is the soft label. The expression is as follows:

[0128]

[0129] Where, 、 For the Second and The node label vector at the iteration, is the propagation coefficient, which is used to balance the propagation effect of neighbor sample labels and the anchoring effect of the initial label vector. It is manually set according to the process scenario. is the similarity matrix, generated based on the cosine similarity calculated in the middle layer construction, is the degree matrix, which is calculated from the similarity matrix. The sum of each row element is used as the diagonal value, and the off-diagonal elements are all 0 to normalize the similarity matrix to avoid the propagation deviation caused by the difference in the number of sample neighbors. is the initial label vector;

[0130] Because the characteristics of unknown defects are complex and difficult to identify with traditional models, the advanced layer uses a deep neural network and a known process detection model. The first three convolutional layers are frozen, and two fully connected layers are added to process cross-modal fusion features and output defect confidence. A hybrid loss function is designed, combining the cross-entropy loss of expert-labeled samples and the KL divergence loss of pseudo-labeled samples. Only the parameters of the newly added fully connected layer are updated to improve the detection accuracy of unknown defects.

[0131] Considering the differences in detection requirements at different stages, different weights are assigned to each layer to achieve adaptive balance in the network, such as 70% for the primary layer, 20% for the intermediate layer, and 10% for the advanced layer, giving priority to intercepting known defects.

[0132] Each week, the proportion of correctly detected samples in each layer is normalized to obtain the proportional adjustment coefficient. The proportional adjustment coefficient is multiplied by the preset adjustment step size and added to the original weight for normalization to obtain the adjusted weight. At the same time, weight constraints are set to gradually shift from focusing on intercepting known defects to exploring unknown defects based on the accumulation of samples.

[0133] The advanced layer is based on the known process detection model, retains the first three convolutional layers to extract common features, adds two fully connected layers to process cross-modal fusion features, and outputs defect confidence;

[0134] At the same time, a dynamic weight allocation mechanism is set up to dynamically adjust the weights every week based on the proportion of correctly detected samples in each layer to ensure that detection resources are tilted towards the advantageous layer.

[0135] Step S3: Set three-dimensional quantitative indicators to judge the transition period. When all indicators meet the preset requirements, it automatically switches to the full model detection mode, marking the end of the transition period. This avoids the subjectivity of traditional methods that rely on manual experience and judgment. The quantitative indicators are used to objectively evaluate the detection system's ability to identify defects in the new process and its process adaptability.

[0136] Specifically, the steps for determining the transition period include:

[0137] Because the traditional reliance on manual experience to determine the end of the transition period is highly subjective and inefficient, a three-dimensional quantitative indicator system is constructed to objectively evaluate the detection system's adaptability to the new process. For the labeled samples in the unknown defect candidate set, the number of correctly identified samples in the multi-level hierarchical detection network is counted. Based on the total number of labeled samples, the detection accuracy is calculated. At the same time, the ratio of the standard deviation to the mean of the key process parameters collected during the transition period is calculated, which is defined as the parameter fluctuation coefficient to reflect parameter stability. The number of process-end adjustments triggered by the detection results is counted to evaluate the stability of the process parameters and the effectiveness of the detection strategy, measuring the degree of adaptation between the system and the process. The detection accuracy of the updated model for known defects is calculated using a known defect test set, and the update amplitude of the new process-specific layer parameters is counted to ensure that the model does not destroy the original detection capability when learning the new process. At this time, a three-dimensional indicator score table is output, including quantitative scores for detection accuracy, process adaptability, and model stability. Among them, correct identification requires that the detection result is consistent with the defect type annotated by the expert and the confidence level is higher than the set threshold, reflecting the system's ability to identify defects.

[0138] If any dimension fails to meet the standard, the multi-level hierarchical detection strategy will continue to be implemented, and the amount of active learning sample screening will be increased;

[0139] When the three-dimensional quantitative indicators simultaneously meet the preset indicator thresholds, the transition period end signal is triggered, the primary layer rule engine is deactivated, and only the emergency defect interception function is retained. The weights of the intermediate and advanced layers are merged, the detection layer weights and scanning frequencies are adjusted, and the system automatically switches to full model detection mode. This frees up the resources of the primary layer rule engine and focuses on in-depth detection of unknown defects, avoiding the lag of manual intervention. At the same time, a transition period end judgment report and a full model detection mode configuration file are output;

[0140] In the full model detection phase, the unknown defect candidate set is directly detected by the high-level deep neural network. The middle layer performs a secondary check on the confidence fuzzy samples and synchronizes the detection results to the process fingerprint vector generation step to update the normal process feature space boundary.

[0141] If continuous missed detection occurs, the Bayesian active learning of adaptive detection is triggered in reverse to supplement the training data;

[0142] Finally, the full model detection result report and process fingerprint vector update record are output to achieve continuous evolution of detection capabilities.

[0143] Example 2:

[0144] See also Figure 4 , another embodiment provided by the present invention: an online detection device for activated carbon molding defects, comprising: a data acquisition module, an identification module, and a transition detection module;

[0145] The data acquisition module is used to deploy a sensor matrix at key workstations in the activated carbon molding production line to collect process parameters in real time, pre-process the raw signals, achieve time alignment of multi-source signals through a time synchronization bus, and use edge computing nodes to perform sliding median filtering and Fourier transform denoising on the raw signals to generate a multi-dimensional time series data set with timestamps, providing basic data support for subsequent testing.

[0146] The identification module includes a process identification unit and a defect identification unit;

[0147] The process identification unit is used to calculate the similarity between the current parameter sequence and the historical process library through the dynamic time warping algorithm, identify process mode switching, extract time domain features, frequency domain features and causal features from the multi-dimensional time series data set, integrate and generate a process fingerprint vector, match the known process rule library through the rule engine, calculate the rule matching score, determine whether the current process is a known or unconfirmed process, initiate a pre-response mechanism for the unconfirmed process, encrypt and store the full data and build an isolation forest model, calculate the outlier score to screen high-confidence unknown processes, and simultaneously match the historical template based on the dynamic time warping algorithm to trigger the similarity review mechanism or deep intervention mechanism to complete the identification and confirmation of the new process;

[0148] The defect recognition unit is used to map multi-dimensional time series data to the equipment's three-dimensional coordinate system, calculate the MI value of parameter pairs to generate a cross-modal correlation matrix, generate a fused feature vector, filter known defects by comparing them with a known defect library, detect feature outlier samples using the isolation forest algorithm and the LOF algorithm, filter abnormal samples to generate a set of unknown defect candidates; dynamically adjust the screening threshold based on the sample size, optimize the quality of the unknown defect candidate set, label samples by risk level and store them in a structured manner, and provide target samples for transitional detection;

[0149] The transition detection module includes a layered detection unit and a transition management unit;

[0150] The hierarchical detection unit is used to build a multi-level hierarchical detection network consisting of primary, intermediate, and advanced layers. It adopts a dynamic weight allocation mechanism. Initially, rule-based detection is used to quickly intercept known defects. As samples accumulate, the weight of the deep network is improved. Based on Bayesian active learning, the sample uncertainty score is calculated, high-value samples are screened and labeled, and pseudo-labeling technology is used to update the local parameters of the model to form a detection model that adapts to new processes. This enables adaptive detection of unknown defects, and implements differentiated response strategies for samples based on risk levels. The process end is linked to adjust parameters and update process fingerprint vectors.

[0151] The transition management unit is used to set three-dimensional quantitative indicators of detection accuracy, process adaptability, and model stability, and quantitatively evaluate the adaptability of the detection system to the new process; when the indicators meet the preset indicator threshold, it automatically switches to the full model detection mode, disables the primary layer rule engine, and increases the weight of the advanced layer to determine the end of the transition period. The full model detection stage uses deep neural networks and semi-supervised models for collaborative detection, combined with a knowledge closed-loop mechanism to continuously optimize detection capabilities, and output process fingerprint vector update records and detection result reports.

[0152] Working principle and effect:

[0153] A sensor matrix is ​​deployed at key workstations on the production line to collect real-time data. After pre-processing, the data is compared with the historical process library using a dynamic time warping algorithm to identify process mode switching and trigger an unknown defect pre-identification mechanism. The normal process feature space boundary is constructed using the isolation forest algorithm, and the similarity is calculated using weighted Euclidean distance. Outlier samples are screened to form a position defect candidate set, and the quality of the candidate set is optimized using the KD tree algorithm and multi-level threshold adjustment. The constructed multi-level hierarchical detection network initially focuses on primary layer rule detection. As samples accumulate, the model is updated through Bayesian active learning and pseudo-labeling technology, and the weights of each layer are dynamically adjusted. The system adaptability is quantitatively evaluated using three-dimensional indicators: detection accuracy, process adaptability, and model stability. Once the standards are met, the system switches to full model detection mode.

[0154] Achieve self-perception recognition of new processes in seconds, eliminate the quality control vacuum caused by manual judgment lag, and significantly shorten the first identification time of new defects from hundreds of sample annotation cycles; through the combination of isolation forest and multi-level hierarchical detection network, complete the initial screening of unknown defects under zero-sample conditions, and accurately locate new defects such as microwave thermal stress cracks; the three-dimensional quantitative indicator system objectively determines the transition period, avoids human subjectivity, and improves judgment efficiency; Bayesian active learning and process parameter association mechanism only require a small number of samples to iterate the model, while reducing the defect rate during the transition period of the new process, ensuring the quality of activated carbon molding and performance stability.

[0155] The above description is merely a preferred embodiment of the present invention. The scope of protection of the present invention is not limited to the above embodiment. All technical solutions based on the concept of the present invention are within the scope of protection of the present invention. It should be noted that for those skilled in the art, various improvements and modifications that do not depart from the principles of the present invention should also be considered within the scope of protection of the present invention.

Claims

1. An online detection method for activated carbon molding defects, characterized in that: include: Real-time collection and preprocessing of process parameters to generate parameter sequences for self-perception identification of new processes. Once a new process is identified, the unknown defect pre-identification mechanism is immediately triggered to identify new anomalies that do not conform to known defects and generate a candidate set of unknown defects. Constructing a multi-level hierarchical detection network, performing adaptive detection on the unknown defect candidate set, establishing a benchmark model, screening samples and updating the benchmark model to form a defect recognition model adapted to the new process; Set three-dimensional quantitative indicators to judge the transition period. When all indicators meet the preset requirements, it automatically switches to the full model detection mode.

2. The method for online detection of activated carbon molding defects according to claim 1, characterized in that: The steps of self-perception identification include: Establish a multi-dimensional sensor matrix, collect and pre-process data, and generate a multi-dimensional time series data set; Set up a sliding window and calculate multi-dimensional key indicators within each window; Through fast Fourier transform, the dominant frequency is extracted and the energy proportion of each frequency band is calculated; For any two types of parameters, a bivariate VAR model is constructed, the Granger causal coefficient is calculated, and a causal matrix is ​​generated based on the causal threshold; The extracted features are integrated to generate feature vectors, and the process fingerprint vector is obtained by dimensionality reduction through principal component analysis; Define the necessary features and mutually exclusive conditions of known processes to form a production rule base. Use the rule engine to pre-screen the process fingerprint vector, calculate the rule matching score, set the matching threshold, and determine whether the current process is a known process. If the rule matching score is greater than the matching threshold, it is determined to be a known process; otherwise, it is determined to be a process to be confirmed, and a pre-response mechanism is triggered.

3. The method for online detection of activated carbon molding defects according to claim 2, characterized in that: Pre-response mechanisms include: For the process to be confirmed, the full amount of data in real time and historical periods is automatically collected, encrypted and saved, and an initial label is generated at the same time; The historical traditional process steady-state data is obtained as a training set, and the normal process feature space boundary is constructed based on the isolation forest algorithm; Calculate the outlier value of the process to be confirmed. If the outlier value exceeds the boundary of the normal process feature space, it is determined to be a high-confidence unknown process. Otherwise, it is temporarily stored as a low-risk anomaly and continuously monitored.

4. The method for online detection of activated carbon molding defects according to claim 3, characterized in that: Pre-response mechanisms also include: Calculating the similarity between the process to be confirmed and the historical templates in the process library to be confirmed using weighted Euclidean distance; If the similarity exceeds the preset similarity threshold, the similarity review mechanism is triggered, including: matching similar templates, determining the cluster to which the current sample belongs through K-means clustering, calculating the historical defect rate of templates within the cluster, generating a pre-classification report, and sending it to experts for review to classify it as a new process or abnormal fluctuation; Otherwise, a deep intervention mechanism is triggered, including: calculating the contribution of the process fingerprint vector to the outlier to generate a ranking diagram, explaining the impact of simulation parameter adjustment based on counterfactuals, analyzing whether the anomaly conforms to the new process mechanism, and classifying it as a new process or abnormal fluctuation based on the set confirmation conditions of the new process; If it is determined to be an abnormal fluctuation, the current data is added to the isolated forest noise sample set, the model boundary is retrained, and an equipment maintenance work order is triggered; If it is determined to be a new process, features are extracted to generate initial rules for injection into the rule base. Historical data is retraced and relabeled to expand the training set. The known process model corresponding to the maximum similarity is selected through the K-nearest neighbor algorithm. Continuous monitoring is performed for three consecutive cycles. If it is determined to be a new process, the new process signal is output. Otherwise, the deep intervention mechanism is returned for re-evaluation.

5. The method for online detection of activated carbon molding defects according to claim 4, characterized in that: The unknown defect pre-identification mechanism includes: Based on the multi-dimensional time series data set, a three-dimensional coordinate system for the device is established to map the sensor detection values ​​to physical locations; For each type of parameter, perform equal interval binning, count the number of samples in each interval, generate marginal histogram, and for parameter pairs, calculate joint histogram; Based on the mutual information formula, the MI value of any parameter pair is calculated, and a mutual information threshold is set, and the parameter pairs with MI values ​​less than the mutual information threshold are eliminated to generate a correlation matrix; Assigning different weights to the parameters in the association matrix based on a preset weight allocation threshold; The filtered parameter features are standardized and linearly combined according to the corresponding weights to generate a fusion feature vector; The cosine similarity between the fused feature vector and the traditional defect template in the known defect library is calculated, and samples with similarity less than the traditional similarity threshold are retained and marked as potential unknown defects.

6. The method for online detection of activated carbon molding defects according to claim 5, characterized in that: The unknown defect pre-identification mechanism also includes: Using the forest model, for the sample of the potential unknown defect, an outlier score is calculated, and a defect outlier threshold is set. Once the calculated outlier score exceeds the defect outlier threshold, it is determined to be a high outlier sample; Use the KD tree algorithm to calculate the LOF value. Once the LOF value is greater than the preset local outlier threshold, the sample is determined to be a local outlier. Constructing an unknown defect candidate set, marking candidate samples in the unknown defect candidate set, and classifying the candidate samples into high risk level, medium risk level, and low risk level by setting a secondary threshold of outlier risk; Set the secondary thresholds for adjusting the screening threshold, namely the first adjustment threshold and the second adjustment threshold. Time period, count the sample size in the unknown defect candidate set; If the sample size is lower than the first adjustment threshold, the screening threshold is lowered and the reduction factor is set to If the sample size is between the first adjustment threshold and the second adjustment threshold, the original screening threshold is maintained; if the sample size is higher than the second adjustment threshold, the screening threshold is increased and the increase coefficient is set to ; At the same time, in the new process learning mode, the relaxation coefficient is set to , set the tightening threshold to , and record the threshold adjustment log.

7. The method for online detection of activated carbon molding defects according to claim 6, characterized in that: The steps of the adaptive detection include: Construct a multi-level hierarchical detection network, including high-level, mid-level, and low-level layers, to generate three-level detection results; Establish a benchmark model using historical process parameters and defect characteristics, calculate the posterior distribution entropy and prediction probability variance of the parameters of the benchmark model, and calculate a comprehensive score through weighted summation; Each week, the top samples, and mark them; For unlabeled samples, if the high-level layer detection confidence exceeds the label threshold and the intermediate layer label is consistent with the high-level layer result, a pseudo label containing the defect type and confidence is generated; The high-level neural network parameters are divided into a frozen general layer and a new process-specific layer that is allowed to be updated. A hybrid loss function is set based on the soft labels annotated by experts. The stochastic gradient descent algorithm is used to update only the parameters of the new process-specific layer and then merge it with the general layer. Formulate a risk-graded response strategy based on the risk level and the three-level detection results; When the number of consecutive occurrences of high-risk samples at the same workstation reaches the threshold and the three-level detection results are consistent, the process-end parameters are automatically adjusted, the process fingerprint vector is recalculated, and the normal process feature space boundary is updated.

8. The method for online detection of activated carbon molding defects according to claim 7, characterized in that: The multi-level hierarchical detection network includes primary layer, intermediate layer and advanced layer: The primary layer uses the constructed defect rule base to calculate the comprehensive confidence; If the comprehensive confidence exceeds the matching threshold, the defect type is output; otherwise, no match is output; After the intermediate layer normalizes the fused feature vector, it constructs an undirected graph based on the cosine similarity between samples and sets the initial label vector; Update all node labels through an iterative formula, take the maximum value of the label vector after propagation for unlabeled samples as the confidence, and the corresponding label as the soft label; The advanced layer uses a deep neural network, freezes the first three convolutional layers based on the known process detection model, and adds two fully connected layers to process cross-modal fusion features and output defect confidence; Design a hybrid loss function that combines the cross entropy loss of expert-labeled samples and the KL divergence loss of pseudo-labeled samples, and only update the parameters of the newly added fully connected layer; A dynamic weight allocation mechanism is set up. The proportion adjustment coefficient is normalized based on the proportion of correctly detected samples in each layer every week. After multiplying it by the preset adjustment step size, it is added to the original weight and normalized to obtain the adjusted weight, and weight constraints are set.

9. The method for online detection of activated carbon molding defects according to claim 8, characterized in that: The steps for determining the transition period include: Construct a three-dimensional quantitative indicator system to calculate the detection accuracy, parameter fluctuation coefficient, number of process adjustments, model detection accuracy, and update amplitude, and output a three-dimensional indicator score table; If any dimension fails to meet the standard, continue to implement the multi-level hierarchical detection strategy and increase the amount of active learning sample screening; When the three-dimensional indicators simultaneously meet the preset indicator thresholds, the transition period end signal is triggered, the primary layer is deactivated, the weights of the intermediate and advanced layers are merged, and the full model detection mode is switched to. The high-level deep neural network directly detects the unknown defect candidate set, and the intermediate layer performs a secondary verification of the confidence fuzzy samples. The detection results are synchronized with the process fingerprint vector generation step to update the normal process feature space boundary; If continuous missed detection occurs, the Bayesian active learning of adaptive detection is triggered in reverse to supplement the training data, and the full model detection result report and process fingerprint vector update record are output.

10. An online detection device for activated carbon molding defects, which is used to implement the online detection method for activated carbon molding defects according to any one of claims 1 to 9, characterized in that: include: Data acquisition module, recognition module, transition detection module; The data acquisition module is used to collect process parameters in real time and perform preprocessing to generate a multi-dimensional time series data set; The identification module includes a process identification unit and a defect identification unit; The process identification unit is used to calculate the similarity between the current parameter sequence and the historical process library through a dynamic time warping algorithm to identify process mode switching; The defect recognition unit is used to perform mutual information analysis and feature fusion on the multi-dimensional time series data, filter out known defects, perform feature outlier analysis using the isolation forest algorithm and the LOF algorithm, screen abnormal samples to generate a set of unknown defect candidates, and dynamically adjust the screening threshold based on the production stage; The transition detection module includes a layered detection unit and a transition management unit; The hierarchical detection unit is used to construct a multi-level hierarchical detection network, adopt a dynamic weight allocation mechanism, combine Bayesian active learning and pseudo-labeling technology to update the baseline model, and perform adaptive detection of unknown defects; The transition management unit is used to set three-dimensional quantitative indicators, determine the end conditions of the transition period, and automatically switch to the full model detection mode.

Images