Method and device for on-line detection of activated carbon molding defects

Abstract

The application discloses an active carbon forming defect online detection method and device, and belongs to the technical field of material defect detection. The method comprises the following steps: collecting process parameters in real time and preprocessing, generating parameter sequences, and self-sensing knowledge identification of new processes, so that the lag start of quality control in the vacuum period is avoided. Once a new process is identified, the unknown defect pre-identification mechanism is triggered immediately, new abnormalities that do not conform to known defects are identified, and an unknown defect candidate set is generated. A multi-level hierarchical detection network is constructed, the unknown defect candidate set is adaptively detected, a benchmark model is established, samples are screened and the benchmark model is updated, a defect identification model suitable for the new process is formed, and the influence of insufficient samples on the detection accuracy is effectively alleviated. Three-dimensional quantitative indicators are set, the transition period is judged, and when the indicators all meet the preset requirements, the full model detection mode is automatically switched to. The quantitative indicators are used for objectively evaluating the identification ability of the detection system for new process defects and the process adaptability.

Description

Technical Field

[0001] This invention relates to a method and apparatus for online detection of defects in activated carbon forming, belonging to the field of material defect detection technology. Background Technology

[0002] The molding process of desulfurization and denitrification activated carbon involves multiple steps such as raw material mixing, extrusion, and drying. It is prone to 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, making it difficult to meet the quality control requirements of large-scale continuous production. With the development of industrial automation and intelligent manufacturing technologies, machine vision inspection has gradually become the mainstream means of industrial defect detection due to its advantages such as high precision, non-contact, and quantifiability.

[0003] However, existing technologies do not consider the unknown defect blind spots caused by the introduction of new processes during the molding of desulfurization and denitrification activated carbon. For example, if the current process flow changes and microwave drying or supercritical fluid activation, which are not present in the production line, are introduced, especially when the new process generates defect types that do not exist in the traditional process, such as radial microcracks caused by microwave thermal stress or pore collapse caused by supercritical fluid activation, the existing detection models rely entirely on training with historical defect samples and lack the ability to identify "new defect patterns" not included in the database, such as internal stress concentration caused by non-contact heating or abnormal pore structure caused by new activation processes. This requires manual re-labeling of samples and model iteration, which can easily lead to a quality control vacuum period during the transition period of introducing new processes. It is impossible to identify and intercept new defective products caused by fluctuations in process parameters in a timely manner, which can easily cause batches of unqualified products to flow into subsequent processes, affecting the molding quality and adsorption performance stability of desulfurization and denitrification activated carbon. In addition, existing technologies have not established a correlation prediction mechanism between process parameters and new defects, and cannot reduce the defect incidence rate through process control during model iteration, further exacerbating the quality control risks during the transition period. Summary of the Invention

[0004] To address the shortcomings of existing technologies, the present invention aims to provide an online detection method and device for activated carbon forming defects. By constructing a multi-level hierarchical detection network and combining dynamic weight allocation, Bayesian active learning, and pseudo-labeling technology, the invention solves the problems of insufficient samples, false alarms, and missed alarms during the transition period of new processes, and realizes the adaptive evolution and quantitative judgment of the detection system.

[0005] To achieve the above objectives, the present invention provides the following technical solution:

[0006] Online detection methods for activated carbon molding defects include:

[0007] Real-time acquisition and preprocessing of process parameters generate parameter sequences for self-sensing identification of new processes. Once a new process is identified, an 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] A multi-level hierarchical detection network is constructed to adaptively detect the unknown defect candidate set, establish a benchmark model, screen samples and update the benchmark model to form a defect identification model adapted to the new process.

[0009] Three-dimensional quantitative indicators are set to determine the transition period. When all indicators meet the preset requirements, the system automatically switches to the full model detection mode.

[0010] Specifically, the self-sensing recognition steps include:

[0011] Establish a multi-dimensional sensor matrix, collect and preprocess data, and generate a multi-dimensional time-series dataset.

[0012] Set up a sliding window and calculate multidimensional key indicators within each window;

[0013] The dominant frequency is extracted by fast Fourier transform, and the energy percentage of each frequency band is calculated.

[0014] For any two types of parameters, construct a bivariate VAR model, calculate the Granger causality coefficient, and generate a causality matrix based on the causality threshold.

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

[0016] Define the reproducible features and mutual exclusion 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, full data from real-time and historical periods are automatically collected, encrypted, and saved, while initial tags are generated.

[0020] Historical steady-state data of traditional processes were obtained as a training set, and the feature space boundary of normal processes was constructed based on the isolated forest algorithm.

[0021] Calculate the outlier values ​​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] The similarity between the process to be confirmed and historical templates in the process to be confirmed library is calculated using weighted Euclidean distance.

[0024] If the similarity exceeds the preset similarity threshold, a 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 pushing it to experts for review, so as 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 outliers to generate a ranking map, interpreting the impact of simulated parameter adjustments based on counterfactual explanations, analyzing whether the anomaly conforms to the new process mechanism, and classifying it as a new process or an abnormal fluctuation based on the set confirmation conditions for the new process.

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

[0027] If it is determined to be a new process, the core features are extracted to generate initial rules and inject them into the rule base. The historical data is re-labeled to expand the training set. The known process model corresponding to the maximum similarity is selected by the K-nearest neighbor algorithm. The monitoring is carried out for 3 consecutive cycles. If it is determined to be a new process in all three cycles, the new process signal is output. Otherwise, the deep intervention mechanism is returned for re-evaluation.

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

[0029] Based on the multi-dimensional time-series dataset, a three-dimensional coordinate system for the device is established, and the sensor detection values ​​are mapped to physical locations.

[0030] For each type of parameter, perform equal-interval binning, count the number of samples in each interval, generate an edge histogram, and for parameter pairs, calculate a joint histogram.

[0031] Based on the mutual information formula, calculate the MI value of any parameter pair, set a mutual information threshold, and remove parameter pairs whose MI value is less than the mutual information threshold to generate an association matrix;

[0032] Based on a preset weight allocation threshold, different weights are assigned to the parameters in the correlation matrix;

[0033] The filtered parameter features are standardized and linearly combined according to their corresponding weights to generate a fused feature vector.

[0034] The cosine similarity between the fused feature vector and the traditional defect template in the known defect library is calculated. 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 a forest model, outlier scores are calculated for the samples with potential unknown defects, 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] The KD tree algorithm is used 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] Construct an unknown defect candidate set, label the candidate samples in the unknown defect candidate set, and divide the candidate samples into high-risk, medium-risk, and low-risk levels by setting a secondary threshold for outlier risk;

[0039] Set two thresholds for adjusting the filtering threshold, namely the first adjustment threshold and the second adjustment threshold, every [percentage missing]. During a given time period, the sample size in the candidate set of unknown defects is calculated.

[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 and second adjustment thresholds, 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] Meanwhile, in the new process learning mode, a relaxation coefficient is set to , and a tightening threshold is set to during stable production, and a threshold adjustment log is recorded.

[0042] Specifically, the adaptive detection steps include:

[0043] A multi-level hierarchical detection network is constructed, including a high-level layer, a mid-level layer, and a low-level layer, to generate three levels of detection results;

[0044] A benchmark model is established using historical process parameters and defect characteristics. The posterior distribution entropy and prediction probability variance of the parameters of the benchmark model are calculated. A comprehensive score is calculated by weighted summation.

[0045] Each week, the top performers are selected based on their overall scores. Each sample was labeled.

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

[0047] The parameters of the high-level neural network 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 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 them with the general layer.

[0048] Based on the comprehensive risk level and the results of the three-level detection, a risk-based response strategy was formulated.

[0049] When a high-risk sample appears consecutively at the same workstation a number of times that threshold is reached and the results of the three-level detection are consistent, the process parameters are automatically adjusted, the process fingerprint vector is recalculated, and the boundary of the normal process feature space is updated.

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

[0051] A multi-level hierarchical detection network includes a primary layer, an intermediate layer, and a high-level layer:

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

[0053] If the overall confidence level exceeds the matching threshold, output the defect type; otherwise, output no match.

[0054] After standardizing the fused feature vectors in the intermediate layer, an undirected graph is constructed based on the cosine similarity between samples, and an initial label vector is set.

[0055] All node labels are updated using an iterative formula. For unlabeled samples, the maximum value of the label vector after propagation is taken as the confidence level, and the corresponding label is the soft label.

[0056] The advanced layer uses a deep neural network. Based on the known process detection model, the first three convolutional layers are frozen, and two new fully connected layers are added to process cross-modal fusion features and output defect confidence.

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

[0058] A dynamic weight allocation mechanism is set up. Each week, the proportion adjustment coefficient is obtained by normalizing the proportion of correctly detected samples in each layer. This coefficient is then multiplied by the preset adjustment step size, added to the original weight, and normalized to obtain the adjusted weight. Weight constraints are also set.

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

[0060] A three-dimensional quantitative indicator system is constructed to calculate the detection accuracy, parameter fluctuation coefficient, number of process adjustments, model detection accuracy, and update range, and output a three-dimensional indicator scoring table.

[0061] If any dimension fails to meet the standard, continue to implement the multi-level hierarchical detection strategy and increase the number of active learning samples screened.

[0062] When all three dimensions of indicators meet the preset threshold, the transition period ends, the primary layer is deactivated, the weights of the intermediate and advanced layers are merged, and the system switches to full model detection mode. The advanced layer deep neural network directly detects the unknown defect candidate set, the intermediate layer performs secondary verification on the confidence fuzzy samples, and the detection results are synchronized to the process fingerprint vector generation step to update the normal process feature space boundary.

[0063] If consecutive missed detections occur, the adaptive detection Bayesian active learning is triggered to supplement the training data, and a 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, perform preprocessing, and generate a multi-dimensional time-series dataset.

[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, and to identify process mode switching.

[0068] The defect identification unit is used to perform mutual information analysis and feature fusion on the multi-dimensional time series data, filter known defects, perform feature outlier analysis through the isolated forest algorithm and LOF algorithm, screen abnormal samples to generate a candidate set of unknown defects, 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, adopts a dynamic weight allocation mechanism, and combines Bayesian active learning and pseudo-label technology to update the baseline model for adaptive detection of unknown defects.

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

[0072] The beneficial effects of this invention are:

[0073] By utilizing a dynamic time warping algorithm, the system achieves second-level self-sensing identification of new processes, eliminating the lag in quality control vacuum periods. This allows the detection system to enter an early warning state at the initial stage of new defect generation, shortening the initial identification time of new defects from the hundreds of sample labeling cycles of traditional methods to a short time after process start-up. Through the combination of the isolated forest algorithm and a multi-level hierarchical detection network, it achieves initial screening of unknown defects under zero-sample conditions, preventing new defects from flowing into subsequent processes. The feature fusion and outlier analysis mechanism based on a three-dimensional coordinate system can accurately locate new defects such as radial microcracks caused by microwave thermal stress, which cannot be identified by traditional models. Simultaneously, a three-dimensional quantitative index system objectively evaluates the system's adaptability to new processes. When detection accuracy, process adaptability, and model stability meet the standards, the system automatically switches to full... The model-based detection mode avoids the subjectivity of human experience-based judgment and improves the efficiency of transition period judgment. Bayesian active learning and pseudo-label technology solve the model iteration problem in small sample scenarios. Only a small number of labeled samples are needed to update the model, avoiding damage to the original detection capabilities. At the same time, a correlation prediction mechanism between process parameters and defects is established through parameter fluctuation coefficients and causal matrices. During model iteration, the defect incidence rate can be reduced through process control. Experimental data shows that the defect incidence rate during the transition period of the new process is effectively reduced, effectively ensuring 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. Attached Figure Description

[0074] Figure 1 This is a schematic diagram of an online detection method for defects in activated carbon molding.

[0075] Figure 2 This is a flowchart of the self-sensing recognition process in this invention;

[0076] Figure 3 This is a flowchart of the unknown defect pre-identification mechanism in this invention;

[0077] Figure 4 This is a structural diagram of an online detection device for defects in activated carbon molding. Detailed Implementation

[0078] The technical solution of the present invention will be described in detail below with reference to 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 thereof. 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 in the figure, this embodiment introduces an online detection method for activated carbon molding defects, including the following steps:

[0081] Step S1: Deploy sensor matrices at key stations in the activated carbon forming production line, such as the extrusion outlet and the drying section inlet, to collect process parameters in real time, such as pressure, temperature, power, and flow rate. Preprocess the collected data to generate a continuous parameter sequence. Use a dynamic time warping algorithm to calculate the similarity between the current parameter sequence and the historical process library to perform self-sensing identification of new processes. This helps identify process mode switching and avoids delayed startup during quality control vacuum periods. It ensures that detection and early warning are initiated at the initial stage of new defects. Once a new process is identified, an unknown defect pre-identification mechanism is immediately triggered to identify new anomalies that do not conform to known defects. Unknown anomaly samples with outliers are screened from the production data stream to form a preliminary unknown defect candidate set. These samples are automatically marked and enter the review process, achieving initial screening of anomalies under zero-sample conditions. This shortens the first identification time of new defects, prevents unknown defects from directly flowing into subsequent processes, and provides an accurate target sample set for transition period detection.

[0082] In this embodiment, due to the difference in physical field between the new process and the traditional process, new defects are easily generated. The detection model relies on historical samples for training and lacks prior knowledge of defects not included in the database, thus forming a transition period. At the same time, the traditional supervised learning model cannot directly generalize to new defect types when facing zero-sample or small-sample scenarios. In the initial detection, it is easy to "misclassify normal products as unknown defects" or "miss new defects". It is necessary to gradually converge the accuracy through at least hundreds of labeled samples. 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-sensing recognition include:

[0084] A multi-dimensional sensor matrix is ​​established at key locations in the activated carbon forming production line. All sensors are precisely aligned in time via a time synchronization bus to ensure that the collected multi-source signals correspond strictly in the time dimension, avoiding feature analysis errors caused by asynchronous acquisition. This allows for the acquisition of raw data that comprehensively reflects the physical changes in the production process. Simultaneously, the collected data is preprocessed, and industrial edge computing nodes are used to perform real-time noise reduction on the raw sensor signals. Impulse noise is removed by sliding median filtering, and power frequency interference is filtered out based on Fourier transform. 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] The key sections include the extrusion molding section, the microwave drying section, and the supercritical activation section. In the extrusion molding section, a ring array pressure sensor is used to monitor the pressure distribution uniformity, and a laser rangefinder is used to measure the fluctuation of the preform diameter in real time to capture the spatial non-uniformity of the material's plastic deformation. In the microwave drying section, a three-dimensional temperature field is constructed using a distributed infrared thermal imager, 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, fiber optic strain gauges and density sensors are used to measure the pore deformation and activation efficiency.

[0086] To adapt to the dynamic changes in the process, a variable-length sliding window is applied to the multi-dimensional time-series dataset. 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 decay coefficient of the parameters, to obtain the characteristics of the parameters in the time series. Among them, 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; kurtosis is used to measure the steepness of the waveform, and the degree to which the signal deviates from the normal distribution is used to determine whether the waveform is abnormal; the autocorrelation decay 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 dataset to obtain the amplitude spectrum, then traverse the frequency domain amplitude spectrum to find the frequency with the largest amplitude. The frequency band is defined as the dominant frequency. The bandwidth is set based on the physical mechanism of the new process. The characteristic frequency band is pre-defined with the dominant frequency as the center. The ratio of the energy in the characteristic frequency band to the energy in the whole frequency band is calculated. In this way, the energy ratio of each frequency band is calculated 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 lag association between the parameters, and the Granger causality coefficient between the parameters is calculated to quantify the driving relationship between the parameters. A causal threshold is set, and once the Granger causality coefficient is greater than the causal threshold, there is a causal influence between the two parameters, thus generating a causal matrix.

[0089] The extracted features are integrated to generate a feature vector, and dimensionality is reduced by principal component analysis to reduce subsequent computational complexity, thereby generating a process fingerprint vector to extract composite information that can characterize process features.

[0090] Define the reproducible features and mutual exclusion conditions of known processes to form a production rule base. Use a rule engine to pre-screen the extracted process fingerprint vectors, count the number of matching rules that match the production rule base, 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, and set a matching threshold to determine whether the current process is a known process for preliminary screening.

[0091] If the rule matching score is greater than the matching threshold, it indicates that the current process matches a known process, and is therefore determined to be a known process. At this time, the rationality of the process parameters is verified and the corresponding detection strategy is initiated. Otherwise, the matching degree between the current process features and the known process rules is insufficient, and it is determined to be 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 process variations or extreme operating conditions. At the same time, the matching threshold is essentially a manually set probability threshold, and the boundary of the threshold is ambiguous. When the score is close to the threshold, insufficient matching may occur due to sensor noise and short-term process fluctuations. Directly judging it as a new process at this time will introduce a pseudo-new process, which is actually the edge state of a known process. Furthermore, the known process will have a gradual feature shift due to factors such as raw material batches and equipment aging, resulting in a rule matching score that is 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 the genuine innovation of process features, rather than data noise or feature extraction errors, to eliminate short-term interference and avoid misjudgment.

[0093] Pre-response mechanisms include:

[0094] For confirmed processes, both current and past processes are automatically saved. The full data within the time period, including raw sensor signals, preprocessed multi-dimensional time-series datasets, and process fingerprint vectors, is stored in HDF5 format with encrypted storage. The file name includes a process identification timestamp and a summary of key features. Initial labels are automatically generated, including suspected process type, anomalous feature summary, and equipment status.

[0095] Historical steady-state data of traditional processes were acquired. This data, manually annotated and defect-free, covers more than three equipment states, such as full load, half load, and startup phases, and is used as the training set. Based on the Isolation Forest algorithm, a normal process feature space boundary is constructed, including randomly selecting feature dimensions and segmentation values. Data is recursively segmented until a single sample is isolated or the maximum depth is reached, constructing 100 isolation trees to form a forest model. The average path length of each sample is calculated, i.e., the number of times the sample is isolated in the isolation tree. The path length is converted into an outlier using the outlier calculation formula. The 95th percentile is selected as the normal process feature space boundary. The 95th percentile is an unknown threshold. This represents the maximum outlier score among the top 5% of the normal samples. If 95% of the normal samples have an outlier score not exceeding 0.3, then... Samples exceeding this value are judged as abnormal. Among them, normal samples are densely distributed in historical data and require more splits to be isolated, while abnormal samples are sparsely distributed and are often isolated within fewer splits.

[0096] Calculate the outliers of the current process to be confirmed. When the outlier exceeds the unknown threshold, it is determined to exceed the boundary of the normal process feature space and is marked as a high-confidence unknown process. Samples that deviate significantly 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, using weighted Euclidean distance, key features defined by definition are assigned a weight of 0.8, such as microwave energy and supercritical pressure, while auxiliary features defined by definition are assigned a weight of 0.2. The similarity between the process to be confirmed and historical templates in the process to be confirmed library is calculated. The process to be confirmed library stores process feature templates that have not been defined in the past 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 completely new unknown process or an extreme anomaly, triggering the deep intervention mechanism, which analyzes whether the anomaly conforms to the mechanism of the new process through interactive feature mining, 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 learning-needed similar templates are matched first. K-means clustering is performed on the similar templates to determine the cluster to which the current sample belongs. The historical defect rate of templates within the cluster is calculated, a pre-classification report is generated, and a review work order is pushed to process experts. This includes a comparison of historical process curves of similar templates and the results of equipment status consistency verification, so as to facilitate expert decision-making and classify it as a new process or abnormal fluctuation.

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

[0101] If there is an abnormal fluctuation, add the current data to the noise sample set of the isolated forest, retrain the model boundary, which is the normal process feature space boundary, trigger the equipment maintenance work order, and return the full data for saving.

[0102] If it is a new process, the algorithm extracts the core features, generates initial rules and injects them into the rule base. At the same time, historical data is backtracked for re-labeling to expand the training sample set. The K-nearest neighbor algorithm is used to select the most similar known process model, and continuous monitoring is carried out for 3 consecutive cycles. If it is determined to be a new process in all three cycles, a new process confirmation signal is output to trigger the subsequent unknown defect pre-identification process. Otherwise, the process returns to the deep intervention mechanism for re-evaluation. In this process, the similarity between the new process and the known process model is calculated, and the known process model with the maximum similarity is defined as the most similar known process model.

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

[0104] Based on real-time acquired multi-dimensional time-series datasets, a three-dimensional coordinate system for the device is established, and sensor detection values ​​are mapped to physical locations to complete spatial coordinate mapping.

[0105] For each type of parameter, equal-interval binning is performed, the number of samples in each interval is counted, and an edge histogram is generated. At the same time, for any two types of parameters, that is, parameter pairs, a joint histogram is calculated. Based on the mutual information formula, the MI value of any parameter pair is calculated, and a mutual information threshold is set. Parameter pairs with MI values ​​less than the mutual information threshold are removed, thereby generating the final cross-modal association matrix.

[0106] Simultaneously, a weight allocation threshold is set, and key parameters with MI values ​​greater than the weight allocation threshold are assigned weights. Weights are assigned to auxiliary parameters whose MI values ​​fall between the mutual information threshold and the weight allocation threshold. The weights were determined using the analytic hierarchy process (AHP), and process experts were invited to score the importance of the parameters.

[0107] The filtered parameter features are standardized and linearly combined according to their corresponding weights to generate a fused feature vector.

[0108] The cosine similarity between the fused feature vector and the traditional defect template in the known defect library is calculated. Samples with similarity less than the traditional similarity threshold are retained and marked as potential unknown defects, and retained in the potential unknown defect set to eliminate interference from known defects.

[0109] Feature outlier analysis and abnormal sample screening are performed to locate novel defect clues. This includes using a forest model trained in the pre-response mechanism to calculate outlier scores for samples with potential unknown defects and setting defect outlier thresholds. Once the calculated outlier score exceeds the defect outlier threshold, it is judged as a high outlier sample.

[0110] Simultaneously use the KD tree algorithm for search The nearest neighbors are used to calculate the reachability distance, thereby calculating the local reachability 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 selected abnormal samples are stored in an industrial standard structure to generate an unknown defect candidate set. Each candidate sample includes corresponding basic information, feature information and evidence chain. The basic information includes sample ID, process type and equipment status. The feature information includes process fingerprint vector, fused feature vector and outlier detection results. The evidence chain includes original signal fragments, spatiotemporal positioning and correlation matrix.

[0112] Candidate samples in the unknown defect candidate set are labeled, and by setting a secondary threshold for outlier risk, the candidate samples are divided into high-risk, medium-risk, and low-risk levels. The files are named and stored according to the rules.

[0113] Different production stages require dynamic adjustment of screening thresholds to avoid overloading or missing detections in the candidate set of unknown defects, thereby improving the effective sample rate and dynamically optimizing the quality of the candidate set of unknown defects. Two secondary thresholds are set for adjusting the screening threshold: a first adjustment threshold and a second adjustment threshold, adjusted every [period]. Sample size in the candidate set of unknown defects during the time period According to the sample size The size of the sample size is adjusted to change the screening threshold; if the sample size is lower than the first adjustment threshold, the screening threshold is lowered, and the reduction factor is set to [value missing]. If the sample size is between the first and second adjustment thresholds, 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 [value missing]. ;

[0114] At the same time, the threshold is relaxed under the new process learning mode, and the relaxation coefficient is set to [value missing]. Tighten the threshold when production is stable, and set the tightening threshold to [value]. And record threshold adjustment logs to ensure the quality of the unknown defect candidate set.

[0115] Step S2: Construct a multi-level hierarchical detection network to adaptively detect unknown defect candidate sets. To address the false positives and false negatives caused by insufficient sample size during the transition period, dynamic weight allocation is adopted. Initially, the primary layer rule detection is the main method to quickly control the risk of obvious defects entering the network. 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 false negatives. At the same time, a baseline model is established using process parameters and defect features from historical data. Samples are screened based on Bayesian active learning. For unlabeled unknown defect candidate samples, soft labels are automatically generated using pseudo-labeling technology to quickly expand the training data and update the baseline model. Only local parameters related to the new process in the model are updated to avoid catastrophic forgetting of the original defect recognition capabilities. The model can be iteratively optimized with a small number of new samples to form an unknown defect recognition model adapted to the new process, effectively mitigating the impact of insufficient samples on detection accuracy.

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

[0117] Since the sample size is small in the early stage of the 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, and output judgment results and suggested response strategies to generate three-level detection results.

[0118] Since there are many unlabeled samples in the candidate set of unknown defects, random labeling is inefficient. Sample screening is carried out through Bayesian active learning. A benchmark model is established using process parameters and defect features in historical data. The posterior distribution entropy of the benchmark model parameters is calculated as cognitive uncertainty. The MC Dropout method is used to calculate the variance of the prediction probability through 100 forward propagations as prediction uncertainty. The comprehensive score is calculated according to the weight threshold of cognitive uncertainty and the weight threshold of prediction uncertainty.

[0119] Each week, the top-ranked teams are selected based on their overall scores. Each sample, along with the 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 the new process, directly training the full benchmark model is prone to overfitting and may damage the original detection capabilities. Therefore, pseudo-labeling technology and local parameter updates are adopted. For unlabeled samples, if the detection confidence of the high-level layer exceeds the preset label threshold and the label of the intermediate layer is consistent with the result of the high-level layer, a pseudo-label containing the defect type and confidence level is automatically generated.

[0121] The parameters of the high-level neural network are divided into a general layer and a new process-specific layer. The general layer shares the same parameters with known defect detection and the parameters are frozen. The new process-specific layer is used to process cross-modal fusion features and allows parameter updates. Based on hard-labeled samples with expert annotations and soft labels with 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, the parameters are merged with the general layer parameters to generate a detection model adapted to the new process.

[0122] To balance risk control and production efficiency during the transition period of the new process, a risk-based response strategy was developed. Based on the risk level and the results of the three levels of testing, for high-risk samples, if the primary layer determines that the defect is a known defect, the corresponding testing strategy is triggered; if the defect is an unknown defect, the process-side shutdown inspection is directly triggered. For medium-risk samples, they are marked as key observation samples, and the testing frequency is increased. For low-risk samples, they are included in a dynamic observation pool, and the trend of characteristic changes is analyzed daily. If there is no worsening of abnormalities within a preset time, the testing priority is reduced.

[0123] When a high-risk defect occurs consecutively at the same workstation a number of times that threshold is reached, and the results of the three-level detection are consistent, the process parameters are automatically adjusted, the process fingerprint vector is recalculated, and the boundary of the normal process feature space is updated.

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

[0125] In the early stages of a new process, known defects account for a high percentage and 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 base. The Rete algorithm is used to optimize the rule matching efficiency, converting feature vectors into fact nodes and performing parallel matching according to rule priority, such as prioritizing high-risk defects. A weighted voting method is used to calculate the overall confidence of rule matching. Once the overall confidence exceeds the preset matching threshold, the match is considered successful. At this time, the defect type is output and the corresponding process adjustment strategy is triggered. Otherwise, the match is considered unsuccessful and no match is output.

[0126] Given the insufficient labeled samples in the early stage of the new process, the intermediate layer uses semi-supervised learning to expand the detection boundary, standardizes the fused feature vector to eliminate the influence of dimensions, constructs the correlation matrix between samples based on the cosine similarity between samples, retains the similarity exceeding the preset value as edges, and constructs an undirected graph with the feature vectors of samples in the unknown defect candidate set as nodes.

[0127] High-risk samples labeled by experts serve as seed nodes, setting initial label vectors. All node labels are updated using an iterative formula. Labels from neighboring samples are propagated to the current sample via a normalized similarity matrix, while retaining the anchoring effect of the initial label vectors. 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 shown below:

[0128]

[0129] In the formula, , For the first Second and third The node label vector at the next iteration. This is the propagation coefficient, used to balance the propagation effect of neighboring sample labels with the anchoring effect of the initial label vector. It is manually set according to the process scenario. The similarity matrix is ​​generated based on the cosine similarity calculated during the construction of the intermediate layer. This is a degree matrix, calculated from the similarity matrix. The sum of elements in each row is used as the diagonal value, and the off-diagonal elements are all 0. This normalizes the similarity matrix and avoids propagation bias caused by differences in the number of sample neighbors. This is the initial label vector;

[0130] Because the characteristics of unknown defects are complex and difficult to identify by traditional models, the advanced layer uses a deep neural network. It utilizes a known process detection model, freezes the first three convolutional layers, adds two fully connected layers to process cross-modal fusion features, outputs defect confidence, and designs a hybrid loss function that combines the cross-entropy loss of expert-annotated samples and the KL divergence loss of pseudo-labeled samples. Only the parameters of the newly added fully connected layers are updated to improve the detection accuracy of unknown defects.

[0131] Considering the different detection requirements at different stages, different weights are assigned to each layer to achieve adaptive balance of the network, such as 70% for the primary layer, 20% for the intermediate layer, and 10% for the advanced layer, prioritizing the interception of known defects;

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

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

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

[0135] Step S3: Set three-dimensional quantitative indicators to determine the transition period. When all indicators meet the preset requirements, automatically switch to the full model detection mode to mark the end of the transition period. This avoids the subjectivity of traditional methods that rely on human experience to make judgments. The quantitative indicators 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 traditional methods of relying on human experience to determine the end of the transition period are prone to subjectivity and inefficiency, a three-dimensional quantitative indicator system is constructed to objectively evaluate the adaptability of the detection system to the new process. For 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 collected key process parameters during the transition period is calculated and defined as the parameter fluctuation coefficient to reflect parameter stability. The number of process-side adjustments triggered by the detection results is counted to evaluate the stability of process parameters and the effectiveness of the detection strategy, and to measure the degree of system and process compatibility. The detection accuracy of the updated model for known defects is calculated using a known defect test set, and the update range of the parameters of the new process-specific layer is counted to ensure that the model does not destroy the original detection capability when learning the new process. At this point, a three-dimensional indicator scoring 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 labeled by experts and the confidence level is higher than a set threshold to reflect the system's defect identification capability.

[0138] If any dimension fails to meet the standard, the multi-level hierarchical detection strategy will continue to be implemented, and the number of active learning samples will be increased.

[0139] When the three-dimensional quantitative indicators simultaneously meet the preset indicator thresholds, the transition period ends, the primary layer rule engine is disabled, only the emergency defect interception function is retained, the weights of the intermediate and advanced layers are merged, the detection layer weights and scanning frequency are adjusted, and the system automatically switches to the full model detection mode. This releases the resources of the primary layer rule engine, focuses on the in-depth detection of unknown defects, avoids the lag of manual intervention, and outputs a transition period end judgment report and a full model detection mode configuration file.

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

[0141] If consecutive missed detections occur, the adaptive detection Bayesian active learning is triggered to supplement the training data.

[0142] Finally, a full model detection result report and process fingerprint vector update records are output, enabling continuous evolution of detection capabilities.

[0143] Example 2:

[0144] Please see Figure 4 Another embodiment of the present invention provides 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 forming production line to collect process parameters in real time, preprocess 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 dataset with timestamps, providing basic data support for subsequent detection.

[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 database through the dynamic time warping algorithm, identify process mode switching, extract time domain features, frequency domain features and causal features from multi-dimensional time series datasets, integrate and generate process fingerprint vectors, match known process rule databases through rule engine, calculate rule matching scores, determine whether the current process is a known or unconfirmed process, start a pre-response mechanism for unconfirmed processes, encrypt and store the full amount of data and build an isolated forest model, calculate outlier scores to filter high-confidence unknown processes, and at the same time match historical templates based on the dynamic time warping algorithm to trigger similarity review mechanism or deep intervention mechanism to complete the identification and confirmation of new processes;

[0148] The defect identification unit maps multi-dimensional time-series data to the device's three-dimensional coordinate system, calculates the MI value of parameter pairs to generate a cross-modal correlation matrix, generates a fused feature vector, filters known defects by comparing with a known defect library, detects outlier samples using the isolated forest algorithm and the LOF algorithm, and filters abnormal samples to generate a candidate set of unknown defects. The unit dynamically adjusts the filtering threshold based on the sample size to optimize the quality of the candidate set of unknown defects, labels samples according to risk level and stores them in a structured manner to provide target samples for transitional period detection.

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

[0150] The hierarchical detection unit is used to construct a multi-level hierarchical detection network containing primary, intermediate, and advanced layers. It adopts a dynamic weight allocation mechanism. Initially, it uses rule-based detection to quickly intercept known defects. As the number of samples accumulates, the weight of the deep network is increased. Based on Bayesian active learning, it calculates the uncertainty score of the samples, selects high-value samples for labeling, and updates the local parameters of the model with pseudo-label technology to form a detection model that adapts to the new process. This enables adaptive detection of unknown defects and implements differentiated response strategies for samples according to risk level. It also links with the process end to adjust parameters and update the process fingerprint vector.

[0151] The transition management unit is used to set three-dimensional quantitative indicators: detection accuracy, process adaptability, and model stability, to quantitatively evaluate the adaptability of the detection system to the new process. When the indicators meet the preset threshold, it automatically switches to the full model detection mode, disables the primary layer rule engine, increases the weight of the advanced layer, and realizes the end of the transition period. In the full model detection stage, deep neural networks and semi-supervised models cooperate in detection, and the detection capability is continuously optimized by combining the knowledge closed loop mechanism, outputting process fingerprint vector update records and detection result reports.

[0152] Working principle and effects:

[0153] Sensor matrices are deployed at key workstations on the production line to collect real-time data. After preprocessing, the data is compared with historical process databases using a dynamic time warping algorithm to identify process mode switching and trigger an unknown defect pre-identification mechanism. The isolated forest algorithm is used to construct the spatial boundary of normal process features, and similarity is calculated by weighted Euclidean distance. Outliers are screened to form a candidate set of location defects, and the quality of the candidate set is then optimized by 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-label technology, and the weights of each layer are dynamically adjusted. The system's adaptability is quantitatively evaluated using three-dimensional indicators: detection accuracy, process adaptability, and model stability. Once the criteria are met, the system switches to full model detection mode.

[0154] Achieving second-level self-sensing recognition of new processes eliminates the quality control vacuum caused by the lag in manual judgment, significantly shortening the initial identification time of new defects from the labeling cycle of hundreds of samples; by combining isolated forests with multi-level hierarchical detection networks, it completes the initial screening of unknown defects under zero-sample conditions, accurately locating new defects such as microwave thermal stress cracks; a three-dimensional quantitative index system objectively judges the transition period, avoiding human subjectivity and improving judgment efficiency; a Bayesian active learning and process parameter association mechanism requires only a small number of samples to iterate the model, while reducing the defect incidence rate during the transition period of new processes, ensuring the quality and performance stability of activated carbon molding.

[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 embodiments. All technical solutions falling within the scope of the present invention's concept are within the scope of protection of the present invention. It should be noted that for those skilled in the art, any improvements and modifications made without departing 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 defects in activated carbon molding, characterized in that, include: Real-time acquisition and preprocessing of process parameters generate parameter sequences for self-sensing identification of new processes. Once a new process is identified, an 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. A multi-level hierarchical detection network is constructed to adaptively detect the unknown defect candidate set, establish a benchmark model, screen samples and update the benchmark model to form a defect identification model adapted to the new process. Set three-dimensional quantitative indicators to determine the transition period. When all indicators meet the preset requirements, automatically switch to the full model detection mode. The self-sensing recognition steps include: Establish a multi-dimensional sensor matrix, collect and preprocess data, and generate a multi-dimensional time-series dataset. Set up a sliding window and calculate multidimensional key indicators within each window; The dominant frequency is extracted by fast Fourier transform, and the energy percentage of each frequency band is calculated. For any two types of parameters, construct a bivariate VAR model, calculate the Granger causality coefficient, and generate a causality matrix based on the causality threshold. The extracted features are integrated to generate a feature vector, and the process fingerprint vector is obtained by dimensionality reduction through principal component analysis. Define the reproducible features and mutual exclusion 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. The multi-level hierarchical detection network includes a primary layer, an intermediate layer, and a high-level layer: The primary layer uses the constructed defect rule base to calculate the overall confidence score; if the overall confidence score exceeds the matching threshold, the defect type is output; otherwise, no match is output. After standardizing the fused feature vectors in the intermediate layer, an undirected graph is constructed based on the cosine similarity between samples, and an initial label vector is set. The labels of all nodes are updated through an iterative formula. For unlabeled samples, the maximum value of the label vector after propagation is taken as the confidence level, and the corresponding label is taken as the soft label. The advanced layer employs a deep neural network. Based on a known process detection model, the first three convolutional layers are frozen, and two new fully connected layers are added to process cross-modal fusion features and output defect confidence. A hybrid loss function combining expert-annotated sample cross-entropy loss and pseudo-labeled sample KL divergence loss is designed, updating only the parameters of the newly added fully connected layers. A dynamic weight allocation mechanism is set up, and each week, the proportion adjustment coefficient is normalized according to the proportion of correctly detected samples in each layer. This coefficient is multiplied by a preset adjustment step size, added to the original weight, and normalized to obtain the adjusted weight. Weight constraints are also set. The steps for determining the transition period include: A three-dimensional quantitative indicator system is constructed to calculate the detection accuracy, parameter fluctuation coefficient, number of process adjustments, model detection accuracy, and update range, and output a three-dimensional indicator scoring table. If any dimension fails to meet the standard, continue to implement the multi-level hierarchical detection strategy and increase the number of active learning samples screened. When all three dimensions of indicators meet the preset threshold, the transition period ends, the primary layer is deactivated, the weights of the intermediate and advanced layers are merged, and the system switches to full model detection mode. The advanced layer deep neural network directly detects the unknown defect candidate set, the intermediate layer performs secondary verification on the confidence fuzzy samples, and the detection results are synchronized to the process fingerprint vector generation step to update the normal process feature space boundary. If consecutive missed detections occur, the adaptive detection Bayesian active learning is triggered to supplement the training data, and a full model detection result report and process fingerprint vector update record are output.

2. The online detection method for activated carbon molding defects according to claim 1, characterized in that: Pre-response mechanisms include: For the process to be confirmed, full data from real-time and historical periods are automatically collected, encrypted, and saved, while initial tags are generated. Historical steady-state data of traditional processes were obtained as a training set, and the feature space boundary of normal processes was constructed based on the isolated forest algorithm. Calculate the outlier values ​​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.

3. The online detection method for activated carbon molding defects according to claim 2, characterized in that: The pre-response mechanism also includes: The similarity between the process to be confirmed and historical templates in the process to be confirmed library is calculated using weighted Euclidean distance. If the similarity exceeds the preset similarity threshold, a 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 pushing it to experts for review, so as 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 outliers to generate a ranking map, interpreting the impact of simulated parameter adjustments based on counterfactual explanations, analyzing whether the anomaly conforms to the new process mechanism, and classifying it as a new process or an abnormal fluctuation based on the set confirmation conditions for the new process. If the fluctuation is determined to be abnormal, the current data will be added to the isolated forest noise sample set, the model boundary will be retrained, and an equipment maintenance work order will be triggered. If it is determined to be a new process, features are extracted to generate initial rules and injected into the rule base. Historical data is backtracked and re-labeled to expand the training set. The known process model corresponding to the maximum similarity is selected by the K-nearest neighbor algorithm. The process is continuously monitored for 3 consecutive cycles. If it is determined to be a new process in all three cycles, a new process signal is output. Otherwise, the process is returned to the deep intervention mechanism for re-evaluation.

4. The online detection method for activated carbon molding defects according to claim 3, characterized in that: The unknown defect pre-identification mechanism includes: Based on the multi-dimensional time-series dataset, a three-dimensional coordinate system for the device is established, and the sensor detection values ​​are mapped to physical locations. For each type of parameter, perform equal-interval binning, count the number of samples in each interval, generate an edge histogram, and for parameter pairs, calculate a joint histogram. Based on the mutual information formula, calculate the MI value of any parameter pair, set a mutual information threshold, and remove parameter pairs whose MI value is less than the mutual information threshold to generate an association matrix; Based on a preset weight allocation threshold, different weights are assigned to the parameters in the correlation matrix; The filtered parameter features are standardized and linearly combined according to their corresponding weights to generate a fused feature vector. The cosine similarity between the fused feature vector and the traditional defect template in the known defect library is calculated. Samples with similarity less than the traditional similarity threshold are retained and marked as potential unknown defects.

5. The online detection method for activated carbon molding defects according to claim 4, characterized in that: The unknown defect pre-identification mechanism also includes: Using a forest model, outlier scores are calculated for the samples with potential unknown defects, 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. The KD tree algorithm is used 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. Construct an unknown defect candidate set, label the candidate samples in the unknown defect candidate set, and divide the candidate samples into high-risk, medium-risk, and low-risk levels by setting a secondary threshold for outlier risk; Set two thresholds for adjusting the filtering threshold, namely the first adjustment threshold and the second adjustment threshold, every [percentage missing]. During a given time period, the sample size in the candidate set of unknown defects is calculated. If the sample size is lower than the first adjustment threshold, then lower the screening threshold and set the reduction coefficient to [value missing]. If the sample size is between the first and second adjustment thresholds, 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 [value missing]. ; At the same time, a relaxation factor is set in the new process learning mode. When production is stable, set the tightening threshold to... And record the threshold adjustment log.

6. The online detection method for activated carbon molding defects according to claim 5, characterized in that: The adaptive detection steps include: A multi-level hierarchical detection network is constructed, including a high-level layer, a mid-level layer, and a low-level layer, to generate three levels of detection results; A benchmark model is established using historical process parameters and defect characteristics. The posterior distribution entropy and prediction probability variance of the parameters of the benchmark model are calculated. A comprehensive score is calculated by weighted summation. Each week, the top performers are selected based on their overall scores. Each sample was labeled. For unlabeled samples, if the confidence level of the high-level detection exceeds the label threshold and the label of the intermediate level is consistent with the result of the high-level level, a pseudo label containing the defect type and confidence level is generated. The parameters of the high-level neural network 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 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 them with the general layer. Based on the comprehensive risk level and the results of the three-level detection, a risk-based response strategy was formulated. When a high-risk sample appears consecutively at the same workstation a number of times that threshold is reached and the results of the three-level detection are consistent, the process parameters are automatically adjusted, the process fingerprint vector is recalculated, and the boundary of the normal process feature space is updated.

7. An online detection device for activated carbon molding defects, used to implement the online detection method for activated carbon molding defects as described in any one of claims 1-6, characterized in that, include: Data acquisition module, recognition module, and transition detection module; The data acquisition module is used to collect process parameters in real time, perform preprocessing, and generate a multi-dimensional time-series dataset. 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, and to identify process mode switching. The defect identification unit is used to perform mutual information analysis and feature fusion on the multi-dimensional time series data, filter known defects, perform feature outlier analysis through the isolated forest algorithm and LOF algorithm, screen abnormal samples to generate a candidate set of unknown defects, 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, adopts a dynamic weight allocation mechanism, and combines Bayesian active learning and pseudo-label technology to update the baseline model for adaptive detection of unknown defects. The transition management unit is used to set three-dimensional quantitative indicators, determine the conditions for the end of the transition period, and automatically switch to the full model detection mode.

Images